text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Parareal convergence for 2D unsteady
flow around a cylinder
arXiv:1509.04252v1 [cs.CE] 14 Sep 2015
Andreas Kreienbuehl1 , Arne Naegel2 , Daniel Ruprecht3 , Andreas Vogel2 ,
Gabriel Wittum2 , and Rolf Krause1
Abstract In this technical report we study the convergence of Parareal
for 2D incompressible flow around a cylinder for different viscosities. Two
methods are used as fine integrator: backward Euler and a fractional step
method. It is found that Parareal converges better for the implicit Euler, likely
because it under-resolves the fine-scale dynamics as a result of numerical
diffusion.
Key words: Parallel-in-time integration, Parareal, Navier-Stokes equations
1 Introduction
The potential of parallel-in-time integration methods to increase the degree
of concurrency in the numerical solution of time-dependent partial differential equations has been widely acknowledged, e.g. in the report Applied
Mathematics Research for Exascale Computing by Dongarra et al. [2014]. A
variety of different methods exists, see e.g. the review by Gander [2015], and
principle efficiency of parallel-in-time integration in large- and extreme-scale
parallel computations has been demonstrated, e.g. in Speck et al. [2012],
Ruprecht et al. [2013], and Gander and Neumueller [2014].
Many problems in computational fluid dynamics require massive computational capacities and suffer from long solution times. Exploring the potential of parallel-in-time methods to speed up such simulations can therefore
be a beneficial endeavour. The performance of the parallel-in-time integration method Parareal, introduced by Lions et al. [2001], when applied to the
1 Institute
of Computational Science, Faculty of Informatics, Università della Svizzera italiana, Lugano, Switzerland · 2 Goethe-Center for Scientific Computing, Goethe-University
Frankfurt am Main, Kettenhofweg 139, DE-60325 Frankfurt am Main, Germany · 3 School
of Mechanical Engineering, University of Leeds, Woodhouse Lane, Leeds LS2 9JT, UK.
1
2
Kreienbuehl et al.
Navier-Stokes equations has been a topic of research since shortly after its introduction. First studies have been conducted by Trindade and Pereira [2004,
2006], including reports of speedup for an MPI implementation; laminar flow
around a cylinder is used as benchmark problem and it is shown that Parareal
can correctly reproduce the Nusselt number. Fischer et al. [2005] investigate
Parareal for spatial discretisations based on finite and spectral element methods, and discuss using fewer spatial degrees-of-freedom for the coarse integrator. The performance of Parareal for simulations of non-Newtonian fluids has
been investigated by Celledoni and Kvamsdal [2009]. Finally, parallel scaling
of Parareal for 3D unsteady flow is investigated by Croce et al. [2014] on up
to 2048 cores.
Based on predictions from linear stability analysis by Gander and Vandewalle [2007], it has been shown by Steiner et al. [2015]) that convergence of
Parareal deteriorates as the Reynolds number increases. However, the studies only analysed a rather simple driven cavity problem, which eventually
approaches a steady-state and thus may underestimate the problem because
of weak transient dynamics towards the end of the simulation. In this report,
we continue this investigation for a different, more complex benchmark involving unsteady flow around a cylinder. It was introduced by Schäfer et al.
[1996] as the case 2D-3 and further analyzed by e.g. John [2004]. Eventually, the here presented benchmarks will be extended to a comprehensive
exploration of Parareal’s performance for 3D flow, including a study of the
influence of spatial resolution.
2 Parareal and model problem
2.1 Parareal
Parareal parallelises the solution of initial value problems
ut = f (t, u(t)),
u(0) = u0 ,
t ∈ [0, T ],
(1)
by decomposing the time domain [0, T ] into time slices [tj−1 , tj ], j =
1, . . . , Npr with Npr equal to the number of processing units. It then iterates between two time integration methods: a coarse integrator C used to
serially propagate corrections, which has to be computationally cheap, and
an accurate integrator F run in parallel. The Parareal iteration reads
k+1
uk+1
) + F(ukj ) − C(ukj ),
j+1 = C(uj
j = 1, . . . , Npr ,
(2)
with k being the iteration index and uj ≈ u(tj ). Note how the computationally expensive computation of the fine method can be done concurrently
Parareal convergence for 2D unsteady flow around a cylinder
3
for all time slices. A detailed presentation including a theoretical model for
projected speedup is given e.g. by Minion [2010].
2.2 Model problem
As model problem, we consider the Navier-Stokes equations
(∂t + U ∇)U = −
∇p
+ ν∇2 U
%
(3)
for an incompressible fluid, i.e. for a fluid with ∇U = 0, at density % :=
1 (kg/m2 ). We focus on the benchmark problem defined in Schäfer et al.
[1996] as 2D-3, which is for unsteady flow around a cylinder in two dimensions,
i.e. in 2D (see also John [2004]). We make use of the definitions
∇ := (∂x , ∂y )T ,
U := U (t, x, y),
U := (u, v)T .
(4)
The Reynolds number for a cylinder with diameter d (the reference length)
located inside a square cuboid with longest edge along the x-coordinate is
d
Rd := ūin ,
ν
(5)
where the reference velocity ūin is chosen to be the mean velocity of inflow
in x-direction and ν the kinematic viscosity. Notice that Rd can be time
dependent, as is the case for the problem considered here.
In the 2D case the mean velocity is
Z h
u(t, 0, y)
2
h
ū(t) :=
dy = u t, 0,
,
v̄(t) := 0,
(6)
h
3
2
0
where h := 0.41 (m). In the 2D-3 case we have the inflow velocity
uin (t, 0, y) := 4uxin sin
π y(h − y)
t
,
8
h2
vin (t, 0, y) := 0,
for which Equation (6) is valid. We choose uxin := 3/2 (m/s) so that
π d d
t
∈ 0,
.
Rd (t) = sin
8 ν
ν
(7)
(8)
Thus, setting d := 0.1 (m), it follows for ν ∈ {0.1, 0.01, 0.001} (m/s) that
Rdx ∈ {1, 10, 100}
defines the maximum-over-time Reynolds number for the chosen ν.
(9)
4
Kreienbuehl et al.
In this report, the primary goal is to outline the performance of Parareal
for the 2D-3 benchmark problem for the three mentioned viscosities, i.e.
ranges of Reynolds numbers.
2.3 Implementation details
The governing equations were implemented using Q2-Q1 finite elements in
the UG4 software toolbox (see Vogel et al. [2013], Vogel [2014]). For the parallelisation in time via Parareal, we used the library Lib4PrM, which was first
applied in Kreienbuehl et al. [2015].1
3 Results
Instead of measuring convergence of Parareal by comparing the discretisation
error with the defect as discussed e.g. by Arteaga et al. [2015], we focus here
on how well Parareal reproduces important characteristic numbers of the
dynamics, namely the drag coefficient Cdr , the lift coefficient Cli and the
pressure difference ∆p between the front and end point of the cylinder over
time. These parameters are respectively defined as follows:
Cdr :=
2Fdr
,
%ū2in d
Cli :=
2Fli
,
%ū2in d
∆p := pfr − pen ,
(10)
where Fdr is the drag force and Fli the lift force, and pfr together with pen
define the pressure at the front and end of the cylinder. Again, we assume
here that the density is 1 (kg/m2 ) and set t ∈ [0, 8] (s) as time domain.
For ν = 0.001 (m/s), Schäfer et al. [1996] report on a maximum-overma
time drag coefficient of Cdr
≈ 2.9500 ± 0.0200, a maximum-over-time lift
ma
coefficient of Cli ≈ 0.4800±0.0100, and a final pressure difference at t = 8 (s)
of ∆fip ≈ −0.1100 ± 0.0050 (kg/s2 ).
3.1 Numerical setup
We use Npr ∈ {2, 4, 8, 16} processors without parallelization in space and
with 13,212 spatial degrees-of-freedom for both the fine and coarse level. For
each ν, we consider the following two Parareal solvers “S” comprised of a
coarse Mco and fine Mfi serial time integration method as well as number of
1
It can be obtained by cloning the Git repository https://scm.ti-edu.ch/repogit/
lib4prm.
Parareal convergence for 2D unsteady flow around a cylinder
5
time steps Nco and Nfi :
(Mco , Nco ) × (Mfi , Nfi ) = (IE, 16) × (IE, 32),
(S1 )
(Mco , Nco ) × (Mfi , Nfi ) = (IE, 16) × (FS, 32),
(S2 )
where “IE” stands for implicit Euler (first-order) and “FS” for fractional
step (second-order). Errors in the three physical quantities discussed above
are measured by
fi
kupa
ph − uph kti
:=
,
(11)
Eph
fi
kuph kti
where the parameter “ph” is in {dr,li,p} for drag and lift coefficient or pressure
difference. We use the l2 -norm over the solutions at the end of all time-slices
kuk2ti :=
Npr
8 X
|u(tn )|2
Npr n=1
(12)
weighted by the time slice length 8/Npr .
3.2 Problem dynamics
Figure 1 shows the flow field at t = 5.25 (s) for the three different viscosities.
As viscosity decreases, the maximum Reynolds number increases and the flow
becomes more turbulent. While for ν = 0.1 (m/s) and ν = 0.01 (m/s) the
flow is essentially laminar, smaller vortices start to form behind the cylinder
for ν = 0.001 (m/s). The resulting evolution over time of the three characteristic numbers for ν = 0.1 (m/s) using serial time integration is shown in
Figure 2. Both fine integrators S1 and S2 produce essentially identical profiles and their profiles closely match the one generated by the corresponding
reference simulation (not shown). Figure 3 shows the same profiles for the
simulation with ν = 0.01 (m/s). Again, both S1 and S2 produce profiles
that match and agree with the results from the corresponding reference (not
shown). Lastly, Figure 4 shows three profiles for ν = 0.001 (m/s), one for the
reference simulation, one for S1 and one for S2 . While the drag coefficient
and pressure difference agree across all three configurations, S1 produces a
lift coefficient profile that is distinctly different from the corresponding reference and S2 . The relatively high numerical diffusion of S1 in combination
with a rather low spatial resolution probably prevents S1 from correctly capturing the more turbulent dynamics in this case. In contrast, although S2
fails to fully reproduce the frequency of oscillations, it still achieves a qualitatively correct representation of the dynamics of the corresponding reference
simulation.
6
Kreienbuehl et al.
Fig. 1: Flow field at t = 5.25 (s) for three different viscosities: ν = 0.1 (m/s)
at the top, ν = 0.01 (m/s) in the middle, and ν = 0.001 (m/s) at the bottom.
3.3 Convergence of Parareal
Here, we analyse how accurately Parareal reproduces the three characteristic values studied above. Figure 5 shows the defect or error according to
Equation (11) in the characteristic values accumulated over all time slices
versus the number of iterations. Here, defect refers to the difference between
the solution computed by Parareal and the solution computed by running
the fine integrator serially. For S1 , the error for all three quantities, i.e. drag
coefficient, lift coefficient and pressure difference, quickly goes to zero, that
is Parareal rapidly produces values identical to ones obtained from the serial
simulation. As the number of time slices is increased, convergence becomes
slower but the increase is not drastic: for Npr = 16 after seven iterations all
three characteristic values have converged up to round-off error. In a production run, where the main goal is to push the defect from Parareal below the
Parareal convergence for 2D unsteady flow around a cylinder
7
25
Pressure difference ∆ p pkg{s2 q
1
Lift coefficient Cli
Drag coefficient Cdr
60
40
20
0
0.5
2
4
6
8
15
10
5
0
0
0
20
0
Time t psq
2
4
6
8
0
Time t psq
2
4
6
8
Time t psq
25
Pressure difference ∆ p pkg{s2 q
1
Lift coefficient Cli
Drag coefficient Cdr
60
40
20
0
0.5
2
4
6
Time t psq
8
15
10
5
0
0
0
20
0
2
4
6
Time t psq
8
0
2
4
6
8
Time t psq
Fig. 2: Drag and lift coefficient, and pressure difference for ν = 0.1 (m/s) for
S1 (top) and S2 (bottom). The curves match the values from the reference
simulation (not shown).
discretisation error (see the discussion in Arteaga et al. [2015]), significantly
fewer iterations will likely suffice.
Decreasing viscosity and thus increasing the Reynolds number range does
negatively affect convergence. For ν = 0.001 (m/s), the simulation with Npr =
16 time slices already requires 13 iterations to converge up to round-off error.
Depending on the desired accuracy, speedup is still possible here but parallel
efficiency will likely be lower than in the more laminar case. Since S1 fails
to resolve the full dynamics of the problem, its convergence behaviour is
probably not representative of the actual physical dynamics.
This is supported by the fact that for S2 and ν = 0.001 (m/s), Parareal
essentially no longer converges. Since S2 does resolve the turbulent dynamics
8
Kreienbuehl et al.
¨10´2
4
6
4
2
Pressure difference ∆ p pkg{s2 q
8
Lift coefficient Cli
Drag coefficient Cdr
8
6
4
2
0
3
2
1
0
0
0
2
4
6
8
0
Time t psq
2
4
6
0
8
2
4
6
8
Time t psq
Time t psq
¨10´2
4
6
4
2
Pressure difference ∆ p pkg{s2 q
8
Lift coefficient Cli
Drag coefficient Cdr
8
6
4
2
0
3
2
1
0
0
0
2
4
6
Time t psq
8
0
2
4
6
Time t psq
8
0
2
4
6
8
Time t psq
Fig. 3: Drag and lift coefficient, and pressure difference for ν = 0.01 (m/s)
for reference (top), S1 (middle) and S2 (bottom).
at least partially, in contrast to S1 , this suggests that the good convergence
of Parareal for S1 and ν = 0.001 (m/s) is an artefact produced by excessive
numerical diffusion. The dynamics of the numerical solution are more laminar than they should be, leading to an unrealistic convergence behaviour
of Parareeal. Supposedly, when using S1 on a significantly finer spatial and
temporal mesh, a similar deterioration of convergence would be observed, as
the numerical solution better resolves the turbulent features of the flow.
Interestingly, this difference between S1 and S2 can already be seen for
ν = 0.1 (m/s), where the physical dynamics are still quite laminar as well.
Although Parareal for S2 does converge, particularly for larger numbers of
Parareal convergence for 2D unsteady flow around a cylinder
2.5
Lift coefficient Cli
2
1
0
Pressure difference ∆ p pkg{s2 q
0.4
3
Drag coefficient Cdr
9
0.2
0
´0.2
´0.4
0
2
4
6
0
8
2
4
6
1.5
1
0.5
0
8
0
Time t psq
¨10´2
Time t psq
2
4
6
8
Time t psq
2.5
3
2
1
Pressure difference ∆ p pkg{s2 q
Lift coefficient Cli
0
Drag coefficient Cdr
2
´0.5
´1
0
2
1.5
1
0.5
0
´1.5
0
2
4
6
8
0
Time t psq
4
6
0
8
2
4
6
8
Time t psq
2.5
3
2
1
Pressure difference ∆ p pkg{s2 q
0.2
Lift coefficient Cli
Drag coefficient Cdr
2
Time t psq
0.1
0
´0.1
0
0
2
4
6
Time t psq
8
0
2
4
6
Time t psq
8
2
1.5
1
0.5
0
0
2
4
6
8
Time t psq
Fig. 4: Drag and lift coefficient, and pressure difference for ν = 0.001 (m/s)
for reference (top), S1 (middle) and S2 (bottom).
10
Kreienbuehl et al.
“2
“4
“8
“ 16
10´6
“2
“4
“8
“ 16
10´8
10´9
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´10
10´11
10´12
10´13
10´14
10´14
1 4 7 10 13 16
1 4 7 10 13 16
1 4 7 10 13 16
Iterations Nit
Iterations Nit
Iterations Nit
“2
“4
“8
“ 16
10´6
10´9
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´8
10´5
Error E p
Npr
Npr
Npr
Npr
10´4
Error Eli
10´6
10´10
10´12
10´15
10´3
Error E p
10´11
10´13
Error Edr
Npr
Npr
Npr
Npr
10´8
Error Eli
Error Edr
10´9
Npr
Npr
Npr
Npr
10´10
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´8
10´11
10´12
10´12
10´14
10´14
10´15
1 4 7 10 13 16
1 4 7 10 13 16
1 4 7 10 13 16
Iterations Nit
Iterations Nit
Iterations Nit
Fig. 5: Fine solver convergence errors for drag and lift coefficient, and pressure
difference for ν = 0.1 (m/s) and S1 (top) and S2 (bottom).
time slices, its rate of convergence is lower than for S1 . The benefit of using an
integrator with damping properties as coarse integrator has been pointed out
before by Bal [2005] but apparently numerical diffusion from the fine method
does help Parareal convergence, too. Since here the fine method realistically
represents the flow features, using a diffusive integrator as fine method in
Parareal when simulating turbulent flow could be an easy way to obtain
decent convergence.
Parareal convergence for 2D unsteady flow around a cylinder
10´5
10´7
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´3
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´5
10´7
10´11
“2
“4
“8
“ 16
10´9
10´12
10´13
10´15
10´15
1 4 7 10 13 16
1 4 7 10 13 16
Iterations Nit
Iterations Nit
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´3
Npr
Npr
Npr
Npr
10´15
“2
“4
“8
“ 16
10´3
10´9
10´12
10´12
10´15
10´15
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´6
Error E p
Error Eli
10´9
1 4 7 10 13 16
Iterations Nit
10´6
10´6
Error Edr
10´9
Npr
Npr
Npr
Npr
10´11
10´13
10´3
Error E p
Error Eli
Error Edr
10´6
10´9
11
10´9
10´12
1 4 7 10 13 16
1 4 7 10 13 16
Iterations Nit
Iterations Nit
10´15
1 4 7 10 13 16
Iterations Nit
Fig. 6: Fine solver convergence errors for drag and lift coefficient, and pressure
difference for ν = 0.01 (m/s) for S1 (top) and S2 (bottom).
4 Conclusions
This report extends the investigation by Steiner et al. [2015] about how a decreasing viscosity in the Navier-Stokes equations affects the convergence behaviour of the Parareal parallel-in-time method. An unsteady 2D flow around
a cylinder is used as a benchmark. Two different configurations are tested,
one using an implicit Euler as coarse and fine method, the other an implicit
Euler as coarse but a second-order fractional step integrator as fine method.
For larger viscosities, both base methods correctly reproduce the evolution of
characteristic quantities like drag and lift coefficient and pressure difference.
However, the numerical diffusion from the backward Euler method as fine
integrator leads to significantly better convergence of Parareal.
12
Kreienbuehl et al.
10´2
10´5
10´9
10´11
10´15
10´14
1 4 7 10 13 16
Iterations Nit
Iterations Nit
“2
“4
“8
“ 16
10´9
10´15
1 4 7 10 13 16
Iterations Nit
10´3
10´2
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
Error Eli
Error Edr
10´6
Npr
Npr
Npr
Npr
10´12
1 4 7 10 13 16
10´3
10´9
10´3
“2
“4
“8
“ 16
10´8
10´12
10´6
Npr
Npr
Npr
Npr
Error E p
“2
“4
“8
“ 16
10´6
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´6
Error E p
Error Edr
10´6
Npr
Npr
Npr
Npr
Error Eli
10´3
10´9
Npr
Npr
Npr
Npr
“2
“4
“8
“ 16
10´10
10´12
10´12
10´14
10´15
1 4 7 10 13 16
1 4 7 10 13 16
Iterations Nit
Iterations Nit
10´15
1 4 7 10 13 16
Iterations Nit
Fig. 7: Fine solver convergence errors for drag and lift coefficient, and pressure
difference for ν = 0.001 (m/s) for S1 (top) and S2 (bottom).
As viscosity decreases, convergence of Parareal becomes slower similarly
to the results in Steiner et al. [2015]. However, while for the fractional step
method Parareal stalls completely, the backward Euler retains reasonable
convergence even for small viscosities. A comparison of the profiles for characteristic numbers with a reference solution suggests that the good convergence for the backward Euler is likely artificial. The fine integrator alone
captures the relatively smooth profiles of the drag coefficient and pressure
difference quite well. However, the high frequency oscillations in the lift coefficient, which are clearly seen in serial runs using the fractional step integrator, are not present when using backward Euler. Most likely, the rather
high numerical diffusion leads to an artificially laminar flow, so that the good
convergence of Parareal is not representative of the used viscosity parameter.
Parareal convergence for 2D unsteady flow around a cylinder
13
Clearly, when assessing Parareal’s convergence for flow problems, care must
be taken to ensure that the fine base method correctly resolves the important
features of the flow. Intrinsic convergence of Parareal alone is not a reliable
indicator.
There are several works proposing strategies to stabilise Parareal for
advection-dominated problems, e.g. by Farhat et al. [2006], Gander and Petcu
[2008], Ruprecht and Krause [2012] or Chen et al. [2014]. An interesting
continuation of the work presented here would be to analyse whether these
strategies improve convergence of Parareal for turbulent flows.
Acknowledgments
We would like to thank Ernesto Casartelli and Luca Mangani from the
Lucerne University of Applied Sciences and Arts (HSLU) in Switzerland for
discussions.
The research of A.K., D.R., and R.K. is funded through the “FUtuRe
SwIss Electrical InfraStructure” (FURIES) project of the Swiss Competence
Centers for Energy Research (SCCER) at the Commission for Technology and
Innovation (CTI) in Switzerland. The research is also funded by the Deutsche
Forschungsgemeinschaft (DFG) as part of the “ExaSolvers” project in the
Priority Programme 1648 “Software for Exascale Computing” (SPPEXA)
and by the Swiss National Science Foundation (SNSF) under the lead agency
agreement as grant SNSF-145271.
References
A. Arteaga, Daniel Ruprecht, and Rolf Krause. A stencil-based implementation of Parareal in the C++ domain specific embedded language
STELLA. Applied Mathematics and Computation, 2015. URL http:
//dx.doi.org/10.1016/j.amc.2014.12.055.
Guillaume Bal. On the convergence and the stability of the parareal algorithm to solve partial differential equations. In Ralf Kornhuber and
et al., editors, Domain Decomposition Methods in Science and Engineering, volume 40 of Lecture Notes in Computational Science and Engineering, pages 426–432, Berlin, 2005. Springer. URL http://dx.doi.org/10.
1007/3-540-26825-1_43.
E. Celledoni and T. Kvamsdal. Parallelization in time for thermo-viscoplastic
problems in extrusion of aluminium. International Journal for Numerical
Methods in Engineering, 79(5):576–598, 2009. URL http://dx.doi.org/
10.1002/nme.2585.
14
Kreienbuehl et al.
Feng Chen, Jan S. Hesthaven, and Xueyu Zhu. On the Use of Reduced Basis
Methods to Accelerate and Stabilize the Parareal Method. In Alfio Quarteroni and Gianluigi Rozza, editors, Reduced Order Methods for Modeling
and Computational Reduction, volume 9 of MS&A - Modeling, Simulation
and Applications, pages 187–214. Springer International Publishing, 2014.
URL http://dx.doi.org/10.1007/978-3-319-02090-7_7.
Roberto Croce, Daniel Ruprecht, and Rolf Krause. Parallel-in-Space-andTime Simulation of the Three-Dimensional, Unsteady Navier-Stokes Equations for Incompressible Flow. In Hans Georg Bock, Xuan Phu Hoang,
Rolf Rannacher, and Johannes P. Schlöder, editors, Modeling, Simulation and Optimization of Complex Processes – HPSC 2012, pages 13–23.
Springer International Publishing, 2014. URL http://dx.doi.org/10.
1007/978-3-319-09063-4_2.
Jack Dongarra et al. Applied Mathematics Research for Exascale Computing.
Technical Report LLNL-TR-651000, Lawrence Livermore National Laboratory, 2014. URL http://science.energy.gov/~/media/ascr/pdf/
research/am/docs/EMWGreport.pdf.
Charbel Farhat, Julien Cortial, C. Dastillung, and H. Bavestrello. Timeparallel implicit integrators for the near-real-time prediction of linear structural dynamic responses. International Journal for Numerical Methods in
Engineering, 67:697–724, 2006. URL http://dx.doi.org/10.1002/nme.
1653.
P. F. Fischer, F. Hecht, and Yvon Maday. A parareal in time semi-implicit
approximation of the Navier-Stokes equations. In Ralf Kornhuber and
et al., editors, Domain Decomposition Methods in Science and Engineering, volume 40 of Lecture Notes in Computational Science and Engineering, pages 433–440, Berlin, 2005. Springer. URL http://dx.doi.org/10.
1007/3-540-26825-1_44.
Martin J. Gander. 50 years of Time Parallel Time Integration. In Multiple
Shooting and Time Domain Decomposition. Springer, 2015. URL {http:
//www.unige.ch/%7Egander/Preprints/50YearsTimeParallel.pdf}.
Martin J. Gander and M. Neumueller. Analysis of a Time Multigrid Algorithm for DG-Discretizations in Time. 2014. URL http://arxiv.org/
abs/1409.5254.
Martin J. Gander and M. Petcu. Analysis of a Krylov Subspace Enhanced
Parareal Algorithm for Linear Problem. ESAIM: Proc., 25:114–129, 2008.
URL http://dx.doi.org/10.1051/proc:082508.
Martin J. Gander and Stefan Vandewalle. On the Superlinear and Linear Convergence of the Parareal Algorithm. In Olof B. Widlund and
David E. Keyes, editors, Domain Decomposition Methods in Science and
Engineering, volume 55 of Lecture Notes in Computational Science and
Engineering, pages 291–298. Springer Berlin Heidelberg, 2007. URL http:
//dx.doi.org/10.1007/978-3-540-34469-8_34.
Volker John. Reference values for drag and lift of a two-dimensional timedependent flow around a cylinder. International Journal for Numerical
Parareal convergence for 2D unsteady flow around a cylinder
15
Methods in Fluids, 44(7):777–788, Mar 2004. doi: 10.1002/fld.679.
Andreas Kreienbuehl, Arne Naegel, Daniel Ruprecht, Robert Speck, Gabriel
Wittum, and Rolf Krause. Numerical simulation of skin transport using
Parareal. Computing and Visualization in Science, Aug 2015. doi: 10.1007/
s00791-015-0246-y. URL http://arxiv.org/abs/1502.03645.
J.-L. Lions, Yvon Maday, and Gabriel Turinici. A ”parareal” in time discretization of PDE’s. Comptes Rendus de l’Académie des Sciences - Series
I - Mathematics, 332:661–668, 2001. URL http://dx.doi.org/10.1016/
S0764-4442(00)01793-6.
Michael L. Minion. A Hybrid Parareal Spectral Deferred Corrections Method.
Communications in Applied Mathematics and Computational Science, 5(2):
265–301, 2010. URL http://dx.doi.org/10.2140/camcos.2010.5.265.
Daniel Ruprecht and Rolf Krause. Explicit parallel-in-time integration of a
linear acoustic-advection system. Computers & Fluids, 59(0):72–83, 2012.
URL http://dx.doi.org/10.1016/j.compfluid.2012.02.015.
Daniel Ruprecht, Robert Speck, Matthew Emmett, Matthias Bolten,
and Rolf Krause. Poster: Extreme-scale space-time parallelism. In
Proceedings of the 2013 Conference on High Performance Computing Networking, Storage and Analysis Companion, SC ’13 Companion,
2013. URL http://sc13.supercomputing.org/sites/default/files/
PostersArchive/tech_posters/post148s2-file3.pdf.
Michael Schäfer, Stefan Turek, Franz Durst, Egon Krause, and Rolf Rannacher. Benchmark Computations of Laminar Flow Around a Cylinder.
In Ernst Heinrich Hirschel, editor, Flow Simulation with High-Performance
Computers II, volume 48 of Notes on Numerical Fluid Mechanics (NNFM),
pages 547–566. Vieweg+Teubner Verlag, 1996. ISBN (13) 9783322898517.
doi: 10.1007/978-3-322-89849-4\ 39.
Robert Speck, Daniel Ruprecht, Rolf Krause, Matthew Emmett, Michael L.
Minion, Mathias Winkel, and Paul Gibbon. A massively space-time parallel
N-body solver. In Proceedings of the International Conference on High
Performance Computing, Networking, Storage and Analysis, SC ’12, pages
92:1–92:11, Los Alamitos, CA, USA, 2012. IEEE Computer Society Press.
URL http://dx.doi.org/10.1109/SC.2012.6.
J. Steiner, Daniel Ruprecht, Robert Speck, and Rolf Krause. Convergence of
Parareal for the Navier-Stokes equations depending on the Reynolds number. In Assyr Abdulle, Simone Deparis, Daniel Kressner, Fabio Nobile, and
Marco Picasso, editors, Numerical Mathematics and Advanced Applications
- ENUMATH 2013, volume 103 of Lecture Notes in Computational Science
and Engineering, pages 195–202. Springer International Publishing, 2015.
URL http://dx.doi.org/10.1007/978-3-319-10705-9_19.
J. M. F. Trindade and J. C. F. Pereira. Parallel-in-time simulation of the
unsteady Navier-Stokes equations for incompressible flow. International
Journal for Numerical Methods in Fluids, 45(10):1123–1136, 2004. URL
http://dx.doi.org/10.1002/fld.732.
16
Kreienbuehl et al.
J. M. F. Trindade and J. C. F. Pereira. Parallel-in-Time Simulation of TwoDimensional, Unsteady, Incompressible Laminar Flows. Numerical Heat
Transfer, Part B: Fundamentals, 50(1):25–40, 2006. URL http://dx.doi.
org/10.1080/10407790500459379.
Andreas Vogel.
Flexible und kombinierbare Implementierung von
Finite-Volumen-Verfahren höherer Ordnung mit Anwendungen für die
Konvektions-Diffusions-, Navier-Stokes- und Nernst-Planck-Gleichungen
sowie dichtegetriebene Grundwasserströmung in porösen Medien. PhD thesis, Johann Wolfgang Goethe-Universität Frankfurt, 2014.
Andreas Vogel, Sebastian Reiter, Martin Rupp, Arne Nägel, and Gabriel
Wittum. UG 4: A novel flexible software system for simulating PDE
based models on high performance computers. Computing and Visualization in Science, 16(4):165–179, Aug 2013.
ISSN 1433-0369.
doi: 10.1007/s00791-014-0232-9. URL http://dx.doi.org/10.1007/
s00791-014-0232-9.
| 5cs.CE
|
Analysis of the Continued Logarithm Algorithm
Pablo Rotondo1,2,3 , Brigitte Vallée2 , and Alfredo Viola3
1
arXiv:1801.10139v2 [cs.DM] 1 Feb 2018
2
IRIF, CNRS and Université Paris Diderot, France
GREYC, CNRS and Université de Caen, France
3
Universidad de la República, Uruguay
Abstract. The Continued Logarithm Algorithm –CL for short– introduced by Gosper in 1978 computes the gcd of two integers; it seems very
efficient, as it only performs shifts and subtractions. Shallit has studied
its worst-case complexity in 2016 and showed it to be linear. We here
perform the average-case analysis of the algorithm : we study its main
parameters (number of iterations, total number of shifts) and obtain precise asymptotics for their mean values. Our “dynamical” analysis involves
the dynamical system underlying the algorithm, that produces continued
fraction expansions whose quotients are powers of 2. Even though this
CL system has already been studied by Chan (around 2005), the presence of powers of 2 in the quotients ingrains into the central parameters a
dyadic flavour that cannot be grasped solely by studying the CL system.
We thus introduce a dyadic component and deal with a two-component
system. With this new mixed system at hand, we then provide a complete
average-case analysis of the CL algorithm, with explicit constants4 .
1
Introduction
In an unpublished manuscript [6], Gosper introduced the continued logarithms,
a mutation of the classical continued fractions. He writes “The primary advantage is the conveniently small information parcel. The restriction to integers of
regular continued fractions makes them unsuitable for very large and very small
numbers. The continued fraction for Avogadro’s number, for example, cannot
even be determined to one term, since its integer part contains 23 digits, only
6 of which are known. (...) By contrast, the continued logarithm of Avogadro’s
number begins with its binary order of magnitude, and only then begins the description equivalent to the leading digits – a sort of recursive version of scientific
notation”.
The idea of Gosper gives rise to an algorithm for computing gcd’s, described
by Shallit in [8]. This algorithm has two advantages: first, it can be calculated
starting from the most representative bits, and uses very simple operations (subtractions and shifts); it does not employ divisions. Second, as the quotients which
intervene in the associated continued fraction are powers of two 2a , we can store
4
Accepted in Latin American Theoretical Informatics 2018.
Thanks to the Dyna3S ANR Project and the AleaEnAmsud AmSud-STIC Project.
2
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
each of them with log2 a bits. Then, the algorithm seems to be of small complexity, both in terms of computation and storage.
Shallit [8] performs the worst-case analysis of the algorithm, and studies the
number of steps K(p, q), and the total number of shifts S(p, q) that are performed
on an integer input (p, q) with p < q: he proves the inequalities
K(p, q) ≤ 2 log2 q + 2,
S(p, q) ≤ (2 log2 q + 2) log2 q ,
(1)
and exhibits instances, namely the family (2n−1 , 1), which show that the previous
bounds are nearly optimal, K(1, 2n−1 ) = 2n − 2, S(1, 2n−1 ) = n(n − 1)/2 + 1 .
In a personal communication, Shallit proposed us to perform the average-case
analysis of the algorithm. In this paper, we answer his question. We consider
the set ΩN which gathers the integer pairs (p, q) with 0 ≤ p ≤ q ≤ N , endowed
with the uniform probability, and we study the mean values EN [K] and EN [S]
as N → ∞. We prove that these mean values are asymptotically linear in the
size log N , and exhibit their precise asymptotic behaviour for N → ∞,
log 3 − log 2
2
log N, EN [S] ∼
EN [K] .
EN [K] ∼
H
2 log 2−log 3
The constant H is related to the entropy of an associated dynamical system and
2
X (−1)k
π
1
+2
− (log 2)(3 log 3−4 log 2) .
(2)
H=
2 log 2−log 3 6
k 2 2k
k≥1
This entails numerical estimates (validated by experiments) for the mean values
EN [K] ∼ 1.49283 log N,
EN [S] ∼ 1.40942 log N .
Then, from (1), the mean number of divisions is about half the maximum.
Our initial idea was to perform a dynamical analysis along the lines described
in [10]. The CL algorithm is defined as a succession of steps, each consisting of a
pseudo-division which transforms an integer pair into a new one. This transformation may be read first on the associated rationals and gives rise to a mapping
T that is further extended to the real unit interval I. This smoothly yields a
dynamical system (I, T ), the CL system, already well studied, particularly by
Chan [3] and Borwein [1]. The system has an invariant density ψ (with an explicit expression described in (8)) and is ergodic. Thus we expected this dynamic
analysis to follow general principles described in [10].
However, the analysis of the algorithm is not so straightforward. The binary
shifts, which make the algorithm so efficient, cause many problems in the analysis. Even on a pair of coprime integers (p, q), the algorithm creates intermediate
pairs (qi+1 , qi ) which are no longer generally coprime, as their gcd is a non-trivial
power of 2. These extra gcd’s are central in the analysis of the algorithm, as they
have an influence on the evolution of the sizes of the pairs (qi+1 , qi ) which may
grow due these extra factors. As these extra factors may only be powers of 2,
they are easily expressed with the dyadic absolute value on Q, at least when
the input is rational. However, when extending the algorithm into a dynamical
system on the unit interval, we lose track of these factors.
Analysis of the Continued Logarithm Algorithm
3
The (natural) idea is thus to add to the usual CL dynamical system (on the unit
interval I) a new component in the dyadic field Q2 . The dyadic component is
just added here to deal with the extra dyadic factors, as a sort of accumulator,
but it is the former real component that dictates the evolution of the system. As
the initial CL system has nice properties, the mixed system inherits this good
behaviour. In particular, the transfer operator of the mixed system presents a
dominant eigenvalue, and the dynamical analysis may be performed successfully.
The constant H in Eqn (2) is actually the entropy of this extended system.
After this extension, the analysis follows classical steps, with methodology mixing tools from analytic combinatorics (generating functions, here of Dirichlet
type), Tauberian theorems (relating the singularities of these generating functions to the asymptotics of their coefficients), functional analysis (which transfers
the geometry of the dynamical system into spectral properties of the transfer
operator).
Plan of the paper and notation. The paper is structured into three sections.
Section 2 introduces the algorithm and its associated dynamical system, as well
as the probabilistic model, the costs of interest and their generating functions.
Then, Section 3 defines the extended dynamical system, allowing us to work
with dyadic costs; it explains how the corresponding transfer operator provides
alternative expressions for the generating functions. Finally, Section 4 describes
the properties of the transfer operator, namely its dominant spectral properties
on a convenient functional space. With Tauberian theorem, it provides the final
asymptotic estimates for the mean values of the main costs of interest.
For an integer q, δ(q) denotes the dyadic valuation, i.e., is the greatest integer k
for which 2k divides q. The dyadic norm | · |2 is defined on Q with the equality
|a/b|2 := 2δ(b)−δ(a) . The dyadic field Q2 is the completion of Q for this norm.
See [7] for more details about the dyadic field Q2 .
2
The CL algorithm and its dynamical system.
We kick off this section with a precise description of the CL algorithm, followed
by its extension to the whole of the unit interval I, giving rise to a dynamical
system, called the CL system, whose inverse branches capture all of our costs
of interest. Then we recall the already known features of the CL system and we
present the probabilistic model, with its generating functions.
Description of the algorithm. The algorithm, described by Shallit in [8], is
a sequence of (pseudo)–divisions: each division associates to a pair (p, q)5 with
p < q a new pair (r, p0 ) (where r is the “remainder”) defined as follows
q = 2a p + r,
5
p0 = 2a p,
with a = a(p, q) := max{k ≥ 0 | 2k p ≤ q} .
Our notations are not the same as in the paper of Shallit as we reverse the roles of
p and q.
4
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
This transformation rewrites the old pair (p, q) in terms of the new one (r, 2a p)
in matrix form,
p
r
0 2−a
0 1
= Na a , with Na =
= 2−a Ma , Ma =
. (3)
q
2 p
1 1
2a 2a
The CL algorithm begins with the input (p, q) with p < q. It lets (q1 , q0 ) := (p, q),
then performs a sequence of divisions
T
T
qi+1 , qi = Nai+1 qi+2 , 2ai+1 qi+1 ,
and stops after k = K(p, q) steps on a pair of the form (0, 2ak qk ). The complete
execution of the algorithm uses the set of matrices Na defined in (3), and writes
the input as
T
T
p, q = Na1 · Na2 · · · Nak 0, 2ak qk .
The rational input p/q is then written as a continued fraction according to the
LFTs (linear fractional transformations) ha associated with matrices Na or Ma ,
p
=
q
2−a1
= ha1 ◦ha2 ◦· · ·◦hak (0) ,
2−a2
1+
2−a3
1+
. 2−ak
1 + ..
1
with ha : x 7→
2−a
. (4)
1+x
Moreover, it is possible to choose the last exponent ak to be 0. (and the last
quotient to be 1). This is a gcd algorithm: as qk is equal to gcd(p, q) up to a
power of 2, the CL algorithm determines the odd part of gcd(p, q) whereas the
even part is directly determined by the dyadic valuations of p and q.
Shallit [8] proves that this algorithm indeed terminates and characterizes the
worst-case complexity of the algorithm. Figure 1 describes the execution of the
algorithm on the pair (31, 75).
i
ai 2ai qi qi+1
0 −
1 1
2 2
3 2
4 1
5 0
6 0
7 −
75
62
52
40
24
16
8
8
31
13
10
12
16
8
8
0
(2ai qi )2 (qi+1 )2 δ(2ai qi ) δ(qi+1 ) δ(b
gi )
1001011
0111110
110100
101000
11000
10000
1000
1000
11111
1101
1010
1100
10000
1000
1000
0
0
1
2
3
3
4
3
3
0
0
1
2
4
3
3
∞
0
0
1
2
3
3
3
3
Fig. 1. Execution for the input (31, 75). Here gbi = gcd(2ai qi , qi+1 ). The dyadic valuation δ(b
gi ) seems to linearly increase with i, with δ(b
gi ) ∼ δ(qi+1 ) ∼ i/2 (i → ∞).
Analysis of the Continued Logarithm Algorithm
5
Dynamical system. The relations
(r, 2a p)T = Na−1 (p, q)T ,
(p, q)T = Na (r, 2a p)T ,
are first transformed into relations on the associated rationals p/q, r/(2a p) via
the LFT’s Ta , ha associated to matrices Na−1 , Na ,
Ta (x) :=
1
− 1,
2a x
ha (x) =
1
,
2a (1 + x)
a ≥ 0.
(5)
They are then extended to the reals of the unit interval I := [0, 1]. This gives
rise to a dynamical system (I, T ), denoted CL in the sequel, defined on the unit
interval I, with fundamental intervals Ia := [2−a−1 , 2−a ], the surjective branches
Ta : Ia → I, and their inverses ha : I → Ia .
The CL system (I, T ) is displayed on the left of the figure below, along with
the shift S : I → I which gives rise to the CL system by induction on the first
branch. The map S is a mix of the Binary and Farey maps, as its first branch
comes from the Binary system, and the second one from the Farey system. On
the right, the usual Euclid dynamical system (defined from the Gauss map) is
derived from the Farey shift by induction on the first branch.
1
1
1
1
1
0
1
1
1
With each k-uple a := ha1 , a2 , . . . , ak i ∈ Nk we associate the matrix Ma :=
Ma1 · · · Mak and the LFT ha := ha1 ◦ ha2 ◦ · · · ◦ hak . Then the set H of the
inverse branches, and the set Hk of the inverse branches of T k (of depth k) are
H := {ha | a ≥ 0},
Hk := {ha | a ∈ Nk }.
As the branches Ta are surjective, the inverse branches are defined on I and the
images ha (I) for a ≥ 0 form a topological partition of I. This will be true at any
depth, and the intervals ha (I), called fundamental intervals of depth k, form a
topological partition for a ∈ Nk .
Properties of the CL system. The Perron Frobenius operator
H[f ](x) :=
X
0
|h (x)| f (h(x)) =
h∈H
1
1+x
2 X
−a
2
f
a≥0
2−a
1+x
.
(6)
describes the evolution of densities : If f is the initial density, H[f ] is the density
after one iteration of the system (I, T ). The invariant density ψ is a fixed point
for H and satisfies the functional equation
ψ(x) =
1
1+x
2 X
a≥0
2−a ψ
2−a
1+x
.
(7)
6
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
Chan [3] obtains an explicit form for ψ
1
1
ψ(x) =
.
log(4/3) (x + 1)(x + 2)
(8)
He also proves that the system is ergodic with respect to ψ, and entropic. However, he does not provide an explicit expression for the entropy. We obtain here
such an expression, with a precise study of the transfer operator of the system.
We introduce two (complex) parameters t, v in (6), and deal with a perturbation
of the operator H, defined by
Ht,v [f ](x) :=
X
|h0 (x)|t d(h)v f (h(x)) =
h∈H
1
1+x
2t X
2 a(v−t) f
a≥0
2−a
1+x
. (9)
Such an operator Ht,v is called a transfer operator. When (t, v) satisfies <(t−v) >
0, we prove the following: the operator Ht,v acts nicely on the space C 1 (I)
endowed with the norm | · |1,1 , defined by |f |1,1 := |f |0 + |f 0 |0 , where | · |0 denotes
the sup norm. In particular, it has a dominant eigenvalue λ(t, v) separated from
the remainder of the spectrum by a spectral gap, for (t, v) close to (1, 0). The
Taylor expansion of λ(t, v) near (1, 0)
λ(t, v) ≈ 1 − A(t − 1) + Dv
involves the two constants A = −∂λ/∂t(1, 1, 0), D = ∂λ/∂v(1, 1, 0) , that are
expressed as mean values with respect to the invariant density ψ,
A = E − D,
E = Eψ [2| log x|],
D = (log 2) Eψ [a] ,
(10)
(here, the function a associates with x the integer defined with the Iverson
bracket a(x) := a · [[x ∈ ha (I)]]. The constants A is the entropy of the system,
and E, D admit explicit expressions
2
X (−1)k
π
1
, D = (log 2) log(3/2) .
+2
(11)
E=
2
k
log(4/3) 6
k 2
log(4/3)
k≥1
Then, with (10) and (11), there is an explicit value for the entropy A, and
.
.
.
A = 1.62352 . . . , D = 0.97693 . . . , E = 2.60045 . . . .
Main costs associated to a truncated expansion. Each real number of
the unit interval admits an infinite continued fraction expansion derived from
the dynamical system, which we call its CLCF expansion. When truncated at a
finite depth, its expansion becomes finite, as in (4), and defines a LFT h := ha .
This expansion gives rise to a rational p/q, (assumed to be irreducible) which is
thus written as p/q = ha (0).
On the other hand, the k-uple a defines a matrix Ma and an integer pair (P, Q),
called the continuant pair, defined by (P, Q)T := Ma (0, 1)T . The equality P/Q =
p/q, holds, but, as the integers P and Q are not necessarily coprime, the pair
(P, Q) does not coincide with the pair (p, q). The integer R(Q) := Q/ gcd(P, Q),
Analysis of the Continued Logarithm Algorithm
7
called the reduced continuant, is an important parameter that actually dictates
the quality of the rational approximation given by the truncation of the CLCF.
We will see that it also plays a central role in the analysis of the CL algorithm.
As gcd(P, Q) divides |det(Ma )| that is a power of two, it is itself a power of two.
It then proves fundamental to deal with dyadic tools.
The main interesting costs associated with a finite expansion, as in (4), are
defined via the LFT h and mainly involve the continuant pair (P, Q) together
with the absolute value of the determinant of the LFT h, denoted as d(h). The
next result describes these costs and provides alternative expressions.
Proposition 1. Consider the function G2 : Q2 → R+ (called the gcd map)
equal to G2 (y) = min(1, |y|−2
2 ), namely
G2 (y) = 1
for |y|2 ≤ 1,
G2 (y) = |y|−2
2
for |y|2 > 1 .
(12)
The main costs associated with the CLCF expansion of a rational h(0)
Q, g(P, Q) := gcd(P, Q), R(P, Q) = Q/gcd(P, Q), |Q|2 .
are all expressed in terms of the quadruple (|h0 (0)|, |h0 (0)|2 , d(h), G2 [h(0)] as
0
Q−2 = |h0 (0)|/d(h), |Q|−2
2 = d(h) |h (0)|2 ,
R−2 (Q) = |h0 (0)| |h0 (0)|2 G2 [h(0)],
g 2 (P, Q) = d(h) |h0 (0)|2 G2 [h(0)] .
Proof. One has (by definition)
P/Q = h(0),
Q−2 = |h0 (0)|/d(h),
r(Q)−2 = g 2 (P, Q)/Q2 .
As g(P, Q) is a power of 2, and using the function G2 defined in (12), one has
−2
−2
g 2 (P, Q) = min(|P |2 , |Q|2 )−2 = |Q|−2
2 min(1, |P/Q|2 ) = |Q|2 G2 (P/Q).
0
We conclude with the equalities : |Q|−2
2 = |h (0)|2 /|d(h)|2 , d(h) · |d(h)|2 = 1.
Any cost C of Proposition 1 admits an expression of the form
|h0 (0)|t |h0 (0)|u2 d(h)v G2 [h(0]z .
The quadruple (t, u, v, z) associated with the cost C is denoted as γC . Moreover, as these costs C are expected to be of exponential growth with respect
to the depth of the CF, we will work with their logarithms c = log C. Figure 2
summarizes the result.
Generating functions. We deal with sets of coprime6 integer pairs
Ω := {(p, q) | 0 < p < q, gcd(p, q) = 1},
ΩN := Ω ∩ {(p, q) | q ≤ N }.
The set ΩN is endowed with the uniform measure, and we wish to study the
mean values EN [c] of parameters c on ΩN . We focus on parameters which describe the execution of the algorithm and are “read” from the CF(p/q) built by
the algorithm as in (4). They are defined in Proposition 1 and depend on the
continuant pair (P, Q); as already explained, the reduced continuant R(P, Q)
plays a fundamental role here.
6
This restriction can be easily removed and our analysis extends to the set of all
integer pairs.
8
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
Cost C
c = log C
d(h)
σ
Q
2
q
2
%
2
R (P, Q)
r
|Q|−2
2
q2
g (P, Q)
Quadruple γC
Constant M (c) Numerical value of M (c)
.
= 0.97693 . . .
(0, 0, 1, 0)
D
.
= 2.60045 . . .
(−1, 0, 1, 0)
A+D
.
(0, 1, 1, 1)
B+D
= 1.26071 . . .
.
(−1, −1, 0, −1)
A−B
= 1.33973 . . .
.
(0, 1, 1, 0)
B+D
= 1.26071 . . .
Fig. 2. Main costs of interest, with their quadruple, and the constants which intervene
in the analysis of their mean values. (see Thm 1).
We deal with analytic combinatorics methodology and work with (Dirichlet)
generating functions (dgf in short). Here is the plain Dirichlet generating function
X 1
ζ(2s − 1)
S(s) :=
=
.
(13)
q 2s
ζ(2s)
(p,q)∈Ω
There are also two generating functions that are associated with a cost C : Ω →
R+ (and its logarithm c), namely the bivariate dgf and the cumulative dgf,
SC (s, w) :=
X ewc(p,q)
X c(p, q)
∂
bC (s) :=
SC (s, w)
,
S
=
2s
2s
q
q
∂w
(p,q)∈Ω
(p,q)∈Ω
. (14)
w=0
The expectation EN [c] is now expressed as a ratio which involves the sums
ΦN (S), ΦN (SbC ) of the first N coefficients of the Dirichlet series S(s) and SbC (s),
EN [c] = ΦN [SbC ]/ΦN [S] .
(15)
From principles of Analytic Combinatorics, we know that the dominant singularity of a dgf (here its singularity of largest real part) provides precise information
(via notably its position and its nature) about the asymptotics of its coefficients, here closely related to the mean value EN [c] via Eqn (15). Here, in the
Dirichlet framework, this transfer from the analytic behaviour of the dgf to the
asymptotics of its coefficients is provided by Delange’s Tauberian Theorem [5].
We now describe an alternative expression of these series, from which it is possible to obtain information regarding the dominant singularity, which will be
transfered to the asymptotics of coefficients.
Proposition 2. The Dirichlet generating S(s) and its bivariate version SC (s, w)
relative to a cost C : Ω → R, admit alternative expressions7
X
S(s) = SC (s, 0),
SC (s, w) =
ewC(h) |h0 (0)|s |h0 (0)|s2 Gs2 ◦ h(0) .
h∈H? ·J
7
We recall that the last exponent is 0 by convention, and the last LFT is thus J = h0 .
Analysis of the Continued Logarithm Algorithm
9
For any cost C described in Figure 2, the general term of SC (s, w) is of the form
|h0 (0)|t |h0 (0)|u2 d(h)v Gz2 ◦ h(0) ,
and involves a quadruple of exponents (t, u, v, z), denoted as γC (s, w), that is
expressed with the quadruple γC defined in Figure 2 as
γC (s, w) = s(1, 1, 0, 1) + w γC .
(16)
Proof. By definition, the denominator q equals R(P, Q). With Figure 2, the
quadruple relative to q −2s is then s(1, 1, 0, 1), whereas the quadruple relative to
ewc = C w is just wγC .
We have thus described the general framework of our paper. We now look for
an alternative form for the generating functions: in dynamical analysis, one expresses the dgf in terms of the transfer operator of the dynamical system which
underlies the algorithm. Here, it is not possible to obtain such an alternative
expressions if we stay in the real “world”. This is why we will add a component to the CL system which allows us to express parameters with a dyadic
flavour. It will be possible to express the dgf’s in term of a (quasi-inverse) of
an (extended) transfer operator, and relate their dominant singularity to the
dominant eigenvalue of this extended transfer operator.
We then obtain our main result, precisely stated in Theorem 1, at the end of the
paper: we will prove that the mean values EN [log C] associated with our costs
of interest are all of order Θ(log N ), and satisfy precise asymptotics that involve
three constants A, B, D: the constants D and A come from the real word, and
have been previously defined in (11) and (10), but there arises a new constant
B that comes from the dyadic word.
3
The extended dynamical system.
In this section, we extend the CL dynamical system, adding a new component to
study the dyadic nature of our costs. We then introduce transfer operators, and
express the generating functions in terms of the quasi-inverses of the transfer
operators.
Extension of the dynamical system. We will work with a two-component
dynamical system: its first component is the initial CL system, to which we add
a second (new) component which is used to “follow” the evolution of dyadic
phenomena during the execution of the first component.
We consider the set I := I × Q2 . We define a new shift T : I → I from the
characteristics of the old shift T defined in (5). As each branch Ta , or its inverse
ha , is a LFT with rational coefficients, it is well-defined on Q2 ; it is moreover a
bijection from Q2 ∪ {∞} to Q2 ∪ {∞}. Then, each branch T a of the new shift T
is defined via the equality T a (x, y) := (Ta (x), Ta (y)) on the fundamental domain
I a := Ia × Q2 , and the shift T a is a bijection from I a to I := I × Q2 whose
inverse branch ha : (x, y) 7→ (ha (x), ha (y)) is a bijection from I to I a .
10
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
Measures. We consider the three domains
B := Q2 ∩ {|y|2 < 1}, U := Q2 ∩ {|y|2 = 1}, C := Q2 ∩ {|y|2 > 1}.
There exists a Haar measure ν0 on Q2 which is finite on each compact of Q2 ,
and can be normalized with ν0 (B) = ν0 (U) = 1/3 (see [7]). We will deal with
the measure ν with density G2 wrt to ν0 , for which ν(C) = 1/3. The measure
ν[2k U] equals (1/3)2−|k| for any k ∈ Z and ν is a probability measure on Q2 .
On I, we deal with the probability measure ρ := µ × ν where µ is the Lebesgue
measure on I and ν is defined on Q2 as previously.
For integrals which involve a Haar measure, there is a change of variables formula8 . As ν0 is a Haar measure, and dν = G2 dν0 , this leads to the following
change of variables formula, for any F ∈ L1 (Q2 , ν),
Z
Z
G2 (h(y))
0
dν(y) =
F (y) dν(y) .
(17)
|h (y)|2 F (h(y))
G2 (y)
Q2
Q2
Density transformer and transfer operator. We now consider the operator H defined as a “density transformer” as follows : with a function F ∈ L1 (I, ρ),
it associates a new function defined by
X
G2 (h(y))
0
0
H [F ](x, y) :=
|h (x)| |h (y)|2 F (h(x), h(y))
.
G2 (y)
h∈H
When F is a density in L1 (I, ρ), then H[F ] is indeed the new density on I
after one iteration of the shift T . This just follows from the change of variables
formula (17) applied to each inverse branch h ∈ H.
Proposition 2 leads us to a new operator that depends on a quadruple (t, u, v, z),
z
X
G2 (h(y))
Ht,u,v,z [F ](x, y) :=
|h0 (x)|t |h0 (y)|u2 d(h)v F (h(x), h(y))
. (18)
G2 (y)
h∈H
We will focus on costs described in Figure 2: we thus deal with operators associated with quadruples γC (s, w) defined in Proposition 2, and in particular with
the quadruple (s, s, 0, s), and its associated operator Hs := Hs,s,0,s .
Alternative expressions of the Dirichlet generating functions. We start
with the expressions of Proposition 1, consider the three types of dgf defined in
(13) and (14), use the equality G2 (0) = 1, and consider the operator Js relative
to the branch J used in the last step. For the plain dgf in (13), we obtain
X
S(s) =
|h0 (0)|s |h0 (0)|s2 Gs2 ◦ h(0) = Js ◦ (I − Hs )−1 [1](0, 0) ,
(19)
h∈H?·J
We now consider the bivariate dgf’s defined in (14). For the depth K, one has
SK (s, w) = ew Js ◦ (I − ew Hs )−1 [1](0, 0) ;
8
This general result can be found for instance in Bourbaki [2], chapitre 10, p.36.
Analysis of the Continued Logarithm Algorithm
11
For costs C of Figure 2, the bivariate dgf involves the quasi-inverse of HγC (s,w) ,
SC (s, w) = JγC (s,w) ◦ (I − HγC (s,w) )−1 [1](0, 0) ,
w
except for C = |Q|−2
2 , where the function 1 is replaced by the function G2 .
The dgf SbC (s) defined in (14) is obtained with taking the derivative of the bivariate dgf wrt w (at w = 0); it is thus written with a double9 quasi inverse which
involves the plain operator Hs , separated “in the middle” by the cumulative
operator Hs,(C) , namely
SbC (s) Js ◦ (I − Hs )−1 ◦ Hs,(C) ◦ (I − Hs )−1 [1](0, 0) ,
and the cumulative operator is itself defined by Hs,(C) :=
4
∂
Hγ (s,w)
∂w C
(20)
.
w=0
Functional Analysis
This section deals with with a delicate context, which mixes the specificities of
each world –the real one, and the dyadic one–. It is devoted to the study of
the quasi-inverses (I − Hs )−1 intervening in the expressions of the generating
functions of interest. We first define an appropriate functional space on which we
prove the operators to act and admit dominant spectral properties. This entails
that the quasi-inverse (I − Hs )−1 admits a pole at s = 1, and we study its
residue, which gives rise to the constants that appear in the expectations of our
main costs.
Functional space. The delicate point of the dynamical analysis is the choice
of a good functional space, that must be a subset of L1 (I, ρ). Here, we know
that, in the initial CL system, the transfer operator Hs acts in a good way on
C 1 (I). Then, for a function F defined on I, the main role will be played by
the family of “sections” F̃y : x 7→ max(1, log |y|2 ) F (x, y) which will be asked to
belong to C 1 (I), under the norm | · |1,1 , defined as |F̃y |1,1 := |F̃y |0 + |F̃y |1 with
|F̃y |0 := supx∈I |F̃ (x, y)|,
|F̃y |1 := supx∈I
We work on the Banach space
n
F := F : I → C | Fy ∈ C 1 (I),
∂
∂x F̃ (x, y)
o
y 7→ F̃y bounded ,
endowed with the norm ||F || := ||F ||0 + ||F ||1 , with
Z
Z
||F ||0 :=
|F̃y |0 dν(y), ||F ||1 :=
|F̃y |1 dν(y) .
Q2
.
(21)
Q2
The next Propositions 3, 4 and 5 will describe the behaviour of the operator
Ht,u,v,z on the functional space F. Their proofs are quite technical and are
omitted here.
9
There is another term which involves only a quasi-inverse. It does not intervene in
the analysis.
12
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
The first result exhibits a subset of quadruples (t, u, v, z) which contains (1, 1, 0, 1)
for which the resulting operator Ht,u,v,z acts on F.
Proposition 3. The following holds:
(a) When the complex triple (t, u, v) satisfies the constraint <(t−v −|u−1|) > 0,
the operator Ht,u,v,u acts on F and is analytic with respect to the triple (t, u, v).
(b) The operator Hs := Hs,s,0,s acts on F for <s > 1/2, and the norm || · ||0 of
the operator Hs satisfies ||Hs ||0 < 1 for <s > 1.
Dominant spectral properties of the operator. The next result describes
some of the main spectral properties of the operator on the space F. Assertion
(a) entails that the k-th iterate of the operator behaves as a true k-th power of
its dominant eigenvalue. Then, as stated in (c), its quasi-inverse behaves as a
true quasi-inverse which involves its dominant eigenvalue.
Proposition 4. The following properties hold for the operator Ht,u,v,u , when
the triple (t, u, v) belongs to a neighborhood V of (1, 1, 0).
(a) There is a unique dominant eigenvalue, separated from the remainder of the
spectrum by a spectral gap, and denoted as λ(t, u, v), with a (normalized)
dominant eigenfunction Ψt,u,v and a dominant eigenmeasure ρt,u,v for the
dual operator.
(b) At (t, u, v, u) = (1, 1, 0, 1), the operator Ht,u,v,u coincides with the density
transformer H1 . At (1, 1, 0) the dominant eigenvalue λ(t, u, v) equals 1, the
function Ψt,u,v is the invariant density Ψ and the measure ρt,u,v equals the
measure ρ.
(c) The estimate holds for any function F ∈ L1 (I, ρ) with ρ[F ] 6= 0,
λ(t, u, v)
Ψt,u,v (x, y) ρt,u,v [F ] .
(I − Ht,u,v,u )−1 [F ](x, y) ∼
1 − λ(t, u, v)
(d) For <s = 1, s 6= 1, the spectral radius of Hs,s,0,s is strictly less than 1.
The third result describes the Taylor expansion of λ(t, u, v) at (1, 1, 0), and makes
precise the behaviour of the quasi-inverse described in (c).
Proposition 5. The Taylor expansion of the eigenvalue λ(t, u, v) at (1, 1, 0),
written as λ(t, u, v) ∼ 1 − A(t − 1) + B(u − 1) + Dv, involves the constants
A = −∂λ/∂t(1, 1, 0), B = ∂λ/∂u(1, 1, 0), D = ∂λ/∂v(1, 1, 0)
(a) The constants A and D already appear in the context of the plain dynamical
system, and are precisely described in (11) and (10). In particular A−D is equal
to the integral E := EΨ [2| log x|];
(b) The constant B is defined with the extension of the dynamical system and its
invariant density Ψ = Ψ1,1,0 . The constant B + D is equal to the dyadic analog
E2 of the integral E, namely, B + D = E2 := EΨ [2 log |y|2 ];
(c) The constant A − B is the entropy of the extended dynamical system.
Analysis of the Continued Logarithm Algorithm
13
Final result for the analysis of the CL algorithm. We then obtain our
final result:
Theorem 1. The mean values EN [c] for c ∈ {K, σ, q, %, r, q2 } on the set ΩN are
all of order Θ(log N ) and admit the precise following estimates,
EN [K] ∼
2
log N,
H
EN [c] ∼ M (c) · EN [K],
for c ∈ {σ, q, %, r, q2 } .
The constant H is the entropy of the extended system. The constants H and M (c)
are expressed with a scalar product that involves the gradient ∇λ of the dominant
eigenvalue at (1, 1, 0) and the beginning γ
bC of the quadruple γC associated with
the cost c. More precisely
H = −h∇λ, (1, 1, 0)i,
M (c) = h∇λ, γ
bC i .
The constants M (c) are exhibited in Figure 2.
Proof. Now, the Tauberian Theorem comes into play, relating the behaviour of
a Dirichlet series F (s) near its dominant singularity with asymptotics for the
sum ΦN (F ) of its first N coefficients. Delange’s Tauberian Theorem is stated as
follows (see [5]):
P
Consider for σ > 0 a Dirichlet series F (s) := n≥1 an n−2s with non negative
coefficients which converges for <s > σ. Assume moreover:
(i) F (s) is analytic on {<s, s 6= σ},
(ii) near σ, F (s) satisfies F (s) ∼ A(s)(s − σ)−(k+1) for some integer k ≥ 0 .
Then, as N → ∞, the sum ΦN (F ) of its first N coefficients satisfies
P
ΦN (F ) := n≤N an ∼ 2k A(σ) [σΓ (k + 1)]−1 N 2σ logk N .
We now show that the two dgf’s S(s) and SbC (s) satisfy the hypotheses of the
Tauberian theorem. The two expressions obtained in (19) and (20) involve quasiinverses (I − Hs )−1 , a simple one in (19), a double one in (20).
First, Propositions 3(b) and 4(d) prove that such quasi-inverses are analytic on
<s ≥ 1, s 6= 1. Then Proposition 4(c), together with Eqn (20), shows that S(s)
and SbC (s) have a pole at s = 1, of order 1 for S(s), of order 2 for SbC (s).
We now evaluate the dominant constants: first, the estimate holds,
1 − λ(s, s, 0) ∼ (A − B)(s − 1) = H(s − 1), with H = −h∇λ, (1, 1, 0)i.
Second, with Proposition 4(c), the dgf’s S(s) and SbC (s) admit the following
estimates which both involve the constant a = J[Ψ ](0, 0), namely,
i
h
a
a
S(s) ∼
, SbC (s) ∼ 2
ρ
H
[Ψ
]
.
1,(C)
H(s − 1)
H (s − 1)2
We now explain the occurence of the constant M (c): we use the definition of the
triple γ
bC (s, w), the definition of the cumulative operator H1,(C) as the derivative
of the bivariate operator HγC (s,w) at (s, w) = (1, 0), and the fact that H1,1,0,1 =
H1 is the density transformer. This entails the sequence of equalities,
h
i
∂
ρ H1,(C) [Ψ ] =
λ(b
γC (1, w))
= h∇λ, γ
bC i = M (c) .
∂w
w=0
14
Pablo Rotondo, Brigitte Vallée, and Alfredo Viola
About the constant B. The invariant density Ψ –more precisely the function
Ψb := Ψ · G2 – satisfies a functional equation of the same type as the invariant
function ψ, (described in Eqn (7)), namely,
Ψb(x, y) =
1
1+x
2
1
1+y
2
X
2 a≥0
Ψb
2−a 2−a
,
1+x 1+y
.
Comparing to Eqn (7), we “lose” the factor 2−a in the sum, and so we have not
succeed in finding an explicit formula for Ψ . We do not know how to evaluate
the integral E2 defined in Proposition 5(b).
However, we conjecture10 the equality D − B = log 2, from experiments of the
same type as those described in Figure 1. This would entail an explicit value for
the entropy of the extended system,
2
X (−1)k
π
1
.
− (log 2)(3 log 3−4 log 2) = 1.33973 . . .
+2
2 log 2−log 3 6
k 2 2k
k≥1
Conclusions and Extensions. We have studied the Continued Logarithm
Algorithm and analyzed in particular the number of pseudo divisions, and the
total number of shifts. It would be nice to obtain an explicit expression of the
invariant density, that should entail a proven expression of the entropy of the
dynamical system. It is also surely possible to analyze the bit complexity of the
algorithm, notably in the case when one eliminates the rightmost zeroes when are
shared by the two qi0 s (as suggested by Shallit). Such a version of this algorithm
may have a competitive bit complexity that merits a further study.
There exist two other gcd algorithm that are based on binary shifts, all involving
a dyadic point of view: the Binary Algorithm, and “the Tortoise and the Hare”
algorithm, already analyzed in [9] and [4]; however, the role of the binary shifts
is different in each case. The strategy of the present algorithm is led by the most
significant bits, whereas the strategy of the “Tortoise and the Hare” is led by
the least significant bits. The Binary algorithm adopts a mixed strategy, as it
performs both right-shifts and subtractions. We have the project to unify the
analysis of these three algorithms, and better understand the role of the dyadic
component in each case.
References
1. Jonathan M. Borwein, Kevin G. Hare and Jason G. Lynch. Generalized Continued
Logarithms and Related Continued Fractions, Journal of Integer Sequences, vol.
20 , 2017.
2. Nicolas Bourbaki, Variétés différentielles et analytiques. Springer 2007
3. Hei-Chi Chan. The Asymptotic Growth Rate of Random Fibonacci Type Sequences. Fibonacci Quarterly, vol. 43, no. 3, pp. 243–255, 2005.
10
This will be explained in the long paper.
Analysis of the Continued Logarithm Algorithm
15
4. Benoit Daireaux, Véronique Maume-Deschamps, Brigitte Vallée. The Lyapounov
Tortoise and the Dyadic hare. Proceedings of AofA’05, DMTCS, pp. 71–94, 2005.
5. Hubert Delange. Généralisation du Théorème d’Ikehara, Ann. Sc. ENS 71, pp
213–242, 1954
6. Bill Gosper. Continued fraction arithmetic, Unpublished manuscript, ca. 1978.
7. Neal Koblitz. p-adic Numbers, p-adic analysis and Zeta functions. 2nd edition,
Springer Verlag, 1984
8. Jeffrey Shallit. Length of the continued logarithm algorithm on rational inputs.
https://arxiv.org/abs/1606.03881v2 arXiv:1606.03881v2, 2016.
9. Brigitte Vallée. Dynamics of the Binary Euclidean Algorithm: Functional analysis
and operators, Algorithmica, vol. 22, no. 4, pp. 660–685, 1998.
10. Brigitte Vallée. Euclidean Dynamics. Discrete and Continuous Dynamical Systems,
vol. 15, no. 1, pp. 281–352, 2006.
| 8cs.DS
|
Abstract
arXiv:cs/0508117v1 [cs.NE] 26 Aug 2005
We report the first results of simulating the coupling of neuronal,
astrocyte, and cerebrovascular activity. It is suggested that the dynamics of the system is different from systems that only include neurons. In the neuron-vascular coupling, distribution of synapse strengths
affects neuronal behavior and thus balance of the blood flow; oscillations are induced in the neuron-to-astrocyte coupling.
1
Long-term neuronal behavior caused by two
synaptic modification mechanisms
Xi Shen (1), Philippe De Wilde (2)
((1) Department of Electrical and Electronic Engineering,
Imperial College London, United Kingdom,
(2) Department of Computer Science,
Heriot-Watt University, United Kingdom)
April 4, 2018
Recently, more details have emerged of the interaction between neurons,
glial cells, and the cerebrovascular network [1, 2, 3]. Most of this work is
on the micro-level, involving only a few cells and capillaries or arterioles.
Numerous investigations have confirmed the following discoveries: The modulation of synaptic efficacy will affect emergent behavior of neuronal assemblies [4, 5]; Dilation of capillaries is highly related to the activity of nearby
neurons [6]; Astrocytes are very sensitive to the level of neuronal activity
because of their position and their sensitivity to activity-dependent changes
in the chemical environment they share with neurons [7]; Intercellular Calcium waves between astrocytes are the main signalling mechanism within
glial cell networks [8, 9]. In this letter, we will build numerical models based
on those physiological findings, because micro-descriptions using simplified
models of firing [10, 11] and wave propagation can be inserted into larger
scale simulations. Here we show a modification of the synapse strengths that
allows the neuronal firing and the cerebrovascular flow to be compatible on a
meso-scale; with astrocyte signalling added, limit cycles exist in the coupled
networks.
Neurons are associated with capillaries in the brain, and according to
physiological discoveries, the neuronal activity can dilate the capillaries that
supply them. Our first model contains 2400 neurons and 30 branches of
capillaries, each of which supplies 80 neurons. Each neuron has two states:
’1’ indicates that at that time the neuron was firing and ’-1’ means the neuron
was not firing. All the synapses between these neurons are described by a
2
2400 by 2400 matrix S, in which each component Si,j represents the strength
of the synapse from neuron j to neuron i. This definition is according to the
Hopfield artificial neural network model [12]. On average, each neuron has a
certain number of synapses which are linking to other neurons. We suppose
that the synapse number follows a normal distribution that peaks at Save .
As a result, most of the neurons have a similar number of synapses. This is
based on the assumption that a small area in the brain is uniform.
These 2400 neurons can be viewed as a very small part of the brain, in
which each neuron has synapses connected with other neurons within this
part as well as outside of this part. Here we define another parameter ploc ,
which indicates the locality of the synapses, i.e., how many of one neuron’s
synapses are connected to the neurons within this area. Consequently, the
input signal of each neuron consists of two parts: one part is from the local
neurons which are connected with it via synapses and the other part is from
the neurons outside, this part can be regarded as Gaussian noise. The capillary model is simple, the original blood flow is set as follows: the blood flow
of the first layer (top two branches) is 800 µm3 /ms , and where the branches
split, both the directions have half this flow, 400µm3/ms; this rule is applied
all the way to the bottom layer, where all the 16 branches have a flow of
100µm3 /ms, which is the low velocity of flow in the capillaries quoted from
data of blood circulation [13]. If the number of firing neurons associated
with this capillary exceeds a certain threshold at time T , the capillary will
dilate at the time, which means in our model that the blood volume is four
times as large as the original one. In addition, the dilation period is set to
dc , because the change of capillary width is slower than that of the neuron
potentials. We also define a parameter which represents the compatibility as
follows:
Cmp =
j=14
X
|Bldf lwj − Bldf lw2j+1 − Bldf lw2j+2 |
(1)
j=1
Compatibility is the summation of the blood flow difference over all 14
capillary joint nodes; the capillary network is illustrated in Fig. 1. We call a
network compatible if the compatibility (Eq.(1)) is zero.
The dilation of each capillary is caused by the fluctuation of the firing
number associated with it, i.e., if more than Nf iring neurons that this capillary
supplies are firing, this capillary is going to be dilated after ∆t time steps
and this dilation will last dc time steps. In a first set of simulations, all
the synapse strengths are uniformly distributed between -1 and 1, the blood
flow turned out to be highly incompatible, since the blood flow difference
fluctuated between 3000 and 8000 µm3 /ms (also see Fig. 2).
3
Figure 1: Thirty branches of capillaries.
Figure 2: Compatibility for the model with synaptic weights uniformly distributed between -1 and 1. The blood flow is not compatible.
4
Figure 3: Average firing number per branch (top): We calculated the firing
number associated with each branch of capillary, evaluated the mean value
and plotted in this figure. Compatibility (bottom): The incompatible time is
really short because the neurons quickly change from ’resting’ to ’excitatory’.
In the second simulation, we manually set these strengths to be uniformly
distributed between 2µ − 1 and 1, i.e., the mean value of them is µ, which
we call local-shift. Initially, the neuron states are set according to the rule
that 1/4 of the neurons are firing. The simulation results are shown in
Fig. 3. We found that this initialization had no effect on the behavior
of the system; neurons will reach their resting state, which is determined
by µ. Over a certain short time period (from T=101 to T=105 in this
simulation), we increased the Gaussian noise in this area so this positive
input forces a number of neurons to fire. Then the neurons’ state will be
changed to ’excitatory’, and at the same time the high level of neuron firing
will trigger the dilation of capillaries. The results imply that this temporary
external noise triggers the ’excitatory’ state and that a local shift µ helps
the maintenance of this state until the next inhibitory signal arrives. This
quick change from ’resting’ to ’excitatory’ makes the incompatible time really
short, so in this way, these coupled networks function well.
Based on these results, we can make a few hypotheses about the neurovascular coupling in the brain. The first hypothesis is that the local synapses
are likely to have positive strengths. This enables the cluster of neurons
5
to maintain either the ’excitatory’ state or the ’resting’ state, and external
signals cause the switch between these two states. From simulations we can
see that this switch is fast, although the capillaries react with time delay,
the incompatible time is so short that it can be ignored. Secondly, since the
local shift of the synapse weights is so tiny, we could assume that this is
associated with the neuronal communication. The communication between
neurons could slightly change the strength of synapses connected them, and
this change is on the same time scale as the dilation of the capillaries. Because the neuron state is determined by this local-shift, changing of strength
could result in changing of state, and at the same time balance the blood
flow.
It is now believed that glial cells (astrocytes) also have their own network
(via Calcium waves) in the brain and that this network also plays a distinct
role in information processing. Calcium waves can both propagate among
networks of astrocytes globally and send specific Calcium signals to a small
number of nearby astrocytes [14]. In other words, glial signalling also has
its preferred routes. Communication between glial cells and neurons is bidirectional and complex. Astrocytes are very sensitive to the level of neuronal
activity because of their position and their sensitivity to activity-dependent
changes in the chemical environment shared by neurons and astrocytes [7, 15].
Astrocytic Calcium waves can be triggered by synaptically-released neurotransmitters and in addition the frequency of these Calcium oscillations can
change according to the level of synaptic activity [16]. It is observed that
activated astrocytes can control the synaptic transmission by regulating the
release of neurotransmitter from the nerve terminal. This regulation can
be either excitatory (by secreting the same neurotransmitter) or inhibitory
(by absorbing the neurotransmitter). If regulation of synaptic activities is
the short-term effect of astrocytes, they can also change long-term synaptic
strengths by releasing signalling molecules that cause the axon to increase
or decrease the amount of neurotransmitter [14]. To sum up briefly, it is
now possible to model the neuron-to-astrocyte coupling because many recent experiments confirm the signalling pathways among these two types of
networks. Neurons communicate with neurons while astrocytes ’listen to’ the
neuronal communication. Basing upon what they ’hear’, astrocytes regulate
neuronal activities and communicate with other astrocytes. In order to quantify these regulations, we will define rules dependent on several parameters
and then simulate this model.
The neuronal network contains 900 neurons, each of which has either a
’1’ state (firing) or a ’0’ state (resting). There are 3600 astrocytes in the
glial network. In order to simplify, we suppose these astrocytes are in a
60 × 60 matrix G. The state of an astrocyte is gi,j and it could be zero
6
or a positive number. These 3600 astrocytes have their own connections,
too. Because Calcium signals are more likely to propagate among nearby
astrocytes, we assume here that the connection between astrocyte (i1, j1)
and (i2, j2) is given by probability
ρ, which is inversely proportional to the
q
2
distance between them ( (i1 − i2) + (j1 − j2)2 ). Since these connections
are in fact invisible Calcium signal pathways, we did not give them weights.
Unlike electrical signals travelling between neurons, Calcium signals need
more time to travel from one astrocyte to another, and we set this time
delay proportional to the distance.
The coupling between the neuronal network and the glial network is complicated. We will assume that each neuron has four astrocytes nearby which
are all sensitive to this neuron’s activity. At the same time, because in the
brain a single astrocyte can cover a large number of synapses [9], in our
model we allocate to each astrocyte 1/4 of all the afferent synapses of this
neuron, so that four nearby astrocytes can cover all afferent synapses without
overlapping. Astrocytes can sense the spike frequency along each synapse,
regulate the release of neurotransmitters, and change the synaptic strength.
A simplified illustration of the position of neurons and astrocytes can be
found in Fig. 4. There are two factors which affect the state of an astrocyte.
The first one is high frequency effective spikes along synapses that this astrocyte is sensing, and ’effective’ means either a positive spike making the
target neuron fire or a negative spike making it rest. The second factor is the
Calcium signals from other astrocytes, and this part can accumulate as well.
If the input signal of an astrocyte exceeds a certain threshold , this astrocyte
will become active, which means that the state of the astrocyte is nonzero.
Since activated astrocytes can have different levels of activation, we define
that when activated, an astrocyte’s state equals its input.
As mentioned previously, activated astrocytes can regulate the release of
neurotransmitters over short periods of time (< 50s) [17], in other words;
they can change the input signal of the nearby neuron within a few seconds.
Because the conditions for excitation and inhibition are unknown, we will
make the following assumptions. The neuron’s input will be increased by
an activated astrocyte proportionally to its state, but if this astrocyte was
over-activated (its state exceeded a certain value), it would inhibit the input
of the neuron. In the brain this local rule may be more complicated, and
we believe that this interaction between neurons and astrocytes is crucial to
short-term memory and information processing.
Meanwhile, as a long-term effect, astrocytes can also monitor the synaptic
activities and gradually modify the strengths of synaptic connections (efficacies). This mechanism is also called adapting synapses, and has been studied
7
Figure 4: The position of one neuron (black circle) and 4 nearby astrocytes
(rectangles). To make it clear, only afferent synapses from other neurons
(shadowed circle) are drawn. The arms of astrocytes (ellipses) surround
some synapses to sense and regulate their activities.
before [18]. Because our model is based on discrete time, the update rule of
synaptic strengths is quite simple. For a positive synapse, we look back 20
time steps and calculate the spike frequency of this synapse and the firing frequency of the neuron. If both of the two frequencies are high and correlated,
in other words, the high firing rate is triggered by episodes of high synaptic
activity through this connection (we call it excitation success), we will make
this connection stronger. On the other hand, if this synapse is negative, it
will be enhanced if the spike frequency is high but the neuron is firing at a
low rate (inhibition success). The strength can also be weakened if high spike
frequency along a positive synapse does not cause high firing rate (excitation failure) or high spike frequency along a negative synapse does not cause
low firing rate (inhibition failure). Although due to limitations of computing
time, our simulating time can not be long enough to represent the gradual
changing of efficacy; we still include this procedure in our program. We believe that the distribution of synaptic strengths heavily affects the emergent
behavior of neuron-to-astrocyte coupling, so it is crucial in understanding
the brain function, especially long-term memory and learning [19, 20].
4
900
X
X
1
1
np (t + 1) = sgn( Tqp (t)nq (t) +
g4p+r (t) − ϕp (t) + η) +
2
2
r=1
q=1
8
(2)
Figure 5: Top figure: Average number of firing neurons in each row in our
simulation of pure neurons. Bottom figure: average number of firing neurons
in each row in our simulation of neuron-to-astrocyte coupling.
Finally, we will explain the time resolution of this model. Based on the
absolute refractory period of neurons, we suppose that each time step in our
simulation equals 1 ms in the real brain. Since neurons can not be firing
at two consecutive time steps, the maximum firing frequency is 500 Hz. At
each time step, we loop over all astrocytes, all neurons and all synapses and
change their states according to the rules we have defined above. The update
rule for a neuron can be found in Eq.(2), in which Tqp (t) represents the timeP
dependent strength, 4r=1 g4p+r (t) means contributions from astrocytes and
ϕp (t) and η are the threshold and noise, respectively.
Our results (Fig.5) imply that in the pure neuronal network, neuronal
activity was quite random within the whole 500 time steps. While in the
simulation with both neurons and astrocytes, it is obvious that there was
an attractor in the neuronal network, after T=300, the neuronal network
9
just oscillated between two states. If we compare the two figures concerning
average firing number in Fig. 5, it is easy to find out that the firing frequency
of neurons in the coupled network is much higher than that in the pure
neuronal network. According to our simulations, on average 17 percent of all
neurons were firing in the pure neuronal network while in the coupled network
it was 39.2 percent. This is mainly because astrocytes changed the strengths
of many synapses. At the same time, activated astrocytes enhancing the
afferent spikes play a small part in this as well.
We have shown that a modification of the synapse strengths can allow the
neuronal firing and the cerebrovascular flow to be compatible on a meso-scale.
With astrocyte signalling added, limit cycles exist in the coupled networks.
This is a first step towards a better understanding of the coupling of neuronal, astrocyte and cerebrovascular networks in the brain. This coupling is
essential in any modern description of the brain, and has to be investigated
using methods of complex systems. They will ultimately show the right scale
for the models, and whether critical phenomena exist.
References
[1] P. G. Haydon, Nat. Neurosci. 2, 185 (March 2001).
[2] S. J. Mulligan and B. A. MacVicar, Nature 431, 195 (2004).
[3] C. Iadecola, Nat. Neurosci. 5, 347 (May 2004).
[4] L. F. Abbott and W. G. Regehr, Nature 431, 796 (2004).
[5] D. B. Chklovskii, B. W. Mel and K. Svoboda, Nature 431, 782 (2004).
[6] G. A. Cowan, Working Papers from Santa Fe Institute (1988).
[7] R. D. Fields and B. Stevens-Graham, Science 298, 556 (2002).
[8] K. Braet, L. Cabooter, K. Paemeleire and L. Leybaert, Biol. Cell 96, 79
(2004).
[9] G. Perea and A. Araque, J. Physiol. - Paris 96, 199 (2002).
[10] J. J. Hopfield, Proc. Natl. Acad. Sci. USA 79, 2554 (1982).
[11] J. M. Casado, Phys. Rev. Lett. 91, 208102 (2004).
[12] P. De Wilde, Neural Network Models (Springer-Verlag, 1997).
10
[13] http://www.uoguelph.ca/∼dstevens/teach/320/problems.htm.
[14] R. D. Fields, Sci. Am. 54, (Apr 2004).
[15] T. Fellin and G. Carmignoto, J. Physiol. 559, 1 (2004).
[16] M. Zonta and G. Carmignoto, J. Physiol. - Paris 96, 193 (2002).
[17] J. Kang, S. A. Goldman and M. Nedergaard, Nat. Neurosci. 1, 683
(1998).
[18] D. W. Dong and J. J. Hopfield, Networks 3, 267 (1992).
[19] A. Destexhe and E Marder, Nature 431, 789 (2004).
[20] M. Tsodyks and C. Gilbert, Nature 431, 775 (2004).
11
| 5cs.CE
|
Conjunctions of Among Constraints
Vı́ctor Dalmau
arXiv:1706.05059v1 [cs.AI] 15 Jun 2017
Dept. of Information and Communication Technologies, Universitat Pompeu Fabra
Abstract. Many existing global constraints can be encoded as a conjunction of among constraints. An among constraint holds if the number
of the variables in its scope whose value belongs to a prespecified set,
which we call its range, is within some given bounds. It is known that
domain filtering algorithms can benefit from reasoning about the interaction of among constraints so that values can be filtered out taking into
consideration several among constraints simultaneously. The present paper embarks into a systematic investigation on the circumstances under
which it is possible to obtain efficient and complete domain filtering algorithms for conjunctions of among constraints. We start by observing that
restrictions on both the scope and the range of the among constraints are
necessary to obtain meaningful results. Then, we derive a domain flowbased filtering algorithm and present several applications. In particular,
it is shown that the algorithm unifies and generalizes several previous
existing results.
1
Introduction
Global constraints play a major role in constraint programming. Very informally,
a global constraint is a constraint, or perhaps more precisely, a family of constraints, which is versatile enough to be able to express restrictions that are
encountered often in practice. For example, one of the most widely used global
constraints is the ’All different’ constraint, AllDiff(S) where S = {x1 , . . . , xn }
is a set of variables, which specifies that the values assigned to the variables in
S must be all pairwise different. This sort of restriction arises naturally in many
areas, such as for example scheduling problems, where the variables x1 , . . . , xn
could represent n activities that must be assigned different times of a common
resource.
Besides is usefulness in simplifying the modeling or programming task, global
constraints also improve greatly the efficiently of propagation-search based solvers.
This type of solver performs a tree search that constructs partial assignments
and enforces some sort of propagation or local consistency that prunes the space
search. Different forms of consistency, including (singleton) bounds consistency,
(singleton, generalized) arc-consistency, path consistency and many others, can
be used in the propagation phase. One of the most commonly used forms of local consistency is domain consistency, also called generalized arc-consistency. A
domain consistency algorithm keeps, for every variable v, a list, L(v), of feasible
values, which is updated, by removing a value d from it, when some constraint
in the problem guarantees that v cannot take value d in any solution. One of
the key reasons of the success of global constraints is that they enable the use
of efficient filtering algorithms specifically tailored for them.
Several constraints studied in the literature including AllDiff[24], GCC[27],
Symmetric-GCC [21], Sequence [6], GlobalSequencing [34], OrderedDistribute [28], and CardinalityMatrix [33] can be decomposed as the
conjunction of a simpler family of constraints, called among constraints [6]. An
among constraint has the form Among(S, R, min, max) where S is again a set
of variables called the scope, R is a subset of the possible values, called the range,
and min, max are integers This constraint specifies that the number of variables
in S that take a value in R must be in the range {min, . . . , max}. For example,
the constraint AllDiff(S) can be expressed as the conjunction of constraints
Among(S, {d1 }, 0, 1), . . . , Among(S, {dk }, 0, 1) where d1 , . . . , dk are the set of
all feasible values for the variables in S.
Besides encoding more complex global cardinality constraints, conjunctions
of among constraints, (CAC), appear in many problems, such as Sudoku or latin
squares. In consequence, CACs have been previously studied [10,32,38], specially
the particular case of conjunctions of AllDiff constraints [2,3,9,13,22,23,25].
Although deciding the satisfiability of an arbitrary conjunction of among constraints is NP-complete [32] this body of work shows that sometimes there
are benefits in reasoning about the interaction between the among constraints.
Hence, it is important to understand under which circumstances among constraints can be combined in order to endow CSP solvers with the ability to
propagate taking into consideration several among constraints simultaneously.
The aim of the present paper is to contribute to this line of research. To this
end we first observe that restrictions on both the scope and the range of the
among constraints are necessary to obtain meaningful results. Then we embark
in a systematic study of which such restrictions guarantee efficient propagation
algorithms. In particular, we introduce a general condition such that every CAC
satisfying it admits an efficient and complete domain filtering algorithm. This
condition basically expresses that the matrix of a system of linear equations
encoding the CAC instance belongs to a particular class of totally unimodular matrices known as network matrices. This allows to reformulate the domain
filtering problem in terms of flows in a network graph and apply the methodology derived by Régin [30,31]. The algorithm thus obtained, although simple,
unifies and generalizes existing domain filtering algorithms for several global constraints, including AllDiff, GCC, Sequence, Symmetric-GCC, OrderedDistribute as well as for other problems expressed as conjunctions of among
constraints in [29,32]. A nice feature of our approach is that it abstracts out the
construction of the network flow problem, so that when exploring a new CAC
one might leave out the usually messy details of the design of the network graph
and reason purely in combinatorial terms.
Several filtering methods have been obtained by decomposing a global constraint into a combination of among constraints. For example the first polynomialtime filtering algorithm for the Sequence constraint [38] is obtained explicitely
2
in this way. However, there have been very few attempts to determine sistematically which particular conjunctions of among constraints allow efficient filtering
algorithms. The seminal paper in this direction is [31] which identifies several
combinations of among and GCC constraints that admit a complete and efficient domain filtering algorithm (see Section 6 for more details). Our approach
is more general as it subsumes the tractable cases introduced in [31]. Another
closely related work is [29] where two tractable combinations of boolean CACs,
called TFO and 3FO, are identified. The approach in [29] differs from ours in
two aspects: it deals with optimization problem and also considers restricions on
the min and max parameters of the among constraints while we only consider
restriction on the scope and range. A different family of CACs has been investigated in [10] although the work in [10] focuses in bound consistency instead of
domain consistency.
Other approaches to the design of filtering algorithms for combinations (but
not necessarily conjunctions) of global (but not necessarily among) constraints
are described in [4,7,8]. The method introduced in [4] deals with logical combinations of some primitive constraints but differs substantially from ours in the
sense that it cannot capture a single among constraint. The work reported in
[7,8] does not guarantee tractability.
Several proofs are omiitted due to space restrictions. They can be found in
the Appendix.
2
Preliminaries
A conjunction of among constraints, (CAC) is a tuple (V, D, L, C) where V is
a finite set whose elements are called variables, D is a finite set called domain,
L : V → 2D is a mapping that sends every variable v to a subset of D, which we
call its list, and C is a finite set of constraints where a constraint is an expression
of the form Among(S, R, min, max) where S ⊆ V is called the scope of the
constraint, R ⊆ D is called range of the constraint, and min, max are integers
satisfying 0 ≤ min ≤ max ≤ |S|.
A solution of (V, D, L, C) is a mapping s : V → D such that s(v) ∈ L(v) for
every variable v ∈ V and min ≤ |{v ∈ S | s(v) ∈ R}| ≤ max for every constraint
Among(S, R, min, max) in C.
Example 1. (GCC and AllDiff constraints) The global cardinality constraint1 ,
GCC [27] corresponds to instances (V, D, L, C) where all the constraints have
the form Among(V, {d}, min, max) with d ∈ D. The AllDiff constraint is the
particular case obtained when, additionally, min = 0 and max = 1.
Let I = (V, D, L, C) be a CAC. We say that a value d ∈ D is supported for a
variable v ∈ V if there is a solution s of I with s(v) = d. In this paper we focus in
the following computational problem, which we will call domain filtering: given
a CAC, compute the set of all the non supported values for each of its variables.
1
We want to stress here that a global constraint is not a single constraint but, in fact,
a family of them.
3
This definition is motivated by the following scenario: think of (V, D, L, C) as
defining a constraint which is part of a CSP instance that is being solved by a
search-propagation algorithm that enforces domain consistency. Assume that at
any stage of the execution of the algorithm, L encodes the actual feasible values
for each variable in V . Then, the domain filtering problem is basically the task
of identifying all the values that need to be pruned by considering the constraint
encoded by (V, D, L, C).
3
Network hypergraphs
An hypergraph H is a tuple, (V (H), E(H)), where V (H) is a finite set whose
elements are called nodes and E(H) is set whose elements are subsets of V (H),
called hyperedges. An hypergraph is totally unimodular if its incidence matrix M
is totally unimodular, that is, if every square submatrix of M has determinant
0, +1, or −1. In this paper we are concerned with a subset of totally unimodular
hypergraphs called network hypergraphs. In order to define network hypergraph
we need to introduce a few definitions.
An oriented tree T is any directed tree obtained by orienting the edges of
an undirected tree. A path p in T is any sequence x1 , e1 , x2 , . . . , en−1 , xn where
x1 , . . . , xn are different vertices of T , e1 , . . . , en−1 are edges in T and for every
1 ≤ i < n, either ei = (xi , xi+1 ) or ei = (xi+1 , xi ). The polarity of an edge
e ∈ E(T ) wrt. p is defined to be +1 (or positive) if e = (xi , xi+1 ) for some
1 ≤ i < n, −1 (or negative) if e = (xi+1 , xi ) for some 1 ≤ i < n, and 0 if e
does not appear in p. A path p has positive (resp. negative) polarity if all its
edges have positive (resp. negative) polarity. Paths with positive polarity are also
called directed paths. Since an oriented tree does not contain symmetric edges,
we might represent a path by giving only its sequence of nodes x1 , . . . , xn .
We say that an oriented tree T defines an hypergraph H if we can associate
to every hyperedge h ∈ E(H) an edge eh ∈ E(T ) and to every node v ∈ V (H) a
directed path pv in T such that for every v ∈ V (H) and h ∈ E(V ), v ∈ h if and
only if eh belongs to pv . We say that an H is a network hypergraph if there is
an oriented tree that defines it.
Example 2. The hypergraph H with variable-set {v1 , . . . , v6 } and hyperedge-set
{h1 , . . . , h5 } given in Figure 1a is a network hypergraph as it is defined by tree
T given in Figure 1b where we have indicated, using labels on the edges, the
edge in T associated to every hyperedge in H. We associate to every variable vi ,
1 ≤ i ≤ 6 the directed path s(i−1 mod 2) , r, t(i−1 mod 3) in T . It can be readily
checked that under this assignment T defines H.
Sometimes, it will be convenient to assume that the tree T defining H is
minimal in the sense that no tree with fewer nodes defines H. Minimal trees
have the nice property that every edge e in T is associated with some hyperedge
of H. Indeed, assume that some edge e = (x, y) is not associated to any hyperedge
in H, then one could find an smaller tree T defining H by contracting edge e,
that is, by merging x and y into a new node z that has as in-neighbours the
4
h1
h3
h2
h5
v4
v6
v2
h4
v1
v3
v5
s0
h4
h5
t0
h1
h2
r
t1
h3
s1
t2
Fig. 1a: Hypergraph H
Fig. 1b: Oriented tree T
union of all in-neighbours of x and y and, as out-neighbours, the union of all
out-neighbours of x and y.
Since the vast majority of the trees defined in this paper will be oriented
we shall usually drop ’oriented’. So, unless, otherwise explicitly stated, a tree is
always an oriented tree. Finally, we note that one can decide whether a given
hypergraph is a network hypergraph in time O(e3 v 2 ) where e is the number of
hyperedges and v is the number of nodes (see chapter 20 in [35] for example).
4
Restricting only the scope or the range
It has been shown by Régin [32] that the domain filtering problem for CACs
is NP-hard. Still, efficient algorithms are known for some particular cases. It
seems natural to start by asking which tractable subcases of the problem can
be explained by considering only the scopes of the constraints. This question
has a close similarity to the study of the so-called structural restrictions of the
CSP (see, for example [18] for a survey) and, not surprisingly, it can be solved
by applying results developed there. Indeed, it follows easily from a result of
Färnquivst and Jonsson [12] that, modulo some mild technical assumptions, if
one allows arbitrary ranges in constraints, then the domain filtering problem
is solvable in polynomial time if and only if the hypergraph of the scopes of
the constraints has bounded tree-width (see Appendix A for precise statement
and the proof). This result, although delineates exactly the border between
tractability and intractability, turns out to be not very useful in explaining the
tractability of global constraints. This is due to the fact that global cardinality
constraints defined by conjunctions of among constraints usually have constraints
with large scopes and the cardinality of the scope in a constraint is a lower bound
on the tree-width of its scope hypergraph.
One can also turn the attention to the range of constraints and inquiry
whether there are tractable subcases of the problem that can be explained only
by the range of the constraints. Here, again the response is not too useful. Indeed, it is very easy to show (see again Appendix A) that as soon as we allow
some non-trivial range R (that is some range different than the empty set and
than the whole domain) and arbitrary scopes in the among constraints, then the
domain filtering problem becomes NP-complete.
In view of this state of affairs it is meaningful to consider families of conjunctions of among constraints that are obtained by restricting simultaneously
5
the scope and the range of the constraints occurring in them. This is done in the
next section.
5
A flow-based algorithm
Let I = (V, D, L, C) be conjunction of among constraints. We will deal first with
the case in which D is a boolean, say D = {0, 1}. Hence, we can assume that
every constraint Among(S, R, min, max) in C satisfies R = {1} since, if R = {0}
it can be reformulated as Among(S, {1}, |S| − max, |S| − min). We also assume
that L(v) = {0, 1} for every v ∈ V since if L(v) 6= {0, 1} we could obtain easily
an equivalent instance without variable v.
It is easy to construct a system of linear equations whose feasible integer
solutions encode the solutions of I. Let v1 , . . . , vn be the variables of I and let
Cj = (Sj , {1}, minj , maxj ), j = 1, . . . , m, be its constraints. The system has
variables xi (1 ≤ i ≤ n), yj (1 ≤ j ≤ m) and the following equations:
X
yj +
xi = maxj
j = 1, . . . , m
vi ∈Sj
0 ≤ yj ≤ maxj − minj
0 ≤ xi ≤ 1
j = 1, . . . , m
i = 1, . . . , n
which we express in matrix form as
Mz = a
0≤z≤c
with z T = (x1 , . . . , xn , y1 , . . . , ym ) (see Example 3).
If M is totally unimodular then one can perform domain filtering in polynomial time. Indeed, for every vi ∈ V and d ∈ L(v), we might decide whether d is
a supported value for v as follows: add equation xi = d to the system and decide
whether there exists a feasible solution of its linear relaxation using a LP solver.
It follows from total unimodularity (see Theorem 19.1 in [35] for example) that
such a feasible solution exists if and only if d is a support for v.
However, this approach implies invoking O(n) times a LP solver, which might
be too expensive to be practical, since, in addition, a propagation-based algorithm might call a domain filtering algorithm many times during its execution.
To overcome this difficulty we shall require further conditions on the matrix
M . To this end, we define the hypergraph associated to instance I to be the
hypergraph H with V (H) = V and E(H) = {Sj | 1 ≤ j ≤ m}.
Now, assume that H is a network hypergraph defined by a tree T . In this
case, one can use specific and more efficient methods like the network simplex
algorithm (see for example [1]) instead of a general purpose LP solver. However,
it is still possible to do better (and avoid the O(n) calls to the network simplex
algorithm) by transforming it into a maximum flow problem. This idea has been
used in [26] to obtain a domain filtering algorithm for the Sequence constraint.
6
More precisely, [26] deals with the particular case of network matrices defined
by a directed path. Our approach draws upon [26] and generalizes it to network
matrices defined by arbitrary trees. This is done as follows.
Let P be the incidence matrix of T . That is, let t1 , . . . , tm+1 be an arbitrary
ordering of nodes in T and define P to be the ((m + 1) × m)-matrix where Pi,j
is +1 if edge eSj starts at ti , −1 if eSj ends at ti , and 0 otherwise.
Let r be a m-ary (column) vector and let p be a path in T . We say that r is
the indicator vector of p if for every j = 1, . . . , m, rj is the polarity of eSj wrt.
p. The next observation follows directly from the definitions.
Observation 1 Let p be a path in T and let r be its indicator vector. Then, the
ith entry, (P r)i , of P r, is +1 if ti is the first node in p, −1 if ti is the last node
in p, and 0 otherwise.
The next two lemmas follow directly from the previous observation.
Lemma 1. P has full rank.
Proof. Let P ′ be the (m × m) matrix obtained by removing the last row
(corresponding to vertex tm+1 ) and consider the (m × m)-matrix Q such that
for every i = 1, . . . , m, the ith column of Q, which we shall denote as Q∗,i , is
the indicator vector of the unique path in T starting at ti and entding at tm+1 .
It follows from Observation 1 that P ′ Q is the identity matrix.
Then, since P has full rank we can obtain an equivalent system P M z = P a
by multiplying both sides of M z = a by P . Let N = P M and b = P a (see
Example 3).
Lemma 2. In every column of N one entry is +1, one entry is −1, and all the
other entries are 0.
Proof. It is only necessary to show that every column, M∗,k , k = 1, . . . , m+n of
M is the indicator vector of some directed path in T . If the variable corresponding
to the k-column is xi for some 1 ≤ i ≤ n then by construction M∗,k is the
indicator column of the path associated to vi in T . Otherwise, if the variable
corresponding to column k is yj for some 1 ≤ j ≤ m then M∗,k is the indicator
vector of the directed path containing only edge eSj .
Hence, matrix N is the incidence matrix of a directed graph G. Note that, by
definition, G contains an edge ek , k = 1, . . . , m + n for each variable zk and
a node uj , j = 1, . . . , m + 1 for each row in N (that is, for every node in T ).
Define the capacity of every edge ek to be ck . Then, feasible solutions of the
system correspond precisely to flows where every node uj has a supply/demand
specified by bj (more precisely, node uj has a demand of bj units if bj > 0 and
a suply of −bj units if bj < 0). It is well know that this problem can be reduced
to the (standard) maximum flow problem by adding new source and sink nodes
s, t and edges from s to uj with capacity bj whenever bj > 0 and from uj to t
with capacity −bj whenever bj < 0.
7
Example 3. Let I be a boolean instance with variables {v1 , . . . , v6 } and constraints: C1 = Among({v1 , v4 }, {1}, 0, 1), C2 = Among({v2 , v5 }, {1}, 0, 1),
C3 = Among({v3 , v6 }, {1}, 0, 1), C4 = Among({v1 , v3 , v5 }, {1}, 1, 1),
C5 = Among({v2 , v4 , v6 }, {1}, 1, 1).
The specific values for M , a and c of the ILP formulation of I are:
y1
C1 1
C2 0
M = C3
0
C4 0
C5 0
cT =
y2
0
1
0
0
0
y3
0
0
1
0
0
x1
1
0
0
1
0
x2
0
1
0
0
1
x3
0
0
1
1
0
x4
1
0
0
0
1
x5
0
1
0
1
0
1 1 1 1
1
1
1
1
x6
0
0
1
0
1
1
1
1
a=
1
1
1
Note that since mini = maxi for i = 4, 5 we did not need to add the slack
variables y4 and y5 . The hypergraph of this instance is precisely the hypergraph
H in Example 2. In particular, hi is the hypererdge corresponding to the scope
of constraint Ci for i = 1, . . . , 5. The matrix P obtained from the tree T defining
H is:
t0
t1
t2
s0
s1
r
h1
−1
0
0
0
0
1
h2
0
−1
0
0
0
1
h3
0
0
−1
0
0
1
h4
0
0
0
1
0
−1
h5
0
0
0
0
1
−1
Multiplying M and c by P we obtain:
y1
t0−1
t1 0
N = P M = t2
0
s0
0
s1 0
r
1
y2
0
−1
0
0
0
1
y3
0
0
−1
0
0
1
x1
−1
0
0
1
0
0
x2
0
−1
0
0
1
0
x3
0
0
−1
1
0
0
x4
−1
0
0
0
1
0
x5
0
−1
0
1
0
0
x6
0
0
−1
0
1
0
−1
−1
b = Pa =
−1
1
1
1
The feasible solutions of the previous LP correspond to the feasible flows of
the network in Figure 2a, where nodes s0 , s1 and r have a supply of one unit of
flow and nodes y0 , y1 , y2 have a demand of one unit of flow. Figure 2b contains
the result of transforming the network in Figure 2a to a (standard) max flow
problem. In both networks all edges have capacity 1.
It follows from this construction that for every 1 ≤ i ≤ n and every d = {0, 1},
there is a solution s of I with s(vi ) = d if and only if there is a saturating flow
(that is, a flow where all the edges leaving s or entering t are at full capacity)
such that the edge associated to xi carries d units of flow. Régin [30,31] has
shown that this later condition can be tested simultaneously for all 1 ≤ i ≤ n
and d ∈ {0, 1} by finding a maximal flow and computing the strongly connected
components of its residual graph. Finding a maximal flow of a network with
8
−1 s0
t0 1
−1 s1
t1 1
−1 r
t2 1
s
Fig. 2a: Network with
suppies/demands
s0
t0
s1
t1
r
t2
t
Fig. 2b: Standard flow
network obtained from (2a)
integral capacities can be done in time O(min(v 2/3 , e1/2 )e log(v 2 /e) log u) using
Goldberg and Rao’s algorithm [17] where v is the number of vertices, e is the
number of edges, and u is the maximum capacity of an edge. Computing the
strongly connected components of the residual graph takes O(v + e) time using
Tarjan’s algorithm [36]. By construction, the network derived by our algorithm
satisfies e ≤ n + 2m and v ≤ m + 1. Furthermore, it is not difficult to see that
u ≤ mn. Indeed, note that the capacity of any edge is either some entry, ci , of
vector c or the absolute value of some entry, bi , of vector b. It follows directly
from the definition of c, that all its entries are at most n. As for b, the claim
follows from the fact that b = P a where, by construction, a has m entries where
every entry is in the range {1, . . . , n} and every entry in P is in {−1, 0, 1}.
Hence, if we define f (n, m) to be min(m2/3 , (n + m)1/2 )(n + m) log(m2 /(n +
m)) log mn) we have:
Lemma 3. There is a domain filtering algorithm for conjunctions of boolean
among constraints whose associated hypergraph is a network hypergraph, which
runs in time O(f (n, m)) where n is the number of variables and m is the number
of constraints, assuming the instance is presented as a network flow problem.
It is customary, when analizying the time complexity of domain filtering to
report, additionally, the so-called time complexity ’down a branch of a search
tree’ which consists in the aggregate time complexity of successive calls to the
algorithm, when at each new call, the list of some of the variables has been
decreased (as in the execution of a propagation-search based solver). It was
observed again by Régin [31] that, in this setting, it is not necessary, to solve
the flow problem from scratch at each call, leading to a considerable redution
in total time. Applying the scheme in [31], we obtain that the time complexity
down a branch of a search tree of our algorithm is O(n(n + m)). We ommit the
details because they are faily standard (see [31]).
There are minor variants (leading to the same asymptotic complexity) obtained by modifying the treatment of the slack variables. Here we will discuss two of them. In the first variant, used in [26], one
P encodes a constraint
Cj = (Sj , {1}, minj , maxj ) with two equations yj + vi ∈Sj xi = maxj , and
P
−zj + vi ∈Sj xi = minj where yj and zj are new slack variables satisfying
0 ≤ yj , zj . This encoding produces a network that has m more nodes and
edges. P
In a second variant, one encodes a constraint Cj with the equation
−yj + vi ∈Sj xi = 0 where yj satisfies minj ≤ yj ≤ maxj . Under this encoding, our approach produces a network problem where, instead of having nodes
9
with specified suply or demand, we have edges with minimum demand. The
asymptotic time bounds in all variants are identical.
If the instance is not presented as a network flow problem then one would
need to add the cost of transforming the instance into it. However this cost
would be easily amortized as a domain filtering algorithm is invoked several
times during the execution of a constraint solver2 . Furthermore, in practical
scenarios, the conjunction of among constraints will encode a global constraint
from a catalog of available global constraints. Hence, it is reasonable to assume
that the formulation of the global constraint as a network flow problem can be
precomputed.
This approach can be generalized to non-boolean domains via boolean encoding. The choice of boolean encoding might depend on the particular instance
at hand but for concreteness we will fix one. The canonical booleanization (see
Example 4) of a conjunction I = (V, D, L, C) of among constraints with |D| ≥ 3
is the boolean instance (V × D,{0,1}, Lb, Cb ) where Lb (v, d) = {0, 1} if d ∈ L(v)
and {0} otherwise, and Cb contains:
– Among(S×R, {1}, min, max) for every constraint Among(S, R, min, max) ∈
C, and
– Among({v}×D, {1}, 1, 1) for every variable v ∈ V . This family of constraints
are called non-empty assignment constraints.
That is, the intended meaning of the encoding is that (v, d) ∈ V × D is true
whenever v takes value d.
We define the hypergraph H associated to I to be the hypergraph associated to the canonical booleanization of I. That is, V (H) is V × D and E(H)
contains hyperedge {(v, d) | v ∈ S, d ∈ R} = S × R for every constraint
Among(S, R, min, max) in C, and hyperedge {(v, d) | d ∈ D} = {v} × D for
every variable v ∈ V . Thus, for arbitrary domains, we have:
Corollary 1. There is a domain filtering algorithm for conjunctions (V, D, L, C)
of among constraints whose associatedP
hypergraph is a network hypergraph, which
runs in time O(f (n, m)) where n = v∈V |L(v)| and m = |C| + |V |, assuming
the instance is presented as a network flow problem.
Proof. It just follows from
Pobserving that the canonical booleanization of instance (V, D, L, C) has n = v∈V |L(v)| variables and m = |C| + |V | constraints.
6
Some applications
The aim of this section is to provide evidence that the kind of CACs covered
by the approach developped in Section 5 are often encountered in practice. To
2
In fact, as shown in [30,31], it is only necessary to solve the max flow problem during
the first invocation so it could be argued that a more realistic bound on the running
time of the algorithm is O(n + m).
10
this end, we shall revisit several families of CACs for which domain filtering
have been previously introduced and show how they can solved and, in some
cases generalized, using our algorithm. Furthermore, we shall compare, whenever
possible, the time complexity bounds of our algorithm with other state-of-the-art
algorithms for the same problem. Like other flow-based algorithms, the algorithm
proposed here has very good time complexity down a branch of the search tree.
However, we will only consider in our comparison the cost of calling the algorithm
just once. This is due to the fact that, once the size of the network produced
is under a certain threshold, the total cost down a branch of the search tree is
dominated by the cost of the incremental updates and, hence, it cannot be used to
assess the comparative quality of different flow-based algorithms. Furthermore,
we will try to compare, whenever possible, the parameteres of the obtained
network flow problem (number of nodes, edges, capacities of the edges) instead
of the actual running time since the latter is dependend on the choice of the maxflow algorithm. Somewhat surprisingly, in many of the cases, even if we did not
attempt any fine-tuning, the network produced by the algorithm is essentially
equivalent to the network produced by specific algorithms.
6.1
Disjoint constraints
As a warm up we shall consider the GCC and AllDiff constraints. We have
seen in Example 1 that both can be formulated as a conjunction (V, D, L, C)
of among constraints of the form Cd = Among(V, {d}, mind, maxd ), d ∈ D.
Note that both have the same associated hypergraph H with node-set, V × D,
and edge set E(H) containing hyperedge hv = {v} × D for every v ∈ V , and
hyperedge hd = V × {d}, for every d ∈ D.
It is not difficult to see (see Example 4) that H is defined by the tree T
defined as follows: The node-set of T consists of {r} ∪ V ∪ D where r is a new
node. The edge-set of T contains an edge from every v ∈ V to r (associated
to hv ) and from r to every node d ∈ D (associated to hd ). Consequently, both
GCC and AllDiff are solvable by our algorithm. The network for the GCC
using the abovementioned tree T has e = O(|V ||D|) edges, v = |V | + |D| + 3
nodes, and the maximum capacity, u, of an edge is at most |D||V |. Hence, it
follows that the total running time of our algorithm for the GCC constraint
is O(min(v 2/3 , e1/2 )e log(v 2 /e) log u). Régin’s algorithm [31] has a O(|V |2 |D|)
complexity which is better when |V | ∈ O(|D|2/3 ) but the comparison between the
two bounds is not very meaningfull because it mainly reflects a different choice
max flow algorithm. Indeed, the network produced by both algorithms are very
similar. In particular, the network obtained using the second variant discussed
after Lemma 3 is essentially the same described in [31]. The only difference
is that the network obtained by our algorithm contains one extra node. In the
particular case of the AllDiff constraint, [30] shows how to produce a bipartite
matching problem that can be solved using specialized algorithms, such as [20],
leading to a total time complexity of O(|V |5/2 ) which is better than ours.
11
Example 4. Consider the constraint AllDiff(s0 , s1 ) where the list of each variable contains the following three values: t0 , t1 , t2 . Then, AllDiff(s0 , s1 ) is encoded as a CAC I with the following constraints: Among({s0 , s1 }, {t0 }, 0, 1),
Among({s0 , s1 }, {t1 }, 0, 1), Among({s0 , s1 }, {t2 }, 0, 1).
The canonical booleanization of I has variables {s0 , s1 } × {t0 , t1 , t2 } and
constraints C1 = Among({s0 , s1 } × {t0 }, {1}, 0, 1), C2 = Among({s0 , s1 } ×
{t1 }, {1}, 0, 1), C3 = Among({s0 , s1 } × {t2 }, {1}, 0, 1), C4 = Among({s0 } ×
{t0 , t1 , t2 }, {1}, 1, 1), and C5 = Among({s1 }×{t0 , t1 , t2 }, {1}, 1, 1). Observe that
this instance is, under the renaming vi 7→ (si−1 mod 2 , ti−1 mod 3 ), the same
instance than we have considered previously in Example 3. The network flow
problem that our algorithm derives for this instance (see Example 3) is almost
identical to the one derived in [31]. Indeed, the network obtained in [31] does
not have node r and, instead, requires that the demand of nodes t0 , t1 , t2 is at
most one (instead of exactly one).
A simple analysis reveals that the same approach can be generalized to instances (V, D, L, C) satisfying the following disjointedness condition: for every
pair of constraints Among(S, R, min, max) and Among(S ′ , R′ , min′ , max′ ) in
C, (S × R) ∩ (S ′ × R′ ) = ∅. The tractability of such instances was, to the best
of our knowledge, not known before. The particular case in which R ∩ R′ = ∅
has been previously shown in [32] using a different approach. The proof given in
[32] does not construct a flow problem nor gives run-time bounds so we ommit
a comparison.
6.2
Domains consisting of subsets
Consider the following generalization of our setting where in a CAC (V, D, L, C)
every variable v must be assigned to a subset of L(v) (instead of a single element).
In this case, the semantics of the among constraint need to be generalized as
well. Instead, we will say a constraint
Among(S, R, min, max) is satisfied by a
P
mapping s : V → 2D if min ≤ v∈S |s(v) ∩ R| ≤ max. To avoid confusion we
shall refer to this variant of the among constraint as set among constraint.
For example, the Symmetric-GCC constraint [21] is precisely a conjunction (V, D, L, C) of set among constraints of the form Among(V, {d}, min, max)
where d is a singleton which, additionally, might contain constraints of the form
Among({v}, D, min, max) restricting the size of the image of a variable v.
It is fairly easy to reduce a conjunction of set among constraints I = (V, D, L, C)
to a conjunction of (ordinary) among constraints over a boolean domain. Indeed, one only needs to construct the instance (V × D,{0,1}, Lb, Cb ) where
Lb (v, d) = {0, 1} if d ∈ L(v) and {0} otherwise, and Cb contains Among(S ×
R, {1}, min, max) for every constraint Among(S, R, min, max) ∈ C. Note that
the instance thus constructed corresponds exactly to the result of removing the
non-empty assignment constraints to the canonical booleanization of I (now
regarded as a conjunction of ordinary among constraints). It is then easy to
observe that if (V, D, L, C) encodes a Symmetric-GCC constraint then the resulting boolean instance has the same hypergraph, H, than the GCC constraint
12
and hence, H is a network matrix. The algorithm in [21] follows closely that
in [31] for the GCC constraints, and, in particular, has the same time bounds.
Consequently the network flow derived by our algorithm is obtained, again, by
adding one extra node with small capacities in the edges to the network introduced in [21].
6.3
The sequence constraint
The Sequence constraint [6] corresponds to instances ({v1 , . . . , vn }, D, L, C)
with constraints Among({vi , . . . , vi+k }, R, min, max), i = 1, . . . , n − k for some
fixed integers min, max, k, and fixed R ⊆ D. It is not difficult to see that the
hypergraph of the canonical booleanization of the Sequence constraint is not
a network hypergraph. However, as shown in [26] one obtains an equivalent instance with a network hypergraph using a different encoding in which for every
original variable vi ∈ V , we have a boolean variable xi which is intended to
be true whenever vi takes a value in R and false otherwise. Under this alternative encoding we obtain a boolean instance I which consists of constraints
Among({xi , . . . , xi+k }, {1}, min, max), i = 1, . . . , n − k. It is shown in [26] that
the hypergraph H of the boolean instance I obtained under this encoding satisfies the so-called consecutive-ones property which implies that H is defined
by a tree T consisting of a single directed path. Indeed, the network flow obtained by our approach is identical to the one derived in [26] if one encodes
Among constraints using the first variant discussed after Lemma 3. Applying
Lemma 3 and noting that, in the particular case of the Sequence constraint,
we have m = O(n) we obtain the bound O(n3/2 log2 n). By inspecting closely
the proof of Lemma 3 this bound can be slightly improved (see Appendix C)
to O(n3/2 log n log max) coinciding with the bound given in [26], which is not
surprising since both networks are essentially equivalent. To the best of our
knowledge O(n3/2 log n log max) is the best bound among all complete domain
consisteny algorithms for the problem, jointly with the algorithm proposed in
[37] which, with time complexity O(n2k ), offers gives better bounds when k ≪ n.
6.4
TFO model
The TFO model was introduced by Razgon et al. [29] as a generalization of
several common global constraints. Formally, a TFO model is a triple (V, F1 , F2 )
where V is a finite set of vertices and F1 and F2 are nonempty families of subsets
of V such that two sets that belong to the same family are either disjoint or
contained in each other. Each set Y in F1 ∪ F2 is associated with two nonnegative integers minY , maxY ≤ |Y |. A subset X of V is said to be valid if
minY ≤ X ∩ Y ≤ maxY for every Y ∈ F1 ∪ F2 . The task is to find the largest
valid subset. Although the methods introduced in the present paper can be
generalized to deal as well with optimization version we will consider only now
the feasibility problem consisting in finding a valid subset (or report that none
exists).
13
First, note that the existence of a valid subset in a TFO model can be
formulated naturally as a satisfiability problem for a combination of among constraints. Indeed, there is a one-to-one correspondence between the valid subsets
of (V, F1 , F2 ) and the solutions of the instance (V, {0, 1}, L, C) where C contains
the constraints Among(Y, {1}, minY , maxY ), Y ∈ F1 ∪ F2 , and the list L(v) of
every variable v ∈ V is {0, 1}. The hypergraph H associated to this instance is
(V, F1 ∪ F2 ). It follows directly from Lemmas 4 and 5 (see Appendix) that H is a
network hypergraph and hence one can use our approach to decide the existence
of a feasible solution of a TFO model. It turns out that the network introduced
in [29] is essentially equivalent to the network flow problem that would be obtained by our approach using the second variant described after Lemma 3. It is
not meaningfull to compare the running time of our algorithm with that of [29]
since it deals with an optimization variant.
6.5
Conjunction of among constraints with full domain
Some global constraints studied in the literature correspond to conjunctions
(V, D, L, C) of among constraints where the scope of every constraint is the full
set V of variables. This class contains, of course, the GCC constraint and also
several others, since we do not require R to be a singleton. For example, the
OrderedDistribute constraint introduced by Petit and Régin [28] can be
encoded as conjunction (V, D, L, C) of among constraint where where the domain
D has some arbitrary (but fixed) ordering d1 , . . . , d|D| and in every constraint
Among(S, R, min, max), S = V and R is of the form {di , . . . , d|D| }.
We shall show that the hypergraph of the conjunction of among constraints
defining OrderedDistribute is a network hypergraph. Indeed, with some extra work we have managed to completely characterize all CAC instances containing only constraints with full scope that have an associated network hypergraph.
Theorem 1. Let I = (V, D, L, C) be a conjunction of among constraints with
|D| ≥ 3 such that the scope of each constraint is V . Then, the following are
equivalent:
1. The hypergraph of the canonical booleanization of I is a network hypergraph.
2. For every pair of constraints in C, their ranges are disjoint or one of them
contained in the other.
In the particular case of OrderedDistribute constraint, the network obtained by our approach is very related to the network introduced in section ([28],
Section V.A). More precisely, the abovementioned network is essentially equivalent to the network that would be obtained by our approach using the second
variant described after Lemma 3. However, our algorithm is far from optimal.
In particular, a complete filtering algorithm with time complexity O(|V | + |D|)
is given also in [28].
14
6.6
Adding new among constraints to a GCC constraint
Let (V, D, L, C) be a conjunction of among constraints encoding the GCC constraint (see Example 1) and assume that we are interested in adding several new
among constraints to it. In general, we might end up with a hard instance but
depending on the shape of the new constraints we might perhaps still preserve
tractability. Which among constraint we might safely add? This question has
been addressed by Régin [32]. In particular, [32] shows that the domain filtering
problem is still tractable whenever:
(a) every new constraint added has scope V and, furthermore, the ranges of
every pair of new constraints are disjoint, or
(b) every new constraint added has range D and furthermore, the scopes of every
pair of new constraints are disjoint.
We can explore this question by inquiring which families of constraints can
be added to an instance (V, D, L, C) encoding GCC such that its associate hypergraph is still a network hypergraph. Somewhat surprisingly we can solve
completely this question (see Theorem 2). This is due to the fact that the presence of the global cardinality constraint restricts very much the shape of the tree
defining the hypergraph of the instance.
Theorem 2. Let I = (V, D, L, C) be a conjunction of among constraints containing a global cardinality constraint with scope V with |D| ≥ 3. Then the
following are equivalent:
1. The hypergraph of the canonical booleanization of I is a network hypergraph.
2. In every constraint in C, the scope is a singleton or V , or the range is a
singleton or D. Furthermore, for every pair Among(S1 , R1 , min1 , max1 ),
Among(S2 , R2 , min2 , max2 ) of constraints in C the following two conditions
hold:
(a) If S1 = S2 = V or S1 = S2 = {v} for some v ∈ V then R1 and R2 are
disjoint or one of them is contained in the other.
(b) If R1 = R2 = D or R1 = R2 = {d} for some d ∈ D then S1 and S2 are
disjoint or one of them is contained in the other.
Note that the previous theorem covers cases (a) and (b) from [32] described
at the beginning of this section. The network produced in [32] for the case (a)
is essentially equivalent to the one derived by our approach using the second
variant described after Lemma 3. For the case (b) [32] does not construct a flow
problem nor gives run-time bounds so we ommit a comparison.
Acknowledgments
The author would like to thank the anonimous referees for many useful comments. This work was supported by the MEIC under grant TIN2016-76573-C21-P and the MECD under grant PRX16/00266.
15
References
1. Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Network flows theory, algorithms and applications. Prentice Hall, 1993.
2. Gautam Appa, Dimitris Magos, and Ioannis Mourtos. LP relaxations of multiple
all different predicates. In Proceedings of CPAIOR’04, pages 364–369, 2004.
3. Gautam Appa, Dimitris Magos, and Ioannis Mourtos. On the system of two
all different predicates. Inf. Process. Lett., 94(3):99–105, 2005.
4. Fahiem Bacchus and Toby Walsh. Propagating logical combinations of constraints.
In Proceedings of IJCAI’05, pages 35–40, 2005.
5. R. Balakrishnan and K Ranganathan. A Textbook of Graph Theory. SpringerVerlag (New York), 2000.
6. N. Beldiceanu and E. Contejean. Introducing global constraints in chip. Mathematical and Computer Modelling, 12:97–123, 1994.
7. Christian Bessiere, Emmanuel Hebrard, Brahim Hnich, Zeynep Kiziltan, and Toby
Walsh. SLIDE: A useful special case of the CARDPATH constraint. In Proceedings
of ECAI’08, pages 475–479, 2008.
8. Christian Bessiere, Emmanuel Hebrard, Brahim Hnich, Zeynep Kiziltan, and Toby
Walsh. Range and roots: Two common patterns for specifying and propagating
counting and occurrence constraints. Artif. Intell., 173(11):1054–1078, 2009.
9. Christian Bessiere, George Katsirelos, Nina Narodytska, Claude-Guy Quimper, and
Toby Walsh. Propagating conjunctions of alldifferent constraints. In Proceedings
of AAAI’10, 2010.
10. Gilles Chabert and Sophie Demassey. The conjunction of interval among constraints. In Proceedings of CPAIOR’12, pages 113–128, 2012.
11. R. Dechter and J. Pearl. Tree clustering for constraint networks. Artificial Intelligence, 38:353–366, 1989.
12. Tommy Färnqvist and Peter Jonsson. Bounded tree-width and csp-related problems. In Proceedings of ISAAC’07, pages 632–643, 2007.
13. Michael R. Fellows, Tobias Friedrich, Danny Hermelin, Nina Narodytska, and
Frances A. Rosamond. Constraint satisfaction problems: Convexity makes alldifferent constraints tractable. Theor. Comput. Sci., 472:81–89, 2013.
14. Jörg Flum and Martin Grohe. Parameterized Complexity Theory. Texts in Theoretical Computer Science. An EATCS Series. Springer, 2006.
15. E.C. Freuder. Complexity of k-structurated constraint satisfaction problems. In
Proceedings of AAAI-90, pages 4–9, 1990.
16. M. R. Garey and David S. Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. W. H. Freeman, 1979.
17. Andrew V. Goldberg and Satish Rao. Beyond the flow decomposition barrier. J.
ACM, 45(5):783–797, 1998.
18. Georg Gottlob, Nicola Leone, and Francesco Scarcello. A comparison of structural
CSP decomposition methods. Artif. Intell., 124(2):243–282, 2000.
19. Martin Grohe. The complexity of homomorphism and constraint satisfaction problems seen from the other side. J. ACM, 54(1):1:1–1:24, 2007.
20. John E. Hopcroft and Richard M. Karp. An n5/2 algorithm for maximum matchings in bipartite graphs. SIAM J. Comput., 2(4):225–231, 1973.
21. Waldemar Kocjan and Per Kreuger. Filtering methods for symmetric cardinality
constraint. In Proceedings of CPAIOR’04, pages 200–208, 2004.
22. Martin Kutz, Khaled M. Elbassioni, Irit Katriel, and Meena Mahajan. Simultaneous matchings: Hardness and approximation. J. Comput. Syst. Sci., 74(5):884–897,
2008.
16
23. Frédéric Lardeux, Eric Monfroy, and Frédéric Saubion. Interleaved alldifferent
constraints: CSP vs. SAT approaches. In Proceedings of AIMSA’08, pages 380–
384, 2008.
24. Jean-Louis Laurière. A language and a program for stating and solving combinatorial problems. Artif. Intell., 10(1):29–127, 1978.
25. Dimitris Magos, Ioannis Mourtos, and Gautam Appa. A polyhedral approach to
the alldifferent system. Math. Program., 132(1-2):209–260, 2012.
26. Michael J. Maher, Nina Narodytska, Claude-Guy Quimper, and Toby Walsh. Flowbased propagators for the SEQUENCE and related global constraints. In Proceedings of CP’08, pages 159–174, 2008.
27. A. Oplobedu, J. Marcovitch, and Y. Toubier. Charme: Un langage industriel de
programmation par contraintes, illustrṕar une application chez renault. In Proceedings of 9th International Workshop on Expert Systems and their Applications,
pages 55–70, 1989.
28. Thierry Petit and Jean-Charles Régin. The ordered distribute constraint. International Journal on Artificial Intelligence Tools, 20(4):617–637, 2011.
29. Igor Razgon, Barry O’Sullivan, and Gregory M. Provan. Generalizing global constraints based on network flows. In Proceedings of CSCLP’07, pages 127–141, 2007.
30. Jean-Charles Régin. A filtering algorithm for constraints of difference in csps. In
Proceedings of AAAI’94, pages 362–367, 1994.
31. Jean-Charles Régin. Generalized arc consistency for global cardinality constraint.
In Proceedings of AAAI’96, pages 209–215, 1996.
32. Jean-Charles Régin. Combination of among and cardinality constraints. In Proceedings of CPAIOR’05, pages 288–303, 2005.
33. Jean-Charles Régin and Carla P. Gomes. The cardinality matrix constraint. In
Proceedings of CP’04, pages 572–587, 2004.
34. Jean-Charles Régin and Jean-Francois Puget. A filtering algorithm for global sequencing constraints. In Proceedings of CP’97, pages 32–46, 1997.
35. Alexander Schrijver. Theory of Linear and Integer Programming. Wiley, 1998.
36. Robert Endre Tarjan. Depth-first search and linear graph algorithms. SIAM J.
Comput., 1(2):146–160, 1972.
37. Willem Jan van Hoeve, Gilles Pesant, Louis-Martin Rousseau, and Ashish Sabharwal. Revisiting the sequence constraint. In Proceedings of CP’06, pages 620–634,
2006.
38. Willem Jan van Hoeve, Gilles Pesant, Louis-Martin Rousseau, and Ashish Sabharwal. New filtering algorithms for combinations of among constraints. Constraints,
14(2):273–292, 2009.
Appendix A: Tractable subcases of the domain filtering
problem restricting only the scope or the range in the
constraints
If I is a set of conjunctions of among constraints, we shall denote by DomFilter(I)
the restriction of the domain filtering problem to instances in I. Our ultimate
goal would be to characterize precisely for which sets I, DomFilter(I) has
efficient algorithms.
The first question that we address is the following: which subcases of the
problem can be explained by considering only the scopes of the constraints? The
17
scopes occurring in an instance can be characterized by an hypergraph. More
precisely, let I = (V, D, L, C) be a conjunction of among constraints. The scope
hypergraph of I is the hypergraph H with node set V (H) = V and that contains
an hyperedge for every scope occurring in a constraint in C. Formally,
E(H) = {S | Among(S, R, min, max) ∈ C}
Let H be a (possibly infinite) set of hypergraphs. We denote by Scope(H) the
collection of instances I of the domain filtering problem whose scope hypergraph
belongs to H. Our question can be formalized in the following manner: for which
sets, H, of hypergraphs, is DomFilter(Scope(H)) efficiently solvable? Note that
if I is an instance whose scope hypergraph H is a subhypergraph of some H ′ ∈ H
then we can construct, by adding superfluous new constraints to I, an equivalent
new instance whose scope hypergraph is H ′ . Hence, we can assume that H is
closed under taking subhypergraphs.
We shall solve completely this question assuming some mild technical assumptions. In order to state our result we need a few definitions from graph
theory and parameterized complexity. The Gaifman graph of a hypergraph H,
denoted Gaifman(H) is the graph where the node-set is V (H) and the edge-set
contains all pairs {u, v} such that there is an hyperedge h in H with {u, v} ⊆ h.
A tree-decomposition of a graph G is a pair (T, β) where T is an (ordinary,
not oriented) tree and β : V (T ) → 2V (G) is a mapping such that the following
conditions are satisfied:
1. For every node v ∈ V (G), the set {x ∈ V (T ) | v ∈ β(x)} is non-empty and
connected in T .
2. For every edge {u, v} ∈ E(G), there is a node x ∈ V (T ) such that {u, v} ∈
β(x).
The width of a tree-decomposition (T, β) is max{|β(x)| − 1 | x ∈ V (T )} and
the tree-width of G is defined to be the minimum w such that G has a treedecomposition of width w.
We will also need some notions and basic facts from parameterized complexity theory. A parameterized problem over some alphabet Σ is a pair (P, κ)
consisting of a problem P ⊆ Σ ∗ and a polynomial time mapping κ : Σ ∗ → N,
called its parameter. A parameterized problem (P, κ) over Σ is fixed-parameter
tractable if there is a computable function f : N → N and an algorithm that
decides if a given instance x ∈ Σ ∗ belongs to P in time f (κ(x)) · |x|O(1) . FPT
denotes the class of fixed-parameter tractable problems. Hence, the notion of
fixed-parameter tractability relaxes the classical notion of polynomial-time solvability, by admitting running times that are exponential in the parameter, which
is expected to be small.
The analogous of NP in parameterized complexity is the class W[1] which is
conjectured to contain strictly FPT. We will omit the definition of W[1] since it
is not needed in our proofs and refer the reader to [14].
Theorem 3. Assume FPT 6= W[1]. For every recursively class of hypergraphs
H, DomFilter(Scope(H)) is polynomial-time solvable if and only if there exists
18
some natural number w such that the tree-width of the Gaifman graph of every
hypergraph in H is at most w.
Proof. It is well know [11,15] that the set of all CSP instances whose scope
hypergraph has tree-width at most w for some fixed w is solvable in polynomial
time. Hence it only remains to show the ’only if’ part. This follows from a result
of Färnquivst and Jonsson [12]. In order to state it, we need to introduce some
background.
In the list homomorphism problem (for graphs), we are given two graphs G,
J and a mapping L : V (G) → 2V (J) called list. The goal is to decide whether
there exists a mapping h : V (G) → V (J) satisfying the following two conditions:
1. h(v) ∈ L(v) for every v ∈ V (G).
2. (h(v1 ), h(v2 )) ∈ E(J) for every (v1 , v2 ) ∈ E(G).
Such a mapping is called a solution of (G, J, L). Färnquivst and Jonsson have
shown that the parameterized version of the problem (parameterized by the size
of V (G)) is W[1]-hard. Indeed, the problem is W[1]-hard even if the input graph,
G, is guaranteed to belong to a previously fixed G of graphs, provided G has unbounded tree-width. Formally, let G be a set of graphs and define p-LHom(G, )
to be the problem:
– INPUT: graphs G, J with G ∈ G, and a mapping L : V (G) → 2V (J) .
– PARAMETER: |V (G)|.
– GOAL: Decide whether (G, J, L) has a solution.
We are finally ready to state the theorem from [12] that we shall use.
Theorem 4. (Lemma 3 in [12]) Let C be a recursively class of graphs that does
not have bounded tree-width. Then p-LHom(C, ) is W[1]-hard.
We note here that, in order to simplify the exposition, we have taken the
liberty to adapt the statement in [12]. We are now ready to complete our proof.
Assume, towards a contradiction, that H is a set of hypergraphs with unbounded
tree-width such that DomFilter(Scope(H)) is solvable in polynomial time. Let
G = {Gaifman(H) | H ∈ H}. It follows directly by the assumptions on H that G
has unbounded tree-width and is recursively enumerable. We shall give an FPT
algorithm for p-LHom(G, ).
The algorithm is as follows. Let (G, J, L) be any instance of the list homomorphism problem with G ∈ G. Enumerate the hypergraphs in H until
finding an hypergraph H ∈ H whose Gaifman graph is G. Construct the instance I in Scope(H) where the set of variables is V (H)(= V (G)), the domain is
V (H) × V (J), the list of every node v ∈ V (H) is {v} × L(v), and the constraints
are defined as follows:
– For every (v1 , v2 ) ∈ E(G), and for every (a1 , a2 ) 6∈ E(H), include the constraint Among(h, R, 0, 1) where h is any hyperedge in H containing {v1 , v2 }
and R = {(v1 , a1 ), (v2 , a2 )}.
19
Note that, by construction, the scope hypergraph of I is a subhypergraph of H
and, hence, I is an instance of Scope(H). Since DomFilter(Scope(H)) is, by assumption, solvable in polynomial time then one can also decide the satisfiability
of I is polynomial time. Finally, return ’yes’ if I is satisfiable and ’no’ otherwise. Note that the time required in finding H depends only on G whereas the
time required in constructing and solving I is polynomial on the size of (G, J, L).
Hence, the algorithm just defined is FPT. It is easy to see that it correctly solves
(G, J, L). Indeed, let s be any mapping s : V (H) → V (H) × V (J) which we can
write as s(v) = (s1 (v), s2 (v)) with s1 : V (H) → V (H) and s2 : V (H) → V (J).
It follows directly from the construction of I that s is a solution of I if and only
if s1 is the identity (formally, s(v) = v for every v ∈ V (H)) and s2 is a solution
of instance (G, J, L). Hence, I is satisfiable if and only if so is (G, J, L).
It is not difficult to show, following [19], that the condition FPT 6= W[1]
cannot be weakened (for example by requiring, instead, only P6=NP). Indeed, if
FPT = W[1] then there exists a family H of hypergraphs of unbounded treewidth such that DomFilter(H) is polynomial-time solvable.
Finally, note that the hardness part of Theorem 3 holds even for conjunctions
of among constraint whose range has cardinality 2. Note, that the hardness
direction does not hold any more if one requires that the among constraints
have range of cardinality 1. Indeed, the AllDiff constraint is encoded by a
conjunction of among constrains whose scope hypergraph can have arbitrary
large tree-width and the range of every constraint is a singleton.
Secondly, we turn our attention to the range and investigate which restrictions on the range of among constraints guarantee that the domain filtering
problem is solvable in polynomial time. To this end we can define the range
hypergraph in a similar way to the scope hypergraph and define, for every set H
of hypergraphs, Range(H) to be the set of all conjunctions of among constraints
whose range hypergraph belongs to H. The next theorem, which is straightforward, shows that all non-trivial hypergraphs give rise to hard problems.
Theorem 5. Assume P6=NP. For every hypergraph H, DomFilter(Range(H))
is polynomial-time solvable if and only every hypergraph H ∈ H has only trivial
hyperedges (that is, if for every H ∈ H and every h ∈ E(H), h = ∅ or h =
V (H)).
Proof. The ’if’ direction is trivial. Indeed, if the range hypergraph of an instance
I has only trivial hyperedges it follows that every among constraint in it is either
superfluous (in the sense that it does not enforce any restriction) or unsatisfiable.
For the ’only if’ direction, assume that H has an hypergraph H with a non
trivial hyperedge h. We define a reduction from One-In-Three Sat which is
the following NP-complete [16] problem:
– INPUT: An hypergraph J where all the hyperedges have cardinality 3.
– GOAL: Decide whether there exists some X ⊆ V (J) such that h ∩ X = 1
for every h ∈ E(J).
20
The reduction is as follows: given an instance J, of One-In-Three Sat, construct the instance I = (V, D, L, C) ∈ Range(H) where V = V (J), D = V (H),
L(v) = V (H) for every v ∈ V , and C contains the constraints (S, h, 1, 1) for
every S ∈ E(J). It is easy to see that the range hypergraph of I is H and that
J is satisfiable if and only if so is I.
Appendix B: Proofs of section 6
Basic concepts and results about network hypergraphs
If T is an oriented tree and x ∈ V (T ) we denote by T − (x) (respectively T + (x))
the subtree of T containing all those nodes appear in some directed path ending
(respectively, starting) at x.
A rooted tree is an oriented tree that is obtained from un undirected tree by
fixing a node r, called the root, and orienting all the edges away from the root
or all the edges towards the root.
Lemma 4. For every hypergraph H the following are equivalent:
1. Every pair of hyperedges in H are either disjoint or contained in each other.
2. There exists a rooted tree T that defines H such that the path associated to
every variable ends at the root.
Furthermore, when condition (2) holds we can assume that in every edge e =
(x, y) associated to a minimal hyperedge of H, x has in-degree zero.
Proof. (1 ⇒ 2). Let H be an hypergraph satisfying (1). Let T be the oriented
tree defined in the following way. The node-set of T is E(H) ∪ {r} where r is a
fresh node. For every h ∈ H(T ), T contains an edge from h to h′ where h′ = r if
h is not contained in any other hyperedge of H and h′ is the smallest hyperedge
in H containing h otherwise. Note that T has root r and that if e = (x, y) is an
edge associated to a minimal hyperedge of H, then x has in-degree zero. It is not
difficult to verify that T satisfies condition (2). (2 ⇒ 1). Assume that T is a tree
satisfying condition (2). Let h and h′ be hyperedges in H and let eh = (x, y) and
eh′ = (x′ , y ′ ) be their associated edges in T . Consider three cases: (a) there is a
directed path from x to x′ , (b) there is a directed path from x′ to x, (c) none of
the previous holds. In case (a) it follows from the fact that the path associated
to every variable ends at the root that every path containing eh contains also
eh′ and hence h ⊆ h′ . By the same argument, case (b) implies that h′ ⊆ h.
Finally, in case (c) there is no directed path containing both eh and eh′ and,
consequently, h ∩ h′ = ∅.
Oriented trees can be combined by gluing some of their nodes. Formally,
let T1 and T2 be trees and let r1 and r2 be nodes in T1 and T2 respectively.
Assume, renaming nodes if necessary, that V (T1 ) ∩ V (T2 ) = ∅. Then, the result
of gluing r1 and r2 is obtained by, first, computing the disjoint union of T1 and
T2 and, then, merging r1 and r2 into a new node w that has as in-neighbours
the union of all in-neighbours of r1 and r2 and, as out-neighbours, the union of
21
all out-neighbours of r1 and r2 . It is not difficult to see that the result is again
an oriented tree.
The following technical lemma, which follows directly from the definitions,
will be useful.
Lemma 5. For i = 1, 2, let Hi be an hypergraph, let Ti be a tree defining Hi
and let ri ∈ V (Ti ). Assume that V (H1 ) = V (H2 ) and that for every v ∈ V (H1 )
the following holds: if v ∈ h1 ∩ h2 with h1 ∈ H1 and h2 ∈ H2 then the last node
of the path associated to v in T1 is r1 and the first node of the path associated
to v in T2 is r2 . Then, the result of gluing r1 and r2 defines H1 ∪ H2 .
Proof. Straightforward.
Lemma 6. Let H be an network hypergraph defined by tree T and let J be a
subhypergraph of H with the property that for every a, b ∈ E(J) there exists some
c ∈ J with a ∩ c 6= ∅ and b ∩ c 6= ∅. Then there exists an element r ∈ V (T ) such
that for every h ∈ E(J), eh ∈ E(T − (r)) ∪ E(T + (r))
Proof. For every h ∈ H1 ∪ H2 , let eh = (x, y) be its associated edge in T and
let us define Th to be the subtree of T with node-set V (T − (x)) ∪ V ((T + (y)).
It follows that V (Ta ) ∩ V (Tb ) 6= ∅ for every a, b ∈ J. Indeed, let c be the
hyperedge in J such that a ∩ c 6= ∅ and b ∩ c 6= ∅. Then both endpoints of the
edge ec associated to c belong to V (Ta ) ∩ V (Tb ).
Next we shall use the 2-Helly property of the subtrees of a tree (see for
example [5]).
Lemma 7. (2-Helly property of subtrees) Let T1 , . . . , Tn be a collection of subtrees of an (undirected)
tree T such that for every 1 ≤ i, j ≤ n, V (Ti )∩V (Tj ) 6= ∅.
T
It follows that 1≤i≤n V (Ti ) 6= ∅.
Note that the 2-Helly property stated deals with undirected trees (instead
of oriented trees). However it easily implies
that the same property holds for
T
V
(Th ) 6= ∅. To complete the proof
oriented trees. Hence, it follows
that
h∈J
T
note that any vertex r in h∈J V (Th ) satisfies the conditions of the Lemma.
Lemma 8. Let I = (V, D, L, C) be a conjunction of among constraints with
|D| ≥ 3, let H be the hypergraph of its canonical booleanization, let H1 be a set
of at least 2 hyperedges in E(H) where every hyperedge, S × R, in H1 satisfies
S = V and let H2 to be the subset of E(H) containing, for every v ∈ V , the
hyperedge {v} × D. Then, if H is a network hypergraph then there exists a tree
T defining H and a node r ∈ V (T ) such that for every h ∈ H1 , eh ∈ E(T − (r))
and for every h ∈ H2 , eh ∈ E(T + (r)).
Proof. Assume that H is a network hypergraph and let T be a tree defining
it. We shall use the following claim.
Claim 1 Let x1 , e1 , x2 , . . . , en−1 , xn be a directed path in T , and let ei , ej , ek ,
i < j < k be different edges such two of their associated hyperedges belong to H2
and the remaining one to H1 . Then, the edge whose associated hyperedge belongs
to H1 is ej .
22
Proof. Let hi , hj , hk be the hyperedges associated to ei , ej , and ek respectively
and assume, towards a contradiction, that hi ∈ H1 (the case hk ∈ H1 is symmetric). Let (v, d) ∈ hi ∩ hk and let p be its associated path. Clearly, p, contains
both ei and ek and hence it must contain also ej in contradiction with the fact
that hj ∩ hk = ∅
For every directed path p = x1 , . . . , xn define T (p) to be the subtree of T
induced by V (T − (x1 ))∪{x1 , . . . , xn }∪V (T + (xn )). We define J − (resp. J + ) to be
the set containing all h ∈ H1 ∪H2 with eh ∈ E(T − (x1 )) (resp. eh ∈ E(T + (xn ))).
The proof proceeds by showing that there exists a tree T defining H and a
directed path p in T satisfying all the following properties:
1.
2.
3.
4.
eh ∈ E(T (p)) for every h ∈ H1 ∪ H2 .
J − ∩ H1 = ∅ or J − ∩ H2 = ∅. Also, J + ∩ H1 = ∅ or J + ∩ H2 = ∅.
J − and J + are non empty.
J + = H2 .
Note that if T and p satisfy (1) − (4) then T and r = xn satisfy the Lemma.
Let us show the existence of T and p satisfying the above-mentioned properties in increasing order:
(1). First, note that J = H1 ∪ H2 satisfies the hypothesis of Lemma 6 and,
hence, it follows that there exists an element r in T such that for every h ∈
H1 ∪ H2 , eh belongs to T − (r) or T + (r). Hence, the path p consisting of single
node r satisfies (1).
(1) → (2). We claim that every path p = x1 , . . . , xn of maximal length
satisfying (1) must satisfy (2) as well. Assume, towards a contradiction, that
ai ∈ J − ∩ Hi for i = 1, 2 (the proof for J + is analogous). For every h ∈ J − let
qh be a directed path in T containing eh and ending at x1 . It follows that the
last edge in qa1 and qa2 must be identical since otherwise there could not be a
directed path containing both ea1 and ea2 , which would imply that a1 ∩ a2 = ∅.
Let (y, x1 ) be the common last edge of qa and qb . Then, applying the same
argument it follows that (y, x1 ) is also the last edge of qh for every h ∈ J − .
Hence, by adding y at the beginning of y, x1 , . . . , xn we obtain another path p
satisfying (1) contradicting the maximality of p.
(1, 2) → (3). It is easy to see that every path p satisfying (1) and (2) has a
subpath that satisfies, additionally, (3).
(1, 2, 3) → (4). It follows from Claim 1 that J − ∩ H2 6= ∅ or J + ∩ H2 6= ∅.
We can assume, by reversing the direction of the edges in T if necessary, that
J + ∩ H2 6= ∅. Since, by (2), J + ∩ H1 = ∅, in order to show (4) it is only
necessary to prove that H2 ⊆ J + . Assume towards a contradiction that there
exists h ∈ H2 \ J + . It follows again by Claim 1 and the fact that H1 is nonempty
that h ∈ J − and hence that H1 ∩ J − = ∅. Let a = V × A, b = V × B be two
different hyperedges in H1 . It follows that ea and eb appear in path p and we
can assume wlog. that ea appears before eb in p. We shall prove that B ⊆ A.
Let d ∈ B and let v ∈ V be such that h = {v} × D. Consider the directed path
q associated to (v, d). Clearly q contains eh and eb and hence q contains ea as
well. Hence (v, d) belongs to (V, A), and hence, d ∈ A. A similar reasoning, using
23
now any arbitrary hyperedge in J + shows that A ⊆ B. Hence we have A = B,
a contradiction. This completes the proof of (4).
Proof of Theorem 1
Let H be the hypergraph of the canonical booleanization of I. Define H1 to
be the set of all hyperedges S × R in H with S = V . Define H2 to be the
hypergraph associated to the non-empty assignment constraints. That is, H2
contains for every v ∈ V , the hyperedge {v} × D. Note that h1 ∩ h2 6= ∅ for every
h1 ∈ H1 and h2 ∈ H2 .
(1) ⇒ (2). In what follows we shall assume that |H1 | > 1 since otherwise
(2) follows directly. Assume that H is a network hypergraph. It follows that H1
and H2 satisfy the hypothesis of Lemma 8. Then, let T the tree defining H and
r the node in T given by Lemma 8. Note that for every node occurring in any
hyperedge in H1 , its associated path p must contain some edge in T + (r) and,
hence, must necessarily contain r as well. Then, condition (2) follows by applying
Lemma 4 to T − (r).
(2) ⇒ (1). We note that this is a particular case of direction (2) ⇒ (1) in
Theorem 2. Still we include a proof since we think it might help the reader to
understand the basic idea before embarking in the more complicated proof in
2. This direction follows easily from Lemma 4 and Lemma 5. Assume that (2)
holds. Then it follows from Lemma 4 that there exists a rooted tree T1 defining
H1 such that, additionally, the path associated to every variable in V (H) ends
at the root, r1 , of T1 . Also, note that every two hyperedges in H2 have an empty
intersection which implies again by Lemma 4 that there exists a rooted tree T2
defining H2 such that the path associated to every variable v ∈ V (H) starts at
the root, r2 , of T2 . Then, it follows from Lemma 5 that by gluing r1 and r2 in
T1 and T2 we obtain the tree defining H.
Proof of Theorem 2
Let H be the hypergraph of the canonical booleanization of I.
(1) ⇒ (2). Define H1 to be the set containing, for every d ∈ D, the hyperedge
V ×{d}. Every such hyperedge is in H because we are assuming that the instance
contains a global cardinality constraint. Furthermore, define H2 to be the set
containing all non-empty assignment constraints. That is H2 contains for every
v ∈ V , the hyperedge {v} × D. Notice that if a = V × {d} belongs to H1 and
b = {v} × D belongs to H2 then a ∩ b is non empty as it contains (v, d). Assume
that H is a network hypergraph. It follows that H1 and H2 satisfy the hypothesis
of Lemma 8 and let T be the tree defining H and r the vertex in V (T ) given by
Lemma 8.
We claim that for every hyperedge h ∈ H, eh belongs to E(T − (r))∪E(T − (r))
Indeed, if it is not the case then there exists some hyperedge h = S ×R such that
24
its associated edge eh has one endpoint y in V (T − (r))∪V (T + (r)) and the other,
x, outside. Assume that y ∈ V (T − (r)) (the case y ∈ V (T + (r)) is symmetric). It
follows that eh = (y, x) since otherwise eh would be included in E(T − (r)). Let
(v, d) ∈ S × R and let p be the direct path in T associated to (v, d). This path
must include eh and also the edge associated to V × {d}, but this is impossible
since both edges must appear with different polarity because the edge associated
to V × {d} belongs to V (T + (r)).
For every d ∈ D, let (xd , yd ) be the edge in T associated to hyperedge V ×{d}
and let Td be T − (xd ). Also, define T1 to be the subtree of T that contains all
the nodes x that belong to a path starting at xd for some d ∈ D and ending at
r.
For every v ∈ V , let (xv , yv ) be the edge in T associated to hyperedge {v}×D
and let Tv be T + (yv ). Also, define T2 to be the subtree of T that contains all the
nodes x that belong to a path starting at r and ending at yv for some v ∈ D.
Let (v, d) ∈ V × D and let p be the directed path in T associated to (v, d).
Clearly p must contain the edges associated to V × {d} and {v} × D and, hence,
also r. It also follows that for every hyperedge, h ∈ E(H), containing (v, d),
eh belongs to Ti with i ∈ {d, v, 1, 2}. It follows that every edge e ∈ E(T ) is
contained in some of the trees Ti , i ∈ D ∪ V ∪ {1, 2} we have just defined and
hence we can infer the shape of H by considering separately the shape of the
hypergraphs defined by each one of the trees. Then (1) ⇒ (2) follows from the
following lemma.
Lemma 9. Let e (resp. e′ ) be an edge in T , let h = S × R (resp. h′ = S ′ × R′ )
be the hyperedge in H associated to e (resp. e′ ). Then the following holds:
1. If there exists d ∈ D such that e, e′ ∈ E(Td ) then R = R′ = {d}. Furthermore, S and S ′ are disjoint or contained one in another.
2. If e, e′ ∈ E(T1 ) then S = S ′ = V . Furthermore, R and R′ are disjoint or
contained one in another.
3. If there exists v ∈ V such that e, e′ ∈ E(Tv ) then S = S ′ = {v}. Furthermore,
R and R′ are disjoint or contained one in another.
4. If e, e′ ∈ E(T2 ) then R = R′ = D. Furthermore, S and S ′ are disjoint or
contained one in another.
Proof. By symmetry we only need to prove the first two cases:
1. We first show that R = {d} (the same argument shows that R′ = {d}. Let
(v, d′ ) be any node in h and let p be its associated directed path in T . Since
p contains r, it must contain the edge associated to V × {d}, which implies
that d′ = d. Hence, R = {d}. The fact that S and S ′ are disjoint or contained
one in another follows from applying Lemma 4 to Td . Note that we use that
the path associated to every node in V × {d} must necessarily include xd .
2. We first show that S = V (the same argument shows that S ′ = V ). Let p
be the directed path associated to any variable (v, d) ∈ S × R. This path
contains xd and r which implies that every path associated to a variable
in V × {d} contains also eh . It follows that V × {d} ⊆ S × R and, hence,
25
S = V . The fact S and S ′ are disjoint or contained one in another follows
by applying Lemma 4 to T1 . Note that we use that the associated path of
every node in V × D must necessarily include r.
(1) ⇐ (2). This follows easily from Lemma 4 and Lemma 5. For every d, let
Hd be the subhypergraph of H containing all the hyperedges S × R ∈ E(H)
with R = {d} and S 6= V . It follows from Lemma 4 that there is a rooted tree
Td defining Hd such that, additionally, all the paths associated to variables in
V (H) finish at the root of Td , which we will denote by rd . Similarly, let H1
be the subhypergraph of H containing all the hyperedges S × R ∈ E(H) with
S = V . Again, by Lemma 4 there is a rooted tree T1 with root, say, r1 defining
H1 . For every d ∈ D, let us denote by (xd , yd ) the edge in T1 associated to
V × {d}. Lemma 4 guarantees that xd has in-degree 0 which implies that the
paths associated to every node in V × {d} start at xd and end at
S r1 . It follows
then from Lemma 5 that we can obtain a tree defining H1 ∪ d∈D Hd (that
is, the subgraph of H containing all hyperedges S × R ∈ E(H) where S = V
or R is a singleton) by taking the disjoint union of T1 along with all the trees
Td and gluing rd with xd for every d ∈ D. The obtained tree, which we will
call T1′ is rooted at r1 and has the additional property that every path in T1′
associated to a variable in V (H) ends at the root r1 . A symmetric argument
shows that there exists a tree T2′ with root r2 that defines the subhypergraph of
H containing all hyperedges S × R ∈ E(H) where R = D or V is a singleton
such that, additionally, every path in T1′ associated to a variable in V (H) starts
at the root r2 . Then, it follows that by gluing r1 and r2 in T1′ and T2′ we obtain
a tree defining H. This finishes the proof.
We note that direction (1 ⇒ 2) in Theorems 1 and 2 does not hold for
boolean domains due to the fact that the proof assumes the canonical encoding.
However, the direction (2 ⇒ 1) holds also for boolean domains.
Appendix C: Proof of the O(n3/2 log n log max) bound for
the Sequence constraint
Recall from Section 6.3 that a Sequence constraint is encoded as a boolean
CAC with constraints: Among({xi , . . . , xi+k }, {1}, min, max) where i = 1, . . . , n−
k and min, max, k are fixed integers. Our goal is to improve the bound O(n3/2 log2 n)
given in Section 6.3 to O(n3/2 log n log max).To this end we note, inspecting the
proof of Lemma 3, that in the log mn factor appearing in the function f (n, m),
mn is has been obtained by bounding the quantity u (where u is the maximum
capacity of the edges in the network constructed by the algorithm) by O(log n).
We shall show that in the particular case of the Sequence constraint we can
obtain a better bound. In particular, we will see that u ≤ max (and hence
log u ≤ log max), which suffices to obtain our desired bound.
Recall the definitions of a, b, c, T , and P from the proof of Lemma 3 and
recall also that the capacity of every edge in the network constructed by our
26
algorithm is either an entry of vector c, which in the particular case of the
Sequence constraint, is either max − min or 1 (and, hence, at most max),
or the absolute value of an entry of vector b = P a. In the particular case of
the Sequence constraint, all entries of a are max. Also, recall that the tree T
defining a the hypergraph of a Sequence constraint is a directed path which
implies that every row of P has at most one +1, at most one −1, and the rest
of entries are 0. It follows that every entry in b is in {−max, 0, max} and we are
done.
27
| 2cs.AI
|
An approximate solution of the MHD Falkner-Skan flow by
Hermite functions pseudospectral method
K. Parand1,∗, A.R. Rezaei, S.M. Ghaderi
arXiv:1008.2322v2 [math-ph] 16 Aug 2010
Department of Computer Sciences, Shahid Beheshti University, G.C., Tehran, Iran
Abstract
Based on a new approximation method, namely pseudospectral method, a solution for the
three order nonlinear ordinary differential laminar boundary layer Falkner-Skan equation
has been obtained on the semi-infinite domain. The proposed approach is equipped by the
orthogonal Hermite functions that have perfect properties to achieve this goal. This method
solves the problem on the semi-infinite domain without truncating it to a finite domain and
transforming domain of the problem to a finite domain. In addition, this method reduces
solution of the problem to solution of a system of algebraic equations. We also present
the comparison of this work with numerical results and show that the present method is
applicable.
Keywords: Falkner-Skan, Pseudospectral method, Semi-infinite, Nonlinear ODE, Hermite
functions, Boundary layer flow, Magnetohydrodynamics (MHD) flow
PACS: 02.70.Jn, 47.11.-j, 02.60.Cb, 02.60.Lj, 02.70.-c
1. Introduction
There are many problems in science and engineering arising in unbounded domains.
Spectral methods are famous ways to solve these kinds of problems. The most common
approach on spectral methods, that is used in this paper too, is through the use of polynomials/functions that are orthogonal over unbounded domains, such as the Hermite and the
Laguerre polynomials/functions [1–8].
Guo [9–11] proposed a method that proceeds by mapping the original problem in an
unbounded domain to a problem in a bounded domain, and then using Jacobi polynomials
to approximate the resulting problems.
Another method is replacing infinite domain with [−L, L] and semi-infinite interval with
[0, L] by choosing L, sufficiently large. This method is called domain truncation [12].
∗
Corresponding author. Tel:+98 21 22431653; Fax:+98 21 22431650.
Email addresses: [email protected] (K. Parand), [email protected] (A.R. Rezaei),
[email protected] (S.M. Ghaderi)
1
Member of research group of Scientific Computing.
February 5, 2018
Using rational approximation is another effective direct approach for solving such problems. Christov [13] and Boyd [14, 15] developed some spectral methods on unbounded
intervals by using mutually orthogonal systems of rational functions. Boyd [15] defined a
new spectral basis, named rational Chebyshev functions on the semi-infinite interval, by
mapping to the Chebyshev polynomials. Guo et al. [16] introduced a new set of rational
Legendre functions which are mutually orthogonal in L2 (0, +∞). They applied a spectral
scheme using the rational Legendre functions for solving the Korteweg-de Vries equation on
the half line. Boyd et al. [17] applied pseudospectral methods on a semi-infinite interval
and compared rational Chebyshev, Laguerre and mapped Fourier sine.
Parand et al. [18–23] applied spectral method to solve nonlinear ordinary differential
equations on semi-infinite intervals. Their approach was based on rational Tau and pseudospectral methods.
The concept of boundary layer flows [24] of an incompressible fluid has several engineering applications, such as aerodynamic extrusion of plastic sheets and cooling of a metallic
plate in a cooling bath, cooling and heating of the plate by free convection current [25].
Such flows are also encountered in the glass and polymer industries. The effect of thermal radiation absorption on an unsteady free convective flow past a vertical plate has been
studied in the presence of a magnetic field and constant wall heat flux by [25] also Makinde
[26] investigated numerically the hydromagnetic mixed convection flow of an incompressible viscous electrically conducting fluid and mass transfer over a vertical porous plate with
constant heat flux embedded in a porous medium by Newton-Raphson shooting method
along with fourth-order Runge-Kutta integration algorithm. The case of mixed-convective
boundary layer flow past a vertical porous plate embedded in a saturated porous medium
with a constant heat flux and mass transfer in the presence of a magnetic field and with
heat absorption has been studied by [27].
The steady laminar flow passing a fixed wedge was first analyzed in the early 1930s by
Falkner and Skan [28] to illustrate the application of Prandtl’s boundary layer theory. The
Magnetohydrodynamic (MHD) flows as boundary layer flows have a great significance both
from a mathematical as well as a physical standpoint. Such flows are very important in
electromagnetic propulsion. The MHD systems are used effectively in many applications
including power generators, pumps, accelerators, electrostatic filters, droplet filters, the
design of heat exchangers, the cooling of reactors etc. [29].
1.1. MHD flows
Consider the boundary layer flow of an electrically conducting viscous fluid. A magnetic
field B(x) acts transversely to the flow. The induced magnetic field is neglected by choosing
small magnetic Reynolds number assumption. Furthermore, the electric field is absent. The
relevant problem is
∂u ∂v
+
= 0,
∂x ∂y
u
∂u
du
∂ 2 v σB 2
∂u
+v
=U
+ν 2 −
(u − U),
∂x
∂y
dx
∂y
ρ
2
(1)
(2)
u = 0,
v = 0 at y = 0,
u = U(x),
as y → ∞,
(3)
where [24]
U(x) = axm
(4)
and [30]
B(x) = B0 x(m−1)/2 ,
(5)
in which u and v are the velocity components, U is the inherent characteristic velocity, ν is
a kinematic viscosity, σ is the electrical conductivity, ρ is the fluid density, B and B0 are
the magnetic field and externally imposed magnetic field in the y-direction respectively.
Defining
r
r
r
m+1 U
2 √
τ=
y, ψ =
νxU f (τ ),
(6)
2
νx
m+1
u = Uf ′ (τ ),
r
v=−
m+1
2
r
νU
m−1 ′
[f +
τ f ],
x
m+1
(7)
the continuity equation is identically satisfied and Eq. 2 and boundary conditions 3 reduce
to the following form
d2 f
df
df
d3 f
+
f
+ β[1 − ( )2 ] − M 2 (
− 1) = 0,
3
2
dτ
dτ
dτ
dτ
(8)
f (0) = f ′ (0) = 0,
(9)
f ′ (+∞) = 1,
2m
where β = m+1
and M 2 = 2σB02 /ρa(1 + m). Now our interest is to find solution of Eq. (8)
for wedge in the accelerated flow (m > 0, β > 0) and decelerated flow (m < 0, β < 0) with
separation.
MHD boundary layer flows have been studied by several researchers. Yih [31] and Ishak
et al. [32] transformed the partial differential boundary layer equations into the nonsimilar
boundary layer equations and a system of ordinary differential equations respectively, then
they used Keller box method to solve them. Hayat et al. [33], Rashidi [34] and Abbasbandy
et al. [35, 36] solved MHD boundary layer flow by modified Adomian decomposition method,
DTM-Padé, Homotopy analysis method and Hankel-Padé respectively.
The purpose of this paper is to develop a pseudospectral method that is more direct and
simpler than the other spectral methods. The method of this paper differs from previous
methods that have been used because it approximates the solution directly. Thus, it does
not truncate the semi-infinite domain to a finite domain, does not introduce the unknown
finite boundary, does not impose an asymptotic condition at this unknown boundary, and
does not use transformation to map the physical domain.
3
This paper is organized as follows:
In section 2, the properties of Hermite functions and the way to construct the pseudospectral
approach for this type equation are described. In section 3 the proposed method is applied
to solve MHD Falkner-Skan equation, and a comparison is made with existing numerical
solutions that were reported in other researches, and we have a discussion in the final
section.
2. Hermite functions pseudospectral method
Spectral methods have been successfully applied in the approximation of differential
boundary value problems defined in unbounded domains. For problems whose solutions
are sufficiently smooth, they exhibit exponential rates of convergence/spectral accuracy.
We can apply different spectral methods that are used to solve problems in semi-infinite
domains. One of these approaches is using Laguerre and Hermite polynomials/functions
[2, 4–8, 37–39]. Guo [5] suggested a Laguerre-Galerkin method for the Burgers equation and
Benjamin-Bona-Mahony (BBM) equation on a semi-infinite interval. In [7] Shen proposed
spectral methods using Laguerre functions and analyzed for model elliptic equations on regular unbounded domains. Siyyam [8] applied two numerical methods for solving initial value
problem differential equations using the Laguerre Tau method. Maday et al. [6] proposed a
Laguerre type spectral method for solving partial differential equations. Funaro and Kavian
[2] considered some algorithms by using Hermite functions. Recently Guo [4] developed
the spectral method by using Hermite polynomials. However, it is not easy to perform the
quadratures in unbounded domains, which are used in the Hermite spectral approximations.
So the Hermite pseudospectral method is more preferable in actual calculations. Guo [39]
developed the Hermite pseudospectral method for the Burgers equation on the whole line.
Guo et al. [37] considered spectral and pseudospectral approximations using Hermite functions for PDEs on the whole line to approximate the Dirac equation. Bao and Shen [38]
proposed a generalized-Laguerre-Hermite pseudospectral method for computing symmetric
and central vortex states in Bose-Einstein condensates (BECs) in three dimensions with
cylindrical symmetry.
Pseudospectral methods has become increasingly popular for solving differential equations and also it is very useful in providing highly accurate solutions to differential equations.
In this paper, we employ the Hermite functions pseudospectral method which we denote HFP
to solve MHD Falkner-Skan initial value problem directly.
2.1. Properties of Hermite functions
In this section, we detail the properties of the Hermite functions that will be used to
construct the HFP method. First we note that the Hermite polynomials are generally not
suitable in practice due to their wild asymptotic behavior at infinities [40, 41].
Hermite polynomials Hn (x), n ≥ 0, are the eigenfunctions of the singular Sturm-Liouville
problem in
′′
′
Hn (x) − 2xHn (x) + 2nHn (x) = 0
(10)
4
Hermite polynomials with large n can be written in direct formula as follow:
√
nπ
Γ(n + 1) x2 /2
e
cos ( 2n + 1x −
)
Γ(n/2 + 1)
2
√
nπ
2
).
∼ nn/2 ex /2 cos( 2n + 1x −
2
Hn (x) ∼
Hence, we shall consider the so called Hermite functions. The normalized Hermite functions
of degree n are defined by
e n (x) = √ 1 e−x2 /2 Hn (x),
H
2n n!
n ≥ 0, x ∈ R.
e n } is an orthogonal system in L2 (R), i.e.,
Clearly, {H
Z +∞
√
e n (x)H
e m (x)dx = πδmn .
H
−∞
where δnm is the Kronecker delta function.
In contrast to the Hermite polynomials, the Hermite functions are well behaved with the
decay property:
e n (x)| −→ 0,
|H
as |x| −→ ∞,
and the asymptotic formula with large n is
√
e n (x) ∼ n− 14 cos( 2n + 1x − nπ )
H
2
The three-term recurrence relation of Hermite polynomials implies
q
p n e
2 e
en+1 (x) = x
H
H (x) − n+1
Hn−1 (x), n ≥ 1,
n+1 n
√
e 0 (x) = e−x2 /2 , H
e 1 (x) = 2xe−x2 /2 .
H
Using recurrence relation of Hermite polynomials and the above formula leads to
√
e n′ (x) = 2nH
e n−1 (x) − xH
e n (x)
H
r
r
ne
n+1 e
=
Hn−1 (x) −
Hn+1 (x).
2
2
and this implies
√
n(n−1)π
−
,
2
Z
√π(n + 1 ),
′
2
e n′ (x)H
em
√
H
(x)dx =
(n+1)(n+2)π
R
,
−
2
0,
m = n − 2,
m = n,
m = n + 2,
otherwise.
5
Let us define
2
PeN := {u : u = e−x /2 v, ∀v ∈ PN }.
where PN is the set of all Hermite polynomials of degree at most N.
We now introduce the Gauss quadrature associated with the Hermite functions approach.
Let {xj }N
j=0 be the Hermite-Gauss nodes and define the weights
√
π
w
ej =
, 0 ≤ j ≤ N.
e 2 (xj )
(N + 1)H
N
Then we have
Z
N
X
p(x)dx =
p(xj )w
ej , ∀p ∈ Pe2N +1 .
R
j=0
For a more detailed discussion of these early developments, see the [42, 43].
2.2. Approximations by Hermite functions
Let us define Λ := {x| − ∞ < x < ∞} and
e 0 (x), H
e1 (x), ..., H
e N (x)}
HN = span{H
The L2 (Λ)-orthogonal projection ξ˜N : L2 (Λ) −→ HN is a mapping in a way that for any
v ∈ L2 (Λ),
< ξ˜N v − v, φ >= 0, ∀φ ∈ HN
or equivalently,
ξ˜N v(x) =
N
X
l=0
e l (x).
ṽl H
To obtain the convergence rate of Hermite functions we define the space HAr (Λ) defined by
HAr (Λ) = {v|v is measurable on Λ and kvkr,A < ∞},
and equipped with the norm kvkr,A = kAr vk. For any r > 0, the space HAr (Λ) and its norm
are defined by space interpolation. By induction, for any non-negative integer r,
r
X
Ar v(x) =
(x2 + 1)(r−k)/2 pk (x)∂xk v(x),
k=0
where pk (x) are certain rational functions which are bounded uniformly on Λ. Thus,
!1/2
r
X
kvkr,A ≤ c
k (x2 + 1)(r−k)/2 pk (x)∂xk v k
.
k=0
Theorem : For any v ∈ HAr (Λ), r ≥ 1 and 0 ≤ µ ≤ r,
k ξ˜N v − vkµ ≤ cN 1/3+(µ−1)/2 k v kr,A .
(11)
Proof. A complete proof is given by Guo et al. [37]. also same theorem has been proved by
Shen et al. [40].
6
2.3. Hermite functions transform
As mentioned before, Falkner-Skan equation is defined on the interval (0, +∞); but we
know properties of Hermite functions are derived in the infinite domain (−∞, +∞).
Also we know approximations can be constructed for infinite, semi-infinite and finite intervals. One of the approaches to construct approximations on the interval (0, +∞) which is
used in this paper, is to use a mapping, that is a change of variable of the form
ω = φ(z) =
1
ln(z).
k
(12)
where k is a constant.
The basis functions on (0, +∞) are taken to be the transformed Hermite functions,
b n (x) ≡ H
e n (x) ◦ φ(x) = H
e n (φ(x)).
H
e n (x) ◦ φ(x) is defined by H
e n (φ(x)). The inverse map of ω = φ(z) is
where H
z = φ−1 (ω) = ekω .
(13)
Thus we may define the inverse images of the spaced nodes {xj }+∞
−∞ as
Γ = {φ−1 (t) : −∞ < t < +∞} = (0, +∞)
and
x̃j = φ−1 (xj ) = ekxj ,
j = 0, 1, 2, ...
Let w(x) denote a non-negative, integrable, real-valued function over the interval Γ. We
define
L2w (Γ) = {v : Γ → R | v is measurable and k vkw < ∞}
where
k vkw =
Z
∞
21
| v(x) | w(x)dx ,
2
0
is the norm induced by the inner product of the space L2w (Γ),
Z ∞
< u, v >w =
u(x)v(x)w(x)dx.
(14)
0
b n (x)}n∈N denotes a system which is mutually orthogonal under (14), i.e.,
Thus {H
√
b n (x), H
bm (x) >w(x) = πδnm ,
<H
where w(x) = 1/x and δnm is the Kronecker delta function. This system is complete in
L2w (Γ). For any function f ∈ L2w (Γ) the following expansion holds
7
f (x) ∼
=
N
X
k=0
with
fk =
b k (x),
fk H
b k (x) >w(x)
< f (x), H
.
b k (x)k2
kH
w(x)
Now we can define an orthogonal projection based on transformed Hermite functions as
below:
Let
bN = span{H
b 0 (x), H
b1 (x), ..., H
b n (x)}
H
bN is a mapping in a way that for any
The L2 (Γ)-orthogonal projection ξˆN : L2 (Γ) −→ H
2
y ∈ L (Γ),
< ξˆN y − y, φ >= 0,
or equivalently,
ξˆN y(x) =
N
X
i=0
bN
∀φ ∈ H
b i (x).
âi H
(15)
2.4. Domain scaling
It has already been mentioned in [44] that when using a spectral approach on the whole
real line R one can possibly increase the accuracy of the computation by a suitable scaling
of the underlying time variable t. For example, if y denotes a solution of the ordinary
differential equation, then the rescaled function is ỹ(t) = y( tl ), where l is constant.
Domain scaling is used in several of the applications presented in the next section. For more
detail we refer the reader to [45].
3. Solving Falkner-Skan wedge flow equation with Hermite functions
Here, we solve Falkner-Skan wedge flow equation for m = −3/5 with M = 5, 10, 15, 20
and m = 2 with M = 5, 10, 50, 100 by pseudospectral method.
We note that the Hermite functions are not differentiable at the point x = 0, therefore
to approximate the solution of Eq. (8) with the boundary conditions Eq. (9) we multiply
operator Eq. (15) by x2 /(x + 1) to satisfy f (0) = f ′ (0) = 0 in Eq. (9) and also we construct
a function p(x) that satisfy f ′ (+∞) = 1 in Eq. (9). This function is given by
p(x) =
x2
,
x+λ
(16)
8
where λ is constant to be determined.
Therefore, the approximate solution of f (x), in Eq. (8) with boundary conditions Eq. (9)
is represented by
x2 ˆ
ξN f (x),
ξbN f (x) = p(x) +
x+1
To apply the pseudospectral scheme, we construct the residual function by substituting f (x)
by ξbN f (x) in the Falkner-Skan Eq. (8):
!
d2 ξb f (x/l)
d3 ξbN f (x/l) b
N
Resl (x) =
(17)
+ ξN f (x/l)
dx3
dx2
!
!
bN f (x/l)
d
ξ
dξbN f (x/l) 2
) − M2
−1 ,
(18)
+β 1−(
dx
dx
where l is a constant that is defined in domain scaling description, before. ξbN f (x) will be a
good approximation of function f (x), if the residual function is zero on the whole domain.
In other words, we should select coefficients ai s so that the residual function tends to zero
on most of the domain. The pseudospectral scheme for Falkner-Skan equation is to find
ξbN f (x) such that
Resl (xj ) = 0,
j = 0, . . . , N + 1,
(19)
where the xj s are N + 2 transformed Hermite-Gauss nodes by Eq. 13. This generates a set
of N + 2 nonlinear equations that can be solved by Newton method for unknown coefficients
ai s and λ.
The physical quantities of interest which is represented by value of f ′′ (0) is the skin
friction coefficient. The approximations of the f ′′ (0) obtained by this method and numerical
value [46] for m = −3/5 with M = 5, 10, 15, 20 and m = 2 with M = 5, 10, 50, 100 are listed
in Tables 1 and 2 respectively.
The velocity boundary layer of the wedge with m = −3/5 and m = 2 for various various
M are shown in figures 1 and 2 respectively. The value of f ′′ (0) can be seen to increase with
M. Hence, the presence of a magnetic field also increases the skin friction.
Logarithmic graphs of absolute coefficients |ai | of Hermite function in the approximate
solutions for m = −3/5 with M = 50 and m = 2 with same M are shown in Figures 3 and
4 respectively. The graphs illustrate that the method has an appropriate convergence rate.
4. Conclusion
The FalknerSkan non-linear ordinary differential equation arises in the study of laminar
boundary layers. This problem is the subject of an extensive research to solve because of
its importance in boundary layer theory, a solution for the non-linear ordinary differential
laminar boundary layer Falkner-Skan equation has been obtained by pseudospectral method
on semi-infinite domain, also in this work we concentrated on the boundary conditions of
9
f (0) = f ′ (0) = 0, f ′ (+∞) = 1, which is corresponding to a fixed and impermeable wedge
flow. The physical quantities of interest which is represented by the value of f ′′ (0) is the
skin friction coefficient. According to the tables and figures value of f ′′ (0) can be seen to
increase with M. Hence, the presence of a magnetic field also increases the skin friction.
Most numerical methods reported in the literature thus far are based on transforming
maps of the physical domain [0, +∞) to the finite domains, shooting methods or finitedifference methods obtained by first truncating the semi-infinite physical domain of the
problem to a finite domain at an unknown finite boundary, which is determined as part of
the solution by imposing an “asymptotic boundary condition” at this boundary.
The method presented in this paper used a set of Hermite functions and solved this
problem on the semi-infinite domain without truncating it to a finite domain, imposing the
asymptotic condition transforming and transforming domain of the problem. These functions are proposed to provide an effective but simple way to improve the convergence of the
solution by pseudospectral method. The validity of the method is based on the assumption that it converges by increasing the number of Gauss points. Through the comparisons
among the numerical solutions of Asaithambi and the current work, it has been shown that
the present work has provided acceptable approach for Falkner-Skan equation; also it was
confirmed by the theorem and logarithmic figures of absolute coefficients that this approach
has exponentially convergence rate. In total, an important concern of spectral methods is
the choice of basis functions; the basis functions have three number of properties: easy to
computation, rapid convergence and completeness, which means that any solution can be
represented to arbitrarily high accuracy by taking the truncation N sufficiently large.
Acknowledgments
The research was supported by a grant from Shahid Beheshti University.
References
[1] O. Coulaud, D. Funaro, O. Kavian, Laguerre spectral approximation of elliptic problems in exterior
domains, Comput. Method. Appl. Mech. Eng. 80 1-3 (1990)451-458.
[2] D. Funaro, O. Kavian, Approximation of some diffusion evolution equations in unbounded domains by
Hermite functions, Math. Comp. 57 (1991)597-619.
[3] D. Funaro, Computational aspects of pseudospectral Laguerre approximations, Appl. Numer. Math. 6
6 (1990)447-457.
[4] B.Y. Guo, Error estimation of Hermite spectral method for nonlinear partial differential equations,
Math. Comput. 68 227 (1999)1067-1078.
[5] B.Y. Guo, J. Shen, Laguerre-Galerkin method for nonlinear partial differential equations on a semiinfinite interval, Numer. Math. 86 4 (2000)635-654.
[6] Y. Maday, B. Pernaud-Thomas, H. Vandeven, Reappraisal of Laguerre type spectral methods, La.
Rech. Aerospatiale. 6 (1985)13-35.
[7] J. Shen, Stable and efficient spectral methods in unbounded domains using Laguerre functions, SIAM.
J. Numer. Anal. 38 4 (2000)1113-1133.
[8] H.I. Siyyam, Laguerre Tau methods for solving higher order ordinary differential equations, J. Comput.
Anal. Appl. 3 2 (2001)173-182.
10
[9] B.Y. Guo, Gegenbauer approximation and its applications to differential equations on the whole line,
J. Math. Anal. Appl. 226 1 (1998)180-206.
[10] B.Y. Guo, Jacobi spectral approximation and its applications to differential equations on the half line,
J. Comput. Math. 18 1 (2000)95-112.
[11] B.Y. Guo, Jacobi approximations in certain Hilbert spaces and their applications to singular differential
equations, J. Math. Anal. Appl. 243 2 (2000)373-408.
[12] J.P. Boyd, Chebyshev and Fourier Spectral Methods, second ed., Dover., New York, 2000.
[13] CI. Christov, A complete orthogonal system of functions in L2 (−∞, ∞) space, SIAM. J. Appl. Math.
42 (1982)1337-1344.
[14] J.P.Boyd, Spectral methods using rational basis functions on an infinite interval, J. Comput. Phys. 69
1 (1987)112-142.
[15] J.P. Boyd, Orthogonal rational functions on a semi-infinite interval, J. Comput. Phys. 70 1 (1987)63-88.
[16] B.Y. Guo, J. Shen, Z.Q. Wang, A rational approximation and its applications to differential equations
on the half line, J. Sci. Comput. 15 2 (2000)117-147.
[17] J.P. Boyd, C. Rangan, P.H. Bucksbaum, Pseudospectral methods on a semi-infinite interval with application to the Hydrogen atom: a comparison of the mapped Fourier-sine method with Laguerre series
and rational Chebyshev expansions, J. Comput. Phys. 188 1 (2003)56-74.
[18] K. Parand, M. Razzaghi, Rational Chebyshev Tau method for solving Volterra’s population model,
Appl. Math. Comput. 149 3 (2004)893-900.
[19] K. Parand, M. Razzaghi, Rational Chebyshev Tau method for solving higher-order ordinary differential
equations, Int. J. Comput. Math. 81 1 (2004)73-80.
[20] K. Parand, M. Razzaghi, Rational Legendre approximation for solving some physical problems on
semi-infinite intervals, Phys. Scripta. 69 (2004)353-357.
[21] K. Parand, M. Shahini, Rational Chebyshev pseudospectral approach for solving Thomas-Fermi equation, Phys. Lett. A. 373 (2009)210-213.
[22] K. Parand, A. Taghavi, Rational scaled generalized Laguerre function collocation method for solving
the Blasius equation, J. Comput. Appl. Math. 233 4 (2009)980-989.
[23] K. Parand, M. Shahini, M. Dehghan, Rational Legendre pseudospectral approach for solving nonlinear
differential equations of Lane-Emden type, J. Comput. Phys. 228 23 (2009)8830-8840.
[24] H. Schlichting, Boundary Layer Theory, 6th ed., McGraw-Hill, New York, 1979.
[25] A. Ogulu, O.D. Makinde, Unsteady hydromagnetic free Convection flow of a dissipative and radiating
fluid past a vertical plate with constant heat flux, Chem. Eng. Comm. 196 (2009) 454462.
[26] O.D. Makinde, On MHD boundary-layer flow and mass transfer past a vertical plate in a porous medium
with constant heat flux, Int. J. Numer. Method Heat Fluid Flow. 19 3/4 (2009) 546-554.
[27] O.D. Makinde, P. Sibanda, Magnetohydrodynamic Mixed-Convective Flow and Heat and Mass Transfer
Past a Vertical Plate in a Porous Medium With Constant Wall Suction, J. Heat Transfer. 130 (2008)
1-8.
[28] V.M. Falkner, S.W. Skan, Some approximate solutions of the boundary-layer equations, Phiols. Mag.
12 (1931)865-896.
[29] G.W. Sutton, A. Sherman, Engineering Magnetohydrodynamics, McGrawHill, New York, 1965.
[30] T.C. Chiam, Hydromagnetic flow over a surface stretching with a power-law velocity, Int. J. Eng. Sci.
33 (1995)429-435.
[31] K.A. Yih, MHD forced convection flow adjacent to a non-isothermal wedge, Int. Comm. Heat. Mass.
Transfer. 26 6 (1999)819-827.
[32] A. Ishak, R. Nazar, I. Pop, MHD boundary-layer flow of a micropolar fluid past a wedge with constant
wall heat flux, Commun. Nonlinear. Sci. Numer. Simul. 14 (2009)109-118.
[33] T. Hayat, Q. Hussain, T. Javed, The modified decomposition method and Padé approximants for the
MHD flow over a non-linear stretching sheet, Nonlinear. Anal. Real. World. Appl. 10 (2009)966-973.
[34] M.M. Rashidi, The modified differential transform method for solving MHD boundary-layer equations,
Comput. Phys. Commun. 180, (2009)2210-2217.
[35] S. Abbasbandy, T. Hayat, Solution of the MHD Falkner-Skan flow by homotopy analysis method,
11
Commun. Nonlinear. Sci. Numer. Simulat. 14 (2009)3591-3598.
[36] S. Abbasbandy, T. Hayat, Solution of the MHD FalknerSkan flow by HankelPadé method, Phys. Lett.
A. 373 (2009)731-734.
[37] B.y. Guo, J. Shen, J., C.l. Xu, Spectral and pseudospectral approximations using Hermite functions:
application to the Dirac equation, Adv. Comput. Math. 19 1-3 (2003)35-55.
[38] W. Bao, J. Shen, A generalized-Laguerre-Hermite pseudospectral method for computing symmetric
and central vortex states in Bose-Einstein condensates, J. Comput. Phys. 227 23 (2008)9778-9793.
[39] B.Y. Guo, C.L. Xu, Hermite pseudospectral method for nonlinear partial differential equations, Math.
Model. Numer. Anal. 34 4 (2000)859-872.
[40] J. Shen, L.L. Wang, Some Recent Advances on Spectral Methods for Unbounded Domains, Commun.
Comput. Phys. 5 2-4 (2009)195-241.
[41] G. Szegö. Orthogonal Polynomials, 4th edition, volume 23. AMS Coll. Publ, 1975.
[42] J. Shen, T. Tang, High Order Numerical Methods and Algorithms, Chinese Science Press, 2005.
[43] J. Shen, T. Tang, L-L. Wang, Spectral Methods Algorithms, Analyses and Applications, Springer, First
edition, 2010.
[44] Y. Liu, L. Liu, T. Tang, The numerical computation of connecting orbits in dynamical systems: a
rational spectral approach, J. Comput. Phys. 111 2 (1994)373-380.
[45] T. Tang, The Hermite spectral method for Gaussian-type functions, SIAM. J. Sci. Comput. 14 3
(1993)594-606.
[46] N.S. Asaithambi, A numerical method for the solution of the Falkner-Skan equation, Appl. Math.
Comput. 81 2-3 (1997)259-264.
12
Table 1: The comparison of HFP approximate solution of f ′′ (0) with numerical results [46] when m=-3/5
M
5
10
15
20
50
N
20
15
15
15
20
k
2
2
1
1
2
l
1.658
1.296
1.089
1
1.336
Present method
4.60075494
9.80646420
14.87167484
19.90393701
49.96165233
Numerical
4.60075494
9.80646420
14.87167484
19.90393701
49.96165233
Table 2: The comparison of HFP approximate solution of f ′′ (0) with numerical results [46] when m = 2
M
5
10
50
100
N
30
30
30
30
k
3
2
2
2
l
1.194
1.112
0.904
0.616
Present method
5.19095945
10.09677545
50.01944071
100.00972170
13
Numerical
5.19095945
10.09677545
50.01944071
100.00972170
Figure 1: The velocity boundary layer of the wedge with m = −3/5 and various M
Figure 2: The velocity boundary layer of the wedge with m = 2 and various M
14
Figure 3: Logarithmic graph of absolute coefficients |ai | of Hermite functions in the approximate solution
for m = −3/5 and M = 50
Figure 4: Logarithmic graph of absolute coefficients |ai | of Hermite functions in the approximate solution
for m = 2 and M = 50
15
| 5cs.CE
|
arXiv:1611.01364v2 [cs.GT] 17 Jan 2018
A Generalization of the Minisum and Minimax
Voting Methods
Shankar N. Sivarajan
Undergraduate, Department of Physics
Indian Institute of Science
Bangalore 560 012, India
[email protected]
Faculty Advisor: Prof. Y. Narahari
Department of Computer Science and Automation
Abstract
In this paper, we propose a family of approval voting-schemes for electing
committees based on the preferences of voters. In our schemes, we calculate the vector of distances of the possible committees from each of the
ballots and, for a given p-norm, choose the one that minimizes the magnitude of the distance vector under that norm. The minisum and minimax
methods suggested by previous authors and analyzed extensively in the
literature naturally appear as special cases corresponding to p = 1 and
p = ∞, respectively. Supported by examples, we suggest that using a
small value of p, such as 2 or 3, provides a good compromise between the
minisum and minimax voting methods with regard to the weightage given
to approvals and disapprovals. For large but finite p, our method reduces
to finding the committee that covers the maximum number of voters, and
this is far superior to the minimax method which is prone to ties. We also
discuss extensions of our methods to ternary voting.
I.
Introduction
In this paper, we consider the problem of selecting a committee of k members
out of n candidates based on preferences expressed by m voters. The most
common way of conducting this election is to allow each voter to select his
favorite candidate and vote for him/her, and we select the k candidates with the
most number of votes. While this system is easy to understand and implement,
upon scrutiny, there arise certain unfavorable aspects.
Suppose we wish to elect a committee of size k = 1, and that there are two
candidates, a Conservative A, and a Liberal B, with B expected to win. In this
case, it is to candidate A’s benefit to introduce a candidate C whose ideology
1
is similar to that of candidate B in order to draw votes away from B, thereby
ensuring his own victory. This may be accomplished in practice by providing
campaign funding, for example. This is illustrated by the following example.
Assume the preferences of the voters for the various candidates are as indicated
in Table 1. The first entry (A, B, C) = 30% means that 30% of the voters rank
the candidates in decreasing order of preference as A followed by B, and then
by C. As a result of these preferences, in an election between candidates A and
B alone, B would win 56% of the votes. But in a three-way election, A would
win, with the votes distributed as in Table 2. Similar problems can arise in the
selection of committees with k > 1 as well.
Preference vector
(A, B, C)
(A, C, B)
(B, A, C)
(B, C, A)
(C, A, B)
(C, B, A)
Percentage of votes
30
10
6
25
4
25
Table 1: Example of a ballot with full preference list indicated.
Candidate
A
B
C
Percentage of votes
40
31
29
Table 2: The ballot from Table 1 with only the top preference indicated.
One way to counteract this is to conduct preliminary elections within the
party, in our case the Liberals, to select candidate B to represent them, and
then conducting an election between candidates A and B. But we seek methods
to make a decision based on a single election.
The reason candidate B is preferable to A is that if we compare the two of
them, a larger percentage of the voters would prefer B. In fact, if we compare
any two candidates, B is preferred. This type of candidate is called a Condorcet
winner, and this criterion, that she be preferred by a majority in any pairwise
comparison is called the Condorcet criterion. However, every election need not
have a Condorcet winner. A voting method that uses the Condorcet criterion
to select a winner is called a Condorcet method.
To check whether the Condorcet criterion is satisfied, we may ask each voter
to supply his/her full list of preferences, as in Table 1, and if there is a Condorcet
winner, we can elect her to the committee. However, if there is no Condorcet
winner, we still need a method to choose the winner. We can compare the
2
candidates pairwise and choose the candidate who wins the most pairwise comparisons, and this is called the Copeland Method.
To elect a single winner, we can also use approval voting [1]. We ask each
voter to approve two candidates, and we select the candidate with the most
number of approvals. In our example, assuming that each voter approves of his
top two preferences, A is approved by 50% of the voters, B is approved by 86%,
and C, by 64%. This elects candidate B, as desired.
Approval voting is not a Condorcet method, in general. However, in large
elections with, say, a million voters, and about twenty candidates, it is impractical to require the voters to provide a full list of preferences, and hence Condercet
or Copeland methods are infeasible but approval voting is eminently feasible.
Motivated by this feature, and also because it is amenable to the analytical
methods we employ, we consider only approval voting and some of its variants
in this paper.
II.
Shortcomings of the Minisum and
Minimax Voting Methods
Suppose we have to elect a committee of k representatives from a pool of n
candidates. Further, suppose there are m voters and each of them can approve
of an arbitrary number of candidates.
We will consider the example used in [3] shown in Table 3. In this case
n = 5, m = 4 and k = 2. We represent the ballot of voter 1 by the n-dimensional
binary vector (1, 1, 0, 0, 1), and similarly for all the other voters.
1
2
3
4
A
1
1
1
0
B
1
1
1
0
C
0
0
1
1
D
0
0
1
1
E
1
0
0
1
Table 3: The example from [3] showing the approvals of the voters 1, 2 . . . 4 for
the candidates A, B . . . E. An approval is indicated by a “1”.
Now, we wish to select a committee of size k = 2, and the most common
method is called the minisum method. This means we simply select the top
k candidates with the most votes, in this case A and B. Like ballots, we also
represent committees by binary vectors of length n with a 1 denoting membership in the committee. Thus we represent the committee consisting of A and B
by the binary vector (1, 1, 0, 0, 0). This method is called the minisum strategy
because, if you take the Hamming distances of this committee-vector to each of
the m ballot-vectors, in this case 1, 0, 2, 5, and add them, this is the committee
vector that minimizes this sum. We now prove this result.
3
Let the candidates be numbered from 1 to n, and let the number of approvals
for them be a1 , a2 , . . . an , respectively. Without loss of generality, assume a1 ≥
a2 ≥ . . . ≥ an . Let the committee that minimizes the sum of Hamming distances
be (i1 , i2 , . . . ik ). This minimum value is equal to
(a1 + a2 + . . . + an ) − (ai1 + ai2 + . . . aik ),
which, neglecting ties for the k th position, is obviously minimized when
(i1 , i2 , . . . ik ) = (1, 2, . . . k).
This result allows us to consider the problem of finding the minisum committee as an optimization problem since we have a computable number, the sum of
Hamming distances, that enables us to compare any two committees and choose
the better one. But this method does not weigh disapprovals enough (as seen
earlier in Table 2) and so, a method called the minimax was introduced [6].
This method chooses the committee which minimizes the maximum Hamming
distance to the ballots. Under this, we elect the committee of A and C, i.e., the
vector (1, 0, 1, 0, 0).
We will demonstrate the inadequacies of these methods bv considering the
example in Table 4, with m = 1000 voters to elect a committee of k = 2
members. We name the n = 4 candidates A1 , A2 , B1 and B2 to indicate the
correlation of ballots between candidates A1 and A2 and between B1 and B2 .
For simplicity, in the rest of the paper, we restrict each of the voters to approve
the same number of candidates as the size of the committee, in this case, two.
Candidate pair
{A1 , A2 }
{A1 , B1 }
{A1 , B2 }
{A2 , B1 }
{A2 , B2 }
{B1 , B2 }
Votes
500
100
10
20
20
350
Table 4: An example of a ballot which exhibits correlations of votes, possibly
between candidates with the same ideology.
Using the minisum method, we would elect candidates A1 and A2 , as they
have the most number of approvals, as shown in Table 5. However, this leaves
35% of the voters without either of their choices in the committee.
The minimax method
is even worse because it does not narrow the choices at
all; all of the nk = 42 = 6 committees yield the same minimax distance and it
results in all ballots tied at a Hamming distance of 4. Such tie is almost certain
whenever the number of voters is large but the number of possible committees is
small. For example, when there are a million voters and a committee of size 2 to
be chosen from 10 candidates, it is highly likely that there is one of the million
4
Candidate
A1
A2
B1
B2
Number of approvals
500 + 100 + 10 = 610
500 + 20 + 20 = 540
100 + 20 + 350 = 470
10 + 20 + 350 = 380
Table 5: The number of approvals per candidate, using the ballots from Table 4.
voters who is against both candidates in every possible one of the 10
= 45
2
committees. In this case, the minimax method will yield the same minimax
distance of 4 for every one of the 45 possible committees. Related papers [2, 3] do
not elaborate upon this shortcoming but instead focus on efficient computation
of the minimax solution.
Another problem with the minimax method is that it weighs disapprovals
too highly. A single voter can change the outcome of the entire election. Suppose 1000 voters are choosing a committee of size two from four candidates,
{A1 , A2 , B1 , B2 }. Further suppose that the first 998 vote for A1 , A2 , the 999th
voter for {A1 , B2 } and the 1000th voter for {B1 , B2 }. Then the maximum
distance to {A1 , A2 } is 4 due to the 1000th voter’s ballot, but the maximum
distance to {A1 , B2 } is only two, and therefore the minimax method elects the
committee {A1 , B2 }. Candidate A2 who was approved by 998 of the 1000 voters
is not in the committee!
An intuitively superior choice for the committee based on the ballots in
Table 4 would be {A1 , B1 }, where 98% of the voters get at least one of their
choices on the committee. This seems fairer because a committee where I get
one choice and you get one choice should be preferred over another committee
where one of us gets two choices, and the other none. The minisum method does
not prefer one of these committees over the other but we will now demonstrate
a method that does so, and chooses the committee {A1 , B1 } in this example.
III.
The p-norm Minimization Method
We represent each of the possible committees as a binary vector (c1 , c2 , . . . , cn )
with ci = 1 if candidate i is in the committee. Each of the m approval ballots
is a vector of the same form, (b1 , b2 , . . . , bn ). For each candidate committeevector, we calculate the vector of Hamming distances of the ballot-vectors from
it. This vector has size m, the number of voters. The components of this vector
of Hamming distances are all even (This is a property of binary vectors of equal
1-norm.) In the case of the committee {A1 , A2 }, this vector will have a 1000
components, with 500 of these being 0, 150 of these being 2, and 350 of these
being 4. For each of the candidate committee-vectors, this distance information
is summarized in Table 6.
Suppose for a certain committee, the number of ballots at a Hamming dis-
5
Committee
{A1 , A2 }
{A1 , B1 }
{A1 , B2 }
{A2 , B1 }
{A2 , B2 }
{B1 , B2 }
# of ballots
at distance 0
500
100
10
20
20
350
# of ballots
at distance 2
150
880
970
970
880
150
# of ballots
at distance 4
350
20
20
10
100
500
Table 6: Number of ballots at different distances from the possible committees,
using the ballots from Table 4.
tance of i is νi . The p-norm of the Hamming distance vector for this committee
is, therefore,
!1/p
X
p
.
νi · i
i
We then find the committee that minimizes this p-norm.
We investigate the result of using intermediate values of p in Table 7. Minimizing the 1-norm yields the same solution as minisum; minimizing the ∞-norm
yields the same solution as minimax. Notice that as the value of p → ∞, the
value of the distance tends to 4, the maximum Hamming distance.
Committee
{A1 , A2 }
{A1 , B1 }
{A1 , B2 }
{A2 , B1 }
{A2 , B2 }
{B1 , B2 }
1
1700
1840
2020
1980
2160
2300
Value of p in
2
3
78.74
28.68
61.97 20.26
64.81
20.83
63.56
20.33
71.55
23.78
92.74
32.14
the p-norm
4
10
17.42 7.19
11.77 5.42
11.99 5.42
11.60 5.08
14.11 6.34
19.00 7.45
100
4.24
4.12
4.12
4.09
4.19
4.26
Table 7: p-norm of the m-vector for all the possible committees for different
values of p, using the ballots from Table 4. The value in bold is the minimizer
for its column.
At higher values of p, the number of voters who have not voted for a given
committee is weighted higher, compared to the number of voters who have voted
for at least one member of the committee. We suggest that using this method,
choosing a small value of p such as 2 or 3, is a good compromise between the
minisum and minimax solution, since it weighs disapprovals more than minisum,
but is much less likely to tie than minimax.1
1 The
p-norm method will produce a tie for finite p, if the number of ballots at each distance
6
Note that for p, equal to 2 or 3, this method chooses the committee {A1 , B1 },
which contains at least one choice of 98% of the voters. This is the intuitively
superior choice we identified earlier.
It is clear that this method generalizes for larger values of k, the committee size, with the distances ranging across all the even integers from 0 to 2k,
assuming n ≥ 2k.
Maximum Coverage Problem (p → ∞)
In this example, note that the optimal solution for p ≥ 3 always produces the
committee {A2 , B1 }. Observe that this is the committee with the least number of
voters who did not approve of any member in it, i.e., 1%. It is easy to show this
is true in general: For sufficiently large but finite p, the p-norm method produces
the solution with the least number of voters at distance 2k, and therefore the
committee with the “maximum coverage”. Intuitively, this method corresponds
to weighing disapprovals more than approvals.
Let the committees C and C ′ have νi and νi′ voters respectively at a Hamming
distance of i. Let the p-norm of the vector of Hamming distances from C be ||C||p .
Since
2k
P
||C||p
= lim i=0
p→∞ ||C ′ ||p
p→∞
2k
P
lim
i=0
νi · i
p
νi′ · ip
1/p
1/p
2k
P
= lim i=0
p→∞
2k
P
i=0
νi ·
νi′ ·
i p
2k
i
2k
p
1/p
1/p
=
ν2k
′
ν2k
1/p
,
we see that for sufficiently large p, the committee with minimum p-norm is the
one which has the minimum number of voters at distance 2k. This committee
minimizes the number of voters who disapprove of every one of its members,
and thereby maximizes the number of voters who approve of at least one of its
members. We can intuitively think of the Maximum Cover method as minimizing disapprovals, and in that sense, the other extreme of minisum, which
maximizes approvals.
The maximum coverage problem [4] is a classic question in computer science
and computational complexity theory, and in its standard form, it is phrased as
follows:
You are given several overlapping sets and a number k. You must
choose at most k of the sets such that the union of the selected sets
is of maximum size.
The voting problem can be interpreted as an adaptation of this. The sets
we are given are the voters who approved of a particular candidate. We are to
is the same for two candidate committees. In this case, there will be a tie between these two
committees for all values of p. This is much less likely than a tie in the minimax case where
two committees will be tied if there one ballot at the maximum possible distance from each
of them.
7
choose k candidates in order to maximize the number of voters with at least
one of their approved candidates on the committee. We will treat our example
in Table 4 as an instance of this problem.
It is easy to see that the maximum cover in this case will be the committee
{A2 , B1 }, covering 99% of the voters. This is the same result when p is large
but finite. While it is known that the maximum coverage problem is NP-hard,
if the committee size k is small, the number of possible committees, nk is also
modest, and the NP-hardness of the maximum cover problem does not pose a
computational problem. In such cases, the maximum cover committee can be
computed exactly as we have done. There may be many elections where this is
true and where the maximum cover committee can be computed without undue
effort. For example, one could consider electing both the senators from a US
state2 from a set of aspirants through state-wide approval voting, and choosing
the maximum cover committee.
It is known that the greedy algorithm is often a good approximation to the
solution of the maximum coverage problem. [7]. Further, the greedy algorithm
is computationally simpler than calculating the Hamming distance vector for
each of the possible committees, and minimizing the p-norm. Therefore, the
greedy algorithm applied to the problem of ensuring that the maximum number
of voters get at least one of their approved candidates on the committee can be
a computationally efficient, as well as reasonable, voting method, if the committee size k is so large that the maxiumum cover problem cannot be solved
exactly. Applying the greedy algorithm to the example in Table 4, the first
candidate we choose is A1 , since we can cover 61% of the voters with this. After
choosing A1 , we can cover the most number of additional voters by choosing
B1 , yielding the committee {A1 , B1 }. This is the same result we get with the
p-norm minimization method, with p = 2 and p = 3.
The case where p → 0
We now investigate the effect of letting p → 0. In this case, it is the number of
votes that the committee as a whole wins that matters, and not the number of
votes of each of the constituent members. This can be shown using the same
technique we used in the case of p → ∞. Using the same notation as before, we
get
2k
1/p
P
p
1/p
νi · i
m − ν0
||C||p
i=0
=
lim
=
lim
.
1/p
p→0 ||C ′ ||p
p→0 P
m − ν0′
2k
′
νi · i p
i=0
So this method chooses the committee with the least value of m − ν0 or
the committee with the highest value of ν0 , namely approvals. In the previous
2 This would require amending the class system where each senator from a state is elected
for different but overlapping terms.
8
example from Table 4, the result with p = 1 coincides with the one with p → 0;
we will therefore use a different example which is shown in Table 8.
Candidate pair
{A1 , A2 }
{A1 , B1 }
{A1 , B2 }
{A2 , B1 }
{A2 , B2 }
{B1 , B2 }
Votes
300
250
150
40
60
200
Table 8: An example of a ballot to demonstrate the result of p → 0.
We can see that the candidates with the most votes are A1 and B1 , as
shown in Table 9, and according to the minisum method, i.e., with p = 1, the
committee that will be elected is {A1 , B1 }. However, reducing the value of p,
we see, from Table 10 that the elected committee changes to {A1 , A2 }, the pair
that obtained the most votes, even though candidate A2 received the fewest
number of votes in total.
Candidate
A1
A2
B1
B2
Number of approvals
300 + 250 + 150 = 700
300 + 40 + 60 = 400
250 + 40 + 200 = 490
150 + 60 + 200 = 410
Table 9: The number of approvals per candidate, using the ballots from Table 8.
IV.
Ternary Voting
It has not escaped our notice that this method of choosing the committee that
minimizes the p-norm of the vector of Hamming distances to the ballots, can
be generalized to include ternary voting [5]. In this form of election, the voters
can express their opinion about a fixed number of candidates, each of whom
can either be approved (represented by a 1) or rejected (represented by a −1).
Consider the example from Table 1 and let the voters be allowed two opinions
or non-zero entries. The voters who have preference vector (A, B, C) have three
ways to express their preference using ternary voting, namely, (1, 1, 0), (1, 0, −1),
or (0, −1, −1).
In the case of only three representatives, the voters can actually indicate
their full preference vector through ternary voting, but with more candidates,
3 Strictly speaking, this should not be termed a norm for p < 1 since the triangle inequality
is reversed.
9
Committee
{A1 , A2 }
{A1 , B1 }
{A1 , B2 }
{A2 , B1 }
{A2 , B2 }
{B1 , B2 }
1
1800
1620
1780
2220
2380
2200
Value of p in the p-norm
0.5
0.1
0.001
1107.11 765.63
700.62
1095.81
808.45
750.56
1225.51
914.08
850.62
1445.51
1040.44
960.77
1475.81
1026.70
940.83
1307.11
880.50
800.76
Table 10: p-norm3 of the m-vector, raised to the power p, for all the possible
committees, for different values of p, using the ballots from Table 8. We raise
the p-norm to the power p to keep the values bounded for small p. The value
in bold is the minimizer for its column.
the full preference list is too complex to express. Moreover, it may be that the
voter is actually neutral about most of the candidates, but rejects a few, and
this form of ternary voting allows them to express this.
Under ternary voting, we need to replace the Hamming distance between a
candidate committee, (c1 , c2 , . . . , cn ) and a ballot, (b1 , b2 , . . . , bn ) by some suitable metric. Since the Hamming metric corresponds to the L1 -metric, namely,
P
n
i=1 |ci − bi |, it is natural to consider the same metric under ternary voting.
However, we must revise our vector representation of the committees. If we
represent the hypothetical committee consisting of the first two out of four candidates as (1, 1, 0, 0), then the distance to the ballot (1, 1, 0, 0) will be 0, whereas
the distance to the ballot (0, 0, −1, −1) will be 4. This is a bad model since a
voter disapproving of the last two candidates is implicitly approving the first
two. Therefore, this candidate committee is better represented by (1, 1, −1, −1)
in which case the distance to both the ballots will be 2, in accordance with our
intuition.
Our generalization of approval voting from binary to ternary voting, to elect
a committee of size k from n candidates, is the following. We represent each
of the possible committees as a ternary vector (c1 , c2 , . . . , cn ) with ci = 1 if
candidate i is in the committee, and ci = −1, otherwise. Each of the m approval
ballots is a ternary vector (b1 , b2 , . . . , bn ) where bi = 1 if the voter approves
candidate i, bi = −1 if the voter rejects candidate i, and bi = 0 otherwise (the
voter has no opinion). Each voter must express k opinions and therefore make
k entries in his ballot as 1 or −1; the rest of the entries must be zero. For the
sake of mathematical simplicity (or elegance!), we have arbitrarily chosen to
weigh approvals and rejections equally—a voter uses up one choice to approve
a candidate, or to reject a candidate.
It is interesting to ask what the properties of the p-norm committees are, for
various values of p, under ternary voting. Suppose candidate i has ai approvals,
ri rejections, and ni = m − ai − ri neutral votes. Then, candidate i contributes
ni + 2ri = m − ai − ri + 2ri = m − (ai − ri )
10
to the 1-norm if he is in the committee, and
ni + 2ai = m − ai − ri + 2ai = m + (ai − ri )
if he is not in the committee. Therefore, the minimum 1-norm committee consists of the k candidates with the k least values of m − (ai − ri ), or equivalently,
the k highest values of ai − ri . This is an intuitively pleasing extension of the
minisum method to ternary voting: the 1-norm committee minimizes the sum
of the net approvals of its members, when rejections (explicit disapprovals) are
allowed.
Similarly, we can extend our result of maximum coverage to ternary voting
as well. Suppose we wish to elect a committee of k members from a set of n.
Let each voter choose k candidates to either approve or reject. We see that the
L1 -distance from the committees to the ballots ranges from n − k to n + k, and
using the same method as before, we see that as p → ∞, the committee that
is selected is the one with the least number of ballots at distance n + k. It is
clear that a ballot is only at this distance when none of the candidates that are
approved are on the committee, and all the candidates that are rejected are on
the committee. Therefore, in the context of the Maximum Coverage Problem,
we consider a voter to be covered if any of the candidates he approves is on the
committee, or if any of the candidates he rejects is not on the committee. Then,
in this case also, choosing the committee that covers the maximum number of
voters is better than using the minimax method because of the preponderance
of ties that arise using the latter method.
We conclude this section with an example. Consider the example of Table 4
but with one change: the 500 voters who approved {A1 , A2 } now choose to
approve A1 and reject B1 . Perhaps B1 evokes strong emotions and the votaries
of A1 and A2 would prefer to sacrifice approving A2 and instead reject B1 . All
other votes remain the same. The results of this election are shown in Table 11
for various values of the p-norm.
Committee
{A1 , A2 }
{A1 , B1 }
{A1 , B2 }
{A2 , B1 }
{A2 , B2 }
{B1 , B2 }
1
3700
3840
3020
4980
4160
4300
Value
2
130.38
123.29
101.39
161.12
133.27
147.65
of p in the p-norm
3
4
10
44.68
26.59 10.79
39.46
22.42
8.57
33.76 19.82 8.38
51.97
29.73 11.21
42.74
24.41
9.65
49.38
28.84 11.18
100
6.36
6.18
6.18
6.39
6.28
6.38
Table 11: p-norm of the m-vector for all the possible committees for different
values of p, under ternary voting, using the ballots from Table 4 except that the
500 voters who approved A1 and A2 now choose to approve A1 and reject B1 .
The value in bold is the minimizer for its column.
For every value of p, the committee elected is {A1 , B2 }. Compare this result
11
with Table 7 where this particular committee is not elected for any value of p.
Moreover, B1 was elected in the binary case for all p ≥ 2 but is never elected
in this ternary case. The option to reject a candidate, as opposed to approve
another candidate, can dramatically change the result of an election!
V.
Summary
We see that all three known approval voting methods—the committee with the
most approvals, the minisum committee and the minimax committee—are all
special cases of our p-norm method for p → 0, p = 1 and p → ∞ respectively.
However, using an intermediate value of p such as 2 or 3 is probably a better choice than any of these methods as a compromise between approvals and
disapprovals.
The minimax method is particularly unsuitable for this type of election given
the preponderance of ties, especially when the number of voters is large, and
should be logically replaced by the Maximum Cover method, which too is a
special case of our p-norm method for large, but finite, p, and intuitively corresponds to minimizing disapprovals.
We also showed that our p-norm method can be generalized to the problem
of ternary voting, and the interpretations of the results when p = 1 and p → ∞
are natural analogs of the binary case.
Acknowledgement
I would like to thank Professor Y. Narahari for giving me the opportunity to
work on this summer project, and Palash Dey for his guidance. I would also
like to acknowledge the KVPY program at IISc for enabling me to pursue a
research-oriented undergraduate program. I would like to thank the anonymous
referee for his or her insightful review, and both the referee and the editor for
their thoughtful suggestions that have improved this paper.
References
[1] S. J. Brams, Approval voting, in The Encyclopedia of Public Choice, C. K.
Rowley and F. Schneider, eds., Springer US, 2004, pp. 344–346.
[2] S. J. Brams, D. M. Kilgour, and M. R. Sanver, A minimax procedure
for electing committees, Public Choice, 132 (2007), pp. 401–420.
[3] I. Caragiannis, D. Kalaitzis, and E. Markakis, Approximation algorithms and mechanism design for minimax approval voting, in AAAI Conference on Artificial Intelligence, 2010.
[4] R. Cohen and L. Katzir, The generalized maximum coverage problem,
Inf. Process. Lett., 108 (2008), pp. 15–22.
12
[5] D. S. Felsenthal and M. Machover, Ternary voting games, Int. J.
Game Theory, 26 (1997), pp. 335–351.
[6] R. LeGrand, E. Markakis, and A. Mehta, Some results on approximating the minimax solution in approval voting, in Proceedings of the 6th
International Joint Conference on Autonomous Agents and Multiagent Systems, AAMAS ’07, ACM, 2007, pp. 198:1–198:3.
[7] G. Nemhauser, L. Wolsey, and M. Fisher, An analysis of approximations for maximizing submodular set functions – i, Mathematical Programming, 14 (1978), pp. 265–294.
13
| 8cs.DS
|
Provable Burer-Monteiro factorization
for a class of norm-constrained matrix problems
Dohyung Park1 , Anastasios Kyrillidis1 , Srinadh Bhojanapalli2
Constantine Caramanis1 , and Sujay Sanghavi1
arXiv:1606.01316v3 [stat.ML] 1 Oct 2016
1
The University of Texas at Austin
{dhpark, anastasios, constantine}@utexas.edu,
[email protected]
2
Toyota Technological Institute at Chicago
[email protected]
October 4, 2016
Abstract
We study the projected gradient descent method on low-rank matrix problems with a strongly
convex objective. We use the Burer-Monteiro factorization approach to implicitly enforce lowrankness; such factorization introduces non-convexity in the objective. We focus on constraint sets
that include both positive semi-definite (PSD) constraints and specific matrix norm-constraints.
Such criteria appear in quantum state tomography and phase retrieval applications.
We show that non-convex projected gradient descent favors local linear convergence in the
factored space. We build our theory on a novel descent lemma, that non-trivially extends recent
results on the unconstrained problem. The resulting algorithm is Projected Factored Gradient
Descent, abbreviated as ProjFGD, and shows superior performance compared to state of the art on
quantum state tomography and sparse phase retrieval applications.
1
Introduction
We consider matrix problems of the form:
min
X∈Rn×n
f (X) subject to X 0, X ∈ C 0 .
(1)
f is assumed to be strongly convex and have Lipschitz continuous gradients. The constraint set contains
PSD and additional C 0 ⊆ Rn×n convex constraints on X.
There are several algorithmic solutions for (1), operating on the variable space X ∈ Rn×n . We
mention [27, 7, 4, 9, 45, 32, 30, 28, 43, 24, 47, 34, 6, 13, 5, 17, 46] and point to references therein. Most
of these schemes focus on the matrix sensing / matrix completion problem, and, thus, are designed for
specific instances of f . Moreover, the majority does not directly handle additional constraints.
More importantly, these methods often involve computationally expensive eigen-value/vector computations –at least once per iteration– in order to satisfy the PSD constraint. This constitutes their
computational bottleneck in large-scale settings. Thus, it is desirable to find algorithms that scale well
in practice.
1
Our approach. One way to avoid this is by positive semi-definite factorization X = U U > . In
particular, we solve instances of (1) in the factored form as follows:
minimize f (U U > )
subject to U ∈ C.
U ∈Rn×r
(2)
This parametrization, popularized by Burer and Monteiro [11, 12], naturally encodes the PSD constraint,
removing the expensive eigen-decomposition projection step. r can be set to rank(X ? ) := r? , where X ?
is the optimal solution to (1), but is often set to be much smaller than r? : in that case, factor U ∈ Rn×r
contains much less variables to maintain and optimize than X = U U > . Thus, such parametrization
also makes it easier to update and store the iterates U . By construction, the U U > is a PSD solution,
which could also be low-rank.
C ⊆ Rn×r is a compact convex set, that models well C 0 in (1). While in practice we can assume any
such constraint C with tractable Euclidean projection operator1 , in our theory we mostly focus on sets
C that satisfy the following assumption.
Assumption 1.1. For X 0, there is U ∈ Rn×r and r ≤ n such that X = U U > . Then, C 0 ∈ Rn×n
is endowed with constraint set C ⊆ Rn×r that (i) for each X ∈ C 0 , there is an subset in C where each
U ∈ C satisfies X = U U > (see Section 3.1 for more details), and (ii) its projection operator, say ΠC (V )
for V ∈ Rn×r , is an entrywise scaling operation on the input V .
Part (i) in Assumption 1.1 is required in our analysis in order to claim convergence also in the X
space, through the factored U space. In other case, our theory still leads just to convergence in U . We
defer this discussion to Section 3.
Criteria of the form (1) appear in applications from diverse research fields. As an exemplar, consider
density matrix estimation of quantum systems [1, 23, 29], where C 0 is defined as Tr(X) ≤ 1 and satisfies
Assumption 1.1. We also experimentally consider problems that only approximately satisfy the above
assumption, such as sparse phase retrieval applications [26, 42, 38], and sparse PCA [31]: in these cases,
C 0 contains `1 -norm constraints on X; see Section 4 for more details.
Contributions. Our aim is to broaden the results on efficient, non-convex recovery for constrained
low-rank matrix problems. Our developments maintain a connection with analogous results in convex
optimization, where standard assumptions are made: we consider the common case where f is (restricted)
smooth and (restricted) strongly convex [2]. We provide experimental results for two important tasks
in physical sciences: quantum state density estimation and sparse phase retrieval.
Some highlights of our results are the following:
• A key property for proving convergence in convex optimization is the notion of descent. I.e., given
current, next and optimal points in the X space, say Xt , Xt+1 , X ? , respectively, and the recursion
Xt+1 = Xt − η∇f (Xt ), the condition hXt − Xt+1 , Xt − X ? i ≥ C –where C > 0 depends on the
gradient norm– implies that Xt+1 “moves" towards the correct direction.2 In this work, we present a
novel descent lemma that non-trivially extends such conditions for the constrained case in (2), under
the assumptions mentioned above. We hope that this result will trigger more attempts towards more
generic convex sets.
• We propose ProjFGD, a non-convex projected gradient descent algorithm that solves instances of (2).
ProjFGD has favorable local convergence guarantees when f is (restricted) smooth and (restricted)
strongly convex. We also present an initialization procedure with guarantees in the supplementary
material.
In general, one could artificially introduce a constraint C in (2) –even when no C 0 constraint is present in (1)– for
better interpretation of results.
2
To see this, observe that X ? − Xt is the best possible direction to follow, while Xt − Xt+1 is the direction we actually
follow. Then, such a condition implies that there is a non-trivial positive correlation between these two directions.
1
2
• Finally, we extensively study the performance of ProjFGD on two problem cases: (i) quantum state
tomography and (ii) sparse phase retrieval. Our findings show significant acceleration when ProjFGD
is used, as compared to state of the art.
1.1
Related work
The work of [18] proposes a first-order algorithm for (2), where the nature of C is more generic, and
depends on the problem at hand. The authors provided a set of conditions (local descent, local Lipschitz,
and local smoothness) under which one can prove convergence to an ε-close solution with O(1/ε) or
O(log(1/ε)) iterations. While the convergence proof is general, checking whether the three conditions
hold is a non-trivial problem and requires different analysis for each problem. We believe this paper
complements [18]: in the latter, the closest to our constraints are these of “max-norm" incoherence
constraints; however, in our case, the objective function needs to only satisfy standard strongly convex
and smoothness assumptions.
[8] proposes the Factored Gradient Descent (FGD) algorithm for (2), where C ≡ Rn×r . FGD is also a
first-order scheme. Key ingredient for convergence is a novel step size selection that can be used for any
f , as long as it is Lipschitz gradient smooth (and strongly convex for faster convergence). However, [8]
cannot accommodate any constraints on U .
Concurrently, [48] presents a new analysis that handles non-square cases in (2). In that case, we look
for a factorization X = U V > ∈ Rn×p . The idea is based on the inexact first-order oracle, previously
used in [3]. Similarly to [8], the proposed theory does not handle any constraints.
Roadmap. Section 2 contains some basic definitions and assumptions that are repeatedly used in the
main text. Section 3 describes ProjFGD and its theoretical guarantees. In Section 4, we motivate the
necessity of ProjFGD via some applications; due to space limitations, only one application is described
in the main text (the second application is included in the supplementary material). This paper
concludes with a discussion on future directions in Section 5. Supplementary material contains further
experiments, all proofs of theorems in main text, and a proposed initialization procedure.
2
Preliminaries
Notation. For matrices X, Y ∈ Rn×n , hX, Y i = Tr X > Y represents their inner product. X 0
denotes X is a positive semi-definite (PSD) matrix. We use kXkF and σ1 (X) for the Frobenius and
spectral norms of a matrix, respectively; we also use kXk2 to denote the spectral norm. Moreover, we
denote as σi (X) the i-th singular value of X. Xr denotes the best rank-r approximation
of X. For X
>
>
>
>
such that X = U U , the gradient of f with respect to U is ∇f (U U ) + ∇f (U U ) U . If f is also
symmetric, i.e., f (X) = f (X > ), then ∇f (X) = 2∇f (X) · U .
An important issue in optimizing f over the factored space is the existence of non-unique possible
factorizations. We use the following rotation invariant distance metric:
Definition 2.1. Let matrices U, V ∈ Rn×r . Define:
Dist (U, V ) := min kU − V RkF ,
R:R∈O
where O is the set of r × r orthonormal matrices R.
Assumptions. We consider applications that can be described by strongly convex functions f with
gradient Lipschitz continuity.3 We state these standard definitions below for the square case.
3
Our ideas can be extended in a similar fashion to the case of restricted smoothness and restricted strong convexity [2].
3
Definition 2.2. Let f : Rn×n → R be convex and differentiable. f is µ-strongly convex if ∀X, Y ∈ Rn×n
f (Y ) ≥ f (X) + h∇f (X) , Y − Xi +
µ
2
kY − Xk2F .
(3)
Definition 2.3. Let f : Rn×p → R be a convex differentiable function. f is gradient Lipschitz continuous
with parameter L (or L-smooth) if ∀X, Y ∈ Rn×n
k∇f (X) − ∇f (Y )kF ≤ L · kX − Y kF .
(4)
For our proofs, we will also make the faithfulness assumption, as in [18]:
Definition
Let E denote the set of equivalent factorizations that lead to a rank-r matrix X ? ∈ Rn×n ;
2.4.
?
i.e., E := U ∈ Rn×r : X ? = U ? U ?> . Then, we assume E ⊆ C, i.e., the resulting convex set C in
(2) (from C 0 in (1)) respects the structure of E.
This assumption is necessary for arguments regarding the quality of solution obtained in the factored
U space, w.r.t. the original X space.
3
The Projected Factored Gradient Descent (ProjFGD) algorithm
Let us first describe the ProjFGD algorithm, a projected, first-order scheme. The discussion in this part
holds for any constraint set C; later in the text, in order to obtain theoretical guarantees, we make
further assumptions—such as Assumption 1.1.
The pseudocode is provided in Algorithm 1. Let ΠC (V ) denote the projection of an input matrix
V ∈ Rn×r onto the convex set C. For initialization, the starting point is computed as follows: we first
compute X0 := 1/Lb · Π+ (−∇f (0)), where Π+ (·) denotes the projection onto the set of PSD matrices
b represents an approximation of L. Then, ProjFGD requires a top-r SVD calculation, only once,
and L
n×r
>
e0 , in
e
e
e
e
to compute U0 ∈ R , such that X0 = U0 U0 ; using U0 , the initial point U0 satisfies U0 = ΠC U
order to accommodate constraints C.
The main iteration of ProjFGD applies the simple rule:
Ut+1 = ΠC Ut − η∇f (Ut Ut> ) · Ut ,
with step size:
η≤
1
.
128(LkX0 k2 +k∇f (X0 )k2 )
(5)
b to approximate L.
Here, one can use L
Key ingredients to achieve provable convergence are initialization–so that initial point U0 leads
to Dist(U0 , U ? ) sufficiently small– and the step size selection. For the initialization, apart from the
procedure mentioned above, we could also use more specialized spectral methods –see [18, 49]– or
even run algorithms on (1) for only a few iterations –this requires further full or truncated SVDs [44].
The discussion regarding our initialization and what type of guarantees one obtains is deferred to the
supplementary material.
3.1
When constrained non-convex problems can be scary?
In stark contrast to the convex projected gradient descent method, proving convergence guarantees for
(2) is not a straightforward task. First, if we are interested in quantifying the quality of the solution
in the factored space w.r.t. X ? , C should be faithful, according to Definition 2.4. Furthermore, there
should exist a mapping U 7→ X that relates the constraint set C 0 , in the original variable space (see (1)),
4
Algorithm 1 ProjFGD method
1: Input: Function f , target rank r, # iterations T .
b · Π+ (−∇f (0)).
2: Compute X0 := 1/L
n×r
e
e e>
3: Set U0 ∈ R
such that
X0 = U0 U0 .
e0 .
4: Compute U0 = ΠC U
Set step size η as in (5).
for t = 0 to T − 1 do
7:
Ut+1 = ΠC Ut − η∇f (Ut Ut> ) · Ut .
8: end for
9: Output: X = UT UT> .
5:
6:
to the factored one C (see (2)). In that case, claims about convergence to a point U ? , in the factored
space, can be “transformed" into claims about convergence to a point close to X ? , in the original space,
that also satisfies the constraints. This is the case for the following constraint case: for any X = U U > ,
Tr(X) ≤ λ ⇔ kU k2F ≤ λ, and, thus, satisfying kU k2F ≤ λ, for any U , guarantees that Tr(X) ≤ λ for
X = U U > . Apart from the example above, other characteristic cases include Schatten norms.
0
∈ Rn×n : kXk1 ≤ λ0 }, where, kXk1 =
P Contrary to this example, consider the case C := {X
n×r : kU k ≤ λ}, for λ, λ0 > 0; however,
1
ij |Xij |. A natural choice for C would be C := {U ∈ R
depending on the selection of λ, points in U ∈ C might result into points X = U U > that X 6∈ C 0 . In
this case, U ? of (2) could be 6∈ E and, thus, convergence guarantees to U ? might have no meaning in
the convergence in X space. However, as we show in Section 6.1, in this case C “simulates" well C 0 : if U
is sparse enough, then X = U U > could also be sparse, so proper selection of λ plays a key role. Even
in this case, ProjFGD performs competitively compared to state-of-the-art approaches.
Second, the projection step itself complicates considerably the non-convex analysis, as we show in
the supplementary material. In our theory, we focus on convex sets C that satisfy (6) where ΠC (V ) can
be equivalently seen as scaling the input. E.g., when C ≡ {U ∈ Rn×r : kU kF ≤ λ}, ΠC (V ) = ξ(V ) · V
where ξ(V ) := kVλkF , for V 6∈ C. Our theory highlights that, even for this simple case, proving
convergence is not a straightforward task.
3.2
Theoretical guarantees of ProjFGD for C := {U ∈ Rm×r : kU kF ≤ λ}
We provide theoretical guarantees for ProjFGD in the case where the constraint satisfies
V
if V ∈ C,
2
1
ΠC (V ) = argmin 2 kU − V kF =
ξ(V ) · V if V 6∈ C,
U ∈C
(6)
i.e., the projection operation is an entry-wise scaling. Such settings include the Frobenius norm
constraint C = {U ∈ Rm×r : kU kF ≤ λ}, which appears in quantum state tomography. Moreover, for
this case, the constraint has one-to-one correspondence with the trace constraint in the original X
space; thus any argument in the U space applies for the X space also.
We assume the optimum X ? satisfies rank(X ? ) = r? . For our analysis, we will use the following
step sizes:
ηb =
1
,
128(LkXt k2 +kQUt Q>
U ∇f (Xt )k2 )
?
1
128(LkX ? k2 +k∇f (X ? )k2 ) ,
η =
t
where QA is a basis for column space of A. By Lemma A.5 in [8], we know that ηb ≥ 56 η and
10 ?
11 ?
b: this is equivalent
11 η ≤ η ≤ 10 η . Due to such relationships, in our proof we will work with step size η
–up to constants– to the original step size η, used in the algorithm. Thus, any results below will
automatically imply similar results hold for η, by using the bounds between step sizes.
5
Theorem 3.1 ((Local) Convergence rate for restricted strongly convex and smooth f ). Let C ⊆ Rn×r
be a convex, compact, and faithful set, with projection operator satisfying (6). Let Ut ∈ C be the
current estimate and Xt = Ut Ut> . Assume current point Ut satisfies Dist(Ut , U ? ) ≤ ρ0 σr (U ? ), for
(X ? )
1
ρ0 := c · Lµ · σσr1 (X
? ) , c ≤ 200 , and given ξt (·) & 0.78 per iteration, the new estimate of ProjFGD,
Ut+1 = ΠC Ut − ηb∇f (Ut Ut> ) · Ut = ξt · Ut − ηb∇f (Ut Ut> ) · Ut satisfies
(7)
Dist(Ut+1 , U ? )2 ≤ α · Dist(Ut , U ? )2 ,
where α := 1 −
µ·σr (X ? )
550(LkX ? k2 +k∇f (X ? )k2 )
< 1. Further, Ut+1 satisfies Dist(Ut+1 , U ? ) ≤ ρ0 σr (U ? ).
The complete proof of the theorem is provided in the supplementary material. The assumption
Dist(Ut , U ? ) ≤ ρ0 σr (U ? ) only leads to a local convergence result. [18] provide some initialization
procedures for different applications, where we can find an initial point U0 such that Dist(U0 , U ? ) ≤
ρ0 σr (U ? ) is satisfied. In the supplementary material, we present a similar generic initialization procedure
that results in exact recovery of the optimum, under further assumptions. We borrow such procedure
in Section 4 for our experiments.
ξt (·) requirement. The assumption ξt (·) & 0.78 implies the iterates of ProjFGD (before the projection
step) are retained relatively close to the set C.4 For some cases, this can be easily satisfied by setting
the step size small enough, as indicated below; the proof can be found in Section 7.
Corollary 3.2. If C = {U ∈ Rn×r : kU kF ≤ λ}, then ProjFGD inherently satisfies
every t. I.e., it guarantees (7) without assumptions on ξt (·).
128
129
≤ ξt (·) ≤ 1, for
We conjecture that the lower bound on ξt (·) could possibly be improved with a different analysis.
Key lemma. The proof of the above theorem primarily depends on the following “descent" lemma for
ProjFGD.
et+1 = Ut − ηb∇f (X > ) · Ut . For f L-smooth and µ-strongly convex,
Lemma 3.3 (Descent lemma). Let U
t
and under the same assumptions with Theorem 3.1, the following inequality holds true:
?
et+1 k2F ≥
2b
η ∇f (Ut Ut> ) · Ut , Ut − U ? RU
+ kUt+1 − U
t
ηb2 k∇f (Ut Ut> )Ut k2F +
3.3
3b
ηµ
10
· σr (X ? ) · Dist(Ut , U ? )2 .
Main differences with [8]
In this subsection, we highlight the main differences with the analysis of [8].
As we already mentioned, the proof in [8] does not handle constraints. In particular, one key factor
is the gradient at an optimal point does not vanish and thus the following descent lemma bound (Lemma
6.1 in [8]) does not hold:
?
2b
η ∇f (Ut Ut> ) · Ut , Ut − U ? RU
≥
t
4b
η2
>
2
3 k∇f (Ut Ut )Ut kF
+
3b
ηµ
20
· σr (X ? ) · Dist(Ut , U ? )2 .
(8)
To see this, in the unconstrained case, if Ut ≡ U ? (up to some rotation), the following holds for [8]
?
0 := 2b
η ∇f (X ? ) · U ? , U ? − U ? RU
≥
t
+
3b
ηµ
20
· σr (X ? ) · Dist(U ? , U ? )2 =: 0,
4b
η2
?
? 2
3 k∇f (X )U kF
since at the optimum we have ∇f (X ? )U ? = 0. When the latter does not hold, this descent lemma
does not hold. As a simple example where this happens, consider the noisy matrix sensing setting:
4
Intuitively, we expect the estimates Ut , before the projection, to be further from C during the first steps of ProjFGD;
as the number of iterations increases, the sequence of solutions gets closer to U ? and thus ξt (·) → 1.
6
Rank r = 1 - Samples m = 3 " n " r
Rank r = 1 - Samples m = 3 " n " r
AccUniPDGrad
ProjFGD
AccUniPDGrad
ProjFGD
Rank r = 20 - Samples m = 3 " n " r
AccUniPDGrad
ProjFGD
10 0
-2
200
400
600
10 -1
10 -2
10
20
30
0
10 -1
2000 4000 6000 8000 100001200014000
2000 4000 6000 8000 10000 12000
40
Cumulative time (sec)
Cumulative time (sec)
10-2
-2
0
SparseApproxSDP
10
800 1000
Number of iterations
10
10-1
RSVP
SparseApproxSDP
b!X ? kF
kX
kX ? kF
10
b!X ? kF
kX
kX ? kF
RSVP
SparseApproxSDP
b!X ?kF
kX
kX ? kF
10
-1
b!X ?kF
kX
kX ? kF
10-1
10 0
b!X ?kF
kX
kX ? kF
-1
b!X ? kF
kX
kX ? kF
10
0
Number of iterations
200
400
10-2
600
Cumulative time (sec)
0
100
200
300
400
Cumulative time (sec)
Figure 1: Left and middle panels: Convergence performance of algorithms under comparison w.r.t.
?
b
kX−X
kF
kX ? kF
vs. (i) the total number of iterations (left) and (ii) the total execution time. Both cases correspond to Csam = 3,
r = 1 (pure setting) and q = 12 (i.e., n = 4096). Right panel: Almost pure state (r = 20). Here, Csam = 3.
y = A(X ? ) + w, where w is a non-negligible noise term. Then, for objective f (X) = ky − A(X)k22 ,
observe that
∇f (X ? ) = −2A∗ (y − A(X ? )) = −2A∗ (w) 6= 0,
where A∗ is the adjoint operator for A. Finally, in [8], Lemma 6.3 (pp.23) assumes zero gradient at the
optimum, while our Lemma 5.3 (supp.material) follows a different approach.
4
Applications
We present two characteristic applications. For each application, we define the problem, enumerate
state-of-the-art algorithms and provide numerical results. We refer the reader to Section 6 for additional
experiments.
4.1
Quantum state tomography
Building on Aaronson’s work on quantum state tomography (QST) [1], we are interested in learning
the (almost) pure 5 q-bit state of a quantum system –known as the density matrix– via a limited
set of measurements. In math terms, the problem can be cast as follows. Let us define the density
matrix X ? ∈ Cn×n of a q-bit quantum system as an unknown Hermitian, positive semi-definite matrix
that satisfies rank(X ? ) = r and is normalized as Tr(X ? ) = 1 [23]; here, n = 2q . Our task is to
recover X ? from a set of QST measurements y ∈ Rm , m n2 , that satisfy y = A(X ? ) + η. Here,
(A(X ? ))i = Tr(Ei X ? ) and ηi can be modeled as independent, zero-mean normal variables. The
operators Ei ∈ Rn×n are typically the tensor product of the 2 × 2 Pauli matrices6 [35].
The above lead to the following non-convex problem formulation7 :
minimize kA(X) − yk2F
X0
subject to rank(X) = r, Tr(X) ≤ 1.
(10)
5
Purity is a structural property of the density matrix: A quantum systems is pure if its density matrix is rank one
and, almost pure if it can be well-approximated by a low rank matrix.
6
[35] showed that, for almost all such tensor constructions –of m = O(rn logc n), c > 0, Pauli measurements– satisfy
√
the so-called rank-r restricted isometry property (RIP) for all X ∈ {X : X 0, rank(X) ≤ r, kXk∗ ≤ rkXkF }:
(1 − δr ) kXk2F ≤ kA(X)k2F ≤ (1 + δr ) kXk2F ,
(9)
where k · k∗ is the nuclear norm (i.e., the sum of singular values), which reduces to Tr(X) since X 0.
7
As pointed out in [29], it is in fact advantageous in practice to choose Tr(X) 6= 1, as it improves the robustness to
noise. Here, we force Tr(X) ≤ 1.
7
State-of-the-art approaches. One of the first provable algorithmic solutions for the QST problem
was through convexification [40]: this includes nuclear norm minimization approaches [23], as well as
proximal variants, as follows:
minimize kA(X) − yk2F + λkXk∗ .
X0
(11)
Here, k · k∗ reduces to Tr(X) since X 0. This approach is considered in the seminal work [23] and is
both tractable and amenable to theoretical analysis. The approach does not include any constraint on
X.8 As one of the most recent algorithms, we mention the work of [46] where a universal primal-dual
convex framework is presented, with the QST problem as application.
From a non-convex perspective, [25] presents SparseApproxSDP algorithm that solves (10), when
the objective is a generic gradient Lipschitz smooth function. SparseApproxSDP solves (10) by updating a putative low-rank solution with rank-1 refinements, coming from the gradient. This way,
SparseApproxSDP avoids computationally expensive operations per iteration, such as full SVDs. In
theory, at the r-th iteration, SparseApproxSDP is guaranteed
to compute a 1r -approximate solution,
1
with rank at most r, i.e., achieves a sublinear O ε convergence rate. However, depending on ε,
SparseApproxSDP might not return a low rank solution. Finally, [7] propose Randomized Singular Value
Projection (RSVP), a projected gradient descent algorithm for (10), which merges gradient calculations
with truncated SVDs via randomized approximations for computational efficiency.
Since the size of these problems grows exponentially with the number of quantum bits, designing
fast algorithms that minimize the computational effort required for (10) or (11) is mandatory.
Numerical results. In this case, the factorized version of (10) can be described as:
minimize kA(U U > ) − yk2F
subject to kU k2F ≤ 1.
U ∈Rn×r
(12)
We compare ProjFGD with the algorithms described above; as a convex representative implementation,
we use the efficient scheme of [46]. We consider two settings: X ? ∈ Rn×n is (i) a pure state (i.e.,
rank(X ? ) = 1) and, (ii) an almost pure state (i.e., rank(X ? ) = r, for some r > 1). For all cases,
Tr (X ? ) = 1 and y = A(X ? ) + η, with kηk = 10−3 . We use Pauli operators for A, as described in [35].
The number of measurements m satisfy m = Csam · r · n log(n), for various values of Csam .
For all algorithms, we used the correct rank input and trace constraint parameter. All methods that
require an SVD routine use lansvd(·) from the PROPACK software package. Experiments and algorithms
are implemented on Matlab environment; we used non-specialized and non-mexified code parts for all
algorithms. For initialization, we use the same starting point for all algorithms, which is either specific
(Section 8) or random. We set the tolerance parameter to tol := 5 · 10−6 .
Convergence plots. Figure 1 (two-leftmost plots)
illustrates the iteration and timing complexities of each
Setting: q = 13, Csam = 3.
algorithm under comparison, for a pure state density
?
b
kX−X
kF
Algorithm
Time
kX ? kF
recovery setting (r = 1). Here, q = 12 which correAccUniPDGrad
7.4151e-02
2354.4552
sponds to a n(n+1)
= 8, 390, 656 dimensional problem;
2
ProjFGD
8.6309e-03
1214.0654
moreover, we assume Csam = 3 and thus the number
of measurements are m = 12, 288. For initialization,
we use the proposed initialization in Section 8 for all Figure 2: Comparison results for reconstruction
and efficiency. Time reported is in seconds.
algorithms: we compute −A∗ (y), extract factor U0 as
the best-r PSD approximation of −A∗ (y), and project
U0 onto C.
It is apparent that ProjFGD converges faster to a vicinity of X ? , compared to the rest of the
algorithms; observe also the sublinear rate of SparseApproxSDP in the inner plots, as reported in [25].
8
E.g., in order to take the trace constraint Tr(X) = 1 into account, either λ should be precisely tuned to satisfy this
constraint or the final estimator is normalized heuristically to satisfy this constraint [20].
8
8
6
4
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
2
0
4
3
2
1
-2
-4
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
5
Total time
Total time (log-scale)
6
10
6
8
10
q-bit (n = 2q )
0
12
1
5
20
Rank r
Figure 4: Timing bar plot: y-axis shows total execution time (log-scale) and x-axis corresponds to different q
values. Left panel corresponds to r = 1 and Csam = 6; right panel corresponds to q = 10 and Csam = 6.
Figure 2 contains recovery error and execution time results for the case q = 13 (n = 8096); in this
case, we solve a n(n+1)
= 33, 558, 528 dimensional problem. For this case, RSVP and SparseApproxSDP
2
algorithms were excluded from the comparison. Appendix provides extensive results, where similar
performance is observed for other values of q, Csam .
Figure 1 (rightmost plot) considers the more general case where r = 20 (almost pure state density)
and q = 12. In this case, m = 245, 760 for Csam = 3. As r increases, algorithms that utilize an SVD
routine spend more CPU time on singular value/vector calculations. Certainly, the same applies for
matrix-matrix multiplications; however, in the latter case, the complexity scale is milder than that of
the SVD calculations. Further metadata are also provided in Figure 3.
For completeness, in the appendix we also
provide results (for the noiseless case) that
Setting: r = 5.
Setting: r = 20.
illustrate the effect of random initialization:
?
?
b
b
kX−X
kF
kX−X
kF
Algorithm
Time
Time
kX ? kF
kX ? kF
Similar to above, ProjFGD shows competitive
behavior by finding a better solution faster,
RSVP
5.15e-02
0.78
1.71e-02
0.38
SparseApproxSDP
3.17e-02
3.74
5.49e-02
4.38
irrespective of initialization point.
AccUniPDGrad
2.01e-02
0.36
1.54e-02
0.33
Timing evaluation (total and per iteraProjFGD
1.20e-02
0.06
7.12e-03
0.04
tion). Figure 4 highlights the efficiency of
our algorithm in terms of time complexity,
Figure 3: Results for reconstruction and efficiency. Time for various problem configurations. Our algoreported is in seconds. For all cases, Csam = 3 and q = 10.
rithm has fairly low per iteration complexity
(where the most expensive operation for this
problem is matrix-matrix and matrix-vector multiplications). Since our algorithm shows also fast
convergence in terms of # of iterations, this overall results into faster convergence towards a good
approximation of X ? , even as the dimension increases. Figure 4(right) shows how the total execution
time scales with parameter r.
Overall performance. ProjFGD shows a competitive performance, as compared to the state-of-the-art
algorithms; we would like to emphasize also that projected gradient descent schemes, such as [7],
are also efficient in small- to medium-sized problems, due to their fast convergence rate. Moreover,
convex approaches might show better sampling complexity performance (i.e., as Csam decreases). For
more experimental results (under noiseless settings), we defer the reader to Appendix, due to space
restrictions.
4.2
Sparse phase retrieval
Consider the sparse phase retrieval (SPR) problem [16, 14, 33]: we are interested in recovering a (sparse)
unknown vector x? ∈ Cn , via its lifted, rank-1 representation X ? = x? x?H ∈ Cn×n , from a set of
9
Original image
10 1
Wirtiger flow
ProjFGD
10 0
ATime = 140.1129 (sec)
10 -1
b!X ? kF
kX
kX ? kF
Time = 134.646 (sec)!
10
ATime = 206.6236 (sec)
-2
Time = 202.8971 (sec)!
ATime = 272.2579 (sec)
10 -3
Time = 268.1068 (sec)!
10 -4
10
-5
10
-6
0
20
40
60
80
100
120
140
Iterations
Figure 5: Left panel: Original image for the sparse phase retrieval problem. The dimension of the
image is 883 × 1280 = 1, 130, 240 variables. Right panel: Convergence performance of algorithms under
?k
b
F
comparison w.r.t. kX−X
vs. the total number of iterations; markers on top of curves indicate the
kX ? kF
total execution time until that point.
quadratic measurements:
yi = Tr(aH
i Xai ) + ηi ,
i = 1, . . . , m.
Here, ai ∈ Cn are given measurement vectors (often Fourier vectors) and ηi is an additive error term.
The above description leads to the following non-convex optimization criterion:
minimize
X0
kA(X) − yk2F ,
subject to rank(X) = 1, (kXk1 ≤ λ) .
(13)
?
Here, A : Cn×n → Cm such that (A(X))i = Tr(Φi X) where Φi = ai aH
i . In the case where we know x
is sparse [33, 39], we can further constrain the lifted variable X to satisfy kXk1 ≤ λ, λ > 0; this way
we implicitly also restrict the number of non-zeros in its factors and can recover X ? from a limited set
of measurements.
Transforming (13) into a factored formulation. Given the rule X = uuH , where u ∈ Cn , one
can consider the factored problem re-formulation:
minimize
n
u∈C
kA(uuH ) − yk2F
subject to kuk1 ≤ λ0 .
(14)
for some λ0 > 0.
Remark 1. In contrast to the QST problem, where there is a continuous map between the constraints
in the original X space and in the factored U space ( i.e., Tr(X) ≤ λ ⇔ kU k2F ≤ λ), this is not true
for the SPR problem: As we state in the main text, points U ∈ C can result into X 6∈ C 0 , depending on
the selection of λ, λ0 values (i.e., C is unfaithful). In this case, the convergence theorem 3.1 in the U
factor space only proves convergence to a point U ? in the factored space, which is not necessarily related
to the optimal point X ? in the original space. However, as we show next, in practice, even in this case
ProjFGD returns a competitive (if not better) solution, compared to state-of-the-art approaches.
10
Wirtiger .ow - 80 iterations - PSNR = 37.3687 (dB)
Wirtiger .ow - 100 iterations - PSNR = 54.4213 (dB)
ProjFGD - 80 iterations - PSNR = 42.8286 (dB)
ProjFGD - 100 iterations - PSNR = 62.7705 (dB)
Figure 6: First two figures: Reconstructed image by using Wirtiger flow algorithm after 80 iterations
(left panel) and 100 iterations (right panel). Last two figures: Reconstructed image by using ProjFGD
algorithm after 80 iterations (left panel) and 100 iterations (right panel).
State-of-the-art approaches. One of the most widely used methods for the phase retrieval problem
comes from the seminal work of Gerchberg-Saxton [22] and Fienup [19]: they propose a greedy scheme
that alternates projections on the range of {ai }m
i=1 and on the non-convex set of vectors b such that
b = |Ax|. Main disadvantage of such greedy methods is that often they get stuck to locally minimum
points.
An popularized alternative to these greedy methods is via semidefinite relaxations. [14] proposes
PhaseLift, where the rank constraint is replaced by the nuclear norm surrogate. However, it is
well-known that such SDP relaxations can be computationally prohibitive, when solved using off-theself software packages, even for small problem instances; some specialized and more efficient convex
relaxation algorithms are given in [21].
In [15], the authors present Wirtinger Flow algorithm, a non-convex scheme for solving phase
retrieval problems. Similar to our approach, Wirtinger Flow consists of three components: (i) a
careful initialization step using a spectral method, (ii) a specialized step size selection and, (iii) a
recursion where gradient steps on the factored variable space are performed. Other approaches include
Approximate Message Passing algorithms [41] and ADMM approaches [39].
Numerical results. We test our algorithm on image recovery, according to the description given in
[15, Section 4.2]. Here, we consider grayscale images that are by nature also sparse (Figure 5 - left
panel). This way, we can also consider `1 -norm constraints, as in the criterion (14). We generate L = 21
random octanary patterns and, using these 21 samples, we obtain the coded diffraction patterns using
the grayscale image as input. As dictated by [15, Section 4.2], we perform 50 power method iterations
for initialization.
For this experiment, we highlight (i) how our algorithm ProjFGD performs in practice, and (ii) how
the additional sparsity constraint could lead to better performance. Figure 5 (right panel) depicts the
?
kF
relative error kX−X
w.r.t. the iteration count for two algorithms: (i) Wirtiger flow [15], and (ii)
kX ? kF
ProjFGD. We observe that ProjFGD shows a slightly better performance, compared to Wirtiger flow,
both in terms of iterations –i.e., we reach to a better solution within the same number of iterations–
and in terms of execution time –i.e., given a time wall, ProjFGD returns an estimate of better quality
within the same amount of time. We note that both algorithms used step sizes that were slightly
different in values, while ProjFGD further performs also a projection step.9 Figure 6 shows some
reconstructed images returned by the algorithms under comparison, during their execution. In all cases,
both algorithms perform appealingly, finding a good approximation of the original image in less than
5 minutes; comparing the two algorithms, we note that ProjFGD returns a solution, within the same
number of iterations, with at least 5 dB higher Peak Signal to Noise Ration (PSNR), in less time.
b
9
The step size in Wirtiger flow satisfies η :=
µt
kU0 k2
F
n
o
, for µt = min 1 − et/t0 , 0.4 and t0 ≈ 330.
11
5
Discussion
We consider a class of low-rank matrix problems where the solution is assumed PSD and further
constrained with some matrix constraints, described in the text. This paper proposes ProjFGD, a
non-convex projected gradient descent algorithm that operates on the factors of the PSD putative
solution. When the objective function is smooth and strongly convex in the original variable space,
ProjFGD has (local) linear rate convergence guarantees (which can become global, given a proper
initialization).
Our current analysis restricts to the Assumption 1.1; extending the proof for more complex constraints
sets is one possible research direction for future work, where an analogous of gradient mapping [37]
might be required. Furthermore, considering barrier functions in the objective function, in order to
accommodate the constraints, could be a possible extension. We hope this work will trigger future
attempts along these directions.
References
[1] S. Aaronson. The learnability of quantum states. In Proceedings of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, volume 463, pages 3089–3114, 2007.
[2] A. Agarwal, S. Negahban, and M. Wainwright. Fast global convergence rates of gradient methods for
high-dimensional statistical recovery. In Advances in NIPS, pages 37–45, 2010.
[3] S. Balakrishnan, M. Wainwright, and B. Yu. Statistical guarantees for the EM algorithm: From population
to sample-based analysis. arXiv preprint arXiv:1408.2156, 2014.
[4] L. Balzano, R. Nowak, and B. Recht. Online identification and tracking of subspaces from highly incomplete
information. In Communication, Control, and Computing (Allerton), 2010 48th Annual Allerton Conference
on, pages 704–711. IEEE, 2010.
[5] S. Becker, J. Bobin, and E. Candès. NESTA: A fast and accurate first-order method for sparse recovery.
SIAM Journal on Imaging Sciences, 4(1):1–39, 2011.
[6] S. Becker, E. Candès, and M. Grant. Templates for convex cone problems with applications to sparse signal
recovery. Mathematical Programming Computation, 3(3):165–218, 2011.
[7] S. Becker, V. Cevher, and A. Kyrillidis. Randomized low-memory singular value projection. In 10th
International Conference on Sampling Theory and Applications (Sampta), 2013.
[8] S. Bhojanapalli, A. Kyrillidis, and S. Sanghavi. Dropping convexity for faster semi-definite optimization. In
29th Annual Conference on Learning Theory, pages 530–582, 2016.
[9] N. Boumal and P.-A. Absil. RTRMC: A Riemannian trust-region method for low-rank matrix completion.
In Advances in neural information processing systems, pages 406–414, 2011.
[10] S. Bubeck. Theory of convex optimization for machine learning. arXiv preprint arXiv:1405.4980, 2014.
[11] S. Burer and R. Monteiro. A nonlinear programming algorithm for solving semidefinite programs via
low-rank factorization. Mathematical Programming, 95(2):329–357, 2003.
[12] S. Burer and R. Monteiro. Local minima and convergence in low-rank semidefinite programming. Mathematical Programming, 103(3):427–444, 2005.
[13] J. Cai, E. Candès, and Z. Shen. A singular value thresholding algorithm for matrix completion. SIAM
Journal on Optimization, 20(4):1956–1982, 2010.
[14] E. Candes, Y. Eldar, T. Strohmer, and V. Voroninski. Phase retrieval via matrix completion. SIAM Review,
57(2):225–251, 2015.
[15] E. Candes, X. Li, and M. Soltanolkotabi. Phase retrieval via Wirtinger flow: Theory and algorithms.
Information Theory, IEEE Transactions on, 61(4):1985–2007, 2015.
12
[16] E. Candes, T. Strohmer, and V. Voroninski. Phaselift: Exact and stable signal recovery from magnitude
measurements via convex programming. Communications on Pure and Applied Mathematics, 66(8):1241–
1274, 2013.
[17] Y. Chen, S. Bhojanapalli, S. Sanghavi, and R. Ward. Coherent matrix completion. In Proceedings of The
31st International Conference on Machine Learning, pages 674–682, 2014.
[18] Y. Chen and M. Wainwright. Fast low-rank estimation by projected gradient descent: General statistical
and algorithmic guarantees. arXiv preprint arXiv:1509.03025, 2015.
[19] C. Fienup and J. Dainty. Phase retrieval and image reconstruction for astronomy. Image Recovery: Theory
and Application, pages 231–275, 1987.
[20] S. Flammia, D. Gross, Y.-K. Liu, and J. Eisert. Quantum tomography via compressed sensing: Error
bounds, sample complexity and efficient estimators. New Journal of Physics, 14(9):095022, 2012.
[21] F. Fogel, I. Waldspurger, and A. d’Aspremont. Phase retrieval for imaging problems. arXiv preprint
arXiv:1304.7735, 2013.
[22] R. Gerchberg. A practical algorithm for the determination of phase from image and diffraction plane
pictures. Optik, 35:237, 1972.
[23] D. Gross, Y.-K. Liu, S. Flammia, S. Becker, and J. Eisert. Quantum state tomography via compressed
sensing. Physical review letters, 105(15):150401, 2010.
[24] M. Hardt and M. Wootters. Fast matrix completion without the condition number. In Proceedings of The
27th Conference on Learning Theory, pages 638–678, 2014.
[25] E. Hazan. Sparse approximate solutions to semidefinite programs. In LATIN 2008: Theoretical Informatics,
pages 306–316. Springer, 2008.
[26] K. Jaganathan, S. Oymak, and B. Hassibi. Sparse phase retrieval: Convex algorithms and limitations. In
Information Theory Proceedings (ISIT), pages 1022–1026. IEEE, 2013.
[27] P. Jain, R. Meka, and I. Dhillon. Guaranteed rank minimization via singular value projection. In Advances
in Neural Information Processing Systems, pages 937–945, 2010.
[28] P. Jain, P. Netrapalli, and S. Sanghavi. Low-rank matrix completion using alternating minimization. In
Proceedings of the 45th annual ACM symposium on Symposium on theory of computing, pages 665–674.
ACM, 2013.
[29] A. Kalev, R. Kosut, and I. Deutsch. Quantum tomography protocols with positivity are compressed sensing
protocols. Nature partner journals (npj) Quantum Information, 1:15018, 2015.
[30] A. Kyrillidis and V. Cevher. Matrix recipes for hard thresholding methods. Journal of mathematical imaging
and vision, 48(2):235–265, 2014.
[31] S. Laue. A hybrid algorithm for convex semidefinite optimization. In Proceedings of the 29th International
Conference on Machine Learning (ICML-12), pages 177–184, 2012.
[32] K. Lee and Y. Bresler. ADMiRA: Atomic decomposition for minimum rank approximation. Information
Theory, IEEE Transactions on, 56(9):4402–4416, 2010.
[33] X. Li and V. Voroninski. Sparse signal recovery from quadratic measurements via convex programming.
SIAM Journal on Mathematical Analysis, 45(5):3019–3033, 2013.
[34] Z. Lin, M. Chen, and Y. Ma. The augmented Lagrange multiplier method for exact recovery of corrupted
low-rank matrices. arXiv preprint arXiv:1009.5055, 2010.
[35] Y.-K. Liu. Universal low-rank matrix recovery from Pauli measurements. In Advances in Neural Information
Processing Systems, pages 1638–1646, 2011.
[36] L. Mirsky. A trace inequality of John von Neumann. Monatshefte für Mathematik, 79(4):303–306, 1975.
[37] Y. Nesterov. Introductory lectures on convex optimization: A basic course, volume 87. Springer Science &
Business Media, 2013.
13
[38] H. Ohlsson, A. Yang, R. Dong, and S. Sastry. CPRL – an extension of compressive sensing to the phase
retrieval problem. In Advances in Neural Information Processing Systems, pages 1376–1384, 2012.
[39] H Ohlsson, A. Yang, R. Dong, and S. Sastry. CPRL – an extension of compressive sensing to the phase
retrieval problem. In Advances in Neural Information Processing Systems 25, pages 1376–1384, 2012.
[40] B. Recht, M. Fazel, and P. Parrilo. Guaranteed minimum-rank solutions of linear matrix equations via
nuclear norm minimization. SIAM review, 52(3):471–501, 2010.
[41] P. Schniter and S. Rangan. Compressive phase retrieval via generalized approximate message passing. Signal
Processing, IEEE Transactions on, 63(4):1043–1055, 2015.
[42] Y. Shechtman, A. Beck, and Y. Eldar. GESPAR: Efficient phase retrieval of sparse signals. Signal Processing,
IEEE Transactions on, 62(4):928–938, 2014.
[43] J. Tanner and K. Wei. Normalized iterative hard thresholding for matrix completion. SIAM Journal on
Scientific Computing, 35(5):S104–S125, 2013.
[44] S. Tu, R. Boczar, M. Soltanolkotabi, and B. Recht. Low-rank solutions of linear matrix equations via
Procrustes flow. arXiv preprint arXiv:1507.03566, 2015.
[45] Z. Wen, W. Yin, and Y. Zhang. Solving a low-rank factorization model for matrix completion by a nonlinear
successive over-relaxation algorithm. Mathematical Programming Computation, 4(4):333–361, 2012.
[46] A. Yurtsever, Q. Tran-Dinh, and V. Cevher. A universal primal-dual convex optimization framework. In
Advances in Neural Information Processing Systems 28, pages 3132–3140. 2015.
[47] D. Zhang and L. Balzano. Global convergence of a grassmannian gradient descent algorithm for subspace
estimation. arXiv preprint arXiv:1506.07405, 2015.
[48] T. Zhao, Z. Wang, and H. Liu. A nonconvex optimization framework for low rank matrix estimation. In
Advances in Neural Information Processing Systems 28, pages 559–567. 2015.
[49] Q. Zheng and J. Lafferty. A convergent gradient descent algorithm for rank minimization and SDP from
random linear measurements. In Advances in NIPS, pages 109–117, 2015.
Additional experiments
5.1
Quantum state tomography – more results
Figures 7-8 show further results regarding the QST problem, where r = 1 and q = 10, 12, respectively.
For each case, we present both the performance in terms of number of iterations needed, as well as
f0 ) such that
what is the cumulative time required. For all algorithms, we use as initial point U0 = ΠC (U
>
∗
e
e
X0 = U0 U0 where X0 = Π+ (−A (y)) and Π+ (·) is the projection onto the PSD cone. Configurations
are described in the caption of each figure. Table 1 contains information regarding total time required
for convergence and quality of solution for all these cases. Results on almost pure density states, i.e.,
r > 1, are provided in Figure 9.
For completeness, we also provide results that illustrate the effect of initialization. In this case, we
consider a random initialization and the same initial point is used for all algorithms. Some results are
illustrated in Figure 10; table 2 contains metadata of these experiments. Similar to above, ProjFGD
shows competitive behavior by finding a better solution faster, irrespective of initialization point.
14
800 1000
0
20
40
60
100
AccUniPDGrad
ProjFGD
10 0
10
-2
RSVP
SparseApproxSDP
10 -2
10 -4
800 1000
200
20
40
100
400
60
80
0
10
10
10 0
20
30
0
10 -2
10
10
0
10
20
30
40
50
0
10
RSVP
SparseApproxSDP
-2
200 400 600 800 10001200
10-4
40
AccUniPDGrad
ProjFGD
10-2
RSVP
SparseApproxSDP
500
1000 1500 2000
Cumulative time (sec)
Cumulative time (sec)
-4
800 1000
Number of iterations
Rank r = 1 - Samples m = 10 " n " r
AccUniPDGrad
ProjFGD
600
600
-4
10 -4
200
Cumulative time (sec)
400
Number of iterations
Number of iterations
Rank r = 1 - Samples m = 6 " n " r
10-2
RSVP
SparseApproxSDP
b!X ?kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10-2
600
10
0
Number of iterations
Rank r = 1 - Samples m = 3 " n " r
400
10 0
Number of iterations
10-4
b!X ? kF
kX
kX ? kF
10
200
AccUniPDGrad
ProjFGD
b!X ?kF
kX
kX ? kF
600
10 -2
b!X ?kF
kX
kX ? kF
400
Number of iterations
-4
Rank r = 1 - Samples m = 10 " n " r
100
10-2
RSVP
SparseApproxSDP
b!X ?kF
kX
kX ? kF
-2
200
10 0
10-2
RSVP
SparseApproxSDP
b!X ?kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10
AccUniPDGrad
ProjFGD
b!X ? kF
kX
kX ? kF
10 0
10-2
Rank r = 1 - Samples m = 6 " n " r
100
b!X ? kF
kX
kX ? kF
AccUniPDGrad
ProjFGD
b!X ?kF
kX
kX ? kF
Rank r = 1 - Samples m = 3 " n " r
0
b!X ? kF
kX
kX ? kF
10
10-4
0
20
Cumulative time (sec)
40
60
0
80
20
40
60
80
100
Cumulative time (sec)
Cumulative time (sec)
Figure 7: Quantum state tomography: Convergence performance of algorithms under comparison
?k
b
F
w.r.t. kX−X
vs. (i) the total number of iterations (top) and (ii) the total execution time (bottom).
kX ? kF
First, second and third column corresponds to Csam = 3, 6 and 10, respectively. For all cases, r = 1 (pure
f0 ) such that X0 = U
e0 U
e > where X0 = Π+ (−A∗ (y)).
state setting) and q = 10. Initial point is U0 = ΠC (U
0
Rank r = 1 - Samples m = 6 " n " r
AccUniPDGrad
ProjFGD
AccUniPDGrad
ProjFGD
10-2
10
-1
10 -2
200
400
600
10
b!X ? kF
kX
kX ? kF
RSVP
SparseApproxSDP
b!X ?kF
kX
kX ? kF
800 1000
Number of iterations
10
0
RSVP
SparseApproxSDP
-1
10 -2
20
30
10-4
0
40
20
40
100
AccUniPDGrad
ProjFGD
200
400
Cumulative time (sec)
10
40
600
60
AccUniPDGrad
ProjFGD
-2
RSVP
SparseApproxSDP
-2
10
-1
10 -2
10 -3
10-4
Cumulative time (sec)
0
100
10
10 -1
b!X ?kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
b!X ?kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10 -2
2000 4000 6000 8000 10000 12000
10
20
Number of iterations
Rank r = 1 - Samples m = 10 " n " r
10 0
RSVP
SparseApproxSDP
10-2
10 -1
-2
0
10 0
RSVP
SparseApproxSDP
300
-4
60
Number of iterations
Rank r = 1 - Samples m = 6 " n " r
200
Number of iterations
10
10 0
10-1
RSVP
SparseApproxSDP
100
Number of iterations
Number of iterations
Rank r = 1 - Samples m = 3 " n " r
AccUniPDGrad
ProjFGD
0
10 -2
b!X ? kF
kX
kX ? kF
10
10
10-2
100 200 300 400 500
10-2
0
AccUniPDGrad
ProjFGD
b!X ?kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10 0
b!X ?kF
kX
kX ? kF
10-1
Rank r = 1 - Samples m = 10 " n " r
100
b!X ?kF
kX
kX ? kF
100
b!X ? kF
kX
kX ? kF
Rank r = 1 - Samples m = 3 " n " r
5000
10000
15000
Cumulative time (sec)
0
500
1000
Cumulative time (sec)
1500
5000
10-4
10000
15000
Cumulative time (sec)
0
1000
2000
3000
Cumulative time (sec)
Figure 8: Quantum state tomography: Convergence performance of algorithms under comparison
?k
b
F
w.r.t. kX−X
vs. (i) the total number of iterations (top) and (ii) the total execution time (bottom).
kX ? kF
First, second and third column corresponds to Csam = 3, 6 and 10, respectively. For all cases, r = 1 (pure
f0 ) such that X0 = U
e0 U
e > where X0 = Π+ (−A∗ (y)).
state setting) and q = 12. Initial point is U0 = ΠC (U
0
15
q = 6, Csam = 3.
q = 6, Csam = 6.
q = 6, Csam = 10.
Algorithm
?
b
kX−X
kF
kX ? kF
Total time
?
b
kX−X
kF
kX ? kF
Total time
?
b
kX−X
kF
kX ? kF
Total time
RSVP
5.1496e-05
0.7848
1.8550e-05
0.3791
6.6328e-06
0.1203
SparseApproxSDP
4.6323e-03
3.7404
2.2469e-03
4.3775
1.4776e-03
3.8536
AccUniPDGrad
4.0388e-05
0.3634
2.4064e-05
0.3311
1.9032e-05
0.4911
ProjFGD
2.4116e-05
0.0599
1.6052e-05
0.0441
1.1419e-05
0.0446
q = 8, Csam = 3.
q = 8, Csam = 6.
q = 8, Csam = 10.
RSVP
1.5774e-04
5.7347
5.2470e-05
3.8649
2.9583e-05
4.6548
SparseApproxSDP
4.1639e-03
16.1074
2.2011e-03
33.7608
1.7631e-03
85.0633
AccUniPDGrad
3.5122e-05
1.1006
2.4634e-05
1.8428
1.7719e-05
3.9440
ProjFGD
2.4388e-05
0.6918
1.5431e-05
0.8994
1.0561e-05
1.8804
q = 10, Csam = 3.
q = 10, Csam = 6.
q = 10, Csam = 10.
RSVP
4.6056e-04
379.8635
1.8017e-04
331.1315
9.7585e-05
307.9554
SparseApproxSDP
3.6310e-03
658.7082
2.1911e-03
1326.5374
1.7687e-03
2245.2301
AccUniPDGrad
3.0456e-05
33.3585
1.9931e-05
56.9693
4.5022e-05
88.2965
ProjFGD
9.2352e-06
13.9547
5.8515e-06
19.3982
1.0460e-05
49.4528
q = 12, Csam = 3.
q = 12, Csam = 6.
q = 12, Csam = 10.
RSVP
4.7811e-03
14029.1525
1.0843e-03
15028.2836
5.6169e-04
15067.7249
SparseApproxSDP
3.1717e-03
13635.4238
3.6954e-03
15041.6235
5.0197e-03
15051.4497
AccUniPDGrad
8.8050e-05
461.2084
5.2367e-05
904.0507
4.5660e-05
1759.6698
ProjFGD
8.4761e-06
266.8203
4.7399e-06
440.7193
1.1871e-05
1159.2885
Table 1: Quantum state tomography: Summary of comparison results for reconstruction and
efficiency. As a stopping criterion, we used kXi+1 − Xi k2/kXi+1 k2 ≤ 5 · 10−6 , where Xi is the estimate at
f0 ) such that X0 = U
e0 U
e>
the i-th iteration. Time reported is in seconds. Initial point is U0 = ΠC (U
0
∗
where X0 = Π+ (−A (y)).
Rank r = 5 - Samples m = 3 " n " r
10 -2
1000
200
10-4
50
100
SparseApproxSDP
AccUniPDGrad
10 -1
10 -2
400
600
800 1000
Number of iterations
0
10 0
b!X ?kF
kX
kX ? kF
10 -1
150
Number of iterations
10
200
2000
10-1
10-2
3000
SparseApproxSDP
AccUniPDGrad
10 0
10 -1
10-4
Cumulative time (sec)
-4
AccUniPDGrad
ProjFGD
RSVP
ProjFGD
200
400
600
SparseApproxSDP
10 0
b!X ?kF
kX
kX ? kF
10-2
SparseApproxSDP
AccUniPDGrad
b!X ?kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10 0
b!X ? kF
kX
kX ? kF
10-2
Rank r = 20 - Samples m = 3 " n " r
Rank r = 20 - Samples m = 3 " n " r
100
RSVP
ProjFGD
RSVP
ProjFGD
b!X ? kF
kX
kX ? kF
0
b!X ?kF
kX
kX ? kF
10
b!X ? kF
kX
kX ? kF
Rank r = 5 - Samples m = 3 " n " r
100
10 -1
800 1000
Number of iterations
2000 4000 6000 8000 100001200014000
Cumulative time (sec)
10-2
0
200
400
600
0
Cumulative time (sec)
10
20
30
40
Number of iterations
50
0
100
200
300
400
Cumulative time (sec)
Figure 9: Quantum state tomography: Convergence performance of algorithms under comparison
?k
b
F
w.r.t. kX−X
vs. (i) the total number of iterations (left) and (ii) the total execution time (right).
kX ? kF
The two left plots correspond to the case r = 5 and the two right plots to the case r = 20. In all cases
f0 ) such that X0 = U
e0 U
e > where X0 = Π+ (−A∗ (y)).
Csam = 3 and q = 10. Initial point is U0 = ΠC (U
0
16
10-2
b!X ?kF
kX
kX ? kF
SparseApproxSDP
-1
10 -2
200
400
600
SparseApproxSDP
10-4
10
10 -1
10 -2
800 1000
200
Number of iterations
400
600
400
600
800
Number of iterations
Rank r = 1 - Samples m = 3 " n " r
100
0
100
200
100
200
300
10 -2
400
10
0
10 -2
200
400
100
150
200
400
250
10
0
SparseApproxSDP
10
-2
500
1000 1500 2000
Cumulative time (sec)
10
0
600
200
10 -1
Cumulative time (sec)
600
-4
0
Cumulative time (sec)
Cumulative time (sec)
800 1000
AccUniPDGrad
ProjFGD
RSVP
200 400 600 80010001200
10-4
0
50
10-2
SparseApproxSDP
600
600
Number of iterations
Rank r = 1 - Samples m = 10 " n " r
100
0
10 -1
Cumulative time (sec)
400
-4
500
b!X ?kF
kX
kX ? kF
b!X ?kF
kX
kX ? kF
10 -1
400
AccUniPDGrad
ProjFGD
RSVP
10-2
SparseApproxSDP
200
-2
Number of iterations
Number of iterations
Rank r = 1 - Samples m = 6 " n " r
b!X ? kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10 0
10-4
10
800 1000
10
AccUniPDGrad
ProjFGD
RSVP
10-2
SparseApproxSDP
10 -1
b!X ? kF
kX
kX ? kF
200
10 0
-2
Number of iterations
10-4
0
AccUniPDGrad
ProjFGD
RSVP
b!X ?kF
kX
kX ? kF
10
10 0
b!X ? kF
kX
kX ? kF
b!X ? kF
kX
kX ? kF
10 0
Rank r = 1 - Samples m = 10 " n " r
100
AccUniPDGrad
ProjFGD
RSVP
b!X ?kF
kX
kX ? kF
10-2
Rank r = 1 - Samples m = 6 " n " r
100
b!X ?kF
kX
kX ? kF
AccUniPDGrad
ProjFGD
RSVP
b!X ? kF
kX
kX ? kF
Rank r = 1 - Samples m = 3 " n " r
100
200
400
600
Cumulative time (sec)
Figure 10: Quantum state tomography: Convergence performance of algorithms under comparison
?k
b
F
w.r.t. kX−X
vs. (i) the total number of iterations (left) and (ii) the total execution time (right). All
kX ? kF
results correspond to executions starting from a random initialization (but common to all algorithms).
In all cases r = 1 and q = 10.
q = 10, Csam = 3.
Algorithm
?
b
kX−X
kF
kX ? kF
q = 10, Csam = 6.
q = 10, Csam = 10.
Total time
?
b
kX−X
kF
kX ? kF
Total time
?
b
kX−X
kF
kX ? kF
Total time
RSVP
4.5667e-04
545.5525
1.8550e-05
0.3791
1.5774e-04
5.7347
SparseApproxSDP
3.7592e-03
646.3486
2.2469e-03
4.3775
4.1639e-03
16.1074
AccUniPDGrad
3.6465e-05
24.8531
2.4064e-05
0.3311
3.5122e-05
1.1006
ProjFGD
7.0096e-06
19.5502
1.6052e-05
0.0441
2.4388e-05
0.6918
Table 2: Quantum state tomography: Summary of comparison results for reconstruction and
efficiency for random initialization. As a stopping criterion, we used kXi+1 − Xi k2/kXi+1 k2 ≤ 5 · 10−6 ,
where Xi is the estimate at the i-th iteration. Time reported is in seconds.
17
Proofs of local convergence of the ProjFGD
Here, we present the full proof of Theorem 3.1. For clarity, we re-state the problem settings: We
consider problem cases such as
subject to X 0, X ∈ C 0 .
minimize f (X)
X∈Rn×n
(15)
We assume the optimum X ? satisfies rank(X ? ) = r? . For our analysis, we assume we know r? and set
r? ≡ r. We solve (15) in the factored space, by considering the criterion:
minimize f (U U > )
subject to U ∈ C.
U ∈Rn×r
(16)
By faithfulness of C (Definition 2.4), we assume that E ⊆ C. This means that the feasible set C in (16)
contains all matrices U ? that lead to X ? = U ? U ?> in (15). Moreover, we assume both C, C 0 are convex
sets and there exists a “mapping" of C 0 onto C, such that the two constraints are “equivalent": for any
U ∈ C, we are guaranteed that X = U U > ∈ C 0 . We restrict our discussion on norm-based sets for C
such that (6) is satisfied. As a representative example, in our analysis consider the case where, for any
X = U U > , Tr(X) ≤ 1 ⇔ kU k2F ≤ 1.
For our analysis, we will use the following step sizes:
ηb =
1
,
128(LkXt k2 +kQUt Q>
U ∇f (Xt )k2 )
t
η? =
1
128(LkX ? k2 +k∇f (X ? )k2 ) .
11 ?
?
By Lemma A.5 in [8], we know that ηb ≥ 56 η and 10
11 η ≤ η ≤ 10 η . In our proof, we will work with step
size ηb, which is equivalent –up to constants– to the original step size η in the algorithm.
For ease of exposition, we re-define the sequence of updates: Ut is the current estimate in the
et+1 = Ut − ηb∇f (Xt )Ut is the putative solution after the gradient step (observe that
factored space, U
e
et+1 ) is the projection step onto C. Observe that for the
Ut+1 might belong in C), and Ut+1 = ΠC (U
et+1 ) = ξt (U
et+1 ) · U
et+1 , where ξt (·) ∈ (0, 1); in
constraint cases we consider in this paper, Ut+1 = ΠC (U
the case ξt (·) = 1, the algorithm boils down to the Fgd algorithm. For simplicity, we drop the subscript
and the parenthesis of the ξ parameter; these values are apparent from the context.
We assume that ProjFGD is initialized with a “good” starting point X0 = U0 U0> , such that:
(A1)
U0 ∈ C
and
Dist(U0 , U ? ) ≤ ρ0 σr (U ? ) for ρ0 := c ·
µ
L
·
σr (X ? )
σ1 (X ? ) ,
where c ≤
1
200 .
By the assumptions above, X0 = U0 U0> ∈ C 0 . Next, we show that the above lead to a local convergence
result. A practical initialization procedure is given in Section 5.5 and follows from [8]; this also is used
in the experimental section 4.
5.2
Proof of Theorem 3.1
For our analysis, we make use of the following lemma [10, Chapter 3], which
characterizes the effect of projections onto convex sets w.r.t. to inner products,
as well as provides a type-of triangle inequality for such projections; see also
Figure 11 for a simple illustration.
Lemma 5.1. Let U ∈ C ⊆ Rn×r and V ∈ Rn×r where V ∈
/ C. Then,
hΠC (V ) − U, V − ΠC (V )i ≥ 0.
18
C
U
⇧C (V )
V
(17) Figure 11: Illustration of
Lemma 5.1
Proof of Theorem 3.1. We start with the following series of (in)equalities:
Dist (Ut+1 , U ? )2 = min kUt+1 − U ? Rk2F
R∈O
(i)
? 2
≤ kUt+1 − U ? RU
k
t F
(ii)
? 2
et+1 + U
et+1 − U ? RU
= kUt+1 − U
k
t F
2
?
? 2
et+1 kF + kU
et+1 − U RU
= kUt+1 − U
k
t F
D
E
et+1 , U
et+1 − U ? R? ,
+ 2 Ut+1 − U
Ut
? := argmin
?
2
where (i) is due to the fact RU
R∈O kUt − U RkF , (ii) is obtained by adding and subtracting
t
et+1 .
U
et+1 to obtain:
Focusing on the second term of the right hand side, we substitute U
? 2
>
? 2
et+1 − U ? RU
kU
k
=
kU
−
η
b
∇f
U
U
Ut − U ? RU
k
t
t
F
t
t F
t
? 2
2
>
= kUt − U ? RU
k
+
η
b
k∇f
U
U
Ut k2F
t
F
t
t
D
E
?
− 2b
η ∇f Ut Ut> Ut , Ut − U ? RU
t
Then, our initial equation transforms into:
et+1 k2F + Dist (Ut , U ? )2 + ηb2 k∇f Ut U > Ut k2
Dist (Ut+1 , U ? )2 ≤ kUt+1 − U
t
F
D
E
?
− 2b
η ∇f Ut Ut> Ut , Ut − U ? RU
t
D
E
?
?
et+1 , U
et+1 − U R
+ 2 Ut+1 − U
Ut
Focusing further on the last term of the expression above, we obtain:
D
E D
E
et+1 , U
et+1 − U ? R? = Ut+1 − U
et+1 , U
et+1 − Ut+1 + Ut+1 − U ? R?
Ut+1 − U
Ut
Ut
D
E
et+1 , U
et+1 − Ut+1
= Ut+1 − U
E
D
et+1 , Ut+1 − U ? R?
+ Ut+1 − U
Ut
et+1 ≡ Ut+1 for all t, i.e., the iterates are always within C before
Observe that, in the special case where U
the projection step, the above equation equals to zero and the recursion is identical to that of [8][Proof
et+1 6≡ Ut+1 for some t—thus U
et+1 6∈ C.
of Theorem 4.2]. Here, we are more interested in the case where U
>
? ∈ C and X ? = U ? R?
? ?
By faithfulness (Definition 2.1), observe that U ? RU
= U ? U ?> .
Ut U RUt
t
et+1 ): Then, according to Lemma 5.1 and focusing on eq. (17), for U := U ? R?
Moreover, Ut+1 = ΠC (U
Ut
et+1 , the last term in the above equation satisfies:
and V := U
D
E
?
et+1 , Ut+1 − U ? RU
Ut+1 − U
≤ 0,
t
and, thus, the expression above becomes:
D
E
et+1 , U
et+1 − U ? R? ≤ −kUt+1 − U
et+1 k2 .
Ut+1 − U
F
Ut
19
Therefore, going back to the original recursive expression, we obtain:
et+1 k2 + Dist (Ut , U ? )2 + ηb2 k∇f Ut U > Ut k2
Dist (Ut+1 , U ? )2 ≤ −kUt+1 − U
F
t
F
D
E
?
− 2b
η ∇f Ut Ut> Ut , Ut − U ? RU
t
For the last term, we use the descent lemma 3.3 in the main text; the proof is provided in Section
5.3. Thus, we can conclude that:
ηµ
?
Dist (Ut+1 , U ? )2 ≤ 1 − 3b
·
σ
(X
)
· Dist(Ut , U ? )2 .
r
10
The expression for α is obtained by observing ηb ≥ 56 η and
C
1
Then, for η ? ≤ LkX ? k +k∇f
(X ? )k and C = /128, we have:
2
10 ?
11 η
≤η≤
11 ?
10 η ,
from Lemma 20 in [8].
2
1−
10 5 ?
3 · 11
· 6η µ
3b
ηµ
· σr (X ? ) ≤ 1 −
· σr (X ? )
10
10
15
= 1 − η ? µ · σr (X ? )
66
15
µ · σr (X ? )
=1−
66 128(LkX ? k2 + k∇f (X ? )k2 )
µ · σr (X ? )
=: α
≤1−
550(LkX ? k2 + k∇f (X ? )k2 )
where α < 1.
Concluding the proof, the condition Dist(Ut+1 , U ? )2 ≤ ρ0 σr (U ? ) is naturally satisfied, since α <
1.
5.3
Proof of Lemma 3.3
First we recall the definition of restricted strong convexity:
Definition 5.2. Let f : Rn×n → R be convex and differentiable. Then, f is (µ, r)-restricted strongly
convex if:
f (Y ) ≥ f (X) + h∇f (X) , Y − Xi +
2.2.
µ
2
kY − Xk2F ,
∀X, Y ∈ Rn×n , rank-r matrices.
(18)
The statements below apply also for standard µ-strong convex functions, as defined in Definition
et+1 = Ut − ηb∇f (Xt )Ut and define ∆ := Ut − U ? R? . Before presenting the proof, we need
Recall U
Ut
the following lemma that bounds one of the error terms arising in the proof of Lemma 3.3. This is a
variation of Lemma 6.3 in [8]. The proof is presented in Section 5.4.
Lemma 5.3. Let f be L-smooth and (µ, r)-restricted strongly convex. Then, under the assumptions of
Theorem 3.1 and assuming step size ηb = 128(LkXk +k∇f1 (X )Q Q> k ) , the following bound holds true:
2
D
t
E
∇f (Xt ), ∆∆> ≥ − η5b k∇f (Xt )Ut k2F −
Now we are ready to present the proof of Lemma 3.3.
20
Ut
Ut 2
µσr (X ? )
10
· Dist(Ut , U ? )2 .
(19)
Proof of Lemma 3.3. First we rewrite the inner product as shown below.
D
E
?
? ?
>
∇f (Xt )Ut , Ut − U ? RU
=
∇f
(X
),
X
−
U
R
U
t
t
Ut t
t
D
E
?
>
U
= 12 h∇f (Xt ), Xt − X ? i + ∇f (Xt ), 12 (Xt + X ? ) − U ? RU
t t
D
E
= 12 h∇f (Xt ), Xt − X ? i + 12 ∇f (Xt ), ∆∆> ,
(20)
which follows by adding and subtracting 21 X ? .
Let us focus on bounding the first term on the right hand side of (20). Consider points Xt = Ut Ut>
> ; by assumption, both X and X
and Xt+1 = Ut+1 Ut+1
t
t+1 are feasible points in (16). By smoothness of
f , we get:
f (Xt ) ≥ f (Xt+1 ) − h∇f (Xt ), Xt+1 − Xt i −
(i)
≥ f (X ? ) − h∇f (Xt ), Xt+1 − Xt i −
L
2
L
2
kXt+1 − Xt k2F
kXt+1 − Xt k2F ,
(21)
where (i) follows from optimality of X ? and since Xt+1 is a feasible point (Xt+1 0, ΠC 0 (Xt+1 ) = Xt+1 )
for problem (15).
Moreover, by the (µ, r)-restricted strong convexity of f , we get,
f (X ? ) ≥ f (Xt ) + h∇f (Xt ), X ? − Xt i +
µ
2
kX ? − Xt k2F .
(22)
Combining equations (21), and (22), we obtain:
h∇f (Xt ), Xt − X ? i ≥ h∇f (Xt ), Xt − Xt+1 i −
L
2
kXt+1 − Xt k2F +
µ
2
kX ? − Xt k2F
(23)
By the nature of the projection ΠC (·) step, it is easy to verify that
>
>
,
Xt+1 = ξ 2 · Xt − ηb∇f (Xt )Xt Λt − ηbΛ>
X
∇f
(X
)
t
t
t
n×n and Q Q> denoting the projection onto the column space
where Λt = I − η2b QUt Q>
Ut Ut
Ut ∇f (Xt ) ∈ R
of Ut . Notice that, for step size ηb, we have
Λt 0,
σ1 (Λt ) ≤ 1 + 1/256 and σn (Λt ) ≥ 1 − 1/256.
Using the above Xt+1 characterization in (23), we obtain:
h∇f (Xt ), Xt − X ? i −
µ
2
(i)
kX ? − Xt k2F +
L
2
kXt − Xt+1 k2F
≥ ∇f (Xt ), 1 − ξ 2 Xt + 2b
η · ξ 2 · h∇f (Xt ), ∇f (Xt )Xt Λt i
(ii)
≥ 1 − ξ 2 · h∇f (Xt )Ut , Ut i + 2b
η · ξ 2 · Tr(∇f (Xt )∇f (Xt )Xt ) · σn (Λt )
η ·ξ 2
2
(24)
≥ 1 − ξ 2 · h∇f (Xt )Ut , Ut i + 255·b
128 k∇f (Xt )Ut kF ,
where: (i) follows from symmetry of ∇f (Xt ) and Xt and, (ii) follows from the sequence equalities an
inequalites:
Tr(∇f (Xt )∇f (Xt )Xt Λt ) = Tr(∇f (Xt )∇f (Xt )Ut Ut> ) − η2b Tr(∇f (Xt )∇f (Xt )Ut Ut> ∇f (Xt ))
≥ 1 − η2b kQUt Q>
∇f
(X
)k
k∇f (Xt )Ut k2F
t
2
Ut
≥ (1 − 1/256) k∇f (Xt )Ut k2F .
21
?
Combining the above in the expression we want to lower bound: 2b
η ∇f (Xt ) · Ut , Ut − U ? RU
+
t
2
et+1 k , we obtain:
kUt+1 − U
F
?
et+1 k2
2b
η ∇f (Xt ) · Ut , Ut − U ? RU
+ kUt+1 − U
F
t
D
E
et+1 k2F
= ηb h∇f (Xt ), Xt − X ? i + ηb ∇f (Xt ), ∆∆> + kUt+1 − U
η 2 ·ξ 2
≥ 1 − ξ 2 · ηb h∇f (Xt )Ut , Ut i + 255·b
k∇f (Xt )Ut k2F
128
+
−
+
2
ηbµ
ηbL
?
2 kX − Xt kF − 2 kXt −
ηbµσr (X ? )
ηb2
2
5 k∇f (Xt )Ut kF −
10
2
e
kUt+1 − Ut+1 kF
Xt+1 k2F
· Dist(Ut , U ? )2
(25)
et+1 for some ξ ∈ (0, 1),
et+1 = ξ · U
For the last term in the above expression and given Ut+1 = ΠC U
we further observe:
et+1 k2F = kξ · U
et+1 − U
et+1 k2F
kUt+1 − U
= (1 − ξ)2 · kUt k2F + (1 − ξ)2 ηb2 · k∇f (Xt )Ut k2F
− 2 (1 − ξ)2 · ηb · h∇f (Xt )Ut , Ut i
Combining the above equality with the first term on the right hand side in (25), we obtain:
2
2
1 − ξ 2 · ηb h∇f (Xt )Ut , Ut i + (1 − ξ) · kUt k2F + (1 − ξ) ηb2 · k∇f (Xt )Ut k2F
2
h
− 2 (1 − ξ) · ηb · h∇f (Xt )Ut , Ut i =
i
2
2
2
1 − ξ 2 − 2 (1 − ξ) · ηb h∇f (Xt )Ut , Ut i + (1 − ξ) · kUt k2F + (1 − ξ) ηb2 · k∇f (Xt )Ut k2F =
2
2
(3ξ − 1) (1 − ξ) · ηb h∇f (Xt )Ut , Ut i + (1 − ξ) · kUt k2F + (1 − ξ) ηb2 · k∇f (Xt )Ut k2F =
2
2
(3ξ−1)2
3ξ−1
·
U
+
(1
−
ξ)
·
η
b
∇f
(X
)
·
U
+
(1
−
ξ)
−
kUt k2F .
t
t
t
2
4
F
Focusing on the first term, let Θt := I +
the definition of ηb and the fact that ηb ≤
3ξ−1
2
2(1−ξ)
3ξ−1
· ηb · ∇f (Xt )QUt Q>
Ut ; then, σn (Θt ) ≥ 1 −
1
. Then:
128k∇f (X )Q Q> k
· Ut + (1 − ξ) · ηb∇f (Xt ) · Ut
t
2
F
Ut
=
≥
≥
2(1−ξ)
3ξ−1
·
1
128 ,
by
Ut 2
3ξ−1
2 Θt
(3ξ−1)2
4
(3ξ−1)2
4
·
· Ut
2
F
2
kUt kF ·
· 1−
σn (Θt )2
2(1−ξ)
3ξ−1
Combining the above, we obtain the following bound:
et+1 k2
1 − ξ 2 · ηb h∇f (Xt )Ut , Ut i + kUt+1 − U
F
(3ξ−1)2
2
·
1
−
1−
≥ (1 − ξ) −
4
·
1
128
2(1−ξ)
3ξ−1
·
2
1
128
· kUt k2F
2
· kUt k2F
The above transform (25) as follows:
?
et+1 k2F
2b
η ∇f (Xt ) · Ut , Ut − U ? RU
+ kUt+1 − U
t
?
2
2
1
r (X )
≥ 255·ξ
b2 k∇f (Xt )Ut k2F + ηb2µ kX ? − Xt kF − ηbµσ10
· Dist(Ut , U ? )2
128 − 5 · η
2
2
2
2(1−ξ)
1
+ (1 − ξ)2 − (3ξ−1)
·
1
−
1
−
·
· kUt k2F − ηb2L kXt − Xt+1 kF
4
3ξ−1
128
22
(26)
Let us focus on the term
ηbL
2
kXt − Xt+1 k2F =
=
(i)
kXt − Xt+1 k2F ; this can be bounded as follows:
ηbL
2
ηbL
ηbL
>
>
2
>
>
2 kUt Ut − Ut+1 Ut+1 kF = 2 kUt Ut − Ut Ut+1
>
ηbL
>
2
2 kUt (Ut − Ut+1 ) + (Ut − Ut+1 ) Ut+1 kF
>
>
+ Ut Ut+1
− Ut+1 Ut+1
k2F
>
≤ ηbL · kUt (Ut − Ut+1 )> k2F + k (Ut − Ut+1 ) Ut+1
k2F
(ii)
≤ ηbL kUt+1 k22 + kUt k22 · kUt+1 − Ut k2F .
where (i) is due to the identity kA + Bk2F ≤ 2kAk2F + 2kBk2F and (ii) is due to the Cauchy-Schwarz
inequality. By definition of Ut+1 , we observe that:
(i)
(ii)
2
kUt+1 k22 = kξ · (Ut − ηb∇f (Xt )Ut ) k22 ≤ ξ 2 · kUt k22 · kI − ηb∇f (Xt )QUt Q>
Ut k2 ≤ 1 +
where (i) is due to Cauchy-Schwarz and (ii) is obtained by substituting ηb ≤
1 2
128
1
128k∇f (Xt )QUt Q>
U k2
ηbL
2
≤
=
=
≤
1
1+ 128
2
+1
· kUt+1 − Ut k2F
128
1 2
1+ 128 +1
et+1 − Ut k2F
· kξ · U
128
2
1
1+ 128 +1
· k(ξ − 1)Ut − ξ · ηb∇f (Xt ) · Ut k2F
128
1 2
1 2
1+ 128 +1
1+ 128 +1
(1 − ξ)2 ·
· kUt k2F +
· ξ2
64
64
where in the last inequality we substitute ηb; observe that ηb ≤
(26), we obtain:
1
128LkXt k2 .
· ηb2 · k∇f (Xt ) · Ut k2F
Combining this result with
?
et+1 k2F
2b
η ∇f (Xt ) · Ut , Ut − U ? RU
+ kUt+1 − U
t
!
2
≥
255·ξ 2
128
−
1
5
−
1
1+ 128
+1
64
+
(1 − ξ)2 ·
1−
(i)
≥ ηb2 k∇f (Xt )Ut k2F +
+
(ii)
and
t
kXt − Xt+1 k2F can be further bounded as follows:
2
ηbL
1 2
≤
η
b
L
·
1
+
+
1
· kUt k22 · kUt+1 − Ut k2F
kX
−
X
k
t
t+1
F
2
128
1 2
= ηbL · 1 + 128 + 1 · kXt k2 · kUt+1 − Ut k2F
since ξ ∈ (0, 1). Thus,
· kUt k22 .
2
(1 − ξ) ·
η
bµ
2
1−
≥ ηb2 k∇f (Xt )Ut k2F +
η
bµ
2
· ξ2
1
1+ 128
2
· ηb2 k∇f (Xt )Ut k2F +
+1
!
−
64
2
kX ? − Xt kF −
!
2
1
1+ 128
+1
2
kX ? − Xt kF −
(3ξ−1)
4
2
kX ? − Xt kF −
· 1− 1−
2(1−ξ)
3ξ−1
·
1
128
η
bµσr (X ? )
10
!
2
· Dist(Ut , U ? )2
· kUt k2F
η
bµσr (X ? )
10
−
64
2
η
bµ
2
(3ξ−1)2
4
· Dist(Ut , U ? )2
· 1 − 1 − 2(1−ξ)
3ξ−1 ·
η
bµσr (X ? )
10
1
128
!
2
· Dist(Ut , U ? )2
1
1+ 128
· kUt k2F
(27)
2
+1
where (i) is due to the assumption ξ & 0.78 and thus
− −
· ξ 2 ≥ 1; see also Figure
64
12 (left panel), and (ii) is due to the non-negativity of the constant in front of kUt k2F ; see also Figure
12 (right panel).
Finally, we bound ηb2µ kX ? − Xt k2F using the following Lemma by [44]:
255·ξ 2
128
23
1
5
1 2
) )
128
1
1.2
1
0.9
0.8
(39!1)2
4
9 : 0:78 %
0.8
0.7
0.6
0.5
0
-0.2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
9
1 2
) +1
(1+ 128
)
64
0.2
0.4
(1 ! 9)2 " (1 !
1
5
0.4
255"9 2
128
0.6
!
!
!
1 2
) +1
(1+ 128
64
1.4
2(1!9)
39!1
" 92
1.6
" (1 ! (1 !
"
1.8
0.2
0.3
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
9
Figure 12: Behavior of constants, depending on ξ, in expression (27).
Lemma 5.4. For any U, V ∈ Rn×r , we have:
kU U > − V V > k2 ≥ 2 ·
√
2 − 1 · σr (U )2 · Dist(U, V )2 .
Thus,
ηbµ
2
kX ? − Xt k2F ≥ ηbµ ·
√
2 − 1 · σr (X ? ) · Dist(Ut , U ? )2 ,
and can thus conclude:
?
et+1 k2F
2b
η ∇f (Xt ) · Ut , Ut − U ? RU
+ kUt+1 − U
t
√
?
r (X )
2 − 1 · σr (X ? ) · Dist(Ut , U ? )2 − ηbµσ10
≥ ηb2 k∇f (Xt )Ut k2F + ηbµ ·
· Dist(Ut , U ? )2
√
1
= ηb2 k∇f (Xt )Ut k2F +
2 − 1 − 10
· ηbµ · σr (X ? ) · Dist(Ut , U ? )2
= ηb2 k∇f (Xt )Ut k2F +
3b
ηµ
10
· σr (X ? ) · Dist(Ut , U ? )2
This completes the proof.
5.4
Proof of Lemma 5.3
Proof. We can lower bound ∇f (Xt ), ∆∆> as follows:
D
E (i) D
E
>
∇f (Xt ), ∆∆> = Q∆ Q>
∆ ∇f (Xt ), ∆∆
>
>
≥ − Tr Q∆ Q∆ ∇f (Xt )∆∆
(ii)
>
≥ −kQ∆ Q>
∆ ∇f (Xt )k2 Tr(∆∆ )
(iii)
>
? 2
? Q ? ∇f (Xt )k2 Dist(Ut , U ) .
∇f
(X
)k
+
kQ
≥ − kQUt Q>
t
2
U
Ut
U
(28)
Note that (i) follows from the fact ∆ = Q∆ Q>
∆ ∆ and (ii) follows from | Tr(AB)| ≤ kAk2 Tr(B), for
PSD matrix B (Von Neumann’s trace inequality [36]). For the transformation in (iii), we use that fact
that the column space of ∆, Span(∆), is a subset of Span(Ut ∪ U ? ), as ∆ is a linear combination of Ut
? .
and U ? RU
t
24
For the second term in the parenthesis above, we first derive the following inequalities; their use is
apparent later on:
(i)
k∇f (Xt )U ? k2 ≤ k∇f (Xt )Ut k2 + k∇f (Xt )∆k2
(ii)
≤ k∇f (Xt )Ut k2 + k∇f (Xt )Q∆ Q>
∆ k2 k∆k2
(iii)
>
?
≤ k∇f (Xt )Ut k2 + k∇f (Xt )QUt Q>
k
+
k∇f
(X
)Q
Q
k
?
t
U
Ut 2
U 2 k∆k2
(iv)
>
?
1
?
≤ k∇f (Xt )Ut k2 + k∇f (Xt )QUt Q>
k
+
k∇f
(X
)Q
Q
k
?
t
U
Ut 2
U 2 200 σr (U )
(v)
1
· 1 k∇f (Xt )Ut k2 +
1
(1− 200
) 200
200
1
≤ 199
k∇f (Xt )Ut k2 + 200
k∇f (Xt )U ? k2 .
≤ k∇f (Xt )Ut k2 +
?
1
200 k∇f (Xt )U k2
? = U − ∆, (ii) is due to generalized Cauchy-Schwarz
where (i) is due to triangle inequality on U ? RU
t
t
inequality; we denote as Q∆ Q∆ the projection matrix on the column span of ∆ matrix, (iii) is due to
triangle inequality and the fact that the column span of ∆ can be decomposed into the column span of
Ut and U ? , by construction of ∆, (iv) is due to the assumption Dist(Ut , U ? ) ≤ ρ0 · σr (U ? ) and
k∆k2 ≤ Dist(Ut , U ? ) ≤
1 σr (X ? )
200 σ1 (X ? )
· σr (U ? ) ≤
1
200
· σr (U ? ).
Finally, (v) is due to the facts:
?
>
?
k∇f (Xt )U ? k2 = k∇f (Xt )QU ? Q>
U ? U k2 ≥ k∇f (Xt )QU ? QU ? k2 · σr (U ),
and
>
k∇f (Xt )Ut k2 = k∇f (Xt )QUt Q>
Ut U k2 ≥ k∇f (Xt )QUt QUt k2 · σr (Ut )
?
1
≥ k∇f (Xt )QU Q>
U k2 · 1 − 200 · σr (U ),
by the proof of (a variant of) Lemma A.3 in [8]. Thus, for the term k∇f (Xt )QU ? Q>
U ? k2 , we have
k∇f (Xt )QU ? Q>
U ? k2 ≤
≤
≤
?
1
σr (U ? ) k∇f (Xt )U k2
1
201
σr (U ? ) 199 k∇f (Xt )Ut k2
201σ1 (U ? ) 201
>
200σr (U ? ) 199 k∇f (Xt )QUt QUt k2 .
Using (29) in (28), we obtain:
D
E
∇f (Xt ), ∆∆> ≥ − kQUt Q>
Ut ∇f (Xt )k2 +
?
201σ1 (U ? ) 201
>
200σr (U ? ) 199 kQUt QUt ∇f (Xt )k2
(29)
Dist(Ut , U ? )2
? 2
≥ − 21·τ10(U ) kQUt Q>
Ut ∇f (Xt )k2 Dist(Ut , U )
We remind that the step size we use here is: ηb =
21·τ (U ? )
10
1
.
128(LkXt k2 +kQUt Q>
U ∇f (Xt )k2 )
Then, we have:
t
? 2
· kQUt Q>
Ut ∇f (Xt )k2 · Dist(Ut , U )
≤
21·τ (U ? )
10
? 2
· ηb · 128LkXt k2 kQUt Q>
Ut ∇f (Xt )k2 · Dist(Ut , U )
+
21·τ (U ? )
10
2
? 2
· ηb · 128 · kQUt Q>
Ut ∇f (Xt )k2 · Dist(Ut , U )
25
(30)
To bound the first term on the right hand side, we observe that kQUt Q>
Ut ∇f (Xt )k2 ≤
kQUt Q>
Ut ∇f (Xt )k2 ≥
21·τ (U ? )
10
µσr (Xt )
.
21·τ (U ? )
·10
10
µσr (Xt )
21·τ (U ? )
·10
10
or
This results further into:
? 2
· ηb · 128LkXt k2 kQUt Q>
Ut ∇f (Xt )k2 · Dist(Ut , U )
21·τ (U ? )
·128·b
η ·LkXt k2 ·µσr (Xt )
10
· Dist(Ut , U ? )2 ,
≤ max
21·τ (U ? )
·10
10
ηb
≤
?
21·τ (U )
10
2
· 128 ·
2
10κτ (Xt )kQUt Q>
Ut ∇f (Xt )k2
? 2
· Dist(Ut , U )
128·b
η ·LkXt k2 ·µσr (Xt )
10
· Dist(Ut , U ? )2
2
?
2
? 2
+ ηb 21·τ10(U ) · 128 · 10κτ (Xt )kQUt Q>
Ut ∇f (Xt )k2 · Dist(Ut , U ) ,
where κ :=
L
µ
and τ (X) :=
σ1 (X)
σr (X)
21·τ (U ? )
· kQUt Q>
Ut ∇f (Xt )k2
10
(i)
(Xt )
· Dist(Ut , U ? )2
≤ µσr10
for a rank-r matrix X. Combining the above with (30):
· Dist(Ut , U ? )2
+ 10κτ (Xt ) ·
(ii)
≤
(iii)
µσr (Xt )
10
· Dist(Ut , U ? )2
+ 11κτ (X ? ) ·
21·τ (U ? )
10
+1 ·
21·τ (U ? )
10
2
? 2
· 128 · ηbkQUt Q>
Ut ∇f (Xt )k2 · Dist(Ut , U )
21·τ (U ? )
10
+1 ·
21·τ (U ? )
10
2
0 2
?
· 128 · ηbkQUt Q>
Ut ∇f (Xt )k2 · (ρ ) σr (X )
12·212
102
≤
µσr (Xt )
10
· Dist(Ut , U ? )2 +
≤
µσr (Xt )
10
· Dist(Ut , U ? )2 + η5b k∇f (Xt )Ut k22
(iii)
· κ · τ (X ? )2 · 128 · ηbk∇f (Xt )Ut k22 ·
11·(ρ0 )2
10
1
where (i) follows from ηb ≤ 128LkX
, (ii) is due to Lemma A.3 in [8] and using the bound Dist(Ut , U ? ) ≤
t k2
?
0
ρ σr (U ) by the hypothesis of the lemma, (iii) is due to σr (X ? ) ≤ 1.1σr (Xt ) by Lemma A.3 in [8], due to
?
?
2
>
2
? 21·τ (U ) +1) ≤ 12κτ (X ? )· 21·τ (U ) ,
the facts σr (Xt )kQUt Q>
Ut ∇f (Xt )k2 ≤ kUt ∇f (Xt )kF and (11κτ (X )·
10
10
1
1
and τ (U ? )2 = τ (X ? ). Finally, (iv) follows from substituting ρ0 := c · κ1 · τ (X
? ) for c = 200 and using
1
Lemma A.3 in [8] (due to the factor 200
, all constants above lead to bounding the term with the
1
constant 5 ).
Thus, we can conclude:
∇f (Xt ), ∆∆> ≥ −
η
b
2
5 k∇f (Xt )Ut kF
+
µσr (X ? )
10
· Dist(Ut , U ? )2 .
This completes the proof.
5.5
Proof of Corollary 3.2
We have
et+1 kF ≤ kUt kF + ηb · k∇f (Xt )Ut kF
kU
≤ kUt kF + ηb · k∇f (Xt )QUt Q>
Ut k2 · kUt kF
= (1 + ηb · k∇f (Xt )QUt Q>
Ut k 2 ) · λ
≤ (1 +
1
128 )
·λ
where the first inequality follows from the triangle inequality, the second holds by the property kABkF ≤
kAk2 · kBkF , and the third follows because the step size is bounded above by ηb ≤ 128k∇f (X 1)Q Q> k .
t
et+1 ) =
Hence, we get ξ(U
λ
et+1 kF
kU
≥
128
129 .
26
Ut
Ut 2
Initialization
In this section, we present a specific initialization strategy for the ProjFGD. For completeness, we repeat
the definition of the optimization problem at hand, both in the original space:
minimize f (X)
X∈Rn×n
subject to X ∈ C 0 .
(31)
and the factored space:
minimize f (U U > )
subject to U ∈ C.
U ∈Rn×r
(32)
For our initialization, we restrict our attention to the full rank (r = n) case. Observe that, in this case,
C 0 is a convex set and includes the full-dimensional PSD cone, as well as other norm constraints, as
described in the main text. Let us denote ΠC 0 (·) the corresponding projection step, where all constraints
are satisfied simultaneously. Then, the initialization we propose follows similar motions with that in [8]:
We consider the projection of the weighted negative gradient at 0, i.e., − L1 · ∇f (0), onto C 0 .10 I.e.,
X0 = U0 U0> = ΠC 0
−1
L
(33)
· ∇f (0) .
Assuming a first-oracle model, where we access f only though function evaluations and gradient
calculations, (33) provides a cheap way to find an initial point with some approximation guarantees as
follows11 :
Lemma 5.5. Let U0 ∈ Rn×n be such that X0 = U0 U0> = ΠC 0 −1
L · ∇f (0) . Consider the problem
in (32) where f is assumed to be L-smooth and µ-strongly convex, with optimum point X ? such that
rank(X ? ) = n. We apply ProjFGD algorithm with U0 as the initial point. Then, in this generic case, U0
satisfies:
Dist(U0 , U ? ) ≤ ρ0 · σr (U ? ),
where ρ0 =
q
1−µ/L
√
2( 2−1)
· τ 2 (U ? ) ·
p
srank(X ? ) and srank(X) =
kXkF
kXk2
.
Proof. To show this, we start with:
kX0 − X ? k2F = kX ? k2F + kX0 k2F − 2 hX0 , X ? i .
Recall that X0 = U0 U0> = ΠC 0
by Lemma 5.1, we get
−1
L
(34)
· ∇f (0) by assumption, where ΠC 0 (·) is a convex projection. Then,
X0 − X ? , − L1 · ∇f (0) − X ? ⇒
0
−1
L ∇f (0), X
− X ∗ ≥ X 0, X 0 − X ? .
(35)
Observe that 0 ∈ Rn×n is a feasible point, since it is PSD and satisfy any common symmetric norm
constraints, as the ones considered in this paper. Hence, using strong convexity of f around 0, we get,
f (X ? ) −
µ
kX ? k2F ≥ f (0) + h∇f (0), X ? i
2
(i)
= f (0) + h∇f (0), X0 i + h∇f (0), X ? − X0 i
(ii)
≥ f (0) + h∇f (0), X0 i + hL · X0 , X0 − X ? i .
10
(36)
As in [8], one can approximate easily L, if it is unknown.
As we show in the experiments section, a random initialization performs well in practice, without requiring the
additional calculations involved in (33). However, a random initialization provides no guarantees whatsoever.
11
27
where (i) is by adding and subtracting h∇f (0), X0 i, and (ii) is due to (35). Further, using the
smoothness of f around 0, we get:
f (X0 ) ≤ f (0) + h∇f (0), X0 i + L2 kX0 k2F
(i)
≤ f (X ? ) − µ2 kX ? k2F + hL · X0 , X ? i − L2 kX0 k2F
≤ f (X0 ) − µ2 kX ? k2F + hL · X0 , X ? i − L2 kX0 k2F .
where (i) follows from (36) by upper bounding the quantity f (0) + h∇f (0), X0 i, (ii) follows from the
assumption that f (X ? ) ≤ f (X0 ). Hence, rearranging the above terms, we get:
hX0 , X ? i ≥ 12 kX0 k2F +
Combining the above inequality with (34), we obtain,
q
?
kX0 − X kF ≤ 1 −
µ
? 2
2L kX kF .
µ
L
· kX ? kF .
Given, U0 such that X0 = U0 Uo> and U ? such that X ? = U ? U ?> , we use Lemma 5.4 from [44] to obtain:
q √
>
? ?>
kU0 U0 − U U kF ≥ 2( 2 − 1) · σr (U ? ) · Dist(U0 , U ? ).
Thus:
kX0 −X ? kF
√
2( 2−1)·σr (U ? )
ρ0 · σr (U ? )
Dist(U0 , U ? ) ≤ √
≤
where ρ0 =
q
1−µ/L
√
2( 2−1)
· τ 2 (U ? ) ·
p
· kX ? kF
srank(X ? ).
Such initialization, while being simple, introduces further restrictions on the condition number τ (X ? ),
and the condition number of function f . Finding such simple initializations with weaker restrictions
remains an open problem; however, as shown in [8, 44, 18], one can devise specific deterministic
initialization for a given application.
As a final comment, we state the following: In practice, the projection ΠC 0 (·) step might not be easy
to compute, due to the joint involvement of convex sets. A practical solution would be to sequentially
project − L1 · ∇f (0) onto the individual constraint sets. Let Π+ (·) denote the projection onto the PSD
cone. Then, we can consider the approximate point:
e0 = U
e0 U
e0> = Π+ X
e0 ;
X
e0 , we can perform an additional step:
Given U
e0 ,
U0 = ΠC U
to guarantee that U0 ∈ C.
28
| 8cs.DS
|
Unsupervised Domain Adaptation with Similarity Learning
arXiv:1711.08995v1 [cs.CV] 24 Nov 2017
Pedro O. Pinheiro
Element AI
Montreal, QC, Canada
[email protected]
Abstract
The objective of unsupervised domain adaptation is to
µ3
leverage
µ1 features from a labeled source domain and learn
a classifier for an unlabeled target domain, with a similar but different data distribution. Most deep learning approaches to domain adaptation consist of two steps: (i)
learn features that preserve a low risk on labeled samples
(source domain)
and (ii) make the features from both doµ2
mains to be as indistinguishable as possible, so that a classifier trained on the source can also be applied on the target domain. In general, the classifiersµin
step (i) consist of
4
fully-connected layers applied directly on the indistinguishable features learned in (ii). In this paper, we propose a
different way to do the classification, using similarity learning. The proposed method learns a pairwise similarity function in which classification can be performed by computing
distances between prototype representations of each category. The domain-invariant features and the categorical
prototype representations are learned jointly and in an endto-end fashion. At inference time, images from the target
domain are compared to the prototypes and the label associated with the one that best matches the image is outputed.
The approach is simple, scalable and effective. We show
that our model achieves state-of-the-art performance in different large-scale unsupervised domain adaptation scenarios.
µ3
µ1
µ2
f (xti )
µ4
Figure 1: Illustrative example of our similarity-based classifier. A query target-domain image representation is compared to a set of prototypes, one per category, computed
from source-domain images. The label of the most similar
prototype is assigned to the test image.
most commonly used approach to transfer learned features
is to further modify them through a process called “finetuning”. In this case, the features are adapted by training
the network with labeled samples from the new data distribution. In many cases, however, acquiring labeled data can
be expensive.
Unsupervised Domain Adaption deals with the domain
shift problem. We are interested in learning representations
that are invariant to domains with different data distributions. In this scenario, the machine has access to a labeled
dataset (called source domain) and an unlabeled one (with a
similar but different data distribution, called target domain),
and the objective is to correctly infer the labels on the latter.
Most current approaches are based on deep learning methods and consist of two steps: (i) learn features that preserve
a low risk on labeled samples (source domain) and (ii) make
the features from both domains to be as indistinguishable as
1. Introduction
Convolutional Neural Networks (ConvNets) [30] based
methods achieve excellent results in large-scale supervised
learning problems, where a lot of labeled data exists [28,
25]. Moreover, these features are quite general and can be
used in a variety of vision problems, such as image captioning [52], object detection [31] and segmentation [24].
However, direct transfer of features from different domains do not work very well in practice, as the data distributions of domains might change. In computer vision, this
problem is sometimes referred to as domain shift [47]. The
1
possible, so that a classifier trained on the source can also
be applied on the target domain.
Unsupervised domain adaptation can also be used to
train models with synthetic data (e.g. 3D renders or game
engines). Assuming we are able to handle the domain shift
problem, we can make use of virtually unlimited number
of labeled (synthetic) images and adapt the model to natural images. Indeed, in certain areas of research, such as
robotics vision or reinforcement learning, acquiring training sample is very expensive. Training in a synthetic domain and transferring the learned features to real-world environments can be a solution to alleviate the high cost of
acquiring and labeling training data.
Theoretical studies in domain adaptation [6, 5] suggest
that a good cross-domain representation is one in which the
algorithms are not able to identify from which domain the
original input comes from. Most current approaches to domain adaptation achieve this goal by mapping cross-domain
features into a common space using deep learning methods.
This is generally achieved by minimizing some measure of
domain variance (such as the Maximum Mean Discrepancy
(MMD) [22, 50, 34, 36]), or by matching moments of the
two distributions [46].
Another way to deal with the domain variation problem
is to make use of adversarial training [1, 15, 48, 49, 8].
In this scenario, the domain adaptation problem is cast as
a minimax game between a domain classifier and feature
learning (see Figure 2a). A neural network learns features
that are, at the same time, as discriminative as possible
(in the source domain) and as indistinguishable as possible
(among the domains).
In general, the classifier used in the source domain is a
simple fully-connected network followed by a softmax over
the categories (as in standard supervised learning). While
this is, in principle, a good idea (given the representations
are trained to be indistinguishable), it leaves the shared representation vulnerable to contamination by noise that is correlated with the underlying shared distribution [44, 8].
In this paper, we propose a different way to do classification, while keeping the adversarial domain-confusion component of [15]. Instead, we propose a similarity-based classifier in which each image (from either the source or target
domain) is compared to a set of prototypes (or centroides).
The label associated to the prototype that best matches the
query image is given to it. See Figure 1. Prototypes are
vector representations that are representative of each category that appears in the dataset. They are learned at the
same time as the image embeddings, and the whole system
is backpropagable.
More precisely, we are interested in learning embeddings
for the inputs (the source/target domain images), the embedding of each prototype (one per category category), and
a pairwise similarity function that compares the inputs and
the prototypes (see Figure 2b). At the same time, adversarial training is performed on the inputs to force domainconfusion. All these components are learned jointly and in
an end-to-end fashion.
At inference time, each embedding prototype is computed a priori (by averaging the embedding over multiple
source images). A test image (from target domain) is then
compared to each of the prototypes and the label of the nearest prototype is assigned to it.
We show empirically that the proposed similarity-based
classification approach is more robust to the domain shift
between two datasets. Our method is able to over perform the commonly used classifier and achieve new state
of the art in multiple domain adaptation scenarios. We
show results in two important large-scale domain adaptation datasets: Office-31 [42], which contains images of office objects in three different domains and VisDA [39], a
large-scale dataset focused on simulation-to-reality shift.
The paper is organized as follows: Section 2 presents related work, Section 3 describes our proposed method and
architecture choices, and Section 4 describes our experiments in different datasets. We conclude in Section 5.
All source code for reproducing the methods in this paper will be released.
2. Related Works
Similarity Learning: Many previous work focus on
learning a similarity measure that is also a metric, like the
case of the positive semi definite matrix that defines the Mahalanobis distance. See [4] for a survey.
Multiple authors (e.g. [54, 16, 17, 53, 37]) used a learned
similarity measure together with nearest neighbors classifier. [43, 11, 41, 10] propose methods in which features are
learned along with the metric learning. Chechik et al. [10]
uses similarity measure in large-scale image search. Similar to us, they define similarity as a bilinear operator but
removes the positivity and symmetry constraints to favor
scalability.
Mensink et al. [37] and Snell et al. [2], similar to this
work, propose to learn metric space in which classification
is performed by computing distances to prototype representation of each category (mean of its examples). However,
the problem setting and the functions parametrization of
these works are very different form ours.
Unsupervised Domain Adaptation: Non-deep approaches for unsupervised domain adaptation usually consist of matching the feature distribution between the source
and target domain [23, 12]. These methods can roughly be
divided into two categories: (i) sample re-weighting [27, 26,
18] and (ii) feature space transformation [38, 21, 3, 45, 29].
Recently, an increasing interest in ConvNets [30] approaches has emerged, due to its ability to learn powerful features. These methods, in general, are trained
to minimize a classification loss and maximize domainconfusion. The classification loss is in general computed
through a fully-connected network trained on labeled data.
The domain-confusion is achieved in two different ways:
(i) discrepancy loss, that reduces the shift between the two
domains [50, 34, 36, 46] or (ii) adversarial loss, which encourages a common feature space with respect to a discriminator [48, 14, 15, 1, 49].
In Deep Domain Confusion (DDC) [50], the authors propose a domain-confusion loss based on MMD [22] applied
on the final representation of a network. Long et al. [34]
(DAN) consider a sum of multiple MMDs between several layers. They consider multiple kernels for adapting the
representations, achieving more robust results than previous
work. In Join Adaptation Networks (JAN) [36], the authors
propose an extension to DAN in which they use the joint
distribution discrepancy over deep features (instead of their
sum). DeepCORAL [46] achieves the domain-confusion by
aligning the second-order statistics of the learned feature
representations.
Ganin et al. [14] and Ajakan et al. [1], and subsequently [15], impose domain confusion through an adversarial objective with respect to a domain discriminator with
the reverse gradient algorithm (RevGrad). They treat domain invariance as a binary classification problem, but directly maximizes the loss of the domain classifier by reversing its gradient. This allows the model to learn features that
are discriminative in the source domain and indiscriminate
with respect to the domain shift. ADDA [49] proposes a
general framework for adversarial deep domain adaptation
and uses an inverted label GAN [20] loss to achieve the domain confusion. The DSN [8] model divides the space of
images into two: one that captures information of the domain, and other that captures information shared between
domains.
Certain methods incorporate generative modeling (usually based on GANs [20]) into the feature learning process. Coupled Generative Adversarial Networks [33] consists of a tuple of GANs each corresponding to one of the
domains that learns a joint distribution of multi-domain images. In [7], the authors propose a model that generates images (conditioned on source-domain images) that are similar to images from target domain.
All methods mentioned in this section use the same classifier component: a fully-connected network. In this paper, contrary to them, we use a different classifier based on
similarity learning. In our method, we follow the approach
of [15] to impose domain-confusion (due to its simplicity
and efficiency). We note, however, that our similarity-based
classifier could be applied to other approaches mentioned in
this section.
3. Method
In domain adaptation, we have access to labeled images
s
Xs = {(xsi , yis )}N
i=0 drawn from a source domain distribut
tion ps (x, y) and target images Xt = {(xti , yit )}N
i=0 drawn
from a target distribution pt (x, y). In the unsupervised setting, we do not have any information about the label on the
target domain.
We address the problem of unsupervised domain adaptation using a similarity-based classifier. Our model, which
we call SimNet, is composed of two different components (see Figure 2b): (i) the domain-confusion component, which forces the features of both domains, f (Xs ) and
f (Xt ), to be as indistinguishable as possible and (ii) a classifier based on a set of prototypes, µc (one for each category
c ∈ {1, 2, ..., C}). Both components are trained jointly and
in an end-to-end fashion.
This approach is based on the assumption that it exists an
embedding for each category such that all the points of the
category cluster around it, independent of its domain. Inference is then performed in a test image by simply finding the
most semantically similar prototype.
We next describe the proposed similarity-based classifier, followed by the training and inference procedures and
implementation details.
3.1. Similarity-Based Classifier
Our classifier is composed of C different prototypes, one
per category. Each prototype represents a general embedding for a category, incorporating all its variations. We assume that there exists a representation space in which all
samples of a given category can be clustered around its corresponding prototype.
Each prototype is represented by a m-dimensional vector µc ∈ Rm parametrized by a ConvNet g(·), with trainable
parameters θg . The prototypes are computed by the average
representation of all source samples belonging to the category c:
µc =
1 X
g(xsi ) ,
|Xc | s c
(1)
xi ∈X
where Xc is the set of all images in the source domain labeled with category c. Similarly, the input images (from
either domain) are represented as a n-dimensional vector
fi = f (xi ) ∈ Rn , thanks to a ConvNet f (·) parametrized
by θf .
By leveraging the powerful representations of convolutional neural networks, we propose a simple model that can
predict which of the prototypes (and therefore categories)
best describes a given input. For this purpose, a similarity
metric between images and prototypes is learned, as illustrated in Figure 2b. The similarity between an input image
f (·)
class
fi
g(·)
µ1
g(·)
µ2
...
disc
f (·)
disc
...
f (·)
g(·)
(a) standard adversarial approach
µC
(b) similarity learning approach
Figure 2: Architectures for adversarial unsupervised domain adaptation. (a) Standard approaches usually consist of two
heads: disc, which imposes features do be indiscriminate w.r.t. the domain and class, that uses a fully-connected network to
classify source domain images. (b) Our method keeps the same disc component but uses instead a similarity-based classifier.
It compares the embedding of an image with a set of prototypes. The label that best matches the image is given. All
representations are learned jointly and in an end-to-end fashion.
xi and prototype µc is defined simply as a bilinear operation:
h(xi , µc ) = fiT Sµc ,
(2)
with S ∈ Rn×m being the trainable parameters. S is an
unconstrained bilinear similarity operator, and it is not required to be positive or symmetric.
Note that the two ConvNets, f and g, do not share
the same parameters. This is particularly important in the
domain adaptation scenario, in which the representations
(from input and from prototypes) have different roles in
the classification. On the one hand, the domain features fi
should, at the same time, be domain invariant and match one
of the prototypes µc . On the other hand, the embedding prototypes should be as close as possible to the source domain
images that represents its category. In the case of single
domain classification, it would make sense to use the same
network for f and g to reduce the capacity of the model,
since there is no shift in the domain. In this case, the model
would be similar to Siamese Networks [11].
The model is trained to discriminate the target prototype
µc from all other prototypes µk (with k 6= c), given a labeled image. We interpret the output of the network as class
conditional probabilities by applying a softmax [9] over the
bilinear operator:
eh(xi ,µc )
pθ (c|xi , µ1 , ..., µC ) = P h(x ,µ ) .
i
k
ke
(3)
θ = {θf , θg , S} represents the set of all trainable parame-
ters of the model. Learning is achieved by minimizing the
negative log-likelihood (with respect to θ), over all labeled
samples (xi , yi ) ∈ Xs :
"
#
X
X
Lclass (θ) = −
h(xi , µyi )−log
eh(xi ,µk ) +γR .
(xi ,yi )
k
(4)
R is a regularization term that encourage the prototypes to
encode different aspects of each category. At each training iteration, the prototypes are approximated by choosing
a random subset of examples for each class.
The regularizer is modeled as a soft orthogonality constraint. Let Pµ be a matrix whose rows are the prototypes,
we write the regularization term as:
R = ||PTµ Pµ − I||2F ,
(5)
where || · ||2F is the squared Frobenius norm and I is the
identity matrix.
3.2. Training and Inference
We are interested in training a classifier on the source
domain (labeled samples) and apply it to the target domain
(unlabeled samples). To achieve this goal, the model learns
features that maximize the domain confusion, while preserving a low risk on the source domain.
The domain-invariant component is responsible to minimize the distance between the empirical source and target
feature representation distributions, f (Xs ) and f (Xt ). Assuming this is the case, the classifier trained on the source
feature representation can thus be directly applied to the target representation.
Domain confusion is achieved with a domain discriminator D, parametrized by θd . The discriminator classifies
whether a data point is drawn from the source or the target
domain, and it is optimized following a standard classification loss:
Ldisc (θ, θd ) = −
−
Ns
X
log D(f (xsi )) +
i=0
Nt
X
i=0
(6)
log(1 −
D(f (xti )))
.
We achieve domain confusion by applying the Reverse
Gradient (RevGrad) algorithm [15], which optimizes the
features to maximize the discriminator loss directly. We
point the reader to [15] for more details about this component.
The model is trained to jointly maximize the domain
confusion (between source and target) and infer the correct category on the source (labeled) samples, through the
similarity-based classifier described in the previous section.
Therefore, our final goal is to optimize the following minimax objective:
min max Lclass (θf , θg , S) + λLdisc (θf , θd ) ,
θf ,θg ,S
θd
(7)
where λ is a balance parameter between the two losses.
The objective is optimized using stochastic gradient descent
with momentum.
At inference time, the prototypes are computed a priori,
following Equation 1, and stored in memory. The similarity between a target-domain test image and each prototype
is computed, and the label that best matches the query is
outputed.
3.3. Implementation Details
The parameters of networks f and g are initialized with
a ResNet-50 [25] that was pre-trained to perform classification on the ImageNet dataset [13], and the classification
layer is removed. The discriminator network and the bilinear classifier are initialized randomly from a uniform distribution. We set the balance parameter λ = 0.5 and the regularization coefficient γ = 0.1 (we observed that the model
is robust to this hyperparameter). We use a learning rate of
10−5 , with a weight decay of 10−5 and momentum of 0.99.
Since the similarity matrix and the discriminator are trained
from scratch, we set their learning rate to be 10 times that
of the other layers.
During training, the images (from both domains) are resized such that the shorter dimension is of size 300 pixels
and a patch of 224 × 224 is randomly sampled. Each minibatch has 32 images from each domain. At each training
iteration, the prototypes are approximated by picking one
random sample for each class. We noticed that the training
converges with this design choice and it is more efficient.
The discriminator is a simple fully-connected network.
It contains two layers, each of dimension 1024 and ReLU
non-linearity, followed by the domain classifier. It receives
as input the output of network f and outputs the probability
of which domain the input comes from (this probability is,
again, modeled by a softmax).
We parametrize the bilinear operation with a low-rank
approximation, S = UT V (U, V ∈ Rn×m , m = 512) and
Equation 2 becomes:
h(xi , µc ) = (Ufi )T · (Vµc ) .
(8)
This parametrization brings multiple benefits. First, it
allows us to control the capacity of the model. Second, it
can trivially be implemented in any modern deep learning
framework and benefits from an efficient implementation.
Finally, it also provides fast inference, as the right side of h
is independent of the input image; it can be computed only
once and stored in memory.
At inference time, the shorter dimensions of the test image is resized to 300 pixels, as in the training stage. The
model is applied densely at every location, resulting in an
output with spatial dimensions bigger than 1. The output is
averaged over the spatial dimensions to get one-dimensional
vector. Because the similarity measure is a bilinear operation, the averaging can be seen as an ensemble over different
spatial locations of the test image.
4. Experimental Results
We evaluate the performance of our model, SimNet, in
two important unsupervised domain adaptation datasets and
across different domain shifts. Figure 3 illustrates image
samples from different datasets and domains.
In the experiments, we use all labeled source images and
all unlabeled target images, following the standard evaluation protocol for unsupervised domain adaptation [15, 36].
All hyperparameters were chosen via cross-validation. We
evaluate the performance of all methods with classification
accuracy metric.
4.1. Office-31 Results
Office-31 [42] is a standard dataset used for domain
adaptation. The dataset contains 4652 images of 31 different categories. The categories are everyday office objects, such as backpack, pen and phone. The images on this
dataset belong to three distinct domains: (i) Amazon (A),
which contains images from the Amazon website, (ii) DSLR
(D) and (iii) Webcam (W), which contains photos taken by
a high quality camera and a webcam, respectively (see Figure 3, left). We evaluate the performance of our method in
amazon
Synthetic
dslr
Real
webcam
Office-31
VisDA
Figure 3: We evaluate our model on unsupervised domain adaptation across different domains in two dataset. The first setting
is over all 6 combinations of domains in Office-31 dataset. The second is on the synthetic-to-real domain shift on VisDA
dataset.
Method
Source Only
TCA [38]
GFK [19]
DDC [50]
DAN [34]
RTN [35]
RevGrad [15]
JAN [36]
JAN-A [36]
SimNet
A→W
D→W
W→D
A→D
D→A
W→A
average
60.9 ± 0.0
61.0 ± 0.0
61.5 ± 0.5
62.8 ± 0.2
64.8 ± 0.3
67.4 ± 0.5
70.0 ± 0.4
70.7 ± 0.5
77.6
77.5
78.3
80.4
81.6
82.2
84.3
84.6
68.4 ± 0.2
96.7 ± 0.1
99.3 ± 0.1
68.9 ± 0.2
62.5 ± 0.3
60.7 ± 0.3
76.1
88.6 ± 0.5
98.2 ± 0.2
99.7 ± 0.2
85.3 ± 0.3
73.4 ± 0.8
71.8 ± 0.6
86.2
72.7 ± 0.0
72.8 ± 0.0
75.6 ± 0.2
80.5 ± 0.4
84.5 ± 0.2
82.0 ± 0.4
85.4 ± 0.3
86.0 ± 0.4
96.7 ± 0.0
95.0 ± 0.0
96.0 ± 0.2
97.1 ± 0.2
96.8 ± 0.1
96.9 ± 0.2
97.4 ± 0.2
96.7 ± 0.3
99.6 ± 0.0
98.2 ± 0.0
98.2 ± 0.1
99.6 ± 0.1
99.4 ± 0.1
99.1 ± 0.1
99.8 ± 0.2
99.7 ± 0.1
74.1 ± 0.0
74.5 ± 0.0
76.5 ± 0.3
78.6 ± 0.2
77.5 ± 0.3
79.7 ± 0.4
84.7 ± 0.3
85.1 ± 0.4
61.7 ± 0.0
63.4 ± 0.0
62.2 ± 0.4
63.6 ± 0.3
66.2 ± 0.2
68.2 ± 0.4
68.6 ± 0.3
69.2 ± 0.4
Table 1: Classification accuracy evaluation (%) on Office-31 dataset for unsupervised domain adaptation with different
domain pairs. All models utilize ResNet-50 as base architecture.
all 6 possible transfer tasks, using the same experimental
protocol as [36].
Table 1 compares the performance of our method with
other approaches. The first row shows results without any
adaptation, where the network is fine-tuned on the source
domain and directly applied to the target domain. This
serves as a lower bound. Other results are reported from
Long et al. [36]. For a fair comparison, all models use the
same base architecture, ResNet-50. SimNet beats the previous state of the art in most settings, while being similar in
the remaining ones.
4.2. VisDA Results
VisDA dataset [39] is focused on simulation-to-real domain shift.1 It contains two very distinct domains: (i) Synthetic, which contains synthetic renderings of 3D models
from different angles and with different lightning conditions and (ii) Real, which contains natural images cropped
either from COCO [32] (validation) or Youtube-BB [40]
(test) datasets (see Figure 3, right). It contains over 280K
images across 12 categories in the combined training, validation and test domains.
Comparison to other methods: Table 3 shows the performance of different methods on the VisDA dataset, for
1 This dataset also contains segmentation labels, but on this paper we
focus on classification task only.
aeroplane
bicycle
bus
car
horse
knife
motocycle
person
plant
skateboard
train
truck
average
Source Only
67.7
36.6
48.4
68.2
76.9
5.3
65.8
38.0
72.5
29.1
82.1
3.73
49.5
RevGrad-ours
SimNet-shared
SimNet-no-reg
SimNet
SimNet-152
75.9
85.7
91.2
92.0
94.3
70.5
51.7
78.7
83.4
82.3
65.3
58.5
69.2
68.4
73.5
17.3
53.4
38.6
42.0
47.2
72.8
74.5
82.5
87.1
87.9
38.6
23.9
22.9
14.5
49.2
58.0
69.7
60.6
74.4
75.1
77.2
68.9
73.6
78.7
79.7
72.5
54.5
78.2
85.7
85.3
40.4
50.9
66.0
77.1
68.5
70.4
76.8
76.3
77.5
81.1
44.7
21.2
39.0
38.2
50.3
58.6
57.5
64.7
68.3
72.9
Train on target
99.5
91.9
97.3
96.8
98.3
98.5
94.1
96.2
99.0
98.2
97.9
82.3
95.8
Table 2: Per-class classification accuracy (%) for different variants of SimNet on VisDA-val.
Figure 4: Confusion matrix for source only, SimNet and oracle (trained on target domain) models on the synthetic-to-real
adaptation on VisDA-val dataset. Best viewed in color.
the synthetic to real (validation set) domain shift.2 Once
again, all the experiments are computed using ResNet-50
backbone for fair comparison of the models. The results
reported are the average per-class accuracy. Results from
other models are taken from Long et al. [36]3 .
RevGrad-ours reports our implementation of the
RevGrad method [15] (with a ResNet-50 architecture), in
which SimNet shares the adversarial learning component.
As RevGrad-ours and SimNet were trained with same training setup and discriminator function, we can disentangle the performance improvements achieved by the proposed similarity-based classifier. SimNet achieves almost
10% boost in performance. This result gives us a hint
2 Some non-published results can be found on-line. However, no detail
of the methods nor how many tricks were used are mentioned.
3 These values were presented by the authors on the conference where
the published paper was presented.
that similarity-based classifiers are more robust than typical fully-connected networks when features are subject to
domain invariance.
VisDA is more challenging than Office-31. The objects
in the target domain contains more variability and the domain shift between the two domains is much bigger. Results in Table 3 show that the model scales nicely with
the data. SimNet achieves a bigger improvement in performance, compared to previously published state of the
art [36], when the scale of the dataset increases. These
results are achieved without any inference trick (such as
model ensembling, multi-scale inference, etc.).
SimNet, when trained with VisDA-test as the target domain and without bells and whistles, achieves a similar performance: 68.66% average per-class accuracy on the test
set.
Model variants: Table 2 shows the performance of dif-
Method
Per-class
avg. acc.
Source Only
49.51
DAN [34]
RTN [35]
RevGrad [15]
JAN [36]
JAN-A [36]
53.02
53.56
55.03
61.06
61.62
RevGrad-ours
SimNet
58.62
68.25
terms of appearance and semantics are closer in the t-SNE
space: ‘bicycle’ and ‘motorcycle’ are very close and ‘truck’
is located between ‘car’ and ‘bus’. Interestingly, SimNet
seems to confuse the categories ‘knife’ and ‘skateboard’.
source domain
Table 3: Per-class average classification accuracy (%) on
VisDA-val dataset for unsupervised domain adaptation. All
the models utilize ResNet-50 as base architecture.
(a)
(b)
truck
train
skboard
plant
person
moto
knife
(c)
horse
car
bus
bicycle
aeropln
ferent versions of SimNet in each of the 12 categories of the
VisDA-val. As before, Source Only shows the results when
trained on the source images only (this is a lower bound).
Train on Target shows results when the model is trained
with the labels of the target domain (this can be interpreted
as an upper bound).
We also show that sharing the weight between the two
networks, f and g, hurts the performance (SimNet-shared).
The results of SimNet-no-reg emphasis the importance of
the regularizer described in Equation 5. The penalty term
imposes the prototypes to be orthogonal, which eases the
similarity computation. In SimNet-152, we show the results
using a ResNet-152 (pre-trained on ImageNet) as base architecture, which pushes the performance even further. This
results shows that proposed bilinear classifier is also able to
scale with the quality of the features.
For more insight on the performance of different categories, we plot different confusion matrices in Figure 4.
Feature visualization: Finally, we use t-SNE [51] to visualize feature representations from different domains and
at different adaptation stages. Figure 5(a-b) show features
from source (blue) and target (red) domains before and after
adaptation, respectively. The features become much more
domain invariant after adaptation. We observe a strong correspondence in terms of classification performance (on the
target domain) and the overlap between the domain distributions.
Figure 5(c-d) show features after adaptation from source
and target domains, respectively. Each category is encoded
by a color and the prototypes are shown as black triangles.
As expected, the images on the source domain are separated better than the images from the target domain. We
can observe few interesting facts about these plots. First,
the similarity measure is well learned for certain categories
(‘aeroplane’,‘horse’, ‘plant’). Categories that are similar in
target domain
(d)
Figure 5: (a-b) Image features from source (blue) and target (red) domains, respectively. The adaptation makes the
the distribution of features much closer. (c-d) Image features after adaptation from source and target domain, respectively. Each category is encoded by a color and the prototypes are shown as black triangles. Best viewed in color.
5. Conclusion
In this paper we propose SimNet, a neural network model fur unsupervised domain adaptation using a
similarity-based classifier. We show that similarity learning,
together with feature learning, can outperform by a large
margin the ‘standard’ fully-connected classifier in the problem of domain adaptation. The method is simple and highly
effective. We demonstrate its efficacy by applying it in
different unsupervised adaptation problems, achieving new
state-of-the-art performance in multiple scenarios. Future
works include applying the model in different modalities
of domain adaptation (e.g. semantic segmentation, object
detection) and with different domain discrepancy reduction
algorithms (such as MMD and its variants). It would also
be interest to explore how the similarity-based classifier can
scale to a larger number of categories and training samples.
Acknowledgements. I thank Negar Rostamzadeh,
Thomas Bosquet and Ishmael Belghazi for helpful discussions and encouragement, and Philippe Mathieu and Jean
Raby for help with computational infrastructure.
References
[1] H. Ajakan, P. Germain, H. Larochelle, F. Laviolette,
and M. Marchand. Domain-adversarial neural networks. 2014. 2, 3
[2] J. S. andKevin Swersky andRichard S. Zemel. Prototypical networks for few-shot learning. 2017. 2
[3] M. Baktashmotlagh, M. T. Harandi, B. C. Lovell, and
M. Salzmann. Unsupervised domain adaptation by domain invariant projection. In ICCV, 2013. 2
[4] A. Bellet, A. Habrard, and M. Sebban. A survey on
metric learning for feature vectors and structured data.
arXiv, 2013. 2
[5] S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza,
F. Pereira, and J. W. Vaughan. A theory of learning
from different domains. Machine Learning, 2010. 2
[6] S. Ben-david, J. Blitzer, K. Crammer, and O. Pereira.
Analysis of representations for domain adaptation. In
NIPS, 2007. 2
[7] K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and
D. Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. In CVPR.
3
[8] K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and D. Erhan. Domain separation networks. In
NIPS, 2016. 2, 3
[9] J. Bridle. Probabilistic interpretation of feedforward
classification network outputs, with relationships to
statistical pattern recognition. Neurocomputing: Algorithms, Architectures and Applications, 1990. 4
[10] G. Chechik, V. Sharma, U. Shalit, and S. Bengio. Large scale online learning of image similarity
through ranking. Journal of Machine Learning Research, 2010. 2
[11] S. Chopra, R. Hadsell, and Y. LeCun. Learning a similarity metric discriminatively, with application to face
verification. In CVPR, 2005. 2, 4
[12] G. Csurka. A comprehensive survey on domain adaptation for visual applications. In Domain Adaptation
in Computer Vision Applications. 2017. 2
[13] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and
L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 5
[14] Y. Ganin and V. S. Lempitsky. Unsupervised domain
adaptation by backpropagation. In ICML, 2015. 3
[15] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain,
H. Larochelle, F. Laviolette, M. Marchand, and V. S.
Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 2016.
2, 3, 5, 6, 7, 8
[16] A. Globerson and S. Roweis. Metric learning by collapsing classes. In NIPS, 2005. 2
[17] J. Goldberger, G. E. Hinton, S. T. Roweis, and R. R.
Salakhutdinov. Neighbourhood components analysis.
In NIPS. 2005. 2
[18] B. Gong, K. Grauman, and F. Sha. Connecting the dots
with landmarks: Discriminatively learning domaininvariant features for unsupervised domain adaptation.
In ICML, 2013. 2
[19] B. Gong, Y. Shi, F. Sha, and K. Grauman. Geodesic
flow kernel for unsupervised domain adaptation. In
CVPR, 2012. 6
[20] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In NIPS, 2014. 3
[21] R. Gopalan, R. Li, and R. Chellappa. Domain adaptation for object recognition: An unsupervised approach. In ICCV, 2011. 2
[22] A. Gretton, K. Borgwardt, M. Rasch, B. Schölkopf,
and A. Smola. A kernel two-sample test. Journal of
Machine Learning Research, 2012. 2, 3
[23] A. Gretton, A. Smola, J. Huang, M. Schmittfull,
K. Borgwardt, and B. Schölkopf. Covariate shift and
local learning by distribution matching. 2009. 2
[24] K. He, G. Gkioxari, P. Dollár, and R. Girshick. Mask
R-CNN. In ICCV, 2017. 1
[25] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual
learning for image recognition. In CVPR, 2016. 1, 5
[26] J. Huang, A. Gretton, K. M. Borgwardt, P. B.
Schölkopf, and A. J. Smola. Correcting sample selection bias by unlabeled data. In NIPS, 2007. 2
[27] J. Jiang and C. Zhai. Instance weighting for domain
adaptation in nlp. In ACL, 2007. 2
[28] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural
networks. In NIPS, 2012. 1
[29] B. Kulis, K. Saenko, and T. Darrell. What you saw is
not what you get: Domain adaptation using asymmetric kernel transforms. In CVPR, 2011. 2
[30] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner.
Gradient-based learning applied to document recognition. In Proceedings of the IEEE, 1998. 1, 2
[31] T. Lin, P. Goyal, R. B. Girshick, K. He, and P. Dollár.
Focal loss for dense object detection. In ICCV, 2017.
1
[32] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona,
D. Ramanan, P. Dollr, and C. L. Zitnick. Microsoft
coco: Common objects in context. In ECCV, 2014. 6
[33] M.-Y. Liu and O. Tuzel. Coupled generative adversarial networks. In NIPS, 2016. 3
[34] M. Long, Y. Cao, J. Wang, and M. I. Jordan. Learning
transferable features with deep adaptation networks.
In ICML, 2015. 2, 3, 6, 8
[35] M. Long, H. Zhu, J. Wang, and M. I. Jordan. Unsupervised domain adaptation with residual transfer networks. In NIPS, 2016. 6, 8
[36] M. Long, H. Zhu, J. Wang, and M. I. Jordan. Deep
transfer learning with joint adaptation networks. In
ICML, 2017. 2, 3, 5, 6, 7, 8
[37] T. E. J. Mensink, J. Verbeek, F. Perronnin, and
G. Csurka. Distance-based image classification: Generalizing to new classes at near zero cost. PAMI, 2013.
2
[38] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang.
Domain adaptation via transfer component analysis.
Transactions on Neural Networks, 2011. 2, 6
[39] X. Peng, B. Usman, N. Kaushik, J. Hoffman, and
D. Wang. Visda: The visual domain adaptation challenge. In arXiv, 2017. 2, 6
[40] E. Real, J. Shlens, S. Mazzocchi, X. Pan, and V. Vanhoucke. Youtube-boundingboxes: A large highprecision human-annotated dataset for object detection in video. 2017. 6
[41] O. Rippel, M. Paluri, P. Dollr, and L. D. Bourdev.
Metric learning with adaptive density discrimination.
ICLR, 2015. 2
[42] K. Saenko, B. Kulis, M. Fritz, and T. Darrell. Adapting visual category models to new domains. In ECCV,
2010. 2, 5
[43] R. Salakhutdinov and G. E. Hinton. Learning a nonlinear embedding by preserving class neighbourhood
structure. In AISTATS, 2007. 2
[44] M. Salzmann, C. H. Ek, R. Urtasun, and T. Darrell. Factorized orthogonal latent spaces. In AISTATS,
2010. 2
[45] B. Sun, J. Feng, and K. Saenko. Return of frustratingly
easy domain adaptation. In AAAI, 2016. 2
[46] B. Sun and K. Saenko. Deep CORAL: correlation
alignment for deep domain adaptation. In ECCV 2016
Workshops, 2016. 2, 3
[47] A. Torralba and A. A. Efros. Unbiased look at dataset
bias. In CVPR, 2011. 1
[48] E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultaneous deep transfer across domains and tasks. In
ICCV, 2015. 2, 3
[49] E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversarial discriminative domain adaptation. 2017. 2,
3
[50] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and
T. Darrell. Deep domain confusion: Maximizing for
domain invariance. arXiv, 2014. 2, 3, 6
[51] L. van der Maaten and G. Hinton. Visualizing data
using t-SNE. Journal of Machine Learning Research,
2008. 8
[52] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show
and tell: A neural image caption generator. In CVPR,
2015. 1
[53] K. Q. Weinberger and L. K. Saul. Distance metric
learning for large margin nearest neighbor classification. Journal of Machine Learning Research, 2009.
2
[54] E. P. Xing, A. Y. Ng, M. I. Jordan, and S. Russell.
Distance metric learning, with application to clustering with side-information. In NIPS, 2003. 2
| 1cs.CV
|
ON A CONJECTURE BY KALAI
arXiv:1212.3726v1 [math.AC] 15 Dec 2012
GIULIO CAVIGLIA, ALEXANDRU CONSTANTINESCU, AND MATTEO VARBARO
Abstract. We show that monomial ideals generated in degree two satisfy a conjecture by
Eisenbud, Green and Harris. In particular we give a partial answer to a conjecture of Kalai
by proving that h-vectors of flag Cohen-Macaulay simplicial complexes are h-vectors of CohenMacaulay balanced simplicial complexes.
1. Introduction
An unpublished conjecture of Gil Kalai, recently verified by Frohmader [Fr], states that for
any flag simplicial complex ∆ there exists a balanced simplicial complex Γ with the same f vector. Here a (d − 1)-simplicial complex is balanced if you can use d colors to label its vertices
so that no face contains two vertices of the same colour. Kalai’s conjecture has also a second
part which is still open: If ∆ happens to be Cohen-Macaulay (CM), then Γ is required to be
CM as well.
In this note we show that for any CM flag simplicial complex ∆ there exists a CM balanced
simplicial complex Γ with the same h-vector. Such a result has been proved in [CV, Theorem
3.3] under the additional assumption that ∆ is vertex decomposable. Other recent developments
concerning Kalai’s conjecture and related topics can be found in [CN, BV]. To this purpose we
will show a stronger statement, Theorem 2.1, namely that the Eisenbud-Green-Harris conjecture
(EGH) holds for quadratic monomial ideals.
Let S = K[x1 , . . . , xn ] be a polynomial ring over a field K. The EGH conjecture, in the
general form, states that for every homogeneous ideal I of S containing a regular sequence
f1 , . . . , fr of degrees d1 ≤ · · · ≤ dr there exists a homogeneous ideal J ⊆ S, with the same
Hilbert function as I (i.e. HFI = HFJ ) and containing xd11 , . . . , xdr r . Furthermore, by a theorem
of Clements and Lindstöm [CL], the ideal J, when it exists, can be chosen to be the sum of the
ideal (xd11 , . . . , xdr r ) and a lex-segment ideal of S. We refer to [EGH1] and [EGH2] for the original
formulation of this conjecture. The only large classes for which the EGH conjecture
P is known
are: when f1 , . . . , fr are Gröbner basis (by a deformation argument), when di > j<i (dj − 1)
for all i = 2, . . . , r ([CM]) and when each fi factors as product of linear forms ([Ab, Corollary
4.3] for the case r = n, and [Ab] together with [CM, Proposition 10] for the general case).
2. The result
Theorem 2.1. Let I ⊆ S = K[x1 , . . . , xn ] be a monomial ideal generated in degree 2, of htI = g.
There exists a monomial ideal J ∈ S, such that (x21 , . . . , x2g ) ⊆ J and
HFI = HFJ .
Furthermore J can be chosen with the same projective dimension as I.
Proof. Since the Hilbert function and the projective dimension are invariant with respect to
field extension, we can assume without loss of generality that K is infinite We will prove that
Date: November 17, 2017.
2010 Mathematics Subject Classification. 13D40, 13F55, 05E45.
Key words and phrases. h-vector and face vector and Cohen Macaulay simplicial complex and flag simplicial
complex, EGH conjecture, regular sequences .
The work of the first author was supported by a grant from the Simons Foundation (209661 to G. C.).
The authors thank the Mathematical Sciences Research Institute, Berkeley CA, where this work was done, for
support and hospitality during Fall 2012 .
1
2
G. CAVIGLIA, A. CONSTANTINESCU, AND M. VARBARO
I contains a regular sequence of the form x1 ℓ1 , . . . , xg ℓg , where ℓi is a linear form for every
i ∈ [g] = {1, . . . , g}. Then we will infer the theorem by a result in [Ab].
Without loss of generality we may assume that (x1 , . . . , xg ) is a minimal prime of I. Thus,
we may decompose the degree 2 component of I as
I2 = x1 V1 ⊕ . . . ⊕ xg Vg ,
where each Vi is a linear space generated by indeterminates. Our goal is to find g linear forms
ℓi ∈ Vi , such that:
(*) for all A ⊆ [g], the K-vector space hxi : i ∈ Ai + hℓi : i ∈ [g] \ Ai has dimension g.
To see that (*) is equivalent to x1 ℓ1 , . . . , xg ℓg being a S-regular sequence (from now on we will
just write regular sequence for S-regular sequence), consider the following short exact sequence
(where C = (x1 ℓ1 , . . . , xg ℓg )):
0 → K[x1 , . . . , xn ]/(C : ℓg )(−1) → K[x1 , . . . , xn ]/C → K[x1 , . . . , xn ]/(C + (ℓg )) → 0.
By induction on the number of variables, both the extremes of the above exact sequence
are graded modules of Krull dimension ≤ n − g. By looking at the Hilbert polynomials,
K[x1 , . . . , xn ]/C has dimension ≤ n − g too, that is possible only if C is a complete intersection.
So we
Q have to seek ℓi ∈ Vi satisfying (*). Let A = {i1 , . . . , ia } be a subset of [g]. We define
UA ⊆ i∈A Vi to be the following set:
Y
UA = {(vi1 , . . . , via ) ∈
Vi : hvi1 , . . . , via i + hxj : j ∈ [g] \ Ai has dimension g}.
i∈A
As the condition of linear dependence is obtained by imposing certain determinantal relations to
Q
eA below is a Zariski open set of Qg Vi
be zero, UA is a Zariski open set of i∈A Vi . Thus the U
i=1
eA = UA ×
U
Y
Vi ⊆
g
Y
Vi .
i=1
i∈[g]\A
This construction can be done for every A ⊆ [g], and thus we can define the open set
U=
\
A⊆[g]
eA ⊂
U
g
Y
Vi .
i=1
AnyQelement (ℓ1 , . . . , ℓg ) ∈ U will automatically satisfy (*), so our goal is to show that U 6= ∅.
As gi=1 Vi is irreducible, it is enough to show that all the open sets UA ’s are nonempty. For
any A ⊆ [g] we have
(1)
dimK
X
i∈A
Vi +
X
hxj i ≥ g,
j∈[g]\A
P
P
otherwise ( i∈A Vi + j∈[g]\A hxj i) would be a prime ideal containing I of height < g.
Given A ⊆ [g], we define a bipartite graph GA . The vertex set of GA has the following
partition: V (GA ) = {x1 , . . . , xn } ∪ {1, . . . , g}, and the edge set of GA is given by:
xi ∈ Vj , if j ∈ A
{xi , j} ∈ E(GA ) ⇐⇒
i=j
, if j ∈
/A
We fix A and prove that GA satisfies the hypothesis of the Marriage Theorem. For a subset
B ⊆ V (GA ), we denote by N (B) the set of vertices adjacent to some vertex in B. Choose now
B ⊆ {1, . . . , g}. By applying (1) to the set A ∩ B ⊆ [g] we can deduce that
X
X
dimK
Vi +
hxj i ≥ |B|,
i∈A∩B
j∈([g]\A)∩B
ON A CONJECTURE BY KALAI
3
Furthermore notice that the dimension of the above vector space is |N (B)|, thus we can apply
the Marriage Theorem and infer the existence of a matching in GA of the form {xij , j}j∈[g] .
Therefore UA is nonempty as it contains (xij : j ∈ A).
So we found a regular sequence of quadrics f1 , . . . , fg in I consisting of products of linear
forms.
Let pd(I) be the projective dimension of I and assume that pd(I) = p − 1. By applying
a linear change of coordinates, we may assume that xp+1 , . . . , xn is a S/I-regular sequence.
Going modulo (xp+1 , . . . , xn ), the image I ′ ⊆ K[x1 , . . . , xp ] of I may not be monomial, but still
contains a regular sequence of quadrics which are products of linear forms, namely the image of
f1 , . . . , fg . So we find J ′ ⊆ K[x1 , . . . , xp ] containing (x21 , . . . , x2g ) with the same Hilbert function
of I ′ by [Ab, Corollary 4.3] and [CM, Proposition 10]. Clearly pd(J ′ ) ≤ p − 1, but we can
actually choose J ′ such that pd(J ′ ) = p − 1 by [CS, Theorem 4.4]. Defining J = J ′ K[x1 , . . . , xn ]
we have (x21 , . . . , x2g ) ⊆ J, pd(J) = pd(I) and HFI = HFJ .
The following example shows that the above proof cannot be extended to prove EGH for all
monomial ideals.
Example 2.2. The ideal I = (x21 x2 , x22 x3 , x1 x23 ) ⊆ K[x1 , x2 , x3 ] does not contain a regular
sequence of the form ℓ1 ℓ2 ℓ3 , q1 q2 q3 , where all ℓi and qj are linear forms. Elementary direct
computations allow one to see that the generators are the only products of three linear forms,
which are contained in I. Clearly any choice of two of them does not produce a regular sequence.
The following corollary is the main motivation for this note.
Corollary 2.3. For any CM flag simplicial complex ∆ there exists a CM balanced simplicial
complex Γ with the same h-vector.
Proof. Let g be the height of the Stanley-Reisner ideal I∆ . By Theorem 2.1, there exists a
CM ideal J ⊆ S, containing (x21 , . . . , x2g ) and with the same Hilbert function as I∆ . Since
J is unmixed, it has to be the extension to S of a monomial ideal J ′ ⊆ K[x1 , . . . , xg ]. Hence
HFK[x1,...,xg ]/J ′ equals the h-vector of ∆. The CM balanced Γ is the simplicial complex associated
to the polarization of J ′ .
References
[Ab] A. Abedelfatah, On the Eisenbud-Green-Harris conjecture, arXiv:1212.2653v1 [math.AC], 2012.
[BV] J. Biermann, A. Van Tuyl, Balanced vertex decomposable simplicial complexes and their h-vectors,
arXiv:1202.0044 [math.AC], 2012.
[CM] G. Caviglia, D. Maclagan, Some cases of the Eisenbud-Green-Harris conjecture, Math. Res. Lett. 15 (2008),
no. 3, 427-433.
[CS] G. Caviglia, E. Sbarra, The lex-plus-power inequality for local cohomology modules, arXiv:1208.2187, 2012.
[CL] G. F. Clements, B. Lindström, A generalization of a combinatorial theorem of Macaulay, J. Combinatorial
Theory 7, 230-238, 1969.
[CN] D. Cook II, U. Nagel, Cohen-Macaulay Graphs and Face Vectors of Flag Complexes, SIAM J. Discrete Math.
26, 89-101, 2012.
[CV] A. Constantinescu, M. Varbaro, On the h-vectors of Cohen-Macaulay Flag Complexes, Math. Scand., to
appear.
[EGH1] D. Eisenbud, M. Green and J. Harris, Higher Castelnuovo Theory, Asterisque 218, 187-202, 1993.
[EGH2] D. Eisenbud, M. Green and J. Harris, Cayley-Bacharach theorems and conjectures, Bull. Amer. Math.
Soc. (N.S.) 33, no. 3, 295-324, 1996.
[Fr] A. Frohmader, Face Vectors of Flag Complexes, Israel J. Math. 164, 153-164, 2008.
Department of Mathematics, Purdue University 150 N. University Street, West Lafayette, IN
47907-2067
E-mail address: [email protected]
Institut de Mathématiques, Universitè de Neuchâtel Bâtiment UniMail, Rue Emile-Argand 11
2000 Neuchâtel - Switzerland
E-mail address: [email protected]
Dipartimento di Matematica, Università degli Studi di Genova, Via Dodecaneso 35, 16146, Italy
E-mail address: [email protected]
| 0math.AC
|
The Incredible Shrinking Neural Network: New Perspectives on Learning
Representations through the Lens of Pruning
arXiv:1701.04465v2 [cs.NE] 25 Nov 2017
Aditya Sharma
Nikolas Wolfe
Bhiksha Raj
Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213 USA
Abstract
How much can pruning algorithms teach us about
the fundamentals of learning representations in
neural networks? And how much can these fundamentals help while devising new pruning techniques? A lot, it turns out. Neural network pruning has become a topic of great interest in recent
years, and many different techniques have been
proposed to address this problem. The decision
of what to prune and when to prune necessarily
forces us to confront our assumptions about how
neural networks actually learn to represent patterns in data. In this work, we set out to test several long-held hypotheses about neural network
learning representations, approaches to pruning
and the relevance of one in the context of the
other. To accomplish this, we argue in favor of
pruning whole neurons as opposed to the traditional method of pruning weights from optimally
trained networks. We first review the historical literature, point out some common assumptions it makes, and propose methods to demonstrate the inherent flaws in these assumptions.
We then propose our novel approach to pruning
and set about analyzing the quality of the decisions it makes. Our analysis led us to question
the validity of many widely-held assumptions behind pruning algorithms and the trade-offs we
often make in the interest of reducing computational complexity. We discovered that there is a
straightforward way, however expensive, to serially prune 40-70% of the neurons in a trained
network with minimal effect on the learning representation and without any re-training. It is to be
noted here that the motivation behind this work
is not to propose an algorithm that would outperThis work is being published on arXiv only. Copyright and ownership of the work remains with the authors. Please use appropriate
citation when distributing and/or referencing this work.
ADITYAS 2@ ALUMNI . CMU . EDU
NWOLFE @ CS . CMU . EDU
BHIKSHA @ CS . CMU . EDU
form all existing methods, but to shed light on
what some inherent flaws in these methods can
teach us about learning representations and how
this can lead us to superior pruning techniques.
1. Introduction and Literature Review
Representation of Learning has always been a topic of curiosity in the world of Artificial Neural Networks. It is
something that certainly warrants more attention and analysis than it currently receives. It seems very intuitive that
the knowledge of how learning is represented in a network of neurons should prove to be useful in developing
algorithms. Unfortunately a vast majority of work done
in the highly popular fields of Artificial Neural Networks
and Deep Learning doesn’t focus on leveraging this knowledge. One category of algorithms that can certainly leverage this information and yet surprisingly hasn’t is pruning
techniques. Neural Network Pruning gained some popularity in the early 90s and after a period of dormancy has
recently regained some momentum. In the current work we
take a deep dive into existing literature on the topic, both
historical and modern, and recognize some areas of failure arising from fundamentally flawed assumptions about
learning representations. We then argue in favor of a novel
approach to pruning and demonstrate through experiments
the superiority of this new way of thinking. We would like
to note here that this work does not aim to propose an exact new superior pruning algorithm, but a new way of approaching the subject which isn’t flawed in its assumptions
and hence promises to perform better, which we empirically demonstrate.
We begin our exploratory journey in 1989, jumping off
from an insightful though largely forgotten observation by
(Mozer & Smolensky, 1989a) concerning the nature of neural network learning. This observation is best summarized
in a quotation from (Segee & Carter, 1991) on the notion
of fault-tolerance in multi-layer perceptron networks:
The Incredible Shrinking Neural Network
Contrary to the belief widely held, multilayer networks are not inherently fault tolerant.
In fact, the loss of a single weight is frequently
sufficient to completely disrupt a learned function approximation. Furthermore, having a large
number of weights does not seem to improve
fault tolerance. [Emphasis added]
Essentially, (Mozer & Smolensky, 1989b) observed that
during training, neural networks do not distribute the learning representation evenly or equitably across hidden units.
By making this observation alone, this work put itself miles
ahead of a lot of work in pruning that followed in the coming two decades. It was a very straightforward observation
and yet raised a lot of interesting questions, many of which
remain unanswered even today. It largely talked about the
relevance of individual neurons in a network and argued
that it is only a few, elite neurons that learn an approximation of the input-output function, and hence the remaining useless neurons could be trimmed to produce a skeleton
network. They however, do not go into the details of why
this happens or what exactly is the role of these other noisy
neurons, if any. We will go a little deeper into this discussion later in this work.
1.1. Compressing Neural Networks
Pruning algorithms, as comprehensively surveyed by
(Reed, 1993), are a useful set of heuristics designed to identify and remove elements from a neural network which are
either redundant or do not significantly contribute to the
output of the network. This is motivated by the observed
tendency of neural networks to overfit to the idiosyncrasies
of their training data given too many trainable parameters
or too few input patterns from which to generalize, as stated
by (Chauvin, 1990). (Baum & Haussler, 1989) demonstrate
that there is no way to precisely determine the appropriate
size of a neural network a priori given any random set of
training instances. Using too few neurons seems to inhibit
learning, and so while using pruning techniques in practice it is common to attempt to over-parameterize networks
initially using a large number of hidden units and weights,
and then prune or compress these fully-trained networks as
necessary. Of course, as the old saying goes, there’s more
than one way to skin a neural network.
The idea of Skeletonization inspired others to look into
the world of faster and better generalization that pruning
techniques promised. However in doing so, they somehow lost the essence of the most important result put forward by (Mozer & Smolensky, 1989a) regarding learning
representations and instead focused on improving generalization and memory usage through fast and easy pruning of network elements. Optimal Brain Damage (LeCun
et al., 1989) was one such algorithm that quickly became
very popular, as did its later variants such as Optimal Brain
Surgeon (Hassibi & Stork, 1993). These algorithms took
a weight centric pruning approach which quickly became
the de facto standard for almost all the work that followed,
while the original ideas put forward by (Mozer & Smolensky, 1989a) were largely forgotten.
In the years to come, pruning algorithms paved the way for
model compression algorithms. Model compression often
has benefits with respect to generalization performance and
the portability of neural networks to operate in memoryconstrained or embedded environments. Without explicitly
removing parameters from the network, weight quantization allows for a reduction in the number of bytes used to
represent each weight parameter, as investigated by (Balzer
et al., 1991), (Dundar & Rose, 1994), and (Hoehfeld &
Fahlman, 1992).
A recently proposed method for compressing recurrent
neural networks (Prabhavalkar et al., 2016) uses the singular values of a trained weight matrix as basis vectors
from which to derive a compressed hidden layer. (Øland &
Raj, 2015) successfully implemented network compression
through weight quantization with an encoding step while
others such as (Han et al., 2016) have tried to expand on
this by adding weight-pruning as a preceding step to quantization and encoding.
Since (Mozer & Smolensky, 1989a), other heuristics affecting the generalization behavior of neural networks have
been well studied too, and have been used to avoid overfitting, such as dropout (Srivastava et al., 2014), maxout (Goodfellow et al., 2013), and cascade correlation
(Fahlman & Lebiere, 1989), among others. Of course,
while cascade correlation specifically tries to construct
minimal networks, many techniques that try to improve
network generalization do not explicitly attempt to reduce
the total number of parameters or the memory footprint of
a trained network per se.
In summary, we can see how since the ideas put forth by
(Mozer & Smolensky, 1989a) back in 1989 the research in
this area went on a tangential path which led to most of our
popular pruning techniques today. In subsequent sections
of this paper, we will look deeper into the original ideas
from 1989, try to answer some questions that were left
unanswered, demonstrate how not leveraging these original ideas introduced fundamental flaws in today’s pruning techniques, and finally make a case for using learning representations to drive decision making in pruning
algorithms. Through our experiments we not only concretely validate the theory put forth by (Mozer & Smolensky, 1989b) but we also successfully build on it to show
that it is possible to prune networks to 40 to 70 % of their
original size without any major loss in performance.
The Incredible Shrinking Neural Network
2. Pruning Neurons to Shrink Networks
2.1. Why Prune Neurons Instead of Pruning Weights?
It is surprising that despite the promising ideas presented
by (Mozer & Smolensky, 1989a) regarding the role of individual neurons in the overall network performance, it was
weight pruning that became popular instead. The more one
thinks about it, the more one begins to realize why pruning neurons sounds more logical than pruning weights. We
look into some of these reasons here.
First is the original argument itself made by (Mozer &
Smolensky, 1989a). Not all neurons are equal, some of
them contribute more than the others, so it makes logical sense to free up space by completely getting rid of
these unimportant neurons. Contrast this with eliminating
weights, where the pruning criteria relies on some function
of actual weight values. At no point during weight-pruning
you are getting rid of whole neurons. One might argue
that unimportant neurons will end up being removed automatically in a weight-pruning approach as their weights
will have low values, but it was demonstrated by (Mozer
& Smolensky, 1989b) that this is very much not the case.
The actual values of the weights emanating from a neuron have nothing to do with its contribution to the overall
performance of the network. Even if a weight-pruning algorithm managed to completely remove a neuron, it might
end up being an important neuron which would disrupt the
network completely.
Second, again a quite obvious reason, pruning neurons as
opposed to individual weights frees up more space since
you eliminate all incoming and outgoing weights from the
pruned neuron in one go. The removal of a single weight
from a large network is a drop in the bucket in terms of
reducing a network’s core memory footprint. Somewhat
tied to this is the third reason, pruning neurons instead of
weights means that you have far fewer (orders of magnitude
less) number of elements to consider during the pruning
process itself, which makes it much cheaper computationally. Pruning neurons has a practical advantage in terms of
quickly reaching a hypothetical target reduction in memory
consumption.
Fourth, as touched upon in the first reason, the values of
individual weights have nothing to do with the relevance of
the corresponding neurons. The neurons represent learning, the weights are just a means of storing this learning
in memory by splitting it into smaller pieces used to interact with other neurons. Weights provide little information
about learning individually and it is their interaction with
connected neurons that makes their existence meaningful, a
fact that most weight-pruning algorithms tend to overlook.
In fact, this is a major drawback in the case of quantizationbased algorithms as these methods completely disregard
the individual role of neurons and focus solely on quantizing weight values.
Lastly, if you look at it in terms of features, each neuron
learns exactly one feature, with the features becoming more
complex as you move to layers closer to the output. As
in any Machine Learning problem, many of these features
are redundant, the “Train Problem” discussed in (Mozer &
Smolensky, 1989b) is a very good basic example of this.
In practice, these redundant features should be removed
(which is usually done in a “feature engineering” step in
classic ML techniques) as they add noise, increase training
time and lead to wrong generalization. In neural networks,
which have a more end-to-end architecture as compared to
classic ML techniques, this can only be achieved by removing whole neurons. In contrast, pruning weights only modifies such features thereby removing their redundant components, but doesn’t take into account the overall relevance
of the feature and hence never completely removes it.
2.2. Setting up the experiments
Now that we have an understanding of how pruning algorithms evolved over the years, some faulty assumptions
they made about learning representations, and finally, the
benefits of pruning whole neurons instead of individual
weights, we are ready to look deeper into the mechanics
of popular pruning algorithms. We will look at some other
assumptions they make and experimentally analyze their
performance when compared to a very simple brute-force
based approach which carefully takes into consideration the
role of learning representations while pruning networks.
Since we have already discussed the benefits of pruning
neurons over weights, all experiments presented in this paper will use neuron pruning. We will also not experimentally evaluate algorithms that quantize weights and compress networks because one, they are not exactly pruning
algorithms and two, their biggest fundamental flaw has already been discussed while making the argument against
weight pruning earlier in this section.
It is beyond the scope of this paper to do an empirical analysis of every single pruning algorithm out there. We could
have selected a top few popular algorithms for our comparisons but that would defeat the central purpose of this
work, which is to look for major flaws in the fundamental
assumptions made by these algorithms. This requires an
approach that is representative of the central ideas used by
most, if not all of these existing algorithms.
2.2.1. L INEAR AND Q UADRATIC A PPROXIMATION
A PPROACH
The overall structure of most pruning algorithms is pretty
much the same: you rank individual elements (weights or
neurons) based on some criteria (usually related to the el-
The Incredible Shrinking Neural Network
ement’s impact on the network error) and keep trimming
away the least relevant elements until you see a drop in performance. The overall impact of an element on the network
performance in these algorithms is typically represented by
a linear or quadratic approximation of change in error. We
will compare both these approaches to our simple bruteforce approach. The experiments presented in this work
will follow a ranking-based procedure very similar to most
popular pruning techniques. The linear and quadratic error
estimates were derived using a Taylor Series representation of the network error, very similar to what (Mozer &
Smolensky, 1989a) did in their original work. The math
involved for the quadratic estimate is somewhat complicated, as is its implementation which includes gradients
collected from second-order back-propagation and hence,
we will not go into the details of either in the main body
of the paper. All these details however, are provided in the
Supplementary Material as an Appendix for readers who
might be interested in pursuing and implementing this.
2.2.2. B RUTE F ORCE R EMOVAL A PPROACH
This is perhaps the most intuitive, naive and yet the most
accurate method for pruning a network, but surprisingly,
not very popular, perhaps due to its time complexity. It
does not make any assumptions about the learning representations in the network, nor does it make any approximate estimates about the impact of individual elements on
the network’s output, which make it a perfect candidate to
serve as the ground truth for our experiments. This method
explicitly evaluates each neuron in the network. The idea
is to manually check the effect of every single neuron on
the output. This is done by running a forward propagation
on the validation set K times (where K is the total number
of neurons in the network), turning off exactly one neuron each time (keeping all other neurons active) and noting
down the change in error.
2.2.3. N EURON I NTER - DEPENDENCIES
Something that was briefly touched upon in the Section 1
was the fact that while the findings of (Mozer & Smolensky, 1989a) about learning representations were of great
significance, one important question remained unanswered:
Do the noisy/unimportant neurons have a role to play once
the network is trained? Although there is no empirical evidence to support their role during the training process itself,
one can still easily imagine the necessity to include these
extra neurons to aid better generalization. However, their
role after the network is trained is something we would
like to explore. (Mozer & Smolensky, 1989a) proved that
these neurons don’t really contribute to the network output (which is why we prune them), but does their existence
impact the other neurons in some way? In order to find
out, we begin our experiments exactly where (Mozer &
Smolensky, 1989a) left off, and we don’t assume these unnecessary neurons to have any impact on each other or the
relevant neurons. We use Algorithm 1 to verify how correct
this assumption is. We will analyze our observations and
follow them up with with further experiments in the next
section. The idea here is to generate a single ranked list
based on the values of ∆Ek , the Taylor Series approximation of the change in the overall error of the network due to
the neuron k. This involves a single pass of second-order
back-propagation (without weight updates) to collect the
gradients for each neuron. The neurons from this rank-list
(with the lowest values of ∆Ek ) are then pruned according
to the stopping criterion decided.
Algorithm 1 Single Overall Ranking
Data: An optimally trained network, training dataset
Output:A pruned network
Initialize and define stopping criterion;
Perform forward propagation over the training set;
Perform second-order back-propagation without updating weights and collect linear and quadratic gradients;
Rank the remaining neurons based on ∆Ek ;
while stopping criterion is not met do
Remove the last ranked neuron;
end while
3. Experimental Results & Discussion
For most experiments presented in this section, the MNIST
database of Handwritten Digits by (LeCun & Cortes, 2010)
was used. It is worth noting that due to the time taken
by the brute force algorithm we used a 5000 image subset
of the MNIST database in which we normalized the pixel
values between 0 and 1.0, compressed the image sizes to
20x20 images rather than 28x28, and used a 1 or 2-layer
fully connected feedforward neural network with a total of
100 neurons and sigmoid activation, so the starting test accuracy reported here appears higher than those reported by
LeCun et al. This modification does not affect the interpretation of the presented results because the fundamental
learning problem does not change with a larger dataset or
input dimension or network architecture. Also, as has been
pointed out multiple times before, the aim of this work is
not to propose a superior algorithm, but to point out flaws in
the assumptions that existing algorithms make by means of
the better performing brute force method, hence the choice
of dataset does not impact our results. Wherever necessary, some other toy datasets have also been used to empirically verify some novel theoretical insights presented
in the previous sections of this work. The sigmoid activation function was chosen since it makes estimating gradients easy (as compared to a ReLU activation function), and
more importantly because it has been the popular choice
The Incredible Shrinking Neural Network
historically, which helps us make sure our experiments are
representative of the older pruning techniques as well.
3.1. Neuron Inter-dependencies
We start off by trying to answer the question we put forth at
the end of the Section 2. Do the irrelevant neurons in a network still have a role to play after the training is completed?
As mentioned before, we first start off by assuming that the
answer to the question is no, simply because we saw no
historical evidence that suggested otherwise. We present
the results for a single-layer neural network in Figure 1,
using Single Overall Ranking (Algorithm 1) as proposed in
Section 2. After training, each neuron is assigned its permanent ranking based on the three criteria discussed previously: A brute force “ground truth” ranking, and the linear
and quadratic error approximations that represent popular
existing techniques in principle.
8000
6000
5000
4000
3000
2000
1000
0
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
Sum of Squared Errors
Sum of Squared Errors
8000
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
6000
5000
4000
3000
2000
1000
0
20
40
60
Percentage of Neurons Removed
80
100
0
0
20
40
60
Percentage of Neurons Removed
80
100
Figure 1. Degradation in squared error after pruning a 1-layer network (left) and a 2-layer network (right) using The Single Overall
Ranking algorithm (Network 1: 1 layer, 100 neurons, starting test
accuracy: 0.998; Network 2: 2 layers, 50 neurons/layer, starting
test accuracy: 1.000)
As can be observed, none of the three approaches perform
well at all. The performance starts dropping as soon as
neurons are removed according to the ranked list. This is
a very important result, although not unexpected. This behavior implies that there is a flaw in our assumption that
the irrelevant neurons have no role to play in the network’s
performance post-training. We performed the same experiment on other toy datasets and we got the same results.
These results answer the question posed at the beginning
of this section. These irrelevant neurons do have a role to
play post-training. Their existence is somehow linked to
the performance of the relevant neurons, and as (Segee &
Carter, 1991) assert, the removal of any single hidden unit
at random can actually result in a network fault. We have
empirically, if not numerically, proved the role of these irrelevant units but this makes us wonder if it is possible,
without going into the mathematics of the inter-neural relationships, to come up with a way to still prune networks
efficiently.
What we did wrong in the experiment above was ignore
the fact that after a neuron is removed, no matter how irrelevant its existence was to the overall performance, it
does have an impact on the performance of the other neurons, some of which might be highly relevant in the overall context. Keeping that, and the intriguing fault tolerance (or lack thereof) of trained neural networks in mind,
we decided to modify our algorithm. In this greedy variation of the algorithm (Algorithm 2), after each neuron
removal, the remaining network undergoes a single forward and backward pass of second-order back-propagation
(without weight updates) and the rank list is formed again.
Hence, each removal involves a new pass through the network. This method takes into account the dependencies
the neurons might have on one another each time the rank
list is re-formed and relies on the low fault tolerance of the
network which disallows large error drops that might come
with the removal of a neuron, but only as long as all the
relevant neurons are still intact.
Algorithm 2 Iterative Re-Ranking
Data: An optimally trained network, training dataset
Output:A pruned network
Initialize and define stopping criterion;
while stopping criterion is not met do
Perform forward propagation over the training set;
Perform second-order back-propagation without updating weights and collect linear and quadratic gradients;
Rank the remaining neurons based on ∆Ek ;
Remove the worst neuron based on the ranking;
end while
3.2. Example Regression Problem
Even though the results above decisively prove the significance of irrelevant neurons, they still do not prove the fact
that neuron pruning is actually possible. So far we have
considered the results from (Mozer & Smolensky, 1989a)
as the gospel truth. Since these results have been long forgotten, and since most of the reasons cited by us in favor of
the novel approach of pruning whole neurons in Section 2
were largely theoretical, one might argue for more empirical evidence. Hence, to get some heuristic validation we
present results from a simple regression experiment. With
this experiment we hope to prove both the existence of nonuniform learning distribution across neurons as well as the
efficiency of a neuron-pruning approach.
We trained two networks to learn the cosine function, with
one input and one output. This is a task which requires
no more than 11 sigmoid neurons to solve entirely, and in
this case we don’t care about overfitting because the cosine
function has a precise definition. Furthermore, the cosine
function is a good toy example because it is a smooth continuous function and, as demonstrated by (Nielsen, 2015),
if we were to tinker directly with the parameters of the net-
The Incredible Shrinking Neural Network
2000
3000
2500
2000
1500
1000
1000
500
0
0
5
10
Number of Neurons Removed
15
20
0
0
20
40
60
Number of Neurons Removed
80
100
Figure 2. Degradation in squared error after pruning a two-layer
network trained to compute the cosine function (Left Network:
2 layers, 10 neurons each, starting test accuracy: 0.9999993,
Right Network: 2 layers, 50 neurons each, starting test accuracy:
0.9999996)
Figure 7 shows two graphs. Both graphs demonstrate the
use of Algorithm 2 and the comparative performance of the
brute-force pruning method (in blue), the first order method
(in green), and the second order method (in red). The graph
on the left shows the performance of these methods on a
network with two layers of 10 neurons (20 total), and the
graph on the right shows a network with two layers of 50
neurons (100 total).
In the left graph, we see that the brute-force method shows
a graceful degradation, and the error only begins to rise
sharply after 50% of the total neurons have been removed.
The error is basically constant up to that point. In the first
and second order methods, we see evidence of poor decision making in the sense that both made mistakes early on,
which disrupted the output function approximation. The
first order method made a large error early on, though we
see after a few more neurons were removed this error was
corrected somewhat (though it only got worse from there).
This is direct evidence of the lack of fault tolerance in a
trained neural network once a relevant unit is removed.
This phenomenon is even more starkly demonstrated in the
second order method. After making a few poor neuron removal decisions in a row, the error signal rose sharply, and
then went back to zero after the 6th neuron was removed.
This is due to the fact that the neurons it chose to remove
were trained to cancel each others’ influence within a localized part of the network. After the entire group was elimi-
This experiment gives us some very important results. One,
it empirically verifies the fact that learning is not equally or
equitably distributed in a neural network. Two, it proves
that pruning whole neurons without a loss in performance
is actually possible and three, it shows us some early indications of the inefficient nature of popular first and second order error approximation techniques. Clearly, in the
case of the brute force or oracle method, which makes no
faulty assumptions about the nature of learning representations, up to 90% of the network can be completely extirpated before the output approximation even begins to show
any signs of degradation. We will explore this further in the
next sections.
3.3. Experiments on MNIST and Toy Datasets
8000
6000
5000
4000
3000
2000
0
20
40
60
Percentage of Neurons Removed
8000
80
0.8
0.6
0.4
0.0
100
0
20
1.2
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
40
60
Percentage of Neurons Removed
5000
4000
3000
2000
80
100
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.0
6000
0.8
0.6
0.4
0.2
1000
0
1.0
0.2
1000
0
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.2
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
Classification Accuracy
3000
Classification Accuracy
3500
This phenomenon is even more definitely demonstrated in
the graph on the right. Here we see the first order method
got “lucky” in the beginning and made decent decisions
up to about the 40th removed neuron. The second order
method had a small error in the beginning which it recovered from gracefully and proceeded to pass the 50 neuron
point before finally beginning to unravel. The brute force
method, in sharp contrast, shows little to no increase in error at all until 90% of the neurons in the network have been
obliterated. Clearly first and second order methods have
some value in that they do not make completely arbitrary
choices, but the brute force method is far better at this task.
Sum of Squared Errors
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
4000
Sum of Squared Errors
Sum of Squared Errors
4500
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
4000
nated, the approximation returned to normal. This can only
happen if the output function approximation is not evenly
distributed over the hidden units in a trained network.
Sum of Squared Errors
work, we could allocate individual units within the network
to be responsible for constrained ranges of inputs, similar
to a basis spline function with many control points. This
would distribute the learned function approximation evenly
across all hidden units, and thus we have presented the network with a problem in which it could productively use as
many hidden units as we give it. In this case, a pruning algorithm would observe a fairly consistent increase in error
after the removal of each successive unit. In practice however, regardless of the number of experimental trials, this is
not what happens. The network will always use 10-11 hidden units and leave the rest to cancel each others influence,
as we will see.
0
20
40
60
Percentage of Neurons Removed
80
100
0.0
0
20
40
60
Percentage of Neurons Removed
80
100
Figure 3. Degradation in squared error (left) and classification accuracy (right) after pruning a single-layer network (top) and 2layer network (bottom) using Algorithm 2 (Top Network: 1
layer, 100 neurons, starting test accuracy: 0.998; Bottom Network: 2 layers, 50 neurons/layer, starting test accuracy: 1.000)
With encouraging observations from our previous experiments, in Figure 9 we present our results using Algorithm
The Incredible Shrinking Neural Network
1.2
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
15000
10000
5000
1.0
0.9
0.8
0.7
0.6
0.5
0
0
20
40
60
Number of Neurons Removed
30000
0
20
1.1
40
60
Number of Neurons Removed
Classification Accuracy
15000
10000
80
100
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.0
20000
0.9
0.8
0.7
0.6
0.5
0.4
5000
0
0.4
100
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
25000
Sum of Squared Errors
80
0.3
0
20
40
60
Number of Neurons Removed
80
100
0
20
40
60
Number of Neurons Removed
80
Sum of Squared Errors
Sum of Squared Errors
103
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
4000
102
3000
2000
1000
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
Sum of Squared Errors
104
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
9000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
103
0
10
7000
102
10
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
8000
Sum of Squared Errors
101
6000
5000
4000
3000
2000
1000
101
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
104
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
9000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
103
0
10
8000
7000
102
10
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
6000
5000
4000
3000
2000
1000
101
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
0
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
Figure 5. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the brute force, first order and second order
criteria
3.4. Visualization of Error Surface & Pruning
Decisions
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.1
Classification Accuracy
Sum of Squared Errors
20000
5000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
Sum of Squared Errors
In order to ensure that our observations were not just a result of idiosyncrasies of the MNIST dataset, we did the
same experiment on two toy datasets (Figure 18). These
datasets were simple pattern recognition problems where
the network’s job was to predict whether a given point
would lie inside a given shape on the Cartesian plane. The
performance of the 2nd-order method was found to be exceptionally good and produced results very close to the
brute force method, although it demonstrated a very irregular decline past the tipping point. The 1st-order method,
as expected, performed poorly here as well.
104
Sum of Squared Errors
2 (the Iterative Re-ranking Algorithm) on single layer and
2-layer networks trained on the MNIST dataset. There are
a few key observations here. Using the brute force ranking criteria, almost 60% of the neurons in the network can
be pruned away without any major loss in performance in
the case of a single-layered network. The other noteworthy
observation here is that the 2nd order Taylor Series approximation of the error performs consistently better than its 1st
order version. It is clear that it becomes harder to remove
neurons 1-by-1 with a deeper network, which makes sense
because the neurons have more inter-dependencies in that
case. Even with a more complex network, it is possible
to remove up to 40% of the neurons with no major loss in
performance which is clearly illustrated by the brute force
curve. The more important observation here is the clear potential of an ideal pruning technique and the inconsistency
of 1st and 2nd order Taylor Series approximations of the
error as ranking criteria.
100
Figure 4. Degradation in squared error (left) and classification accuracy (right) after pruning a 2-layer network using Algorithm 2
on a toy “diamond” shape dataset (top) and a toy “random shape”
dataset (below); (Network: 2 layers, 50 neurons/layer, 10 outputs, logistic sigmoid activation, starting test accuracy: 0.992(diamond); 0.986(random shape)
The graphs in Figure 10 are a visualization of the error surface of the network output with respect to the neurons chosen for removal using each of the 3 ranking criteria, represented in intervals of 10 neurons. Between 0.0 and 1.0,
we are graphing the literal effect of turning the neuron off
(α = 0), and when α > 1.0 we are simulating a boosting of the neuron’s influence in the network. It can be observed easily that for certain neurons in the network, even
doubling, tripling, or quadrupling the scalar output of the
neuron has no effect on the overall error of the network,
indicating the remarkable degree to which the network has
learned to ignore the value of certain parameters. In other
cases, we can get a sense of the sensitivity of the network’s
output to the value of a given neuron when the curve rises
steeply after the red 1.0 line. This indicates that the learned
value of the parameters emanating from a given neuron are
relatively important, and this is why we should ideally see
sharper upticks in the curves for the later-removed neurons
in the network, that is, when the neurons crucial to the
learning representation start to be picked off. As can be
easily observed, these functions are too complex to be ap-
The Incredible Shrinking Neural Network
proximated by straight lines or parabolas, which explains
why the popular Taylor Series based methods fail so miserably in our experiments.
4. Conclusions & Future Work
While this work had never promised proposing a fullfledged pruning algorithm to outperform the existing ones,
it had three major goals, all of which it successfully managed to achieve:
1. Attaining new and deeper insights about learning representations in neural networks.
2. Studying existing pruning techniques to figure out the
flaws in the assumptions they make and analyzing
these flaws empirically.
3. Using this knowledge to propose a novel paradigm
of pruning neural networks and both theoretically and
empirically demonstrating its fundamental advantages
and correctness over existing techniques.
Through our experiments and exploratory arguments, we
were successful in verifying and reinforcing some forgotten
observations, were able to answer some previously unanswered questions about learning representations, were able
to analyze flaws in existing pruning techniques and finally,
were able to propose a novel approach to pruning along
with demonstrating its benefits over traditional approaches.
We started our exploratory journey by looking at historical literature in the field of neural network pruning. We
discovered that the work done by (Mozer & Smolensky,
1989a) was not only the first successful attempt at pruning but it also put forth an important insight regarding the
unequal distribution of learning representations in trained
neural networks, which was largely forgotten over time.
We looked at the development of other popular pruning
algorithms over the years and made the observation that
these algorithms make assumptions which are not necessarily true. We found that some of the most common mistakes
these algorithms made were:
• Ignoring the fact that in a trained neural network,
learning is distributed unequally across neurons.
• Ignoring the fact that neurons in a trained network,
irrespective of their overall relevance, rely on each
other and hence should not be treated independently,
thereby assuming that pruning can be done in a serial
fashion.
• Assuming that the error function with respect to each
individual neuron can be approximated with a straight
line or more precisely with a parabola.
We proposed a novel approach to neural network pruning
as we argued in favor of pruning whole neurons instead of
individual weights in Section 2, citing multiple theoretical
and logical reasons. We were later able to confirm the feasibility and correctness of this approach empirically through
our experiments in Section 3.
One major accomplishment of this work was to answer the
previously unanswered question about the relevance of unnecessary neurons in a trained network and the existence
of inter-dependencies between relevant and irrelevant neurons. We were able to demonstrate empirically in Section
3.1 that these irrelevant neurons do have a role to play in
a fully-trained network as the relevant neurons of the network rely on them. Without going into the mathematics
of this relationship we were able to demonstrate that the
removal of one irrelevant neuron changes the dynamics of
the network, following which the contributions of all the
remaining neurons need to be reassessed, something that
most popular pruning algorithms largely ignore.
Through our experiment on the Example Regression Problem in Section 3.2, we were able to conclusively and empirically verify the original observation made by (Mozer &
Smolensky, 1989a) regarding the unequal and inequitable
distribution of learning representations in a trained network. We also saw a strong evidence in support of this
during our experiments on visualizing the error surfaces.
Through our experiments in Section 3.3, we were able
to demonstrate the flaws in popular pruning techniques
which use a linear or quadratic error approximation function. These techniques performed poorly as compared to
the simple brute-force pruning method which did not make
any faulty assumptions about learning representations or
error approximations. We were also able to analyze actual
error surfaces which gave us a deeper insight into how these
algorithms make bad pruning decisions by approximating
complex functions using straight lines and parabolas. We
saw how easy was it for these methods to get hopelessly
lost after making a bad pruning decision resulting in faults.
Furthermore, we found that the brute-force algorithm does
surprisingly well, despite being computationally expensive. We argue that further investigation is warranted to
make this algorithm computationally tractable. We have
observed that pruning whole neurons from an optimally
trained network without major loss in performance is not
only possible but also enables compressing networks to 4070% of their original size. However expensive, it would be
extremely easy to parallelize the brute-force method, or potentially approximate it using a subset of the training data
to decide which neurons to prune. This avoids the problem
of trying to approximate the importance of a unit and potentially making a mistake. It is also important to note that
pruning is only a one-time step in the context of the over-
The Incredible Shrinking Neural Network
all Machine Learning problem and hence its computational
complexity can be a good tradeoff for the compression efficiency it brings to the table.
LeCun, Yann, Denker, John S, Solla, Sara A, Howard,
Richard E, and Jackel, Lawrence D. Optimal brain damage. In NIPs, volume 89, 1989.
Finally, we encourage the readers of this work to take these
results into consideration when making decisions about
pruning methods. We also encourage them to leverage this
knowledge and insights about learning representations to
propose better approximations to the brute force neuronpruning algorithm. It should be remembered that various
heuristics may perform well in practice for reasons which
are in fact orthogonal to the accepted justifications given by
their proponents.
Mozer, Michael C and Smolensky, Paul. Skeletonization:
A technique for trimming the fat from a network via relevance assessment. In Advances in neural information
processing systems, pp. 107–115, 1989a.
References
Balzer, Wolfgang, Takahashi, Masanobu, Ohta, Jun, and
Kyuma, Kazuo. Weight quantization in boltzmann machines. Neural Networks, 4(3):405–409, 1991.
Baum, Eric B and Haussler, David. What size net gives
valid generalization? Neural computation, 1(1):151–
160, 1989.
Chauvin, Yves. Generalization performance of overtrained
back-propagation networks. In Neural Networks, pp. 45–
55. Springer, 1990.
Dundar, Gunhan and Rose, Kenneth. The effects of quantization on multilayer neural networks. IEEE transactions on neural networks/a publication of the IEEE Neural Networks Council, 6(6):1446–1451, 1994.
Fahlman, Scott E and Lebiere, Christian. The cascadecorrelation learning architecture. 1989.
Goodfellow, Ian J, Warde-Farley, David, Mirza, Mehdi,
Courville, Aaron, and Bengio, Yoshua. Maxout networks. arXiv preprint arXiv:1302.4389, 2013.
Han, Song, Mao, Huizi, and Dally, William J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv
preprint arXiv:1510.00149v5, 2016.
Hassibi, Babak and Stork, David G. Second order derivatives for network pruning: Optimal brain surgeon. Morgan Kaufmann, 1993.
Hoehfeld, Markus and Fahlman, Scott E.
Learning
with limited numerical precision using the cascadecorrelation algorithm. IEEE Transactions on Neural Networks, 3(4):602–611, 1992.
LeCun, Yann and Cortes, Corinna. MNIST handwritten
digit database. 2010.
Mozer, Michael C and Smolensky, Paul. Using relevance to
reduce network size automatically. Connection Science,
1(1):3–16, 1989b.
Nielsen,
Michael
A.
Neural
networks
and
deep
learning.
URL:
http://neuralnetworksanddeeplearning.com/chap4.html.(visited:
01.11. 2014), 2015.
Øland, Anders and Raj, Bhiksha. Reducing communication overhead in distributed learning by an order of
magnitude (almost). In IEEE International Conference
on Acoustics, Speech and Signal Processing, pp. 2219–
2223, 2015.
Prabhavalkar, Rohit, Alsharif, Ouais, Bruguier, Antoine,
and McGraw, Lan. On the compression of recurrent neural networks with an application to lvcsr acoustic modeling for embedded speech recognition. In 2016 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5970–5974. IEEE, 2016.
Reed, Russell. Pruning algorithms-a survey. Neural Networks, IEEE Transactions on, 4(5):740–747, 1993.
Segee, Bruce E and Carter, Michael J. Fault tolerance
of pruned multilayer networks. In Neural Networks,
1991., IJCNN-91-Seattle International Joint Conference
on, volume 2, pp. 447–452. IEEE, 1991.
Srivastava, Nitish, Hinton, Geoffrey, Krizhevsky, Alex,
Sutskever, Ilya, and Salakhutdinov, Ruslan. Dropout:
A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):
1929–1958, 2014.
The Incredible Shrinking Neural Network
The second derivative of the sigmoid is easily derived from
the first derivative:
A. Second Derivative Back-Propagation
σ 0 (x) = σ(x) (1 − σ(x))
d
σ(x) (1 − σ(x))
σ 00 (x) =
dx |{z} | {z }
(1)
P xj
σ(·)
(0)
(0) (1)
cji = wji oj
(0)
P xi
f (x)
σ(·)
E
Figure 6. A computational graph of a simple feed-forward network illustrating the naming of different variables, where σ(·) is
the nonlinearity, MSE is the mean-squared error cost function and
E is the overall loss.
(m)
oi
(m)
) xi
0
=
X
(m) (m+1)
j
(1)
(11)
d
=
The first and second derivatives of the cost function with
respect to the outputs:
∂E
(0)
∂oi
(0)
= oi
− ti
(2)
∂ E
(0) 2
=1
d2 oi
(m) 2
(3)
∂oi
=
dxi
(m)
σ 0 (x) = σ(x) (1 − σ(x))
(4)
σ 00 (x) = σ 0 (x) (1 − 2σ(x))
(5)
(m)
= σ(xi
(m)
doi
(m)
dxi
)
(m)
dxi
(12)
(13)
(m)
(m)
= oi
1 − oi
(14)
!
(15)
(m)
(m)
(m)
1 − oi
1 − 2oi
= oi
(m)
= σ 00 xi
(16)
(17)
Derivative of the error with respect to the ith neuron’s input
(0)
xi in the output layer:
∂E
(0)
(0)
=
∂E ∂oi
(0)
(18)
(0)
∂o ∂xi
i
(0)
(0)
(0)
= oi − t i σ x i
1 − σ xi
| {z } |
{z
}
from (2)
The first and second derivatives of the sigmoid function in
forms depending only on the output:
(m)
oi
d
∂xi
2
dx
i
(m)
(m)
= oi (m) (m+1)
1 − oi
(m)
cji = wji oj
(m)
= σ 0 xi
(m)
A.1. First and Second Derivatives
(9)
(10)
(m)
wji oj
Superscripts represent the index of the layer of the network
in question, with 0 representing the output layer. E is the
(m)
squared-error network cost function. oi is the ith output
in layer m generated by the activation function σ, which in
(m)
this paper is is the standard logistic sigmoid. xi is the
weighted sum of inputs to the ith neuron in the mth layer,
(m)
and cji is the contribution of the jth neuron in the m + 1
layer to the input of the ith neuron in the mth layer.
0
σ 00 (x) = σ 0 (x)(1 − 2σ(x))
(m)
(m)
00
(8)
0
σ (x) = σ (x) − 2σ(x)σ (x)
doi
dxi
= σ(xi
0
And for future convenience:
Name and network definitions:
1 X (0)
(o − ti )2
2 i i
00
g(x)
σ (x) = σ (x)(1 − σ(x)) − σ(x)σ (x)
MSE
E=
0
(7)
σ (x) = f (x)g(x) + f (x)g 0 (x)
(0)
oi
00
(6)
from (4)
(0)
(0)
=
− t i oi
1 − oi
(0)
(0)
= oi − ti σ 0 xi
(0)
oi
(19)
(20)
(21)
Second derivative of the error with respect to the ith neu-
The Incredible Shrinking Neural Network
(0)
ron’s input xi
(0)
contribution cji :
in the output layer:
∂2E
(0) 2
=
∂xi
(0)
∂E ∂oi
(0)
(0)
∂oi ∂xi
∂
(0)
∂xi
(0)
(0)
(0) 2
∂
=
(0)
∂cji
∂cji
∂E ∂ 2 oi
+
(0)
∂xi
(0)
∂oi
(0) 2
∂
=
∂xi
(0)
∂cji
(23)
2
∂ E
=
(0)
(0)
∂xi ∂oi
(0)
oi
|
1−
{z
(0)
oi
+
}
from (4)
∂E
(0)
(0)
0
(34)
=
o
−
t
σ
x
i
i
i
(0)
∂cji
{z
}
|
from (32)
(0)
∂oi
∂xi ∂oi
∂2E
(22)
(0)
∂2E
=
!
|
(0)
oi
− ti
{z } |
(0)
oi
from (2)
(0)
(0)
σ xi
− t i σ 0 xi
X
∂
(0)
(0)(m) (m+1)
1=− oi(0) σ1− 2oiwji oj
∂cji{z
}
j
from (5)
(35)
− ti σ 0
X
(m) (m+1)
wji oj
j
(36)
(24)
X
X (0)
∂
(0)
0
∂ ∂E
∂
∂2E
(0)
(0) =
(0) σ
c
−
t
σ
cji
i
ji
(0)
=
=
oi − t i =
= oi
1−
(0)
(0)
(0)
(0)
(0)
∂cojii
j
j
∂xi ∂oi
∂xi ∂oi
∂xi | {z }
|
{z
}
(37)
from (2)
from (4)
(25)
∂
(0)
(0)
0
2
=
(0) σ cji + k − ti σ cji + k (38)
∂2E
(0)
(0)
(0)
(0)
(0)
(0)
{z
}|
{z
}
= oi
1 − oi
+ o i − t i oi
1 − oi
1∂c
−ji2oi|
(0)
(0)
(0) 2
f
c
g
c
∂xi
ji
ji
(26)
2
We now make use of the abbreviations f and g:
(0)
(0)
(0)
= σ 0 xi
+ oi − ti σ 00 xi
(0)
(0)
(0)
(0)
(27)
= f 0 cji g cji + f cji g 0 cji
(39)
(0)
(0)
(0)
(0)
= σ 0 cji + k σ 0 cji + k + σ cji + k − ti σ 00 cji + k
First derivative of the error with respect to a single input
(40)
(0)
contribution cji from neuron j to neuron i with weight
2
(0)
(0)
(0)
(0)
= σ 0 cji + k + oi − ti σ 00 cji + k
w in the output layer:
!
(0)
∂oi
(0)
∂xi
ji
∂E
(0)
∂cji
(0)
=
∂E ∂oi
(0)
∂oi
(0)
∂xi
(0)
∂xi
(0)
(0)
= oi − ti oi
| {z } |
from (2)
(m)
∂xi
(m)
∂
=
∂cji
(m)
∂cji
x(m)
=
i
(0)
1 − oi
{z
∂x(0)
from (4)
}
i
(0)
∂cji
X
(m) (m+1)
wji oj
2
∂ E
=
∂E
(0)
(0)
= σ 0 xi
|
(m)
c∂ji2 E + k ∂=2 E
1
=
2
(0)
(0) 2
∂cji
∂xi
=
∂
(m)
∂cji
(31)
X
(m) (m+1)
wji oj
(0)
= xi
j
2
(0)
(0)
+ oi − ti σ 00 xi
{z
}
(43)
from (27)
(44)
A.1.1. S UMMARY O F O UTPUT L AYER D ERIVATIVES
(32)
∂E
(0)
∂oi
from (21)
(0)
∂cji
(0) 2
∂cji
(30)
∂E
(0)
(0)
(0)
=
o
−
t
o
1
−
o
i
i
i
i
(0)
∂cji
(0)
(0)
= oi − ti σ 0 xi
|
{z
}
∂E
(0)
cji =
(42)
(29)
j
X
j
!
c(0)
ji + k =
(28)
(0)
∂cji
(41)
(0)
= oi
∂2E
− ti
(0) 2
=1
(45)
∂oi
(33)
∂xi
∂E
(0)
∂xi
Second derivative of the error with respect to a single input
(0)
(0)
= oi − t i σ 0 x i
∂2E
(0) 2
2
(0)
(0)
= σ 0 xi
+ oi − ti σ 00
∂xi
(46)
The Incredible Shrinking Neural Network
∂E
(0)
∂cji
∂2E
∂E
=
(0) 2
(0)
∂xi
∂2E
=
∂cji
(0) 2
(47)
∂xi
2
∂ E
(1) 2
∂oj
∂
=
X (0) (1)
X (0) (1)
(0)
σ
wji oj wji
wji oj − ti σ 0
X
(1)
∂oj
i
j
j
{z
|
A.1.2. H IDDEN L AYER D ERIVATIVES
(1)
f oj
The first derivative of the error with respect to a neuron
(1)
with output oj in the first hidden layer, summing over all
partial derivative contributions from the output layer:
=
X
f0
}|
{z
(1)
g oj
}
(55)
(1)
(1)
(1)
(1)
oj
g oj
+ f oj
g 0 oj
i
(56)
(0)
X
X
X
(0)
(0)
X ∂E ∂o ∂x ∂cji
X (0)
(0) (1)
(0)
(0) (1)
(0)
∂E
(0)
(0)
i
i
wji oj wji + . . .
σ0
wji oj wji σ 0
wji =
=
=
oi − ti σ 0 xi
(1)
(0)
(0)
(0)
(1)
j
i
j
∂oj
∂cji ∂oj
{z
}
i ∂oi ∂xi
i |
from (21)
(57)
(48)
X
X (0) (1)
X (0) (1) (0) 2
(m)
σ
wji oj − ti σ 00
wji oj wji
∂cji
∂
(m)
(m) (m+1)
(m)
=
c
=
w
o
=
w
ji
ji
j
ji
i
j
j
(m+1)
(m+1)
∂oj
∂oj
(58)
(49)
2
X (0) 2 (0) 2 (0)
(0)
(0)
X ∂E (0)
∂E
=
σ 0 xi
+ oi − ti σ 00 xi
wji
wji
=
w
(50)
(1)
(0) ji
i
∂oj
i ∂xi
(59)
2
X (0) 2 (0)
(0)
(0)
σ 0 xi
+ oi − ti σ 00 xi
=
wji
Note that this equation does not depend on the specific form
i |
{z
}
of ∂E
(0) , whether it involves a sigmoid or any other activafrom (27)
∂xi
(60)
tion function. We can therefore replace the specific indexes
with general ones, and use this equation in the future.
Summing up, we obtain the more general expression:
∂E
(m+1)
=
∂oj
X ∂E
i
(m)
w
(m) ji
∂xi
∂2E
(51)
(1) 2
=
∂oj
The second derivative of the error with respect to a neuron
(1)
with output oj in the first hidden layer:
(1) 2
=
∂oj
=
∂
(1)
(1)
∂oj
=
∂
(1)
∂oj
(52)
(1)
∂oj ∂oj
∂ X ∂E
(0)
w
(0) ji
i ∂xi
X (0)
(0)
(0)
oi − ti σ 0 xi
wji
(53)
(54)
i
(0)
(0)
If we now make use of the fact, that oi = σ xi
=
P
(0) (1)
σ
, we can evaluate the expression furj wji oj
ther.
(0)
wji
2
(61)
∂xi
i
∂xx
with general ones:
(m+1) 2
∂E
(0) 2
Note that the equation in (61) does not depend on the form
2
E
of ∂ (0)
2 , which means we can replace the specific indexes
∂2E
∂2E
X ∂2E
=
∂oj
X ∂2E
i
(m) 2
(m)
wji
2
(62)
∂xi
At this point we are beginning to see the recursion in the
form of the 2nd derivative terms which can be thought of
analogously to the first derivative recursion which is central
to the back-propagation algorithm. The formulation above
which makes specific reference to layer indexes also works
in the general case.
(m)
Consider the ith neuron in any layer m with output oi
(m)
and input xi . The first and second derivatives of the error
E with respect to this neuron’s input are:
∂E
(m)
∂xi
(m)
=
∂E ∂oi
(m)
∂oi
(m)
∂xi
(63)
The Incredible Shrinking Neural Network
=
∂xi
=
=
(m) 2
∂xi
∂E
(m)
∂xi
(64)
(m)
(m)
∂
∂E ∂oi
(m)
∂xi
A.1.3. S UMMARY O F H IDDEN L AYER D ERIVATIVES
∂xi
!
(m)
∂xi
(m)
∂2E
∂oi
+
(m)
(m)
(m)
∂xi ∂oi ∂xi
∂
∂E
∂E
(65)
(m)
∂oi
(m+1)
(m)
∂E ∂ 2 oi
2
(m)
∂oi ∂x(m)
i
!
(m)
(m)
∂E ∂oi
∂oi
(m)
(m)
(m)
∂oi ∂xi
∂xi
=
∂oj
X ∂E
∂2E
(m)
w
(m) ji
∂xi
i
(m+1) 2
=
∂oj
X ∂2E
(m) 2
2
∂xi
i
(73)
(66)
=
+
∂E
Note the form of this equation is the general form of what
was derived for the output layer in (27). Both of the above
first and second terms are easily computable and can be
stored as we propagate back from the output of the network
to the input. With respect to the output layer, the first and
second derivative terms have already been derived above.
In the case of the m + 1 hidden layer during back propagation, there is a summation of terms calculated in the mth
layer. For the first derivative, we have this from (51).
∂E
(m+1)
∂oj
=
X ∂E
i
(m)
∂xi
(m)
wji
(70)
4500
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
4000
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
4000
3500
3000
2000
3000
2500
2000
1500
1000
1000
500
0
0
5
10
Number of Neurons Removed
15
20
0
0
20
40
60
Number of Neurons Removed
80
100
Figure 7. Degradation in squared error after pruning a two-layer
network trained to compute the cosine function (Left Network:
2 layers, 10 neurons each, 1 output, logistic sigmoid activation,
starting test accuracy: 0.9999993, Right Network: 2 layers, 50
neurons each, 1 output, logistic sigmoid activation, starting test
accuracy: 0.9999996)
And the second derivative for the jth neuron in the m + 1
layer:
B.2. Results on MNIST Dataset
(m+1)
00
For all the results presented in this section, the MNIST
=
σ
x
j
(m+1)
(m+1) 2
(m+1) 2
database of Handwritten Digits by (LeCun & Cortes, 2010)
∂oj
∂xj
∂oj
was used. It is worth noting that due to the time taken by the
(71)
brute force algorithm we rather used a 5000 image subset
of the MNIST database in which we have normalized the
We can replace both derivative terms with the forms which
pixel values between 0 and 1.0, and compressed the image
depend on the previous layer:
sizes to 20x20 images rather than 28x28, so the starting test
X ∂ 2 E (0) 2 (m+1) 2 X ∂Eaccuracy
∂2E
here appears higher than those reported
(m) reported
(m+1)
0
00
=
w
σ
x
+
w
σ
x
ji
j
ji
(m)
by LeCun
et al.j We do not believe that this affects the inter(m+1) 2
(0) 2
∂xj
i ∂xi
i ∂xi
|
{zpretation
} of the presented results because the basic learning
|
{z
}
from problem
(51)
from (62)
does not change with a larger dataset or input di(72)
mension.
∂2E
(m)
wji
(m)
(m)
00 ∂E
∂E ∂oi
x
σ
i
(m)
(m)
(m)
=
(74)
(m)
(m)
(m)
∂oi
∂xi
∂oi
∂xi
∂oi ∂xi
(67)
∂ 2 E 0 (m+1) 2
∂E
∂2E
!
(m+1)
00
(m)
(m)
x
x
=
σ
+
σ
2
j
j
(m+1)
∂ E
∂E
∂oi ∂oi
(m+1) 2
(m+1) 2
(m)
∂oj
∂xj
∂oj
=
+ (m) σ 00 xi
(m)
(m)
(m) 2
∂xi ∂xi
∂oi
∂oi
(75)
(68)
∂E
∂ 2 E 0 (m) 2
(m)
B. Additional Experimental Results
σ xi
+ (m) σ 00 xi
=
2
(m)
∂oi
∂oi
B.1. Example Regression Problem
(69)
=
∂2E
∂
Sum of Squared Errors
(m) 2
Sum of Squared Errors
∂2E
∂2E
2
(m+1)
σ 0 xj
+
∂E
And this horrible mouthful of an equation gives you a general form for any neuron in the jth position of the m + 1
layer. Taking very careful note of the indexes, this can be
more or less translated painlessly to code. You are welcome, world.
B.3. Pruning a 1-Layer Network
The network architecture in this case consisted of 1 layer,
100 neurons, 10 outputs, logistic sigmoid activations, and
a starting test accuracy of 0.998.
The Incredible Shrinking Neural Network
B.3.1. S INGLE OVERALL R ANKING A LGORITHM
We first present the results for a single-layer neural network
in Figure 8, using the Single Overall algorithm (Algorithm
1) as proposed in Section 2.
8000
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
0.8
6000
Classification Accuracy
5000
4000
3000
2000
0.6
0.4
0.2
1000
0
20
40
60
Percentage of Neurons Removed
80
0.0
100
0
20
40
60
Percentage of Neurons Removed
80
100
Figure 8. Degradation in squared error (left) and classification accuracy (right) after pruning a single-layer network using The Single Overall Ranking algorithm (Network: 1 layer, 100 neurons,
10 outputs, logistic sigmoid activation, starting test accuracy:
0.998)
B.3.2. I TERATIVE R E -R ANKING A LGORITHM
8000
Classification Accuracy
Sum of Squared Errors
6000
5000
4000
3000
2000
1.0
B.3.4. V ISUALIZATION OF BRUTE FORCE P RUNING
D ECISIONS
0.8
0.6
0.4
0.2
1000
0
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.2
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
Remember that lower is better in terms of the height of the
curve and minimal (or negative) horizontal change between
the vertical red line at 1.0 (neuron on, α = 1.0) and 0.0
(neuron off, α = 0.0) is indicative of a good candidate
neuron to prune, i.e. there will be minimal effect on the
network output when the neuron is removed.
0
20
40
60
Percentage of Neurons Removed
80
100
0.0
0
20
40
60
Percentage of Neurons Removed
80
104
100
Figure 9. Degradation in squared error (left) and classification accuracy (right) after pruning a single-layer network the iterative
re-ranking algorithm (Network: 1 layer, 100 neurons, 10 outputs,
logistic sigmoid activation, starting test accuracy: 0.998)
B.3.3. V ISUALIZATION OF E RROR S URFACE &
P RUNING D ECISIONS
These graphs are a visualization of the error surface of the
network output with respect to the neurons chosen for removal using each of the 3 ranking criteria, represented in
intervals of 10 neurons. In each graph, the error surface
of the network output is displayed in log space (left) and
in real space (right) with respect to each candidate neuron
chosen for removal. We create these plots during the pruning exercise by picking a neuron to switch off, and then
multiplying its output by a scalar gain value α which is adjusted from 0.0 to 10.0 with a step size of 0.001. When
the value of α is 1.0, this represents the unperturbed neuron output learned during training. Between 0.0 and 1.0,
we are graphing the literal effect of turning the neuron off
(α = 0), and when α > 1.0 we are simulating a boosting
of the neuron’s influence in the network, i.e. inflating the
value of its outgoing weight parameters.
We graph the effect of boosting the neuron’s output to
demonstrate that for certain neurons in the network, even
5000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
103
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
4000
Sum of Squared Errors
0
Sum of Squared Errors
Sum of Squared Errors
1.0
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
doubling, tripling, or quadrupling the scalar output of the
neuron has no effect on the overall error of the network,
indicating the remarkable degree to which the network has
learned to ignore the value of certain parameters. In other
cases, we can get a sense of the sensitivity of the network’s
output to the value of a given neuron when the curve rises
steeply after the red 1.0 line. This indicates that the learned
value of the parameters emanating from a given neuron are
relatively important, and this is why we should ideally see
sharper upticks in the curves for the later-removed neurons in the network, that is, when the neurons crucial to
the learning representation start to be picked off. Some
very interesting observations can be made in each of these
graphs.
102
3000
2000
1000
101
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
0
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
Figure 10. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the brute force criterion; (Network: 1 layer,
100 neurons, 10 outputs, logistic sigmoid activation, starting test
accuracy: 0.998)
In Figure ??, we notice how low to the floor and flat most of
the curves are. It’s only until the 90th removed neuron that
we see a higher curve with a more convex shape (clearly a
more sensitive, influential piece of the network).
B.3.5. V ISUALIZATION OF 1 ST O RDER
A PPROXIMATION P RUNING D ECISIONS
It can be seen in Figure 11 that most choices seem to have
flat or negatively sloped curves, indicating that the first order approximation seems to be pretty good, but examining
the brute force choices shows they could be better.
The Incredible Shrinking Neural Network
7000
5000
4000
3000
6000
5000
4000
3000
2000
0
1000
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
0
10
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
103
7000
102
6000
5000
4
6
Gain value (1 = normal output, 0 = off)
8
10
20
40
60
Percentage of Neurons Removed
80
0.0
100
8000
3000
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
20
40
60
Percentage of Neurons Removed
10
Figure 12. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the 2nd order Taylor Series error approximation criterion; (Network: 1 layer, 100 neurons, 10 outputs,
logistic sigmoid activation, starting test accuracy: 0.998)
B.3.6. V ISUALIZATION OF 2 ND O RDER
A PPROXIMATION P RUNING D ECISIONS
The method in Figure 12 looks similar to the brute force
method choices, though clearly not as good (they’re more
spread out). Notice the difference in convexity between the
2nd and 1st order method choices. It’s clear that the first
order method is fitting a line and the 2nd order method is
fitting a parabola in their approximation.
B.4. Pruning A 2-Layer Network
The network architecture in this case consisted of 2 layers,
50 neurons per layer, 10 outputs, logistic sigmoid activation, and a starting test accuracy of 1.000.
B.4.1. S INGLE OVERALL R ANKING A LGORITHM
Figure 13 shows the pruning results for Algorithm 1 on a
2-layer network. The ranking procedure is identical to the
one used to generate Figure 8.
Unsurprisingly, a 2-layer network is harder to prune because a single overall ranking will never capture the interdependencies between neurons in different layers. It makes
sense that this is worse than the performance on the 1-layer
1.2
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
0
0
80
100
B.4.2. ITERATIVE RE - RANKING A LGORITHM
Sum of Squared Errors
2
0
4000
1000
0
0.2
Figure 13. Degradation in squared error (left) and classification
accuracy (right) after pruning a 2-layer network using the Single Overall Ranking algorithm; (Network: 2 layers, 50 neurons/layer, 10 outputs, logistic sigmoid activation, starting test
accuracy: 1.000)
2000
101
0.4
network, even if this method is already known to be bad,
and we’d likely never use it in practice.
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
8000
Sum of Squared Errors
Sum of Squared Errors
9000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
0.6
10
Figure 11. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the 1st order Taylor Series error approximation criterion; (Network: 1 layer, 100 neurons, 10 outputs,
logistic sigmoid activation, starting test accuracy: 0.998)
104
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
0.8
1000
2000
101
1.0
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
7000
6000
5000
4000
3000
2000
0.8
0.6
0.4
0.2
1000
0
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.0
Classification Accuracy
102
6000
8000
Sum of Squared Errors
103
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
8000
Classification Accuracy
9000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
Sum of Squared Errors
Sum of Squared Errors
104
0
20
40
60
Percentage of Neurons Removed
80
100
0.0
0
20
40
60
Percentage of Neurons Removed
80
100
Figure 14. Degradation in squared error (left) and classification
accuracy (right) after pruning a 2-layer network using the iterative
re-ranking algorithm; (Network: 2 layers, 50 neurons/layer, 10
outputs, logistic sigmoid activation, starting test accuracy: 1.000)
Figure 14 shows the results from using Algorithm 2 on a 2layer network. We compute the same brute force rankings
and Taylor series approximations of error deltas over the
remaining active neurons in the network after each pruning
decision used to generate Figure 9. Again, this is intended
to account for the effects of canceling interactions between
neurons.
B.4.3. V ISUALIZATION OF E RROR S URFACE &
P RUNING D ECISIONS
As seen in the case of a single layered network, these
graphs are a visualization the error surface of the network
output with respect to the neurons chosen for removal using
each algorithm, represented in intervals of 10 neurons.
B.4.4. V ISUALIZATION OF BRUTE FORCE P RUNING
D ECISIONS
In Figure 15, it is clear why these neurons got chosen, their
graphs clearly show little change when neuron is removed,
are mostly near the floor, and show convex behaviour of
error surface, which argues for the rationalization of using
2nd order methods to estimate difference in error when they
The Incredible Shrinking Neural Network
2500
102
2000
Sum of Squared Errors
103
104
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
3000
1500
1000
7000
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
103
5000
102
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
0
10
4000
3000
2000
500
101
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
6000
Sum of Squared Errors
3500
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
Sum of Squared Errors
Sum of Squared Errors
104
1000
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
101
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
0
10
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
Figure 15. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the brute force criterion; (Network: 2
layers, 50 neurons/layer, 10 outputs, logistic sigmoid activation,
starting test accuracy: 1.000)
Figure 17. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the 2nd order Taylor Series error approximation criterion; (Network: 2 layers, 50 neurons/layer, 10 outputs, logistic sigmoid activation, starting test accuracy: 1.000)
are turned off.
first layer. What is interesting to note is that from some
other experiments conducted on toy datasets (predicting
whether a given point would lie inside a given shape on the
Cartesian plane), the performance of the 2nd-order method
was found to be exceptionally good and produced results
very close to the brute force method. The 1st-order method,
as expected, performed poorly here as well. Some of these
results are illustrated in Figure 18.
B.4.5. V ISUALIZATION OF 1 ST O RDER
A PPROXIMATION P RUNING D ECISIONS
7000
103
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
6000
5000
102
4000
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
0
0
2
4
6
Gain value (1 = normal output, 0 = off)
8
10
Sum of Squared Errors
2000
1000
101
1.2
20000
3000
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
15000
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.1
Classification Accuracy
Neuron: 0
Neuron: 10
Neuron: 20
Neuron: 30
Neuron: 40
Neuron: 50
Neuron: 60
Neuron: 70
Neuron: 80
Neuron: 90
Sum of Squared Errors
Sum of Squared Errors
104
10000
5000
1.0
0.9
0.8
0.7
0.6
0.5
Drawing a flat line at the point of each neurons intersection
with the red vertical line (no change in gain) shows that the
1st derivative method is actually accurate for estimation of
change in error in these cases, but still ultimately leads to
poor decisions.
B.4.6. V ISUALIZATION OF 2 ND O RDER
A PPROXIMATION P RUNING D ECISIONS
Clearly these neurons are not overtly poor candidates for
removal (error doesn’t change much between 1.0 & zerocrossing left-hand-side), but could be better (as described
above in the brute force Criterion discussion).
B.5. Experiments on Toy Datasets
As can be seen from the experiments on MNIST, even
though the 2nd-order approximation criterion is consistently better than 1st-order, its performance is not nearly as
good as brute force based ranking, especially beyond the
0
0
20
40
60
Number of Neurons Removed
30000
80
0
20
1.1
40
60
Number of Neurons Removed
Classification Accuracy
15000
10000
80
100
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
1.0
20000
0.9
0.8
0.7
0.6
0.5
0.4
5000
0
0.4
100
Brute Force
1st Order / Skeletonization
2nd Order Taylor Approx
25000
Sum of Squared Errors
Figure 16. Error surface of the network output in log space (left)
and real space (right) with respect to each candidate neuron chosen for removal using the 1st order Taylor Series error approximation criterion; (Network: 2 layers, 50 neurons/layer, 10 outputs,
logistic sigmoid activation, starting test accuracy: 1.000)
0.3
0
20
40
60
Number of Neurons Removed
80
100
0
20
40
60
Number of Neurons Removed
80
100
Figure 18. Degradation in squared error (left) and classification
accuracy (right) after pruning a 2-layer network using the iterative
re-ranking algorithm on a toy “diamond” shape dataset (top) and
a toy “random shape” dataset (below); (Network: 2 layers, 50
neurons/layer, 10 outputs, logistic sigmoid activation, starting test
accuracy: 0.992(diamond); 0.986(random shape)
| 9cs.NE
|
NEURAL IDEAL PRESERVING HOMOMORPHISMS
arXiv:1612.06150v1 [math.AC] 19 Dec 2016
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
Abstract. The neural ideal of a binary code C ⊆ Fn2 is an ideal in F2 [x1 , . . . , xn ] closely
related to the vanishing ideal of C. The neural ideal, first introduced by Curto et al, provides an algebraic way to extract geometric properties of realizations of binary codes. In
this paper we investigate homomorphisms between polynomial rings F2 [x1 , . . . , xn ] which
preserve all neural ideals. We show that all such homomorphisms can be decomposed into
a composition of three basic types of maps. Using this decomposition, we can interpret
how these homomorphisms act on the underlying binary codes. We can also determine
their effect on geometric realizations of these codes using sets in Rd . We also describe how
these homomorphisms affect a canonical generating set for neural ideals, yielding an efficient
method for computing these generators in some cases.
1. Introduction
Given a binary code C ⊆ Fn2 , a realization of C is a collection of sets U = {U1 , . . . , Un } in
Rd such that C records intersection patterns of the sets in U. More precisely, an element c is
in C exactly when there is a point in Rd lying in only the Ui ’s with ci = 1. Every code has a
realization in Rd for d ≥ 1 if we place no restrictions on the sets in U. However, if we require
that the sets in U satisfy certain geometric properties such as connectedness or convexity
not every code has a realization. Such restrictions naturally arise in neuroscientific contexts,
as addressed in [3].
The question of when a code is realizable is difficult to answer in these restricted settings; however, algebraic techniques have been succesfully used to gain insight into obstructions to codes being realizable. Given a code C, the authors in [3] define the neural ideal
JC ⊆ F2 [x1 , . . . , xn ] and the associated neural ring, which algebraically process and store the
information in a neural code. Each neural ideal JC has a particular set of generators, its
canonical form, which reveals the fundamental relationships between sets in any realization
of C.
In recent work [4], the authors use the ideal-variety correspondence to relate homomorphisms between neural rings to maps between their associated codes, and focus particularly
on those maps between codes which have a natural interpretation in neuroscience. We examine ring homomorphisms from F2 [x1 , ..., xn ] to F2 [x1 , ..., xm ] which preserve neural ideals
in the sense that the image of any neural ideal is again a neural ideal. In Theorem 2.10
we completely classify these homomorphisms, and give a method of decomposing any such
homomorphism into a composition of three basic types. Theorem 2.13 gives a precise description of how these homomorphisms act on the underlying codes of neural ideals. Theorem
2.15 describes how these maps transform realizations of the underlying codes.
The structure of the paper is as follows. In Section 2, we provide pertinent definitions
and describe in detail the main results of our work. Sections 3 and 4 are devoted to proving
Date: December 20, 2016.
2010 Mathematics Subject Classification. 05C15, 05C30, 05C31.
1
2
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
Theorems 2.10 and 2.13 respectively. In Section 5 we prove Theorem 2.15. In Section 6 we
build on Theorem 2.10 by describing how homomorphisms preserving neural ideals affect the
canonical form of neural ideals.
2. Preliminaries and Main Results
We start by presenting the fundamental definitions for the theory we will be exploring,
beginning with the formal definition of a neural code.
Definition 2.1. A code or neural code is a set C ⊆ Fn2 of binary vectors. The vectors in C
are called codewords.
Throughout we will let [n] denote the set {1, 2, . . . , n}. For any vector v ∈ Fn2 the support
of v is the set supp(v) := {i ∈ [n] | vi = 1}.
Definition 2.2. Let U = {U1 , . . . , Un } be a collection of sets in a space X ⊆ Rd . The code
of U is the code defined by
!
\
[
Ui
Uj 6= ∅ ,
C(U) := v ∈ Fn2
\
vi =1
vj =0
where we adopt the convention that the empty intersection is X and the empty union is ∅.
Given C ⊂ Fn2 , we say C is realizable if there exists a collection U such that C = C(U), and
call U a realization of the code in the space X. If a code C has a realization consisting of
convex open sets then we say C is a convex code.
Classifying which codes are convex is an open problem which has been considered by
many researchers [1, 3, 7] . Many partial results exist along with tools for approaching this
task. In this paper we build on algebraic tools introduced in [3] that have been pivotal in
understanding obstructions to convex realizability of codes; see for example [1, 2, 6]. As a
code C is a subset of Fn2 , we will work over the polynomial ring F2 [x1 , . . . , xn ]; for convenience
we will use F2 [n] to denote F2 [x1 , . . . , xn ].
Definition 2.3 ([3]). A pseudomonomial is a polynomial f ∈ F2 [n] of the form
Y Y
f=
xi (1 − xj ).
i∈σ
j∈τ
where σ, τ ⊆ [n] and σ ∩ τ = ∅.
Note that every pseudomonomial has degree at most n. For any vector v ∈ Fn2 , its
indicator pseudomonomial, denoted ρv , is the degree n pseudomonomial with σ = supp(v)
and τ = [n] \ supp(v). Using indicator pseudomonomials, the authors in [3] construct a
unique ideal associated to a binary code, called the neural ideal.
Definition 2.4 ([3]). Let C ⊆ Fn2 be a neural code. The neural ideal of C, denoted JC , is
the ideal
JC := hρv | v ∈
/ Ci.
Here we adopt the convention that the ideal generated by the empty set is the zero ideal.
An ideal in F2 [n] is called a neural ideal if it is equal to JC for some code C.
NEURAL IDEAL PRESERVING HOMOMORPHISMS
3
An ideal in F2 [n] is a neural ideal if and only if it has a generating set consisting of pseudomonomials [5]. This characterization of neural ideals will be useful in proving that certain
homomorphisms preserve neural ideals, and in classifying all such homomorphisms. We now
formally introduce our objects of interest: the class of homomorphisms which preserve neural
ideals.
Definition 2.5. Let φ : F2 [n] → F2 [m] be a homomorphism of rings. We say that φ preserves
neural ideals if the image of any neural ideal in F2 [n] under φ is a neural ideal in F2 [m].
That is, for every code C ⊆ Fn2 there is a code D ⊆ Fm
2 so that
φ(JC ) = JD .
Because neural ideals in F2 [n] and codes on n bits are in bijective correspondence we can
think of such a map as defining a process for transforming codes on n bits to codes on m
bits. Note immediately that composing any two maps which preserve neural ideals yields a
map which again preserves neural ideals.
Example 2.6. Consider the surjective map φ : F2 [6] → F2 [3] defined by
x1 7→ 0
x2 7→ 1 − x3
x3 7→ 1
x4 →
7 1 − x1
x5 →
7 1
x6 7→ x2
extended algebraically to all of F2 [6]. Let J be the neural ideal generated by {x1 x2 (1 −
x3 ), x4 (1 − x2 ), x5 x6 (1 − x1 )}. If we apply φ to this set of generators then we obtain a set of
generators for the ideal φ(J):
φ(J) = h0, (1 − x1 )x3 , x2 i = hx3 (1 − x1 ), x2 i.
Since φ(J) is generated by pseudomonomials it is a neural ideal. We will see shortly as a
result of Theorem 2.10 that the image of any neural ideal under φ is again a neural ideal.
The following definitions give three useful classes of maps that preserve neural ideals. For
the proof that these maps preserve neural ideals, see Lemma 3.2.
Definition 2.7. Let λ be a permutation of [n]. Then the map induced by sending xi to xλ(i)
is called a permutation map of F2 [n] and is denoted simply by λ.
Definition 2.8. For any i ∈ [n] the i-th bit flip is the map δi : F2 [n] → F2 [n] induced by
(
xj
j 6= i
δi (xj ) =
1 − xj j = i.
We will call any composition of such maps a bit flipping map.
Definition 2.9. Let 1 ≤ m ≤ m′ ≤ n and define ωm,m′ : F2 [n] → F2 [m] to be the map
induced by
xi i ≤ m
ωm,m′ (xi ) = 0 m < i ≤ m′
1 m′ < i ≤ n
Such a map is called a restriction map from F2 [n] to F2 [m].
4
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
Having described these three classes of neural ideal preserving homomorphisms we can
state our first main result, which tells us that every neural ideal preserving homomorphism
can be expressed as a composition of these maps.
Theorem 2.10. Let φ : F2 [n] → F2 [m] be a homomorphism. Then φ preserves neural ideals
if and only if φ can be expressed as a composition of the following types of maps:
(i) Bit flipping (Definition 2.8),
(ii) Permutation (Definition 2.7), and
(iii) Restriction (Definition 2.9).
Moreover, we can write φ = ω ◦ λ ◦ δ where δ is a bit flipping map, λ is a permutation, and
ω is a restriction.
The proof of Theorem 2.10 can be found in Section 3. Theorem 2.10 provides a complete
characterization of homomorphisms that preserve neural ideals. By decomposing these homomorphisms into three basic types we gain an understanding of their structure and obtain
a compact way to describe any such homomorphism.
Example 2.11. Recall the map φ described in Example 2.6. We can illustrate this map
and its decomposition given by Theorem 2.10 by considering the action of φ on variables.
This illustration is provided in Figure 1, where solid line segments connect a variable to its
image (for example x1 7→ 0 or x6 7→ x2 ) while dashed lines indicate mappings of the form
xi 7→ 1 − xj .
Figure 1. The homomorphism φ written as the composition of permutation,
bit flipping, and restriction maps. Dashed lines indicate places where xi 7→
1 − xj .
We now describe three transformations of codes, which we will see are related naturally
to the three types of homomorphisms given in Definitions 2.7, 2.8, and 2.9.
Definition 2.12. Let C be a code on n bits.
(i) Let λ be a permutation of [n]. We define the permutation of C by λ to be the code
λ(C) := {u ∈ Fn2 | supp(u) = λ(supp(c)) for some c ∈ C}.
NEURAL IDEAL PRESERVING HOMOMORPHISMS
5
(ii) For any i ∈ [n] the i-th bit flip of C, denoted δi (C), is the code on n bits defined by
δi (C) := {u ∈ Fn2 | supp(u) = supp(c) ⊕ {i} for some c ∈ C}.
where ⊕ denotes symmetric difference.
(iii) Let m and m′ be integers so that 1 ≤ m ≤ m′ ≤ n, and let σ = [n] \ [m′ ]. The
restriction of C to (m, m′ ) is the code
rest(C, m, m′ ) := {u ∈ Fm
2 | supp(u) ∪ σ = supp(c) for some c ∈ C}.
Restricting codes generalizes the notion of taking the link of a face in a simplicial complex.
In particular, when m = m′ and the supports of codewords in C forms an abstract simplicial
complex ∆, then the supports of vectors in rest(C, m, m′ ) form the simplicial complex Lkσ (∆)
where σ = [n] \ [m]. This connection is important since taking links in the simplicial complex
of a code can be used to understand the convexity of that code, as described in [2].
The following theorem allows us to translate precisely between homomorphisms preserving
neural ideals and transformations of neural codes.
Theorem 2.13. Let C be a neural code. Then
(i) λ(JC ) = Jλ(C) ,
(ii) δi (JC ) = Jδi (C) , and
(iii) ωm,m′ (JC ) = Jrest(C,m,m′ ) .
The proof of Theorem 2.13 can be found in Section 4. Finally, we give a geometric
interpretation of the behavior of permutation, bit flipping, and restriction maps. We first
require a definition.
Definition 2.14. Let U = {U1 , . . . , Un } be a collection of sets and let 1 ≤ m ≤ m′ ≤ n.
The compatible region of (m, m′ ) in U is the set
!
!
[
\
Uj .
Ui \
m′ +1≤i≤n
m+1≤j≤m′
Theorem 2.15. Let C be a code with a realization U = {U1 , . . . , Un } in a space X.
(i) Let λ be a permutation of [n]. Then the set {Uλ(1) , . . . , Uλ(n) } is a realization of λ(C)
in X.
(ii) For any i the collection of sets {U1 , . . . , (X \ Ui ), . . . , Un } is a realization of δi (C) in
X.
(iii) Let m and m′ be integers so that 1 ≤ m ≤ m′ ≤ n and let X ′ be the compatible region
of (m, m′ ) in U. The collection {Ui ∩ X ′ | i ∈ [m]} is a realization of rest(C, m, m′ )
in the space X ′ .
The proof of Theorem 2.15 is given in Section 5. Together with Theorem 2.10, Theorem
2.15 allows us to describe any homomorphism preserving neural ideals in terms of operations
on the realizations of codes.
We have completely classified all homomorphisms preserving neural ideals and described
how they affect the underlying codes. We have also described how these homomorphisms act
geometrically on realizations of the associated codes. The correspondence we have described
6
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
can be summarized as follows.
Permutation maps
λ(xi ) = xλ(i)
Permuting bits in
←→ codewords: C 7→ λ(C)
Permuting labels on
←→ realization: U 7→ λ(U)
Bit flipping
( maps
1 − xj j = i;
δi (xj ) =
xj
i=
6 j
Flipping i-th bit in
←→ all codewords of C:
C 7→ δi (C)
Replacing Ui with
←→ its complement
X \ Ui
Restriction maps
ωm,m′ : F2 [n] → F2 [m]
Intersecting all sets in
Restricting to compatrealization with the
←→
←→ ible codewords:
compatible region for
C 7→ rest(C, m, m′ )
(m, m′ )
The canonical form, an object introduced in [3], provides a concise presentation of JC and
yields information about the underlying code and its realizations. In general it is nontrivial
to compute the canonical form of a neural ideal. However, homomorphisms that preserve
neural ideals provide a computational shortcut for translating between the related canonical
forms.
Definition 2.16. Let JC be a neural ideal. The canonical form of JC is the collection of
pseudomonomials in JC which are minimal with respect to division, and is denoted CF (JC ).
For any homomorphism φ preserving neural ideals, the canonical forms CF(JC ) and
CF (φ(JC )) are related as follows.
Theorem 2.17. Let φ : F2 [n] → F2 [m] be a homomorphism preserving neural ideals. Then
CF (φ(JC )) ⊆ φ(CF(JC )) for any neural ideal JC , with equality when φ is a composition of a
permutation and bit flipping map.
This result suggests that we can obtain CF (φ(JC )) from CF (JC ). In fact we can: if we
apply φ to every pseudomonomial in CF(JC ) and select from the result everything which is
minimal with respect to division, the resulting set will be precisely CF (φ(JC )). This result
is proven in Section 6.
3. Classifying homomorphisms that preserve neural ideals
In order to establish Theorem 2.10 we first show that permutation, bit flipping, and
restriction all preserve neural ideals. To this end, we give a preliminary characterization of
homomorphisms that preserve neural ideals.
Lemma 3.1. Let φ : F2 [n] → F2 [m] be a homomorphism. Then φ preserves neural ideals
if and only if φ is surjective and for any pseudomonomial f ∈ F2 [n], φ(f ) is either a pseudomonomial or zero.
Proof. Let φ be a homomorphism which preserves neural ideals. To see that φ is surjective,
note that F2 [n] is a neural ideal containing 1, so its image must also be a neural ideal containing 1. Now, let f ∈ F2 [n] be a pseudomonomial and recall that hf i is a neural ideal,
and φ(hf i) = hφ(f )i. Suppose that φ(f ) is not a pseudomonomial. Every factor of a pseudomonomial is again a pseudomonomial, so hφ(f )i does not contain any pseudomonomials.
The only neural ideal containing no pseudomonomials is the zero ideal and so φ(f ) = 0.
NEURAL IDEAL PRESERVING HOMOMORPHISMS
7
Conversely, let JC be a neural ideal in F2 [n]. Note that since φ is surjective φ(JC ) is an
ideal of F2 [m]. Recall that JC = hρv | v ∈
/ Ci, and so we will have that
φ(JC ) = hφ(ρv ) | v ∈
/ Ci.
Removing the φ(ρv ) which are zero from the set of generators above we obtain a generating
set for φ(JC ) consisting only of pseudomonomials. This implies that φ(JC ) is a neural ideal
and the result follows.
Lemma 3.2. Permutation maps, bit flipping maps, and restriction maps all preserve neural
ideals.
Proof. Note that permutation maps and bit flipping maps are both automorphisms of F2 [n],
and so are surjective. The image of a pseudomonomial under a permutation map is again
a pseudomonomial since we are simply permuting the variables. Under a bit flipping map
δi any occurrence of xi in a pseudomonomial will be changed to (1 − xi ) and vice versa,
while the remaining factors are unchanged, so the image of any pseudomonomial is still a
pseudomonomial. Thus permutation maps and bit flipping maps satisfy the conditions of
Lemma 3.1 and must preserve neural ideals.
Let ωm,m′ : F2 [n] → F2 [m] be a restriction map and f ∈ F2 [n] a pseudomonomial. Note
that ωm,m′ (xi ) ∈ {0, 1, xi } for all i, and ωm,m′ (1−xj ) ∈ {0, 1, 1−xj } for all j. Hence applying
ωm,m′ to f yields either zero (if one of the factors vanishes) or another pseudomonomial (when
all factors map to either 1 or themselves). Furthermore ωm,m′ is surjective since it acts as
the identity on F2 [m] as a subring of F2 [n]. Thus by Lemma 3.1 the map ωm,m′ preserves
neural ideals.
Next we describe how a homomorphism that preserves neural ideals affects variables in
F2 [n]. This will be critical in decomposing homomorphisms that preserve neural ideals into
a composition of the three basic homomorphisms.
Lemma 3.3. Let φ : F2 [n] → F2 [m] be a map preserving neural ideals. Then
(i) for each j ∈ [m] there exists a unique i ∈ [n] so that φ(xi ) ∈ {xj , 1 − xj }, and
(ii) for all i ∈ [n] we have φ(xi ) ∈ {0, 1, xj , 1 − xj }.
Proof. The set {xi | i ∈ [n]} generates F2 [n] as an algebra, and since φ is surjective the set
{φ(xi ) | i ∈ [n]} must generate F2 [m]. This implies that the set
σ = {i ∈ [n] | φ(xi ) is not constant}
has at least m elements. Consequently the pseudomonomial
!
Y
Y
φ
xi =
φ(xi )
i∈σ
i∈σ
has degree at least m. But no pseudomonomial in F2 [m] can have degree larger than m
and so this pseudomonomial has exactly m linear factors. Each φ(xi ) must be exactly one
of these factors, and so for each j ∈ [m] we see that there is a unique i ∈ [n] for which
φ(xi ) ∈ {xj , 1 − xj }. The remainder of the xi ’s map to constants, but the only constants are
0 and 1 and so the second statement of the lemma follows.
With Lemma 3.3 in hand we prove Theorem 2.10.
8
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
Proof of Theorem 2.10. If φ is a composition of restriction, permutation, and bit flipping
maps then by Lemma 3.2 it preserves neural ideals. To prove the converse, let φ : F2 [n] →
F2 [m] be a homomorphism that preserves neural ideals. We will decompose φ directly as the
composition of a permutation, bit flipping map, and restriction map. First define four sets
of indices:
α0 = {i ∈ [n] | φ(xi ) = 0},
β0 = {i ∈ [n] | φ(xi ) = xj },
α1 = {i ∈ [n] | φ(xi ) = 1},
β1 = {i ∈ [n] | φ(xi ) = 1 − xj }.
Note by Lemma 3.3 that these four sets completely partition [n]. Now define a permutation,
bit flipping map, and restriction map as follows:
• Let δ : F2 [n] → F2 [n] be the bit flipping map which replaces xi by 1 − xi for all i ∈ β1 .
• Let λ to be a permutation of n which maps α0 to the set of indices between m + 1
and m + |α0 | inclusive, maps α1 to the set of indices between m + |α0 | + 1 and n
inclusive, and maps i to j where φ(xi ) ∈ {xj , 1 − xj } for all other indices.
• Let ω : F2 [n] → F2 [m] be the restriction map which sends xi to zero whenever
m + 1 ≤ i ≤ m + |α0 | and to one whenever m + |α0 | + 1 ≤ i ≤ n.
We claim that φ = ω ◦ λ ◦ δ. To prove this it suffices to show that the image of all variables
in F2 [n] under ω ◦ λ ◦ δ is the same as under φ. We consider four cases.
• If i ∈ α0 then the image of xi is 0 since δ does not affect xi , while by construction λ
maps xi to an index which is sent to 0 by ω.
• If i ∈ α1 then the image of xi is 1. Again δ does not affect xi , and λ maps xi to a
variable mapped to 1 by ω.
• If i ∈ β0 then δ does not affect it. The map λ then sends xi to φ(xi ) = xj by
construction, and ω leaves the result unchanged. Thus the image of xi is exactly
φ(xi ) as desired.
• Finally, if i ∈ β1 then δ(xi ) = 1 − xi . Applying λ, we obtain 1 − xj which by
construction is φ(xi ). Applying ω has no affect on this quantity and so again the
image of xi is exactly φ(xi ).
Since the image under ω ◦ λ ◦ δ of each variable is the same as its image under φ we conclude
that φ = ω ◦ λ ◦ δ, and the result follows.
Corollary 3.4. Suppose that φ : F2 [n] → F2 [m] is a map preserving neural ideals. Then for
every neural ideal JD in F2 [m] there exists a neural ideal JC in F2 [n] so that φ(JC ) = JD .
Proof. We first show that every pseudomonomial f has a preimage fˆ under φ such that
fˆ is a pseudomonomial. By Theorem 2.10 it suffices to prove this for bit flipping maps,
permutations, and restrictions. For bit flipping maps and permutations this is clear since
they are automorphisms whose inverses also preserve neural ideals. For restrictions, the
pseudomonomial itself serves as its own preimage.
Now let {f1 , . . . , fk } be a generating set of pseudomonomials for JD . We can find pseudomonomials {fˆ1 , . . . , fˆk } in F2 [n] so that φ(fˆi ) = fi , and letting JC be the neural ideal
generated by {fˆ1 , . . . , fˆk } we see that φ(JC ) = JD .
4. The effects of homomorphisms on codes
Given that we can decompose any homomorphism preserving neural ideals into three basic
building blocks (permutation, bit flipping, and restriction) it is natural to ask how these
NEURAL IDEAL PRESERVING HOMOMORPHISMS
9
building blocks affect the underlying codes of neural ideals. In particular, if φ(JC ) = JD ,
how are the codes C and D related in terms of the decomposition of φ? Theorem 2.13 gives
a complete answer to this question.
Proof of Theorem 2.13. For any v ∈ Fn2 define λ(v) to be the vector whose support is
λ(supp(v)). Then note that λ(ρv ) = ρλ(v) for any vector v and permutation λ. With this we
compute directly that
λ(JC ) = hρλ(v) | v ∈
/ Ci = hρv | v ∈
/ λ(C)i = Jλ(C) .
For the bit flipping map δi we may define δi (v) to be the vector whose support is supp(v)⊕{i}
so that δi (C) = {δi (c) | c ∈ C}. Then note that δi (ρv ) = ρδi (v) and so we can compute
δi (JC ) = hρδi (v) | v ∈
/ Ci = hρv | v ∈
/ δi (C)i = Jδi (C) .
Finally, let ωm,m′ be a restriction map and let σ = [n] \ [m′ ]. For simplicity we write ω for
ωm,m′ and let D = rest(C, m, m′ ). We then wish to show that ω(JC ) = JD . Recall from
Definition 2.12 that u ∈ D if and only if supp(u) ∪ σ ∈ C. We will argue that ω(JC ) and JD
contain the same indicator pseudomonomials. Let ρu ∈ JD be an indicator pseudomonomial,
and let v be the n-bit vector whose support is supp(u) ∪ σ. Note that since u ∈
/ D we
have v ∈
/ C by the definition of restriction, and hence ρv ∈ JC . By construction we have
ω(ρv ) = ρu , and so ρu ∈ ω(JC ).
For the reverse inclusion, observe that since {ρv | v ∈
/ C} is a generating set for JC the
set {ω(ρv ) | v ∈
/ C} generates ω(JC ). All nonzero polynomials in {ω(ρv ) | v ∈
/ C} will be
indicator pseudomonomials in F2 [m], since applying ω removes all variables with indices
greater than m while leaving those with smaller indices fixed. Each neural ideal is uniquely
associated to its generating set of indicator pseudomonomials, and so we conclude that the
nonzero elements of {ω(ρv ) | v ∈
/ C} comprise all indicator pseudomonomials in ω(JC ). Thus
if ρu ∈ ω(JC ) is an indicator pseudomonomial then ρu = ω(ρv ) for some ρv ∈ JC . But
the only indicator ρv whose image under ω is ρu arises from the vector v whose support is
supp(u) ∪ σ. Any other indicator will either disagree on a variable xi with i ∈ [m], or it will
vanish under ω. Thus if ρu ∈ ω(JC ) then supp(u) ∪ σ ∈
/ C, which implies that u ∈
/ D by the
definition of restriction. Then ρu ∈ JD , as desired. Having proven that ω(JC ) and JD are
generated by the same indicator pseudomonomials we conclude that they are equal and the
overall result follows.
5. Geometric interpretation of homomorphisms preserving neural ideals
Finally we give a geometric characterization of permutation maps, bit flipping maps, and
restriction maps. Theorem 2.13 allows us to translate the behavior of these homomorphisms
into the world of neural codes, which we then interpret geometrically in Theorem 2.15.
Definition 5.1. Let U = {U1 , . . . , Un } be a collection of sets and v ∈ Fn2 be a binary vector.
Then the codeword region of v in U is the set
!
\
[
AUv :=
Ui \
Uj .
vi =1
vj =0
The codeword region is the set of points where all the sets Ui for i ∈ supp(v) are present,
and no others. Note that with this definiton we can write C(U) = {v ∈ Fn2 | AUv 6= ∅}.
That is, C(U) is the collection of vectors v so that the codeword region AUv is nonempty.
10
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
This notation is used heavily in the following proof to simplify the translation between a
collection of sets and its associated code.
Proof of Theorem 2.15. Let U = {U1 , . . . , Un } be a realization of the code C. The codeword
region for λ(v) in the realization {Uλ(1) , . . . , Uλ(n) } is exactly AUv since we have simply permuted the indices of U. We conclude that the code of the collection {Uλ(1) , . . . , Uλ(n) } is
{λ(v) | v ∈ C}, as desired. For bit flipping maps, observe that the codeword region AUv is
exactly the codeword region for the vector whose support is supp(v) ⊕ {i} in the realization
{U1 , . . . , X \Ui , . . . , Un } since taking the set difference with Ui is the same as intersecting with
its complement, and intersecting with Ui is the same as taking the difference with its complement. Hence u is in the code of {U1 , . . . , X \Ui , . . . , Un } if and only if supp(u) = supp(c)⊕{i}
for some c ∈ C and the result follows.
Finally we show that {Ui ∩ X ′ | i ∈ [m]} is a realization of rest(C, m, m′ ). For convenience
let V = {Ui ∩X ′ | i ∈ [m]} and σ = [n]\[m′ ] as in Definiton 2.12. We then consider the region
V
AVv for all vectors v ∈ Fm
2 . If v = 00 · · · 0 then the empty intersection in the definition of Av
′
′
V
will be all of X since we are considering this realization in the space X . The region Av is
nonempty if and only if X ′ is not covered by {Ui | i ∈ [m]}, i.e., if and only if supp(c) = σ
for c ∈ C. Hence in this case we have v ∈ C(V) if and only if supp(v) ∪ σ ∈ C. When v is
nonzero we compute directly that
\
[
AVv =
(Ui ∩ X ′ ) \
(Uj ∩ X ′ )
i∈supp(v)
=
\
i∈supp(v)∪σ
j∈[m]\supp(v)
Ui \
[
Uj .
j∈[n]\(supp(v)∪σ)
In this case we also see that v ∈ C(V) if and only if supp(v) ∪ σ ∈ C. By definition this
implies that C(V) = rest(C, m, m′ ), proving the result.
This geometric interpretation allows us to specify certain classes of maps which preserve
convex neural ideals.
Corollary 5.2. Permutation maps preserve convex neural ideals. When m = m′ the restriction map ωm,m′ preserves convex neural ideals.
Proof. Relabelling sets in a realization does not affect convexity, and so permutation maps
necessarily preserve convex neural ideals. For a restriction map ω with m = m′ , we consider
the realization given by Theorem 2.15. If we begin with a convex realization U of C then the
compatible region X ′ is simply the intersection of a finite number of convex open sets and
hence is itself convex and open. Likewise, all Ui ∩ X ′ are convex and open, so {Ui ∩ X ′ | i ∈
[m]} is a convex realization of rest(C, m, m′ ). By Theorem 2.13 we have ω(JC ) = Jrest(C,m,m′ )
and so ω(JC ) is a convex neural ideal.
Example 5.3. Figure 2 shows a code C on four bits along with a convex realization U =
{U1 , U2 , U3 , U4 } in the plane. We can apply Theorem 2.15 to obtain a realization of the code
rest(C, 2, 2) in the space X ′ = U3 ∩ U4 . As observed in Corollary 5.2, this new realization
consists of convex sets in a convex space, and hence rest(C, 2, 2) is a convex code.
Bit flipping maps are the least well behaved among maps preserving neural ideals in terms
of respecting convexity of the underlying codes. Indeed, taking the complement of a convex
NEURAL IDEAL PRESERVING HOMOMORPHISMS
11
r
Figure 2. A convex realization of a code C, and a convex realization of
rest(C, 2, 2) as given by Theorem 2.15 and Corollary 5.2.
set almost never yields another convex set. Thus in the context of convex codes, bit flipping
does not play a natural role.
It is also important to note that Corollary 5.2 does not completely describe when convexity
is preserved. Corollary 5.2 gives us a guarantee about when φ(JC ) corresponds to a convex
code for all convex codes C ⊆ Fn2 , but there may be cases when JC and φ(JC ) both correspond
to convex codes for other homomorphisms φ. For example, it is possible that δi (JC ) = JC for
a convex code C.
6. Homomorphisms and the canonical form
Lastly we prove Theorem 2.17. As a preliminary ingredient we show that every pseudomonomial can be expressed as a sum of indicator pseudomonomials.
Lemma 6.1. Let f be a pseudomonomial. Then f can be written as a sum of indicator
pseudomonomials.
Proof. We proceed by induction on n − deg(f ). Note that we have n − deg(f ) ≥ 0 since
pseudomonomials never have degree larger than n. In the base case that n − deg(f ) = 0 we
have that deg(f ) = n so f is an indicator and the result follows trivially. If n − deg(f ) > 0
then there is some variable xi on which f does not depend. Then we may apply the inductive
hypothesis to the pseudomonomials xi f and (1 − xi )f . Noticing that f = xi f + (1 − xi )f we
obtain the desired result.
Proof of Theorem 2.17. By Theorem 2.10 it suffices to prove the theorem for permutations,
bit flipping maps, and restrictions. In the case of permutations and bit flipping maps the
result is clear since these are automorphisms of F2 [n] and so they provide a bijective correspondence between pseudomonomials that are minimal with respect to division. Thus if φ
is a composition of bit flipping and permutation maps then CF (φ(JC )) = φ(CF (JC )).
This leaves the case of a restriction P
map ω with parameters 1 ≤ m ≤ m′ ≤ n. In this
case let f ∈ CF (φ(JC )) and write f =
ρv as guaranteed by Lemma 6.1. Recall from the
12
R. AMZI JEFFS, MOHAMED OMAR, AND NORA YOUNGS
proof of Theorem 2.13 that ρv ∈ ω(JC ) if and only if ρu ∈ JC where supp(u) = supp(v) ∪ σ.
Defining a pseudomonomial
Y
Y
(1 − xj )
xi
h=
m′ +1≤i≤n
m+1≤j≤m′
we see
Pthat it is equivalent to state that ρv ∈ ω(JC ) if and only if hρv ∈ JC . We conclude
that
hρv is in JC , noting that this is simply the pseudomonomial hf . Furthermore we can
observe that ω(hf ) = f , and so we have shown that there is some pseudomonomial in JC
which maps to f .
Among all pseudomonomials in JC that map to f , choose fˆ to be minimal with respect to
division. We claim that the pseudomonomial fˆ is in CF (JC ). If g|fˆ for a pseudomonomial g ∈
JC then ω(g) must divide f . Since f ∈ CF (ω(JC )) we see that ω(g) = f . By the minimality
of fˆ we have g = fˆ, and so fˆ ∈ CF (JC ). We have shown that every pseudomonomial f in
CF (ω(JC )) is also in ω(CF(JC )) and so the result follows.
References
[1] Joshua Cruz, Chad Giusti, Vladimir Itskov, and Bill Kronholm. Open and closed convex codes. 2016.
http://arxiv.org/abs/1609.03502.
[2] Carina Curto, Elizabeth Gross, Jack Jeffries, Katherine Morrison, Mohamed Omar, Zvi Rosen, Anne
Shiu, and Nora Youngs. What makes a neural code convex? 2015. http://arxiv.org/abs/1508.00150.
[3] Carina Curto, Vladimir Itskov, Alan Veliz-Cuba, and Nora Youngs. The neural ring: an algebraic tool
for analyzing the intrinsic structure of neural codes. Bulletin of Mathematical Biology, 75(9), 2013.
[4] Carina Curto and Nora Youngs. Neural ring homomorphisms and maps between neural codes. 2015.
http://arxiv.org/abs/1511.00255.
[5] R. Amzi Jeffs. Convexity of neural codes. 2016. https://www.math.hmc.edu/ rjeffs/thesis/rjeffs-2016thesis.pdf.
[6] Caitlin Lienkaemper, Anne Shiu, and Zev Woodstock. Obstructions to convexity in neural codes. 2015.
http://arxiv.org/abs/1509.03328.
[7] M. Tancer. Intersection patterns of convex sets via simplicial complexes: a survey. Thirty essays on
geometric graph theory, pages 521–540, 2013.
Department of Mathematics. Harvey Mudd College, Claremont, CA 91711
E-mail address: [email protected]
Department of Mathematics. Harvey Mudd College, Claremont, CA 91711
E-mail address: [email protected]
Department of Mathematics. Colby College, Waterville, ME
E-mail address: [email protected]
| 0math.AC
|
1
Energy-Efficient Scheduling for Homogeneous
Multiprocessor Systems
arXiv:1510.05567v2 [cs.OS] 12 Nov 2015
Mason Thammawichai, Student Member, IEEE and Eric C. Kerrigan, Member, IEEE
Abstract—We present a number of novel algorithms, based on mathematical optimization formulations, in order to solve a
homogeneous multiprocessor scheduling problem, while minimizing the total energy consumption. In particular, for a system with a
discrete speed set, we propose solving a tractable linear program. Our formulations are based on a fluid model and a global scheduling
scheme, i.e. tasks are allowed to migrate between processors. The new methods are compared with three global energy/feasibility
optimal workload allocation formulations. Simulation results illustrate that our methods achieve both feasibility and energy optimality
and outperform existing methods for constrained deadline tasksets. Specifically, the results provided by our algorithm can achieve up to
an 80% saving compared to an algorithm without a frequency scaling scheme and up to 70% saving compared to a constant frequency
scaling scheme for some simulated tasksets. Another benefit is that our algorithms can solve the scheduling problem in one step
instead of using a recursive scheme. Moreover, our formulations can solve a more general class of scheduling problems, i.e. any
periodic real-time taskset with arbitrary deadline. Lastly, our algorithms can be applied to both online and offline scheduling schemes.
✦
1
I NTRODUCTION
D
Ue to higher computational power demands in modern computing systems, e.g. sensor networks, satellites, multi-robot systems, as well as personal embedded
electronic devices, a system’s power consumption is significantly increase. Therefore, an efficient energy management
protocol is required in order to balance the power consumption and workload requirement of a system. With this motivation, this paper is aimed at using an optimization-based
approach to develop an algorithm for a multiprocessor
real-time scheduling problem with the goal of minimizing
energy consumption.
The dynamic power consumption of a CMOS processor
is often represented as a function of both clock frequency
and supply voltage [1]. Thus, a dynamic voltage and frequency scaling (DVFS) scheme is often applied to reduce
processor power consumption. DVFS is now commonly implemented in various computing systems, both at hardware
and software levels. Examples include Intel’s SpeedStep
technology, AMD’s PowerNow! technology and the Linux
Kernel Power Management Scheme.
1.1 Terminologies and Definitions
This section provides basic terminologies and definitions
used throughout the paper.
Speed s: The speed s of a processor is defined as the
ratio between the operating frequency f and the maximum
system frequency fmax , i.e. s := f /fmax . We denote with
smin the minimum execution speed of a processor.
Task Ti : A task Ti is defined as a triple Ti := (ci , di , pi );
ci is the required number of processor cycles, di is the task’s
•
•
Mason Thammawichai is with the Department of Aeronautics, Imperial
College London, London SW7 2AZ, UK
E-mail: [email protected]
Eric C. Kerrigan is with the Department of Electrical & Electronic
Engineering and the Department of Aeronautics, Imperial College London,
London SW7 2AZ, UK.
E-mail: [email protected]
deadline and pi is the task’s period. If the task’s deadline
is equal to its period, the task is said to have an ‘implicit
deadline’. The task is considered to have a ‘constrained
deadline’ if its deadline is not larger than its period, i.e.
di ≤ pi . In the case that the task’s deadline can be less than,
equal to, or greater than its period, it is said to have an
‘arbitrary deadline’.
Job Tij : A job Tij is defined as the j th instance of task Ti ,
where j ≥ 1. Tij arrives at time (j − 1)pi , has the required
execution cycles ci and a deadline at time (j − 1)pi + di .
Taskset: The taskset is defined as a set composed of realtime tasks.
Minimum Execution Time xi : The minimum execution
time xi of a task Ti is the execution time of the task when
the task is executed at fmax , xi := ci /fmax .
Task Density δi (si ): The task density δi (si ) of a task Ti
executed at a speed si is defined as the ratio between the
task execution time and the minimum of its deadline and its
period, i.e. δi (si ) := ci /(si fmax min{di , pi }). When all tasks
are assumed to have an implicit deadline, this is often called
‘task utilization’.
Taskset Density D(si ): The taskset density D(si ) is
defined as theP
summation of all task densities in the taskset,
i.e. D(si ) := n
i=1 δi (si ), where n is a number of tasks in
the taskset.
The
minimum taskset density D is given by
P
D := ni=1 δi (1).
Feasibility Optimal: Feasibility optimality of a homogeneous multiprocessor system is obtained when the upper
bound on the minimum taskset density D, for which the
algorithm is able to construct a valid schedule such that no
deadlines are missed, equals the total number of processors
in the system.
Scheduling Scheme: Multiprocessor scheduling can be
classified according to the task migration scheme1 : a ‘global
1. A task migration occurs when a task execution is suspended on one
processor and continues on another processor, while pre-emption is used
when the execution of a task on a processor is suspended in order to
start executing another task.
2
scheduling scheme’ allows task migration between processors and a ‘partitioned scheduling scheme’ does not allow
task migration.
Remaining Execution
Time
x_ i
Fluid Schedule Path
1.2 Related Work
There are at least two well-known feasibility optimal homogeneous multiprocessor scheduling algorithms of an implicit deadline taskset that are based on a fluid scheduling
model: Proportionate-fair (Pfair) [2] and Largest Local Remaining Execution Time First (LLREF) [3]. Both Pfair and
LLREF are global scheduling algorithms. A fluid model,
shown in Figure 1, is the ideal schedule path of a task
Ti , where the remaining minimum execution time is represented by a straight line and the slope of the line is the
task execution speed si .
By introducing the notion of fairness, Pfair ensures that
at any instant no task is one or more quanta (time intervals)
away from the task’s fluid path. However, the Pfair algorithm suffers from a significant run-time overhead, because
tasks are split into several segments, incurring frequent
algorithm invocations and task migrations.
To overcome these disadvantages, the LLREF algorithm
pre-empts a task at two scheduling events within each time
interval [3]. One occurs when the remaining time of an
executing task is zero and the other event happens when the
difference between a deadline and a remaining execution
time of a task is zero.
By incorporating a DVFS scheme with a fluid model, [4]
proposed the real-time static voltage and frequency scaling
(RT-SVFS) algorithm, which allows the slope of a fluid
schedule to vary between 0 and 1. To improve the performance of the open-loop static algorithm of [4], a closedloop dynamic algorithm is proposed in [5]. Their dynamic
algorithm considered uncertainty in the task execution time.
By extending [4], an energy-efficient scheduling algorithm
for real-time tasks with unknown arrival time, known as a
sporadic real-time task, was proposed in [6].
Deadline Partitioning (DP) [7] is the technique that partitions time into intervals bounded by two successive task
deadlines, after which each task is allocated the workload
and is scheduled at each time interval. A simple optimal scheduling algorithm, called DP-WRAP, was presented
in [7]. The DP-WRAP algorithm partitions time according
to the DP technique and, at each time interval, the tasks are
scheduled using McNaughton’s wrap around algorithm [8].
McNaughton’s wrap around algorithm aligns all task workloads along a number line, starting at zero, then splits tasks
into chunks of length 1 and assigns each chunk to the same
processor.
However, the algorithms that are based on the fairness
notion [3]–[5], [7], [9], [10] are schedulability optimal, but
have hardly been applied in a real system, since they suffer
from high scheduling overheads, i.e. task pre-emptions and
migrations. Recently, two schedulability optimal algorithms
that are not based on the notion of fairness have been
proposed. One is the RUN algorithm [11], which uses a dualization technique to reduce the multiprocessor scheduling
problem to a series of uniprocessor scheduling problems.
The other is U-EDF [12], which generalises the earliest deadline first (EDF) algorithm to multiprocessors by reducing the
problem to EDF on a uniprocessor.
Practical Schedule Path
di
Non-Execution
Interval
Time
Execution
Interval
Ti
Fig. 1. Fluid Schedule and a Practical Schedule [3]
Alternatively to the above methods, the multiprocessor
scheduling problem can also be formulated as an optimization problem. However, since the problem is NP-hard, in
general [13], polynomial-time heuristic methods are often
used. An example of these approaches can be found in [14],
which addresses the energy-aware multiprocessor partitioning problem and [15], which proposes an energy and feasibility optimal global framework for a two-type multicore
platform. In general, the tasks are partitioned among the
set of processors, followed with computing the running
frequency. Among all of the feasibility assignments, an optimal energy consumption assignment is chosen by solving
a mathematical optimization problem, where the objective
is to minimize energy. The constraints are to ensure that
all tasks will meet their deadlines and only one processor
is assigned to a task. In partitioned scheduling algorithms,
such as [14], once a task is assigned to a specific processor,
the multiprocessor scheduling problem is reduced to a set
of uniprocessor scheduling problems, which is well studied [16]. However, a partitioned scheduling method cannot
provide an optimal schedule.
1.3 Contribution
In this paper, we propose three mathematical optimization
problems to solve a periodic hard real-time task scheduling problem on homogeneous multiprocessor systems with
DVFS capabilities. First is an MINLP, which adopts the fluid
model used in [2]–[4] to represent a scheduling dynamic.
The MINLP relies on the optimal control of a suitablydefined dynamic to globally solve for a valid schedule,
while solutions are obtained by solving each instance of
the time interval obtained using the DP technique [3], [4],
[7]. By determining a fraction of task’s execution time and
operating speed, rather than task assignments, the same
scheduling problem can be formulated as an NLP. Lastly, we
propose an LP for a system with discrete speed levels. Our
work presents homogeneous multiprocessor scheduling algorithms that are both feasibility optimal and energy optimal. Furthermore, our formulations are capable of solving
any periodic tasksets, i.e. implict, constrained and arbitary
deadlines.
3
1.4 Outline of Paper
2.3 Scheduling as an Optimal Control Problem
This paper is organized as follows: Section 2 defines our
scheduling problem in detail. Three mathematical optimization formulations to solve the same multiprocessor scheduling problem are proposed in Section 3. The simulation setup
and results are presented in Section 4. Finally, conclusions
and future work are discussed in Section 5.
The objective is to minimize the total energy consumption
of executing a periodic taskset within the hyperperiod L.
The scheduling problem can therefore be formulated as
the following infinite-dimensional continuous-time optimal
control problem:
XZ L
ℓ(aijk (t), sijk (t))dt
(3a)
minimize
2
xij (·),aijk (·),
i,j,k
sijk (·),i∈I,
j∈Ji ,k∈K
P ROBLEM F ORMULATION
0
subject to
2.1 Task and Processor Model
We consider a set of n periodic real-time tasks that are
to be partitioned on m identical processors, where each
processor’s voltage/speed can be adjusted individually. All
tasks are assumed to start at the same time. The hyperperiod
L is defined as the least common multiple of all task periods.
The tasks can be pre-empted at any time, do not share
resources and do not have any precedence constraints. It
is assumed that D ≤ m in order to guarantee the existence
of a valid schedule.
Below, we will refer to the sets I := {1, . . . , n}, Ji :=
{1, . . . , L/pi } and K := {1, . . . , m}. The remaining minimum execution time of job Tij at time t will be denoted by
xij (t). Note that ∀i, ∀j, ∀k, ∀t will be used as short-hand for
∀i ∈ I, ∀j ∈ Ji , ∀k ∈ K, ∀t ∈ [0, L], respectively.
2.2 Energy Consumption Model
For CMOS processors, the total power consumption is often
simply expressed as an increasing function of the form
P (s) := Pd (s) + Ps = αsβ + Ps ,
(1)
where Pd (s) is the dynamic power consumption due to the
charging and discharging of CMOS gates, α > 0 and β ≥
1 are hardware-dependent constants and the static power
consumption Ps , which is mostly due to leakage current, is
assumed to be either constant or zero [1].
The total energy consumption from executing a task can
be expressed as a summation of the active and idle energy
consumed, i.e. E = Eactive + Eidle , where Eactive is the
energy consumed when the processor is busy executing the
task and Eidle is the energy consumed when the processor
is idle. The energy consumed by executing and completing
a task at a constant speed s is
E(s) = Eactive (s) + Eidle (s)
c (Pactive (s) − Pidle )
=
+ Pidle d
fmax
s
x(Pactive (s) − Pidle )
=
+ Pidle d,
s
(2a)
(2b)
(2c)
where Pactive (s) := Pd (s) + Ps is the power consumption in
the active interval, Pidle is the power consumption during
the idle period. Pidle > 0 and Ps > 0 will be assumed to
be constants. Note that Pactive (s) − Pidle is strictly greater
than zero and that the last term Pidle d is not a function of
the speed.
xij ((j − 1)pi ) = xi ,
xij ((j − 1)pi + di ) = 0,
X
sijk (t)aijk (t),
ẋij (t) = −
X
∀i, j
∀i, j
(3b)
∀i, j, t, a.e.
(3d)
aijk (t) ≤ 1,
∀i, j, t
(3e)
aijk (t) ≤ 1,
∀k, t
(3f)
∀i, j, k, t
∀i, j, k, t
(3h)
(3c)
k
k
X
i,j
smin ≤ sijk (t) ≤ 1,
aijk (t) ∈ {0, 1},
(3g)
where
ℓ(a, s) := a(Pactive (s) − Pidle ),
sijk (t) is the execution speed of job Tij at time t and aijk (t)
is used to indicate processor assignment, i.e. aijk (t) = 1 if
and only if job Tij is active on processor k at time t.
The initial conditions on the minimum execution time of
all jobs are specified in (3b) and job deadline constraints are
specified by (3c). The fluid model of the scheduling dynamic
is given in (3d), where the state is x and the control inputs
are a and s. The constraint that each job is assigned to at
most one processor at a time is ensured by (3e) and (3f)
enforces that each processor is assigned to at most one job at
a time. Upper and lower bounds on the processor speed are
given in (3e). The binary nature of job assignment variables
are given by (3h).
3 S OLVING THE S CHEDULING P ROBLEM WITH
F INITE - DIMENSIONAL M ATHEMATICAL O PTIMIZA TION
This section provides details on three mathematical optimization problems to solve the same scheduling problem
defined in Section 2.3. The original problem (3) will be
discretized by introducing piecewise constant constraints on
the control inputs s and a.
Let {τ0 , τ1 , . . . , τN }, which we will refer to as the major
grid, be the set of time instances corresponding to the
distinct arrival times and deadlines of all jobs within the
hyperperiod L, where 0 = τ0 ≤ τ1 ≤ · · · ≤ τN = L.
3.1 Mixed-integer Nonlinear Program (MINLP-DVFS)
The above scheduling problem, subject to piecewise constant constraints on the control inputs, is most naturally
formulated as an MINLP, as defined below.
4
Though it has been observed by [3] that it is often
the case that tasks might have to be split in order to
make the entire taskset schedulable, the runtime overheads
caused by context switches can jeopardize the performance.
Therefore, a variable discretization time step [17] method
is applied in a so-called minor grid, so that the solution to
our scheduling problem does not depend on the size of the
discretization time step. Let {τµ,0 , . . . , τµ,M } denote the set
of time instances on a minor grid within the time interval
[τµ , τµ+1 ] with τµ = τµ,0 ≤ · · · ≤ τµ,M = τµ+1 , so that
{τµ,1 , . . . , τµ,M−1 } is to be determined for all µ from solving
an appropriately-defined optimization problem.
Let ∀µ and ∀ν be short-hand for ∀µ ∈ U := {0, . . . , N −
1} and ∀ν ∈ V := {0, . . . , M − 1}. Define the notation
[µ, ν] := (τµ,ν ), ∀µ, ν and the discretized state and input
sequences as
xij [µ, ν] := xij (τµ,ν ),
sijk [µ, ν] := sijk (τµ,ν ),
aijk [µ, ν] := aijk (τµ,ν ),
∀i, j, µ, ν
∀i, j, k, µ, ν
∀i, j, k, µ, ν
If sijk (·) and aijk (·) are constant in-between time instances on the minor grid, i.e.
sijk (t) = sijk [µ, ν], ∀t ∈ [τµ,ν , τµ,ν+1 ), µ, ν
aijk (t) = aijk [µ, ν], ∀t ∈ [τµ,ν , τµ,ν+1 ), µ, ν
(5a)
(5b)
then it is easy to show that the solution of the scheduling
dynamic (3d) is given by
xij [µ, ν + 1] = xij [µ, ν]
X
sijk [µ, ν]aijk [µ, ν],
− h[µ, ν]
∀i, j, µ, ν
(6a)
where
h[µ, ν] := τµ,ν+1 − τµ,ν , ∀µ, ν.
(6b)
Let Λ denote the set of all jobs within hyperperiod L, i.e.
Λ := {Tij |i ∈ I, j ∈ Ji }. Define a function ΦA : Λ → U × V
by ΦA (Tij ) := (µ, ν) such that τµ,ν = (j − 1)pi , ∀Tij ∈ Λ
and a function ΦD : Λ → U × V by ΦD (Tij ) := (µ, ν) such
that τµ,ν = (j − 1)pi + di , ∀Tij ∈ Λ.
The original problem (3) subject to the piecewise constant constraints on the inputs (5) is therefore equivalent to
the following finite-dimensional MINLP:
X
h[µ, ν]ℓ(aijk [µ, ν], sijk [µ, ν]) (6c)
minimize
xij [·],sijk [·],
µ,ν,i,j,k
aijk [·],h[·],
i∈I,j∈Ji ,k∈K
subject to (6a) and
∀i, j
∀i, j
(6d)
∀i, j, µ, ν
(6f)
∀k, µ, ν
(6g)
∀i, j, k, µ, ν
∀i, j, k, µ, ν
∀µ, ν
(6h)
∀µ
(6k)
(6e)
k
X
aijk [µ, ν] ≤ 1,
i,j
smin ≤ sijk [µ, ν] ≤ 1,
aijk [µ, ν] ∈ {0, 1},
0 ≤ h[µ, ν],
X
h[µ, ν] = τµ+1 − τµ ,
ν
Theorem 1. If a solution exists to (3b)–(3h) and M ≥
max{n − 1, m − 1}, then a solution exists to problem (6).
Proof: Follows from the fact that if a solution exists
to (3b)–(3h), then the DP-WRAP scheduling algorithm can
find a valid schedule with at most n − 1 context switches
and m − 1 migrations per slice [7, Thm 6].
3.2 Continuous Nonlinear Program (NLP-DVFS)
This section proposes an NLP formulation without integer
variables to solve the problem in Section 2.3. The idea is to
relax the binary constraints in (6i) so that the value of a can
be interpreted as the fraction of a time interval during which
the job is executed on a processor.
Moreover, the number of variables can also be reduced
compared to problem (6), since the processor assignment information does not help in finding the task execution order,
hence k is dropped from the subscripts in the notation. That
is, partitioning time using only the major grid (i.e. M = 1)
is enough to find a feasible schedule if a solution exists to
the original problem (3). Since we only need a major grid,
we define the notation [µ] := (τµ ).
Consider now the following finite-dimensional NLP:
X
(τµ+1 − τµ )ℓ(aij [µ], sij [µ])
(7a)
minimize
xij [·],sij [·],
aij [·],i∈I,j∈Ji
µ,i,j
subject to
k
xij [ΦA (Tij )] = xi
xij [ΦD (Tij )] = 0,
X
aijk [µ, ν] ≤ 1,
where (6j)–(6k) enforce upper and lower bounds on the
discretization time steps.
(6i)
(6j)
xij [ΦA (Tij )] = xi ,
xij [ΦD (Tij )] = 0,
xij [µ + 1] = xij [µ]
− (τµ+1 − τµ )sij [µ]aij [µ],
X
aij [µ] ≤ m,
∀i, j
∀i, j
(7b)
∀i, j, µ
(7d)
∀µ
(7e)
smin ≤ sij [µ] ≤ 1,
0 ≤ aij [µ] ≤ 1,
∀i, j, µ
∀i, j, µ
(7f)
(7c)
i,j
(7g)
where aij [µ] ∈ R is the fraction of the time interval
[τµ , τµ+1 ) for which job Tij is executing on a processor at
speed sij [µ] and (7e) specifies that the total workload in
time interval [τµ , τµ+1 ) should be less than or equal to the
system capacity.
Theorem 2. If a solution exists to (3b)–(3h), then a solution
to problem (7) exists. Furthermore, at least one valid
schedule satisfying (3b)–(3h) can be constructed from a
solution to problem (7) with the same energy consumption.
Proof: A similar argument as Theorem 1 can be used
to prove the existence of a solution. The simplest valid
schedule can be constructed using McNaughton’s wrap
around algorithm [8] for each time interval [τµ , τµ+1 ). During each interval [τµ , τµ+1 ), the taskset is schedulable since
(i) the total density of the taskset does not exceed the total
number of processors, which is guaranteed by contraint (7e),
(ii) no task workload is greater than its deadline, which
is ensured by constraint (7g) and (iii) task migration is
allowed, which is our assumption. The optimal energy
5
consumption relies on the fact that the total energy consumption does not depend on the order of which tasks are
scheduled, but rather depends on the taskset’s density, i.e.
the objective value stays the same regardless of the number
of discretization steps M on the minor grid.
3.3 Linear Program (LP-DVFS)
Suppose now that the set of speed levels is finite, as is the
case with a real processor. We denote with sq the speed of
the processor at level q ∈ Q := {1, . . . , l}, where l is the
total number of speed levels.
Consider now the following finite-dimensional LP:
X
(τµ+1 − τµ )ℓ(aqij [µ], sq )
(8a)
minimize
q
xij [·],aij [·],
i∈I,j∈Ji ,q∈Q
µ,i,j,q
subject to
xij [ΦA (Tij )] = xi
xij [ΦD (Tij )] = 0,
xij [µ + 1] = xij [µ]
− (τµ+1 − τµ )
X
sq aqij [µ],
∀i, j
∀i, j
(8b)
∀i, j, µ
(8d)
(8c)
q
X
aqij [µ]
≤ 1,
aqij [µ] ≤ m,
sk ,k∈K
i,k
X
uik (sk ) ≤ 1,
∀i, j, µ
(8e)
∀µ
(8f)
X
∀i, j, q, µ
(8g)
smin ≤ sk ≤ 1,
sk = 1,
i,j,q
0 ≤ aqij [µ] ≤ 1,
For a system with a continuous speed range, four algorithms were compared: (i) MINLP-DVFS, (ii) NLP-DVFS,
(iii) GP-SVFS, which represents a global energy/feasibilityoptimal workload partitioning with constant frequency scaling scheme and (iv) GP-NoDVFS, which is a global workload allocation without frequency scaling scheme. For a
system with discrete speed levels, three algorithms are compared: (i) LP-DVFS, (ii) GP-NoDVFS and (iii) GP-SDiscrete,
which represents global energy/feasibility-optimal workload allocations with constant discrete frequency scaling.
Note that the formulations of GP-SVFS and GP-SDiscrete are
similar to [14], [15]. Specifically, the GP-SVFS is based on a
constant frequency scaling with global scheduling scheme,
while [14] is a partitioning-based formulation and [15] is
a generalized formulation for a two-type heterogeneous
multi-core system. GP-SDiscrete is an extension to systems
with discrete speed levels. Details on GP-SVFS, GP-NoDVFS
and GP-SDiscrete are given below.
GP-SVFS/GP-NoDVFS: Solve
X
minimize
Lℓ(uik (sk ), sk )
(9a)
subject to
q
X
4.2 Comparison between Algorithms
q
where aij [µ] ∈ R is the fraction of the time interval
[τµ , τµ+1 ) for which job Tij is executing on a processor at
speed sq .
Constraint (8e) assures that a task will not run on more
than one processor at a time. Constraint (8f) guarantees that
a processor’s workload will not exceed its capacity at a time.
Lastly, constraint (8g) provides upper and lower bounds on
a fraction of a job’s execution time at a specific speed level
variable.
Note that, given a solution to the optimization problem (8), one could also employ McNaughton’s wrap around
algorithm [8] to construct a valid schedule of processor
assignments with the same energy consumption. In other
words, Theorem 2 can be applied here as well to prove an
existence of a valid schedule given a solution to problem (8).
∀i
(9b)
∀k
(9c)
k
uik (sk ) ≤ 1,
i
∀k (GP − SV F S)
(9d)
∀k (GP − N oDV F S) (9e)
where uik (sk ) is the task density on processor k , i.e.
uik (sk ) := xi /(sk fmax di ) and sk is the static execution
speed of processor k . A task will not be executed on more
than one processor at the same time due to (9b) and the
assigned workload will not exceed processor capacity due
to (9c). The difference between GP-SVFS and GP-NoDVFS
lies in restriction on operating speed to be either a continous
variable (9d) or fixed at the maximum speed (9e).
GP-SDiscrete: Determine a fraction of the workload yikq
of a task at a specific speed level and a processor speed level
selection zkq such that:
X
minimize
Lℓ(uikq (yikq )zkq , sq )
(10a)
yikq ,zkq ,i∈I,
k∈K,q∈Q i,k,q
subject to
X
yikq zkq = 1,
∀i
(10b)
zkq = 1,
∀k
(10c)
uikq (yikq )zkq ≤ 1,
∀i
(10d)
uikq (yikq )zkq ≤ 1,
∀k
(10e)
∀i, k, q
∀k, q
(10f)
k,q
4
S IMULATION R ESULTS
4.1 System, Processor and Task Models
The performance of solving the above optimization problems is evaluated on two models of practical commercial
processors, namely an XScale and PowerPC 405LP. The
power consumption details of the two commercial processors, which have also been used in [18]–[20], are given in
Table 1. The active power consumption models of the XScale
and PowerPC 405LP shown in Table 2 were obtained by a
polynomial curve fitting to the generic form (1) (details are
given in the appendix B). The plots of the actual data versus
the fitted models are shown in Fig. 2.
X
q
X
k,q
X
i,q
0 ≤ yikq ≤ 1,
zkq ∈ {0, 1},
(10g)
where yikq represents a fraction of the workload of a task
at a specific speed level and zkq is a speed level selection
variable for processor k , i.e. zkq = 1 if a speed level q is
selected and zkq = 0 otherwise. The constraints in (10b)
6
1.6
0.8
Actual Data
Power Model
Power Consumption (Watt)
Power Consumption (Watt)
1.8
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
Operating Speed s
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
Actual Data
Power Model
0.2
(a) XScale
0.4
0.6
Operating Speed s
0.8
1
(b) PowerPC 405LP
Fig. 2. Actual Data versus Fitted Model
TABLE 1
Commercial Processor Details for Simulation
Processor type
Frequency (MHz)
Speed
Voltage (V)
Active Power (mW)
Idle Power (mW)
150
0.15
0.75
80
400
0.4
1.0
170
XScale [21]
600
800
0.6
0.8
1.3
1.6
400
900
40 [18]
1000
1.0
1.8
1600
PowerPC 405LP [22]
33
100
266
333
0.1
0.3
0.8
1.0
1.0
1.0
1.8
1.9
19
72
600
750
12
TABLE 2
Processor Power Consumption Fitted Model
Processor type
XScale
PowerPC 405LP
Active Power Model
Pactive (s) := 1524.92s3.0269 + 75.1092
Pactive (s) := 736.87s2.0990 + 13.1333
TABLE 3
Simulation Tasksets
D
T1
T2
T3
T4
0.4
(0.75,5,10)
(0.75,5,10)
(0.5,10,10)
(0.5,10,10)
0.6
(1,5,10)
(1,5,10)
(1,10,10)
(1,10,10)
0.8
(1.5,5,10)
(1.5,5,10)
(1,10,10)
(1,10,10)
1.0
(2,5,10)
(2,5,10)
(1,10,10)
(1,10,10)
1.2
(2.5,5,10)
(2.5,5,10)
(1,10,10)
(1,10,10)
1.4
(3,5,10)
(3,5,10)
(1,10,10)
(1,10,10)
1.6
(3,5,10)
(3,5,10)
(2,10,10)
(2,10,10)
1.8
(3.5,5,10)
(3.5,5,10)
(2,10,10)
(2,10,10)
2.0
(4,5,10)
(4,5,10)
(2,10,10)
(2,10,10)
Note: The first parameter of a task is xi ; ci can be obtained
by multiplying xi by fmax .
guarantee that the total workload of a task is equal to 1.
The constraints in (10c) ensure that only one speed level
is selected, (10d) assures that a task will be executed on
only one processor at a time, (10f) ensures that a processor
workload capacity is not violated and (10g) emphasises that
the speed level selection variable is binary. Note that GPSVFS and GP-NoDVFS are NLPs and GP-SDiscrete is an
MINLP.
2. See Appendix A.
MAPE2
1.1236
5.2323
4.3 Simulation Setup and Results
For simplicity, we consider the case where four independent
periodic real-time tasks with constrained deadlines need
to be scheduled onto two homogeneous processors. The
total energy consumption of each taskset in Table 3 were
evaluated. For the MINLP-DVFS, NLP-DVFS and LP-DVFS
implementations, the major grid discretization step N := 2
because all tasksets for the simulation only have two distinct
deadlines. For the MINLP-DVFS, we chose the minor grid
discretization step M := 4, since there are at most four
jobs in each major grid, which implies that there will be
at most four preemptions within each major grid. All of our
formulations were modelled using ZIMPL [23] and solved
by SCIP [24].
Simulation results are shown in Figures 3 and 4. The
minimum taskset density D is represented on the horizontal
axis. The vertical axis represents the total energy consumption normalised by the energy used by GP-NoDVFS, i.e.
less than 1 means that an algorithm does better than GPNoDVFS. It can be seen from the plots that for a timevarying workload such as a constrained deadline taskset,
our results from solving MINLP-DVFS, NLP-DVFS and LPDVFS are energy optimal, while GP-SVFS, GP-Discrete and
GP-NoDVFS are not. This is because our formulations are
incorporated with time, which provides benefits on solving
a scheduling problem with both time-varying workload
1
0.8
MINLP−DVFS
NLP−DVFS
GP−SVFS
0.6
0.4
0.2
0
0.5
1
1.5
Minimum Taskset Density
2
Normalised Energy Consumption
Normalised Energy Consumption
7
1
0.8
MINLP−DVFS
NLP−DVFS
GP−SVFS
0.6
0.4
0.2
0
0.5
(a) Xscale
1
1.5
Minimum Taskset Density
2
(b) PowerPC 405LP
1
LP−DVFS
GP−Discrete
0.8
0.6
0.4
0.2
0
0.5
1
1.5
Minimum Taskset Density
2
Normalised Energy Consumption
Normalised Energy Consumption
Fig. 3. Simulation results for scheduling real-time tasks with constrained deadlines on two homogeneous processors with continuous speed range
1
LP−DVFS
GP−Discrete
0.8
0.6
0.4
0.2
0
0.5
(a) Xscale
1
1.5
Minimum Taskset Density
2
(b) PowerPC 405LP
Fig. 4. Simulation results for scheduling real-time tasks with constrained deadlines on two homogeneous processors with discrete speed levels
(constrained deadline taskset) and constant workload (implicit deadline taskset). In general, compared to the constant speed profile, the time-varying speed profile saving
increases as the minimum taskset density increases. It can
also be noticed that the percentage saving of the Xscale
is nonlinear compared to that of the PowerPC’s. This is
because the power consumption model of the Xscale is a
cubic, while the PowerPC’s power consumption model is a
quadratic. However, it has to be mentioned that the energy
saving percentage varies with the taskset, which implies that
the number shown on the plots can be varied, but the significant outcomes stay the same. Lastly, the computation times
to solve NLP-DVFS, LP-DVFS, GP-SVFS and GP-SDiscrete
are extremely fast, i.e. milliseconds using a general-purpose
desktop PC with off-the-shelf optimization solvers, while
the time to solve MINLP-DVFS can be up to an hour in
some cases.
5
C ONCLUSIONS
Three mathematical optimization problems were proposed
to solve a homogeneous multiprocessor scheduling problem
with a periodic real-time taskset. Though our MINLP and
NLP formulations are both energy optimal and feasibility optimal, the computation time is high compared with
heuristic algorithms. However, our LP formulation is computationally tractable and suitable for a system with discrete
speed level set. Moreover, we have shown via simulations
that our formulations are able to solve a more general class
of scheduling problem than existing work in the literature
due to incorporating a scheduling dynamic model in the formulations as well as allowing for a time-varying executing
speed profile. The simulation results illustrate the possibility
that the time-varying speed profile can save energy up to
70% compared to the constant speed profile. Possible future
work could also include developing numerically efficient
methods to solve the various mathematical optimization
problems defined in this paper. One could also extend the
ideas presented here to solve a dynamic scheduling problem
with uncertainty in task execution time and to include slack
reclamation for further energy reduction.
R EFERENCES
[1]
[2]
[3]
J. M. Rabaey, A. P. Chandrakasan, and B. Nikolic, Digital integrated
circuits : a design perspective, 2nd ed., ser. Prentice Hall electronics
and VLSI series. Pearson Education, Jan. 2003.
S. K. Baruah, N. K. Cohen, C. G. Plaxton, and D. A.
Varvel, “Proportionate progress: A notion of fairness in resource
allocation,” in Proceedings of the Twenty-fifth Annual ACM
Symposium on Theory of Computing, ser. STOC ’93. New York, NY,
USA: ACM, 1993, pp. 345–354. [Online]. Available: http://doi.
acm.org/10.1145/167088.167194
H. Cho, B. Ravindran, and E. Jensen, “An optimal real-time
scheduling algorithm for multiprocessors,” in Real-Time Systems
Symposium, 2006. RTSS ’06. 27th IEEE International, Dec 2006, pp.
101–110.
8
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
K. Funaoka, S. Kato, and N. Yamasaki, “Energy-efficient optimal
real-time scheduling on multiprocessors,” in Object Oriented RealTime Distributed Computing (ISORC), 2008 11th IEEE International
Symposium on, May 2008, pp. 23–30.
K. Funaoka, A. Takeda, S. Kato, and N. Yamasaki, “Dynamic
voltage and frequency scaling for optimal real-time scheduling on
multiprocessors,” in Industrial Embedded Systems, 2008. SIES 2008.
International Symposium on, June 2008, pp. 27–33.
D.-S. Zhang, F.-Y. Chen, H.-H. Li, S.-Y. Jin, and D.-K. Guo, “An
energy-efficient scheduling algorithm for sporadic real-time tasks
in multiprocessor systems,” in High Performance Computing and
Communications (HPCC), 2011 IEEE 13th International Conference on,
Sept 2011, pp. 187–194.
G. Levin, S. Funk, C. Sadowski, I. Pye, and S. Brandt, “Dpfair: A simple model for understanding optimal multiprocessor
scheduling,” in Real-Time Systems (ECRTS), 2010 22nd Euromicro
Conference on, July 2010, pp. 3–13.
R. McNaughton, “Scheduling with deadlines and loss function,”
Machine Science, vol. 6(1), pp. 1–12, October 1959.
S. Funk, V. Berten, C. Ho, and J. Goossens, “A global optimal
scheduling algorithm for multiprocessor low-power platforms,”
in Proceedings of the 20th International Conference on Real-Time and
Network Systems, ser. RTNS ’12. New York, NY, USA: ACM,
2012, pp. 71–80. [Online]. Available: http://doi.acm.org/10.1145/
2392987.2392996
F. Wu, S. Jin, and Y. Wang, “A simple model for the energy-efficient
optimal real-time multiprocessor scheduling,” in Computer Science
and Automation Engineering (CSAE), 2012 IEEE International Conference on, vol. 3, May 2012, pp. 18–21.
P. Regnier, G. Lima, E. Massa, G. Levin, and S. Brandt, “Run:
Optimal multiprocessor real-time scheduling via reduction to
uniprocessor,” in Real-Time Systems Symposium (RTSS), 2011 IEEE
32nd, Nov 2011, pp. 104–115.
G. Nelissen, V. Berten, V. Nelis, J. Goossens, and D. Milojevic, “Uedf: An unfair but optimal multiprocessor scheduling algorithm
for sporadic tasks,” in Real-Time Systems (ECRTS), 2012 24th Euromicro Conference on, July 2012, pp. 13–23.
E. Lawler, “Recent results in the theory of machine scheduling,”
in Mathematical Programming The State of the Art, A. Bachem,
B. Korte, and M. Grtschel, Eds. Springer Berlin Heidelberg,
1983, pp. 202–234. [Online]. Available: http://dx.doi.org/10.
1007/978-3-642-68874-4 9
H. Aydin and Q. Yang, “Energy-aware partitioning for multiprocessor real-time systems,” in Parallel and Distributed Processing
Symposium, 2003. Proceedings. International, April 2003, pp. 9 pp.–.
H. S. Chwa, J. Seo, H. Yoo, J. Lee, and I. Shin, “Energy
and feasibility optimal global scheduling framework on
big.little platforms,” Department of Computer Science, KAIST
and Department of Computer Science and Engineering,
Sungkyunkwan University, Republic of Korea, Tech. Rep., 2014.
[Online]. Available: https://cs.kaist.ac.kr/upload files/report/
1407392146.pdf
J.-J. Chen and C.-F. Kuo, “Energy-efficient scheduling for realtime systems on dynamic voltage scaling (DVS) platforms,” in
Embedded and Real-Time Computing Systems and Applications, 2007.
RTCSA 2007. 13th IEEE International Conference on, Aug 2007, pp.
28–38.
M. Gerdts, “A variable time transformation method for mixedinteger optimal control problems,” Optimal Control Applications
and Methods, vol. 27, no. 3, pp. 169–182, 2006. [Online]. Available:
http://dx.doi.org/10.1002/oca.778
R. Xu, C. Xi, R. Melhem, and D. Moss, “Practical pace for
embedded systems,” in Proceedings of the 4th ACM International
Conference on Embedded Software, ser. EMSOFT ’04. New York,
NY, USA: ACM, 2004, pp. 54–63. [Online]. Available: http://doi.
acm.org/10.1145/1017753.1017767
[19] G. Zeng, T. Yokoyama, H. Tomiyama, and H. Takada, “Practical
energy-aware scheduling for real-time multiprocessor systems,” in
Embedded and Real-Time Computing Systems and Applications, 2009.
RTCSA ’09. 15th IEEE International Conference on, Aug 2009, pp.
383–392.
[20] H.-C. Wang, I. Woungang, C.-W. Yao, A. Anpalagan, and
M. S. Obaidat, “Energy-efficient tasks scheduling algorithm for
real-time multiprocessor embedded systems,” J. Supercomput.,
vol. 62, no. 2, pp. 967–988, Nov. 2012. [Online]. Available: http://
dx.doi.org/10.1007/s11227-012-0771-0
[21] I. X. M. Benchmarks, 2005, http://web.archive.org/web/
20050326232506/developer.intel.com/design/intelxscale/
benchmarks.htm.
[22] C. Rusu, R. Xu, R. Melhem, and D. Mosse, “Energy-efficient
policies for request-driven soft real-time systems,” in Real-Time
Systems, 2004. ECRTS 2004. Proceedings. 16th Euromicro Conference
on, June 2004, pp. 175–183.
[23] T. Koch, “Rapid mathematical prototyping,” Ph.D. dissertation,
Technische Universität Berlin, 2004.
[24] T. Achterberg, “SCIP: Solving constraint integer programs,” Mathematical Programming Computation, vol. 1, no. 1, pp. 1–41, July 2009,
http://mpc.zib.de/index.php/MPC/article/view/4.
A PPENDIX A
M EAN A BSOLUTE P ERCENTAGE E RROR (MAPE)
MAPE :=
k
1 X |F (z(i)) − y(i)|
× 100,
k i=1
|y(i)|
(11)
|F (z(i))−y(i)|
is the magnitude of the relative error in
where
|y(i)|
th
the i measurement, z 7→ F (z) is the estimated function,
z is the input data, y is the actual data and k is the total
number of fitted points.
A PPENDIX B
C URVE F ITTING
The following shows the curve fitting formulation to obtain
the active power consumption function of the XScale and
PowerPC 405LP with the objective of minimizing the MAPE:
minimize 100 ×
α,β,Ps
γ
1X
|αsβi + Ps − Pactive (si )|/Pactive (si )
γ i=1
subject to
Ps ≥ 0
α≥0
β≥1
(12a)
(12b)
(12c)
where α, β and Ps are parameters to be fitted, γ is the
total number of speed levels of a processor. The speed
level of a processor si and the active power consumption
of a processor Pactive (si ) are measurement data shown in
Table 1.
| 3cs.SY
|
arXiv:1703.07391v2 [math.AG] 9 May 2017
THE ASSOCIATED GRADED OF THE TEST MODULE
FILTRATION
AXEL STÄBLER
Abstract. We show that the associated graded of the test module filtration
τ (M, f λ )λ≥0 admits natural Cartier structures. If λ is an F -jumping number, then these Cartier structures are nilpotent on τ (M, f λ−ε )/τ (M, f λ ) if
and only if the denominator of λ is divisible by p. We also show that these
Cartier structures coincide with certain Cartier structures that are obtained
by considering certain D-modules associated to M that were used to construct
Bernstein-Sato polynomials.
Moreover, we point out that the zeros of the Bernstein-Sato polynomial
attached to an F -regular Cartier module correspond to its F -jumping numbers.
This generalizes [BS16a, Theorem 5.4]. Finally, we use non-F -pure modules to
prove a weaker connection between Bernstein-Sato polynomials of an arbitrary
Cartier module (M, κ) and certain jumping numbers attached to M .
Introduction
To fix ideas let us consider a hypersurface f inside a polynomial ring R =
Fp [x1 , . . . , xn ]. In order to study the singularities of f one may consider the test
ideal filtration τ (R, f λ )λ≥0 which is a decreasing Q-indexed right-continuous filtration of ideals that is defined in terms of certain p−e -linear maps. One has
τ (R, f 0 ) = R and the smallest λ for which τ (R, f λ ) 6= R is called the F -pure threshold, fpt(f ) for short. It has been known for a long time that if f ∈ Q[x1 , . . . , xn ]
and one considers the various reductions fp of f to positive prime characteristic,
p→∞
then fpt(fp ) −−−→ lct(f ) and fpt(fp ) ≤ lct(f ) for almost all p ([HY03, Theorem
6.8]). Here lct(f ) is the so-called log-canonical threshold which is a similar characteristic zero invariant that is defined using an embedded resolution of singularities.
It is moreover conjectured that fpt(fp ) = lct(f ) for infinitely many p ([MTW04,
Conjecture 3.6]).
It has been observed for quasi-homogeneous hypersurfaces that if the log canonical threshold does not coincide with the F -pure threshold, then the denominator
of the F -pure threshold is a pth power (see [HNnBWZ16, Theorem 3.5] or [Mül16,
Lemma 3.7 (2)]). On the other hand, there are only two known (families of) examples where the F -pure threshold does not coincide with the log canonical threshold
but the denominator of the F -pure threshold is not divisible by p (see [MTW04,
Example 4.5] and [CHSW16, Proposition 2.7, Corollary 2.10]).
From the point of view of birational geometry F -jumping numbers with a denominator divisible by p are special in the sense that the correspondence between
certain Cartier linear maps and certain Q-divisors breaks down in this case (e.g.
Date: January 25, 2018.
2010 Mathematics Subject Classification. Primary 13A35; Secondary 14F10, 14B05.
1
2
AXEL STÄBLER
[Sch09]). This correspondence is central to many applications of test ideals in birational geometry.
The purpose of this note is to further illustrate that F -pure thresholds whose
denominators are divisible by p are special in the following sense. We call a
finitely generated R-module M endowed with a p−1 -linear map, i.e. an R-linear map
κ : F∗ M → M a Cartier module. One can associate to M and f ∈ R a test module
filtration τ (M, f λ )λ≥0 that has similar properties as in
L the case M = R. In particular, we can form the associated graded Grλ M = λ>0 τ (M, f λ−ε )/τ (M, f λ ).
In Section 2 we will attach a natural Cartier module structure to these summands
and show that such a Cartier module is nilpotent if and only if the denominator of
λ is divisible by p. Here nilpotent means that some power of the structural map κ
acts as zero on the module.
This notion of nilpotence is interesting for the following reason. Nilpotent Cartier
modules form a Serre subcategory so that we may consider the attached localized
category of Cartier crystals. This category is then equivalent ([BB11]) to the category of unit R[F ]-modules of Emerton and Kisin ([EK04]) and anti-equivalent to
the category of perverse constructible Fp -sheaves on the étale site associated to
Spec R.
The first main result that we obtain is
Theorem (Theorem 2.3). Let R be essentially of finite type over an F -finite field,
(M, κ) be a Cartier module, f ∈ R and λ an F -jumping number of the test module
filtration of M along f . The Cartier structure on Grλ M defined by κe f ae (λ) is
not nilpotent if and only if ae (λ) = λ(pe − 1) and this quantity is an integer. In
particular, if the denominator of λ is divisible by p, then all these Cartier structures
are nilpotent.
Finally, we also show, extending and vastly simplifying some results of [BS16a,
Section 4], that test modules admit a simple description, akin to the case of an
ideal in a polynomial ring, in many cases. A special case is the following
Theorem (cf. Theorem 2.6). Let R be essentially of finite type over an F -finite
field, (M, κ) an F -regular Cartier module and f ∈ R an M -regular element, then
e
one has τ (M, f λ ) = κe f dλp e M for all e 0.
There is evidence that F -pure thresholds with denominator divisible by p are related to certain arithmetic phenomena like non-ordinarity via the anti-equivalence
defines a smooth quasi-homogeneous
mentioned above. If X = V (f ) ⊆ Pn+1
k
Calabi-Yau hypersurface then the F -pure threshold of f in k[x0 , . . . , xn+1 ] is 1 − hp ,
where 0 ≤ h ≤ n is the order of vanishing of the Hasse-invariant associated to a
certain deformation space of X (see [Mül16] for details). In particular, h 6= 0 if and
only if X is not ordinary. As a special case if dim X = 1, then non-ordinariness
1
1
(X) = 0, where the index 0 indicates the
coincides with the case Hcrys
(X)0 = Hét
slope zero part of crystalline cohomology. In this case, the crystalline cohomology
is concentrated at slope 21 .
Moreover, we will also relate these Cartier structures on Grλ M to certain Cartier
structures obtained from eigenspaces of higher Euler operators that play a crucial role when constructing Bernstein-Sato polynomials in positive characteristic
(see [BS16a]). Similar results were observed by Bitoun ([Bit15]) for the case that
M = R in the (equivalent) framework of unit R[F ]-modules. His proof relies on
formal properties of p-adic expansions. We will show that it is in fact also a formal
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
3
consequence of the machinery of test modules via Theorem 2.3. Our main result
in this direction is Theorem 3.8. Since its formulation requires several D-module
theoretic we only state a partial result in the introduction and refer the reader to
Section 3 for a detailed exposition.
e
Let DR[t]
be the ring of differential operators of order ≤ pe −1. We denote by θi ∈
[i]
e
DR[t]
so-called higher Euler operator ti ∂t . To the data (M, κ, f, λ) one associates
e
DR[t] -modules (for varying e). In order to construct Bernstein-Sato polynomials
one looks at certain eigenspaces of the higher Euler operators associated to these
e
DR[t]
-modules.
Theorem (cf. Theorem 3.8). Let R be regular and essentially of finite type over
an F -finite field, (M, κ) an F -regular Cartier module and f an M -regular element.
Then for e 0 there is an isomorphism F e ! Grλ M → Ei , where Ei is a certain
e
eigenspace associated to a certain DR
[θ1 , . . . , θpe−1 ]-module. This isomorphism ine!
λ
duces a transition map F Gr M → F e ! Grλ M . This transition map is induced by
a non-nilpotent morphism C : Grλ M → F s ! Grλ M if and only if λ(ps − 1) ∈ Z and
s
e, a are multiples of s. Moreover, in this case C is the adjoint of κs f λ(p −1) .
We will obtain a partial generalization of the correspondence between F -jumping
numbers and zero of Bernstein-Sato polynomials for the case that (M, κ) is F regular to arbitrary Cartier modules. The precise statement we prove is
Theorem (cf. Theorem 5.3). Let R be an F -finite regular ring and (M, κ) a Cartier
module. Fix a rational number λ. If for some e 0 such that λ(pe − 1) ∈ Z one
has that λ− pλe is a zero of the Bernstein-Sato polynomial beM,f (s), then Grσλ M 6= 0.
Here Grσ M is the filtration of M by so-called non-F -pure modules. We will study
their basic properties in Section 4 and also explain how they are a generalization of
non-F -pure ideals studied in [FST11]. We will also see that some of the pathologies
that non-F -pure ideals exhibit in comparison to characeristic zero are constrained
to the cases where the “jumps” have a denominator divisible by p.
We start with a short review of test modules in Section 1. In Section 2 we discuss
(non)-nilpotence of Cartier structures defined on the associated graded of the test
module filtration. Then after recalling the necessary setup we relate these Cartier
structures to the construction one uses to obtain Bernstein-Sato polynomials in
Section 3. In Section 4 we introduce the notion of non-F -pure module and study
its basic properties. In the last section we use these to prove Theorem 5.3.
Acknowledgements. I thank Kevin Tucker and Mircea Mustaţă for a useful discussion. Part of this paper was conceived while the author was visiting the University of Utah. I thank Karl Schwede for inviting me and for inspiring discussions.
The author is supported by grant STA 1478/1-1 of the Deutsche Forschungsgemeinschaft (DFG).
1. A brief review of test modules
In this section we review very briefly the necessary facts on test modules that
we need. We refer the reader to [BS16b] for a detailed treatment.
Fix an F -finite ring R. A Cartier module is a pair (M, κ), where M is an Rmodule and κ : F∗e M → M is an R-linear map. In many cases we assume e = 1. It
is however crucial to allow e ≥ 1 in order to obtain meaningful Cartier structures on
4
AXEL STÄBLER
certain quotients. If we assume that Spec R is embeddable into a smooth scheme,
then for given e there is a contravariant functor to the category constructible Fpe sheaves on the étale site. If we localize at nilpotent Cartier modules (to be defined
below), call the resulting category Cartier crystals, then this functor induces an
anti-equivalence between Cartier crystals and perverse constructible Fpe -sheaves
on the étale site (see [Sch16] and references therein).
Given a Cartier module (M, κ) and f ∈ R and λ a non-negative rational number
e
e
we can form R-linear maps κe f dλp e : F∗e M → M given by m 7→ κe (f dλp e m) for
varying e ≥ 1. The collection of these maps with addition induced
Lby the one in M
and multiplication by composition form an N-graded subring C of e≥0 Hom(F∗e M, M ),
where we set C0 = R. It has both a left and a right R-module structure that are
e
related via rϕ = ϕrp for any homogeneous element ϕ of degree e. This ring is a
special case of a so-called Cartier algebra.
1.1.
algebra is an N-graded ring
L Definition. Let R be an F -finite ring. A Cartier
pe
C
with
C
=
R
satisfying
the
relation
rϕ
=
ϕr
for any ϕ ∈ Ce and r ∈ R.
e
0
e≥0
L
h
As usual C+ = e≥1 Ce . We will write C+
for (C+ )h .
e
1.2. Remark. We will mostly use Cartier algebras of the form κe f dλp e in this
article. The reader familiar with test ideals may notice that people also often use
e
algebras of the form κe f dλ(p −1)e . If one computes test modules then both notions
yield the same result (cf. [Stä16, Lemma 3.1]). However, the categories of crystals
are not the same and this will play an important role later on (cf. Section 4).
A C-module means a left module over C. We will moreover always assume that it
a
is finitely generated as an R-module. We call a C-module M nilpotent if C+
M =0
for some (equivalently all) a 0. A morphism ϕ : M → N of C-modules is a
nil-isomorphism if its kernel and cokernel are nilpotent.
2
M ⊇ . . . stabilizes (see
If M is a C-module then the descending chain C+ M ⊇ C+
[Bli13, Proposition 2.13]) and we denote its stable member by M .
1.3. Definition. The test module τ (M, f λ ) is the smallest C-submodule N of M
such that the inclusion Hη0 (N )η ⊆ Hη0 (M )η is a nil-isomorphism for every associated
prime η of the R-module M .
At this point we encourage the reader to take a look at [BS16b, Sections 1 and
2] for further discussion. It is proven in [BS16b, Theorems 3.4 and 3.6] that test
modules exist if R is essentially of finite type over an F -finite field. Moreover, in
this case the test module filtration τ (M, f λ )λ≥0 is a decreasing right-continuous
discrete filtration. Many other formal properties like Briançon-Skoda also hold in
this more general situation (see [BS16b, Section 4]). We call a number λ such that
τ (M, f λ ) 6= τ (M, f λ−ε ) for all ε ≤ λ an F -jumping number.
Finally, we say that a prime η ∈ Spec R of a C-module M is an associated prime
of M if Hη0 (M )η is not nilpotent. These form a subset of the associated primes of
the underlying R-module.
We call a Cartier module (M, κ) F -regular if τ (M, f 0 ) = M . The smallest λ > 0
such that τ (M, f λ ) 6= τ (M, f 0 ) is called the F -pure threshold of f with respect to
e
M . More generally, if we have a C-module N , where Ce = κe f dλp e , then we say
that N is F -regular if τ (N, f λ ) = N .
If N is a C-module and η1 , . . . , ηn its associated primes then we call c1 , . . . , cn a
sequence of test elements if ci ∈
/ ηi and the Hη0i (Nci ) are F -regular. If all associated
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
5
primes of M are minimal then we only need a single test element and this condition
simplifies: If N is a C-module whose associated primes are minimal, then we call
c ∈ R a test element if c is not contained in any minimal prime of N and Nc is
F -regular.
With this notion one has
1.4. Theorem. Let R be essentially of finite type over an F -finite field, (M, κ) a
Cartier module and f ∈ R. Then there exists a sequence of test elements c1 , . . . , cn
e
for the C-module M , where Ce = κe f dλp e , and one has
τ (M, f λ ) =
n
XX
Ce cai i Hη0i (M )
e≥e0 i=1
for any e0 ≥ 0 and any ai ≥ 1. Moreover, if M only has minimal associated primes,
then this simplifies to
X
τ (M, f λ ) =
Ce ca M .
e≥e0
Proof. See [BS16b, Theorem 3.4, Theorem 3.6] for the general case and [Bli13,
Theorem 3.11] for the special case
It is mostly this presentation that we will be used in this article. Also note that
we will prove shortly that, if (M, κ) is F -regular, then one has in fact τ (M, f λ ) =
e
κe f dλp e M (Theorem 2.6 below).
2. Nilpotence of the associated graded
Throughout this section R is a ring essentially of finite type over an F -finite
field. This assumption is imposed to ensure existence of test modules (see [BS16b,
Theorem 3.6]; in our setup this automatically implies existence of a sequence of test
elements – cf. [BS16b, Remark 3.7]) and discreteness of the filtration. Granting
these notions our arguments work for arbitrary F -finite rings.
Fix a Cartier module (M, κ) and f ∈ R. In [Stä16, Proposition 4.5] the author
defined a Cartier structure on the associated graded of the test module filtration.
Namely, if Grλ (M ) = τ (M, f λ−ε )/τ (M, f λ ), then κf dλ(p−1)e operates on this quoe
tient. In fact, more generally κe f dλ(p −1)e operates on Grλ (M ) and one easily
checks if Grλ M is nilpotent with respect to this Cartier structure, then also with
respect to the one above.
While this definition may seem ad hoc we point out that if i : Spec R/(f ) →
Spec R denotes the natural inclusion and f is a non-zero-divisor on R then for any
Cartier module (M, κ) on Spec R the induced Cartier structure on R1 i! M is given
by κf p−1 (cf. [BB13, Example 3.3.12]). By Briançon-Skoda one easily sees that
the support of Grλ (M ) is contained in Spec R/(f ). We will in fact see shortly that
these Cartier structures are very natural.
The next lemma was already proven in [Stä16, Proposition 3.2] for the case that
M has only minimal primes. We give a simplified proof here.
2.1. Lemma. Let (M, κ) be a Cartier module and f ∈ R. Then for all λ ≥ 0 we
λ
have κ(τ (M, f λ )) = τ (M, f p ).
6
AXEL STÄBLER
Proof. By virtue of [BS16b, Theorem 3.4] we have
τ (M, f λ ) =
n X
X
e
κe f dλp e ci Hη0i (M )
i=1 e≥e0
for any e0 ≥ 0, where the ηi are the associated primes of M and the ci form a
sequence of test elements in the sense of [BS16b, Definition 3.1]. We thus have
κ(τ (M, f λ )) =
n X
X
λ
e+1
λ
e
κe+1 f d p p
e
ci Hη0i (M )
i=1 e≥e0
=
n
X
X
λ
κe f d p p e ci Hη0i (M ) = τ (M, f p ).
i=1 e≥e0 +1
2.2. Lemma. Let (M, κ) be a Cartier module and f ∈ R. For any integer ae (λ)
the map given by κe f ae (λ) induces a Cartier structure on Grλ M if and only if
ae (λ) ≥ dλ(pe − 1)e.
Proof. Using Lemma 2.1 and Briançon-Skoda ([BS16b, Proposition 4.1]) one has
κe f ae (λ) τ (M, f λ ) = τ (M, f
λ−ae (λ)
pe
).
For this to induce a Cartier structure on the quotient we must have λ−apee (λ) ≥ λ.
Equivalently, ae (λ) ≥ λ(pe − 1) and since ae (λ) is an integer this is equivalent to
ae (λ) ≥ dλ(pe − 1)e. One similarly checks that in this case κe f ae (λ) τ (M, f λ−ε ) ⊆
τ (M, f λ−ε ) using that the test module filtration is decreasing.
The first main result of this section is
2.3. Theorem. Let (M, κ) be a Cartier module, f ∈ R and λ an F -jumping number
of the test module filtration of M along f . The Cartier structure on Grλ M defined
by κe f ae (λ) is not nilpotent if and only if ae (λ) = λ(pe − 1) and this quantity is
an integer. In particular, if the denominator of λ is divisible by p, then all these
Cartier structures are nilpotent.
Proof. We may write dλ(pe − 1)e = λ(pe − 1) + δ with 0 ≤ δ < 1 and δ = 0 if and
only if λ(pe − 1) is an integer. By Briançon-Skoda ([BS16b, Proposition 4.1]) we
have
κe f λ(p
e
−1)+δ
τ (M, f λ−ε ) = κe τ (M, f λ−ε+λ(p
e
−1)+δ
) = κe τ (M, f λp
e
+δ−ε
).
Now we use Lemma 2.1 and obtain
κe τ (M, f λp
e
+δ−ε
) = τ (M, f λ+
δ−ε
pe
).
0
Note in fact, that we may take any ε such that ε > ε0 > 0 and still have
0
τ (M, f λ−ε ) = τ (M, f λ−ε ). Hence, if δ > 0 this actually forces ε < δ. But then
δ−ε
τ (M, f λ+ pe ) ⊆ τ (M, f λ ) which shows that the Cartier structure is nilpotent. The
same argument also shows nilpotence for any ea (λ) > dλ(pe − 1)e.
2.4. Remark. Note that with the notation of the proof of Theorem 2.3 if µ < λ
is the previous F -jumping number then necessarily λ − µ ≤ δ. In particular, if λ
is the F -pure threshold, then it actually follows that if δ 6= 0 then δ ≥ λ since we
can take any 0 < ε < λ without changing τ (M, f λ−ε ). Still assuming that λ is the
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
7
F -pure threshold and writing δ = dλ(pe −1)e−λ(pe −1) ≥ λ we obtain equivalently
that dλ(pe − 1)e ≥ λpe . From this one easily deduces that λ ∈
/ ( pae , pea−1 ) for any
e
integer 0 ≤ a ≤ p − 1. This is the analogue of [BMS09, Proposition 4.3 (ii)]
(which is proved using similar techniques) for modules. Also note that the result
of [BMS09] requires R to be regular and F -finite while the above argument works
for any F -finite ring.
2.5. Example. Consider the cusp f = x2 + y 3 ∈ Fp [x, y] = R. If λ denotes the F pure threshold then τ (R, f λ ) = (x, y) and for p > 3 one has (see [Mus09, Example
3.4])
(
5
p ≡ 1 mod 3,
λ = 65
1
−
p ≡ 2 mod 3.
6
6p
Moreover, λ = 12 for p = 2 and λ = 23 if p = 3. So if p ≡ 1 mod 3 we can take
e = 1 and obtain 56 (p − 1) ∈ Z. Since the quotient Grλ is just Fp and we know that
the obtained Cartier structure is not nilpotent it has to be κ = id.
Of course, for any p one has Grλ = Fp (as an R-module quotient) so that it
admits the non-nilpotent Cartier structure κ = id.
We end this section by proving a simple description of τ (M, f λ ) in the case where
(M, κ) is F -regular. This is very useful for computations. We will use it in the next
section to extend the relation of F -jumping numbers and zeros of Bernstein-Sato
polynomials ([BS16a]) to F -regular Cartier modules.
2.6. Theorem. Let (M, κ) be a Cartier module and f an M -regular element. If
e
(M, κ) is F -regular, then τ (M, f λ ) = κe f dλp e M for all e 0.
Proof. By F -regularity we have M = τ (M, f 0 ). Hence, by Briançon-Skoda ([BS16b,
Proposition 4.1]) and Lemma 2.1 we get
e
e
e
e
κe f dλp e M = κe f dλp e τ (M, f 0 ) = κe τ (M, f dλp e ) = τ (M, f dλp
e
e
λ ≤ dλp
pe
e
dλp e
Since
τ (M, f
ep−e
).
e
≤ λ+
we conclude by right-continuity that τ (M, f dλp
) for all e sufficiently large.
1
pe
ep−e
)=
2.7. Corollary. Let (M, κ) be a Cartier module and f an M -regular element. Then
e
τ (M, f λ ) = τ (τ (M, f 0 ), f λ ). In particular, we have τ (M, f λ ) = κe f dλp e τ (M, f 0 ).
Proof. Clearly, τ (M, f 0 ) ⊆ M so that τ (τ (M, f 0 ), f λ ) ⊆ τ (M, f λ ) by [BS16b,
Proposition 1.15].
For the other inclusion, by definition τ (M, f 0 ) is the smallest submodule of M for
which the inclusions Hη0 (τ (M, f 0 ))η ⊆ Hη0 (M )η are nil-isomorphisms with respect
to κ for all associated primes η of M . Being a nil-isomorphism here just means
that some
of κ annihilates the cokernel. But then a fortiori some power of
L power
dλpe e
C+ =
acts as zero on this cokernel. By definition of τ (M, f λ ) this
e≥1 κf
λ
shows τ (M, f ) ⊆ τ (M, f 0 ). Now we may apply τ (−, f λ ) to this inclusion to obtain
τ (M, f λ ) ⊆ τ (τ (M, f 0 ), f λ ), where we again use [BS16b, Proposition 1.15] and the
fact that τ (for a fixed Cartier algebra) is idempotent.
The final claim is an immediate application of Theorem 2.6.
2.8. Remark. The assumption on the F -regularity cannot be omitted, that is, if
e
M is not F -regular, then τ (M, f λ ) 6= κe f dλp e M in general. Consider for example
8
AXEL STÄBLER
R = Fp [x, y] and the Cartier module M = Fp [x, y] · y −1 ⊆ j∗ Fp [x, y, y −1 ], where
j : D(y) → Spec R, with Cartier structure induced by localization.
Then M is not F -regular since Fp [x, y] is a proper submodule that generically
agrees with M . It is easy to see that M is F -pure. Moreover, y is a test element
e
for (M, κ) and therefore τ (M, xλ ) = xbλc R while κe xdλp e M is not even contained
in R since y −1 is a fixed point for the Cartier operation.
Similarly, if M is not F -pure and has only minimal associated primes then
e
τ (M, f λ ) 6= κe f dλp e κa cM in general. For example, consider the Cartier module M = k[x] · x−n ⊆ k[x] with n ≥ 2 and take λ = 0 and f = x. Clearly, c = x
is a test element and M = κa k[x] · x−n = k[x] · x−1 . We see that τ (M, x0 ) = k[x]
by the theorem. However, κa xM = k[x] · x−1 for all a 0 so that (since λ = 0)
κe x0 k[x] · x−1 = k[x] · x−1 .
3. Cartier structures on the associated graded induced by
differential operators
Throughout this section we assume that R is a regular ring essentially of finite
type over an F -finite field. Regularity of R is critical since we need that F∗e R is a
flat R-module to ensure that F e ! is exact.
The goal of this section is to show that the Cartier structures defined on the associated graded of τ (M, f λ ) at the beginning of Section 2 correspond to the Cartier
e
[θ1 , θp , . . . , θp−1 ]-modules
structures obtained on quotients of eigenspaces of certain DR
λ
associated to (M, f ) for varying e. In particular, we will recover and generalize
results of Bitoun ([Bit15]).
We start by recalling the necessary D-module theoretic notions (see [Mus09] and
[BS16a] for more elaborate discussions). For any ring R containing Fp the ring of Fp linear differential operators DR ⊆ EndFp (R) in the sense ofSGrothendieck ([GD67,
e
e ∼
Definition 16.8.1]) admits the so-called p-filtration DR = e≥0 DR
, where DR
=
e
e
EndR (F∗ R) with colime≥0 DR = DR (see [Cha74]). In particular, we see that for
e
on F e ! M = HomR (F∗e R, M ).
any R-module M there is a natural right action of DR
3.1. Convention. Unless otherwise specified modules over rings of differential operators will always be right modules.
3.2. Example. If R = k[t1 , . . . , tn ] is a polynomial ring over a perfect field k, then
[1]
[pe −1]
[j]
e
DR
= k[∂ti , . . . , ∂ti
| i = 1, . . . , n], where the ∂ti are divided power operators
that act as follows
ai a1
ai−1 ai −j ai+1
[j]
∂ti • (ta1 1 · · · tann ) =
t · · · ti−1
ti
ti+1 · · · tann .
j 1
[a ]
[a ]
Moreover, one has [∂ti i , ∂tj j ] = 0 for i 6= j.
As an aside we also mention that since k is perfect DR coincides with k-linear
differential operators.
Assume from now on that R is a regular essentially of finite type over an F -finite
field and f ∈ R. We consider the graph embedding γ : Spec R → Spec R[t], t 7→
f . Given a Cartier module M on R we obtain a Cartier structure on γ∗ M via
the natural isomorphism γ∗ F∗ ∼
= F∗ γ∗ which we will from now on suppress from
[a]
e
notation. Note that DR[t]
contains the differential operator θa = ta ∂t for any
e
e
a < pe . In fact, DR[t]
= DR
[t, ∂ta | a < pe ]. We call the θa the (higher divided
power) Euler operators.
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
9
3.3. Lemma. Let (M, κ) be a Cartier module on R and γ : Spec R → Spec R[t], t 7→
f the graph embedding. Then the adjoint of the Cartier structure F e ! γ∗ M →
!
F e+a γ∗ M is given by
HomR[t] (F∗e R[t], γ∗ M ) −→ HomR[t] (F∗a+e R[t], γ∗ M )
ϕ 7−→ κa ◦ F∗e ϕ.
Proof. First of all, note that duality of finite morphisms yields a map
γ∗ M −→ F a ! γ∗ M,
m 7−→ [s 7→ κe (γ # (s)m)],
where γ # denotes the induced map R[t] → R, t 7→ f , corresponding to F∗a γ∗ M →
γ∗ M . Applying F e ! (which is just a hom-functor) we get a map
F e γ∗ M −→ F e ! F a ! γ∗ M,
ϕ 7−→ [r 7→ [s 7→ κa (γ # (r)(ϕ(s)))]].
!
Now by tensor-hom adjunction we have F e ! F a ! ∼
= F e+a which sends a map as
a
#
above to the map [s 7→ κ (ϕ(γ (s)))].
e
Fix a Cartier module M . Recall that F e ! γ∗ M admits a DR[t]
-module structure
e
and hence in particular a DR [θ1 , θp , . . . , θpe−1 ]-module structure. As such it admits
L
a decomposition into generalized eigenspaces F e ! γ∗ M = i∈Fep Ei , where θpj acts
on Ei by multiplication with ij (see [BS16a, Lemma 3.1] for details). As follows
from Lucas’ Theorem (see [Luc78, §XXI]) the projection F e ! γ∗ M → Ei is induced
by π : F∗e R[t] → F∗e R[t], Rtm 7→ 0 unless m = i1 + i2 p + . . . + ie pe−1 in which case
rtm 7→ rtm .
Moreover, this eigenspace decomposition is preserved by morphisms. If we cone−1
e
[θ1 , θp , . . . , θpe−2 ]-module, then these
sider a DR
[θ1 , θp , . . . , θpe−1 ]-module as a DR
eigenspace
decompositions
are
compatible
with
each other in L
the sense that if
L
M = i∈Fpe−1 Ei is an eigenspace decomposition, then the Ei = j∈Fp Ei1 ,...,ie−1 ,j
e−1
are the eigenspaces as DR
[θ1 , θp , . . . , θpe−2 ]-module.
Note that one has a natural map γ∗ F e ! M → F e ! γ∗ M, ϕ 7→ ϕ(γ # ). By abuse of
notation we will write γ∗ C e (M ) for the image of C e (M ) = {[r 7→ κe (rm)] | m ∈ M }
under this natural map.
A key technical result ([BS16a, Corollary 5.3]) yields that for all e 0 the
(i1 , . . . , ie )-eigenspace of the quotient
e
e
Ne := (γ∗ C e (M ))DR
[θ1 , θp , . . . , θpe−1 ]/(γ∗ C e (f M ))DR
[θ1 , θp , . . . , θpe−1 ]
e
is ismorphic as a DR
-module to
Pe := C e (f i1 +i2 p+...+ie p
e−1
e
M ) · DR
/C e (f 1+i1 +i2 p+...+ie p
e−1
e
M ) · DR
.
This isomorphism is given as follows. Given ϕ ∈ Ne that is contained in the
(i1 , . . . , ie )-eigenspace one has ϕ(rtm ) = 0 for all m < pe if m 6= i1 +i2 p+. . .+ie pe−1 .
e−1
The image of ϕ under this isomorphism is the map r 7→ ϕ(rf i1 +i2 p+...+ie p ).
With this notation and under the additional assumption that M is F -regular
it follows that Pe and hence the (i1 , . . . , ie )-eigenspace of Ne above is naturally
e
isomorphic as a DR
-module to
F e ! Gri M := F e ! (τ (M, f (i1 +i2 p+...+ie p
e−1
)/pe
)/τ (M, f (1+i1 +i2 p+...+ie p
e−1
)/pe
))
10
AXEL STÄBLER
for all e 0. For the case that M has only minimal associated primes, see
[BS16a, Corollary 4.8] and note that F e ! is exact since R is regular. In the general F -regular case, we can use Theorem 2.6 above instead. In fact, for n ∈ N we
have C e (f n M ) = {[r 7→ κe (rf n m)] | m ∈ M } ⊆ Hom(F∗e R, M ) while F e ! τ (M ) ⊆
F e ! M = Hom(F∗e R, M ) and the natural inclusion C e (f m M ) → F e ! M induces the
isomorphism.
We stress that [BS16a, Corollary 4.8] is the only test module theoretic input of
[BS16a]. Thus, by virtue of Theorem 2.6 the result [BS16a, Theorem 5.4] immediately generalizes to the more general framework of F -regular Cartier modules in
the sense of [BS16b]. In fact, Theorem 2.6 is more precise, since it guarantees an
m
equality κe f m M = τ (M, f pe ) for any m ∈ Z and e ≥ 1, thereby removing an inaccuracy in the proof of [BS16a, Theorem 5.4] – the application of [BS16a, Corollary
4.8] is not correct since one may have to enlarge the numerator.
We will prove a partial generalization of this result to arbitrary Cartier modules
in the last section.
With these preliminaries we can prove the
3.4. Proposition. Let M be a Cartier module on R and f ∈ R a non-zero!
divisor on M . Then the transition map γ∗ F e ! M → F e+a γ∗ M induces a map
Ne → Ne+a . Furthermore, if Ei is the (i1 , . . . , ie )-eigenspace of Ne and Ei,j the
(i1 , . . . , ie , je+1 , . . . , je+a )- eigenspace of Ne+a then we get an induced map
α : Ei
i
/ Ne+a
/ Ne
/ Ei,j ,
π
where π is the projection onto the eigenspace and i is the natural inclusion.
If, in addition, M is F -regular, then via the isomorphisms Ei → F e ! Gri M ,
!
Ei,j → F e+a Gri,j M and adjunction we get an induced Cartier structure F∗a F e ! Gri M →
e+a−1
F e ! Gri M which is given by ϕ 7→ κa f je+1 +je+2 p+...+je+a p
◦ ϕ.
Proof. Recall that the isomorphism Ei → F e ! Gri is given by
ϕ 7−→ [r 7→ ϕ(rf i1 +i2 p+...ie p
Similarly, the image of α(ϕ) in F
e+a !
e−1
)].
Gri,j M is given by
r 7−→ πκa ϕ(rf i1 +i2 p+...ie p
e−1
r).
We claim that this map coincides with the map
ψj : r 7−→ κa ϕ(rf i1 +i2 p+...ie p
l
e−1
+je+1 pe +...je+a pe+a−1
).
e+a
Note that {rt | r ∈ R with 0 ≤ l ≤ p
− 1} is a set of generators for F∗e+a R[t].
Since Ne+a is the direct sum of its eigenspaces the claim comes down to verifying
that the equality
X
ψj = κa ϕ
j∈Fa
p
holds in Ne+a , where we lift the ψj to Ne+a by taking the natural section of the
projection onto the eigenspace.
Hence, ψj evaluated at
ti1 +i2 p+...+ie p
e−1
coincides with
κa ϕ(f i1 +i2 p+...+ie p
+le+1 pe +...le+a pe+a−1
e−1
+le+1 pe +...le+a pe+a−1
)
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
11
if j = l and is zero otherwise. In particular, we see that the claimed identity holds.
Now since we may view ϕ as a map ϕ : F∗e R → M we have that
ψj (r) = κa f je+1 +je+2 p+...je+a p
a−1
ϕ(f i1 +i2 p+...+ie p
e−1
r)
which shows the claim.
3.5. Remark. We note that taking the projection onto the corresponding eigenspace
in Proposition 3.4 is necessary (see [Mus09, Example 6.15]).
Next, recall that any λ ∈ [0, 1) admits a unique non-terminating p-adic expansion
P
λ = i≥1 cip(λ)
with 0 ≤ ci (λ) ≤ p − 1. Moreover, one obtains for any e ≥ 1 that
i
Pe ci (λ)
dλpe e−1
.
i=1 pi =
pe
dλpe e
Hence, if λ ∈ [0, 1), then we can identify for all e 0 F e ! Gr pe M as a
e
DR
-module with the (c1 (λ), . . . , ce (λ))-eigenspace of Ne . Moreover, since the test
module filtration is discrete we may assume, by choosing a larger e if necessary,
that
dλpe e
dλpe e−1
τ (M, f pe ) = τ (M, f λ ) and τ (M, f pe ) = τ (M, f λ−ε ).
In other words Gr
dλpe e
pe
M = Grλ M .
3.6. Lemma. For e 0 the morphism
F e ! Grλ M
∼
=
/ Ei
α
∼
=
/ Ei,j
/ F e+a ! Grλ M ,
where α is the map in Proposition 3.4, is of the form F e ! C for a morphism C :
Grλ M → F a ! Grλ M if λ(ps − 1) is an integer for some s ∈ Z and both e, a are
multiples of s.
Proof. λ(ps − 1) is an integer if and only if the denominator of λ is not divisible by
p. This is equivalent to the fact that the p-adic expansion of λ is strictly periodic
Ps
(see
with period length dividing s. In this case, one has λ(ps − 1) = ps i=1 cip(λ)
i
0
0
e.g. [HNnBWZ16, Lemma 2.6]). Write e = e s and a = a s Then by Proposition
!
3.4 and the previous observation the Cartier structure F e ! Grλ M → F e+a Grλ M is
given by the adjoint of
κa f dλ(p
a
−1)e
= (κs f dλ(p
s
−1)e a0
)
and we may set
C = [m 7−→ [r 7→ (κs f dλ(p
s
−1)e e0
) (rm)]].
e!
3.7. Remark. Note that the Cartier structure on F Grλ M is only giving us information on Grλ M if we have a map C as in Lemma 3.6. Quite generally, in this
case, C is a nil-isomorphism (see e.g. [Stä17, Lemma 2.2]) so that as crystals Grλ M
and F e ! Grλ M coincide. Otherwise F e ! Grλ M does, at least a priori, not encode
more information than any other faithful functor.
We come to the main result of this section:
3.8. Theorem. Let R be an F -finite regular ring, (M, κ) an F -regular Cartier
module and f an M -regular element. Assume that λ ∈ Q is an F -jumping number
of the test module filtration of M along f . Then for e 0 we consider the map
!
C : F e ! Grλ M → F e+a Grλ M obtained by the isomorphism of Lemma 3.6.
12
AXEL STÄBLER
If λ(ps − 1) ∈ Z, then C is induced by a morphism Grλ → F s ! Grλ M if and only
s
if s | e and s | a. In this case, the morphism is given by the adjoint of κs f λ(p −1)
and is not nilpotent.
If λ(ps − 1) ∈
/ Z, then there is a morphism Grλ M → F s ! Grλ M inducing C if
and only if s | e, s = a and dλ(ps − 1)e is of the form c1 ps−1 + c2 ps−2 + . . . + cs ,
where cp1 + pc22 + . . . + pcss is the truncated p-adic expansion of λ. In this case, the
a
morphism is nilpotent and given by the adjoint of κa f dλ(p −1)e .
Proof. If κ : F∗e M → M is a Cartier structure then its adjoint is given by C :
M → F e ! M, m 7−→ [r 7→ κ(rm)]. In what follows we will not distinguish between
C and κ. The Cartier structure C is of the form κa f b(a) for some b(a) ∈ N. In
particular, the Cartier structure Grλ M → F s ! Grλ M is necessarily also of the form
κs f ? . But then, using Lemma 2.2, it is not nilpotent if and only if ? = λ(ps − 1)
s
with λ(ps − 1) ∈ Z. Clearly, if some power of κs f λ(p −1) induces C then we must
have s | e and s | a. The converse is the assertion of Lemma 3.6.
For the addendum note that it is clearly
P necessary that s | e and s | a. If we
write the p-adic expansion of λ as λ = i≥1 ci p−i , then the existence of such a
morphism κs f us (λ) : Grλ M → F s ! Grλ M is equivalent to a commutative diagram
of the form
κe+a f
κe f
λ
Gr M
pe −1
us (λ) s
p −1
us (λ)
pe+a −1
ps −1
/ F e ! Grλ M
κa f ca +ca−1 p+...+c1 p
(
/ F e+a ! Grλ M.
a−1
By Lemma 2.2 we may write us (λ) = dλ(ps − 1)e + m for some non-negative integer
m. Using this, the diagram is commutative if and only if we have the following
equality
(dλ(ps − 1)e + m)
pe − 1 a
pe+a − 1
p + ca + ca−1 p + . . . + c1 pa−1 = (dλ(ps − 1)e + m) s
.
s
p −1
p −1
This simplifies to
ca + ca−1 p + . . . + c1 pa−1 = (dλ(ps − 1)e + m)
pa − 1
.
ps − 1
Note that dλ(ps −1)e = c1 ps−1 +c2 ps−2 +. . .+cs +ε, where ε ∈ {0, 1} depending on
whether the fractional part is negative or positive. Using this formula and writing
a = sã we may further expand the equation to
ca + ca−1 p + . . . + c1 pa−1 = c1 pa−1 + c2 pa−2 + . . . + cs pa−s + (m + ε)pa−s
+ c1 pa−1−s + c2 pa−2−s + . . . + cs pa−2s + (m + ε)pa−2s
..
.
+ c1 ps−1 + c2 ps−2 + . . . + cs + (m + ε)
where we have ã rows. The summands in the first row of the right hand side except
the last one all occur on the left hand side. Subtracting this from both sides we
observe that the left hand side is < pa−s . Hence, if m + ε is positive, then equality
cannot hold. Thus, we must have m = ε = 0.
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
13
If m = ε = 0, then the last summand of each row vanishes. If moreover a = s,
then equality holds. If s < a, then there must exist 0 ≤ i ≤ s with ci 6= crs+i for
some 1 ≤ r < ã . Hence, equality cannot hold.
3.9. Example. The second case of the theorem occurs for instance for any F jumping number of the form p2ap
−1 with a < p (and necessarily p > 2).
If the denominator of λ is divisible by p then the second case of the theorem can
only occur if λ = pan b with p - b and a > b. Indeed, assume that a ≤ b. Denote the
coefficients of the p-adic expansion of λ by ci and the coefficients of the expansion
of ab by di . Then c1 = . . . = cn = 0 and cn+i = di for i ≥ 1. In particular, the
fractional part of λ(ps − 1) is then non-negative since it is of the form
X ci
X ci
−
.
pi−s
pi
i≥1
i≥1
s
Since λ(p − 1) is not an integer by assumption its fractional part is strictly positive
so that the second case of the theorem cannot occur.
3.10. Question. Are there F -pure thresholds which actually satisfy the second case
of Theorem 3.8?
A possible strategy may be to consider quasi-homogeneous polynomials f ∈
k[x1 , . . . , xn ] whose Jacobian ideal coincides, up to radical, with the irrelevant ideal.
For then a likely candidate for the F -pure threshold is the log canonical threshold
P
deg(xi )
.
deg f
If fpt(f ) 6= lct(f ) then the denominator of the F -threshold has to be a pth power
provided that p ≥ (n − 1) deg f − deg f − n + 1.PSo one would need f with fpt(f ) =
n
deg xi
lct(f ) and f having degree pb (since lct(f ) = i=1
).
deg f
Finally, we show that these results continue to hold if we work with the image of
the natural map γ∗ M → colime≥0 F e ! γ∗ M instead of γ∗ M → F e ! γ∗ M for varying
e. This is the setting of [Mus09] (for M = R). Indeed, note that by [BS16a, Proposition 3.5] one has colime≥0 F e ! M = γ+ colime≥0 F e ! M as right DR[t] -modules. If
M = ωR then colime≥0 F e ! ωR ∼
= ωR and the equivalence of right and left DR[t] modules yields γ+ R = R[t]t−f /R[t] which is what is considered in [Mus09].
Before we proceed, we need to recall one more notion from the theory of Cartier
crystals. If (M, κ) is a Cartier module then we can consider the union of all nilpotent submodules Mnil . This is again a nilpotent Cartier module and we define
M = M/Mnil . Note, in particular, that the natural projection M → M is a nilisomorphism.
We note the following
3.11. Lemma. Let i : Spec R/I → Spec R be a closed immersion and (M, κ) a
Cartier module. Then i∗ M = i∗ M .
3.12. Theorem. Assume that (M, κ) is an F -regular Cartier module satisfying
M = M . Write γ+ M = colime≥0 F e ! γ∗ M and denote the natural maps F e ! γ∗ M →
γ+ M by ϕe . Then for any e the quotients
e
e
ϕ0 (γ∗ C e (M ))DR
[θ1 , θp , . . . , θpe−1 ]/ϕ0 (γ∗ C e (f M ))DR
[θ1 , θp , . . . , θpe−1 ]
14
AXEL STÄBLER
and
e
e
[θ1 , θp , . . . , θpe−1 ]
γ∗ C e (M )DR
[θ1 , θp , . . . , θpe−1 ]/γ∗ C e (f M )DR
e
are naturally isomorphic as DR
[θ1 , θp , . . . , θpe −1 ]-modules. In particular, we obtain
e
DR -isomorphisms of generalized eigenspaces.
e
Proof. First of all, note that the natural map ϕe : F e ! γ∗ M → γ+ M is DR[t]
-linear.
Next, observe that by Lemma 3.11 we have γ∗ M = γ∗ M . It follows that the map
γ∗ M → F e ! γ∗ M is injective for all e (cf. [Stä16, Lemma 6.20]). Hence, the natural
γ∗ C e
ϕe
map ϕ0 is also injective. Note that ϕ0 factors as γ∗ M −−−→ F e ! γ∗ M −→ γ+ M for
e
any e ≥ 1. Since ϕe is DR[t]
-linear we have
e
e
ϕe (γ∗ C e (M )DR
[θ1 , θp , . . . , θpe−1 ]) = ϕ0 (γ∗ C e (M ))DR
[θ1 , θp , . . . , θpe−1 ],
and a similar assertion holds for γ∗ C e (f M ). We conclude that the quotients
e
e
ϕ0 (γ∗ C e (M ))DR
[θ1 , θp , . . . , θpe−1 ]/ϕ0 (γ∗ C e (f M ))DR
[θ1 , θp , . . . , θpe−1 ]
and
e
e
γ∗ C e (M )DR
[θ1 , θp , . . . , θpe−1 ]/γ∗ C e (f M )DR
[θ1 , θp , . . . , θpe−1 ]
e
[θ1 , θp , . . . , θpe −1 ]-modules since ϕ0 is injective. In
are naturally isomorphic as DR
e
particular, we obtain DR isomorphisms of generalized eigenspaces.
3.13. Proposition. Let (M, κM ), (N, κN ) be Cartier modules and f ∈ R M -regular
and N -regular. If ϕ : M → N is a nil-isomorphism then we get an induced nilisomorphism Grλ M → Grλ N with respect to any of the Cartier structures κe? f ae (λ) ,
where ? = M or N and ae (λ) ≥ dλ(pe − 1)e.
Proof. By [Stä17, Theorem 2.8] the restriction of ϕ induces a surjective map τ (ϕ) :
τ (M, f λ ) → τ (N, f λ ). By [Stä17, Lemma 2.1, proof of Theorem 2.8] we may assume
that both M, N are F -pure. In particular, ϕ is surjective. Now the claim follows
from [Stä16, Theorem 5.8].
In particular, given an F -regular Cartier module M we obtain nil-isomorphisms
Grλ M → Grλ M . Since for M working on the eth level or in the colimit induces
natural isomorphisms of the generalized eigenspaces of the quotients by Theorem
3.12 we see that we obtain the same nilpotence results if we construct the quotients
by working with the colimit.
4. Non-F -pure modules
In this section we study a generalization of the non-F -pure ideal or φ-fixed ideal
to modules. These non-F -pure ideals were first introduced in [FST11] and are
further developed and studied in [Sch14] and [HSZ14]. They are the characteristic
p analog of the so-called non-lc ideal and have applications to birational geometry.
The importance of these for us is that we will prove a connection between zeros
of Bernstein-Sato polynomials and certain non-F -pure-modules. These non-F -pure
modules will form a decreasing and discrete filtration of the ambient Cartier module.
However, there will be no continuity properties even if M is F -regular.
The actual connection with Bernstein-Sato polynomials will be discussed in the
next section. Here we just develop the basic theory of non-F -pure modules.
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
15
Let us however briefly motivate why it is interesting to consider non-F -pure
modules in this context. In the situation of the previous section, the generalized
eigenspaces always admit a description of the form
!
e
e ∼
C e (f m M ) · DR
/C e (f m+1 M ) · DR
= F e (κe f m M/κe f m+1 M ).
A first naive guess (based on [BS16a, Lemma 4.3] which asserts that the statement
is true if M is F -regular)L
may be that κe f m M (for e 0) should correspond to
e
e
D+ M = M , where D+ = e≥1 κe f dλp e R with m + 1 = dλpe e.
L
e
In practice, we replace the algebra D with the algebra C = e≥0 κe f dλ(p −1)e R.
Both algebras, C and D yield the same notion of test module (see [Stä16, Lemma
3.1]).
Unless specified otherwise R in this section is an arbitrary F -finite ring.
4.1. Definition. Fix an F -finite ring R. Given a Cartier module (M, κ) and an M e
regular element f we denote by C the Cartier algebra generated by the κe f dλ(p −1)e .
h
M for all h 0.
We define the non-F -pure submodule σ(M, f λ ) as C+
h+1
h
M ⊇ C+
M ⊇ ...
Note that by [Bli13, Proposition 2.13] the descending chain C+
stabilizes for arbitrary Cartier algebras in any F -finite ring.
4.2. Remark. (a) The reason for working with the algebra C rather than D is that
Lemma 4.3 does not hold in this context (cf. Example 4.5 below).
(b) This is not the definition that is used in [FST11]. We will see in Corollary 4.9
below that these definitions agree.
4.3. Lemma. Let R be an F -finite ring and let (M, κ) be a Cartier module. Given
an M -regularL
element f and a rational number λ we may consider the Cartier
e
algebra C+ = e≥0 κe f dλ(p −1)e R acting on M . For all a 0 such that λ(pa −1) ∈
a
Z we have σ(M, f λ ) = κa f λ(p −1) M .
h
Proof. Note that by definition σ(M, f λ ) = M = C+
M for all h 0. We first deal
h
with the inclusion from right to left. Fix a natural number h such that C+
M =
h+1
e
C+ M . Take any e such that λ(p − 1) ∈ Z and set a = eh. Then
κe f λ(p
|
e
−1)
· · · κe f λ(p
{z
e
−1)
h times
M = κa f λ(p
a
−1)
M
}
which shows the desired inclusion. Also note that, say, taking e minimal and
replacing h with h + 1 shows that e(h + 1) also works. So that the inclusion holds
for all a 0 such that λ(pa − 1) ∈ Z.
h
Given a such that λ(pa − 1) ∈ Z we show that for any h ≥ a one has C+
M ⊆
a λ(pa −1)
κ f
M . It suffices to show that
κe1 f dλ(p
e1
−1)e
· · · κeh f dλ(p
eh
−1)e
M ⊆ κa f λ(p
a
−1)
M
for all e1 , . . . , eh ≥ 1.
Write e = e1 + . . . + eh and e = la + r (note that l ≥ 1). Then we have
κe1 f dλ(p
e1
−1)e
· · · κeh f dλ(p
eh
−1)e
M ⊆ κe f dλ(p
la λ(pla −1) r λ(pr −1)+δ
=κ f
κ f
where δ = dλ(pe − 1)e − λ(pe − 1).
e
−1)e
M
la λ(pla −1)
M ⊆κ f
M = κa f λ(p
a
−1)
M,
16
AXEL STÄBLER
4.4. Example. In this example we point out that the assumption that λ(pa −1) is an
integer in Lemma 4.3 cannot be omitted. Let k be an F -finite field of characteristic
2. Consider the Cartier module M = k[x] x1 with Cartier structure induced from
the usual one on k[x] (i.e., xp−1 7→ 1, xi 7→ 0 for 0 ≤ i ≤ p − 2). Explicitly,
κ( xg ) = x1 κ(gxp−1 ).
M is F -pure but not F -regular (a test element is given by x). We filter M along
f = x2 and take λ = 21 . Then for any e ≥ 1 we have
κe f dλ(p
e
−1)e
M = κe f 2
e−1
e
1
k[x] = κe x2 −1 k[x] = k[x].
x
e
2
2
On the other hand, for any e ≥ 2 we have C+
M ⊆ C+
M . Elements of C+
M in
turn are sums of elements of
κa f dλ(p
a
−1)e b dλ(pb −1)e
κ f
m = κa+b f 2
a+b−1
f2
b−1
m ∈ κa+b x2
a+b
−1 2b
x k[x] ⊆ (x),
e
where a, b ≥ 1. We conclude that C+
M ⊆ (x) for e ≥ 2.
AsLremarked earlier, Lemma 4.3 does not hold if we use the Cartier algebra
e
D = e≥1 κe f dλp e R ⊕ R:
4.5. Example. Consider M = k[x] x1 endowed with its usual Cartier structure. Take
L
a
e
h
f = x then if D = e≥1 κe f dλp e R⊕R and λ = 1 one has C+
M ( κa f dλp e M for all
a
a, h 0. Indeed, for any a ≥ 1 the right hand side coincides with κa xp −1 k[x] =
k[x]. For the left hand side note that by [BS16a, Lemma 4.1] one has D+ M =
a
κa f dλp e M for all a 0. Hence, for h 0 we have
a
a
h+2
h
h a p
D+
M = D+
M = D+
κ x κa xp k[x]
1
h a pa
h
= D+
κ x k[x] = D+
(x).
x
Since x is a test element for M one observes that τ (M, x1 ) = (x). In particular,
h
(x) is F -pure with respect to the Cartier algebra D. Hence, D+
(x) = (x).
The following example further illustrates that the filtration obtained via D also
has all the shortcomings of the one obtained via C.
h
4.6. Example. We work in the situation of Example 4.4 but use D+ . Then D+
M 6=
a dλpa e
a λpa
a pa −1
κ f
M for all h, a 0. One has κ f M = κ x
k[x] = k[x]. If we had an
a
a
equality above then in particular D+ k[x] = D+ κa f λp M = κa f λp M = k[x]. But
e λpe
e pe
for any e ≥ 1 we have κ f k[x] = κ x k[x] = (x).
Furthermore, using this filtration there will be no continuity properties. For
instance, if (M, κ) is any F -pure non-F -regular Cartier module whose associated
h
primes are all minimal and f a test element then for λ = 0 one obtains D+
M =M
h
0
while for any 0 < λ 1 one obtains D+ M = τ (M, f ) using [BS16a, Lemma 4.1]
and Corollary 2.7. So the filtration is not right-continuous. On the other hand, if
(M, κ) is F -regular then the filtration coincides with the test module filtration and
is thus right-continuous but not left-continuous.
4.7. Proposition. Let R be F -finite and (M, κ) a Cartier module. Fix an M regular element f ∈ R. Then the filtration σ(M, f λ ) is discrete and decreasing.
Proof. In orderQto show that the filtrationQis decreasing simply observe that for
e
e
e
e
any e one has i=1 κei f d(λ+ε)(p −1)e M ⊆ i=1 κei f dλ(p −1)e M . Discreteness now
follows by the same argument as in [Bli13, Theorem 4.18]
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
17
In [FST11] the authors introduce some other candidates for the non-F -pure ideal
and ask whether they all coincide. In what follows we write C for the Cartier algebra
e
generated in degree e by κe f dλ(p −1)e . The following variants are considered:
(a) For a fixed n 0 define σn (M, f λ ) = (C≥n )h M for h 0.
(b) Suppose that λ(pe − 1) ∈ Z for some sufficientlyL
large and sufficiently divisible
ea λ(pea −1)
e. Write D for the algebra that is given by
. Then set
a≥0 κ f
h
σ 0 (M, f λ ) = D+
M for h 0.
Once again note that by [Bli13, Proposition 2.13] all these notions are welldefined for any F -finite ring R.
We point out that if one studies non-F -pure ideals in the context of non-principal
ideals (i.e. if f is replaced by an ideal a ⊆ R), then one works with the integral
e
closures of the adλ(p −1)e making things more subtle (if R is normal then any principal ideal generated by a non-zero-divisor is integrally closed – see e.g. [BH98,
Proposition 10.2.3]).
As an application of Lemma 4.3 we show that these notions all coincide in our
situation of a principal ideal. First we record an elementary
4.8. Lemma. Let R be F -finite and (M, κ) a Cartier module. Given a rational
ea
number λ such that λ(pa −1) ∈ Z we have that the κea f λ(p −1) M form a decreasing
chain for varying e.
Proof. One has κ2a f λ(p
2a
−1)
M = κa f λ(p
a
−1) a λ(pa −1)
κ f
M ⊆ κa f λ(p
a
−1)
M.
4.9. Corollary. Let R be essentially of finite type over an F -finite field. If the
denominator of λ is not divisible by p, then one has
σ(M, f λ ) = σn (M, f λ ) = σ 0 (M, f λ )
for n 0. In general, one still has σ(M, f λ ) = σn (M, f λ ) for all n 0.
Proof. Note that one clearly has containments σ 0 (M, f λ ) ⊆ σn (M, f λ ) ⊆ σ(M, f λ )
whenever these objects are defined.
Let us first consider the case λ(pa − 1) ∈ Z for some a. It suffices to show
b
that σ(M, f λ ) ⊆ σ 0 (M, f λ ). By Lemma 4.3 we have σ(M, f λ ) = κb f λ(p −1) M for
h
infinitely many b. In particular, if σ 0 (M, f λ ) = D+
M for some fixed h 0 and
L
ea λpea −1
D+ = a≥1 κ f
R then we may arrange for b = eh. Now use Lemma 4.8.
Let us now consider the case of an arbitrary λ. Fix h 0 such that (C≥n )h M =
h
σn (M, f λ ) and C+
M = σ(M, f λ ). An argument just as in the second part of the
a
proof of Lemma 4.3 shows that for any a ≥ n one has C+
M ⊆ C≥n M . We conclude
h
ah
h
that C+ M = C+ M ⊆ C≥n M .
As is well-known to experts one cannot expect any continuity properties for the
filtration σ(M, f λ ) (not even in the case M = R). For completeness we give an
example:
4.10. Example. Consider the case of Example 4.4 (i.e. M = k[x] x1 , f = x2 ) but
e
this time for a field k of characteristic 6= 2. We claim that C+
M = M for any λ ≤ 21
1
e
but C+ M = (x) for any λ > 2 . Since 2 and p are coprime we can use Lemma 4.3
and obtain
e
1
e
C+
M = κe xp −1 k[x] = k[x]
x
18
AXEL STÄBLER
for λ = 21 and hence for all λ ≤ 12 since the filtration is decreasing. If λ > 21 then
λ(pe − 1) > 12 (pe − 1). Choosing e in such a way that the left-hand side is an integer
we get that λ(pe − 1) ≥ 12 (pe − 1) + 1. Using Lemma 4.3 once more we compute
κe f λ(p
e
−1)
M ⊇ κe xp
e
+1
M = (x).
Now assume that char k = 2. We claim that the filtration is not left-continuous
at λ = 12 in this case. In Example 4.4 we saw that σ(M, f λ ) ⊆ (x). We will show
that for all ε > 0 one has σ(M, f λ−ε ) = R. We may assume that (λ − ε)(pa − 1) ∈ Z
for some a 0. Then using Lemma 4.3 we obtain
a
a
a+1
1
σ(M, f λ−ε ) = κa f (λ−ε)(p −1) M = κa x2 −2 ε−1+2ε k[x].
x
Since the exponent is non-zero and ≤ 2a − 1 we get σ(M, f λ−ε ) = R.
Finally, note that the Cartier module (M, κ) is isomorphic to the Cartier module
(k[x], κxp−1 ) via the multiplication by x map k[x] x1 → k[x].
4.11. Proposition. Assume that R is an F -finite ring and (M, κ) a Cartier module.
If the denominator of λ is divisible by p then σ(M, f λ ) = σ(M, f λ+ε ) for all 0 <
ε 1.
Proof. Since the filtration is decreasing by Proposition 4.7 we only need to show
the inclusion from left to right.
h+g
h
h
M = C+
⊆ C≥h . By definition we have σ(M, f λ ) = C+
M⊆
Next note that C+
h
h
C+
C≥g M for all h 0 and all g. Hence, any element of C+
C≥g M may be written
as
h
Y
ei
e0
κei f dλ(p −1)e κe0 f dλ(p −1)e m
i=1
h
M = σ(M, f λ ). If we write dλ(pei − 1)e =
with e0 ≥ g. Fix h such that C+
Ph
λ(pei − 1) + δi with 0 < δi < 1 for 0 ≤ i ≤ h and e = i=0 ei , then we obtain that
h
any element of C+
C≥g is of the form
Ph
Ph
ej
e
j=i+1
λ(p
−1)+
δ
p
i
i=0
κe f
.
Now we set
Ph
ej
j=i+1
δ
p
i=0 i
.
ε0 = lim
eh →∞
pe − 1
Note that this sequence converges to ε0 from above and that ε0 > 0. Thus by
perturbing a little bit we may choose ε > ε0 such that the denominator of λ + ε is
not divisible by p and such that
Ph
Ph
ej
e
e
j=i+1
δi p
e λ(p −1)+
i=0
κ f
m ∈ κe f (λ+ε)(p −1) M = σ(M, f λ+ε )
Ph
holds for all eh sufficiently large and divisible (using Lemma 4.3).
It is observed in [FST11, Remark 14.11] that, contrary to the situation in characteristic zero, one does not have σ(R, f λ ) = τ (R, f λ−ε ) if R is (F -)regular.
However, we have the
4.12. Proposition. Let R be essentially of finite type over an F -finite field, (M, κ)
an F -regular Cartier module and f an M -regular element. If the denominator of λ
is not divisible by p, then σ(M, f λ ) = σ(M, f λ−ε ) = τ (M, f λ−ε ).
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
19
e
Proof. By our assumption on λ we find e 0 such that σ(M, f λ ) = κe f λ(p −1) M .
Moreover, by discreteness we may assume that λ−ε also has denominator not divisible by p and that the same e computes σ(M, f λ−ε ). By F -regularity we have M =
τ (M, f 0 ). Using Briançon-Skoda ([BS16b, Proposition 4.1]) and Lemma 2.1 we
λ−ε
1
e
e
then have κe f λ(p −1) M = τ (M, f λ− pe ) and κe f (λ−ε)(p −1) M = τ (M, f λ−ε− pe )
and by discreteness of τ these coincide.
4.13. Question. Is the filtration σ(M, f λ ) left-continuous if the F -jumping numbers
of τ (M, f λ ) all have a denominator that is not divisible by p?
4.14. Proposition. Let R be essentially of finite type over an F -finite field, (M, κ)
an F -regular Cartier module and f an M -regular element.. Then for all λ ≥ 0 one
has σ(M, f λ+ε ) = τ (M, f λ ) for all 0 < ε 1.
Proof. Since M is F -regular we have τ (M, f 0 ) = M . As the filtration σ is discrete
and decreasing there exists 0 < ε 1 such that σ is constant in (λ, λ + ε). In
particular, we may assume that the denominator of λ + ε is not divisible by p.
Hence, we can apply Lemma 4.3 and obtain
σ(M, f λ+ε ) = κe f (λ+ε)(p
e
−1)
M = τ (M, f λ+ε−
λ+ε
pe
) = τ (M, f λ ),
where we use Briançon-Skoda ([BS16b, Proposition 4.1]) and Lemma 2.1 for the
second equality and right-continuity of τ (and the fact that we may choose a larger
e) for the last equality.
This correspondence breaks down however in the non F -regular case.
4.15. Example. Take M = k[x, y] x1 with the usual Cartier structure, k any F -finite
field of positive characteristic. We filter along f = y. Then τ (M, f 0 ) = k[x, y] and,
in particular, τ (M, f 1 ) = (y). For σ we obtain σ(M, f λ ) = k[x, y] x1 for λ ≤ 1 and
σ(M, f 1+ε ) = (y) x1 .
5. Bernstein-Sato polynomials attached to Cartier modules and the
filtration σ
5.1. Definition. Given an F -finite ring R, a Cartier module (M, κ) and an M regular element f we define Grσλ M = σ(M, f λ )/σ(M, f λ+ε ) for any 0 < ε 1.
5.2. Lemma. Let R be an F -finite ring, (M, κ) a Cartier module and f an M a
regular element. Then κa f dλ(p −1)e defines a Cartier structure on Grσλ M . If
λ
Grσ M 6= 0, then this Cartier structure is non-nilpotent if and only if λ(pa − 1) ∈ Z.
a
h+1
h
h
Proof. We have σ(M, f λ ) = C+
M = C+
M ⊇ κa f dλ(p −1)e C+
M . We may assume
e
λ+ε
e (λ+ε)(pe −1)
that (λ + ε)(p − 1) ∈ Z for some e 0 and σ(M, f
)=κ f
M . Write
dλ(pa − 1)e = λ(pa − 1) + δ. Then we compute
a
κa f dλ(p
−1)e
σ(M, f λ+ε ) = κa+e f λ(p
e+a
−1)+ε(pe −1)+δpe
M.
Replacing ε by 2pa ε0 and possibly choosing a larger e so that
0
κe f (λ+ε )(p
e
−1)
0
M = σ(M, f λ+ε ) = σ(M, f λ+ε )
we obtain that ε0 2pa (pe − 1) ≥ ε0 (pe+a − 1). Hence, we get the desired containment
κa f dλ(p
a
−1)e
σ(M, f λ+ε ) ⊆ σ(M, f λ+ε ).
20
AXEL STÄBLER
e
For the supplement, if λ(pa − 1) ∈ Z, then σ(M, f λ ) = κe f λ(p −1) M for all
a
e 0 such that λ(pe − 1) ∈ Z. So κa f λ(p −1) σ(M, f λ ) = σ(M, f λ ). Conversely,
we may write dλ(pa − 1)e = λ(pa − 1) + δ = (λ + δ 0 )(pa − 1). By construction the
denominator of λ + δ 0 is not divisible by p so that we obtain for suitable e 0
(κa f dλ(p
a
0
−1)e e
) σ(M, f λ ) = κea f (λ+δ )(p
ea
−1)
σ(M, f λ )
0
⊆ κea f (λ+δ )(p
ea
−1)
0
M = σ(M, f λ+δ )
where we use Lemma 4.3 for the last equality. As the filtration is decreasing this
shows nilpotence.
Assume that R is regular and F -finite. Let us now recall from [BS16a] the
notion of Bernstein-Sato polynomial attached to a Cartier module (M, κ) and an
M -regular element f . As in Section 3 we denote by γ : Spec R → Spec R[t] the
e
graph embedding t 7→ f and may consider, for any e ≥ 1, the right DR
[θ1 , . . . , θpe−1 ]module
e
e
Ne := (γ∗ C e (M ))DR
[θ1 , θp , . . . , θpe−1 ]/(γ∗ C e (f M ))DR
[θ1 , θp , . . . , θpe−1 ].
Now let Γef ⊆ Fep be the set of those i = (i1 , . . . , ie ) for which the corresponding
i-eigenspace with respect to the action of the θ1 , θp , . . . , θpe−1 is non-trivial. Then
by lifting Fp to {0, . . . , p − 1} ⊆ Z we define the eth Bernstein-Sato polynomial
beM,f (s) as
Y
ie
i1
(s − ( + . . . + e−1 )).
p
p
e
i∈Γf
5.3. Theorem. Let R be an F -finite regular ring and (M, κ) a Cartier module. Fix
a rational number λ. If λ(pe − 1) is a zero of the Bernstein-Sato polynomial for
e 0 such that λ(pe − 1) ∈ Z, then Grσλ is non-trivial.
Proof. First of all, we may assume that (M, κ) is F -pure (use [Stä17, Lemma 2.1]).
By construction the zeros of the Bernstein-Sato polynomial are in [0, 1]. Fix 0 <
ε 1 such that λ + ε ∈ Z(p) and take e 0 such that (using Lemma 4.3)
e
e
σ(M, f λ ) = κe f λ(p −1) M and σ(M, f λ+ε ) = κe f (λ+ε)(p −1) M . Note that one has
e
e
e
κe f λ(p −1) f M ⊇ κe f λ(p −1 )f ε(p −1) M by possibly choosing a larger e.
e
e
Now Grσλ M is non-trivial if and only if κe f λ(p −1) M 6= κe f (λ+ε)(p −1) M . Since
e!
e
R is regular the functor F (−) = Hom(F∗ R, −) is fully faithful (i.e. Frobenius is
finite flat – [Kun69]) so that this is further equivalent to
F e ! (κe f λ(p
e
−1)
M ) 6= F e ! (κe f (λ+ε)(p
e
−1)
M ).
Using [BS16a, Lemma 4.6], one obtains
F e ! (κe f λ(p
and similarly for F e ! (κe f (λ+ε)(p
C e (f λ(p
e
−1)
e
e
−1)
−1)
M ) = C e (f λ(p
e
−1)
e
M )DR
M ). Hence, Grσλ M 6= 0 if and only if
e
M )DR
6= C e (f (λ+ε)(p
e
−1)
e
M )DR
Let us write λ(pe − 1) = i1 + . . . + ie pe−1 , with ij ∈ {0, . . . , p − 1}, and assume
that λ(pa − 1) is a zero of the Bernstein-Sato polynomial. Then, by (left)exactnes
of F e ! , we conclude that
C e (f i1 +...+ie p
e−1
e
M )DR
6= C e (f 1+i1 +...+ie p
In particular, Grσλ M is non-zero as claimed.
e−1
e
M )DR
⊇ C e (f (λ+ε)(p
e
−1)
e
M )DR
.
THE ASSOCIATED GRADED OF THE TEST MODULE FILTRATION
21
5.4. Example. The following example illustrates that the other implication of
a
Theorem 5.3 is not true. Take M = k[x] x1 , f = x and λ = 1. Then κa xp −1 xM =
a
a
a
κa xp −1 M = k[x] while κa xp −1 xε(p −1) M = (x) as soon as ε(pa − 1) ≥ 2.
5.5. Remark. (a) If (M, κ) is F -regular then all non-nilpotent information is ree
covered. Indeed, for suitable e 0 one has σ(M, f λ+ε ) = κe f (λ+ε)(p −1) M =
λ+ε
1−λ
e
τ (M, f λ+ε− pe ), and κe f λ(p −1) f M = τ (M, f λ+ pe ) using Theorem 2.6. By
right-continuity of τ these two quantities coincide for 0 < ε 1 and e 0
suitable.
(b) Note that in characteristic zero one has that if λ is a jumping number of the
multiplier ideal filtration then it is a zero of the Bernstein-Sato polynomial.
Here the situation is the other way around.
References
[BB11]
M. Blickle and G. Böckle, Cartier modules: Finiteness results, J. reine angew.
Math. 661 (2011), 85–123.
, Cartier crystals, arXiv:1309.1035v1 (2013).
[BB13]
[BH98]
W. Bruns and J. Herzog, Cohen-Macaulay rings, revised edition, Cambridge University Press, 1998.
[Bit15]
T. Bitoun, On a theory of the b-function in positive characteristic,
arXiv:1501.00185 (2015).
[Bli13]
M. Blickle, Test ideals via p−e -linear maps, J. Algebraic Geom. 22 (2013), no. 1,
49–83.
[BMS09]
M. Blickle, M. Mustaţă, and K. Smith, F -thresholds of hypersurfaces, Trans. Amer.
Math. Soc. 361 (2009), no. 12, 6549–6565.
[BS16a]
M. Blickle and A. Stäbler, Bernstein-Sato polynomials and test modules in positive
characteristic, Nagoya Math. J. 222 (2016), no. 1, 74–99.
, Functorial Test Modules, preprint, arXiv:1605.09517 (2016).
[BS16b]
[Cha74]
S. U. Chase, On the homological dimension of algebras of differential operators,
Comm. Algebra 1 (1974), no. 5, 351–363.
[CHSW16]
E. Canton, D. J. Hernandez, K. Schwede, and E. E. Witt, On the behavior of
singularities at the F -pure threshold, arXiv:1508.05427 (2016).
[EK04]
M. Emerton and M. Kisin, The Riemann-Hilbert correspondence for unit F crystals, Asterisque 293 (2004).
[FST11]
Osamu Fujino, Karl Schwede, and Shunsuke Takagi, Supplements to non-lc ideal
sheaves, Higher dimensional algebraic geometry, RIMS Kôkyûroku Bessatsu, B24,
Res. Inst. Math. Sci. (RIMS), Kyoto, 2011, pp. 1–46. MR 2809647
[GD67]
A. Grothendieck and J. Dieudonné, Eléments de géométrie algébrique IV, Quatrième partie, Inst. Hautes Études Sci. Publ. Math., 1967.
[HNnBWZ16] D. J. Hernández, L. Núñez Betancourt, E. E. Witt, and W. Zhang, F -pure thresholds of homogeneous polynomials, Michigan Math. J. 65 (2016), no. 1, 57–87.
[HSZ14]
Jen-Chieh Hsiao, Karl Schwede, and Wenliang Zhang, Cartier modules on toric
varieties, Trans. Amer. Math. Soc. 366 (2014), no. 4, 1773–1795. MR 3152712
[HY03]
N. Hara and K. Yoshida, A generalization of tight closure and multiplier ideals,
Trans. Amer. Math. Soc. 355 (2003), no. 8, 3143–3174.
[Kun69]
E. Kunz, Characterizations of regular local rings of characteristic p, Amer. J. Math.
91 (1969), 772–784.
[Luc78]
E. Lucas, Théorie des Fonctions Numériques Simplement Périodiques, Amer. J.
Math. 1 (1878), no. 2, 197–240.
[MTW04]
M. Mustaţă, S. Takagi, and K.-I. Watanabe, F -thresholds and Bernstein-Sato polynomials, Proceedings of the 4th European congress of mathematics (ECM) (2004),
341–364.
[Mül16]
S. Müller,
The F-pure threshold of quasi-homogeneous polynomials,
arXiv:1601.08086 (2016).
22
[Mus09]
[Sch09]
[Sch14]
[Sch16]
[Stä16]
[Stä17]
AXEL STÄBLER
M. Mustaţă, Bernstein-Sato polynomials in positive characteristic, J. Algebra 321
(2009), no. 1, 128–151.
K. Schwede, F -adjunction, Algebra Number Theory 3 (2009), no. 8, 907–950.
MR 2587408
Karl Schwede, A canonical linear system associated to adjoint divisors in characteristic p > 0, J. Reine Angew. Math. 696 (2014), 69–87. MR 3276163
T. Schedlmeier, Cartier crystals and perverse constructible étale p-torsion sheaves,
PhD thesis (2016).
A. Stäbler, V -filtrations in positive characteristic and test modules, Trans. Amer.
Math. Soc. 368 (2016), no. 11, 7777–7808.
, Test module filtrations for unit F -modules, to appear J. Algebra (2017).
Axel Stäbler, Department of Mathematics, University of Michigan, Ann Arbor, MI
48109, USA
E-mail address: [email protected]
| 0math.AC
|
Published as a conference paper at ICLR 2016
D ELVING D EEPER INTO C ONVOLUTIONAL N ETWORKS
FOR L EARNING V IDEO R EPRESENTATIONS
Nicolas Ballas1 , Li Yao1 , Chris Pal2 , Aaron Courville1
MILA, Université de Montréal.
2
École Polytechnique de Montŕeal.
1
arXiv:1511.06432v4 [cs.CV] 1 Mar 2016
A BSTRACT
We propose an approach to learn spatio-temporal features in videos from intermediate visual representations we call “percepts” using Gated-Recurrent-Unit Recurrent Networks (GRUs). Our method relies on percepts that are extracted from
all levels of a deep convolutional network trained on the large ImageNet dataset.
While high-level percepts contain highly discriminative information, they tend to
have a low-spatial resolution. Low-level percepts, on the other hand, preserve a
higher spatial resolution from which we can model finer motion patterns.
Using low-level percepts, however, can lead to high-dimensionality video representations. To mitigate this effect and control the number of parameters, we
introduce a variant of the GRU model that leverages the convolution operations
to enforce sparse connectivity of the model units and share parameters across the
input spatial locations.
We empirically validate our approach on both Human Action Recognition and
Video Captioning tasks. In particular, we achieve results equivalent to state-of-art
on the YouTube2Text dataset using a simpler caption-decoder model and without
extra 3D CNN features.
1
I NTRODUCTION
Video analysis and understanding represents a major challenge for computer vision and machine
learning research. While previous work has traditionally relied on hand-crafted and task-specific
representations (Wang et al., 2011; Sadanand & Corso, 2012), there is a growing interest in designing
general video representations that could help solve tasks in video understanding such as human
action recognition, video retrieval or video captionning (Tran et al., 2014).
Two-dimensional Convolutional Neural Networks (CNN) have exhibited state-of-art performance in
still image tasks such as classification or detection (Simonyan & Zisserman, 2014b). However, such
models discard temporal information that has been shown to provide important cues in videos (Wang
et al., 2011). On the other hand, recurrent neural networks (RNN) have demonstrated the ability
to understand temporal sequences in various learning tasks such as speech recognition (Graves &
Jaitly, 2014) or machine translation (Bahdanau et al., 2014). Consequently, Recurrent Convolution
Networks (RCN) (Srivastava et al., 2015; Donahue et al., 2014; Ng et al., 2015) that leverage both
recurrence and convolution have recently been introduced for learning video representation. Such
approaches typically extract “visual percepts” by applying a 2D CNN on the video frames and then
feed the CNN activations to an RNN in order to characterize the video temporal variation.
Previous works on RCNs has tended to focus on high-level visual percepts extracted from the 2D
CNN top-layers. CNNs, however, hierarchically build-up spatial invariance through pooling layers (LeCun et al., 1998; Simonyan & Zisserman, 2014b) as Figure 2 highlights. While CNNs tends
to discard local information in their top layers, frame-to-frame temporal variation is known to be
smooth. The motion of video patches tend to be restricted to a local neighborhood (Brox & Malik,
2011). For this reason, we argue that current RCN architectures are not well suited for capturing fine
motion information. Instead, they are more likely focus on global appearance changes such as shot
transitions. To address this issue, we introduce a novel RCN architecture that applies an RNN not
solely on the 2D CNN top-layer but also on the intermediate convolutional layers. Convolutional
1
Published as a conference paper at ICLR 2016
…
…
…
…
…
…
Figure 1: Visualization of convolutional maps on successive frames in video. As we go up in the
CNN hierarchy, we observe that the convolutional maps are more stable over time, and thus discard
variation over short temporal windows.
layer activations, or convolutional maps, preserve a finer spatial resolution of the input video from
which local spatio-temporal patterns are extracted.
Applying an RNN directly on intermediate convolutional maps, however, inevitably results in a drastic number of parameters characterizing the input-to-hidden transformation due to the convolutional
maps size. On the other hand, convolutional maps preserve the frame spatial topology. We propose
to leverage this topology by introducing sparsity and locality in the RNN units to reduce the memory
requirement. We extend the GRU-RNN model (Cho et al., 2014) and replace the fully-connected
RNN linear product operation with a convolution. Our GRU-extension therefore encodes the locality
and temporal smoothness prior of videos directly in the model structure.
We evaluate our solution on UCF101 human action recognition from Soomro et al. (2012) as well as
the YouTube2text video captioning dataset from Chen & Dolan (2011). Our experiments show that
leveraging “percepts” at multiple resolutions to model temporal variation improves performance
over our baseline model with respective gains of 3.4% for action recognition and 10% for video
captioning.
2
GRU: G ATED R ECURRENT U NIT N ETWORKS
In this section, we review Gated-Recurrent-Unit (GRU) networks which are a particular type of
RNN. An RNN model is applied to a sequence of inputs, which can have variable lengths. It defines
a recurrent hidden state whose activation at each time is dependent on that of the previous time.
Specifically, given a sequence X = (x1 , x2 , ..., xT ), the RNN hidden state at time t is defined as
ht = φ(ht−1 , xt ), where φ is a nonlinear activation function. RNNs are known to be difficult to train
due to the exploding or vanishing gradient effect (Bengio et al., 1994). However, variants of RNNs
such as Long Short Term Memory (LSTM) (Hochreiter & Schmidhuber, 1997) or Gated Recurrent
Units (GRU) (Cho et al., 2014) have empirically demonstrated their ability to model long-term
temporal dependency in various task such as machine translation or image/video caption generation.
In this paper, we will mainly focus on GRU networks as they have shown similar performance to
LSTMs but with a lower memory requirement (Chung et al., 2014).
GRU networks allow each recurrent unit to adaptively capture dependencies of different time scales.
The activation ht of the GRU is defined by the following equations:
zt
rt
= σ(Wz xt + Uz ht−1 ),
= σ(Wr xt + Ur ht−1 ),
h̃t
=
tanh(Wxt + U(rt
ht
=
(1 − zt )ht−1 + zt h̃t ,
2
(1)
(2)
ht−1 ),
(3)
(4)
Published as a conference paper at ICLR 2016
…
…
…
…
…
…
…
…
…
Figure 2: High-level visualization of our model. Our approach leverages convolutional maps from
different layers of a pretrained-convnet. Each map is given as input to a convolutional GRU-RNN
(hence GRU-RCN) at different time-step. Bottom-up connections may be optionally added between
RCN layers to form Stack-GRU-RCN.
where is an element-wise multiplication. zt is an update gate that decides the degree to which
the unit updates its activation, or content. rt is a reset gate. σ is the sigmoid function. When a unit
rti is close to 0, the reset gate forgets the previously computed state, and makes the unit act as if
it is reading the first symbol of an input sequence. h̃t is a candidate activation which is computed
similarly to that of the traditional recurrent unit in an RNN.
3
D ELVING D EEPER INTO C ONVOLUTIONAL N EURAL N ETWORKS
This section delves into the main contributions of this work. We aim at leveraging visual percepts
from different convolutional levels in order to capture temporal patterns that occur at different spatial
resolution.
, xL
Let’s consider (x1t , ..., xL−1
t
t )(t=1..T ) , a set of 2D convolutional maps extracted from L layers
at different time steps in a video. We propose two alternative RCN architectures, GRU-RCN, and
Stacked-GRU-RCN (illustrated in Figure 2) that combines information extracted from those convolutional maps.
3.1
GRU-RCN:
In the first RCN architecture, we propose to apply L RNNs independently on each convolutional
map. We define L RNNs as φ1 , ..., φL , such that hlt = φl (xlt , hlt−1 ). The hidden representation of
the final time step h1T , ..., hL
T are then fed to a classification layer in the case of action recognition,
or to a text-decoder RNN for caption generation.
To implement the RNN recurrent function φl , we propose to leverage Gated Recurrent Units (Cho
et al., 2014). GRUs were originally introduced for machine translation. They model input to hiddenstate and hidden to hidden transitions using fully connected units. However, convolutional map
inputs are 3D tensors (spatial dimension and input channels). Applying a GRU directly can lead to
a drastic number of parameters. Let N1 , N2 and Ox be the input convolutional map spatial size and
number of channels. Applying a GRU directly would require input-to-hidden parameters Wl , Wzl
and Wrl to be of size N1 × N2 × Ox × Oh where Oh is the dimensionality of the GRU hidden
representation.
3
Published as a conference paper at ICLR 2016
Fully-connected GRUs do not take advantage of the underlying structure of convolutional maps.
Indeed, convolutional maps are extracted from images that are composed of patterns with strong
local correlation which are repeated over different spatial locations. In addition, videos have smooth
temporal variation over time, i.e. motion associated with a given patch in successive frames will be
restricted in a local spatial neighborhood. We embed such a prior in our model structure and replace
the fully-connected units in GRU with convolution operations. We therefore obtain recurrent units
that have sparse connectivity and share their parameters across different input spatial locations:
zlt
rlt
h̃lt
hlt
= σ(Wzl ∗ xlt + Ulz ∗ hlt−1 ),
=
=
=
σ(Wrl
xlt
l
∗ +
∗ hlt−1 ),
tanh(W ∗ xlt + U ∗ (rlt
(1 − zlt )hlt−1 + zlt h̃lt ,
(5)
Ulr
(6)
hlt−1 ),
(7)
(8)
where ∗ denotes a convolution operation. In this formulation, Model parameters W, Wzl , Wrl and
Ul , Ulz , Ulr are 2D-convolutional kernels. Our model results in hidden recurrent representation that
preserves the spatial topology, hlt = (hlt (i, j)) where hlt (i, j)) is a feature vector defined at the
location (i, j). To ensure that the spatial size of the hidden representation remains fixed over time,
we use zero-padding in the recurrent convolutions.
Using convolution, parameters Wl , Wzl and Wrl have a size of k1 ×k2 ×Ox ×Oh where k1 ×k2 is the
convolutional kernel spatial size (usually 3 × 3), chosen to be significantly lower than convolutional
map size N1 × N2 . The candidate hidden representation h̃t (i, j), the activation gate zk (i, j) and the
reset gate rk (i, j) are defined based on a local neigborhood of size (k1 × k2 ) at the location(i, j)
in both the input data xt and the previous hidden-state ht−1 . In addition, the size of receptive field
associated with hl (i, j)t increases in the previous presentation hlt−1 , hlt−2 ... as we go back further
in time. Our model is therefore capable of characterizing spatio-temporal patterns with high spatial
variation in time.
A GRU-RCN layer applies 6 2D-convolutions at each time-step (2 per GRU gate and 2 for computing the candidate activation). If we assume for simplicity that the input-to-hidden and hidden-tohidden convolutions have the same kernel size and perserve the input dimension, GRU-RCN requires
O(3T N1 N2 k1 k2 (Ox Oh + Oh Oh )) multiplications. GRU-RCN sparse connectivity therefore saves
computation compared to a fully-connected RNN that would require O(3T N1 N2 N1 N2 (Ox Oh +
Oh Oh )) computations. Memorywise, GRU-RCN needs to store the parameters for all 6 convolutions kernels leading to O(3k1 k2 (Ox Oh + Oh Oh )) parameters.
3.2
S TACKED GRU-RCN:
In the second RCN architecture, we investigate the importance of bottom-up connection across
RNNs. While GRU-RCN applies each layer-wise GRU-RNN in an independent fashion, Stacked
GRU-RCN preconditions each GRU-RNN on the output of the previous GRU-RNN at the current
l
time step: hlt = φl (hlt−1 , hl−1
t , xt ). The previous RNN hidden representation is given as an extrainput to the GRU convolutional units:
zll
rlt
h̃lt
hlt
= σ(Wzl ∗ xlt + Wzl l ∗ hl−1
+ Ulz ∗ hlt−1 ),
t
=
=
=
σ(Wrl
∗
xlt
l
tanh(W ∗
(1 −
+
Wrl l
xlt
zlt )hlt−1
∗
hl−1
t
l
+U ∗
+
zlt h̃lt ,
+ Ulr hlt−1 ),
(rt hlt−1 ),
(9)
(10)
(11)
(12)
Adding this extra-connection brings more flexibility and gives the opportunity for the model to
leverage representations with different resolutions.
4
R ELATED W ORK
Deep learning approaches have recently been used to learn video representations and have produced
state-of-art results (Karpathy et al., 2014; Simonyan & Zisserman, 2014a; Wang et al., 2015b; Tran
et al., 2014). Karpathy et al. (2014); Tran et al. (2014) proposed to use 3D CNN learn a video representations, leveraging large training datasets such as the Sport 1 Million. However, unlike image
4
Published as a conference paper at ICLR 2016
classification (Simonyan & Zisserman, 2014b), CNNs did not yield large improvement over these
traditional methods (Lan et al., 2014) highlighting the difficulty of learning video representations
even with large training dataset. Simonyan & Zisserman (2014a) introduced a two-stream framework where they train CNNs independently on RGB and optical flow inputs. While the flow stream
focuses only on motion information, the RGB stream can leverage 2D CNN pre-trained on image
datasets. Based on the Two Stream representation, Wang et al. (2015a) extracted deep feature and
conducted trajectory constrained pooling to aggregate convolutional feature as video representations.
RNN models have also been used to encode temporal information for learning video representations
in conjonction with 2D CNNs. Ng et al. (2015); Donahue et al. (2014) applied an RNN on top of the
the two-stream framework, while Srivastava et al. (2015) proposed, in addition, to investigate the
benefit of learning a video representation in an unsupervised manner. Previous works on this topic
has tended to focus only on high-level CNN “visual percepts”. In contrast, our approach proposes
to leverage visual “percepts” extracted from different layers in the 2D-CNN.
Recently, Shi et al. (2015) also proposed to leverage convolutional units inside an RNN network.
However, they focus on different task (now-casting) and a different RNN model based on an LSTM.
In addition, they applied their model directly on pixels. Here, we use recurrent convolutional units
on pre-trained CNN convolutional maps, to extract temporal pattern from visual “percepts” with
different spatial sizes.
5
E XPERIMENTATION
This section presents an empirical evaluation of the proposed GRU-RCN and Stacked GRU-RCN
architectures. We conduct experimentations on two different tasks: human action recognition and
video caption generation.
5.1
ACTION R ECOGNITION
We evaluate our approach on the UCF101 dataset Soomro et al. (2012). This dataset has 101 action
classes spanning over 13320 YouTube videos clips. Videos composing the dataset are subject to
large camera motion, viewpoint change and cluttered backgrounds. We report results on the dataset
UCF101 first split, as this is most commonly used split in the literature. To perform proper hyperparameter seach, we use the videos from the UCF-Thumos validation split Jiang et al. (2014) as the
validation set.
5.1.1
M ODEL A RCHITECTURE
In this experiment, we consider the RGB and flow representations of videos as inputs. We extract
visual “percept” using VGG-16 CNNs that consider either RGB or flow inputs. VGG-16 CNNs are
pretrained on ImageNet (Simonyan & Zisserman, 2014b) and fine-tuned on the UCF-101 dataset,
following the protocol in Wang et al. (2015b). We then extract the convolution maps from pool2,
pool3, pool4, pool5 layers and the fully-connected map from layer fc-7 (which can be view as a
feature map with a 1 × 1 spatial dimension). Those features maps are given as inputs to our RCN
models.
We design and evaluate three RCN architectures for action recognition. In the first RCN architecture,
GRU-RCN, we apply 5 convolutional GRU-RNNs independently on each convolutional map. Each
convolution in the GRU-RCN has zero-padded 3 × 3 convolutions that preserves the spatial dimension of the inputs . The number of channels of each respective GRU-RNN hidden-representations
are 64, 128, 256, 256, 512. After the RCN operation we obtain 5 hidden-representations for each
time step. We apply average pooling on the hidden-representations of the last time-step to reduce
their spatial dimension to 1 × 1, and feed the representations to 5 classifiers, composed by a linear
layer with a softmax nonlineary. Each classifier therefore focuses on only 1 hidden-representation
extracted from the convolutional map of a specific layer. The classifier outputs are then averaged to
get the final decision. A dropout ratio of 0.7 is applied on the input of each classifiers.
In the second RCN architecture, Stacked GRU-RCN, we investigate the usefulness of bottom-up
connections. Our stacked GRU-RCN uses the same base architecture as the GRU-RCN, consisting of 5 convolutional GRU-RNNs having 64, 128, 256, 256 channels respectively. However, each
5
Published as a conference paper at ICLR 2016
convolutional GRU-RNN is now preconditioned on the hidden-representation that the GRU-RNN
applied on the previous convolution-map outputs. We apply max-pooling on the hidden representations between the GRU-RNN layers for the compatibility of the spatial dimensions. As for the
previous architecture, each GRU-RNN hidden-representation at the last time step is pooled and then
given as input to a classifier.
Finally, in our bi-directional GRU-RCN, we investigate the importance of reverse temporal information. Given convolutional maps extracted from one layer, we run the GRU-RCN twice, considering
the inputs in both sequential and reverse temporal order. We then concatenate the last hiddenrepresentations of the foward GRU-RCN and backward GRU-RCN, and give the resulting vector to
a classifier.
5.1.2
M ODEL T RAINING AND E VALUATION
We follow the training procedure introduced by the two-stream framework Simonyan & Zisserman (2014a). At each iteration, a batch of 64 videos are sampled randomly from the the training set. To perform scale-augmentation, we randomly sample the cropping width and height from
256, 224, 192, 168. The temporal cropping size is set to 10. We then resize the cropped volume to
224 × 224 × 10. We estimate each model parameters by maximizing the model log-likelihood:
L(θ) =
N
1 X
log p(y n | c(xn ), θ),
N n=1
where there are N training video-action pairs (xn , y n ), c is a function that takes a crop at random.
We use Adam Kingma & Ba (2014) with the gradient computed by the backpropagation algorithm.
We perform early stopping and choose the parameters that maximize the log-probability of the validation set.
We also follow the evaluation protocol of the two-stream framework Simonyan & Zisserman
(2014a). At the test time, we sample 25 equally spaced video sub-volumes with a temporal size
of 10 frames. From each of these selected sub-volumes, we obtain 10 inputs for our model, i.e. 4
corners, 1 center, and their horizontal flipping. The final pre- diction score is obtained by averaging
across the sampled sub-volumes and their cropped regions.
5.1.3
R ESULTS
Method
VGG-16
VGG-16 RNN
GRU-RCN
Stacked-GRU RCN
Bi-directional GRU-RCN
Two-Stream Simonyan & Zisserman (2014b)
Two-Stream + LSTM Donahue et al. (2014)
Two-Stream + LSTM + Unsupervised Srivastava et al. (2015)
Improved Two-Stream Wang et al. (2015b)
C3D one network Tran et al. (2014), 1 million videos as training
C3D ensemble Tran et al. (2014), 1 million videos as training
Deep networks Karpathy et al. (2014), 1 million videos as training
RGB
78.0
78.1
79.9
78.3
80.7
72.8
71.1
77.7
79.8
82.3
85.2
65.2
Flow
85.4
84.9
85.7
81.2
76.9
83.7
85.7
-
Table 1: Classification accuracy of different variants of the model on the UCF101 split 1. We report
the performance of previous works that learn representation using only RGB information only.
We compare our approach with two different baselines, VGG-16 and VGG-16 RNN. VGG-16 is
the 2D spatial stream that is described in Wang et al. (2015b). We take the VGG-16 model, pretrained on Image-Net and fine-tune it on the UCF-101 dataset. VGG-16 RNN baseline applied an
RNN, using fully-connected gated-recurrent units, on top-of VGG-16. It takes as input the VGG-16
fully-connected representation fc-7. Following GRU-RCN top-layer, the VGG-16 RNN has hiddenrepresentation dimensionality of 512.
6
Published as a conference paper at ICLR 2016
The first column of Table 1 focuses on RGB inputs. We first report results of different GRU-RCN
variants and compare them with the two baselines: VGG-16 and VGG-16 RNN. Our GRU-RCN
variants all outperform the baselines, showing the benefit of delving deeper into a CNN in order to
learn a video representation. We notice that VGG-16 RNN only slightly improve over the VGG16 baseline, 78.1 against 78.0. This result confirms that CNN top-layer tends to discard temporal
variation over short temporal windows. Stacked-GRU RCN performs significantly lower than GRURCN and Bi-directional GRU-RCN. We argue that bottom-up connection, increasing the depth of
the model, combined with the lack of training data (UCF-101 is train set composed by only 9500
videos) make the Stacked-GRU RCN learning difficult. The bi-directional GRU-RCN performs the
best among the GRU-RCN variant with an accuracy of 80.7, showing the advantage of modeling
temporal information in both sequential and reverse order. Bi-directional GRU-RCN obtains a gain
3.4% in term of performances, relatively to the baselines that focus only the VGG-16 top layer.
Table 1 also reports results from other state-of-art approaches using RGB inputs. C3D Tran et al.
(2014) obtains the best performance on UCF-101 with 85.2. However, it should be noted that C3D is
trained over 1 million videos. Other approaches use only the 9500 videos of UCF101 training set for
learning temporal pattern. Our Bi-directional GRU-RCN compare favorably with other Recurrent
Convolution Network (second blocks), confirming the benefit of using different CNN layers to model
temporal variation.
Table 1 also evaluates the GRU-RCN model applied flow inputs. VGG-16 RNN baseline actually
decreases the performance compared to the VGG-16 baseline. On the other hand, GRU-RCN outperforms the VGG-16 baseline achieving 85.7 against 85.4. While the improvement is less important
than the RGB stream, it should be noted that the flow stream of VGG-16 is applied on 10 consecutive
flow inputs to extract visual “percepts”, and therefore already captures some motion information.
Finally, we investigate the combination of the RGB and flow streams. Following Wang et al. (2015b),
we use a weighted linear combination of their prediction scores, where the weight is set to 2 as for
the flow stream net and 1 for the temporal stream. Fusion the VGG-16 model baseline achieve an
accuracy of 89.1. Combining the RGB Bi-directional GRU-RCN with the flow GRU-RCN achieves
a performance gain of 1.9% over baseline, reaching 90.8. Our model is on part with Wang et al.
(2015b) that obtain state-of-art results using both RGB and flow streams which obtains 90.9.
5.2
V IDEO C APTIONING
We also evaluate our representation on the video captioning task using YouTube2Text video corpus Chen & Dolan (2011). The dataset has 1,970 video clips with multiple natural language descriptions for each video clip. The dataset is open-domain and covers a wide range of topics such
as sports, animals, music and movie clips. Following Yao et al. (2015b), we split the dataset into a
training set of 1,200 video clips, a validation set of 100 clips and a test set consisting of the remaining
clips.
5.2.1
M ODEL S PECIFICATIONS
To perform video captioning, we use the so-called encoder-decoder framework Cho et al. (2014).
In this framework the encoder maps input videos into abstract representations that precondition a
caption-generating decoder.
As for encoder, we compare both VGG-16 CNN and Bi-directional GRU-RCN. Both models have
been fine-tuned on the UCF-101 dataset and therefore focus on detecting actions. To extract an
abstract representation from a video, we sample K equally-space segments. When using the VGG16 encoder, we provide the f c7 layer activations of the each segment’s first frame as the input to
the text-decoder. For the GRU-RCN, we apply our model on the segment’s 10 first frames. We
concatenate the GRU-RCN hidden-representation from the last time step. The concatenated vector
is given as the input to the text decoder. As it has been shown that characterizing entities in addition
of action is important for the caption-generation task Yao et al. (2015a), we also use as encoder a
CNN Szegedy et al. (2014), pretrained on ImageNet, that focuses on detecting static visual object
categories.
As for the decoder, we use an LSTM text-generator with soft-attention on the video temporal
frames Yao et al. (2015b).
7
Published as a conference paper at ICLR 2016
YouTube2Text
Model
VGG-16 Encoder
Bi-directional GRU-RCN Encoder
GoogleNet Encoder
GoogleNet + Bi-directional GRU-RCN Encoder
GoogleNet + Bi-directional GRU-RCN Encoder
GoogleNet + Bi-directional GRU-RCN Encoder
GoogleNet + Bi-directional GRU-RCN Encoder
GoogleNet + HRNE (Pan et al., 2015)
VGG + p-RNN (Yu et al., 2015)
VGG + C3D + p-RNN (Yu et al., 2015)
Soft-attention Yao et al. (2015b)
Venugopalan et al. Venugopalan et al. (2015)
+ Extra Data (Flickr30k, COCO)
Thomason et al. Thomason et al. (2014)
model selection
BLEU
BLEU
BLEU
BLEU
NLL
METEOR
CIDEr
-
BLEU
METEOR
CIDEr
0.3700
0.4100
0.4128
0.4963
0.4790
0.4842
0.4326
0.436
0.443
0.499
0.4192
0.3119
0.3329
0.1368
0.2640
0.2850
0.2900
0.3075
0.3114
0.3170
0.3160
0.321
0.311
0.326
0.2960
0.2687
0.2907
0.2390
0.4330
0.5010
0.4804
0.5937
0.6782
0.6538
0.6801
0.5167
-
Table 2: Performance of different variants of the model on YouTube2Text for video captioning.
Representations obtained with the proposed RCN architecture combined with decoders from Yao
et al. (2015b) offer a significant performance boost, reaching the performance of the other state-ofthe-art models.
5.2.2
T RAINING
For all video captioning models, we estimated the parameters θ of the decoder by maximizing the
log-likelihood:
L(θ) =
N tn
1 XX
n
log p(yin | y<i
, xn , θ),
N n=1 i=1
where there are N training video-description pairs (xn , y n ), and each description y n is tn words
long. We used Adadelta Zeiler (2012) We optimized the hyperparameters (e.g. number of LSTM
units and the word embedding dimensionality, number of segment K) using random search (Bergstra
& Bengio, 2012) to maximize the log-probability of the validation set.
5.2.3
R ESULTS
Table 2 reports the performance of our proposed method using three automatic evaluation metrics.
These are BLEU in Papineni et al. (2002), METEOR in Denkowski & Lavie (2014) and CIDEr
in Vedantam et al. (2014). We use the evaluation script prepared and introduced in Chen et al.
(2015). All models are early-stopped based on the negative-log-likelihood (NLL) of the validation
set. We then select the model that performs best on the validation set according to the metric at
consideration.
The first two lines of Table 2 compare the performances of the VGG-16 and Bi-directional GRURCN encoder. Results clearly show the superiority of the Bi-Directional GRU-RCN Encoder as it
outperforms the VGG-16 Encoder on all three metrics. In particular, GRU-RCN Encoder obtains a
performance gain of 10% compared to the VGG-16 Encoder according to the BLEU metric. Combining our GRU-RCN Encoder that focuses on action with a GoogleNet Encoder that captures visual
entities further improve the performances.
Our GoogleNet + Bi-directional GRU-RCN approach significantly outperforms Soft-attention Yao
et al. (2015b) that relies on a GoogLeNet and cuboids-based 3D-CNN Encoder, in conjunction to a
similar soft-attention decoder. This result indicates that our approach is able to offer more effective
representations. According to the BLEU metric, we also outperform other approaches using more
complex decoder schemes such as spatial and temporal attention decoder (Yu et al., 2015) or a
hierarchical RNN decoder (Pan et al., 2015) Our approach is on par with Yu et al. (2015), without
the need of using a C3D-encoder that requires training on large-scale video dataset.
8
Published as a conference paper at ICLR 2016
6
C ONCLUSION
In this work, we address the challenging problem of learning discriminative and abstract representations from videos. We identify and underscore the importance of modeling temporal variation from
“visual percepts” at different spatial resolutions. While high-level percepts contain highly discriminative information, they tend to have a low-spatial resolution. Low-level percepts, on the other hand,
preserve a higher spatial resolution from which we can model finer motion patterns. We introduce a
novel recurrent convolutional network architecture that leverages convolutional maps, from all levels of a deep convolutional network trained on the ImageNet dataset, to take advantage of “percepts”
from different spatial resolutions.
We have empirically validated our approach on the Human Action Recognition and Video Captioning tasks using the UCF-101 and YouTube2Text datasets. Experiments demonstrate that leveraging
“percepts” at multiple resolutions to model temporal variation improve over our baseline model,
with respective gain of 3.4% and 10% for the action recognition and video captions tasks using
RGB inputs. In particular, we achieve results comparable to state-of-art on YouTube2Text using a
simpler text-decoder model and without extra 3D CNN features.
ACKNOWLEDGMENTS
The authors would like to acknowledge the support of the following agencies for research funding
and computing support: NSERC, Calcul Québec, Compute Canada, the Canada Research Chairs
and CIFAR. We would also like to thank the developers of Theano (Bergstra et al., 2010; Bastien
et al., 2012) , for developing such a powerful tool for scientific computing.
R EFERENCES
Bahdanau, D., Cho, K., and Bengio, Y. Neural machine translation by jointly learning to align and
translate. arXiv preprint arXiv:1409.0473, 2014.
Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Bergstra, James, Goodfellow, Ian J., Bergeron,
Arnaud, Bouchard, Nicolas, and Bengio, Yoshua. Theano: new features and speed improvements.
Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop, 2012.
Bengio, Y., Simard, P., and Frasconi, P. Learning long-term dependencies with gradient descent is
difficult. Neural Networks, IEEE Transactions on, 1994.
Bergstra, James and Bengio, Yoshua. Random search for hyper-parameter optimization. JMLR,
2012.
Bergstra, James, Breuleux, Olivier, Bastien, Frédéric, Lamblin, Pascal, Pascanu, Razvan, Desjardins, Guillaume, Turian, Joseph, Warde-Farley, David, and Bengio, Yoshua. Theano: a CPU
and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), 2010.
Brox, T. and Malik, J. Large displacement optical flow: descriptor matching in variational motion
estimation. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2011.
Chen, David L and Dolan, William B. Collecting highly parallel data for paraphrase evaluation. In
Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, pp. 190–200. Association for Computational Linguistics,
2011.
Chen, Xinlei, Fang, Hao, Lin, Tsung-Yi, Vedantam, Ramakrishna, Gupta, Saurabh, Dollar, Piotr,
and Zitnick, C Lawrence. Microsoft coco captions: Data collection and evaluation server. arXiv
1504.00325, 2015.
Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., and Bengio,
Y. Learning phrase representations using rnn encoder-decoder for statistical machine translation.
arXiv preprint arXiv:1406.1078, 2014.
9
Published as a conference paper at ICLR 2016
Chung, Junyoung, Gulcehre, Caglar, Cho, KyungHyun, and Bengio, Yoshua. Empirical evaluation
of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.
Denkowski, Michael and Lavie, Alon. Meteor universal: Language specific translation evaluation
for any target language. In EACL Workshop, 2014.
Donahue, J., Hendricks, L., Guadarrama, S., Rohrbach, M., Venugopalan, S., Saenko, K., and Darrell, T. Long-term recurrent convolutional networks for visual recognition and description. arXiv
preprint arXiv:1411.4389, 2014.
Graves, A. and Jaitly, N. Towards end-to-end speech recognition with recurrent neural networks. In
ICML, 2014.
Hochreiter, Sepp and Schmidhuber, Jürgen. Long short-term memory. Neural computation, 1997.
Jiang, YG, Liu, J, Roshan Zamir, A, Toderici, G, Laptev, I, Shah, M, and Sukthankar, R. Thumos
challenge: Action recognition with a large number of classes. Technical Report, 2014.
Karpathy, Andrej, Toderici, George, Shetty, Sachin, Leung, Tommy, Sukthankar, Rahul, and Fei-Fei,
Li. Large-scale video classification with convolutional neural networks. In CVPR. IEEE, 2014.
Kingma, Diederik and Ba, Jimmy. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Lan, Zhenzhong, Lin, Ming, Li, Xuanchong, Hauptmann, Alexander G, and Raj, Bhiksha. Beyond gaussian pyramid: Multi-skip feature stacking for action recognition. arXiv preprint
arXiv:1411.6660, 2014.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 1998.
Ng, Joe Yue-Hei, Hausknecht, Matthew, Vijayanarasimhan, Sudheendra, Vinyals, Oriol, Monga,
Rajat, and Toderici, George. Beyond short snippets: Deep networks for video classification.
arXiv preprint arXiv:1503.08909, 2015.
Pan, Pingbo, Xu, Zhongwen, Yang, Yi, Wu, Fei, and Zhuang, Yueting. Hierarchical recurrent neural
encoder for video representation with application to captioning. arXiv preprint arXiv:1511.03476,
2015.
Papineni, Kishore, Roukos, Salim, Ward, Todd, and Zhu, Wei-Jing. Bleu: a method for automatic
evaluation of machine translation. In ACL, 2002.
Sadanand, S. and Corso, J. Action bank: A high-level representation of activity in video. In CVPR.
IEEE, 2012.
Shi, Xingjian, Chen, Zhourong, Wang, Hao, Yeung, Dit-Yan, Wong, Wai-Kin, and Woo, Wangchun. Convolutional lstm network: A machine learning approach for precipitation nowcasting.
arXiv preprint arXiv:1506.04214, 2015.
Simonyan, Karen and Zisserman, Andrew. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, pp. 568–576, 2014a.
Simonyan, Karen and Zisserman, Andrew. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556, 2014b.
Soomro, Khurram, Zamir, Amir Roshan, and Shah, Mubarak. Ucf101: A dataset of 101 human
actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.
Srivastava, N., Mansimov, E., and Salakhutdinov, R. Unsupervised learning of video representations
using lstms. In ICML, 2015.
Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir,
Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions.
arXiv preprint arXiv:1409.4842, 2014.
10
Published as a conference paper at ICLR 2016
Thomason, Jesse, Venugopalan, Subhashini, Guadarrama, Sergio, Saenko, Kate, and Mooney, Raymond. Integrating language and vision to generate natural language descriptions of videos in the
wild. In COLING, 2014.
Tran, D., Bourdev, L., Fergus, R., Torresani, L., and Paluri, M. C3d: generic features for video
analysis. arXiv preprint arXiv:1412.0767, 2014.
Vedantam, Ramakrishna, Zitnick, C Lawrence, and Parikh, Devi. CIDEr: Consensus-based image
description evaluation. arXiv:1411.5726, 2014.
Venugopalan, Subhashini, Xu, Huijuan, Donahue, Jeff, Rohrbach, Marcus, Mooney, Raymond,
and Saenko, Kate. Translating videos to natural language using deep recurrent neural networks.
NAACL, 2015.
Wang, H., Kläser, A., Schmid, C., and Liu, C. Action recognition by dense trajectories. In CVPR.
IEEE, 2011.
Wang, Limin, Qiao, Yu, and Tang, Xiaoou. Action recognition with trajectory-pooled deepconvolutional descriptors. arXiv preprint arXiv:1505.04868, 2015a.
Wang, Limin, Xiong, Yuanjun, Wang, Zhe, and Qiao, Yu. Towards good practices for very deep
two-stream convnets. arXiv preprint arXiv:1507.02159, 2015b.
Yao, Li, Ballas, Nicolas, Cho, Kyunghyun, Smith, John R., and Bengio, Yoshua. Trainable performance upper bounds for image and video captioning. arXiv preprint arXiv:1511.0459, 2015a.
Yao, Li, Torabi, Atousa, Cho, Kyunghyun, Ballas, Nicolas, Pal, Christopher, Larochelle, Hugo,
and Courville, Aaron. Describing videos by exploiting temporal structure. In Computer Vision
(ICCV), 2015 IEEE International Conference on. IEEE, 2015b.
Yu, Haonan, Wang, Jiang, Huang, Zhiheng, Yang, Yi, and Xu, Wei. Video paragraph captioning
using hierarchical recurrent neural networks. arXiv 1510.07712, 2015.
Zeiler, Matthew D. ADADELTA: an adaptive learning rate method. Technical report, 2012.
11
| 9cs.NE
|
Memetic Graph Clustering
Sonja Biedermann1 , Monika Henzinger2 , Christian Schulz3 , and
Bernhard Schuster4
1
2
3
arXiv:1802.07034v1 [cs.NE] 20 Feb 2018
4
University of Vienna, Vienna, Austria
[email protected]
University of Vienna, Vienna, Austria
[email protected]
University of Vienna, Vienna, Austria
[email protected]
University of Vienna, Vienna, Austria
[email protected]
Abstract
It is common knowledge that there is no single best strategy for graph clustering, which justifies a
plethora of existing approaches. In this paper, we present a general memetic algorithm, VieClus,
to tackle the graph clustering problem. This algorithm can be adapted to optimize different
objective functions. A key component of our contribution are natural recombine operators that
employ ensemble clusterings as well as multi-level techniques. Lastly, we combine these techniques
with a scalable communication protocol, producing a system that is able to compute high-quality
solutions in a short amount of time. We instantiate our scheme with local search for modularity
and show that our algorithm successfully improves or reproduces all entries of the 10th DIMACS
implementation challenge under consideration using a small amount of time.
1998 ACM Subject Classification G.2.2 Graph Theory
Keywords and phrases Graph Clustering, Evolutionary Algorithms
1
Introduction
Graph clustering is the problem of detecting tightly connected regions of a graph. Depending
on the task, knowledge about the structure of the graph can reveal information such as voter
behavior, the formation of new trends, existing terrorist groups and recruitment [41] or a
natural partitioning of data records onto pages [16]. Further application areas include the
study of protein interaction [34], gene expression networks [47], fraud detection [1], program
optimization [28, 14] and the spread of epidemics [31]—possible applications are plentiful, as
almost all systems containing interacting or coexisting entities can be modeled as a graph.
It is common knowledge that there is no single best strategy for graph clustering, which
justifies a plethora of existing approaches. Moreover, most quality indices for graph clusterings have turned out to be NP-hard to optimize and are rather resilient to effective
approximations, see e.g. [4, 11, 46], allowing only heuristic approaches for optimization.
The majority of algorithms for graph clustering are based on the paradigm of intra-cluster
density versus inter-cluster sparsity. One successful heuristic to cluster large graphs is the
multi-level approach [10], e.g. the Louvain method for the optimization of modularity [9].
Here, the graph is recursively contracted to obtain smaller graphs which should reflect the
same general structure as the input. After applying an initial clustering algorithm to the
smallest graph, the contraction steps are undone and, at each level, a local search method
XX:2
Memetic Graph Clustering
is used to improve the clustering induced by the coarser level w.r.t some objective function
measuring the quality of the clustering. The intuition behind this approach is that a good
clustering at one level of the hierarchy will also be a good clustering on the next finer level.
Hence, depending on the definition of the neighborhood, local search algorithms are able
to explore local solution spaces very effectively in this setting. However, these methods are
also prone to get trapped in local optima. The multi-level scheme can help to some extent
since local search has a more global view on the problem on the coarse levels and a very
fine-grained view on the fine levels of the multi-level hierarchy. In addition, as with many
other randomized meta-heuristics, several repeated runs can be made in order to improve
the final result at the expense of running time.
Still, even a large number of repeated executions can only scratch the surface of the huge
search space of possible clusterings. In order to explore the global solution space extensively,
we need more sophisticated meta-heuristics. This is where memetic algorithms (MAs),
i.e. genetic algorithms combined with local search [24], come into play. Memetic algorithms
allow for effective exploration (global search) and exploitation (local search) of the solution
space. The general idea behind genetic algorithms is to use mechanisms inspired by biological
evolution such as selection, mutation, recombination, and survival of the fittest. A genetic
algorithm (GA) starts with a population of individuals (in our case clusterings of the graph)
and evolves the population over several generational cycles (rounds). In each round, the
GA uses a selection rule based on the fitness of the individuals of the population to select
good individuals and combines them to obtain improved offspring [20]. When an offspring
is generated an eviction rule is used to select a member of the population to be replaced
by the new offspring. For an evolutionary algorithm it is of major importance to preserve
diversity in the population [5], i.e., the individuals should not become too similar in order to
avoid a premature convergence of the algorithm. This is usually achieved by using mutation
operations and by using eviction rules that take similarity of individuals into account.
In this paper, we present a general memetic algorithm, VieClus (Vienna Graph Clustering), to tackle the graph clustering problem. This algorithm can be adapted to optimize
different objective functions simply by using a local search algorithm that optimizes the
objective function desired by the user. A key component of our contribution are natural
recombine operators that employ ensemble clusterings as well as multi-level techniques. In
machine learning, ensemble methods combine multiple weak classification (or clustering)
algorithms to obtain a strong algorithm for classification (or clustering). More precisely,
given a number of clusterings, the overlay/ensemble clustering is a clustering in which two
vertices belong to the same cluster if and only if they belong to the same cluster in each of
the input clusterings. Our recombination operators use the overlay of two clusterings from
the population to decide whether pairs of vertices should belong to the same cluster [33, 42].
This is combined with a local search algorithm to find further improvements and also embedded into a multi-level algorithm to find even better clusterings. Our general principle is
to randomize tie-breaking whenever possible. This diversifies the search and also improves
solutions. Lastly, we combine these techniques with a scalable communication protocol,
producing a system that is able to compute high-quality solutions in a short amount of
time. In our experimental evaluation, we show that our algorithm successfully improves or
reproduces all entries of the 10th DIMACS implementation challenge under consideration
in a small amount of time. In fact, for most of the small instances, we can improve the
old benchmark result in less than a minute. Moreover, while the previous best result for
different instances has been computed by a variety of solvers, our algorithm can now be used
as a single tool to compute the result.
S. Biedermann et. al.
2
2.1
Preliminaries
Basic Concepts
Let G = (V = {0, . . . , n − 1}, E) be an undirected graph and N (v) := {u : {v, u} ∈ E}
denote the neighbors of v. The degree of a vertex v is d(v) := |N (v)|. The problem that we
tackle in this paper is the graph clustering problem. A clustering C is a partition of the set of
vertices, i.e. a set of disjoint blocks/clusters of vertices V1 ,. . . ,Vk such that V1 ∪ · · · ∪ Vk = V .
However, k is usually not given in advance. A size-constrained clustering constrains the size
of the blocks of a clustering by a given upper bound U . A clustering is trivial if there is only
one block, or all clusters/blocks contain only one element, i.e., are singletons. We identify a
cluster Vi with its node-induced subgraph of G. The set E(C) := E ∩ (∪i Vi × Vi ) is the set
of intra-cluster edges, and E \ E(C) is the set of inter-cluster edges. We set |E(C)| =: m(C)
and |E\E(C)| =: m(C). An edge running between two blocks is called cut edge. There are
different objective functions that are optimized in the literature. We review some of them
in Section 2.3. Our main focus in this work is on modularity. However, our algorithm can
be generalized to optimize other objective functions. The graph partitioning problem is also
looking for a partition of the vertices. Here, a balancing constraint demands that all blocks
have weight |Vi | ≤ (1 + )d |Vk | e =: Lmax for some imbalance parameter . A vertex is a
boundary vertex if it is incident to a vertex in a different block. The objective is to minimize
S
the total cut ω(E ∩ i<j Vi × Vj ).
In general, our input graphs G have unit edge weights and vertex weights. However, even
those will be translated into weighted problems in the course of a multi-level algorithm. In
order to avoid tedious notation, G will denote the current state of the graph before and after
a (un)contraction in the multi-level scheme throughout this paper.
2.2
Ensemble/Overlay Clusterings
In machine learning, ensemble methods combine multiple weak classification algorithms to
obtain a strong classifier. These base clusterings are used to decide whether pairs of vertices
should belong to the same cluster [33, 43]. Given two clusterings, the overlay clustering is
a clustering in which two vertices belong to the same cluster if and only if they belong to
the same cluster in each of the input clusterings. More precisely, given two clusterings C1
and C2 the overlay clustering is the clustering where each block corresponds to a connected
component of the graph GE = (V, E\E) where E is the union of the cut edges of C1 and C2 ,
i.e. all edges that run between blocks in either C1 or C2 . Intuitively, if the input clusters
agree that two vertices belong to the same cluster, the two vertices belong together with
high confidence. It is easy to see that the number of clusters in the overlay clustering cannot
be smaller than the number of clusters in each of the input clusterings.
An overlay clustering can be computed in expected linear-time. More precisely, given
two clusterings {C1 , C2 }, we use the following approach to compute the overlay clustering.
Initially, the overlay clustering O is set to the clustering C1 . We then iterate through the
remaining clusterings and incrementally update the current solution O. To this end, we use
pairs of cluster IDs (i, j) as a key in a hash map H, where i is a cluster ID of O and j
is a cluster ID of the current clustering C. We initialize H to be the empty hashmap and
set a counter c to zero. Then we iterate through the nodes. Let v be the current node.
If the pair (O[v], C[v]) is not contained in H, we set H(O[v], C[v]) to c and increment c by
one. Afterwards, we update the cluster ID of v in O to be H(O[v], C[v]). At the end of the
algorithm, c is equal to the number of clusters contained in the overlay clustering and each
vertex is labeled with its cluster ID in O.
XX:3
XX:4
Memetic Graph Clustering
2.3
Objective Functions
There are a variety of measures used to assess the quality of a clustering, such as coverage [10], performance [45], inter-cluster conductance [23], surprise [3], map equation [37] and
modularity [32]. The most simple index realizing a traditional measure of clustering quality
is coverage. The coverage of a graph clustering C is defined as the fraction of intra-cluster
edges within the complete set of edges cov(C) := m(C)
m . Intuitively, large values of coverage
correspond to a good quality of a clustering. However, one principal drawback of coverage
is that the converse is not necessarily true: coverage takes its largest value of 1 in the trivial
case where there is only one cluster. Modularity fixes this issue by comparing the coverage of
a clustering to the same value after rearranging edges randomly. Performance is the fraction
of correctly classified vertex pairs, w.r.t the set of edges. Inter-cluster conductance returns
the worst (i.e. the thickest) bottleneck created by separating a cluster from the rest of the
graph. Surprise measures the probability that a random graph R has more intra-cluster
edges. Map equation is a flow-based and information-theoretic method to assess clustering
quality. Our focus in this work is on modularity as it is a widely accepted quality function
and has been the main objective function in the 10th DIMACS implementation challenge [7].
For further discussions of these indices we refer the reader to the given references, and simply
state the formal definition of modularity here:
!2
1 X X
m(C)
−
d(v)
Q(C) := cov(C) − E[cov(C)] =
m
4m2
Vi ∈C
2.4
v∈Vi
Related Work
This paper is a summary and extension of the bachelor’s thesis by Sonja Biedermann [8].
There has been a significant amount of research on graph clustering. We refer the reader
to [19, 22] for thorough reviews of the results in this area. Here, we focus on results closely
related to our main contributions. It is common knowledge that there is no single best
strategy for graph clustering. Moreover, most quality indices for graph clusterings have
turned out to be NP-hard to optimize and rather resilient to effective approximations, see
e.g. [4, 11, 46], allowing only heuristic approaches for optimization. Other approaches often
rely on specific strategies with high running times, e.g. the iterative removal of central
edges [32], or the direct identification of dense subgraphs [15]. Provably good methods with
a decent running time include algorithms that have a spectral embedding of the vertices
as the basis for a geometric clustering [12], min-cut tree clustering [18], a technique which
guarantees certain bounds on bottlenecks and an approach which relies on random walks in
graphs staying inside dense regions with high probabilities [45]. The Louvain method is a
multi-level clustering algorithm introduced by Blondel et al. [9] that optimizes modularity
as an objective function. As we use this method in our algorithm to create the initial
population as well as to improve individuals after recombination, we go into more detail for
that method in Section 2.6.
Recently, the 10th DIMACS challenge on graph partitioning and graph clustering compared different state-of-the-art graph clustering algorithms w.r.t optimizing modularity [6].
Most of the best results have been obtained by Ovelgönne and Geyer-Schulz [33] as well as
Aloise et al. [2]. CGGC [33] used an ensemble learning strategy during the challenge. From
several weak clusterings an overlap clustering is computed and a more expensive clustering
algorithm is applied. VNS [2] applies the Variable Neighborhood Search heuristic to the
graph clustering problem. Later, Džamić et al. [17] outperformed VNS by proposing an
ascent-decent VNS. We compare ourselves to the results above in Section 4.
S. Biedermann et. al.
Evolutionary Graph Clustering. Tasgin and Bingol [44] introduce a genetic algorithm
using modularity as a quality measure. They chose an integer encoding for representing the
population clusterings. Individuals are randomly initialized, with some bias for assigning
direct neighbors to the same cluster. Recombination is done one way, i.e. instead of using
mutual exchange, clusters are transferred from a source individual to a target individual.
More precisely, the operation picks a random vertex in the source individual and clusters
all vertices of its cluster together in the destination individual. This is repeated several
times in one recombination operation. As for mutation, the authors chose to pick one vertex
and move it to a random cluster, which is almost guaranteed to decrease the fitness. This
approach does not use local search to improve individuals.
A mutation-less agglomerative hierarchical genetic algorithm is presented by Lipczak
and Milios in [26]. The authors represent each cluster as one individual and use one-point
and uniform crossovers in that representation. Two synthetic networks have been used for
the experimental evaluation. An important advantage of this algorithm is the possibility to
distribute its computations.
2.5
Karlsruhe High-Quality Partitioning
Within this work, we use the open source multi-level graph partitioning framework KaHIP [40]
(Karlsruhe High-Quality Partitioning). More precisely, we employ partitioning tools contained therein to create high-quality partitions of the graphs. We shortly outline its main
components. KaHIP implements many different algorithms, for example, flow-based methods and more-localized local searches within a multi-level framework, as well as several
coarse-grained parallel and sequential meta-heuristics. Recently, specialized methods to
partition social networks and web graphs have been included in the framework [29].
2.6
Multi-level Louvain Method
The Louvain method is a multi-level clustering algorithm introduced by Blondel et al. [9]. It
is an approach to graph clustering that optimizes modularity as an objective function. Since
we instantiate our memetic algorithm to optimize for modularity, we give more detail in order
to be self-contained. The algorithm is organized in two phases: a local movement phase and
contraction/uncontraction phase. The first phase, local movement, works in rounds and
is done as follows: In the beginning, each vertex is a singleton cluster. Vertices are then
traversed in random order and always moved to the neighboring cluster yielding the highest
modularity increase. Computing the best move can be done in time proportional to the
degree of a vertex by storing cumulative vertex degrees of clusters. More precisely, the gain
in modularity by removing a vertex from a cluster can be computed by
P
P
(d(u) + v∈C d(v))2
d(u) ( v∈C d(v))2
s
+
+
−
.
∆Q(u) =
m 4m2
4m2
4m2
A similar formula holds when adding a singleton vertex to a cluster. Once a vertex is moved,
the cumulated degrees of the affected clusters are updated. Hence, the best move can be
found in time proportional to the degree of a vertex. The local movement algorithm stops
when a local maximum of the modularity is attained, that is when no vertex move a yields
modularity gain. The second phase of the algorithm consists in contracting the clustering.
Contracting the clustering works as follows: each block of the clustering is contracted into
a single node. There is an edge between two vertices u and v in the contracted graph if the
two corresponding blocks in the clustering are adjacent to each other in G, i.e. block u and
XX:5
XX:6
Memetic Graph Clustering
block v are connected by at least one edge. The weight of an edge (A, B) is set to the sum of
the weight of the edges that run between block A and block B of the clustering. Moreover,
a self-loop is inserted for each vertex in the contracted graph. The weight of this edge is set
to be the cumulative weight of the edges in the respective cluster. Note that due to the way
the contraction is defined, a clustering of the coarse graph corresponds to a clustering of the
finer graph with the same objective. The algorithm then continues with the local movement
phase on the contracted graph. In the end, the clustering contractions are undone and at
each level local movement improves the current clustering w.r.t. modularity.
3
Memetic Graph Clustering
We now explain the components of our memetic graph clustering algorithm. Our algorithm
starts with a population of individuals (in our case one individual is a clustering of the
graph) and evolves the population into different populations over several rounds. In each
round, the EA uses a selection rule based on the fitness of the individuals (in our case the
objective function of the clustering problem under consideration) of the population to select
good individuals and recombine them to obtain improved offspring. Our selection process is
based on the tournament selection rule [30], i.e. C is the fittest out of two random individuals
R1 , R2 from the population. When an offspring is generated an elimination rule is used to
select a member of the population and replace it with the new offspring. In general, one
has to take both into consideration, the fitness of an individual and the distance between
individuals in the population [35] in order to avoid premature convergence of the algorithm.
We evict the solution that is most similar with respect to the edges that run in between
clusters with the offspring among those individuals in the population that have a worse or
equal objective than the offspring itself. If there is no such individual, then the offspring
is rejected and not inserted into the population. The difference between two individuals is
defined as the size of the symmetric difference between their sets of cut edges. Our algorithm
generates only one offspring per generation.
The core of our algorithm are our novel recombination and mutation operations. We
provide two different kinds of operations, flat- and multi-level recombination operations.
We also define a mutation operator that splits clusters by employing graph partitioning
techniques. In any case, the offspring is typically improved using a local search algorithm
that optimizes for the objective function of the graph clustering problem. We continue this
section by explaining the details of our algorithm with a focus on modularity—the results
are transferable to other clustering problems by using local search algorithms that optimize
the objective function of the problem under consideration.
3.1
Initialization/Creating Individuals
We initialize our population using the following modified Louvain algorithm. We also use
the following algorithm to create an individual in recombine and mutation operations. Depending on the objective of the evolutionary algorithm, the choice of algorithms to initialize
the population may be different. We describe our scheme for modularity.
We mainly use the Louvain method to create clusterings. Due to the non-deterministic
nature of this algorithm—the order of nodes visited is randomized—we obtain different initial
graph clusterings and later on individuals. In order to introduce more diversification, we
modify the approach by using a different coarsening strategy based on the size-constrained
label propagation algorithm [29]. Label propagation works similar to the local movement
phase of the Louvain method. However, the objective of the algorithm is different. Without
S. Biedermann et. al.
size-constraints, it was proposed by Raghavan et al. [36]. Initially, each vertex is in its own
cluster/block, i.e. the initial block ID of a vertex is set to its vertex ID. The algorithm
then works in rounds. In each round, the vertices of the graph are traversed in a random
order. When a vertex v is visited, it is moved to the block that has the strongest connection
to v, i.e. it is moved to the cluster Vi that maximizes ω({(v, u) | u ∈ N (v) ∩ Vi }) whilst
not overloading the target cluster w.r.t to the size-constraint bound U . Ties are broken
randomly. The process is repeated until the process has converged. Here, we perform at
most ` rounds of the algorithm, where ` is a tuning parameter, and stop the algorithm if
less than five percent of the vertices changed its cluster during one round. One LPA round
can be implemented to run in O(n + m) time.
We modify the coarsening stage of the Louvain method using size-constrained label
propagation as follows: For the first λ ∈rnd [0, 4] levels of the multi-level hierarchy we use
size-constraint label propagation to compute the clustering to be contracted instead of the
local movement phase to compute a clustering. We choose U ∈rnd [n/10, n] in the beginning
of the multi-level algorithm. Afterwards, we switch to the local movement phase of the
Louvain method to compute a clustering to be contracted. Note that for λ = 0 the method
is the Louvain method. In any case, in the uncoarsening phase local movement improves
the current clustering by optimizing modularity.
3.2
Recombination
We define flat- and multi-level recombination operations which we are going to explain now.
Both operations use the notion of overlay clustering to pass on good parts of the solutions
to the offspring. Some of our recombination operators ensure that the offspring has nondecreasing fitness. Moreover, our recombine operations can recombine a solution from the
population with an arbitrary clustering of the graph. Due to the fact that our local search
and multi-level algorithms are randomized, a recombine operation performed twice using
the same parents can yield different offsprings.
Flat Recombination
The basic flat recombination operation starts by taking two clusterings C1 , C2 from the population as input and computes its overlay. The overlay is then contracted such that vertices
represent blocks in the overlay clustering and edges represent edges between vertices of two
blocks. Edge weights are equal to the summed weight of the edges that run between the respective blocks, and self-loops are inserted with the edge weight being equal to the summed
internal edge weight of that corresponding block. Note that due to the way contraction is
defined, a clustering of the contracted graph can be transformed into a clustering of the input
network having the same modularity score. Hence, we use the Louvain method to cluster the
contracted graph, which then constitutes the offspring. Also, note that the contraction of
the overlay ensures that vertices that are clustered together in both inputs will belong to the
same cluster in the offspring. Only vertices which are split by one of the input clusterings
will be affected by the Louvain method. What follows are multiple variations of the basic
flat recombination method that ensure non-decreased fitness of the input individual, i.e. the
offspring has fitness at least as good as the better input individual.
Apply Input Clustering. This operator uses the better of the two parent clusterings as a
starting point for the Louvain method on the contracted graph. Due to the way contraction
is defined the objective function of the starting point on the contracted graph is the same
as of the corresponding clustering on the input network. The resulting offspring can be
XX:7
XX:8
Memetic Graph Clustering
expected to be less diverse than when using the plain recombination operator, but may
significantly more improve upon the parent.
Cluster-/Partition Recombination. This operator differs from the previous two operators
insofar that only one of the parents is selected from the population. The other parent is
manufactured on the spot by running either the size-constrained label propagation (SCLP)
or a graph partitioning algorithm on the input graph. The resulting clustering can be very
different from to the other parent. This introduces more diversification which helps local
search to explore a larger search space. When using SCLP clustering for the recombination,
we use the parameters as described in Section 3.1. When using a partition, we use the KaHIP
graph partitioning framework with the fastsocial preconfiguration using k ∈rnd [2, 64] and
∈rnd [0.03, 0.5].
Multi-level Recombination
We now explain our multi-level recombine operator. This recombine operator also ensures
that the solution quality of the offspring is at least as good as the best of both parents. For
our recombine operator, let C1 and C2 again be two individuals from the population. Both
individuals are used as input for the multi-level Louvain method in the following sense. Let
E be the set of edges that are cut edges, i.e. edges that run between two blocks, in either C1
or C2 . All edges in E are blocked during the coarsening phase, i.e. they are not contracted
during the coarsening phase. In other words, these edges cannot be contracted during the
multi-level scheme. We ensure this by modifying the local movement phase of the Louvain
method, i.e. clusters can only grow inside connected components of the overlay (V, E\E).
We stop contracting clusterings when no contractable edge is left.
When coarsening is stopped, we then apply the better out of both input individuals
w.r.t. the objective to the coarsest graph and use this as initial clustering instead of running
the Louvain method on the coarsest graph. However, during uncoarsening local search still
optimizes modularity on each level of the hierarchy. Note that this is possible since we did not
contract any cut edge of both inputs. Also, note that this way we obtain a clustering of the
coarse graph having a modularity score being equal to the better of both input individuals.
Local movement guarantees no worsening of the clustering. Hence, the offspring is at least
as good as the best input individual. Note that the coarsest graph is the same as the graph
obtained in a flat recombination. However, the operation now is able to pass on good parts
of the solution on multi-levels of the hierarchy during uncoarsening and hence has a more
fine-grained view on both individuals.
3.3
Mutation
Our recombination operators can only decrease the number of clusters present in the solution.
To counteract this behavior, we define a mutation operator that selects a subset of clusters
and splits each of them in half. Splitting is done using the KaHIP graph partitioning
framework. The number of clusters to be split is set to ps |C| where ps is the splitting
probability and |C| is the number of clusters that the clustering selected from the population
has. That means that we split a cluster into two balanced blocks such that there are only
a small amount of edges running between them. We do this operation with two individuals
from the population that are found by tournament selection. This results in two output
individuals which are used as input to a multi-level recombination operation. The computed
offspring of that operation is then inserted into the population as described above.
S. Biedermann et. al.
3.4
Parallelization
We now explain the island-based parallelization that we use. We use a parallelization scheme
that has been successfully used in graph partitioning [39]. Each processing element (PE)
basically performs the same operations using different random seeds. First, we estimate
the population size S: each PE creates an individual and measures the time t spent. We
then choose S such that the time for creating S clusterings is approximately ttotal /c where
the fraction c is a tuning parameter and ttotal is the total running time that the algorithm
is given to produce a clustering of the graph. The minimum amount of individuals in the
population is set to 3, the maximum amount of the individuals in the population is set to
100. The lower bound on the population size is chosen to ensure a certain minimum of
diversity, while the upper bound is used to ensure convergence.
Each PE then builds its own population. Afterwards, the algorithm proceeds in rounds
as long as time is left. Either a mutation or recombination operation is performed. Our
communication protocol is similar to randomized rumor spreading which has shown to be
scalable in previous work [39]. Let p denote the number of PEs used. A communication step
is organized in rounds. In each round, a PE chooses a communication partner and sends her
the currently best individual C of the local population. The selection of the communication
partner is done uniformly at random among those PEs to which C not already has been sent
to. Afterwards, a PE checks if there are incoming individuals and if so inserts them into the
local population using the elimination strategy described above. If the best local individual
has changed, all PEs are again eligible. This is repeated log p times. The algorithm is
implemented completely asynchronously, i.e. there is no need for a global synchronization.
4
Experimental Evaluation
System and Methodology
We implemented the memetic algorithm described in the previous section within the KaHIP
(Karlsruhe High Quality Partitioning) framework. The code is written in C++ and MPI.
It has been compiled using g++-5.2 with flags -O3 and OpenMPI 1.6.5. We refer to the
algorithm presented in this paper as VieClus. We plan to further improve the code and then
to release it to make it available to a larger audience. Throughout this section, our main
objective is modularity. All experiments comparing VieClus with competing algorithms
are performed on a cluster with 512 nodes, where each node has two Intel Xeon E5-2670
Octa-Core (Sandy Bridge) processors clocked at 2.6 GHz, 64 GB main memory, 20 MB L3and 8x256 KB L2-Cache and runs RHEL 7.4. We use the arithmetic mean when averaging
over solutions of the same instance and the geometric mean when averaging over different
instances in order to give every instance a comparable influence on the final result. It is well
known that the algorithms that scored most of the points during the 10th DIMACS challenge
compute better results than the Louvain method. Moreover, LaSalle [25] reports results for
his Nerstrand algorithm that are on average equal or slightly better than the Louvain method
on the instances used here, but the best results of this method are consistently worse than
the result computed during the DIMACS challenge. Hence, we refrain from doing additional
experiments with the Louvain method.
XX:9
XX:10
Memetic Graph Clustering
Table 1 Results of our algorithm on the benchmark test set. Columns from left to right: average
modularity achieved by our algorithm, t average time in minutes needed to beat the old challenge
result, the best score computed of our algorithm, the best modularity scores achieved by challenge
participants, running time in minutes needed to create the previous best entry according to [25]
and reference to the solver that achieved the result during the 10th DIMACS challenge. Entries
that improve or reproduce the result of the implementation challenge are highlighted.
Graph
Avg. Q
as-22july06
astro-ph
celegans_metabol
cond-mat-2005
email
PGPgiantcompo
polblogs
power
smallworld
memplus
G_n_pin_pout
caidaRouterLevel
rgg_n17
luxembourg.osm
0,679 391
0,746 285
0,453 248
0,750 065
0,582 829
0,886 853
0,427 105
0,940 975
0,793 186
0,701 242
0,500 457
0,872 804
0,978 448
0,989 665
coAuthorsCiteseer
citationCiteseer
coPapersDBLP
belgium.osm
ldoor
eu-2005
in-2004
333SP
prefAttachment
0,906 804
0,825 518
0,868 019
0,995 062
0,970 521
0,941 575
0,980 684
0,989 316
0,315 843
t [m]
Small
<1
<1
<1
<1
<1
<1
<1
<1
<1
2,6
3,9
5,0
5,0
7,3
Large
3,9
12,9
20,5
29,5
35,1
65,8
237,4
297,1
*
Max. Q
Instances
0,679 396
0,746 292
0,453 248
0,750 171
0,582 829
0,886 853
0,427 105
0,940 977
0,793 187
0,701 275
0,500 466
0,872 828
0,978 454
0,989 672
Instances
0,906 830
0,825 545
0,868 058
0,995 064
0,970 555
0,941 575
0,980 690
0,989 356
0,316 089
Q [7]
tsol [m]
Solver
0,678 267
0,744 621
0,453 248
0,746 254
0,582 829
0,886 564
0,427 105
0,940 851
0,793 042
0,700 473
0,500 098
0,872 042
0,978 324
0,989 621
6,6
11,9
<1
40,9
<1
1,9
<1
<1
16,8
3,2
64,8
81,0
37,5
40,9
CGGC[33]
VNS[2]
VNS[2]
CGGC[33]
VNS[2]
CGGC[33]
VNS[2]
VNS[2]
VNS[2]
CGGC[33]
CGGC[33]
CGGC[33]
VNS[2]
VNS[2]
0,905 297
0,823 930
0,866 794
0,994 940
0,969 370
0,941 554
0,980 622
0,989 095
0,315 994
91,3
77,6
603,3
102,9
485,6
341,5
244,0
976,9
1 353,1
CGGC[33]
CGGC[33]
CGGC[33]
CGGC[33]
ParMod[13]
CGGC[33]
CGGC[33]
ParMod[13]
VNS[2]
Parameters
We did not perform a tuning of the parameters of the algorithm, rather we chose the parameters described above and below to be reasonable and to introduce a large amount of
diversification or we chose parameters that were a good choice in previous evolutionary algorithms [39]. As we our main design goal is to introduce as much diversification as possible,
we use all recombination and mutation operations in our algorithm. The ratio of mutation
to recombine operations has been set to 1:9 as this has been a good choice in previous evolutionary algorithms [39]. When we perform a recombine operation, we pick the recombine
operation uniformly at random and diversify the parameters as described above. When
performing a mutation operation, we use a splitting probability ps uniformly at random
in [0.01, 0.1]. We invest 1/10 of the total time to create the initial population.
Instances
We use the graphs that have been used for the 10th DIMACS implementation challenge on
graph clustering and graph partitioning [7]. A list of the instances as well as the modularity
scores that have been obtained during the challenge can be found in Table 1. We exclude
the instances cage15, audikw1, er-fact1.5-scale25, kron_*, uk-2002, uk-2007-05 from
the challenge testbed in our evaluation, because they are either too large to be feasible for
an evolutionary algorithm or they do not contain a significant cluster structure as indicated
by the reported modularity score [6].
S. Biedermann et. al.
4.1
Evolutionary Graph Clustering
We now run our algorithm on all the DIMACS instances under consideration using the rules
used there, i.e. running time is not an issue but we want to achieve modularity values as
large as possible for each instance. On the small instances, we give our algorithm 2 hours of
time to compute a solution and on the large instances, we set the time limit to 16 hours. In
any case, we use 16 cores of our machine, i.e. one node of the machine. We perform the test
five times with different random seeds. Table 1 summaries the results of our experiment,
and Figure 1 shows convergence plots for a selected subset of the instances.
First of all, on every instance under consideration, our algorithm is able to compute a
result that is better than the currently reported modularity value in literature. More precisely, in 98 out of 115 runs of our algorithm, the previous benchmark result of the 10th
DIMACS implementation challenge is outperformed. In further 15 out of 115 runs, we reproduce the results. This is the case for each of the runs for the graphs celegans_metabolic,
email and polblogs. The two cases in which our algorithm does not beat the previous
best solver, are 2 runs on the graph prefAttachment. The other 3 runs on that graph
outperform the previous result. The time needed to compute a clustering having a similar
score on that graph is roughly 95% of the total running time. Moreover, as the convergence
plots in Figure 1 show this may be fixable by giving the algorithm a larger amount of time
to compute the solution. Overall, the time needed to outperform the previous benchmark
results ranges from less than a minute to a couple of minutes on the small instances. On
the large instances our algorithm needs more time, but on most instances, the previous
benchmark result can be computed in roughly an hour. Table 1 also reports the running
time of the solver that obtained the result during the DIMACS challenge. Our algorithm is
faster in every case compared to the previous solver (eventually by more than an order of
magnitude), however, the machines used for the experiment are different and our algorithm
is a parallel algorithm whereas previous solvers are sequential. Note that the final improvements over the old result are fairly small (<0.1% on average). This is not surprising, as
previous solvers already invested a large amount of time to compute the results. However,
note that the previous result has been computed by different solvers and our evolutionary
algorithm can be seen as a single tool to compute the result.
We now compare the results with recently published results [25, 21, 27, 38]. LaSalle [25]
reports results on all graphs from the DIMACS challenge subset. However, each of the
best computed result is worse than the result computed by the respective best algorithm
during the DIMACS challenge (and hence worse compared to our algorithm). Moreover,
LaSalle [25] reports that his Nerstrand algorithm is on average equal or slightly better
than the Louvain method on the instances used here. Lu et al. [27] present a result for
coPapersDBLP (Q = 0,858 088) and Ryu and Kim [38] report modularity for email (Q =
0,568) which are worse compared to the results that we report here. Hamann et al. [21]
report the result for in-2004 (Q = 0,980) which is comparable to the result that we report
here. Džamić [17] build upon VNS proposing an ascent-decent VNS. Results are reported for
celegans_metabolic (Q = 0.453248), email (Q = 0,582 829), polblogs (Q = 0,427 105)
for which their algorithm computes the same results as all other tools reported in Table 1.
Moreover, the following best results are reported as-22july06 (Q = 0,678 381), astro-ph
(Q = 0,745 246), cond-mat-2005 (Q = 0,747 181), PGPgiantcompo (Q = 0.886647), as well
as power (Q = 0.940974) which are indeed better than the previous best result obtained
during the 10th DIMACS challenge, but still worse than average result of our algorithm in
every case. Hence, overall we consider our algorithm as a new state-of-the-art heuristic for
solving the modularity clustering problem.
XX:11
Memetic Graph Clustering
100
10000
50
0.7925
10000
0.9884
0.9890
333SP
mean modularity
belgium.osm
500 10000
VieClus
Challenge Result
20
200
2000
50000
time t[s]
ldoor
preferentialAttachment
5000
time t[s]
0.28
100
0.30
time t[s]
mean modularity
0.9945
10 100
time t[s]
VieClus
Challenge Result
10
1
time t[s]
VieClus
Challenge Result
5
VieClus
Challenge Result
0.7910
10
smallworld
mean modularity
0.87
0.85
1
0.9948
0.83
VieClus
Challenge Result
0.966
0.960
mean modularity
mean modularity
caidaRouterLevel
mean modularity
XX:12
VieClus
Challenge Result
1
10
1000
time t[s]
Figure 1 Convergence plots for the instances the small instances caidaRouterLevel, smallworld
and the large instances 333SP, belgium.osm, ldoor, preferentialAttachment.
S. Biedermann et. al.
Convergence Plots. We are now interested in how solution quality evolves over time. We
start by explaining how to compute the data for a single instance I, i.e., a clustering of
a graph G. Whenever an algorithm computes a clustering that improves the modularity
score, it reports a pair (t, Q), where the timestamp t is the currently elapsed time and Q
is the modularity score of the currently reported solution. For r repetitions with different
seeds s, these r sequences TsI of pairs are merged into one sequence T I of triples (t, s, Q),
which is sorted by the timestamp t. Since we are interested in the evolution of the solution
I
representing event-based average values. We start by
quality, we compute the sequence Tmax
computing the average objective value Q and the average time t using the first pair (t, Q) of
I
all r sequences TsI and insert (t, Q) into Tmax
. We then sweep through the remaining entries
I
(t, s, Q) of T . Each entry corresponds to a clustering computed at timestamp t using seed
s that improved the solution quality to Q. For each entry we replace the old objective value
of seed s that took part in the computation of Q with the new value Q, recompute Q and
I
I
. Tmax
therefore represents the evolution of the average
insert a new pair (t, Q) into Tmax
solution quality Q for instance I over time.
We present those plots for a selected set of instances in Figure 1. In the plots one can
see that solution quality levels off in the beginning. At this point, the algorithm is still in
the initialization phase, where the population is built using repeated runs of the Louvain
method with different coarsening strategies. However, this is not enough to find significantly
improved solutions. Afterwards, solution quality improves rapidly as the actual evolutionary
operations run.
5
Conclusion
We presented a parallel memetic algorithm, VieClus, that tackles the graph clustering problem. A key component of our contribution are natural recombine operators that employ
ensemble clusterings as well as multi-level techniques. We combine these techniques with a
scalable communication protocol, producing a system that is able to reproduce or improve
previous all entries of the 10th DIMACS implementation challenge under consideration as
well as results recently reported in the literature in a short amount of time. Moreover, while
the previous best result for different instances has been computed by a variety of solvers,
our algorithm can now be used as a single tool to compute the result. Hence, overall we
consider our algorithm as a new state-of-the-art heuristic for solving the modularity clustering problem. Considering the good results of our algorithm, we want to further improve
and release it. In the future, it may be interesting to instantiate our scheme for different
objective functions depending on the application domain or to use a more diverse set of
initial solvers to create the population. We also want to look at distributed memory parallel multi-level algorithms for the problem that can use the depicted algorithm as an initial
clustering scheme on the coarsest level of the hierarchy.
Acknowledgements
The research leading to these results has received funding from the European Research
Council under the European Community’s Seventh Framework Programme (FP7/2007-2013)
/ERC grant agreement No. 340506.
XX:13
XX:14
Memetic Graph Clustering
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
L. Akoglu, H. Tong, and D. Koutra. Graph Based Anomaly Detection and Description: A
Survey. Data Min. Knowl. Discov., 29(3):626–688, May 2015. URL: http://dx.doi.org/
10.1007/s10618-014-0365-y, doi:10.1007/s10618-014-0365-y.
D. Aloise, G. Caporossi, S. Perron, P. Hansen, L. Liberti, and M. Ruiz. Modularity Maximization in Networks by Variable Neighborhood Search. In 10th DIMACS Impl. Challenge
Workshop. Georgia Inst. of Technology, Atlanta, GA, 2012.
V. Arnau, S. Mars, and I. Marín. Iterative Cluster Analysis of Protein Interaction Data.
Bioinformatics, 21(3):364–378, 2004.
G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela, and M. Protasi.
Complexity and Approximation: Combinatorial Optimization Problems and their Approximability Properties. Springer Science & Business Media, 2012.
T. Bäck. Evolutionary Algorithms in Theory and Practice: Evolution Strategies, Evolutionary Programming, Genetic Algorithms. PhD thesis, 1996.
D. Bader, A. Kappes, H. Meyerhenke, P. Sanders, C. Schulz, and D. Wagner. Benchmarking
for Graph Clustering and Partitioning. In Encyclopedia of Social Network Analysis and
Mining. Springer, 2014.
D. Bader, H. Meyerhenke, P. Sanders, and D. Wagner, editors. Proc. of the 10th DIMACS
Impl. Challenge, Cont. Mathematics. AMS, 2012.
S. Biedermann. Evolutionary Graph Clustering. Bachelor’s Thesis, Universität Wien, 2017.
V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. Fast Unfolding of
Communities in Large Networks. Journal of Statistical Mechanics: Theory and Experiment, 2008(10):P10008, 2008. URL: http://stacks.iop.org/1742-5468/2008/i=10/a=
P10008.
U. Brandes. Network Analysis: Methodological Foundations, volume 3418. Springer Science
& Business Media, 2005.
U. Brandes, D. Delling, M. Gaertler, R. Gorke, M. Hoefer, Z. Nikoloski, and D. Wagner. On
Modularity Clustering. IEEE Transactions on Knowledge and Data Engineering, 20(2):172–
188, 2008.
U. Brandes, M. Gaertler, and D. Wagner. Engineering Graph Clustering: Models and
Experimental Evaluation. ACM Journal of Experimental Algorithmics, 12(1.1):1–26, 2007.
Ü. V. Çatalyürek, K. Kaya, J. Langguth, and B. Uçar. A Divisive Clustering Technique for
Maximizing the Modularity. In 10th DIMACS Impl. Challenge Workshop. Georgia Inst. of
Technology, Atlanta, GA, 2012.
J. Demme and S. Sethumadhavan. Approximate Graph Clustering for Program Characterization. ACM Trans. Archit. Code Optim., 8(4):21:1–21:21, January 2012. URL:
http://doi.acm.org/10.1145/2086696.2086700, doi:10.1145/2086696.2086700.
I. Derényi, G. Palla, and T. Vicsek. Clique Percolation in Random Networks. Physical
review letters, 94(16):160202, 2005.
A. A. Diwan, S. Rane, S. Seshadri, and S. Sudarshan. Clustering Techniques for Minimizing
External Path Length. In Proceedings of the 22th International Conference on Very Large
Data Bases, VLDB ’96, pages 342–353, San Francisco, CA, USA, 1996. Morgan Kaufmann
Publishers Inc. URL: http://dl.acm.org/citation.cfm?id=645922.673636.
D. Džamić, D. Aloise, and N. Mladenović. Ascent–descent Variable Neighborhood Decomposition Search for Community Detection by Modularity Maximization. Annals of
Operations Research, Jun 2017. URL: https://doi.org/10.1007/s10479-017-2553-9,
doi:10.1007/s10479-017-2553-9.
G. W. Flake, R. E. Tarjan, and K. Tsioutsiouliklis. Graph Clustering and Minimum Cut
Trees. Internet Mathematics, 1(4):385–408, 2004.
S. Fortunato. Community Detection in Graphs. Physics reports, 486(3):75–174, 2010.
S. Biedermann et. al.
XX:15
20
D. E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine Learning.
Addison-Wesley, 1989.
21
M. Hamann, B. Strasser, D. Wagner, and T. Zeitz. Simple Distributed Graph Clustering
using Modularity and Map Equation. arXiv preprint arXiv:1710.09605, 2017.
22
T. Hartmann, A. Kappes, and D. Wagner. Clustering Evolving Networks. In Algorithm
Engineering, pages 280–329. Springer, 2016.
23
R. Kannan, S. Vempala, and A. Vetta. On clusterings: Good, bad and spectral. Journal
of the ACM (JACM), 51(3):497–515, 2004.
24
J. Kim, I. Hwang, Y. H. Kim, and B. R. Moon. Genetic Approaches for Graph Partitioning: A Survey. In Proceedings of the 13th Annual Genetic and Evolutionary Computation
Conference (GECCO’11), pages 473–480. ACM, 2011.
25
D. LaSalle. Graph Partitioning, Ordering, and Clustering for Multicore Architectures.
2015.
26
M. Lipczak and E. E. Milios. Agglomerative Genetic Algorithm for Clustering in Social
Networks. In Franz Rothlauf, editor, GECCO, pages 1243–1250. ACM, 2009. URL: http:
//dblp.uni-trier.de/db/conf/gecco/gecco2009.html#LipczakM09.
27
H. Lu, M. Halappanavar, and A. Kalyanaraman. Parallel heuristics for scalable community
detection. Parallel Computing, 47:19–37, 2015.
28
S. McFarling. Program Optimization for Instruction Caches. SIGARCH Comput. Archit.
News, 17(2):183–191, April 1989. URL: http://doi.acm.org/10.1145/68182.68200,
doi:10.1145/68182.68200.
29
H. Meyerhenke, P. Sanders, and C. Schulz. Partitioning Complex Networks via SizeConstrained Clustering. In SEA, volume 8504 of Lecture Notes in Computer Science, pages
351–363. Springer, 2014.
30
B. L Miller and D. E Goldberg. Genetic Algorithms, Tournament Selection, and the Effects
of Noise. Evolutionary Computation, 4(2):113–131, 1996.
31
M. E. J. Newman.
68(2):026121, 2003.
32
M. E. J. Newman and M. Girvan. Finding and Evaluating Community Structure in Networks. Physical review E, 69(2):026113, 2004.
33
M. Ovelgönne and A. Geyer-Schulz. An Ensemble Learning Strategy for Graph Clustering.
In Graph Partitioning and Graph Clustering, number 588 in Contemporary Mathematics.
2013.
34
J. B. Pereira-Leal, A. J. Enright, and C. A. Ouzounis. Detection of Functional Modules from
Protein Interaction Networks. Proteins: Structure, Function, and Bioinformatics, 54(1):49–
57, 2004. URL: http://dx.doi.org/10.1002/prot.10505, doi:10.1002/prot.10505.
35
D. C. Porumbel, J.-K. Hao, and P. Kuntz. Spacing Memetic Algorithms. In 13th Annual
Genetic and Evolutionary Computation Conference, GECCO 2011, Proceedings, Dublin,
Ireland, July 12-16, 2011, pages 1061–1068, 2011.
36
U. N. Raghavan, R. Albert, and S. Kumara. Near Linear Time Algorithm to Detect
Community Structures in Large-Scale Networks. Physical Review E, 76(3), 2007.
37
M. Rosvall, D. Axelsson, and C. T. Bergstrom. The Map Equation. The European Physical
Journal-Special Topics, 178(1):13–23, 2009.
38
S. Ryu and D. Kim. Quick Community Detection of Big Graph Data Using Modified
Louvain Algorithm. In High Performance Computing and Communications; IEEE 14th
International Conference on Smart City; IEEE 2nd International Conference on Data Science and Systems (HPCC/SmartCity/DSS), 2016 IEEE 18th International Conference on,
pages 1442–1445. IEEE, 2016.
Properties of Highly Clustered Networks.
Physical Review E,
XX:16
Memetic Graph Clustering
39
40
41
42
43
44
45
46
47
P. Sanders and C. Schulz. Distributed Evolutionary Graph Partitioning. In Proc. of the
12th Workshop on Algorithm Engineering and Experimentation (ALENEX’12), pages 16–
29, 2012.
P. Sanders and C. Schulz. Think Locally, Act Globally: Highly Balanced Graph Partitioning. In 12th International Symposium on Experimental Algorithms (SEA’13). Springer,
2013.
S. E. Schaeffer. Survey: Graph Clustering. Comput. Sci. Rev., 1(1):27–64, August 2007. URL: http://dx.doi.org/10.1016/j.cosrev.2007.05.001, doi:10.1016/
j.cosrev.2007.05.001.
C. L. Staudt and H. Meyerhenke. Engineering High-Performance Community Detection
Heuristics for Massive Graphs. In Proceedings 42nd Conference on Parallel Processing
(ICPP’13), 2013.
C. L. Staudt and H. Meyerhenke. Engineering Parallel Algorithms for Community Detection
in Massive Networks. IEEE Trans. on Parallel and Distributed Systems, 27(1):171–184,
2016. doi:10.1109/TPDS.2015.2390633.
M. Tasgin and H. Bingol. Community Detection in Complex Networks using Genetic
Algorithm. In ECCS ’06: Proc. of the European Conference on Complex Systems, April
2006. URL: http://arxiv.org/abs/cond-mat/0604419, arXiv:cond-mat/0604419.
S. M. Van Dongen. Graph Clustering by Flow Simulation. PhD thesis, 2001.
D. Wagner and F. Wagner. Between Min Cut and Graph Bisection. In Proceedings of the
18th International Symposium on Mathematical Foundations of Computer Science, pages
744–750. Springer, 1993.
Y. Xu, V. Olman, and D. Xu. Clustering Gene Expression Data using a Graph-Theoretic
Approach: an Application of Minimum Spanning Trees. Bioinformatics, 18(4):536–545,
2002.
S. Biedermann et. al.
A
XX:17
Approximate Bound for Modularity
The modularity clustering problems seeks to find a clusterings such that cov(C) - E[cov(C)]
is maximum. However, the problem is NP-complete and our algorithm heuristically finds
solutions. Moreover, modularity is trivially bounded by 1. The true optimum clustering,
however, has typically a value below that. Thus we try to find a bound or a value to
normalize the score such that one gets a clearer picture on how far away from the optimum
score the achieved modularity value is.
P
To do so, let us define v∈Vi d(v) to be the volume vol(Vi ) of the cluster Vi . Note that
P
1
2
have the same volume. But again,
the formula 4m
2
Vi ∈C vol(Vi ) is minimized if all clustersP
finding the optimum value of this equation is hard. As Vi ∈C vol(Vi ) = 2m, the equation
is minimized when vol(Vi )= 2m/k ∀i = 1 . . . k, where k is the number of clusters. Thus it
P
1
2m 2
1
1
1
2
follows that 4m
2
Vi ∈C vol(Vi ) ≥ 4m2 · k · ( k ) = k . This gives an upper of 1 − k for
Q(C) if the number of clusters is an input to the clustering problem. This upper bound is,
however, not very far from the trivial upper bound of 1 for k > 10.
Thus for each graph we also clustered the nodes into cluster Vi using the following
heuristic to try to balance the volumes of the resulting clusters as much as possible: We
sort the nodes in decreasing order of their degree and then assign them step by step to the
cluster having the smallest volume. We break ties by using the cluster with the smallest ID
among those having the smallest volume. For this heuristic, we use the number of clusters
that our algorithm has computed as value for k. For the resulting clustering C ∗ we compute
the value E[cov(C ∗ )] and use B := 1 − E[cov(C ∗ )] as a value to normalize Q(C). Table 2
shows the resulting bound and compares them against our results.
Table 2 Columns from left to right: the best score computed of our evolutionary algorithm, the
approximate bound computed with the heuristic algorithm to balance cluster volumes and the ratio
of the two values.
Graph
as-22july06
astro-ph
celegans_metabol
cond-mat-2005
email
PGPgiantcompo
polblogs
power
smallworld
memplus
G_n_pin_pout
caidaRouterLevel
rgg_n17
luxembourg.osm
coAuthorsCiteseer
citationCiteseer
coPapersDBLP
belgium.osm
ldoor
eu-2005
in-2004
333SP
prefAttachment
Max. Q
B
Small Instances
0,679 396 0,973 684
0,746 292 0,999 070
0,453 248 0,888 889
0,750 171 0,999 468
0,582 829 0,900 000
0,886 853 0,989 796
0,427 105 0,996 168
0,940 977 0,975 610
0,793 187 0,995 652
0,701 275 0,984 127
0,500 466 0,993 865
0,872 828 0,997 758
0,978 454 0,992 537
0,989 672 0,996 283
Large Instances
0,906 830 0,995 575
0,825 545 0,993 289
0,868 058 0,995 283
0,995 064 0,998 358
0,970 555 0,989 796
0,941 575 0,996 564
0,980 690 0,999 136
0,989 356 0,995 726
0,316 089 0,875 000
Q/B
0,697 758
0,746 987
0,509 904
0,750 570
0,647 588
0,895 996
0,428 748
0,964 501
0,796 651
0,712 586
0,503 555
0,874 789
0,985 811
0,993 364
0,910 861
0,831 123
0,872 172
0,996 701
0,980 561
0,944 821
0,981 538
0,993 603
0,361 245
| 9cs.NE
|
arXiv:1705.03597v1 [cs.AI] 10 May 2017
Solving Multi-Objective MDP with
Lexicographic Preference:
An application to stochastic planning with
multiple quantile objective
Yan Li1 and Zhaohan Sun2
1
2
School of Mathematics, Georgia Institute of Technology
H. Milton Stewart School of Industrial and Systems Engineering, Georgia
Institute of Technology
Abstract
In most common settings of Markov Decision Process (MDP), an agent evaluate a policy based on expectation of (discounted) sum of rewards. However in many applications
this criterion might not be suitable from two perspective: first, in risk aversion situation expectation of accumulated rewards is not robust enough, this is the case when
distribution of accumulated reward is heavily skewed; another issue is that many applications naturally take several objective into consideration when evaluating a policy,
for instance in autonomous driving an agent needs to balance speed and safety when
choosing appropriate decision. In this paper, we consider evaluating a policy based on
a sequence of quantiles it induces on a set of target states, our idea is to reformulate
the original problem into a multi-objective MDP problem with lexicographic preference naturally defined. For computation of finding an optimal policy, we proposed an
algorithm FLMDP that could solve general multi-objective MDP with lexicographic
reward preference.
1
Introduction
The most classical MDP problem consider maximizing a scalar reward’s expectation
[3], however in many situation a single scalar objective is not enough to represent an
agent’s objective. For example, in self-autonomous driving one need to balance speed
and safety [2]. A common approach is to use a weight vector and scalarization function to project the multi-objective function to single objective problem. However in
practice it is hard to evaluate and analyze the projected problem since there might be
many viable Pareto optimal solutions to the original problem [1]. On the other hand,
1
in some cases, an agent might have explicit preference over the objectives, that is, an
agent might expect to optimize the higher priority objective over the lower priority
ones when finding optimal policy. For example, in autonomous driving an agent would
consider safety the highest priority, placing speed in the second place.
Several previous studies have considered such multi-objective problem with lexicographical order. Using a technique called Ordinal dynamic programming, Mitten [5]
assumed a specific preference ordering over outcomes for a finite horizon MDP; Sobel
[7] extended this model to infinite horizon MDPs. Ordinal dynamic programming has
been explored under reinforcement learning. Wray et.al [2] also consider a more general
setting when lexicographical order depends on initial state and slack for higher objective value is allowed for improvement over lower priority objective. In their paper they
proposed an algorithm called LVI that tries to approximate optimal policy in infinite
horizon setting, although work empirically well, the algorithm lacks theoretical guarantee, in fact, the performance could be arbitrarily worse if the MDP is adversarially
designed.
Even in the setting that an agent indeed has only one reward, the expectation of
accumulated reward is not always suitable. This is the case when the agent is risk
aversion, for instance in financial market an institutional fund would like to design an
auto-trading system that maximize certain lower quantile. The essential idea of such
strategy is to improve the worst case situation as much as possible. Based on this
motivation, Hugo and Weng [8] proposed quantile based reinforcement learning algorithm which seeks to optimize certain lower/upper quantile on the random outcome.
In their paper they define a set of end states in finite horizon setting, let Pπ (·) be the
probability distribution induced by policy π on the end states, they seek to find the
optimal policy in the sense that the τ -lower quantile of Pπ (·) is maximized. Note that
their objective could be improved by following observation:
1. Among all the policy that achieve the optimal τ -lower quantile, a refined class
of policy could be chosen in the sense that following such policy, the probability of ending at a state the is less preferable than the optimal quantile state is
minimized.
2. Suppose τ1 < τ2 , then after finding policy class that maximize τ1 -quantile, one can
further find policy that maximize τ2 -quantile in this policy class. For situation
when multiple τi -quantile are to be optimized, we can find optimal policy by
repeating the same procedure iteratively.
In general, if τ1 < τ2 < . . . < τL are in consideration, we have a multi-quantileobjective MDP, in this paper, we showed a proper way to transfer this problem into
a pure multi-objective MDP with lexicographic preference. To tackle computation of
an optimal policy, we will introduce an algorithm called FLMDP that not only solve
our multi-quantile-objective MDP, but also generalize multi-objective MDP with finite
states, action, and horizon. Generalization to infinite states or actions to find ǫ-optimal
policy could be done fairly easy with small modification in our algorithm.
2
2
Problem Definition
We consider finite horizon problem here, a multi-objective Markov Decision Process is
described by a tuple (S, A, P, R) where:
• S is finite state space.
• A is finite action space.
• G is finite end state space.
• T is finite horizon.
• P is transition function given by: P (s, a, s′ ) = P(s′ |s, a), i.e., the probability of
transiting from s to s′ after performing action a.
• R = [R1 , R2 , . . . , Rk ] is reward vector, with each component Ri (s, a, s′ ) defining
reward of starting from state s, performing action a and transit to state s′ .
Without loss of generality we may assume G = {g1 , . . . , gn }. On G we may define our
preference as g1 6 g2 6 . . . 6 gn where gi 6 gj denotes gj is preferred over gi . To
enforce end state nature of set G, we further define transition probability and reward
function have following properties:
P (g, a, g) = 1, ∀g ∈ G, ∀a ∈ A
R(g, a, g) = 0, ∀g ∈ G, ∀a ∈ A
That is, whenever the current state is in set G, we remains at state g until process
ends at horizon T, in the meantime receiving no rewards at all. To enforce the process
ends at one of the end state, we define a special end state g0 = t = T and declare
g0 6 gi , ∀i > 1. Let π be any policy, we define the probability distribution Pπ (·)
induced by π induced on set G as Pπ (gi ) = Pπ (sT = gi ). Then we can further define
cumulative distribution function.
X
F π (g) =
Pπ (gi )
gi 6g
The associated τ -lower quantile is given by:
q πτ = min{gi : F π (gi ) > τ }
Finding Optimal Policy
Given τ1 < . . . < τL ∈ [0, 1], following our motivation in Introduction section, our
procedure to find optimal policy is a series of optimization procedure, we will show
later this could be reshaped into multi-objective MDP with lexicographic preference.
Algorithmic Scheme 1
1. Denote Π0 = {all possible policy}.
2. After finding Πi−1 , construct Πi :
q ⋆τ = max {q πτ }.
i
π∈Πi−1
i
Π̂i = {π ∈ Πi−1 : q πτ = q⋆τ }
i
3
i
That is, Π̂i is the set of policy that maximize the τi quantile in Πi . Let pi be the
”biggest” state that is ”smaller” than q ⋆τ . Here ”biggest” and ”smaller” should
i
be interpreted in terms of preference. Then to minimize the probability of ending
at a state that is less preferable than q ⋆τ , we should have Πi as follows:
i
Πi = argmin F π (pi )
(1)
π∈Π̂i
3. Proceed as step 2 until we have found ΠL . Then any policy π that is in ΠL will
be our optimal policy.
Note however it is unclear how to translate such algorithmic scheme into a tractable
algorithm, the problem is that we do not know how to properly ”choose” an policy
from a policy class. We’ll tackle this issue in the next section.
3
Multi-Quantile-Objective MDP
In this section we first present a lemma that generalize the Lemma 1 of Hugo and Weng
[8], this lemma fully characterize the q ⋆τ
i
Lemma 1. For i = 1, . . . , L, let
following condition:
q⋆τ
i
and Πi be defined as before, then q⋆τ satisfies the
i
q ⋆τ = min{g : Fi⋆ (g) > τi }
i
Fi⋆ (g)
= min F π (g), ∀g ∈ G
π∈Πi−1
Proof. We proof by induction:
For i=1: observe that
F1⋆ (g) 6 F π (g),
∀π, ∀g
This follows directly from the definition of F1⋆ (g). Hence the τ1 -quantile of F1⋆ (g)(denoted
as gi1 ) is greater or equal than q πτ for all π.
1
Now by the definition of gi1 , we have F1⋆ (gi1 ) > τ1 and F1⋆ (gi1 −1 ) < τ1 . Then by
definition of F1⋆ (), we have ∃π1 , s.t.:
F π1 (gi1 −1 ) = F1⋆ (gi1 −1 ) < τ1
F π1 (gi1 ) > F1⋆ (gi1 ) > τ1
This means that gi1 is τ1 -quantile of both F1⋆ () and F π1 (). Hence we have gi1 > q πτ , ∀π,
1
and gi1 = q πτ 1 . Thus q ⋆τ = gi1 by definition of q ⋆τ .
1
1
1
Assume the claim holds for i < k:
For i = k: observe that
Fk⋆ (g) 6 F π (g),
∀π ∈ Πk−1 , ∀g
Hence the τk -quantile of Fk⋆ (g)(denoted as gik ) is greater or equal than q πτ for all
k
π ∈ Πk−1 .
4
Now by the definition of gik , we have Fk⋆ (gik ) > τk and Fk⋆ (gik −1 ) < τk . Then by
definition of Fk⋆ (), we have ∃πk ∈ Πk−1 , s.t.:
F πk (gik −1 ) = Fk⋆ (gik −1 ) < τk
F πk (gik ) > Fk⋆ (gik ) > τk
This means that gik is τk -quantile of both Fk⋆ () and F πk (). Hence we have gik > q πτ , ∀π,
k
and gik = qπτ k . Thus q ⋆τ = gik by definition of q⋆τ . By induction, proof complete.
k
k
k
Following the proof of Lemma 1, we could construct Πi as follows:
Algorithmic Scheme 2
1. Let Π0 ={all possible policy}.
2. Suppose Πi−1 has been constructed, then we construct Πi as following:
q ⋆τ = max {q πτ }.
i
π∈Πi−1
i
Let pi the same as before, i.e. pi be the ”biggest state” that is ”smaller” than
q ⋆τ . Then we construct Πi as follows:
i
Πi = argmin F π (pi )
(2)
π∈Πi−1
Note that in equation (2) we construct Πi here directly from Πi−1 instead of from
Π̂i in equation (1), the reason here is that by proof of Lemma 1, the policy π that
minimize F π (pi ) also has q ⋆τ as its τi -quantile.
i
Solving the Algorithmic Scheme mentioned before is hard in general, but giving our
work before we are now ready to formulate the previous Algorithmic Scheme into a
MDP with Lexicographical objective preference. We may now restrict ourself in the
setting that q ⋆τ , q ⋆τ · · · q ⋆τ are known beforehand, and consider the more general case
1
2
L
later.
To do this, we define reward functions {Ri }L
i=1 as follows:
(
1 if st 6∈ G and st+1 = gi , gi > q ⋆τ
i
Ri (st , at , st+1 ) =
0 otherwise
(3)
P
Then it is easy to verify that Eπ [ Tt=0 Ri (st , at , st+1 )] = 1 − F π (pi ). Hence minimizing
F π (pi ) is equivalent
PT to maximizing expected reward of the MDP.
π
π
Define Vi = E [ t=1 Ri (st , at , st+1 )] the expected total reward corresponding to reward function Ri , then equation (2) becomes as:
Πi = argmax Viπ
(4)
π∈Πi−1
We will show in the next subsection, if {q ⋆τ }L
i=1 are known, the procedure described in
i
Algorithmic Scheme 2 exactly corresponds to the procedure of solving a multi-objective
MDP with lexicographic preference.
5
Multi-Objective MDP with Lexicographic Preference
Definition 1. Recall that a point ū is lexicographical larger than 0 if ui = 0 for i =1,2
· · · j and uj > 0 for some 1 6 j 6 n, we write u = (u1 , u2 · · · un ) >l 0. We then define
our lexicographical order index as j, which is the first index in the vector that strictly
larger than zero. Thus say ū is lexicographical larger than v̄ if ū − v̄ >l 0
A multi-objective MDP differs from standard MDP that it has reward vector R(s, a) =
[R1 (s, a), . . . , RL (s, a)] and associated value vector V(s) = [V1 (s), . . . , VL (s)], and a
preference is defined on the value function associated with different rewards, say
V1 (s) > V2 (s) > . . . > VL (s). Classic multi-objective MDP seeks to find a policy
that has Pareto optimal value vector. With lexicographic preference defined on value
vectors, we say a policy π ⋆ is lexicographic optimal if there is no policy π so that
⋆
Vπ (s) >l Vπ (s).
In pure algorithmic scheme, an multi-objective MDP is solved by iteratively finding
the optimal policy class for lower priority value function in the optimal policy class for
higher priority ones. That is, denote Π0 ={any policy}, Πi+1 is found by:
π
(s)
Πi+1 = argmax Vi+1
π∈Πi
In our multi-quantile-objective MDP, (S, A, P, R) is defined as the same as in section
2, the reward functions Ri is defined as in equation (3). With value vector Vπ =
(V1π , . . . , VLπ ), we define lexicographical preference on Vπ as defined in definition 1.
Then with equation (4) replacing equation (2) in Algorithmic 2, it is easy to see that
Algorithmic Scheme 2 now become a procedure of solving multi-objective lexicographic
MDP with parameters (S, A, P, R, V).
4
Solving Multi-Quantile-Objective MDP
Solving Multi-Quantile-Objective MDP lies in general situation of solving multi-objective
MDP with lexicographical preference. A natural one is to shape the original problem to a sequence of constrained case MDP and solve this sequence of constrained
MDP iteratively. In the next subsection we proposed an algorithm that can solve general multi-objective MDP with lexicographic preference directly, thus solving multiquantile-objective MDP here is just a special case.
Constrained MDP formulation
The following procedure reshape a multi-objective MDP with lexicographic preference
to a sequence of constrained MDP problem.
1. At step 1, Π0 ={all possible policy}. Optimize objective V1π , V1⋆ = maxπ∈Π0 V1π .
2. At step i, Optimize objective Viπ with constraints:
minimize
π
Viπ
subject to Vjπ > Vj⋆ , j = 1, . . . , i − 1.
6
3. Proceed as in 2 until step L is finished.
It is easy to see that at step i the constraints in the optimization procedure naturally restrict the algorithm to search policy in the class that is identical to Πi−1 , thus
correctness of this reshape is guaranteed. Altman [6] has shown that an optimal randomized policy could be found in such constrained MDP, Chen and Feinberg [4] also
showed how to find optimal deterministic policy. Note this type of algorithm indeed
does unnecessary work by restarting from searching whole policy space in every step.
In this next subsection, we design a dynamic programming flavor algorithm that finds
an optimal deterministic policy for general lexicographic order MDP.
Lexicographic Markov Decision Process
In this subsection we introduce an algorithm FLMDP that solves general lexicographic
MDP in finite horizon, in particular it can be used to solve our previous formulated
multi-quantile-objective MDP.
π : L × S × T → R be the expected reward obtained by using policy π in decision
Let Vi,t
epochs t, t+1, · · · T, here, for simplicity, we let reward of end state equals zero, thus
π can be represented as
Vi,t
T
X
π
Ri (sn , an )]
Vi,t
(s) = Eπst =s [
n=t
Note that although in our problem Ri () relates to out next state, we can solve this
problem by simply define Ri (st , at ) = E[Ri (st , at , st+1 )] with expectation taken w.r.t
st+1 .
We first define state value function:
Qπi,t (s, a) = Ri (s, a) +
X
π
P r(s′ |s, a)Vi,t+1
(s′ )
s′ ∈S
Then following the definition in constrained MDP, ∀t = T, T −1 · · · 1, and ∀i = 1, 2 · · · L,
we define restricted bellman equation operator Bit as
X
π
π
Bit Vi,t
(s) = max {Ri (s, a) +
P r(s′ |s, a)Vi,t+1
(s′ )}
a∈Ati−1
s′ ∈S
where
Ati+1 (s) = {a ∈ Ati (s)| max Qπi,t (s, a′ ) = Qπi,t (s, a)}
a′ ∈Ati (s)
and
At0 (s) = A(s)
7
Algorithm 1 Finite-horizon Lexicographic MDP - FLMDP
Input Ri (s, a), i = 1, 2 · · · L
π
Set Vi,T
(s) = 0, ∀i = 1, 2 · · · L, ∀s ∈ S
for t = T − 1, T − 2 · · · 1 do
for i = 1, 2 · · · L do
Vi,tπ (s) = Bit Vi,tπ (s)
end for
πt⋆ ∈ AtL
end for
Output π1⋆ , π2⋆ , · · · , πT⋆
Theorem 1. In our algorithm 1, ∀t = T − 1, T − 2, · · · 1, {πt⋆ }t6T −1 are optimal policy
for our Lexicographic MDP problem.
Proof. Before beginning our proof, we need some notations. Recall:
π
Vi,t
(s)
=
T
X
Eπst =s [
t
T
X
⋆
Vi,t
(s) = Eπst⋆=s [
Ri (st , at )]
Ri (st , at )]
t
π
π
(s)]
Vtπ (s) = [V1,t
(s), . . . , VL,t
⋆
⋆
(s)]
(s), . . . , VL,t
Vt⋆ (s) = [V1,t
π (s) defines the value
where {πt⋆ } denotes the policy output by Algorithm 1. Then Vj,t
function associated with reward Ri , starting a tail problem with initial state s at time
t following given policy π. Note that Vtπ (s) is exactly the value vector for full horizon
MDP with initial state s. By our specification of reward function Ri , we naturally have
VTπ (s) = 0 and VT⋆ (s) = 0.
Let 6l , <l , >l , >l denotes lexicographical order relationship on value vector Vtπ (s). We
use backward induction to show that for ∀ π, and for ∀ t = 1, . . . , T − 1, for ∀ s, we have
Vtπ (s) 6l Vt⋆ (s).
For t = T − 1, VTπ −1 (s) 6l VT⋆ −1 (s) is trivial by procedure of our algorithm. A simple
induction on i suffice to give a formal proof, we omit the details here.
Suppose the claim holds for t + 1, . . . , T − 1, now we proceed to prove the claim holds
π (s) < V⋆ (s)
for t: assume Vt+1
l
t+1
π
⋆
Vi,t+1
(s) = Vi,t+1
(s),
Viπt+1 ,t+1 (s)
<l Vi⋆t+1 ,t+1 (s)
We next show that Vtπ (s) <l Vt⋆ (s) also holds:
8
i = 1, . . . , it+1 − 1
π (s) < V ⋆ (s), then we are done.
1. if V1,t
1,t
π (s) = V ⋆ (s), construction of π ⋆ and value iteration for finite horizon MDP
2. if V1,t
1,t
gives us:
X
π
π
V1,t
(s) = R1 (s, π(s)) +
P (s, π(s), j)V1,t+1
(j)
j
⋆
V1,t
(s)
= max R1 (s, a) +
a
X
⋆
P (s, a, j)V1,t+1
(j)
j
π (s) 6
⋆
π
(j), then we must have V1,t
(j) = V1,t+1
By induction hypothesis we have V1,t+1
⋆
V1,t (s). Now since we have equality achieved, by our definition of A1 (s) in our
algorithm, we must have π(s) ∈ A1 (s).
3. We now use induction to show that for if i < it+1 − 1, and
π
⋆
(s) = Vj,t
(s), j = 1, . . . , i
Vj,t
π
⋆
then we must have π(s) ∈ Ai (s) and Vi+1,t
(s) ≤ Vi+1,t
(s). The base case i=1 have
been proved in step 2. Suppose the claim holds for i − 1, then for i:
By induction hypothesis we have π(s) ∈ Ai−1 (s). Construction of π ⋆ and value
iteration for finite horizon MDP gives us:
X
π
π
P (s, π(s), j)Vi,t+1
(j)
(5)
Vi,t
(s) = Ri (s, π(s)) +
j
⋆
Vi,t
(s)
=
max
a∈Ai−1 (s)
Ri (s, a) +
X
⋆
P (s, a, j)Vi,t+1
(j)
(6)
j
π
⋆
π (s) 6
By induction hypothesis we have Vi,t+1
(j) = Vi,t+1
(j), then we must have Vi,t
⋆
Vi,t (s). Now since we have equality achieved, by our definition of Ai (s) in our
algorithm, we must have π(s) ∈ Ai (s). Then replacing i equation (5) with i + 1
we have:
X
π
π
Vi+1,t
(s) = Ri+1 (s, π(s)) +
P (s, π(s), j)Vi+1,t+1
(j)
j
⋆
Vi+1,t
(s) = max Ri+1 (s, a) +
a∈Ai (s)
X
π
P (s, a, j)Vi+1,t+1
(j)
j
π
⋆
By induction hypothesis we have Vi+1,t+1
(j) = Vi+1,t+1
(j), noticing that now
π(s) ∈ Ai (s), then we have:
π
⋆
Vi+1,t
(s) 6 Vi+1,t
(s)
Finally, when i = it+1 − 1, if
π
⋆
Vj,t
(s) = Vj,t
(s), j = 1, . . . , i
π
Then following the argument as before, and utilizing that now Vi+1,t+1
(j) <
⋆
Vi+1,t+1 (j), we must have:
π
⋆
Vi+1,t
(s) < Vi+1,t
(s)
which gives us Vtπ (s) <l Vt⋆ (s)
9
π (s) = V⋆ (s) ⇒
Notice that our previous argument could also be used to prove Vt+1
t+1
π
⋆
Vt (s) 6l Vt (s). Then combining all the ingredients we have, the following statement
holds:
π
⋆
Vt+1
(s) 6l Vt+1
(s) ⇒ Vtπ (s) 6l Vt⋆ (s)
(7)
To conclude our proof, notice we have VTπ −1 (s) 6l VT⋆ −1 (s), apply equation (7) iteratively, we have V1π (s) 6l V1⋆ (s), the optimality of our output policy follows immediately.
Now we return to the general case where optimal quantiles {q ⋆τ }L
is not known before
i i=1
hand. Out idea is to proceed iteratively, at kth iteration, we used bisection to guess
the location of the unknown q ⋆τ , we then solve a lexicographic MDP with k reward
k
[R1 , . . . , Rk ] and preference aligns with our preference for total L reward. Specifically,
at k-th iteration, we maintain u and l such that Fk⋆ (gl−1 ) < τk and Fk⋆ (gu−1 ) > τk . We
successively reduce u − l by half until u − l = 1. Then q ⋆τ = gu−1 .
k
To proceed the k-th iteration, we need to define our reward function as follows:
(
1 if st 6∈ G and st+1 = gi , gi > qτi
q τi
Ri (st , at , st+1 ) =
0 otherwise
The reward vector at k-th iteration is then given by:
q ⋆τ
q ⋆τ
qτ
k−1
R = [R1 1 , . . . , Rk−1
, Rk k ]
where q τ is our guess for q ⋆τ .
k
k
Algorithm 2 Multi-Quantile-Objective(MQO) MDP
π
Set Vi,T
(s) = 0, ∀i = 1, 2 · · · L, ∀s ∈ S
for i = 1, 2 · · · L do
Guess a proper q τ , which should be larger than q ⋆τ , ∀k = i − 1 · · · 1
i
k
Set l be the largest index of {gk } s.t. gk < q τ , Set u ← n
i
repeat
Solve Lexicographic MDP with q ⋆τ , q⋆τ · · · q ⋆τ , j 6 i − 1
1
2
j
Output Vi,tπ (s)
if Vi,tπ (s) 6 1 − τi then
l←i
else
u←i
end if
i ← ⌈ l+u
⌉
2
until u − l = 1
q ⋆τ ← gu−1
i
end for
5
Conclusion
In this paper we consider a multi-quantile-objective MDP problem that combines previous work in quantile objective MDP and multi-objective MDP. Our contribution is
10
two folds, first we formulate the problem into multi-objective MDP problem, the second is that our algorithm to solve this problem could also solve general multi-objective
MDP problem with finite horizon, state space and action space. Extension to infinite
state space or action space could be also done with slight modification.
We note our possible future work here: Pineda et.al [1] has showed that constrained
MDP could be reshaped into a sequence of multi-objective MDP with lexicographic
preference and additional slack variables, thus if one could solve lexicographic MDP
with slack variable efficiently, then solution of constrained MDP follows. For finite
horizon, we believe similar dynamic programming flavor algorithm could be invented
for solving lexicographic MDP with slack variables, we leave it here as an open problem
and our future work.
References
[1] Luis Pineda, Kyle H. Wray, and Shlomo Zilberstein. Revisiting Multi-Objective
MDPs with relaxed Lexicographic Preferences. AAAI Fall Symposium on Sequential Decision Making for Intelligent Agents, pages 63–68,Arlington, Virginia, USA,
November 2015.
[2] Kyle H. Wray, Shlomo Zilberstein, and Abdel-Illah Mouaddib. Multi-Objective
MDPs with Conditional Lexicographic Reward Preferences. In Proceedings of
the Twenty-Ninth Conference on Artificial Intelligence (AAAI), pages 3418–3424,
Austin, TX, USA, January 2015.
[3] Feinberg, Eugene A., and Adam Shwartz, eds. Handbook of Markov decision processes: methods and applications. Vol. 40. Springer Science & Business Media, 2012.
[4] Richard C. Chen, and Eugene A. Feinberg. Non-Randomized Policies for Constrained Markov Decision Processes. Mathematical Methods of Operations Research.
[5] L. G. Mitten. Preference Order Dynamic Programming. Management Science. Volume 21, Issue 1, Pages 43-46, September 1974.
[6] Eitan Altman. Constrained Markov Decision Process. Chapman and Hall/CRC,
1999.
[7] Matthew J. Sobel. Ordinal Dynamic Programming. Management Science. Vol. 21,
No. 9, Theory Series, pp. 967-975, May, 1975.
[8] Hugo Gilbert, Paul Weng. Quantile Reinforcement Learning. In ACM 2016
[9] Sutton, Richard S., and Andrew G. Barto. Reinforcement learning: An introduction. Vol. 1. No. 1. Cambridge: MIT press, 1998.
[10] Weng, Paul. Markov Decision Processes with Ordinal Rewards: Reference PointBased Preferences. In ICAPS. 2011.
11
| 2cs.AI
|
arXiv:1707.09763v1 [math-ph] 31 Jul 2017
DIFFERENTIAL ALGEBRA
AND
MATHEMATICAL PHYSICS
J.-F. Pommaret
CERMICS, Ecole des Ponts ParisTech,
6/8 Av. Blaise Pascal, 77455 Marne-la-Vallée Cedex 02, France
E-mail: [email protected]
URL: http://cermics.enpc.fr/∼pommaret/home.html
ABSTRACT
Many equations of mathematical physics are described by differential polynomials, that is by
polynomials in the derivatives of a certain number of functions. However, up to the knowledge of
the author, differential algebra in a modern setting has never been applied to study the specific
algebraic feature of such equations. The purpose of this short but difficult paper is to revisit a
few domains like general relativity, conformal geometry and contact geometry in the light of a
modern approach to nonlinear systems of partial differential equations, using new methods from
Differential Geometry (D.C. Spencer, 1970), Differential Algebra (J.F. Ritt, 1950 and E. Kolchin,
1973) and Algebraic Analysis (M. Kashiwara, 1970). Identifying the differential indeterminates of
Ritt and Kolchin with the jet coordinates of Spencer, the idea is to study Differential Duality by
using only linear differential operators with coefficients in a differential field. In particular, the
linearized second order Einstein equations are parametrizing the first order Cauchy stress equations
but cannot themselves be parametrized. In the framework of Homological Algebra, this result is
not coherent with the vanishing of certain first and second extension modules. As a byproduct,
we shall prove that gravitation and electromagnetism must only depend on the second order jets
(elations) of the system of conformal Killing equations. Finally, we shall use these new methods in
order to study contact transformations in arbitrary odd dimension and apply these results to the
study of Hamilton-Jacobi equations in mechanics.
KEY WORDS
Nonlinear partial differential equations, Differential algebra, Riemannian stucture, Conformal
structure, Contact structure, Hamilton-Jacobi equation, General relativity, Einstein equations,
Elastic waves, Lie group, Lie pseudogroup, Differential Galois theory, Spencer operator, Algebraic
Analysis, Differential modules, Homological algebra, Extension modules.
1
1) INTRODUCTION
Let us consider a manifold X of dimension n with local coordinates x = (xi ) = (x1 , ..., xn ),
tangent bundle T , cotangent bundle T ∗ , vector bundle Sq T ∗ of q-symmetric covariant tensors and
vector bundle ∧r T ∗ of r-skewsymmetric covariant tensors or r-forms. The group of isometries
y = f (x) of the non-degenerate metric ω with det(ω) 6= 0 on X is defined by the nonlinear first
order system of general finite Lie equations in Lie form:
ωkl (f (x))∂i f k (x)∂j f l (x) = ωij (x)
Linearizing at the identity transformation y = x, we may introduce the corresponding Killing operator T → S2 T ∗ : ξ → Dξ = L(ξ)ω = Ω, which involves the Lie derivative L and provides twice the
so-called infinitesimal deformation tensor of continuum mechanics when ω is the Euclidean metric.
We may consider the linear first order system of general infinitesimal Lie equations in Medolaghi
form, also called system of Killing equations:
Ωij ≡ (L(ξ)ω)ij ≡ ωrj (x)∂i ξ r + ωir (x)∂j ξ r + ξ r ∂r ωij (x) = 0
which is in fact a family of systems only depending on the geometric object ω and its derivatives. Introducing the Christoffel symbols γ, we may differentiate once and add the operator
L(ξ)γ = Γ ∈ S2 T ∗ ⊗ T with the well known Levi-Civita isomorphism j1 (ω) = (ω, ∂x ω) ≃ (ω, γ) in
order to obtain the linear second order system of general infinitesimal Lie equations in Medolaghi
form:
k
k
r
k
Γkij ≡ (L(ξ)γ)kij ≡ ∂ij ξ k + γrj
(x)∂i ξ r + γir
(x)∂j ξ r − γij
(x)∂r ξ k + ξ r ∂r γij
(x) = 0
We have thus linearized a nonlinear differential algebraic system in order to obtain a linear system
with coefficients in the differential field Q(ω, ∂ω, ...) along the idea of E. Vessiot [65]. This system
is formally integrable if and only if ω has a constant Riemannian curvaure.
Similarly, introducing the Jacobian determinant ∆(x) = det(∂i f k (x)), the group of conformal
transformations of the metric ω may be defined by the nonlinear first order system of general finite
Lie equations in Lie form:
2
ω̂kl (f (x))∆− n (x)∂i f k (x)∂j f l (x) = ω̂ij (x)
1
while introducing the metric density ω̂ij = | det(ω) |− n ωij ⇒ det(ω̂) = ±1 as a new geometric
object, rather than by eliminating a conformal factor as usual. The conformal Killing operator
ξ → D̂ξ = L(ξ)ω̂ = Ω̂ may be defined by linearization as above and we obtain the first order
system of general infinitesimal Lie equations in Medolaghi form, also called system of conformal
Killing equations:
2
ω̂ij (x)∂r ξ r + ξ r ∂r ω̂ij (x) = 0
n
as a system with coefficients in the differential field Q(ω̄, ∂ ω̄, ...). We may introduce the trace
−1
tr(Ω) = ω ij Ωij with standard notations and obtain therefore tr(Ω̂) = 0 because Ω̂ij = | det(ω) | n (Ωij −
1
n ωij tr(Ω)). This system becomes formally interable if and only if the corresponding Weyl tensor
vanishes.
Ω̂ij ≡ ω̂rj (x)∂i ξ r + ω̂ir (x)∂j ξ r −
The reader may look at [31,32,33,34,44] for finding examples of Lie groups or Lie pseudogroups
of transformations along the approach initiated by E. Vessiot in 1903 [53].
In classical elasticity, the stress tensor density σ = (σ ij = σ ji ) existing inside an elastic body is
a symmetric 2-tensor density introduced by A. Cauchy in 1822. Integrating by parts the implicit
i
σ rs = f i . When ω is the
summation − 21 σ ij Ωij , we obtain the Cauchy operator σ → ∂r σ ir + γrs
euclidean metric, the corresponding Cauchy stress equations can be written as ∂r σ ir = f i where
the right member describes the local density of forces applied to the body, for example gravitation.
With zero second member, we study the possibility to ”parametrize ” the system of PD equations
∂r σ ir = 0, namely to express its general solution by means of a certain number of arbitrary functions or potentials, called stress functions. Of course, the problem is to know about the number of
2
such functions and the order of the parametrizing operator. For n = 1, 2, 3 one may introduce the
Euclidean metric ω = (ωij = ωji ) while, for n = 4, one may consider the Minkowski metric. A few
definitions used thereafter will be provided later on.
When n = 2, the stress equations become ∂1 σ 11 + ∂2 σ 12 = 0, ∂1 σ 21 + ∂2 σ 22 = 0. Their second
order parametrization σ 11 = ∂22 φ, σ 12 = σ 21 = −∂12 φ, σ 22 = ∂11 φ has been provided by George
Biddell Airy (1801-1892) in 1863 [1]. It can be simply recovered in the following manner:
σ 12
∂1 σ 11 − ∂2 (−σ 12 ) = 0
∂2 σ 22 − ∂1 (−σ 21 ) = 0
= σ 21 ⇒ ∂1 ϕ − ∂2 ψ = 0
∃ϕ, σ 11 = ∂2 ϕ, σ 12 = −∂1 ϕ
∃ψ, σ 22 = ∂1 ψ, σ 21 = −∂2 ψ
∃φ, ϕ = ∂2 φ, ψ = ∂1 φ
⇒
⇒
⇒
We get the linear second order system:
σ 11
−σ 12
σ 22
≡ ∂22 φ = 0
≡ ∂12 φ = 0
≡ ∂11 φ = 0
1
1
1
2
•
•
which is involutive with one equation of class 2, 2 equations of class 1 and it is easy to check that
the 2 corresponding first order CC are just the stress equations.
When constructing a long prismatic dam with concrete as in [12,13] or in the Introduction of [36],
we may transform a problem of 3-dimensional elasticity into a problem of 2-dimensional elasticity
by supposing that the axis x3 of the dam is perpendicular to the river with Ωij (x1 , x2 ), ∀i, j = 1, 2
and Ω33 = 0 because of the rocky banks of the river are supposed to be fixed. We may introduce
the two Lamé constants (λ, µ) and the Poisson coefficient ν = λ/2(λ + µ) in order to describe the
usual constitutive relations of an homogeneous isotropic medium as follows (care: n = 3 ⇒ n = 2):
σ=
1
λ tr(Ω) ω + µ Ω, tr(Ω) = Ω11 + Ω22
2
⇒
µΩ = σ −
λ
tr(σ) ω, tr(σ) = σ 11 + σ 22
2(λ + µ)
even though σ 33 = 21 λ(Ω11 + Ω22 ) = 12 λtr(Ω) ⇒ σ 33 = ν(σ 11 + σ 22 ) 6= 0. Let us consider the right
square of the diagram below with locally exact rows:
2
..
.
2
Killing
−→
3
Riemann
−→
↓↑
Cauchy
←−
3
Airy
←−
1
..
.
1
Taking into account the formula (5.1.4) of [15] for the linearization of the only component of the
Riemann tensor at ω when n = 2 and substituting the Airy parametrization, we obtain:
tr(R) ≡ d11 Ω22 + d22 Ω11 − 2d12 Ω12 = 0
⇒
µ tr(R) ≡
λ + 2µ
∆∆φ = 0
2(λ + µ)
⇒
∆∆φ = 0
where the linearized scalar curvature tr(R) is allowing to define the Riemann operator in the previous diagram, namely the only compatibility condition (CC) of the Killing operator. Using now
the left square of the previous diagram, we may also substitute the Airy parametrization in the
Cauchy stress equations and get, when f~ = ~g (gravity)(care: n=3):
~
∇
~ + µ∆ξ~ = f~ ⇒
~ ∇.
~ ξ)
(λ + µ)∇(
(λ + 2µ)∆tr(Ω) = 0 ⇒ ∆tr(Ω) = 0 ⇒ ∆tr(σ) = 0 ⇒ ∆∆φ = 0
It remains to exhibit an arbitrary homogeneous polynomial solution of degree 3 and to determine
its 4 coefficients by the boundary pressure conditions on the upstream and downstream walls of
the dam. Of course, the Airy potential φ has nothing to do with the perturbation Ω of the metric
ω and the Airy parametrization is nothing else but the formal adjoint of the Riemann operator,
linearization of the Riemann tensor at ω, expressing the second order compatibility conditions (CC)
of the inhomogeneous system Dξ = Ω. Also, as we discover at once, the origin of elastic waves
is shifted by one step backwards, from the right square to the left square of the diagram. Indeed,
~ 2 for a medium with mass ρ per unit volume in the right member
using inertial forces f~ = ρ ∂ 2 ξ/∂t
3
of Cauchy stress equations because of Newton law, we discover the existence of two types of elastic
waves, namely the longitudinal and transversal waves with different speeds vT < vL that are really
existing because they are responsible for earthquakes [36]:
q
~ ξ~ = 0
∇.
⇒
µ△ξ~ = f~
⇒
vT = µρ
q
∇
~ ∧ ξ~ = 0
⇒
(λ + 2µ)△ξ~ = f~
⇒
vL = λ+2µ
ρ
It is this comment that pushed the author to use the formal adjoint of an operator, knowing
already that an operator and its (formal) adjoint have the same differential rank. In the case of the
conformal Killing operator, the second order CC are generated by the Weyl operator, linearization
of the Weyl tensor at ω̂ when n ≥ 4. The particular situation n = 3 will be studied and its corresponding 5 third order CC are not known after one century [44]. Finally, the Bianchi operator
describing the CC of the Riemann operator does not appear in this scheme.
Summarizing what we have just said, the study of elastic waves in continuum mechanics only
depends on group theory because it has only to do with one differential sequence and its formal adjoint, combined together by means of constitutive relations. We have proved in many books [33,34,]
and in [41,42,44] that the situation is similar for Maxwell equations, a result leading therefore to revisit the mathematical foundations of both General Relativity (GR) and Electromagnetism (EM),
thus also of Gauge Theory (GT).
The first motivation for studying the methods used in this paper has been a 1000$ challenge
proposed in 1970 by J. Wheeler in the physics department of Princeton University while the author
of this paper was a student of D.C. Spencer in the closeby mathematics department:
Is it possible to express the generic solutions of Einstein equations in vacuum by means of the
derivatives of a certain number of arbitrary functions like the potentials for Maxwell equations ?.
During the next 25 years and though surprising it may look like, no progress has been made
towards any solution and we found the negative solution of this challenge in 1995 [35]. Then, being
already in contact with M.P. Malliavin as I gave a seminar on the ”Deformation Theory of Algebraic and Geometric Structures ” [24,44], I presented in 1995 a seminar at IHP in Paris, proving
the impossibility to parametrize Einstein equations, a result I just found. One of the participants
called my attention on a recently published translation from japanese of the 1970 master thesis of
M. Kashiwara that he just saw on display in the library of the Institute [21]. This has been the
true starting of the story because I discovered that the duality involved in the preceding approach
to physics was only a particular example of a much more sophisticated framework having to do
with homological algebra [2,7,27,36,50].
Let us explain this point of view by means of an elementary example. With ∂22 ξ = η 2 , ∂12 ξ = η 1
for D, we get ∂1 η 2 −∂2 η 1 = ζ for the CC D1 . Then ad(D1 ) is defined by µ2 = −∂1 λ, µ1 = ∂2 λ while
ad(D) is defined by ν = ∂12 µ1 + ∂22 µ2 but the CC of ad(D1 ) are generated by ν ′ = ∂1 µ1 + ∂2 µ2 .
Using operators, we have the two differential sequences:
ξ
ν
D
−→
ad(D)
←−
ւ
η
µ
D
1
−→
ad(D1 )
←−
ζ
λ
ν′
where D1 generates the CC of D in the upper sequence but ad(D) does not generate the CC of
ad(D1 ) in the lower sequence, even though D1 ◦ D = 0 ⇒ ad(D) ◦ ad(D1 ) = 0, contrary to what
happened in the previous diagram. We shall see that this comment brings the need to introduce
the first extension module ext1 (M ) of the differential module M determined by D.
In the meantime, following U. Oberst [28,29], a few persons were trying to adapt these methods
to control theory and, thanks to J.L. Lions, I have been able to advertise about this new approach
in a european course, held with succes during 6 years [34] and continued for 5 other years in a
4
slightly different form [37]. By chance I met A. Quadrat, a good PhD student interested by control
and computer algebra and we have been staying alone because the specialists of Algebraic Analysis were pure mathematicians, not interested at all by applications. As a byproduct, it is rather
strange to discover that the impossibility to parametrize Einstein equations, that we shall prove
in Section 4, has never been acknowledged by physicists but can be found in a book on control because it is now known that a control system is controllable if and only if it is parametrizable [37,56].
The following example of a double pendulum will prove that this result, still not acknowledged
today by engineers, is not evident at all. For this, let us consider two pendula of respective length
l1 and l2 attached at the ends of a rigid bar sliding horizontally with a reference position x(t). If
the pendula move with a respective (small) angle θ1 (t) and θ2 (t) with respect to the vertical, it is
easy to prove from the Newton principle that the equations of the movements does not depend on
the respective masses m1 and m2 of the pendula but only depend on the respective lengths and
gravity g along the two formulas:
d2 x + l1 d2 θ1 + gθ1 = 0,
d2 x + l2 d2 θ2 + gθ2 = 0
where d = dt is the standard time derivative. It is experimentally visible and any reader can check
it with a few dollars, that the system is controllable, that is the angles can reach any prescribed
(small) values in a finite time when starting from equilibrium, if and only if l1 6= l2 and, in this
case, we have the following (injective) 4th order parametrization:
−l1 l2 d4 φ − g(l1 +2 )d2 φ − g 2 φ = x,
l2 d4 φ + gd2 φ = θ1 ,
l1 d4 φ + gd2 φ = θ2
0f course, if l1 = l2 = l, the system cannot be controllable because, setting θ = θ1 − θ2 , we obtain
by substraction ld2 θ + gθ = 0 and thus θ(0) = 0, dθ(0) = 0 ⇒ θ(t) = 0.
We end this Introduction explaining on a simple example why the second extension module
ext2 (M ) must also be considered, especially in the study of Einstein equations, though surprising
it may look like. To make a comparison, let us consider the following well known Poincaré sequence:
d
d
d
d
d
∧0 T ∗ −→ ∧1 T ∗ −→ ∧2 T ∗ −→ ... −→ ∧n−1 T ∗ −→ ∧n T ∗ → 0
where d : ω = ωI dxI → ∂i ωI dxi ∧ dxI is the exterior derivative. When n = 3, we have:
d
d
d
ad(d)
ad(d)
ad(d)
∧0 T ∗ −→ ∧1 T ∗ −→ ∧2 T ∗ −→ ∧3 T ∗ → 0
0 ← ∧3 T ∗ ←− ∧2 T ∗ ←− ∧1 T ∗ ←− ∧0 T ∗
⇔
⇔
grad
curl
div
div
curl
grad
φ −→ ξ −→ η −→ ζ → 0
0 ← θ ←− ν ←− µ ←− λ
From their definition it follows that div is parametrized by curl while curl is parametrized by
grad. Also, in local coordinates, we have ad(div) = −grad, ad(curl) = curl, ad(grad) = −div and
the adjoint sequence is also the Poincaré sequence up to the sign. Let us nevertheless consider the
new (minimal) parametrization of div obtained by setting ξ 3 = 0, namely [45,46]:
d2 ξ 3 −d3 ξ 2 = η 1 , d3 ξ 1 −d1 ξ 3 = η 2 , d1 ξ 2 −d2 ξ 1 = η 3 ⇒ −d3 ξ 2 = η 1 , d3 ξ 1 = η 2 , d1 ξ 2 −d2 ξ 1 = η 3
If we define the differential rank of an operator by the maximum number of differentially independent second member, this is clearly an involutive differential operator with differential rank
equal to 2 because (ξ 1 , ξ 2 ) can be given arbitrarily and thus (η 1 , η 2 ) can be given arbitrarily or,
equivalently, because the differential rank of div is of course equal to 1 as div has no CC. Now,
the involutive system d3 ξ 2 = 0, d3 ξ 1 = 0, d1 ξ 2 − d2 ξ 1 = 0 canot be parametrized by one arbitrary
function because both ξ 1 and ξ 2 are autonomous in the sense that they both satisfy to at least one
partial differential equation (PDE). Accordingly, we discover that div can be parametrized by the
curl through 3 arbitrary functions (ξ 1 , ξ 2 , ξ 3 ) where ξ 3 may be given arbitrarily, the curl being
itself parametrized by the grad, but div can also be parametrized by another operator with less
arbitrary functions or potentials which, in turn, cannot be parametrized again. Such a situation is
similar to the one met in hunting rifles that may have one, two or more trigger mechanisms that
can be used successively. It happens that the possibility to have one parametrization of div is an
5
intrinsic property described by the vanishing of ext1 (M ) where the differential module M is determined by grad while the property to have two successive parametrizations is an intrinsic property
described by the vanishing of ext1 (M ) as we just said plus the vanishing of ext2 (M ), and so on. It
follows that certain parametrizations are ” better ” than others and no student should even imagine
the minimal parametrization of div that we have presented above. A similar procedure has been
adopted by J.C. Maxwell [25] and G. Morera [26] when they modified the parametrization of the
Cauchy stress equations obtained by E. Beltrami in 1892 (See [3, 9,10,43,47,48] for more details).
We now treat the case dim(X) = 3 as the case dim(X) = n = 2p + 1 ≥ 5 will need much more
work (See [39] for more details). Let us introduce the so-called contact 1-form α = dx1 − x3 dx2
and consider the Lie pseudogroup Γ ⊂ aut(X) of (local) transformations preserving α up to a
function factor, that is Γ = {f ∈ aut(X)|j1 (f )−1 (α) = ρα} where again jq (f ) is a symbolic way
for writing out the derivatives of f up to order q and α transforms like a 1-covariant tensor. It
may be tempting to look for a kind of ”object ” the invariance of which should characterize Γ.
Introducing the exterior derivative dα = dx2 ∧ dx3 as a 2-form, we obtain the volume 3-form
α ∧ dα = dx1 ∧ dx2 ∧ dx3 . As it is well known that the exterior derivative commutes with any
diffeomorphism, we obtain sucessively:
j1 (f )−1 (dα) = d(j1 (f )−1 (α)) = d(ρα) = ρdα + dρ ∧ α ⇒ j1 (f )−1 (α ∧ dα) = ρ2 (α ∧ dα)
As the volume 3-form α ∧ dα transforms through a division by the Jacobian determinant ∆ =
∂(f 1 , f 2 , f 3 )/∂(x1 , x2 , x3 ) 6= 0 of the transformation y = f (x) with inverse x = f −1 (y) = g(y),
the desired object is thus no longer a 1-form but a 1-form density ω = (ω1 , ω2 , ω3 ) transforming
like a 1-form but up to a division by the square root of the Jacobian determinant. We obtain the
nonlinear differential algebraic system of general finite Lie equations in Lie form:
ωk (y)(
∂(y 1 , ..., y n ) − 1 k
) 2 yi = ωi (x)
∂(x1 , ..., xn )
It follows that the infinitesimal contact transformations are vector fields ξ ∈ T = T (X) the tangent
bundle of X, satisfying the 3 so-called first order system of general infinitesimal Lie equations in
Medolaghi form:
Ωi ≡ (L(ξ)ω)i ≡ ωr (x)∂i ξ r − (1/2)ωi (x)∂r ξ r + ξ r ∂r ωi (x) = 0
When ω = (1, −x3 , 0), we obtain the special involutive system (See below for details):
∂3 ξ 3 + ∂2 ξ 2 + 2x3 ∂1 ξ 2 − ∂1 ξ 1 = 0, ∂3 ξ 1 − x3 ∂3 ξ 2 = 0, ∂2 ξ 1 − x3 ∂2 ξ 2 + x3 ∂1 ξ 1 − (x3 )2 ∂1 ξ 2 − ξ 3 = 0
with 2 equations of class 3 and 1 equation of class 2, a result leading thus to only 1 compatibility
conditions (CC) for the second members. Equivalently, we have the system:
Ω3 ≡ ξ31 − x3 ξ32 = 0
Ω2 ≡ ξ21 − x3 ξ22 + 12 x3 (ξ11 + ξ22 + ξ33 ) − ξ 3 = 0
Ω1 ≡ ξ11 − x3 ξ12 − 12 (ξ11 + ξ22 + ξ33 ) = 0
For an arbitrary ω, we may ask about the differential conditions on ω such that all the equations
of order r + 1 are only obtained by differentiating r times the first order equations, exactly like in
the special situation just considered where the system is involutive. We notice that, in a symbolic
way, ω ∧ dω is now a scalar c(x) providing the zero order equation ξ r ∂r c(x) = 0 and the condition
is c(x) = c = cst. The integrability condition (IC) is the Vessiot structure equation:
I(j1 (ω)) ≡ ω1 (∂2 ω3 − ∂3 ω2 ) + ω2 (∂3 ω1 − ∂1 ω3 ) + ω3 (∂1 ω2 − ∂2 ω1 ) = c
involving the only structure constant c like the Riemannian structure.
For ω = (1, −x3 , 0), we get c = 1. If we choose ω̄ = (1, 0, 0) leading to c̄ = 0, we may define
Γ̄ = {f ∈ aut(X)|j1 (f )−1 (ω̄) = ω̄} with infinitesimal transformations satisfying the involutive
system:
∂3 ξ 3 + ∂2 ξ 2 − ∂1 ξ 1 = 0,
6
∂3 ξ 1 = 0,
∂2 ξ 1 = 0
with again 2 equations of class 3 and 1 equation of class 2. The equivalence problem j1 (f )−1 (ω) = ω̄
cannot be solved even locally because this system cannot have any invertible solution. Indeed,
studying the system j1 (g)−1 (ω̄) = ω, we have to solve:
1
∂g 1
∂g 1
∂g 1
∂g 1
∂g 1
3 ∂g
+
y
=
0,
=
0
⇒
=
0,
=
0,
=0
∂y 2
∂y 1
∂y 3
∂y 1
∂y 2
∂y 3
by using crossed derivatives.
Using now the definition of contact transformations, we have the three equations:
(L(ξ)α)i ≡ αr ∂i ξ r + ξ r ∂r αi = ρ(x)αi
Eliminating the arbitrary factor ρ(x), we obtain the two linearly independent infinitesimal Lie
equations:
α2 αr ∂3 ξ r − α3 αr ∂2 ξ r + (α2 ∂r α3 − α3 ∂r α2 )ξ r = 0
α3 αr ∂1 ξ r − α1 αr ∂3 ξ r + (α3 ∂r α1 − α1 ∂r α3 )ξ r
=
0
which are nevertheless not in the Medolaghi form because the 1-form α is not a geometric object.
Multiplying on the left the first equation by the test function λ1 and the second by the test function
λ2 and integrating by part, we obtain for example, separating the terms involving only λ1 from
the terms involving only λ2 :
1
→ −∂3 (α1 α2 λ1 ) + ∂2 (α1 α3 λ1 ) + (α2 ∂1 α3 − α3 ∂1 α2 )λ1
ξ
2
ξ
→ ...
3
ξ
→ −∂3 (α2 α3 λ1 ) + ∂2 ((α3 )2 λ1 ) + (α2 ∂3 α3 − α3 ∂3 α2 )λ1
and:
1
ξ
ξ2
3
ξ
→ −∂1 (α1 α3 λ2 ) + ∂3 ((α1 )2 λ2 ) + (α3 ∂1 α1 − α1 ∂1 α3 )λ2
→ ...
→ −∂1 ((α3 )2 λ2 ) + ∂3 (α1 α3 λ2 ) + (α3 ∂3 α1 − α1 ∂3 α3 )λ2
that we may rewrite respectively as:
−(α1 α2 )∂3 λ1 + (α1 α3 )∂2 λ1 − (α1 (∂2 α3 − ∂3 α2 ) − α2 (∂3 α1 − ∂1 α3 ) − α3 (∂1 α2 − ∂2 α1 ))λ1
...
−(α2 α3 )∂3 λ1 + (α3 )2 ∂2 λ1 + 2α3 (∂2 α3 − ∂3 α2 )λ1
and:
−(α1 α3 )∂1 λ2 + (α1 )2 ∂3 λ2 + 2α1 (∂3 α1 − ∂1 α3 )λ2
...
−(α3 )2 ∂1 λ2 + (α1 α3 )∂3 λ2 + 2α3 (∂3 α1 − ∂1 α3 )λ2
Multiplying each first row on the left by −α3 , then each third row on the left by α1 in order
to eliminate the derivatives of λ, adding and collecting the results, we discover that λ2 strikingly
disappears and we only obtain for the kernel of the adjoint operator:
α3 [α1 (∂2 α3 − ∂3 α2 ) + α2 (∂3 α1 − ∂1 α3 ) + α3 (∂1 α2 − ∂2 α1 )]λ1 = 0
that is α3 I(j1 (α))λ1 =0 and all the possible permutations. As α 6= 0, then one at least of the three
components must not vanish and may even be supposed to be equal to 1 because α is defined uo to
a function factor. We get therefore I(j1 (α))λ = 0, that is cλ = 0 whenever the system is formally
integrable. We let the reader treat directly the standard case with α3 = 0 (care):
α = dx1 − x3 dx2 ⇒ ∂3 ξ 1 − x3 ∂3 ξ 2 = 0, ∂2 ξ 1 − x3 ∂2 ξ 2 + x3 (∂1 ξ 1 − x3 ∂1 ξ 2 )− ξ 3 = 0 ⇒ c = 1 ⇒ λ = 0
Though it is rather surprising at first sight, let us now explain why we shall need non trivial homological algebra in order to understand the previous results. Indeed, the last system R1 is neither
7
formally integrable nor involutive, even though it has an involutive symbol g1 defined by:
ξ31 − x3 ξ32 = 0,
ξ21 − x3 ξ22 + x3 ξ11 − (x3 )2 ξ12 = 0
(1)
It is now easy to check that the system R1 ⊂ J1 (T ) defined by the 3 PD equations:
3
Φ ≡ ξ33 + ξ22 + 2x3 ξ12 − ξ11
Φ2 ≡ ξ31 − x3 ξ32
1
Φ ≡ ξ21 − x3 ξ22 + x3 ξ11 − (x3 )2 ξ12 − ξ 3
=
=
=
0
0
0
1 2
1 2
1 2
3
3
•
is involutive with 2 equations of class 3 and 1 equation of class 2. Taking into account the relations
Φ1 = Ω2 + x3 Ω1 , Φ2 = Ω3 , Φ3 = −2Ω1 and substituting, we obtain the only first order CC:
d3 Φ1 − d2 Φ2 − x3 d1 Φ2 + Φ3 = 0 ⇔ (d2 Ω3 − d3 Ω2 ) − x3 (d3 Ω1 − d1 Ω3 ) + Ω1 = 0
and we recognize the linearization of the Vessiot structure equation, following exactly the same
procedure as the one used previously for the linearization of the constant Riemannian curvature.
However, in this new framework, we shall now prove and illustrate the following Lemma and striking Theorem (See next Sections for the definitions):
LEMMA 1.1: A (formally) surjective linear differential operator D defined over a differential field
K is defining a projective and thus torsion-free differential module M if and only if its (formal)
adjoint is (formally) injective.
Proof: In this specific situation, let us consider the finite free presentation over D = K[d]:
p
D
0 → F1 −→ F0 −→ M → 0
If M is projective, then it is well known that that such a sequence splits (See [2,7,17,27,36,50]
for more details or [23], Lemma 3.3, p 212). Then, applying homD (•, D) we get again the new
splitting sequence:
p∗
D∗
0 ← F1 ∗ ←− F0 ∗ ←− M ∗ ← 0
and obtain ker(ad(D)) = 0 in the operator sense or rather coker(D∗ ) = 0 in the module sense.
Conversely,we have already exhibited the long exact dual sequence:
D∗
p∗
0 ← N ← F1 ∗ ←− F0 ∗ ←− M ∗ ← 0
Accordingly, if N = 0, as the dual F ∗ of a free differential module F is again a free differential
module, thus a projective module, this sequence splits and M ∗ is thus a projective module.
Applying again homD (•, D), we have the commutative and exact diagram:
p
D
F0
↓
−→
D∗∗
F0 ∗∗
−→
0→
F1
↓
−→
0→
F1 ∗∗
−→
p∗∗
M
↓
→0
M ∗∗
→0
∗∗
Using the isomorphism F ≃ F when F has finite rank over D, we obtain an isomorphism
M ≃ M ∗∗ . As M ∗ is projective because F0∗ ≃ F1∗ ⊕ M ∗ , then M ∗∗ is also projective and thus M
is projective.
Q.E.D.
EXAMPLE 1.2: In the preceding contact situation, the system R1 is defined by 2 equa(1)
tions only while the system R1 is defined by 3 equations that we have provided. Accordingly, with K = Q(x1 , x2 , x3 ) and D = K[d] = K[d1 , d2 , d3 ], we obtain the free presentation
0 → D2 → D3 → M → 0 and we have seen that c 6= 0 ⇒ λ = 0. The parametrization:
ξ 1 = φ − x3 ∂3 φ, ξ 2 = −∂3 φ, ξ 3 = ∂2 φ + x3 ∂1 φ ⇒ φ = α(ξ) ⇒ L(ξ)α = ∂1 φ α
8
by means of an arbitrary function φ is well known and proves that M ≃ D. On the contrary, if we
choose ω = (1, 0, 0), then the system of special Medolaghi equations that we have exhibited shows
that ξ 1 is a torsion element and this system cannot be parametrized.
THEOREM 1.3: The possibility to parametrize the system of general Medolaghi equations only
depends on the structure constant c.
Proof: For any geometric object ω of order q and the corresponding system Rq (ω) of general
Medolaghi equations, let us now define an equivalence relation ω̄ ∼ ω ⇔ Rq (ω) = Rq (ω̄). In the
contact situation, we have first to study when we have ω̄r ξir − 12 ω̄i ξrr = 0 ⇔ ωr ξir − 12 ωi ξrr . Though
it looks like to be a simple algebraic problem, one needs an explicit computation or computer
algebra and we prefer to use another more powerful technique ([39], p 688). Introducing the completely skewsymmetrical symbol ǫ = (ǫi1 i2 i3 ) where ǫi1 i2 i3 = 1 if (i1 i2 i3 ) is an even permutation
of (123) or −1 if it is an odd permutation and 0 otherwise, let us introduce the skewsymmetrical
2-contravariant density ω ij = ǫijk ωk . Then one can rewrite the system of general infinitesimal Lie
equations R1 (ω) as :
1
−ω rj (x)ξri − ω ir (x)ξrj − ω ij (x)ξrr + ξ r ∂r ω ij (x) = 0
2
and we may exhibit a section ξri = ω is Ars with Ars = Asr and thus ξrr = 0. It is important to
notice that det(ω) = 0 when n = 2p + 1, contrary to the Riemann or symplectic case and ω cannot
therefore be used in order to raise or lower indices. As we must have R̄10 = R10 where the isotropy
π1
R10 is defined by the short exact sequence 0 → R10 → R1 →0 T → 0, the same section must satisfy
(ω̄ rj ω is + ω̄ ir ω js )Ars = 0, ∀Ars = Asr , and we must have (ω̄ rj ω is + ω̄ ir ω js )+(ω̄ sj ω ir + ω̄ is ω jr ) = 0.
Setting s = j, we get ω̄ rj ω ij = ω̄ ij ω rj ⇒ ω̄ ij (x) = a(x)ω ij (x). Substituting and substracting, we
get ω ij (x)ξ r ∂r a(x) = 0 ⇒ a(x) = a = cst 6= 0 because ω 6= 0 and one of the components at least
must be nonzero. Accordingly, the normalizer N (Θ) = {ξ ∈ T | L(ξ)ω = Aω, Ac = 0} and Θ is
of codimension 1 in its normalizer if c = 0 or N (Θ) = Θ if c 6= 0. For example, in the case of
a contact structure with c = 1, we have N (Θ) = Θ but, when ω = (1, 0, 0) ⇒ c = 0, we have
to eliminate the constant A among the equations ∂3 ξ 3 + ∂2 ξ 2 − ∂1 ξ 1 = −2A, ∂3 ξ 1 = 0, ∂2 ξ 1 = 0
and we may add the infinitesimal generator xi ∂i of a dilatation providing A = − 12 . As we have
already seen, the parametrization is only existing for c 6= 0. This is an ”open property ” because
ω̄ = aω, a = cst ⇒ c̄ = a2 c and thus any nonzero value of c can be reached because a 6= 0.
Q.E.D.
However, no one of the previous results can be extended to an arbitrary n = 2p + 1 ≥ 5.
It is clear from the beginning of this Introduction that an isometry is a solution of a nonlinear
system in Lie form [31,34,44] and that we have linearized this system at the identity transformation
in order to study elastic waves. However, in general, no explicit solution may be known but most
nonlinear systems of OD or PD equations of mathematical physics (constant riemannian curvature
is a good example in [14]) are defined by differential polynomials. This is particularly clear for
riemannian, conformal, complex, contact, symplectic or unimodular structures on manifolds [44].
Hence, in Section 2 we shall provide the main results that exist in the formal theory of systems of
nonlinear PD equations in order to construct a formal linearization. The proof of many results is
quite difficult as it involves delicate chases in 3-dimensional diagrams [31,34,36]. In physics, the
linear system obtained may have coefficients in a certain differential field and we shall need to
revisit differential algebra in Section 3 because Spencer and Kolchin never clearly understood that
their respective works could be combined. It will follow that the linear systems will have coefficients in a differential field K and we shall have to introduce the ring D = K[d] = K[d1 , ..., dn ] of
differential operators with coefficients in K, which is even an integral domain. This fact will be
particularly useful in order to revisit differential duality in Section 4 before applying it to the study
of conformal structures in Section 5, caring separately about the cases n = 3, n = 4 and n ≥ 5,
then to contact structures in Section 6, caring also separately about the cases n = 3 and n ≥ 5,
finally concluding in the last Section 7.
9
2) DIFFERENTIAL GEOMETRY
If X is a manifold with local coordinates (xi ) for i = 1, ..., n = dim(X), let E be a fibered manifold over X with dimX (E) = m, that is a manifold with local coordinates (xi , y k ) for i = 1, ..., n
and k = 1, ..., m simply denoted by (x, y), projection π : E → X : (x, y) → (x) and changes
of local coordinates x̄ = ϕ(x), ȳ = ψ(x, y). If E and F are two fibered manifolds over X with
respective local coordinates (x, y) and (x, z), we denote by E×X F the fibered product of E and
F over X as the new fibered manifold over X with local coordinates (x, y, z). We denote by
f : X → E : (x) → (x, y = f (x)) a global section of E, that is a map such that π ◦ f = idX but
local sections over an open set U ⊂ X may also be considered when needed. Under a change of
coordinates, a section transforms like f¯(ϕ(x)) = ψ(x, f (x)) and the derivatives transform like:
∂ f¯l
∂ψ l
∂ψ l
r
(ϕ(x))∂
ϕ
(x)
=
(x,
f
(x))
+
(x, f (x))∂i f k (x)
i
∂ x̄r
∂xi
∂y k
We may introduce new coordinates (xi , y k , yik ) transforming like:
ȳrl ∂i ϕr (x) =
∂ψ l
∂ψ l
(x,
y)
+
(x, y)yik
∂xi
∂y k
k
We shall denote by Jq (E) the q-jet bundle of E with local coordinates (xi , y k , yik , yij
, ...) = (x, yq )
called jet coordinates and sections fq : (x) → (x, f k (x), fik (x), fijk (x), ...) = (x, fq (x)) transforming
like the sections jq (f ) : (x) → (x, f k (x), ∂i f k (x), ∂ij f k (x), ...) = (x, jq (f )(x)) where both fq and
jq (f ) are over the section f of E. It will be useful to introduce a multi-index µ = (µ1 , ..., µn )
with length | µ |= µ1 + ... + µn and to set µ + 1i = (µ1 ..., µi−1 , µi + 1, µi+1 , ..., µn ). Finally, a
jet coordinate yµk is said to be of class i if µ1 = ... = µi−1 = 0, µi 6= 0. As the background will
always be clear enough, we shall use the same notation for a vector bundle or a fibered manifold and their sets of sections [31,36]. We finally notice that Jq (E) is a fibered manifold over X
with projection πq while Jq+r (E) is a fibered manifold over Jq (E) with projection πqq+r , ∀r ≥ 0 [, , ].
DEFINITION 2.1: A (nonlinear) system of order q on E is a fibered submanifold Rq ⊂ Jq (E)
and a global or local solution of Rq is a section f of E over X or U ⊂ X such that jq (f ) is a section
of Rq over X or U ⊂ X.
DEFINITION 2.2: When the changes of coordinates have the linear form x̄ = ϕ(x), ȳ = A(x)y,
we say that E is a vector bundle over X. Vector bundles will be denoted by capital letters C, E, F
and will have sections denoted by ξ, η, ζ. In particular, we shall denote as usual by T = T (X)
the tangent bundle of X, by T ∗ = T ∗ (X) the cotangent bundle, by ∧r T ∗ the bundle of r-forms
and by Sq T ∗ the bundle of q-symmetric covariant tensors. When the changes of coordinates have
the form x̄ = ϕ(x), ȳ = A(x)y + B(x) we say that E is an affine bundle over X and we define the
associated vector bundle E over X by the local coordinates (x, v) changing like x̄ = ϕ(x), v̄ = A(x)v.
DEFINITION 2.3: If the tangent bundle T (E) has local coordinates (x, y, u, v) changing like
l
∂ψ l
i
k
ūj = ∂i ϕj (x)ui , v̄ l = ∂ψ
∂xi (x, y)u + ∂y k (x, y)v , we may introduce the vertical bundle V (E) ⊂ T (E)
as a vector bundle over E with local coordinates (x, y, v) obtained by setting u = 0 and changes
∂ψ l
k
v̄ l = ∂y
k (x, y)v . Of course, when E is an affine bundle over X with associated vector bundle E
over X, we have V (E) = E ×X E. With a slight abuse of language, we shall set E = V (E) as a
vector bundle over E.
For a later use, if E is a fibered manifold over X and f is a section of E, we denote by f −1 (V (E))
the reciprocal image of V (E) by f as the vector bundle over X obtained when replacing (x, y, v)
by (x, f (x), v) in each chart. A similar construction may also be done for any affine bundle over
E. Loking at the transition rules of Jq (E), we deduce easily the following results:
PROPOSITION 2.4: Jq (E) is an affine bundle over Jq−1 (E) modeled on Sq T ∗ ⊗E E but we shall
not specify the tensor product in general.
10
PROPOSITION 2.5: There is a canonical isomorphism V (Jq (E)) ≃ Jq (V (E)) = Jq (E) of vector
k
bundles over Jq (E) given by setting vµk = v,µ
at any order and a short exact sequence:
q
πq−1
0 → Sq T ∗ ⊗ E → Jq (E) −→ Jq−1 (E) → 0
of vector bundles over Jq (E) allowing to establish a link with the formal theory of linear systems.
PROPOSITION 2.6: There is an exact sequence:
jq+1
D
0 → E −→ Jq+1 (E) −→ T ∗ ⊗ Jq (E)
k
is called the
where Dfq+1 = j1 (fq ) − fq+1 is over fq with components (Dfq+1 )kµ,i = ∂i fµk − fµ+1
i
(nonlinear) Spencer operator. As Jq+1 (E) ⊂ J1 (Jq (E)), there is an induced exact sequence:
jq
D
1
0 → E −→ Jq+1 (E) −→
T ∗ ⊗ Jq (E)/Sq+1 T ∗ ⊗ E
where D1 is called the first Spencer operator.
DEFINITION 2.7: If Rq ⊂ Jq (E) is a system of order q on E, then Rq+1 = ρ1 (Rq ) =
J1 (Rq ) ∩ Jq+1 (E) ⊂ J1 (Jq (E)) is called the first prolongation of Rq and we may define the
subsets Rq+r . In actual practice, if the system is defined by PDE Φτ (x, yq ) = 0 the first
k
∂Φτ /∂yµk = 0. accordingly,
prolongation is defined by adding the PDE di Φτ ≡ ∂i Φτ + yµ+1
i
τ
τ
τ
k
fq ∈ Rq ⇔ Φ (x, fq (x)) = 0 and fq+1 ∈ Rq+1 ⇔ ∂i Φ + fµ+1i (x)∂Φ /∂yµk = 0 as identities on X
or at least over an open subset U ⊂ X. Differentiating the first relation with respect to xi and
substracting the second, we finally obtain:
k
(x))∂Φτ /∂yµk = 0 ⇒ Dfq+1 ∈ T ∗ ⊗ Rq
(∂i fµk (x) − fµ+1
i
(1)
q+r+1
and the Spencer operator restricts to D : Rq+1 → T ∗ ⊗ Rq . We set Rq+r = πq+r
(Rq+r+1 ).
DEFINITION 2.8: The symbol of Rq is the family gq = Rq ∩ Sq T ∗ ⊗ E of vector spaces over
Rq . The symbol gq+r of Rq+r only depends on gq by a direct prolongation procedure. We may
define the vector bundle F0 over Rq by the short exact sequence 0 → Rq → Jq (E) → F0 → 0 and
we have the exact induced sequence 0 → gq → Sq T ∗ ⊗ E → F0 .
Setting aτk µ (x, yq ) = ∂Φτ /∂yµk (x, yq ) whenever | µ |= q and (x, yq ) ∈ Rq , we obtain:
gq = {vµk ∈ Sq T ∗ ⊗ E | aτk µ (x, yq )vµk = 0}, | µ |= q, (x, yq ) ∈ Rq
k
k
⇒ gq+r = ρr (gq ) = {vµ+ν
∈ Sq+r T ∗ ⊗ E | aτk µ (x, yq )vµ+ν
= 0}, | µ |= q, | ν |= r, (x, yq ) ∈ Rq
In general, neither gq nor gq+r are vector bundles over Rq .
On ∧s T ∗ we may introduce the usual bases {dxI = dxi1 ∧ ... ∧ dxis } where we have set
I = (i1 < ... < is ). In a purely algebraic setting, one has:
PROPOSITION 2.9: There exists a map δ : ∧s T ∗ ⊗ Sq+1 T ∗ ⊗ E → ∧s+1 T ∗ ⊗ Sq T ∗ ⊗ E which
restricts to δ : ∧s T ∗ ⊗ gq+1 → ∧s+1 T ∗ ⊗ gq and δ 2 = δ ◦ δ = 0.
k
k
.
dxI } and set (δω)kµ = dxi ∧ ωµ+1
Proof: Let us introduce the family of s-forms ω = {ωµk = vµ,I
i
τµ
τµ k
k
i
2
k
i
j
k
We obtain at once (δ ω)µ = dx ∧ dx ∧ ωµ+1i +1j = 0 and ak (δω)µ = dx ∧ (ak ωµ+1i ) = 0.
Q.E.D.
The kernel of each δ in the first case is equal to the image of the preceding δ but this may no
longer be true in the restricted case and we set:
s
s
s
s
s
DEFINITION 2.10: Let Bq+r
(gq ) ⊆ Zq+r
(gq ) and Hq+r
(gq ) = Zq+r
(gq )/Bq+r
(gq ) with H 1 (gq ) =
1
Hq (gq ) be the coboundary space im(δ), cocycle space ker(δ) and cohomology space at ∧s T ∗ ⊗ gq+r
of the restricted δ-sequence which only depend on gq and may not be vector bundles. The symbol
11
1
s
gq is said to be s-acyclic if Hq+r
= ... = Hq+r
= 0, ∀r ≥ 0, involutive if it is n-acyclic and finite
type if gq+r = 0 becomes trivially involutive for r large enough. In particular, if gq is involutive
and finite type, then gq = 0. Finally, Sq T ∗ ⊗ E is involutive for any q ≥ 0 if we set S0 T ∗ ⊗ E = E.
Having in mind the example of xyx − y = 0 ⇒ xyxx = 0 with rank changing at x = 0, we have:
PROPOSITION 2.11: If gq is 2-acyclic and gq+1 is a vector bundle over Rq , then gq+r is a
vector bundle over Rq , ∀r ≥ 1.
Proof: We may define the vector bundle F1 over Rq by the following ker/coker exact sequence
where we denote by h1 ⊆ T ∗ ⊗ F0 the image of the central map:
0 → gq+1 → Sq+1 T ∗ ⊗ E → T ∗ ⊗ F0 → F1 → 0
and we obtain by induction on r the following commutative and exact diagram of vector bundles
over Rq :
0
↓
0→
0→
0→
0
↓
gq+r+1
↓δ
T ∗ ⊗ gq+r
↓δ
∧2 T ∗ ⊗ gq+r−1
↓δ
∧3 T ∗ ⊗ Sq+r−2 T ∗ ⊗ E
→
→
→
=
Sq+r+1 T ∗ ⊗ E
↓δ
T ∗ ⊗ Sq+r T ∗ ⊗ E
↓δ
∧2 T ∗ ⊗ Sq+r−1 T ∗ ⊗ E
↓δ
∧3 T ∗ ⊗ Sq+r−2 T ∗ ⊗ E
→
→
→
0
↓
Sr+1 T ∗ ⊗ F0
↓δ
T ∗ ⊗ Sr T ∗ ⊗ F 0
↓δ
∧2 T ∗ ⊗ Sr−1 T ∗ ⊗ F0
→
→
0
↓
Sr T ∗ ⊗ F 1
↓δ
T ∗ ⊗ Sr−1 T ∗ ⊗ F1
where all the maps have been given after Definition 2.9. The image of the central map of the
top row is hr+1 = ρr (h1 ) and a chase proves that h1 is (s − 1)-acyclic whenever gq is s-acyclic by
extending the diagram. The proposition finally follows by upper-semicontinuity from the relation:
dim(gq+r+1 ) + dim(hr+1 ) = m dim(Sq+r+1 T ∗ )
Q.E.D.
LEMMA 2.12: If gq is involutive and gq+1 is a vector bundle over Rq , then gq is also a vector
bundle over Rq . In this case, changing linearly the local coordinates if necessary, we may look at
k
the maximum number β of equations that can be solved with respect to vn...n
and the intrinsic
number α = m − β indicates the number of y that can be given arbitrarily.
Using the exactness of the top row in the preceding diagram and a delicate 3-dimensional chase,
we have (See [31] and [36],p336 for the details):
THEOREM 2.13: If Rq ⊂ Jq (E) is a system of order q on E such that gq+1 is a vector bundle
over Rq and gq is 2-acyclic, then there is an exact sequence:
q+r+1
πq+r
κ
r
Sr T ∗ ⊗ F1
Rq+r+1 −→ Rq+r −→
where κr is called the r-curvature and κ = κ0 is simply called the curvature of Rq .
We notice that Rq+r+1 = ρr (Rq+1 ) and Rq+r = ρr (Rq ) in the following commutative diagram:
Rq+r+1
q+r+1
↓ πq+r
(1)
Rq+r
∩
Rq+r
q+r+1
πq+1
−→
πqq+r
−→
πqq+r
−→
12
Rq+1
↓ πqq+1
(1)
Rq
∩
Rq
(1)
(1)
We also have Rq+r ⊆ ρr (Rq ) because we have successively:
(1)
q+r+1
Rq+r = πq+r
(Rq+r+1 ) =
⊆
=
=
q+r+1
πq+r
(Jr (Rq+1 ) ∩ Jq+r+1 (E))
Jr (πqq+1 )(Jr (Rq+1 )) ∩ Jq+r (E)
(1)
Jr (Rq ) ∩ Jq+r (E)
(1)
ρr (Rq )
while chasing in the following commutative 3-dimensional diagram:
Jr (Rq+1 )
−→
Jr (Jq+1 (E))
↓
ր
−→
Jq+r+1 (E)
↓
Jr (Rq )
−→
Jr (Jq (E))
ր
↓
ր
−→
Jq+r (E)
ր
Rq+r+1
↓
Rq+r
with a well defined map Jr (πqq+1 ) : Jr (Jq+1 (E)) → Jr (Jq (E)). We finally obtain the following
crucial Theorem and its Corollary (Compare to [31], p 72-74 or [36], p 340 to [16]):
THEOREM 2.14: Let Rq ⊂ Jq (E) be a system of order q on E such that Rq+1 is a fibered
submanifold of Jq+1 (E). If gq is 2-acyclic and gq+1 is a vector bundle over Rq , then we have
(1)
(1)
Rq+r = ρr (Rq ) for all r ≥ 0.
q+r+1
DEFINITION 2.15: A system Rq ⊂ Jq (E) is said to be formally integrable if πq+r
: Rq+r+1 →
Rq+r is an epimorphism of fibered manifolds for all r ≥ 1 and involutive if it is formally integrable
with an involutive symbol gq . We have the following useful test [16,31,52]:
COROLLARY 2.16: Let Rq ⊂ Jq (E) be a system of order q on E such that Rq+1 is a fibered
submanifold of Jq+1 (E). If gq is 2-acyclic (involutive) and if the map πqq+1 : Rq+1 → Rq is an
epimorphism of fibered manifolds, then Rq is formally integrable (involutive).
This is all what is needed in order to study systems of algebraic ordinary differential (OD) or
partial differential (PD) equations.
13
3) DIFFERENTIAL ALGEBRA
We now present in an independent manner two OD examples and two PD examples, among
the best ones we know, showing the difficulties met when studying differential ideals and ask the
reader to revisit them later on while reading the main Theorems. As only a few results will be
proved, the interested reader may look at [32,34,36] for more details and compare to [20,22,49].
EXAMPLE 3.1: If k = Q, y is a differential indeterminate and dx is a formal derivation, we may
set dx y = yx , dx yx = yxx and so on in order to introduce the differential ring A = k[y, yx , yxx , ...] =
k{y}. We consider the (proper) differential ideal a ⊂ A generated by the differential polynomial
P = yx2 − 4y. We have dx P = 2yx (yxx − 2) and a cannot be a prime differential ideal. Hence,
looking for the ”solutions ” of P = 0, we must have either yx = 0 ⇒ y = 0 or yxx = 0 and thus
y = (x+c)2 where c should be a ”constant ” with no clear meaning. However, we have successively:
P ∈a
⇒
⇒
⇒
⇒
⇒
⇒
⇒
⇒
⇒
⇒
⇒
⇒
⇒
yx (yxx − 2) ∈ a
yx yxxx + yxx (yxx − 2) ∈ a
(yx )2 yxxx ∈ a
yyxxx ∈ a
yyxxxx + yx yxxx ∈ a
yx (yxxx )2 ∈ a
2yx yxxx yxxxx + yxx (yxxx )2 ∈ a ⇒ 2yx yxxx yxxxx = −yxx(yxxx )2 mod(a)
4yxx yxxx yxxxx + 2yx (yxxxx )2 + 2yx yxxx yxxxxx + (yxxx )3 ∈ a
4yxx (yxxx )2 yxxxx + 2yx yxxx (yxxxx )2 + 2yx (yxxx )2 yxxxxx + (yxxx )4 ∈ a
4yxx (yxxx )2 yxxxx + 2yx yxxx (yxxxx )2 + (yxxx )4 ∈ a
3yxx (yxxx )2 yxxxx + (yxxx )4 ∈ a
−6yx yxxx(yxxxx )2 = (yxxx )4 ∈ a
(yxxx )5 ∈ a ⇒ yxxx ∈ rad(a)
and thus a is neither prime nor perfect, that is equal to its radical, but rad(a) is perfect as it
is the intersection of the prime differential ideal generated by y with the prime differential ideal
generated by yx2 − 4y and yxx − 2, both containing yxxx .
EXAMPLE 3.2: With the notations of the previous Example, let us consider the (proper) differential ideal a ⊂ A generated by the differential polynomial P = yx2 − 4y 3 . We have dx P =
2yx (yxx − 6y 2 ) and a cannot be prime differential ideal. Hence, looking for the ”solutions ” of
P = 0, we must have either yx = 0 ⇒ y = 0 or yx2 − 4y 3 = 0 and yxx − 6y 2 = 0. However, we have
successively:
P ∈a
⇒
⇒
⇒
⇒
⇒
yx (yxx − 6y 2 ) ∈ (a) ⇒ (yx )2 (yxx − 6y 2 )2 ∈ a ⇒ 4y 3 (yxx − 6y 2 )2 ∈ a
yxx (yxx − 6y 2 ) + yx (yxxx − 12yyx) ∈ a
yxx (yxx − 6y 2 )2 ∈ a
(yxx )2 (yxx − 6y 2 )2 − 12y 2 yxx (yxx − 6y 2 )2 + 36y 4 (yxx − 6y 2 )2 ∈ a
(yxx − 6y 2 )4 ∈ a ⇒ yxx − 6y 2 ∈ rad(a)
and thus a is neither prime or perfect as before but rad(a) is the prime differential ideal generated
by yx2 − 4y 3 and yxx − 6y 2 .
EXAMPLE 3.3: If k = Q as before, y is a differential indeterminate and (d1 , d2 ) are two formal
derivations, let us consider the differential ideal generated by P1 = y22 − 21 (y11 )2 and P2 = y12 − y11
in k{y}. Using crossed derivatives, we get successively:
P1 , P2 ∈ a ⇒
⇒
⇒
⇒
y112 − y111 ∈ a, y122 − y11 y111 ∈ a, y222 − y11 y111 ∈ a
Q = d2 P2 − d1 P1 + d1 P2 = (y11 − 1)y111 ∈ a
d1 Q = (y111 )2 + (y11 − 1)y1111 ∈ a
((y111 )3 ∈ a ⇒ y111 ∈ rad(a)
and thus a is neither prime nor perfect but rad(a) is a perfect differential ideal and even a prime
differential ideal p because we obtain easily from the last section that the resisual differential ring
k{y}/p ≃ k[y, y1 , y2 , y11 ] is a differential integral domain. Its quotient field is thus the differential
14
field K = Q(k{y}/p) ≃ k(y, y1 , y2 , y11 ) with the rules:
d1 y = y1 , d1 y1 = y11 , d1 y11 = 0, d2 y = y2 , d2 y1 = y11 , d2 y11 = 0
as a way to avoid looking for solutions. The formal linearization is the linear system R2 ⊂ J2 (E)
obtained in the last section where it was defined over R2 , but not over K, by the two linear second
order PDE:
Y22 − y11 Y11 = 0,
Y12 − Y11 = 0
changing slightly the notations for using the letter v only when looking at the symbols. It is at this
point that the problem starts because R2 is indeed a fibered manifold with arbitrary parametric
jets (y, y1 , y2 , y11 ) but R3 = ρ1 (R2 ) is no longer a fibered manifold because the dimension of its
symbol changes when y11 = 1. We understand therefore that there should be a close link existing
between formal integrability and the search for prime differential ideals or differential fields. The
solution of this problem has been provided as early as in 1983 for studying the ”Differential Galois
Theory ” but has never been acknowledged and is thus not known today ([32,34]). The idea is
to add the third order PDE y111 = 0 and thus the linearized PDE Y111 = 0 obtaining therefore
a third order involutive system well defined over K with symbol g3 = 0. We invite the reader to
treat similarly the two previous examples and to compare.
EXAMPLE 3.4: If k = Q as before, y is a differential indeterminate and (d1 , d2 ) are two
formal derivations, let us consider the differential ideal generated by P1 = y22 − 13 (y11 )3 and
P2 = y12 − 21 (y11 )2 in k{y}. Using crossed derivatives, we get successively:
P1 , P2 ∈ a ⇒ d2 P2 − d1 P1 + y11 d1 P2 = 0 ⇒ R2 involutive
⇒ y222 − (y11 )3 y111 = 0, y122 − (y11 )2 y111 = 0, y112 − y11 y111 = 0, ...
and thus dim(gq ) = 1, ∀q ≥ 1. As the symbol g2 is involutive, there is an infinite number of
parametric jets (y, y1 , y2 , y11 , y111 , ...) and thus k{y}/a ≃ k[y, y1 , y2 , y11 , y111 , ...] is a differential
integral domain with d2 y2 = y22 = 13 (y11 )3 , d2 y11 = y112 = y11 y111 , .... It follows that a = p is a
prime differential ideal with rad(p) = p. The second order linearized system is:
Y22 − (y11 )2 Y11 = 0,
Y12 − y11 Y11 = 0
is now well defined over the differential field K = Q(k{y}/p) and is involutive.
DEFINITION 3.5: A differential ring is a ring A with a finite number of commuting derivations
(∂1 , ..., ∂n ) such that ∂i (a + b) = ∂i a + ∂i b, ∂i (ab) = (∂i a)b + a∂i b that can be extended to derivations of the ring of quotients Q(A) by setting ∂i (a/s) = (s∂i a − a∂i s)/s2 , ∀0 6= s, a ∈ A. We shall
suppose from now on that A is even an integral domain and use the differential field K = Q(A).
For example, if x1 , ..., xn are indeterminates over Q, then Q[x] = Q[x1 , ..., xn ] is a differential ring
for the standard (∂1 , ..., ∂n ) with quotient field Q(x).
If K is a differential field as above and (y 1 , ..., y m ) are indeterminates over K, we transform the
polynomial ring K{y} = limq→∞ K[yq ] into a differential ring by introducing as usual the formal
k
∂/∂yµk and we shall set K < y >= Q(K{y}).
derivations di = ∂i + yµ+1
i
DEFINITION 3.6: We say that a ⊂ K{y} is a differential ideal if it is stable by the di , that is if
di a ∈ a, ∀a ∈ a, ∀i = 1, ..., n. We shall also introduce the radical rad(a) = {a ∈ A | ∃r, ar ∈ a} ⊇ a
and say that a is a perfect (or radical) differential ideal if rad(a) = a. If S is any subset of A, we
shall denote by {S} the differential ideal generated by S and introduce the (non-differential) ideal
ρr (S) = {dν a | a ∈ S, 0 ≤| ν |≤ r} in A.
15
LEMMA 3.7: If a ⊂ A is differential ideal, then rad(a) is a differential ideal containing a.
Proof: If d is one of the derivations, we have ar−1 da = 1r dar ∈ {ar } and thus:
(r − 1)ar−2 (da)2 + ar−1 d2 a ∈ {ar } ⇒ ar−2 (da)3 ∈ {ar }, ... ⇒ (da)2r−1 ∈ {ar }
Q.E.D.
LEMMA 3.8: If a ⊂ K{y}, we set aq = a ∩ K[yq ] with a0 = a ∩ K[y] and a∞ = a. We have in
general ρr (aq ) ⊆ aq+r and the problem will be to know when we may have equality.
We shall say that a differential extension L = Q(K{y}/p) is a finitely generated differential
extension of K and we may define the evaluation epimorphism K{y} → K{η} ⊂ L with kernel p
where η or ȳ is the residual image of y modulo p. If we study such a differential extension L/K,
by analogy with Section 2, we shall say that Rq or gq is a vector bundle over Rq if one can find
a certain number of maximum rank determinant Dα that cannot be all zero at a generic solution
of pq defined by differential polynomials Pτ , that is to say, according to the Hilbert Theorem of
Zeros, we may find polynomials Aα , Bτ ∈ K{yq } such that :
X
X
Aα Dα +
Bτ Pτ = 1
α
τ
In particular the following Lemma will be used in the next important Theorem:
LEMMA 3.9: If p is a prime differential ideal of K{y}, then, for q sufficiently large, there is a
polynomial D ∈ K[yq ] such that D ∈
/ pq and :
Dpq+r ⊂ rad(ρr (pq )) ⊂ pq+r ,
∀r ≥ 0
THEOREM 3.10: (Primality test) Let pq ⊂ K[yq ] and pq+1 ⊂ K[yq+1 ] be prime ideals such that
pq+1 = ρ1 (pq ) and pq+1 ∩ K[yq ] = pq . If the symbol gq of the algebraic variety Rq defined by pq is
2-acyclic and if its first prolongation gq+1 is a vector bundle over Rq , then p = ρ∞ (pq ) is a prime
differential ideal with p ∩ K[yq+r ] = ρr (pq ), ∀r ≥ 0.
COROLLARY 3.11: Every perfect differential ideal of {y} can be expressed in a unique way as
the non-redundant intersection of a finite number of prime differential ideals.
COROLLARY 3.12: (Differential basis) If r is a perfect differential ideal of K{y}, then we have
r = rad(ρ∞ (rq )) for q sufficiently large.
EXAMPLE 3.13: As K{y} is a polynomial ring with an infinite number of variables it is not
noetherian and an ideal may not have a finite basis. With K = Q, n = 1 and d = dx , then
a = {yyx, yx yxx , yxx yxxx , ...} ⇒ (yx )2 + yyxx ∈ a ⇒ rad(a) = {yx } is a prime differential ideal.
PROPOSITION 3.14: If ζ is differentially algebraic over K < η > and η is differentially algebraic over K, then ζ is differentially algebraic over K. Setting ξ = ζ − η, it follows that, if L/K
is a differential extension and ξ, η ∈ L are both differentially algebraic over K, then ξ + η, ξη and
di ξ are differentially algebraic over K.
If L = Q(K{y}/p), M = Q(K{z}/q) and N = Q(K{y, z}/r) are such that p = r ∩ K{y} and
q = r ∩ K{z}, we have the two towers K ⊂ L ⊂ N and K ⊂ M ⊂ N of differential extensions and
we may therefore define the new tower K ⊆ L ∩ M ⊆< L, M >⊆ N . However, if only L/K and
M/K are known and we look for such an N containing both L and M , we may use the universal
property of tensor products an deduce the existence of a differential morphism L⊗K M → N by
setting d(a ⊗ b) = (dL a) ⊗ b + a ⊗ (dM b) whenever dL | K = dM | K = ∂. The construction of
an abstract composite differential field amounts therefore to look for a prime differential ideal in
L⊗K M which is a direct sum of integral domains [32].
16
DEFINITION 3.15: A differential extension L of a differential field K is said to be differentially
algebraic over K if every element of L is differentially algebraic over K. The set of such elements
is an intermediate differential field K ′ ⊆ L, called the differential algebraic closure of K in L. If
L/K is a differential extension, one can always find a maximal subset S of elements of L that
are differentially transcendental over K and such that L is differentially algebraic over K < S >.
Such a set is called a differential transcedence basis and the number of elements of S is called the
differential transcendence degree of L/K.
THEOREM 3.16: The number of elements in a differential basis of L/K does not depent on the
generators of L/K and his value is dif f trd(L/K) = α. Moreover, if K ⊂ L ⊂ M are differential
fields, then dif f trd(M/K) = dif f trd(M/L) + dif f trd(L/K).
THEOREM 3.17: If L/K is a finitely generated differential extension, then any intermediate
differential field K ′ between K and L is also finitely generated over K.
EXAMPLE 3.18: With k = Q, let us introduce the manifolds X with local coordinate x and
Y with local coordinates (y 1 , y 2 ). We may consider the algebraic Lie pseudogroup Γ ⊂ aut(Y ) of
(local, invertible) transformations of Y preserving the 1-form y 2 dy 1 , that is to say made up by
transformations ȳ = g(y) solutions of the Pfaffian system ȳ 2 dȳ 1 = y 2 dy 1 . Equivalently, we have
to look for the invertible solutions of the algebraic first order involutive system R1 ⊂ J1 (Y × Y )
defined over k(y 1 , y 2 ) by the first order involutive system of algebraic PD equations in Lie form:
ȳ 2
∂ ȳ 1
= y2,
∂y 1
ȳ 2
∂ ȳ 1
=0
∂y 2
⇒
∂(ȳ 1 , ȳ 2 )
=1
∂((y 1 , y 2 )
By chance one can obtain the generic solution ȳ 1 = g(y 1 ), ȳ 2 = y 2 /(∂g(y 1 )/∂y 1 ) where g is an
arbitrary function of one variable. Now, if we introduce a function y = f (x) and consider the corresponding transformations of the jets (y 1 , y 2 , yx1 , yx2 , ...), we obtain the only generating differential
invariant Φ ≡ ȳ 2 ȳx1 = y 2 yx1 . Hence, setting K = k < y 2 yx1 > and L = k < y 1 , y 2 >, we have the
tower of differential extensions k ⊂ K ⊂ L. As any intermediate differential field K ⊂ K ′ ⊂ L is
finitely generated, let us consider K ′ = k < y 2 yx1 , yx2 >. Then:
ȳx2
∂ ȳ 2
∂ ȳ 2
∂ ȳ 2 1 ∂ ȳ 2
y
+
⇒
=
0,
= 1 ⇒ ȳ 1 = y 1 + cst,
x
∂y 1
∂y 2
∂y 1
∂y 2
ȳ 2 = y 2
allows to define a Lie subpseudogroup Γ′ ⊂ Γ with generating differential invariants yx1 , y 2 in such a
way that, if we set K” = k < yx1 , y 2 >, we have the strict inclusions K ⊂ K ′ ⊂ K” and it does not
seem possible to obtain a differential Galois correspondence between algebraic subpseudogroups
and intermediate differential fields, similar to the classical one. We have explained in [32] how to
overcome this problem but this is out of the scope of this paper. It is finally important to notice
that the fundamental differential isomorphism [4,5,32]:
Q(L⊗K L) ≃ Q(L⊗k(y) k[Γ])
is the Hopf dual of the projective limit of the action graph isomorphisms between fibered manifolds:
Aq ×X Aq ≃ Aq ×Y Rq
of fibered dimension 2(q + 2). The corresponding automorphic system y 2 yx1 = ω in Lie form where
ω is a geometric object as in the Introduction and its prolongations has been introduced as early
as in 1903 by E. Vessiot [53,54] as a way to study principal homogeneous spaces (PHS) for Lie
pseudogroups, namely if y = f (x) is a solution and ȳ = f¯(x) is another solution, then there exists
one and only one transformation ȳ = g(y) of Γ such that f¯ = g ◦ f .
This is all what is needed in order to study systems of infinitesimal Lie equations defined, like
the classical and conformal Killing systems, over Q < ω > where ω is a geometric object solution of
a system of algebraic Vessiot structure equations (constant riemannian curvature, zero Weyl tensor).
17
4) DIFFERENTIAL DUALITY
Let A be a unitary ring, that is 1, a, b ∈ A ⇒ a + b, ab ∈ A, 1a = a1 = a and even an integral
domain (ab = 0 ⇒ a = 0 or b = 0) with field of fractions K = Q(A). However, we shall not
always assume that A is commutative, that is ab may be different from ba in general for a, b ∈ A.
We say that M = A M is a left module over A if x, y ∈ M ⇒ ax, x + y ∈ M, ∀a ∈ A or a right
module MB over B if the operation of B on M is (x, b) → xb, ∀b ∈ B. If M is a left module over
A and a right module over B with (ax)b = a(xb), ∀a ∈ A, ∀b ∈ B, ∀x ∈ M , then we shall say that
M = A MB is a bimodule. Of course, A = A AA is a bimodule over itself. We define the torsion
submodule t(M ) = {x ∈ M | ∃0 6= a ∈ A, ax = 0} ⊆ M and M is a torsion module if t(M ) = M or
a torsion-free module if t(M ) = 0. We denote by homA (M, N ) the set of morphisms f : M → N
such that f (ax) = af (x). We finally recall that a sequence of modules and maps is exact if the
kernel of any map is equal to the image of the map preceding it.
When A is commutative, hom(M, N ) is again an A-module for the law (bf )(x) = f (bx) as we
have (bf )(ax) = f (bax) = f (abx) = af (bx) = a(bf )(x). In the non-commutative case, things are
more complicate and, given A M and A NB , then homA (M, N ) becomes a right module over B for
the law (f b)(x) = f (x)b.
DEFINITION 4.1: A module F is said to be free if it is isomorphic to a (finite) power of A
called the rank of F over A and denoted by rkA (F ) while the rank rkA (M ) of a module M is the
rank of a maximum free submodule F ⊂ M . It follows from this definition that M/F is a torsion
module. In the sequel we shall only consider finitely presented modules, namely finitely generated
p
d1
F0 −→ M −→ 0 where F0 and F1 are free
modules defined by exact sequences of the type F1 −→
modules of finite ranks m0 and m1 often denoted by m and p in examples. A module P is called
projective if there exists a free module F and another (projective) module Q such that P ⊕ Q ≃ F .
f
g
PROPOSITION 4.2: For any short exact sequence 0 → M ′ −→ M −→ M ” → 0, we have the
relation rkA (M ) = rkA (M ′ ) + rkA (M ”), even in the non-commutative case.
The following proposition will be used many times in Section 5, in particular for exhibiting the
Weyl tensor from the Riemann tensor ([2],p 73)([50],p 33) :
PROPOSITION 4.3: We shall say that the following short exact sequence splits if one of the
following equivalent three conditions holds:
u
′
v
←−
f
←−
g
0 −→ M −→ M −→ M ′′ −→ 0
• There exists a monomorphism v : M ′′ → M called lift of g and such that g ◦ v = idM ′′ .
• There exists an epimorphism u : M → M ′ called lift of f and such that u ◦ f = idM ′ .
• There exist isomorphisms ϕ = (u, g) : M → M ′ ⊕ M ′′ and ψ = f + v : M ′ ⊕ M ′′ → M that are
inverse to each other and provide an isomorphism M ≃ M ′ ⊕ M ′′ with f ◦ u + v ◦ g = idM and
thus ker(u) = im(v).
These conditions are automatically satisfied if M ” is free or projective.
Using the notation M ∗ = homA (M, A), for any morphism f : M → N , we shall denote by
f : N ∗ → M ∗ the morphism which is defined by f ∗ (h) = h ◦ f, ∀h ∈ homA (N, A) and satisfies
rkA (f ) = rkA (im(f )) = rkA (f ∗ ), ∀f ∈ homA (M, N )(See [37], Corollary 5.3, p 179). We may take
∗
d
d
1
2
F0 −→ 0 and apply homA (•, A) in
F1 −→
out M in order to obtain the deleted sequence ... −→
d∗
2
d∗
1
order to get the sequence ... ←− F1∗ ←− F0∗ ←− 0.
PROPOSITION 4.4: The extension modules ext0A (M ) = ker(d∗1 ) = homA (M, A) = M ∗ and
exti (M ) = extiA (M ) = ker(d∗i+1 )/im(d∗i ), ∀i ≥ 1 do not depend on the resolution chosen and are
torsion modules for i ≥ 1.
18
Let A be a differential ring, that is a commutative ring with n commuting derivations {∂1 , ..., ∂n },
that is ∂i ∂j = ∂j ∂i = ∂ij , ∀i, j = 1, ..., n while ∂i (a+b) = ∂i a+∂i b and ∂i (ab) = (∂i a)b+a∂i b, ∀a, b ∈
A. We shall use thereafter a differential integral domain A with unit 1 ∈ A whenever we shall
need a differential field Q ⊂ K = Q(A) of coefficients, that is a field (a ∈ K ⇒ 1/a ∈ K) with
∂i (1/a) = −(1/a2 )∂i a, in order to exhibit solved forms for systems of partial differential equations
as in the preceding section. Using an implicit summation on multi-indices, we may introduce the
(noncommutative) ring of differential operators D = A[d1 , ..., dn ] = A[d] with elements P = aµ dµ
such that | µ |< ∞ and di a = adi + ∂i a. The highest value of |µ| with aµ 6= 0 is called the order of
the operator P and the ring D with multiplication (P, Q) −→ P ◦ Q = P Q is filtred by the order
q of the operators with the filtration 0 = D−1 ⊂ D0 ⊂ D1 ⊂ ... ⊂ Dq ⊂ ... ⊂ D∞ = D. Moreover,
it is clear that D, as an algebra, is generated by A = D0 and T = D1 /D0 with D1 = A ⊕ T if
we identify an element ξ = ξ i di ∈ T with the vector field ξ = ξ i (x)∂i of differential geometry, but
with ξ i ∈ A now. It follows that D = D DD is a bimodule over itself, being at the same time a
left D-module D D by the composition P −→ QP and a right D-module DD by the composition
P −→ P Q with Dr Ds = Dr+s , ∀r, s ≥ 0 in any case.
k
to
If we introduce differential indeterminates y = (y 1 , ..., y m ), we may extend di yµk = yµ+1
i
d
i
k
+ ∂i aτk µ yµk for τ = 1, ..., p. Therefore, setting Dy 1 + ... + Dy m =
di Φτ ≡ aτk µ yµ+1
Φτ ≡ aτk µ yµk −→
i
m
Dy ≃ D and calling I = DΦ ⊂ Dy the differential module of equations, we obtain by residue the
p
differential module or D-module M = Dy/DΦ, introducing the canonical projection Dy −→ M → 0
k
k
and denoting the residue of yµ by ȳµ when there can be a confusion. Introducing the two
free differential modules F0 ≃ Dm0 , F1 ≃ Dm1 , we obtain equivalently the free presentation
p
d1
F0 −→ M → 0 of order q when d1 = D = Φ ◦ jq . It follows that M can be endowed
F1 −→
with a quotient filtration obtained from that of Dm which is defined by the order of the jet coordinates yq in Dq y. We shall suppose that the system Rq = ker(Φ) is formally integrable. We
have therefore the inductive limit 0 = M−1 ⊆ M0 ⊆ M1 ⊆ ... ⊆ Mq ⊆ ... ⊆ M∞ = M with
di Mq ⊆ Mq+1 which is the dual of the projective limit R = R∞ → ... → Rq → R0 → 0 if we set
R = homK (M, K) with Rq = homK (Mq , K) and DRq+1 ⊆ T ∗ ⊗ Rq . This is the main reason for
using a differential field K because homK (•, K) transform any short exact sequence into a short
exact sequence. We have in general Dr Is ⊆ Ir+s , ∀r ≥ 0, ∀s < q with Ir = I ∩ Dr y.
More generally, introducing the successive CC as in the preceding Section while changing slightly
the numbering of the respective operators, we may finally obtain the free resolution of M , namely
p
d1
d2
d3
F0 −→ M −→ 0 where p is the canonical projecF1 −→
F2 −→
the exact sequence
... −→
tion. Also, with a slight abuse of language, when D = Φ ◦ jq is involutive, that is to say when
Rq = ker(Φ) is involutive, one should say that M has an involutive presentation of order q or that
Mq is involutive.
REMARK 4.5: In actual practice, one must never forget that D = Φ ◦ jq acts on the left on
column vectors in the operator case and on the right on row vectors in the module case. For this
reason, when E is a (finite dimensional) vector bundle over X, we may apply the correspondence
J∞ (E) ↔ D⊗K E ∗ : Jq (E) ↔ Dq ⊗K E ∗ with πqq+1 : Jq+1 (E) → Jq (E) ↔ Dq ⊂ Dq+1 and
E ∗ = homK (E, K) between jet bundles and left differential modules in order to be able to use the
double dual isomorphism E ≃ E ∗∗ in both cases. We shall say that D(E) = D ⊗K E ∗ = ind(E ∗ )
is the the left differential module induced by E ∗ . Hence, starting from a differential operator
D∗
D
E −→ F , we may obtain a finite presentation D⊗K F ∗ −→ D⊗K E ∗ → M → 0 and conversely,
keeping the same operator matrix if we act on the right of row vectors. This comment becomes
particularly useful when dealing with the Poincaré sequence in electromagnetism (n = 4) or even
as we already saw in the Introduction (n = 3).
Roughly speaking, homological algebra has been created in order to find intrinsic properties
of modules not depending on their presentations or even on their resolutions and we now exhibit
another approach by defining the formal adjoint of an operator P and an operator matrix D:
ad
DEFINITION 4.6: Setting P = aµ dµ ∈ D ←→ ad(P ) = (−1)|µ| dµ aµ ∈ D, we have ad(ad(P )) =
19
P and ad(P Q) = ad(Q)ad(P ), ∀P, Q ∈ D. Such a definition can be extended to any matrix of
operators by using the transposed matrix of adjoint operators and we get:
< λ, Dξ >=< ad(D)λ, ξ > + div (...)
from integration by part, where λ is a row vector of test functions and <> the usual contraction.
We quote the useful formulas rkD (D) = rkD (ad(D)) as in ([34], p 339-341 or [35]).
The following technical Lemma is crucially used in the next proposition:
LEMMA 4.7: If f ∈ aut(X) is a local diffeomorphisms on X, we may set x = f −1 (y) = g(y) and
we have the identity:
∂
1
(
∂i f k (g(y)) ≡ 0.
∂y k ∆(g(y))
D
PROPOSITION 4.8: If we have an operator E −→ F , we may obtain by duality an operator
∧n T ∗ ⊗ E
∗ ad(D)
←− ∧n T ∗ ⊗ F ∗ .
Now, with operational notations, let us consider the two differential sequences:
D
D
1
ζ
ξ −→ η −→
ad(D)
ad(D1 )
ν ←− µ ←− λ
where D1 generates all the CC of D. Then D1 ◦ D ≡ 0 ⇐⇒ ad(D) ◦ ad(D1 ) ≡ 0 but ad(D) may
not generate all the CC of ad(D1 ) as we already saw in the Introduction. Passing to the module
framework, we just recognize the definition of ext1 (M ) when M is determined by D.
As D = D DD is a bimodule, then M ∗ = homD (M, D) is a right D-module according to
Lemma 3.1 and we may thus define a right module ND by the ker/coker long exact sequence
D∗
0 ←− ND ←− F1∗ ←− F0∗ ←− M ∗ ←− 0 but we have [6,36,43,51]:
THEOREM 4.9: We have the side changing procedures M = D M → MD = ∧n T ∗ ⊗A M and
ND → N = D N = homA (∧n T ∗ , ND ) with D ((MD )) = M and D (ND ) = N .
Now, exactly like we defined the differential module M from D, we may define the differential
module N from ad(D). For any other presentation of M with an accent, we have [23,36]:
THEOREM 4.10: The modules N and N ′ are projectively equivalent, that is one can find two
projective modules P and P ′ such that N ⊕ P ≃ N ′ ⊕ P ′ and we obtain therefore extiD (N ) ≃
extiD (N ′ ), ∀i ≥ 1.
THEOREM 4.11: The operator D is simply parametrizable if ext1 (N ) = 0 and doubly parametrizable if ext1 (N ) = 0 and ext2 (N ) = 0. Moreover, we have the ker/coker long exact sequence:
ǫ
0 → ext1 (N ) → M −→ M ∗∗ → ext2 (N ) → 0
where (ǫ(m))(f ) = f (m) whenever f ∈ M ∗ and we have t(M ) = ext1 (N ) = ker(ǫ).
Proof: We prove first that t(M ) ⊆ ker(ǫ). Indeed, if m ∈ t(M ), then one may find 0 6= P ∈ D
such that P m = 0 and thus f (P m) = P f (m) = 0 ⇒ f (m) = 0 because D = K[d] is an integral
domain and thus t(M ) ⊆ ker(ǫ).
Let us now start with a free presentation of M = cocker(d1 ):
p
d
1
F0 −→ M −→ 0
F1 −→
Applying homD (M, D), we may define ND = coker(d∗1 ) and exhibit the following free resolution
of N by right D-modules:
d∗
d∗
d∗
−1
1
0
∗
∗
←− F−2
0 ←− ND ←− F1∗ ←−
F0∗ ←−
F−1
20
where M ∗ = ker(d∗1 ) = im(d∗0 ) ≃ coker(d∗−1 ). The deleted sequence is:
d∗
d∗
−1
d∗
1
0
∗
∗
0 ←− F1∗ ←−
F0∗ ←−
F−1
←− F−2
Applying again homD (•, D) and using the canonical isomorphism F ∗∗ ≃ F for any free module F
of finite rank, we get the sequence of left D-modules:
d
1
F0
0 −→ F1 −→
↓
M
↓
0
d
0
−→
F−1
ց
↑
ǫ
−→ M ∗∗
↑
0
d−1
−→ F−2
Denoting as usual a coboundary space by B, a cocycle space by Z and the corresponding cohomology by H = Z/B, we get the commutative and exact diagram:
0 −→ B0
↓
0 −→ Z0
−→ F0
k
−→ F0
−→
−→
M
↓ǫ
M ∗∗
−→ 0
An easy chase provides at once H0 = Z0 /B0 = ext1D (N ) ≃ ker(ǫ). It follows that ker(ǫ) is a torsion module and, as we already know that t(M ) ⊆ ker(ǫ) ⊆ M , we finally obtain t(M ) = ker(ǫ).
Also, as B−1 = im(ǫ) and Z−1 ≃ M ∗∗ , we obtain H−1 = Z−1 /B−1 = ext2A (N, A) ≃ coker(ǫ).
Accordingly, a torsion-free (ǫ injective)/reflexive (ǫ bijective) module is described by an operator
that admits respectively a single/double step parametrization.
Q.E.D.
We know turn to the operator framework;
D
DEFINITION 4.12: If a differential operator ξ −→ η is given, a direct problem is to find generatD1
ζ such that Dξ = η ⇒ D1 η = 0. Conversely,
ing compatibility conditions (CC) as an operator η −→
D
D
1
ζ, the inverse problem will be to look for ξ −→ η such that D1 generates the CC of D
given η −→
and we shall say that D1 is parametrized by D if such an operator D is existing. We finally notice
that any operator is the adjoint of a certain operator because ad(ad(P )) = P, ∀P ∈ D and we get:
THEOREM 4.13: (reflexivity test) In order to check whether M is reflexive or not, that is to
find out a parametrization if t(M ) = 0 which can be again parametrized, the test has 5 steps which
are drawn in the following diagram where ad(D) generates the CC of ad(D1 ) and D1′ generates the
CC of D = ad(ad(D)) while ad(D−1 ) generates the CC of ad(D) and D′ generates the CC of D−1 :
η′
D
4
φ
3
θ
−→
ad(D−1 )
←−
D
ξ
−→
ad(D)
ν
←−
5
ζ
1
λ
2
ր
ր
D−1
ζ′
D1′
′
η
µ
D1
−→
ad(D1 )
←−
D1 parametrized by D ⇔ D1 = D1′ ⇔ ext1 (N ) = 0 ⇔ ǫ injective ⇔ t(M ) = 0
D parametrized by D−1 ⇔ D = D′ ⇔ ext2 (N ) = 0 ⇔ ǫ surjective
D
p
1
F0 −→ M → 0 is a fiCOROLLARY 4.14: In the differential module framework, if F1 −→
nite free presentation of M = coker(D1 ) with t(M ) = 0, then we may obtain an exact sequence
D
D
1
F0 −→ E of free differential modules where D is the parametrizing operator. However,
F1 −→
D′
D
1
F0 −→ E ′ called minimal parametrizations such
there may exist other parametrizations F1 −→
′
that coker(D ) is a torsion module and we have thus rkD (M ) = rkD (E ′ ).
21
REMARK 4.15: The following chains of inclusions and short exact sequences allow to compare
the main procedures used in the respective study of differential extensions and differential modules:
K ⊂ K < S >⊂ L ⇒
K ⊂ K′ ⊂ L
0→F →M →T →0
⇒ 0 → t(M ) → M → M ′ → 0
where F is a maximum free submodule of M , T = M/F is a torsion-module and M ′ = M/t(M ) is
a torsion-free module. The next examples open the way towards a new domain of research.
EXAMPLE 4.16: With n = 2, m = 3, K = Q, let us consider the first order nonlinear involutive
system:
P2 ≡ y22 − y 3 y12 = 0
P1 ≡ y21 − y 3 y11 = 0,
This system defines a prime differential ideal p ⊂ K{y} and the differential extension L =
1
2
Q(K{y}/p) is differentially algebraic over K < y 3 > with parametric jets (y 1 , y 2 , y11 , y12 , y11
, y11
, ...).
The linearized system D1 Y = 0 over L is:
d2 Y 1 − y 3 d1 Y 1 − y11 Y 3 = 0,
d2 Y 2 − y 3 d1 Y 2 − y12 Y 3 = 0
Multiplying by test functions (λ1 , λ2 ) and integrating by part, we get ad(D1 )λ = µ in the form:
Y1
Y2
Y3
→ −d2 λ1 + y 3 d1 λ1 + y13 λ1
→ −d2 λ2 + y 3 d1 λ2 + y13 λ2
→
−y11 λ1 − y12 λ2
=
=
=
µ1
µ2
µ3
Using only the parametric jets for y and λ in the PD equations provided, we get:
1
2
−y11 (y 3 d1 λ1 + y13 λ1 ) − (y 3 y11
+ y11 y13 )λ1 − y12 (y 3 d1 λ2 + y13 λ1 ) − (y 3 y11
+ y12 y13 )λ2 = d2 µ3 − y11 µ1 − y12 µ2
1 1
2 2
−y 3 y11 d1 λ1 − y 3 y11
λ − y 3 y12 d1 λ2 − y 3 y11
λ − 2y11 y13 λ1 − 2y12 y13 λ2 = y 3 d1 µ3 + 2y13 µ3
and the only CC ad(D)µ = 0 over L:
−d2 µ3 + y 3 d1 µ3 + y11 µ1 + y12 µ2 + 2y13 µ3 = 0
Multiplying by a test function ξ and integrating by part, we get Dξ = Y over L in the form:
y11 ξ = Y 1 ,
y12 ξ = Y 2 ,
d2 ξ − y 3 d1 ξ + y13 ξ = Y 3
admitting the CC D1 Y = 0 of course but also the additional zero order CC:
ω ≡ y11 Y 2 − y12 Y 1 = 0
which provides a torsion element ω satisfying d2 ω − y 3 d1 ω − y13 ω = 0. Setting Y = δy as the
standard variational notation used by engineer, we obtain easily ω ∧ δω 6= 0 and ω cannot therefore
admit an integrating factor, a result showing that K is its own differential algebraic closure in L.
EXAMPLE 4.17: If α = dx1 − x3 dx2 ∈ T ∗ , the linear system obtained over K = Q(x1 , x2 , x3 ) by
eliminating the factor ρ(x) in the linear system L(ξ)α = ρ(x)α admits the injective parametrization −x3 ∂3 θ + θ = ξ 1 , −∂3 φ = ξ 2 , ∂2 φ − x3 ∂1 φ = ξ 3 ⇒ ξ 1 − x3 ξ 2 = φ. It defines therefore a
free differential module M ≃ D which is thus reflexive and even projective. Any resolution of
this module splits, like the short exact sequence 0 → D2 → D3 → D → 0, and we shall prove in
section 6 that the corresponding differential sequence of operators is locally exact like the Poincaré
sequence ([32], p 684-691).
22
5) CONFORMAL STRUCTURE
We start this section with a general (difficult) result on the actions of Lie groups, covering at
the same time the study of the classical and conformal Killing systems. For this, we notice that the
j1
involutive first Spencer operator D1 : C0 = Rq → J1 (Rq ) → J1 (Rq )/Rq+1 ≃ T ∗ ⊗ Rq /δ(gq+1 ) = C1
of order one is induced by the Spencer operator D : Rq+1 → T ∗ ⊗ Rq : ξq+1 → j1 (ξq ) − ξq+1 =
k
| 0 ≤| µ | q}. Introducing the Spencer bundles Cr = ∧r T ∗ ⊗ Rq /δ(∧r−1 T ∗ ⊗ gq+1 ), the
{∂i ξµk − ξµ+1
i
first order involutive (r + 1)-Spencer operator Dr+1 : Cr → Cr+1 is induced by D : ∧r T ∗ ⊗ Rq+1 →
∧r+1 T ∗ ⊗ Rq : α ⊗ ξq+1 → dα ⊗ ξq + (−1)r α ∧ Dξq+1 . We obtain therefore the canonical linear
Spencer sequence ([34], p 150 or [52]):
jq
D
D
D
D
n
3
2
1
Cn −→ 0
... −→
C2 −→
0 −→ Θ −→ C0 −→
C1 −→
PROPOSITION 5.1: The Spencer sequence for the Lie operator describing the infinitesimal
action of a Lie group G is (locally) isomorphic to the tensor product of the Poincaré sequence by
the Lie algebra G = Te (G) where e ∈ G is the identity element. It follows that Dr+1 generates the
CC of Dr ⇔ ad(Dr ) generates the CC of ad(Dr+1 ), a result not evident at all.
Proof: We may introduce a basis {θτ = θτi (x)∂i } of infinitesimal generators of the action with
τ = 1, ..., dim(G) and the commutation relations [θρ , θσ ] = cτρσ θτ discovered by S. Lie giving the
structure constants c of G (See [34] and [44] for more details). Any element λ ∈ G can be written
λ = {λτ = cst}. ” Gauging ” such an element, that is to say replacing the constants by functions
or, equivalently, introducing a map X → ∧0 T ∗ ⊗ G : (x) → (λτ (x)), we may obtain locally a
map ∧0 T ∗ ⊗ G → T : λτ (x) → λτ (x)θτk (x) or, equivalently, vector fields ξ = (ξ i (x)∂i ) ∈ T of the
form ξ k (x) = λτ (x)θτk (x), keeping the index i for 1-forms. More generally, we can introduce a map :
k
∧r T ∗ ⊗ G → ∧r T ∗ ⊗ Jq (T ) = λ → λ ⊗ jq (θ) = Xq : λτ (x) → λτ (x)∂µ θτk (x) = Xµ,I
(x)dxI
that we can lift to the element λ ⊗ jq+1 (θ) = Xq+1 ∈ ∧r T ∗ ⊗ Jq+1 (T ). It follows from the definitions that Dr Xq = DXq+1 by introducing any element of Cr (T ) through its representative
Xq ∈ ∧r T ∗ ⊗ Jq (T ). We obtain therefore the crucial formula:
Dr Xq = DXq+1 = D(λ ⊗ jq+1 (θ)) = dλ ⊗ jq (θ) + (−1)r λ ∧ Djq+1 (θ) = dλ ⊗ jq (θ)
allowing to identify locally the Spencer sequence for jq with the Poincaré sequence. When the
action is effective, the map ∧0 T ∗ ⊗ G → Jq (T ) is injective. We obtain therefore an isomorphism
∧0 T ∗ ⊗ G → Rq ⊂ Jq (T ) when q is large enough allowing to exhibit an isomorphism between the
canonical Spencer sequence and the tensor product of the Poincaré sequence by G when q is large
enough in such a way that Rq is involutive with dim(Rq ) = dim(G) and gq = 0.
Q.E.D.
We now study what happens when n ≥ 3 because the case n = 2 has already been provided,
proving that conformal geometry must be entirely revisited.
• n = 3: Using the euclidean metric ω, we have 6 components of Ω ∈ F0 = S2 T ∗ with dim(F0 ) =
n(n + 1)/2 = 6 in the case of the classical Killing system/operator and obtain easily the n2 (n2 −
1)/12 = 6 components of the second order Riemann operator, linearization of the Riemann tensor
at ω. We have n2 (n2 − 1)(n − 2)/24 = 3 first order Bianchi identities ([32], p 625). Introducing
the respective adjoint operators while taking into account the last Proposition and the fact that
the extension modules do not depend on the resolution used (a difficult result indeed !), we get the
following diagram where we have set ad(Riemann) = Beltrami for historical reasons [3] and each
operator generates the CC of the next one:
3
0←
3
Killing
−→
Cauchy
←−
6
Riemann
−→
6
Bianchi
−→
3→0
6
Beltrami
6
←−
3
←−
As in the Introduction where Airy = ad(Riemann), the Beltrami operator is now parametrizing
the 3 Cauchy stress equations [3] but it is rather striking to discover that the central second order
23
operator is self-adjoint and can be given
0
0
0
−2d
33
0
2d23
d33
0
−2d23 2d13
d22
−2d12
as follows:
0
2d23
−2d22
−2d13
2d12
0
d33
0
−2d13
0
0
d11
−2d23
2d13
2d12
0
−2d11
0
d22
−2d12
0
d11
0
0
The study of the conformal case is much more delicate. As F̂0 can be described by trace-free
symmetric tensors, we have dim(F̂0 ) = dim(F0 ) − 1 = 5 and it remains to discover the operator
that will replace the Riemann operator. Having in mind the diagram of Proposition 2.11 and the
fact that dim(ĝ2 ) = 3 while ĝ3 = 0 ⇒ ĝ4 = 0, we have successively:
• NO CC order 1: 0 → ĝ2 → S2 T ∗ ⊗ T → T ∗ ⊗ F̂0 → F̂1 ⇒ 0 ⇒ dim(F̂1 ) = 3 − 18 + 15 = 0.
• NO CC order 2: 0 → ĝ3 → S3 T ∗ ⊗ T → S2 T ∗ ⊗ F̂0 → F̂1 ⇒ 0 ⇒ dim(F̂1 ) = 0 − 30 + 30 = 0.
• OK CC order 3: 0 → ĝ4 → S4 T ∗ ⊗ T → S3 T ∗ ⊗ F̂0 → F̂1 ⇒ 0 ⇒ dim(F̂1 ) = 0 − 45 + 50 = 5.
Once again, the central third order operator is self-adjoint as can be easily seen by proving that
the last 5 → 3 operator, obtained in [44] by means of computer algebra, can be chosen to be the
transpose of the first 3 → 5 conformal Killing operator, just by changing columns.
This result can also be obtained by using the fact that, when an operator/a system is formally
integrable, the order of the generating CC is equal to the number of prolongations needed to get a
2-acyclic symbol plus 1 ([34], p 120, [44]). In the present case, neither ĝ1 nor ĝ2 are 2-acyclic while
ĝ3 = 0 is trivially involutive, so that (3 − 1) + 1 = 3.
• n = 4: In the classical case, we may proceed as before for exibiting the 20 components of the
second order Riemann operator and the 20 components of the first order Bianchi operator.
The study of the conformal case is much more delicate and still unknown. Indeed, the symbol ĝ2
is 2-acyclic when n ≥ 4 and 3-acyclic when n ≥ 5. Accordingly, the Weyl operator, namely the
CC for the conformal Killing operator, is second order like the Riemann operator. However, when
n = 4 only (care), the symbol ĥ2 of the Weyl system is not 2-acyclic while its first prolongation ĥ3
becomes 2-acyclic. It follows that the CC for the Weyl operator are second order, ... and so on.
For example, we have the long exact sequence:
0 → ĝ5 → S5 T ∗ ⊗ T → S4 T ∗ ⊗ F̂0 → S2 T ∗ ⊗ F̂1 → F̂2 → 0
and deduce that dim(F̂2 ) = (−0) + (56 × 4) − (35 × 9) + (10 × 10) = 9, a result that can be ckecked
by computer algebra in a few milliseconds but is still unknown.
We shall finally prove below that the Einstein parametrization of the stress equations is neither
canonical nor minimal in the following diagrams:
4
0← 4
Killing
−→
Cauchy
←−
10
k
Riemann
−→
20
↓
Bianchi
10
Einstein
−→
10
−→
div
4
10
k
10
Beltrami
20
↑
10
←−
20
←−
Einstein
←−
−→
20 −→ 6
↓
→
→0
0
obtained by using the fact that the Einstein operator is self-adjoint, where by Einstein operator
we mean the linearization of the Einstein equations at the Minkowski metric, the 6 terms being
exchanged between themselves [35,40]. Indeed, setting Eij = Rij − 12 ωij tr(R) with tr(R) = ω ij Rij ,
it is essential to notice that the Ricci operator is not self-adjoint because we have for example:
ad
λij (ω rs dij Ωrs ) −→ (ω rs dij λij )Ωrs
and ad provides a term appearing in −ωij tr(R) but not in 2Rij because we have, as in (5.1.4) of [15]:
tr(Ω) = ω rs Ωrs
⇒
tr(R) = ω rs drs tr(Ω) − drs Ωrs
24
The upper div induced by Bianchi has nothing to do with the lower Cauchy stress equations,
contrary to what is still believed today while the 10 on the right of the lower diagram has nothing
to do with the perturbation of a metric which is the 10 on the left in the upper diagram. It also
follows that the Einstein equations in vacuum cannot be parametrized as we have the following
diagram of operators recapitulating the five steps of the parametrizability criterion (See [35,37] for
more details or [44,56] for a computer algebra exhibition of this result):
Riemann
20
ր
4
4
Killing
−→
Cauchy
←−
10
10
Einstein
−→
10
Einstein
10
←−
As a byproduct, we are facing only two possibilities, both leading to a contradiction:
Einstein
1) If we use the operator S2 T ∗ −→ S2 T ∗ in the geometrical setting, the S2 T ∗ on the left has
indeed someting to do with the perturbation of the metric but the S2 T ∗ on the right has nothing
to do with the stress.
Einstein
2) If we use the adjoint operator ∧n T ∗ ⊗ S2 T ←− ∧n T ∗ ⊗ S2 T in the physical setting, then
n ∗
∧ T ⊗ S2 T on the left has of course something to do with the stress but the ∧n T ∗ ⊗ S2 T on the
right has nothing to do with the perturbation of a metric.
These purely mathematical results question the origin and existence of gravitational waves.
We may summarize these results, which do not seem to be known, by the following differential
sequences where the order of an operator is written under its arrow:
• n = 3:
• n = 4:
• n = 5:
3 −→ 5 −→ 5 −→ 3 → 0
1
3
1
4 −→ 9 −→ 10 −→ 9 −→ 4 → 0
1
2
2
1
5 −→ 14 −→ 35 −→ 35 −→ 14 −→ 5 → 0
1
2
1
2
1
2
THEOREM 5.2: Recalling that we have F1 = H (g1 ) = Z 2 (g1 ) and F̂1 = H 2 (ĝ1 ) 6= Z 2 (ĝ1 ), we
have the following commutative and exact ”fundamental diagram II ”:
0
0 −→ S2 T ∗
−→
δ
0
↓
0
↓
−→ Z 2 (g1 )
↓
T ∗ ⊗ ĝ2
↓
−→ Z 2 (ĝ1 )
↓
−→ T ∗ ⊗ T ∗
↓
0
δ
δ
−→
∧2 T ∗
↓
0
0
↓
S2 T ∗
↓
−→ H 2 (g1 ) −→ 0
↓
−→ H 2 (ĝ1 ) −→ 0
↓
−→
0
The following theorem will provide all the classical formulas of both Riemannian and conformal
geometry in one piece but in a totally unusual framework not depending on any conformal factor:
THEOREM 5.3: All the short exact sequences of the preceding diagram split in a canonical way,
that is in a way compatible with the underlying tensorial properties of the vector bundles involved.
T ∗ ⊗ T ∗ ≃ S2 T ∗ ⊕ ∧ 2 T ∗
⇒ Z 2 (ĝ1 ) = Z 2 (g1 ) + δ(T ∗ ⊗ ĝ2 ) ≃ Z 2 (g1 ) ⊕ ∧2 T ∗
⇒ H 2 (g1 ) ≃ H 2 (ĝ1 ) ⊕ S2 T ∗
25
Proof: First of all, we recall that:
g1 = {ξik ∈ T ∗ ⊗ T | ωrj ξir + ωir ξjr = 0} ⊂ ĝ1 = {ξik ∈ T ∗ ⊗ T | ωrj ξir + ωir ξjr −
2
ωij ξrr = 0}
n
k
k
r
r
r
0 = g2 ⊂ ĝ2 = {ξij
∈ S2 T ∗ ⊗ T | nξij
= δik ξrj
+ δjk ξri
− ωij ω ks ξrs
}
⇒
k
Now, if (τli,j
) ∈ T ∗ ⊗ ĝ2 , then we have:
k
r
r
r
nτli,j
= δlk τri,j
+ δik τrl,j
− ωli ω ks τrs,j
r
and we may set τri,j
= τi,j 6= τj,i with (τi,j ) ∈ T ∗ ⊗ T and such a formula does not depend on any
conformal factor [53]. We have:
k
k
k
δ(τli,j
) = (τli,j
− τlj,i
) = (ρkl,ij ) ∈ B 2 (ĝ1 ) ⊂ Z 2 (ĝ1 )
with:
Z 2 (ĝ1 ) = {(ρkl,ij ) ∈ ∧2 T ∗ ⊗ ĝ1 ) | δ(ρkl,ij ) = 0} ⇒ ϕij = ρrr,ij 6= 0
X
δ(ρkl,ji ) = (
ρkl,ij = ρkl,ij + ρki,jl + ρkj,li ) ∈ ∧3 T ∗ ⊗ T
(l,i,j)
• The splitting of the lower row is obtained by setting (τi,j ) ∈ T ∗ ⊗ T ∗ → ( 21 (τi,j + τj,i )) ∈ S2 T ∗
in such a way that (τi,j = τj,i = τij ) ∈ S2 T ∗ ⇒ 12 (τij + τji ) = τij .
Similarly, (ϕij = −ϕji ) ∈ ∧2 T ∗ → ( 12 ϕij ) ∈ T ∗ ⊗ T ∗ and ( 12 ϕij − 21 ϕji ) = (ϕij ) ∈ ∧2 T ∗ .
• The most important result is to split the right column. For this, we first need to describe the
monomorphism 0 → S2 T ∗ → H 2 (g1 ) which is in fact produced by a diagonal north-east snake type
k
chase. Let us choose (τij = τi,j = τj,i = τji ) ∈ S2 T ∗ ⊂ T ∗ ⊗ T ∗. Then, we may find (τli,j
) ∈ T ∗ ⊗ ĝ2
r
r
2
k
k
k
by deciding that τri,j = τi,j = τj,i = τrj,i in Z (ĝ1 ) and apply δ in order to get ρl,ij = τli,j
− τk,lj,i
such that ρrr,ij = ϕij = 0 and thus (ρkl,ij ) ∈ Z 2 (g1 ) = H 2 (g1 ). We obtain:
nρkl,ij
r
r
r
r
r
r
= δlk τri,j
− δlk τrj,i
+ δik τrl,j
− δjk τrli
− ω ks (ωli τrs,j
− ωlj τrs,i
)
k
k
ks
= (δi τlj − δj τli ) − ω (ωli τsj − ωlj τsi )
Contracting in k and i while setting simply tr(τ ) = ω ij τij , tr(ρ) = ω ij ρij , we get:
nρij = nτij − τij − τij + ωij tr(τ ) = (n − 2)τij + ωij tr(τ ) = nρji ⇒ ntr(ρ) = 2(n − 1)tr(τ )
Substituting, we finally obtain τij =
ρkl,ij =
n
(n−2) ρij
−
n
2(n−1)(n−2) ωij tr(ρ)
and thus the tricky formula:
1
1
(δik ρlj − δjk ρli ) − ω ks (ωli ρsj − ωlj ρsi )) −
(δ k ωlj − δjk ωli )tr(ρ)
(n − 2)
(n − 1)(n − 2) i
Contracting in k and i, we check that ρij = ρij indeed, obtaining therefore the desired canonical
lift H 2 (g1 ) → S2 T ∗ → 0 : ρki,lj → ρri,rj = ρij . Finally, using Proposition 4.3, the epimorphism
H 2 (g1 ) → H 2 (ĝ1 ) → 0 is just described by the formula:
k
σl,ij
= ρkl,ij −
1
1
(δ k ρlj − δjk ρli − ω ks (ωli ρsj − ωlj ρsi )) +
(δ k ωlj − δjk ωli )tr(ρ)
(n − 2) i
(n − 1)(n − 2) i
r
r
which is just the way to define the Weyl tensor. We notice that σr,ij
= ρrr,ij = 0 and σi,rj
= 0 by
2
2
∗
using indices or a circular chase showing that Z (ĝ1 ) = Z (g1 ) + δ(T ⊗ ĝ2 ). This purely algebraic
result only depends on the metric ω and does not depend on any conformal factor. In actual
practice, the lift H 2 (g1 ) → S2 T ∗ is described by ρkl,ij → ρri,rj = ρij = ρji but it is not evident at
k
k
all that the lift H 2 (ĝ1 ) → H 2 (g1 ) is described by the strict inclusion σl,ij
→ ρkl,ij = σl,ij
providing
r
r
a short exact sequence as in Proposition 4.3 because ρij = ρi,rj = σi,rj = 0 by composition.
Q.E.D.
COROLLARY 5.4: When n ≥ 4, each component of the Weyl tensor is a torsion element killed
by the Dalembertian whenever the Einstein equations in vacuum are satisfied by the metric. Hence,
26
there exists a second order operator Q such that we have an identity:
✷ ◦ W eyl = Q ◦ Ricci
Proof: According to Proposition 4.4, each extension module exti (M ) is a torsion module, ∀i ≥ 1.
It follows that each additional CC in D1′ which is not already in D1 is a torsion element as it
belongs to this module. One may also notice that:
n(n + 1)
n(n − 1)
n(n + 1)
n(n − 1)
−n=
, rkD (Riemann) =
−n=
2
2
2
2
The differential ranks of the Einstein and Riemann operators are thus equal, but this is a pure
coincidence because rkD (Einstein) has only to do with the div operator induced by contracting
the Bianchi identities, while rkD (Riemann) has only to do with the classical Killing operator and
the fact that the corresponding differential module is a torsion module because we have a Lie group
of transformations having n + n(n−1)
= n(n+1)
parameters (translations + rotations). Hence, as
2
2
the Riemann operator is a direct sum of the Weyl operator and the Einstein or Ricci operator
according to the previous theorem, each component of the Weyl operator must be killed by a
certain operator whenever the Einstein or Ricci equations in vacuum are satisfied. A direct tricky
computation can be found in ([8], p 206]) and ([18], exercise 7.7]).
Q.E.D.
rkD (Einstein) =
REMARK 5.5: In a similar manner, the EM wave equations ✷F = 0 are easily obtained when
the second set of Maxwell equations in vacuum is satisfied, avoiding therefore the Lorenz (no ”t”)
gauge condition for the EM potential. Indeed, let us start with the Minkowski constitutive law
with electric constant ǫ0 and magnetic constant µ0 such that ǫ0 µ0 c2 = 1 in vacuum:
F rs =
1 ri sj
ω̂ ω̂ Fij ∼ ω ri ω sj Fij
µ0
−1/n
where ω̂ij = | det(ω) |
ωij ⇒| det(ω̂) |= 1, F ∈ ∧2 T ∗ is the EM field and the induction F is
thus a contravariant skewsymmetric 2-tensor density. From the Maxwell equations we have:
∂r Fij + ∂i Fjr + ∂j Fri = 0,
∇r Fri = 0
⇒
∇r Fri = 0
✷Fij = ∇r ∇r Fij = ∇r (∇i Frj − ∇j Fri ) = 0
⇒
REMARK 5.6: Using Proposition 4.3 and the splittings of Theorem 5.3 for the second column,
we obtain the following commutative and exact diagram:
0
↓
10
↓↑
0
↓
−→
10
k
Riemann
−→
20
↓↑
Bianchi
10
↓
0
Einstein
10
↓
0
−→
−→
−→
div
0
0
↓
↓
16 → 6
↓
k
20 → 6
↓
↓
4
↓
0
→0
→0
→ 0
It follows that the 10 components of the Weyl tensor must satisfy a first order linear system with 16
equations, having 6 generating first order CC. The differential rank of the corresponding operator
is thus equal to 16 − 6 = 10 and such an operator defines a torsion module in which we have to look
separately for each component of the Weyl tensor in order to obtain Corollary 5.4. The situation is
similar to that of the Cauchy-Riemann equations when n = 2. Indeed, any complex transformation
y = f (x) must be solution of the (linear) first order system y22 − y11 = 0, y21 + y12 = 0 of finite Lie
2
2
1
1
= 0, that is y 1 and y 2 are separately killed by
+ y22
= 0, y11
+ y22
equations though we obtain y11
the second order laplacian ∆ = d11 + d22 .
27
6) CONTACT STRUCTURE
Changing slightly the notations whileP
setting α = 1, ..., p and ᾱ = α + p = p + 1, ..., 2p, we may
consider the contact 1-form χ = dxn − pα=1 xᾱ dxα ⇒ χ ∧ (dχ)p = (1 )p+1 p!dx1 ∧ ... ∧ dxn 6= 0
where the exterior power of dχ is taken p times. As before, we obtain the injective parametrization:
∂φ
∂φ
∂φ
∂φ
∂φ
, ξ ᾱ =
+ xᾱ n , ξ n = φ − xβ̄ β̄ ⇒ φ = i(ξ)χ ⇒ L(ξ)χ =
χ
n
∂xᾱ
∂xα
∂x
∂x
∂x
From now on, considering χ as a 1-form density as we did before, we may consider φ as a density
section of a vector bundle E with dim(E) = 1 and we obtain the defining system in the Medolaghi
form with n equations:
ξα = −
χr (x)∂i ξ r −
1
χi ∂r ξ r + ξ r ∂r χi = 0
p+1
We have seen that this system is involutive when n = 3 but let the reader check as a difficult
exercise that this system is not even formally integrable when n ≥ 5.
We may define the linear first order operator C = A ◦ j1 : E → T and the linear first order
operator D = B ◦ j1 : T → F0 by the two rows of the following commutative and exact diagram:
0 → Q2
↓
→ J2 (E)
↓
0 → Q1
→ J1 (E)
ρ1 (A)
→
A
→
B
J1 (T ) → F0
↓
↓
T
→
→0
0
where Q1 = ker(A) ⊂ J1 (E) and its symbol K1 ⊂ T ∗ ⊗E is easily seen to be involutive with
dim(Kr+1 ) = 1. As the parametrizing operator C is injective with a lift ξ → i(ξ)χ = φ, we obtain
(1)
Q1 = π12 (Q2 ) = 0 and thus Q1 is not formally integrable. However, using Theorem 2.14, we have
(1)
(1)
Qr+1 = ρr (Q1 ) = 0 and thus Qr+1 ≃ Kr+1 ⇒ dim(Qr+1 ) = 1. We obtain therefore at once:
dim(F0 ) = 1 − (n + 1)(n + 2)/2 + n(n + 1) = n(n − 1)/2
and dim(F0 ) = 10 when n = 5 instead of the 5 equations we obtained with the 1-form density χ.
Prolonging this diagram r-times by induction, we obtain the following commutative diagram:
0
↓
0
↓
0 → Kr+2
↓
→ Sr+2 T ∗ ⊗ E
↓
0→
Qr+2
↓
→
Jr+2 (E)
↓
0→
Qr+1
→
Jr+1 (E)
↓
0
0
↓
σr+1 (A)
−→
ρr+1 (A)
−→
ρr (A)
−→
Sr+1 T ∗ ⊗ T
↓
Jr+1 (T )
↓
Jr (T )
↓
0
0
↓
σr (B)
−→
ρr (B)
−→
ρr−1 (B)
−→
Sr T ∗ ⊗ F0
↓
Jr (F0 )
↓
Jr−1 (F0 )
↓
0
Chasing in this diagram while cutting it in the middle by setting:
Rr+1 = im(ρr+1 (A)) ⊆ ker(ρr (B)) = ρr (R1 ) ⇒ πrr+1 (Rr+1 ) = Rr ⇒ gr+1 ⊆ ρr (g1 )
we obtain successively:
28
dim(gr+1 ) =
=
=
=
dim(ρr (g1 ))
=
=
=
=
dim(Rr+1 ) − dim(Rr )
(dim(Jr+2 (E)) − dim(Qr+2 )) − (dim(Jr+1 (E) − dim(Qr+1 ))
dim(Jr+2 (E)) − dim(Jr+1 (E))
dim(Sr+2 T ∗ ) = (r + n + 1)!/(r + 2)!(n − 1)!
dim(ker(σr (B)))
dim(im(σr+1 (A))) + 1
(dim(Sr+2 T ∗ ) − 1) + 1
dim(Sr+2 T ∗ ) = dim(gr+1 )
It follows that gr+1 = ρr (g1 ) and thus Rr+1 = ρr (R1 ) by induction on r. Hence R1 is an involutive
first order system because it is already formally integrable and its symbol g1 is involutuve because
K1 is involutive. Extending step by step the previous diagram on the right, we obtain:
THEOREM 6.1: We have the locally exact Janet sequence where all the operators are first order
and involutive but C which is first order but not formally integrable:
C
D
Dn−2
D
D
2
1
... −→ Fn−2 → 0
F1 −→
0 → E −→ T −→ F0 −→
Proof: We recall that a differential sequence is locally exact, like the Poincaré sequence, if any
(local) section of Fr killed by Dr+1 is the image of a (local) section of Fr−1 by Dr . To prove this
result we just need to apply the Spencer operator D to the middle row of the preceding diagram
when r is large enough. As it is known that all the resulting vertical Spencer sequences are locally
exact (See [31,34,36,52] for more details), then the commutative diagram thus obtained is exact
but perhaps the first purely algebraic column on the left which is an induced δ-sequence, exact
because K1 is involutive. For helping the reader we provide the upper part of this diagram:
0
↓
0
↓
0→
C
→
Kr+2
↓ −δ
0 → T ∗ ⊗ Kr+1
0
↓
0
↓
D
E
−→
T
−→
F0
↓ jr+2
↓ jr+1
↓ jr
→
Jr+2 (E)
−→
Jr+1 (T )
−→
Jr (F0 )
↓D
↓D
↓D
→ T ∗ ⊗ Jr+1 (E) → T ∗ ⊗ Jr (T ) → T ∗ ⊗ Jr−1 (F0 )
→ ...
→ ...
→ ...
Comparing to the Poincaré sequence, we get dim(Fr ) = n!/(r + 2)!(n − r − 2)! and it remains to
find the geometric object providing D. For this, we may introduce the dual density ω = (ω ij ) of
α = (αk1 ,...kn−2 ) = χ ∧ (dχ)p−1 in a symbolic way by introducing ǫ ∈ ∧n T with | ǫi1 ,...,in |= 1 and
set ω ij = ǫijk1 ,...,kn−2 αk1 ,...,kn−2 . Using jet notations, we have the nonlinear system of finite Lie
equations in Lie form:
ω kl (y)
1
∂xi ∂xj ∂(y 1 , ..., y n ) − p+1
(
)
= ω ij (x)
k
l
1
n
∂y ∂y ∂(x , ..., x )
We obtain therefore by linearization the involutive system R1 in of general infinitesimal Lie equations in Medolaghi form:
−ω rj (x)∂r ξ i − ω ir (x)∂r ξ j −
1
ω ij (x)∂r ξ r + ξ r ∂r ω ij (x) = 0
p+1
The Vessiot structure equations involve only one constant and become:
χi (∂j χk − ∂k χj ) + χj (∂k χi − ∂i χk ) + χk (∂i χj − ∂j χi ) = c σijk
where the 3-form density σ = (σijk ) is the dual of (ω)p−1 . Finally, as χ is proportional to the dual
density of (ω)p , it may be expressed rationally in terms of ω. Linearizing as we did in Section 2
and with the Riemann tensor, we obtain D1 with dim(F1 ) = n(n − 1)(n − 2)/6 in a coherent way.
Q.E.D.
29
We use the previous results in order to revisit the Hamilton-Jacobi equation and prove the need
to use differential algebra for studying the nonlinear systems involved. This is a difficult problem
indeed, in the sense that no classical approach by means of exterior calculus can be adapted as
formal integrability or involution become crucial tools that cannot be avoided. Let z = f (t, x)
be a solution of the non-linear PD equation zt + H(t, x, z, zx ) = 0 written with jet notations for
the single unknown z. When dealing with applications, t will be time, x will be space, z will be
the action and, as usual, we shall set p = zx for the momentum. It is important to notice that,
in this general setting, H(t, x, z, p) cannot be called Hamiltonian as it involves z (See [32] and in
particular [33], p 506 for more details):
DEFINITION 6.2: A complete integral z = f (t, x; a, b) is a family of solutions depending on two
constant parameters (a, b) in such a way that the Jacobian condition ∂(z, p)/∂(a, b) 6= 0 whenever
p = ∂x f (t, x; a, b). Using the implicit function theorem, we may set
THEOREM 6.3: The search for a complete integral of the PD equation:
zt + H(t, x, z, zx ) = 0
is equivalent to the search for a single solution of the automorphic system A1 with n = 4, m = 3,
obtained by eliminating ρ(t, x, z, p) in the Pfaffian system:
dz − pdx + H(t, x, z, p)dt = ρ(dZ − P dX)
The corresponding Lie pseudogroup is the pseudogroup Γ of contact transformations of (X, Z, P )
that reproduces the contact 1-form dZ − P dX up to a function factor.
Proof: If z = f (t, x; a, b) is a complete integral, we have:
∂f
∂f
da +
db
∂a
∂b
Using the implicit function theorem and the Jacobian condition, we may set:
dz − pdx + H(t, x, z, p)dt =
a = X(t, x, z, p), b = Z(t, x, z, p) ⇒ ρ(t, x, z, p) =
∂f ∂f
∂f
, P (t, x, z, p) =
/
∂b
∂a ∂b
The converse is left to the reader.
For another solution denoted wit a ”bar”, we have:
dz − pdx + H(t, x, z, p)dt = ρ̄(dZ̄ − P̄ dX̄) ⇒ dZ̄ − P̄ dX̄ =
ρ
(dZ − P dX)
ρ̄
Closing this system, we obtain at once:
ρ
dX̄ ∧ dZ̄ ∧ dP̄ = ( )2 dX ∧ dZ ∧ dP
ρ̄
Closing again, we discover that ρ/ρ̄ is in fact a function of (X, Z, P ), a result bringing the Lie pseudogroup of contact transformations and showing that no restriction must be imposed to H(t, x, z, p).
Q.E.D.
It is quite more dificult to exhibit the equations of the above automorphic sytem and the corresponding equations of the Lie pseudogroup Γ in Lie form or even as involutive systems of PD
equations. From what has been said, we obtain at least:
∂ Z̄
∂X
∂ Z̄
∂Z
X̄
− P̄ ∂∂X
X̄
− P̄ ∂∂Z
= −P,
∂ Z̄
∂P
∂ Z̄
∂Z
X̄
− P̄ ∂∂P
X̄
− P̄ ∂∂Z
=0⇒
∂ X̄
∂ Z̄
− P̄
=0
∂P
∂P
for defining R1 , that is to say:
∂ Z̄
∂ X̄
∂ Z̄
∂ X̄
− P̄
+ P(
− P̄
) = 0,
∂X
∂X
∂Z
∂Z
30
∂ Z̄
∂ X̄
− P̄
=0
∂P
∂P
Using now letters (x, z, p) instead of the capital letters (X, Z, P ) and (ξ, η, ζ) for the corresponding vertical bundles, we obtain by linearization the system of first order infinitesimal Lie equations:
∂η
∂ξ
∂η
∂ξ
∂η
∂ξ
−p
− ζ + p(
− p ) = 0,
−p
=0
∂x
∂x
∂z
∂z
∂p
∂p
This system is not involutive as it is not even formally integrable. Using crossed derivatives in x/p,
we obtain the only new first order equation:
∂ξ
∂ζ
∂η
∂η
−
+
+ 2p
=0
∂x ∂z
∂p
∂z
(1)
and the resulting system R1 is involutive with two equations of class x solved with respect to
(1)
∂ξ ∂η
∂ξ
( ∂x
, ∂x ) and one equation of class p solved with respecto ∂p
, that is dimY (R1 ) = (3+3×3)−3 = 9.
Accordingly, the non-linear system of Lie equations must become involutive by adding only one
equation in Lie form, namely:
∂(Z̄,X̄,P̄ )
∂(Z,X,P )
∂ Z̄
X̄ 2
( ∂Z
− P̄ ∂∂X
)
=1
and its linearization just provides:
∂η
∂ξ
∂ζ
∂ξ
∂η
+
+
= 2(
−p )
∂x ∂z
∂p
∂z
∂z
that is exactly the previous equation.
Coming back to the original system and notations, we may suppose
the 7 = 3 + 4 equations:
∂Z
∂z
− P ∂X
∂z 6= 0 and introduce
∂X
∂Z
∂X
∂Z
∂X
∂Z
∂X
∂Z
∂X
∂Z
−P
+ p(
−P
) = 0,
−P
− H(
−P
) = 0,
−P
=0
∂x
∂x
∂z
∂z
∂t
∂t
∂z
∂z
∂p
∂p
∂(Z, X, P )
∂Z
∂X 2
∂(Z, X, P ) ∂H ∂Z
∂X 2
−(
−P
) = 0,
−
(
−P
) = 0, ...
∂(z, x, p)
∂z
∂z
∂(z, p, t)
∂p ∂z
∂z
Starting now, the next results canot be obtained by exterior calculus and are therefore not
known. Indeed, developping the 3 × 3 Jacobian determinant, the fourth equation provided can be
written as:
∂Z
∂X 2
∂Z ∂(X, P ) ∂Z ∂((X, P ) ∂Z ∂(X, P )
.
−
.
+
.
−(
−P
) =0
∂x ∂(x, p)
∂x ∂(z, p)
∂p ∂(z, x)
∂z
∂z
Using the previous equations in order to eliminate
∂Z
∂x
and
∂Z
∂p ,
we obtain:
∂Z
∂X ∂((X, P )
∂X ∂(X, P )
∂X ∂(X, P )
∂Z ∂(X, P )
.
+ p(
−P
).
−P
.
+P
.
=
∂x ∂(x, p)
∂z
∂z
∂(z, p)
∂x ∂(z, p)
∂p ∂(z, x)
(
∂X ∂(X, P )
∂(X, P )
∂Z
∂X 2
∂Z
−P
)(
+p
)=(
−P
)
∂z
∂z
∂(x, p)
∂(z, p)
∂z
∂z
and thus:
∂(X, P )
∂Z
∂X
∂(X, P )
+p
)−(
−P
)=0
∂(x, p)
∂(z, p)
∂z
∂z
which is nothing else than the first order equation that can be obtained from the first and third
(1)
among the previous 7 equations by using crossed derivatives in x/p. It follows that A1 may be
(1)
defined by 6 equations only and we have thus dimX (A1 ) = (3 + 4 × 3) − 6 = 9. This result proves
(1)
(1)
that the involutive system A1 is an automorphic system for the involutive Lie groupoid R1 .
31
We finally show the link which is existing with differential algebra and the differential Galois
theory because the Lie pseudogroup of contact transformations is an algebraic Lie pseudogroup.
For this, using jet notations, let us consider the chain of strict inclusions of differential fields:
K =Q<
Z x − P Xx Z t − P Xt Z p − P Xp
,
,
>, L = Q < X, Z, P > ⇒ Q ⊂ K ⊂ L
Z z − P Xz Z z − P Xz Z z − P Xz
Using the chain rule for derivatives, we let the reader prove as an exercise that each fraction is a
differential invariant for the Lie pseudogroup Γ of contact transformations. Accordigly, L/K is a
differential automorphic extension in the sense that the corresponding infinite dimensional model
differential variety is a principal homogeneous space (PHS) for Γ. It is not so evident that:
∂(X, Z, P ) ∂Z
∂X 2
/(
−P
) ∈K
∂(x, z, p)
∂z
∂z
because it is also a differential invariant of Γ. The intermediate differential field K ⊂ K ′ ⊂ L with
K ′ = K < Zz − P Xz > is the differential field of invariants of the Lie subpseudogroup Γ′ ⊂ Γ
of strict or unimodular contact transformations preserving the contact form dZ − P dX and thus
the volume 3-form dZ ∧dX ∧dP . We let the reader adapt the previous results to this particular case.
32
7) CONCLUSION
Whenever Rq ⊆ Jq (E) is an involutive system of order q on E, we may define the Janet bundles
Fr for r = 0, 1, ..., n by the short exact sequences:
0 → ∧r T ∗ ⊗ Rq + δ(∧r−1 T ∗ ⊗ Sq+1 T ∗ ⊗ E) → ∧r T ∗ ⊗ Jq (E) → Fr → 0
We may pick up a section of Fr , lift it up to a section of ∧r T ∗ ⊗ Jq (E) that we may lift up to a
section of ∧r T ∗ ⊗ Jq+1 (E) and apply D in order to get a section of ∧r+1 T ∗ ⊗ Jq (E) that we may
project onto a section of Fr+1 in order to construct an operator Dr+1 : Fr → Fr+1 generating the
CC of Dr in the canonical linear Janet sequence ([34], p 145):
D
D
D
D
n
2
1
Fn −→ 0
... −→
F1 −→
0 −→ Θ −→ E −→ F0 −→
If we have two involutive systems Rq ⊂ R̂q ⊂ Jq (E), the Janet sequence for Rq projects onto the
Janet sequence for R̂q and we may define inductively canonical epimorphisms Fr → F̂r → 0 for
r = 0, 1, ..., n by comparing the previous sequences for Rq and R̂q .
A similar procedure can also be obtained if we define the Spencer bundles Cr for r = 0, 1, ..., n by
the short exact sequences:
0 → δ(∧r−1 T ∗ ⊗ gq+1 ) → ∧r T ∗ ⊗ Rq → Cr → 0
We may pick up a section of Cr , lift it to a section of ∧r T ∗ ⊗Rq , lift it up to a section of ∧r T ∗ ⊗Rq+1
and apply D in order to construct a section of ∧r+1 ⊗ Rq that we may project to Cr+1 in order to
construct an operator Dr+1 : Cr → Cr+1 generating the CC of Dr in the canonical linear Spencer
sequence which is another completely different resolution of the set Θ of (formal) solutions of Rq :
jq
D
D
D
D
n
3
2
1
Cn −→ 0
... −→
C2 −→
0 −→ Θ −→ C0 −→
C1 −→
However, if we have two systems as above, the Spencer sequence for Rq is now contained into the
Spencer sequence for R̂q and we may construct inductively canonical monomorphisms 0 → Cr → Ĉr
for r = 0, 1, ..., n by comparing the previous sequences for Rq and R̂q .
When dealing with applications, we have set E = T and considered systems of finite type Lie
equations determined by Lie groups of transformations and ad(Dr ) generates the CC of ad(Dr+1 )
while ad(Dr ) generates the CC of ad(Dr+1 ). We have obtained in particular Cr = ∧r T ∗ ⊗ Rq ⊂
∧r T ∗ ⊗ R̂q = Ĉr when comparing the classical and conformal Killing systems, but these bundles
have never been used in physics. Therefore, instead of the classical Killing system R2 ⊂ J2 (T )
defined by Ω ≡ L(ξ)ω = 0 and Γ ≡ L(ξ)γ = 0 or the conformal Killing system R̂2 ⊂ J2 (T )
defined by Ω ≡ L(ξ)ω = A(x)ω and Γ ≡ L(ξ)γ = (δik Aj (x) + δjk Ai (x) − ωij ω ks As (x)) ∈ S2 T ∗ ⊗ T ,
we may introduce the intermediate differential system R̃2 ⊂ J2 (T ) defined by L(ξ)ω = Aω with
A = cst and Γ ≡ L(ξ)γ = 0, for the Weyl group obtained by adding the only dilatation with
infinitesimal generator xi ∂i to the Poincaré group. We have R1 ⊂ R̃1 = R̂1 but the strict inclusions R2 ⊂ R̃2 ⊂ R̂2 and we discover exactly the group scheme used through this paper, both
with the need to shift by one step to the left the physical interpretation of the various differential
D1
T ∗ ⊗ R̂2 is induced by the
sequences used. Indeed, as ĝ2 ≃ T ∗ , the first Spencer operator R̂2 −→
D
r
r
r
∗
r
− 0) and thus
, ∂i ξrj
usual Spencer operator R̂3 −→ T ⊗ R̂2 : (0, 0, ξrj , ξrij = 0) → (0, ∂i 0 − ξri
∗
∗
∗
projects by cokernel onto the induced operator T → T ⊗ T . Composing with δ, it projects
d
therefore onto T ∗ → ∧2 T ∗ : A → dA = F as in EM and so on by using the fact that D1 and d are
both involutive or the composite epimorphisms Ĉr → Ĉr /C̃r ≃ ∧r T ∗ ⊗ (R̂2 /R̃2 ) ≃ ∧r T ∗ ⊗ ĝ2 ≃
δ
∧r T ∗ ⊗ T ∗ −→ ∧r+1 T ∗ . The main result we have obtained is thus to be able to increase the order
and dimension of the underlying jet bundles and groups, proving therefore that any 1-form with
value in the second order jets ĝ2 (elations) of the conformal Killing system (conformal group) can
be decomposed uniquely into the direct sum (R, F ) where R is a section of the Ricci bundle S2 T ∗
and the EM field F is a section of ∧2 T ∗ as in [41,42](Compare to [55]).
The mathematical structures of electromagnetism and gravitation only depend on second order jets.
33
REFERENCES
[1] Adler, F.W.: Über die Mach-Lippmannsche Analogie zum zweiten Hauptsatz, Anna. Phys.
Chemie, 22, 578-594 (1907).
[2] Airy, G.B.: On the Strains in the Interior of Beams, Phil. Trans. Roy. Soc.London, 153, 1863,
49-80 (1863).
[3] Arnold, V.: Méthodes Mathématiques de la Mécanique Classique, Appendice 2 (Géodésiques
des métriques invariantes à gauche sur des groupes de Lie et hydrodynamique des fluides parfaits),
MIR, Moscow (1974,1976).
[4] Assem, I.: Algèbres et Modules, Masson, Paris (1997).
[5] Beltrami, E.: Osservazioni sulla Nota Precedente, Atti Reale Accad. Naz. Lincei Rend., 5,
141-142 (1892).
[6] Birkhoff, G.: Hydrodynamics, Princeton University Press (1954).
[7] Bjork, J.E. (1993) Analytic D-Modules and Applications, Kluwer (1993).
[8] Bourbaki, N.: Algèbre, Ch. 10, Algèbre Homologique, Masson, Paris (1980).
[9] de Broglie, L.: Thermodynamique de la Particule isolée, Gauthiers-Villars, Pris 1964).
[10] Choquet-Bruhat, Y.: Introduction to General Relativity, Black Holes and Cosmology, Oxford
University Press (2015).
[11] Chyzak, F.,Quadrat, A., Robertz, D.: Effective algorithms for parametrizing linear control
systems over Ore algebras, Appl. Algebra Engrg. Comm. Comput., 16, 319-376, 2005.
[12] Chyzak, F., Quadrat, A., Robertz, D.: OreModules: A symbolic package for the study of
multidimensional linear systems, Springer, Lecture Notes in Control and Inform. Sci., 352, 233264, 2007.
http://wwwb.math.rwth-aachen.de/OreModules
[13] Cosserat, E., & Cosserat, F.: Théorie des Corps Déformables, Hermann, Paris, 1909.
[14] Eisenhart, L.P.: Riemannian Geometry, Princeton University Press, Princeton (1926).
[15] Foster, J., Nightingale, J.D.: A Short Course in General relativity, Longman (1979).
[16] Gröbner, W.: Über die Algebraischen Eigenschaften der Integrale von Linearen Differentialgleichungen mit Konstanten Koeffizienten, Monatsh. der Math., 47, 247-284 (1939).
[17] Hu,S.-T.: Introduction to Homological Algebra, Holden-Day (1968).
[18] Hughston, L.P., Tod, K.P.: An Introduction to General Relativity, London Math. Soc. Students Texts 5, Cambridge University Press (1990).
[19] Janet, M.: Sur les Systèmes aux Dérivées Partielles, Journal de Math., 8, 65-151 (1920).
[20] Kashiwara, M.: Algebraic Study of Systems of Partial Differential Equations, Mémoires de la
Société Mathématique de France, 63 (1995) (Transl. from Japanese of his 1970 Masters Thesis).
[21] Kolchin, E.R.: Differential Algebra and Algebraic groups, Academic Press, New York (1973).
[22] Kumpera, A., & Spencer, D.C.: Lie Equations, Ann. Math. Studies 73, Princeton University
Press, Princeton (1972).
[23] Kunz, E.: Introduction to Commutative Algebra and Algebraic Geometry, Birkhaser (1985).
[24] Lippmann, G.: Extension du Principe de S. Carnot à la Théorie des Pénomènes électriques,
C. R. Acad/ Sc. Paris, 82, 1425-1428 (1876).
[25] Lippmann, G.: Über die Analogie zwischen Absoluter Temperatur un Elektrischem Potential,
Ann. Phys. Chem., 23, 994-996 (1907).
[26] Macaulay, F.S.: The Algebraic Theory of Modular Systems, Cambridge (1916).
[27] Mach, E.: Die Geschichte und die Wurzel des Satzes von der Erhaltung der Arbeit, p 54, Prag:
Calve (1872).
[28]Mach, E.: Prinzipien der Wärmelehre, 2, Aufl., p 330, Leipzig: J.A. Barth (1900).
[29] Maxwell, J.C.: On Reciprocal Figures, Frames and Diagrams of Forces, Trans. Roy. Soc.
Ediinburgh, 26, 1-40 (1870).
[30] Morera, G.: Soluzione Generale della Equazioni Indefinite dellEquilibrio di un Corpo Continuo, Atti. Reale. Accad. dei Lincei, 1, 137-141+233(1892).
[31] Nordström, G.: Einstein’s Theory of Gravitation and Herglotz’s Mechanics of Continua, Proc.
Kon. Ned. Akad. Wet., 19, 884-891 (1917).
[32] Northcott, D.G.: An Introduction to Homological Algebra, Cambridge university Press (1966).
[33] Northcott, D.G.: Lessons on Rings Modules and Multiplicities, Cambridge University Press
(1968).
[34] Oberst, U.: Multidimensional Constant Linear Systems, Acta Appl. Math., 20, 1-175 (1990).
34
[35] Oberst, U.: The Computation of Purity Filtrations over Commutative Noetherian Rings of
Operators and their Applications to Behaviours, Multidim. Syst. Sign. Process. (MSSP) 26,
389-404 (2013).
http://dx.doi.org/10.1007/s11045-013-0253-4
[36] Ougarov, V.: Théorie de la Relativité Restreinte, MIR, Moscow, 1969, (french translation,
1979).
[37] Poincaré, H.: Sur une Forme Nouvelle des Equations de la Mécanique, C. R. Académie des
Sciences Paris, 132 (7) (1901) 369-371.
[38] Pommaret, J.-F.: Systems of Partial Differential Equations and Lie Pseudogroups, Gordon
and Breach, New York, 1978; Russian translation: MIR, Moscow, 1983.
[39] Pommaret, J.-F.: Differential Galois Theory, Gordon and Breach, New York, 1983.
[40] Pommaret, J.-F.: Lie Pseudogroups and Mechanics, Gordon and Breach, New York, 1988.
[41] Pommaret, J.-F.: Partial Differential Equations and Group Theory, Kluwer, 1994.
http://dx.doi.org/10.1007/978-94-017-2539-2
[42] Pommaret, J.-F.: François Cosserat and the Secret of the Mathematical Theory of Elasticity,
Annales des Ponts et Chaussées, 82, 59-66 (1997) (Translation by D.H. Delphenich).
[43] Pommaret, J.-F.: Group Interpretation of Coupling Phenomena, Acta Mechanica, 149 (2001)
23-39.
http://dx.doi.org/10.1007/BF01261661
[44] Pommaret, J.-F.: Partial Differential Control Theory, Kluwer, Dordrecht, 2001.
[45] POMMARET, J.-F.: Algebraic Analysis of Control Systems Defined by Partial Differential
Equations, in ”Advanced Topics in Control Systems Theory”, Springer, Lecture Notes in Control
and Information Sciences 311 (2005) Chapter 5, pp. 155-223.
[46] Pommaret, J.-F.: Arnold’s Hydrodynamics Revisited, AJSE-mathmatiques, 1, 1, 2009, pp.
157-174.
[47] Pommaret, J.-F.: Parametrization of Cosserat Equations, Acta Mechanica, 215 (2010) 43-55.
http://dx.doi.org/10.1007/s00707-010-0292-y
[48] Pommaret, J.-F.: Macaulay Inverse Systems revisited, Journal of Symbolic Computation, 46,
1049-1069 (2011).
[49] Pommaret, J.-F.: Spencer Operator and Applications: From Continuum Mechanics to Mathematical Physics, in ”Continuum Mechanics-Progress in Fundamentals and Engineering Applications”, Dr. Yong Gan (Ed.), ISBN: 978-953-51-0447–6, InTech, 2012, Available from:
http://dx.doi.org/10.5772/35607
[50] Pommaret, J.-F.: The Mathematical Foundations of General Relativity Revisited, Journal of
Modern Physics, 4 (2013) 223-239.
http://dx.doi.org/10.4236/jmp.2013.48A022
[51] Pommaret, J.-F.: The Mathematical Foundations of Gauge Theory Revisited, Journal of Modern Physics, 5 (2014) 157-170.
http://dx.doi.org/10.4236/jmp.2014.55026
[52] Pommaret, J.-F.: Relative Parametrization of Linear Multidimensional Systems, Multidim.
Syst. Sign. Process., 26, 405-437 2015).
DOI 10.1007/s11045-013-0265-0
[53] Pommaret,J.-F.:From Thermodynamics to Gauge Theory: the Virial Theorem Revisited, pp.
1-46 in ”Gauge Theories and Differential geometry,”, NOVA Science Publisher (2015).
[54] Pommaret, J.-F.: Airy, Beltrami, Maxwell, Einstein and Lanczos Potentials revisited, Journal
of Modern Physics, 7, 699-728 (2016).
http://dx.doi.org/10.4236/jmp.2016.77068
[55] Pommaret, J.-F.: Deformation Theory of Algebraic and Geometric Structures, Lambert Academic Publisher (LAP), Saarbrucken, Germany (2016). A short summary can be found in ”Topics
in Invariant Theory ”, Séminaire P. Dubreil/M.-P. Malliavin, Springer Lecture Notes in Mathematics, 1478, 244-254 (1990).
http://arxiv.org/abs/1207.1964
[56] Pommaret, J.-F. and Quadrat, A.: Localization and Parametrization of Linear Multidimensional Control Systems, Systems & Control Letters, 37, 247-260 (1999).
[57] Pommaret, J.-F., Quadrat, A.: Algebraic Analysis of Linear Multidimensional Control Systems, IMA Journal of Mathematical Control and Informations, 16, 275-297 (1999).
[58] Quadrat, A., Robertz, D.: Parametrizing all solutions of uncontrollable multidimensional lin35
ear systems, Proceedings of the 16th IFAC World Congress, Prague, July 4-8, 2005.
[59] Quadrat, A.: An Introduction to Constructive Algebraic Analysis and its Applications, Les
cours du CIRM, Journees Nationales de Calcul Formel, 1(2), 281-471 (2010).
[60] Quadrat, A., Robertz, R.: A Constructive Study of the Module Structure of Rings of Partial
Differential Operators, Acta Applicandae Mathematicae, 133, 187-234 (2014).
http://hal-supelec.archives-ouvertes.fr/hal-00925533
[61] Rotman, J.J.: An Introduction to Homological Algebra, Pure and Applied Mathematics, Academic Press (1979).
[62] Schneiders, J.-P.: An Introduction to D-Modules, Bull. Soc. Roy. Sci. Liège, 63, 223-295
(1994).
[63] Spencer, D.C.: Overdetermined Systems of Partial Differential Equations, Bull. Am. Math.
Soc., 75 (1965) 1-114.
[64] Teodorescu, P.P.: Dynamics of Linear Elastic Bodies, Abacus Press, Tunbridge, Wells (1975)
(Editura Academiei, Bucuresti, Romania).
[65] Vessiot, E.: Sur la Théorie des Groupes Infinis, Ann. Ec. Norm. Sup., 20, 411-451 (1903)
(Can be obtained from http://numdam.org).
[66] Weyl, H.: Space, Time, Matter, Springer, 1918, 1958; Dover, 1952.
[67] Zerz, E.: Topics in Multidimensional Linear Systems Theory, Lecture Notes in Control and
Information Sciences (LNCIS) 256, Springer (2000).
[68] Zou, Z., Huang, P., Zang ,Y., Li, G.: Some Researches on Gauge Theories of Gravitation,
Scientia Sinica, XXII, 6, 628-636 (1979).
36
| 0math.AC
|
Fast Deterministic Selection
Andrei Alexandrescu
arXiv:1606.00484v2 [cs.DS] 4 Aug 2016
The D Language Foundation
[email protected]
Abstract
lection algorithm most used in practice [8, 27, 31]. It can
be thought of as a partial Q UICKSORT [16]: Like Q UICK SORT , Q UICKSELECT relies on a separate routine to divide
the input array into elements less than or equal to and greater
than or equal to a chosen element called the pivot. Unlike
QUICKSORT which recurses on both subarrays left and right
of the pivot, Q UICKSELECT only recurses on the side known
to contain the kth smallest element. (The side to choose is
known because the partitioning computes the rank of the
pivot.)
The pivot choosing strategy strongly influences Q UICK SELECT’s performance. If each partitioning step reduces the
array portion containing the kth order statistic by at least a
fixed fraction, Q UICKSELECT runs in O(n) time; otherwise,
Q UICKSELECT is liable to run in quadratic time. Therefore,
strategies for pivot picking are a central theme in Q UICKSE LECT. Some of the most popular and well-studied heuristics
choose the median out of a small (1–9) number of elements
(either from predetermined positions or sampled uniformly
at random) and use that as a pivot [3, 8, 14, 32]. However,
such heuristics cannot provide strong worst-case guarantees.
Selection in guaranteed linear time remained an open
problem until 1973, when Blum, Floyd, Pratt, Rivest, and
Tarjan introduced the seminal “Median of Medians” algorithm [4] (also known as BFPRT after the initials of its authors), an ingenious pivot selection method that works in
conjunction with Q UICKSELECT. § 3 discusses both Q UICK SELECT and M EDIAN O F M EDIANS in detail.
Although widely described and studied, M EDIAN O F M E DIANS is seldom used in practice. This is because its pivot
finding procedure has run time proportional to the input size
and is relatively intensive in both element comparisons and
swaps. In contrast, sampling-based pivot picking only does a
small constant amount of work. In practice, the better quality
of the pivot found by M EDIAN O F M EDIANS does not justify its higher cost. Therefore, the state of the art in selection with partitioning has been Q UICKSELECT in conjunction with simple pivot choosing heuristics.
It would seem that heuristics-based selection has won
the server room, leaving deterministic selection to the classroom. However, fast deterministic selection remains desirable for several practical reasons:
The Median of Medians (also known as BFPRT) algorithm,
although a landmark theoretical achievement, is seldom used
in practice because it and its variants are slower than simple approaches based on sampling. The main contribution
of this paper is a fast linear-time deterministic selection algorithm Q UICKSELECTA DAPTIVE based on a refined definition of M EDIAN O F M EDIANS. The algorithm’s performance brings deterministic selection—along with its desirable properties of reproducible runs, predictable run times,
and immunity to pathological inputs—in the range of practicality. We demonstrate results on independent and identically distributed random inputs and on normally-distributed
inputs. Measurements show that Q UICKSELECTA DAPTIVE
is faster than state-of-the-art baselines.
Categories and Subject Descriptors I.1.2 [Symbolic and
Algebraic Manipulation]: Algorithms
1.
Introduction
The selection problem is widely researched and has many
applications. In its simplest formulation, selection is finding the kth smallest element (also known as the kth order
statistic) of an array: given an unstructured array A, a nonstrict order ≤ over elements of A, and an index k, the task
is to find the element that would be at position A[k] if the
array were fully sorted. A variant that is the focus of this
paper is partition-based selection: in addition to finding the
kth smallest element, the algorithm must also permute elements of A such that A[i] ≤ A[k] ∀i, 0 ≤ i < k, and
A[k] ≤ A[i] ∀i, k ≤ i < n. Sorting A solves the problem in
O(n log n), so the challenge is to achieve selection in O(n)
time.
Selection often occurs as a step in more complex algorithms such as shortest path, nearest neighbor, and ranking.
Q UICKSELECT, invented by Hoare in 1961 [15], is the se-
[Copyright notice will appear here once ’preprint’ option is removed.]
1
2016/7/27
• Immunity to pathological inputs: Deterministic sampling
dard library, where it can serve as a reference and basis for
porting to other languages and systems.
In the following sections we use the customary pseudocode and algebraic notations to define algorithms. Divisions of integrals are integral with truncation as in many programming languages, e.g. n/k is b nk c. When real division
is needed, we write real(n)/real(k) to emphasize conversion of operands to real numbers prior to the division. Arrays
are zero-based so as to avoid minute differences between the
pseudocode algorithms and their implementations available
online. The length of array A is written as |A|. We denote
with A[a : b] (if a < b) a “slice” of A starting with A[a] and
ending with (and including) A[b − 1]. The slice A[a : b] is
empty if a = b. Elements of the array are not copied—the
slice is a bounded view into the existing array.
The next section discusses related work. § 3 reviews the
algorithm definitions we start from. § 4 introduces the improvements we propose to M EDIAN O F M EDIANS and R E PEATED S TEP , which are key to better performance. § 5 discusses adaptation for M EDIAN O F M EDIANS and derivatives.
§ 6 includes a few engineering considerations. § 7 presents
experiments and results with the proposed algorithms. A
summary concludes the paper.
heuristics (such as “median of first, middle, and last element”) are all susceptible to pathological inputs that
make Q UICKSELECT’s run time quadratic. These inputs
are easy to generate and contain patterns that are plausible in real data [26, 27, 34]. In order to provide linearity guarantees, many current implementations of Q UICK SELECT include code that detects and avoids such situations, at the cost of sometimes falling back to a slower
(albeit not quadratic) algorithm, such as M EDIAN O F M EDIANS itself. The archetypal example is I NTROSE LECT [27]. A linear deterministic algorithm that is also
fast would avoid these inefficiencies and complications
by design.
• Deterministic behavior: Randomized pivot selection
leaves the input array in a different configuration after
each call (even for identical inputs), which makes it difficult to reproduce runs (e.g. for debugging purposes).
In contrast, deterministic selection always permutes elements of a given array the same way.
• Predictability of running time: In real-time systems
(e.g. that compute the median of streaming packets) a
randomized pivot choice may cause problems. The first
1–3 pivot choices have a high impact on the overall run
time of quickselect with randomized pivots. This makes
for a large variance in individual run times [9, 18], even
against the same input. Deterministic algorithms have a
more predictable run time.
2.
Related Work
The selection problem has a long history, starting with Lewis
Carrol’s 1883 essay on the fairness of tennis tournaments (as
recounted by Knuth [20, Vol. 3]).
Hoare [15] created the Q UICKSELECT algorithm in 1961,
which still is the preferred choice of practical implementations, in conjunction with various pivot choosing strategies. Sophisticated running time analyses exist for quickselect [13, 18, 19, 23, 28]. Martı́nez, Panario, and Viola [24]
analyze the behavior of Q UICKSELECT with small data sets
and propose stopping Q UICKSELECT’s recursion early and
using sorting as an alternative policy below a cutoff, essentially a simple multi-strategy Q UICKSELECT. Same authors [25] propose several adaptive sampling strategies for
Q UICKSELECT that take into account the index searched.
Blum, Floyd, Pratt, Rivest, and Tarjan created the M EDI AN O F M EDIANS algorithm [4]. Subsequent work provided
variants and improved on its theoretical bounds [5, 11, 12,
21, 36, 37]. Chen and Dumitrescu [6] propose variants of
M EDIAN O F M EDIANS that group 3 or 4 elements at a time
(the original uses groups of 5 or more elements). Most variants offer better bounds and performance than the original,
but to date neither has been shown to be competitive against
heuristics-based algorithms.
Battiato et al. [2] describe Sicilian Median Selection,
a fast algorithm for computing an approximate median. It
could be considered the transitive closure of Chen and Dumitrescu’s R EPEATED S TEP algorithm. In spite of good average performance, the algorithm’s worst-case pivot quality
is insufficient for guaranteeing exact selection in linear time.
We seek to improve the state of the art in fast deterministic selection, specifically aiming practical large-scale applicability. First, we propose a refined definition of M EDI AN O F M EDIANS. The array permutations necessary for finding the pivot also contribute to the partitioning process, thus
reducing both overall comparisons and data swaps. We apply
this idea to Chen and Dumitrescu’s recent R EPEATED S TEP
algorithm [6], a M EDIAN O F M EDIANS variant particularly
amenable to our improvements, and we obtain a competitive
baseline.
Second, we add adaptation to M EDIAN O F M EDIANS.
The basic observation is that M EDIAN O F M EDIANS is specialized in finding the median, not some arbitrary order
statistic. That makes the performance of M EDIAN O F M E DIANS degrade considerably when selecting order statistics
away from the median (e.g. 25th percentile). We devise simple and efficient specialized algorithms for searching for indexes away from the middle of the searched array. The best
algorithm to use is chosen dynamically.
The resulting composite algorithm Q UICKSELECTA DAP TIVE was measured to be faster than relevant state-of-the-art
baselines, which makes it a candidate for solving the selection problem in practical libraries and applications. We
propose a generic open-source implementation of Q UICK SELECTA DAPTIVE for inclusion in the D language’s stan2
2016/7/27
3.
Background
3.1
Q UICKSELECT and M EDIAN O F M EDIANS
Most pivot selection schemes are based on heuristics:
they choose an array element likely to be not too close to
the minimum or maximum, and then use H OARE PARTI TION [16] (Algorithm 2), which partitions the array around
that element in linear time. Some popular and well researched heuristics include:
We provide a quick overview of the Q UICKSELECT and
M EDIAN O F M EDIANS algorithms [8, 20] with an emphasis
on practical implementations.
Algorithm 1 illustrates Q UICKSELECT. It takes as parameters not only the data (array A and index k), but also a partitioning primitive called PARTITION, in a higher-order function fashion. PARTITION(A, k) returns an index p (called
pivot) with 0 ≤ p < |A| and also partitions A such that
A[j] ≤ A[p] ∀j, 0 ≤ j < p, and A[p] ≤ A[j] ∀j, p < j <
|A|. Q UICKSELECT, in a divide and conquer manner, uses
the pivot to either reduce the portion of the array searched
from the left when p < k, reduce it from the right when
k < p, or finish the search when p = k. Overall, Q UICKSE LECT uses an approximate partitioning primitive repeatedly
to obtain a precise partition around the specific index k.
• Random pivot [8]: the pivot is chosen uniformly at ran-
dom from A;
• Median of 3 [32]: the pivot is the median of A [0],
A [|A|/2], and A [|A| − 1];
• Median of 3 randomized [17, 32]: the pivot is the median
of three elements chosen uniformly at random from A;
• Tukey’s ninther [3, 33]: Choose A [0], A [|A|/8],
A [2|A|/8],
A [3|A|/8],
A [4|A|/8],
A [5|A|/8],
A [6|A|/8], A [7|A|/8], and A [|A| − 1], then compute the median of first three, middle three, and last
three, and finally the pivot is the median of the three
medians;
Algorithm 1: Q UICKSELECT
Data: PARTITION, A, k with |A| > 0, 0 ≤ k < |A|
Result: Puts kth smallest element of A in A[k] and
partitions A around it.
1 while true do
2
p ← PARTITION(A, k);
3
if p = k then
4
return;
5
end
6
if p > k then
7
A ← A[0 : p];
8
else
9
k ← k − p − 1;
10
A ← A[p + 1 : |A|];
11
end
12 end
• Tukey’s ninther randomized [34]: Similar to the ninther,
but the nine samples are chosen uniformly at random
from A.
Algorithm 2: H OARE PARTITION
Data: A, p with |A| > 0, 0 ≤ p < |A|
Result: p0 , the new position of A[p]; A partitioned at
A[p0 ]
1 S WAP (A[p], A[0]);
2 a = 1; b = |A| − 1;
3 loop: while true do
4
while true do
5
if a > b then break loop;
6
if A[a] ≥ A[0] then break;
7
a ← a + 1;
8
end
9
while A[0] < A[b] do b ← b − 1;
10
if a ≥ b then break;
11
S WAP(A[a], A[b]);
12
a ← a + 1;
13
b ← b − 1;
14 end
15 S WAP (A[0], A[a − 1]);
16 return a − 1;
The running time is linear if PARTITION(A, k) is linear
and able to return a pivot p within some fixed fraction f <
1 from either end of A, i.e. p > f |A| and p < (1 −
f )|A|, which allows Q UICKSELECT to eliminate at least
f |A| elements at each pass. In that case, the time complexity
T (n) (depending on n = |A|) of Q UICKSELECT satisfies:
T (n) ≤ T ((1 − f )n) + O(n)
(1)
which results in T (n) = O(n) by the Master Theorem [30].
There is no required relationship between k and p, and
in fact many partitioning routines do not use k. It may be
considered a hint. Ideally, the partitioning primitive would
return a pivot equal to k, or close to k (greater than it if
k < |A|
2 , smaller otherwise). That allows Q UICKSELECT to
either end the search or eliminate many elements at each
pass. Contemporary approaches prevalently ignore k and
attempt to partition close to the median, strategy shown
recently to be suboptimal [25].
No heuristics-based approach can provide a good worstcase run time guarantee. However, the constant-time speed
of computing the pivot compensates on average for its poor
quality.
In contrast, M EDIAN O F M EDIANS spends more time to
systematically guarantee good pivot choices. Algorithm 3 illustrates the prevalent implementation of M EDIAN O F M E DIANS [7, 10, 29, 35].
3
2016/7/27
Algorithm 3: BFPRTBASELINE
Data: A, k (k not used)
Result: Pivot 0 ≤ p < |A|; A partitioned at A[p]
1 if |A| < 5 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 i ← 0; j ← 0;
5 while i + 4 < |A| do
6
M EDIAN 5(A, i, i + 1, i + 2, i + 3, i + 4);
7
S WAP(A[i + 2], A[j]);
// median to 1st quintile
8
i ← i + 5;
9
j ← j + 1;
10 end
11 Q UICKSELECT (BFPRTBASELINE , A[0 : j], j/2);
12 return H OARE PARTITION (A, j/2);
Algorithm 4: M EDIAN 5
Data: A, a, b, c, d, e
Result: Puts median of A[a], A[b], A[c], A[d], A[e] in
A[c] and leaves the smaller items in A[a], A[b]
and the larger ones in A[d], A[e].
1 if A[c] < A[a] then S WAP (A[a], A[c]);
2 if A[d] < A[b] then S WAP (A[b], A[d]);
3 if A[d] < A[c] then
4
S WAP(A[c], A[d]);
5
S WAP(A[a], A[b]);
6 end
7 if A[e] < A[b] then S WAP (A[b], A[e]);
8 if A[e] < A[c] then
9
S WAP(A[c], A[e]);
10
if A[c] < A[a] then S WAP(A[a], A[c]);
11 else
12
if A[c] < A[b] then S WAP(A[b], A[c]);
13 end
The algorithm first computes medians of groups of 5 elements at a time, for a total of |A|
5 groups. The primitive
M EDIAN 5 takes an array and five distinct indexes, computes
their median with a small specialized algorithm, and places
it in the third index received. (The last |A| mod 5 elements
of the array are ignored for simplicity.) Computing (recursively) the median of these medians yields a pivot with a
desirable property. There are |A|
10 elements less than or equal
to the pivot, but each of those is the median of some 5 distinct elements, so for each of those |A|
10 elements there are 2
others also less than or equal to the pivot. That guarantees at
least 3|A|
10 elements less than or equal to the pivot. By symmetry, at least 3|A|
10 elements in the array are greater than or
equal to the pivot.
Selection is initiated by invoking Q UICKSE LECT(BFPRTBASELINE , A, k). Line 11 passes the
name of the algorithm (BFPRTBASELINE) as an argument
to Q UICKSELECT. This makes the two algorithms mutually
recursive: BFPRTBASELINE is the partitioning primitive
used by Q UICKSELECT, and in turn uses Q UICKSELECT
itself.
M EDIAN 5 is used intensively by the algorithm and therefore deserves a careful implementation. A good balance
needs to be found between comparisons, swaps, and function size. We chose a function with 6 comparisons and 0–7
swaps (only 3.13 on average against distinct random numbers), shown in Algorithm 4. (To simplify notation we use
A[a] < A[b] as a shortcut for ¬(A[b] ≤ A[a]).) The function
is also idempotent: if the five slots are already partitioned, it
performs no swaps.
To prove linearity for M EDIAN O F M EDIANS, let us look
at the number of comparisons C(n) that BFPRTBASELINE
executes as a function of n = |A|. In the worst case, we
have:
C(n) ≤ C
n
5
+C
7n
10
+
6n
+n
5
(2)
where the first term accounts for the recursive call to compute the median of medians, the second is the recursive call
in Q UICKSELECT (in the worst case 3n
10 elements are eliminated in one partitioning step so 7n
are
left), the third is the
10
cost of computing the medians of five, and the last is the cost
of H OARE PARTITION. Consequently C(n) ≤ 22n.
The number of swaps S(n) is also of interest. That satisfies the recurrence:
S(n) ≤ S
n
5
+S
7n
10
+
n
7n n − 10
+
5
2
(3)
The terms correspond to those for C(n). Our median of
five routine uses at most 7 swaps. Also in the last term we
n
eliminate the first 10
elements of the array. Those are known
to be no greater than the median so they will not be swapped.
Consequently S(n) ≤ 37n
2 .
Although Algorithm 3 uses groups of 5, any greater constant works as well (for the implementation and the linearity
proof above); in fact, Blum et al. discuss group sizes of 15
and 21 elements. Group size 5 is prevalently used today because it is the most effective in practice.
3.2 The R EPEATED S TEP Algorithm
Recently Chen and Dumitrescu [6] proposed M EDIAN O F M EDIANS variants that use groups of 3 or 4 elements
yet have linear worst-case running time, disproving longstanding conjectures to the contrary. Algorithm 5 shows the
pseudocode of their R EPEATED S TEP algorithm with a group
size of 3.
4
2016/7/27
of three (up to three comparisons per group of three), computing the medians of medians of three, and the partitioning.
Consequently C(n) ≤ 21n.
We now calculate a bound for S(n). Each median of three
uses up to two swaps, to which we add one for placing the
median at the front of the array. The partitioning costs at
most n2 swaps in general, but in this case we already have
n
elements of the array known to be no greater
the first 18
than the pivot, so they will not be swapped. We obtain:
Algorithm 5: R EPEATED S TEP
Data: A, k (k not used)
Result: Pivot 0 ≤ p < |A|; A partitioned at A[p]
1 if |A| < 9 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 i ← 0; j ← 0;
5 while i + 2 < |A| do
6
M EDIAN 3(A, i, i + 1, i + 2);
7
S WAP(A[i + 1], A[j]);
// median to 1st tertile
8
i ← i + 3;
9
j ← j + 1;
10 end
11 i ← 0; m ← 0;
12 while i + 2 < j do
13
M EDIAN 3(A, i, i + 1, i + 2);
14
S WAP(A[i + 1], A[m]);
// median to 1st 9th-ile
15
i ← i + 3;
16
m ← m + 1;
17 end
18 Q UICKSELECT (R EPEATED S TEP , A[0 : m], m/2);
19 return H OARE PARTITION (A, m/2);
S(n) ≤ S
n
9
+C
7n
9
+
3n 3n
+
+n
3
9
9
+S
7n
9
+n+
n
n n − 18
+
3
2
(5)
which solves to S(n) ≤ 65n
4 . Both bounds are better than
those of BFPRTBASELINE.
4.
Improving Layout
We now start improving these algorithms, first by laying out
data in a more efficient manner. Note that BFPRTBASELINE
(Algorithm 3) already embeds an important optimization—
it reuses the first quintile of A for storing the medians.
This approach avoids the additional overhead of creating and
using a separate temporary array to store the medians of 5
and is in keeping with today’s most frequently encountered
implementations [7, 10, 29, 35].
One key insight motivates a different layout choice: we
aim to make the comparisons and swaps performed during
the medians of 5 computation also count toward the partitioning. M EDIAN O F M EDIANS organizes the array in groups
of 5 and computes the median of each. In addition, in each
group of 5, the 2 smaller elements are to the left, and the 2
larger ones are to the right of the median. That imparts quite
a non-trivial implicit structure onto the array in addition to
computing the pivot. However, that structure is of no use to
H OARE PARTITION. Ideally, that structure should be embedded in the array in a form favorable to the subsequent partitioning step; conversely, the partitioning step should make
maximal use of the implicit information established in the
array by the medians of 5 stage. Indeed, the original M EDI AN O F M EDIANS paper [4] provides an optimized algorithm
called P ICK 1 that makes use of the implicit structuring of
the input, but is still not efficient in practice.
Our approach (BFPRTI MPROVED shown in Algorithm 6) is to make the groups of 5 non-contiguous and lay
them out in a manner that is advantageous for the partitioning step. We place the subarray of medians in the third quintile, the smaller values to its left, and the larger values to
its right. This is efficient because the third quintile is in the
middle of the entire array. After that quintile is partitioned
around its own median (by the recursive call to Q UICKSE LECT), it already is globally partitioned properly around the
pivot; there is no more need to visit it. That way, the medians
of 5 computation step contributes one fifth of the final result
at no additional cost. The subsequent partitioning step only
The algorithm is similar to BFPRTBASELINE with group
size changed to 3, with one important addition: the median
of medians step is repeated, essentially choosing the pivot
as the median of medians of medians of three (sic), instead
of median of medians of three. This degrades the quality of
the pivot because of the double approximation: the pivot is
guaranteed to be within 2n
9 elements from either edge of the
array (weaker than n3 without the repeated step, or 3n
10 if using groups of 5). However, this loss in quality comes with an
improvement in speed: the median of medians computation
only needs to recurse on n9 elements instead of n3 .
Intuitively, trading off some pivot quality for faster processing in M EDIAN O F M EDIANS is a good idea, since pivot
heuristics that provide relatively poor pivot estimates very
fast perform so well in practice. There are advantages specific to using groups of 3—computing the median of 3 is
simpler and takes 2–3 comparisons whereas the median of 5
requires 6 comparisons. So it is expected that R EPEATED S TEP improves performance over BFPRTBASELINE. To assess that, let us estimate the number of comparisons C(n)
performed by R EPEATED S TEP from the recurrence:
C(n) ≤ C
n
(4)
where the first term is the cost of recursion for computing the
median of medians of medians, the second is the worst-case
time spent processing the remaining elements, and the last
three terms account respectively for computing the medians
5
2016/7/27
needs to partition quintiles 1, 2, 4, and 5. Better yet, quintiles 1 and 2 contain the statistically smaller elements and
quintiles 4 and 5 contain the statistically larger elements so
the medians of 5 computation also saves on swapping during
partitioning.
In order to distribute the groups of 5 appropriately, we
take advantage of the fact that the M EDIAN 5 primitive (Algorithm 4) accepts any five distinct indexes, not only consecutive ones. So we can choose any convenient grouping
and iteration schedule. Let f = |A|
5 such that quintiles start
at A[0], A[f ], A[2f ], A[3f ], and A[4f ]. Then we choose the
first group as M EDIAN 5(A, 0, 1, 2f, 3f, 3f + 1). These are
respectively the leftmost two elements to the left of the third
quintile, the first element of the third quintile, and the leftmost two elements to the right of the third quintile. Iteration proceeds with median calls following the schedule M E DIAN 5(A, i, i + 1, 2f + j, 3f + i, 3f + i + 1), with i iterating
with step 2 and j iterating with step 1. The entire array is
covered when j = f .
Computing medians of five using the iteration strategy
above and then computing the median of the third quintile
leaves the array with the following layout:
Algorithm 6: BFPRTI MPROVED
Data: A, k (k not used)
Result: Pivot p, 0 ≤ p < |A|
1 if |A| < 5 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 f ← |A|/5; i ← 0;
5 for j ← 2f through 3f − 1 do
6
M EDIAN 5(A, i, i + 1, j, 3f + i, 3f + i + 1);
7
i ← i + 2;
8 end
9 Q UICKSELECT (BFPRTI MPROVED , A[2f : 3f ], f /2);
return
E XPAND PARTITION(A, 2f, 2f + f /2, 3f − 1);
We now derive worst-case bounds for BFPRTI MPROVED
depending on n = |A|. The only difference in computing
C(n) is one fifth of the elements do not need to be compared
any more, which changes the last term:
• The elements likely less than or equal to the pivot (at least
3n
10 )
C(n) ≤ C
are already to the left of A’s middle.
• The elements likely larger than or equal to the pivot
(again at least
3n
10 )
S(n) ≤ S
and are already partitioned around the pivot, so there
is no more need for the partitioning step to compare n5
elements.
0
≤ A[p] A[p] ≥ A[p]
i→ a
p
?
b ←j
n
+S
5
7n
10
+
6n 4n
+
5
5
(6)
5
7n
10
+
7n 2n
+
5
10
(7)
The last term changed because there are at most 2n
10 elements on either side to swap across the pivot, and the algorithm puts two elements in the right place with one swap
whenever possible. The equation results in S(n) ≤ 16n. By
these estimates, BFPRTI MPROVED reduces comparisons by
about 9% and swaps by about 14% over BFPRTBASELINE.
This leaves the array well suited for partitioning with only
light swapping. The subarrays A[0 : 2f ] and A[3f : |A|]
need to be visited and partitioned properly. This work is
carried by E XPAND PARTITION (not shown as it is trivial
yet replete with detail), a slightly modified H OARE PARTI TION that takes into account the already-partitioned subarray
around the pivot. The call E XPAND PARTITION(A, a, p, b)
proceeds by the following scheme, starting with i = 0 and
j = |A| − 1 and moving them toward a and b, respectively:
?
+C
which leads to C(n) ≤ 20n. For swaps:
are already to the right of A’s middle.
• The medians of five are in the third quintile of the array
≤ A[p]
n
4.1 Improving Layout for R EPEATED S TEP
Let us now apply the same layout idea to the R EPEATED S TEP algorithm. The result is shown in Algorithm 7. The
loop starting at line 5 computes medians of 3 noncontiguous
elements using M EDIAN 3 (trivial, not shown): first from the
first tertile, second from the middle tertile, and third from the
last tertile. The three indexes move in lockstep, so only one
iteration variable is needed. Similarly, the next loop uses the
same iteration schedule to compute medians of three for the
middle tertile. This leaves the medians of medians of medians of 3 in the fifth 9th ile (middle “novile”) of the array.
The call to Q UICKSELECT partitions the mid 9th ile
around its (exact) median. Right after line 11 the following
conditions are satisfied:
≥ A[p]
The procedure swaps as many elements as possible between A[i : a] and A[b + 1 : j + 1], because that is the most
efficient use of swapping—one swap puts two elements in
their final slots. Often there will be some asymmetry (one of
i and j reaches its limit before the other) so the pivot position will need to shift to the left or to the right. E XPAND PARTITION returns the final position of the pivot, which
BFPRTI MPROVED forwards to the caller. E XPAND PARTI TION (A, a, b) performs max(a, |A| − b) comparisons and at
most max(a, |A| − b) swaps.
• Pivot value is in A[p];
• A[4f : p] contains only elements less than or equal to the
pivot;
6
2016/7/27
These bounds are adequate for proving linearity and useful as proxies, but not necessarily tight (producing worstcase inputs for M EDIAN O F M EDIANS remains an open problem); the measured speed improvements are larger. Although
algorithms with better upper bounds (both comparisons and
swaps) for selection have been devised [5, 21], they are more
complex and entail large hidden costs; to date neither of
those has been shown to be efficient in practice.
This layout change has been crucial for obtaining efficiency from M EDIAN O F M EDIANS. The improved layout
not only drastically improves performance, but also allows
further optimizations to build upon it. R EPEATED S TEP I M PROVED is our core algorithm for fast deterministic selection.
Algorithm 7: R EPEATED S TEP I MPROVED
Data: A, k (k not used)
Result: Pivot p, 0 ≤ p < |A|; A partitioned at p
1 if |A| < 9 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 f ← |A|/9;
5 for i ← 3f through 6f − 1 do
6
M EDIAN 3(A, i − 3f, i, i + 3f );
7 end
8 for i ← 4f through 5f − 1 do
9
M EDIAN 3(A, i − f, i, i + f );
10 end
11 Q UICKSELECT (R EPEATED S TEP I MPROVED , A[4f :
5f ], f /2);
12 return E XPAND PARTITION (A, 4f, 4f + f /2, 5f − 1);
5.
The performance of R EPEATED S TEP I MPROVED compares
favorably with state-of-the-art baselines when searching the
median. However, performance degrades quickly as the index sought migrates from the middle of the array toward either of its ends. This is because R EPEATED S TEP I MPROVED
(like all M EDIAN O F M EDIANS variations discussed so far)
is specialized in finding a good approximate median, not a
good approximate kth order statistic for any k. We acknowledged that by mentioning “k not used” in each algorithm’s
parameters section.
However, k may be very informative especially when far
from the middle. For an extreme example, consider searching a position k only a few slots from one edge of the array.
Then any heuristic will produce a good-quality pivot very
quickly simply because the range of good pivots is large. In
contrast, a M EDIAN O F M EDIANS algorithm would do considerable work to eliminate about half the array. This continues through the end of the search—heuristics eliminate
swaths of the array at constant cost, whereas M EDIAN O F M EDIANS pedantically reduces the array by about half at
every pass at linear cost. The average time complexity is linear for either approach, but heuristics-based algorithms will
complete faster.
Our approach to introducing adaptation is two-pronged.
First, we use simple interpolation to improve the speed of
R EPEATED S TEP I MPROVED with off-median indexes. Second, we define specialized algorithms for indexes close to
one edge of the array. In the following we focus on small indexes (i.e. less than |A|
2 ). This is without loss of generality;
large indexes work symmetrically.
• A[p+1 : 5f ] contains only elements greater than or equal
to the pivot;
• At least
2n
9
to the pivot;
elements in A[0 : 4f ] are less than or equal
elements in A[5f : |A|] are greater than or
equal to the pivot.
• At least
2n
9
Let us derive bounds for C(n) and S(n) for this algorithm. The number of comparisons in R EPEATED S TEP I M PROVED satisfies the recurrence:
C(n) ≤ C
n
9
+C
7n
9
+n+
n 8n
+
3
9
(8)
The only difference from the basic algorithm is the last
term, which is slightly smaller in this case because we don’t
revisit the middle 9th ile during partitioning. This solves to
C(n) ≤ 20n (5% better than R EPEATED S TEP).
Each M EDIAN 3 call performs at most two swaps. Due to
the layout choice the partitioning step already leaves at least
2n
9 elements less than or equal to the pivot to the pivot’s left,
and at least as many elements greater than or equal to the
pivot to its right, so in the worst case the algorithm needs
5n
to swap 5n
18 elements from the left side and 18 elements
from the right. However, the swaps don’t sum from the two
sides because one swap operation takes care of two elements
wherever possible. So the total number of swaps performed
by E XPAND PARTITION is at most 5n
18 .
S(n) ≤ S
n
9
+S
Consequently S(n) ≤
algorithm.
PEATED S TEP
7n
9
21n
2 ,
+
2n 2n 5n
+
+
3
9
18
Adaptation
5.1 Making R EPEATED S TEP I MPROVED Adaptive
Line 11 of R EPEATED S TEP I MPROVED computes the
median of the fourth 9th ile with the call Q UICKSE LECT(R EPEATED S TEP I MPROVED , A[4f : 5f ], f /2). The
last argument divides the subarray in half, computing an approximate median of the entire array. However, if k 6= |A|
2 ,
(9)
over 35% fewer than R E -
7
2016/7/27
we are more interested in a pivot close to k than an approximate median.
To make R EPEATED S TEP I MPROVED adaptive, we adopt
a simple interpolation approach: replace f /2 with kf /|A| in
the call above, idea first proposed and analyzed by Martı́nez
et al. [25] with the name “proportional-of-s”. That way, k is
proportionally represented in the fraction by which the subarray is partitioned. There is no change if k = |A|/2; as k
gets smaller the pivot computed will move left, and as k gets
larger the pivot will move right. Algorithm 8 shows the pseudocode of the resulting algorithm R EPEATED S TEPA DAP TIVE , which is the first discussed here to use k.
particular interest to us because it partitions the array asymmetrically. Consider taking the lower median of 4 in the first
step, then the median of 3 in the second step and use that
as the pivot. We do so with the help of the trivial routine
L OWER M EDIAN 4(A, a, b, c, d) (not shown), which places
the lower median of A[a], A[b], A[c], A[d] in A[b] and the
minimum in A[a]. In that case, after the two partition steps
(with group sizes 4 then 3) are finished, there are at least |A|
6
elements less than or equal to the pivot on the left, and |A|
4
elements greater than or equal to the pivot on the right. (If
the upper median of 4 were used, the numbers would be reversed.) The partitioned subarray is asymmetrically placed
at A [|A|/3 : 5|A|/12], which reduces swaps if the pivot is
also on the left.
Algorithm 9 defines R EPEATED S TEP L EFT. The call to
Q UICKSELECT uses the same interpolation approach as R E PEATED S TEPA DAPTIVE .
Algorithm 8: R EPEATED S TEPA DAPTIVE
Data: A, k
Result: Pivot p, 0 ≤ p < |A|; A partitioned at A[p]
1 if |A| < 9 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 f ← |A|/9;
5 for i ← 3f through 6f − 1 do
6
M EDIAN 3(A, i − 3f, i, i + 3f );
7 end
8 for i ← 4f through 5f − 1 do
9
M EDIAN 3(A, i − f, i, i + f );
10 end
11 Q UICKSELECT (R EPEATED S TEPA DAPTIVE ,
A[4f : 5f ], kf /|A|);
12 return
E XPAND PARTITION(A, 4f, 4f + kf /|A|, 5f − 1);
5.2
Algorithm 9: R EPEATED S TEP L EFT
Data: A, k
Result: Pivot p, 0 ≤ p < |A|; A partitioned at A[p]
1 if |A| < 12 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 f ← |A|/4;
5 for i ← 0 through f − 1 do
6
L OWER M EDIAN 4(A, i, i + f, i + 2f, i + 3f );
7 end
0
8 f ← f /3;
0
9 for i ← f through f + f − 1 do
0
10
M EDIAN 3(A, i, i + f , i + 2f 0 );
11 end
12 Q UICKSELECT (R EPEATED S TEP L EFT ,
A[f : f + f 0 ], kf 0 /|A|);
13 return
E XPAND PARTITION(A, f, f + kf 0 /|A|, f + f 0 − 1);
Searching for Indexes Left of Median
Performance of R EPEATED S TEPA DAPTIVE is better than
R EPEATED S TEP I MPROVED when k is off of |A|
2 . However,
as k gets further away from the middle, the number of swaps
performed by R EPEATED S TEPA DAPTIVE increases. This is
because the 9th ile used for recursion is still in the middle of
the array, so it is suboptimally placed; E XPAND PARTITION
needs to shift elements from that middle 9th ile toward the
pivot, which is off the middle.
When k is far away from the median, a better way to improve performance is to partition the array asymmetrically.
If the index sought is to the left of the median, placing the
subarray of medians on the left and ensuring a larger margin
on the right is better. That way the search has less swapping
to do and is more likely to eliminate a large subarray. To get
asymmetric margins, we modify the group size and placement as follows.
Let us start by noting that R EPEATED S TEPA DAPTIVE
does not necessarily need to use groups of 3. It may use
different group sizes for the first and the repeated step, as
long as linearity can be proven. An even group size is of
The number of comparisons in the worst case obeys:
C(n) ≤ C
n
5n
4n 3n
+C
+
+
+n
12
6
4
12
(10)
with the same term positions as in Eq. 4 and the note that
lower/upper median of 4 requires at most 4 comparisons. For
this asymmetric algorithm, C(n) is also linear and bounded
by 21n, even if inefficiently used for all values of k.
R EPEATED S TEP L EFT is best used with indexes k below
A’s middle (by a threshold we established experimentally).
A similarly-defined routine R EPEATED S TEP R IGHT is to be
used with indexes above the middle of A. For sought indexes
close to the median, R EPEATED S TEPA DAPTIVE offers best
results.
8
2016/7/27
5.3
Searching for Indexes Far Left of Median
the subarray passed down to Q UICKSELECT is placed at
A [2|A|/3 : 3|A|/4].
Consider now searching for very small indexes. In that case
the motivation is high to not find a pivot to the left of
the index, because that would only eliminate very little of
the input. Martı́nez et al. discuss this risk for their related
proportional-of-3 search strategy [25]. So we design an algorithm to find p such that p ≥ k.
A different schema is useful here. R EPEATED S TEP FAR L EFT (Algorithm 10) takes the lower median of 4 in the first
step, like R EPEATED S TEP L EFT. Then, instead of the median of 3 in the second step, it takes the minimum of 3.
5.4
How to tie these specialized algorithms together? We defined
Q UICKSELECT as a higher-order function parameterized by
the partitioning primitive used. That made it convenient to
discuss and analyze a variety of partitioning algorithms using the same Q UICKSELECT skeleton.
At this point, in order to implement full adaptation,
we need to dynamically choose the partitioning algorithm from among R EPEATED S TEPA DAPTIVE, R EPEATED S TEP L EFT, R EPEATED S TEP R IGHT, R EPEATED S TEP FAR L EFT, and R EPEATED S TEP FAR R IGHT. A good place to decide strategy is the Q UICKSELECT routine itself, which has
access to all information needed and controls iteration. Before each partitioning step, a partitioning algorithm is chosen depending on the relationship between |A| and k. After
partitioning, both A and k are modified and a new decision
is made, until the search is over. Q UICKSELECTA DAPTIVE
(Algorithm 11) embodies this idea.
1.0
is dictated by the lower margin of R E The cutoff 12.0
PEATED S TEP FAR L EFT and R EPEATED S TEP FAR R IGHT ; as
mentioned, in those cases we never want the pivot to be between the sought index and the closest array edge. The cutoff
7.0
16.0 has been chosen experimentally.
Algorithm 10: R EPEATED S TEP FAR L EFT
Data: A, k
Result: Pivot p, 0 ≤ p < |A|; A partitioned at A[p]
1 if |A| < 12 then
2
return H OARE PARTITION(A, |A|/2);
3 end
4 f ← |A|/4;
5 for i ← f through 2f − 1 do
6
L OWER M EDIAN 4(A, i − f, i, i + f, i + 2f );
7 end
0
8 f ← f /3;
0
9 for i ← f through f + f − 1 do
0
10
if A[i + f ] < A[i] then S WAP(A[i + f 0 ], A[i]);
11
if A[i + 2f 0 ] < A[i] then S WAP(A[i + 2f 0 ], A[i]);
12 end
13 Q UICKSELECT (R EPEATED S TEP FAR L EFT ,
A[f : f + f 0 ], kf 0 /|A|);
14 return
E XPAND PARTITION(A, f, f + kf 0 /|A|, f + f 0 − 1);
6.
n
5n
4n 3n
+C
+
+
+n
12
6
4
12
Engineering Considerations
The deterministic partitioning algorithms presented so far
do not “cheat” (by estimating from samples, making assumptions about data distribution etc): they look at the entire data set and provide guaranteed margins. The absolute
cost of a search is front-loaded, during the first few iterations of Q UICKSELECTA DAPTIVE, when the data examined
is largest. For large data sets with usual distribution statistics,
a good sampling strategy that finds a good pivot without inspecting the entire data may make a large speed difference.
It is therefore very tempting for the optimization-minded engineer to do away with rigorous determinism and try a few
shots at sampling when the data size is large.
However, we resisted the lure of switching to one of the
classic heuristics. When data is large, indeed the reward of
a cheap good pivot is excellent, but conversely the cost of a
bad pivot choice is high, and all heuristics may compute an
arbitrarily bad pivot.
We adopted an engineering solution that offers good
speed improvements on average for only moderate slowdown in the worst case and a minor code change. Each
search starts optimistically with a “sampling mode” flag on.
As long as sampling is enabled, the first median step in all
R EPEATED S TEP algorithm variations is skipped. For example, in Algorithm 7, the loop at line 5 is not executed if sampling is on. This makes the pivot to be evaluated only from
the mid tertile; the first and last tertile are not inspected at
In that case, there are at least |A|
12 elements less than
or equal to the pivot on the left, and 3|A|
elements
8
greater than or equal to the pivot on the right. (If the upper median was chosen, the numbers would be reversed.)
The subarray passed down to Q UICKSELECT is placed at
A [|A|/4 : |A|/3]. C(n) obeys:
C(n) ≤ C
Choosing Strategy Dynamically in Q UICKSELECT
(11)
with the same term positions as in Eq. 4. For this algorithm,
C(n) is also linear and bounded by 21n. However, if R E |A|
PEATED S TEP FAR L EFT is used only when k ≤ 12 , then the
pivot is guaranteed on the right so the bound gets better because 3|A|
8 or more elements are discarded.
The algorithm R EPEATED S TEP FAR R IGHT (not shown)
is defined symmetrically with R EPEATED S TEP FAR L EFT: it
takes the upper medians of 4 and places them in the 3rd
quartile, then within that quartile it takes the maximum of 3.
|A|
The margins are 3|A|
8 on the left and 12 on the right, and
9
2016/7/27
7.
Algorithm 11: Q UICKSELECTA DAPTIVE
Data: A, k with 0 ≤ k < |A|
Result: Puts kth smallest element of A in A[k] and
partitions A around it.
1 while true do
2
r ← real(k)/real(|A|);
3
if |A| < 12 then
4
p ← H OARE PARTITION(A, |A|/2);
5
else if r ≤ 7.0/16.0 then
6
if r ≤ 1.0/12.0 then
7
p ← R EPEATED S TEP FAR L EFT(A, k);
8
else
9
p ← R EPEATED S TEP L EFT(A, k);
10
end
11
else if r ≥ 1.0 − 7.0/16.0 then
12
if r ≥ 1.0 − 1.0/12.0 then
13
p ← R EPEATED S TEP FAR R IGHT(A, k);
14
else
15
p ← R EPEATED S TEP R IGHT(A, k);
16
end
17
else
18
p ← R EPEATED S TEP I MPROVED(A, k);
19
end
20
if p = k then return;
21
if p > k then
22
A ← A[0 : p];
23
else
24
i ← k − p − 1;
25
A ← A[p + 1 : |A|];
26
end
27 end
Experiments and Results
We tested performance on a HP ENVY 810qe Desktop (Intel
Core i7 3.6 GHz) machine with 32GB RAM, against arrays
of various sizes of 64-bit floating point data (results on 32and 64-bit integrals are similar). The code and tests were
written in the D programming language [1]. The compiler
used was the LLVM-based ldc [22] version 0.14.0 (we obtained similar results and rankings with dmd 2.069). We ran
each experiment five times (with different random data sets)
and took the median of the five timings obtained. The data
sets used are:
• Uniformly-distributed random floating-point numbers in
the range −5 000 000 through 5 000 000.
• Numbers generated as above, but sorted prior to testing.
• Normally-distributed data with mean 0 and standard de-
viation 3 333 333.
Many pivot choosing strategies have been proposed in the
literature. We chose the most competitive and in prevalent
industrial use, also mentioned in the introduction:
• M EDIAN 3: Median of 3 deterministic;
• M EDIAN 3R ANDOMIZED: Median of 3 randomized;
• N INTHER: Tukey’s ninther deterministic; and
• N INTHER R ANDOMIZED: Ninther randomized.
The deterministic heuristics are particularly relevant on
almost sorted data, where they perform very well. We tried
but did not plot two more heuristics: single random pivot
performed poorly, and medians of 5 did not perform better
than the ninther.
7.1 Finding the Median
First we focus on finding the median. Fig. 1 plots the absolute run times of the algorithms tested for finding the median
in arrays of uniformly-distributed floating point numbers. To
avoid clutter, we did not plot M EDIAN 3 and N INTHER because their behavior on random data is similar to their respective randomized versions.
To give a clearer image of the relative performance of
the algorithms, Fig. 2 shows the relative speedup of Q UICK SELECTA DAPTIVE over all other algorithms. Fig. 3 shows
speedup only over heuristics-based algorithms.
Ranking is similar on Gaussian-distributed floating point
numbers. Fig. 4 shows the speedup of Q UICKSELECTA DAP TIVE over the others, and Fig. 5 shows its speedup only over
heuristics-based algorithms.
One interesting aspect is that R EPEATED S TEPA DAPTIVE
performed better than R EPEATED S TEP even on this task
(searching for the median), where adaptation is not ostensibly involved. The reason is that adaptation still helps, just not
in the first iteration: after one or more iterations in Q UICK SELECTA DAPTIVE ’s main loop, the index, the array limits,
all, so this mechanism effectively implements sampling. The
rest of the function proceeds the same. The pivot now den
grades in quality from within 2n
9 to within 9 of either end of
the array. However, a quality check step in Q UICKSELECTA DAPTIVE turns the sampling mode flag off as soon as one
partitioning step did not meet its margin guarantees. Therefore, there may be at most one bad pivot choice per search,
and even that is not arbitrarily bad. The time complexity is
not affected. This sampling strategy makes Q UICKSELECTA DAPTIVE adaptive to both k and the data distribution and
takes its performance from passable to compelling.
The implementation also contains a few microoptimizations not shown in the pseudocode: some
computations are memoized and sentinel techniques
are used in some loops to reduce overheads. The performance gains from most of these optimizations are minor but
measurable.
10
2016/7/27
BFPRTBASELINE
R EPEATED S TEP BASELINE
M EDIAN O F 3R ANDOMIZED
A DAPTIVE N O S AMPLING
Time [ms]
4,000
BFPRTI MPROVED
R EPEATED S TEP I MPROVED
N INTHER R ANDOMIZED
Q UICKSELECTA DAPTIVE
2,000
0
10
25
40
55
70
Input size [millions of elements]
85
100
Figure 1. Run times of various selection algorithms (64-bit uniformly distributed floating point numbers)
BFPRTI MPROVED
R EPEATED S TEP I MPROVED
N INTHER R ANDOMIZED
M EDIAN O F 3
M EDIAN O F 3R ANDOMIZED
Speedup of Q UICKSELECTA DAPTIVE [%]
Speedup of Q UICKSELECTA DAPTIVE [%]
BFPRTBASELINE
R EPEATED S TEP BASELINE
M EDIAN O F 3R ANDOMIZED
300
200
100
0
40
30
20
10
0
10
10
25
40
55
70
85
Input size [millions of elements]
100
N INTHER
N INTHER R ANDOMIZED
25
40
55
70
85
Input size [millions of elements]
100
Figure 3. Detail: Relative speedup of Q UICKSELECTA DAPTIVE over heuristics-based algorithms (64-bit uniformly distributed floating point numbers).
Figure 2. Relative speedup of Q UICKSELECTA DAPTIVE
over other algorithms (64-bit uniformly distributed floating
point numbers).
DIAN O F 3R ANDOMIZED and N INTHER R ANDOMIZED also
perform well on this highly structured input.
and their relationship changes. For example, if the first partition pass finds a good approximate pivot, the second pass
will process a k close to one edge of the array, where adaptive algorithms have an advantage.
On sorted data, as expected, the deterministic heuristics (M EDIAN O F 3 and N INTHER) perform the best by
a large margin. This is because they are engineered to
find the exact median if the data is already sorted, so
the cost of Q UICKSELECT is essentially that of one pass
through H OARE PARTITION. The randomized versions M E -
7.2 Finding Indexes Other than the Median
This experiment measures how adaptation helps in searching for indexes other than the median. This time, |A| is fixed
at 1 000 000 and the ratio between k and |A| varies between
5% and 95%. Fig. 8 plots absolute run times for the algorithms tested. Fig. 9 shows the relative speedup of Q UICK SELECTA DAPTIVE over the other algorithms. The improvements over non-adaptive M EDIAN O F M EDIANS derivatives
are strong for all percentiles tested, especially toward the
11
2016/7/27
BFPRTI MPROVED
R EPEATED S TEP I MPROVED
N INTHER R ANDOMIZED
300
200
100
0
10
25
40
55
70
85
Input size [millions of elements]
100
100
0
N INTHER
N INTHER R ANDOMIZED
30
20
10
25
40
55
70
85
Input size [millions of elements]
25
40
55
70
85
Input size [millions of elements]
M EDIAN O F 3
M EDIAN O F 3R ANDOMIZED
Speedup of Q UICK S ELECTA DAPTIVE [%]
Speedup of Q UICK S ELECTA DAPTIVE [%]
200
100
Figure 6. Relative speedup (slowdown if negative) of
Q UICKSELECTA DAPTIVE over other algorithms (64-bit
sorted floating point numbers).
40
10
BFPRTI MPROVED
R EPEATED S TEP I MPROVED
N INTHER
N INTHER R ANDOMIZED
300
10
Figure 4. Relative speedup of Q UICKSELECTA DAPTIVE
over other algorithms (64-bit Gaussian-distributed floating
point numbers).
M EDIAN O F 3
M EDIAN O F 3R ANDOMIZED
BFPRTBASELINE
R EPEATED S TEP BASELINE
M EDIAN O F 3
M EDIAN O F 3R ANDOMIZED
Speedup of Q UICK S ELECTA DAPTIVE [%]
Speedup of Q UICK S ELECTA DAPTIVE [%]
BFPRTBASELINE
R EPEATED S TEP BASELINE
M EDIAN OF 3 R ANDOMIZED
100
Figure 5. Detail: Relative speedup of Q UICKSELECTA DAPTIVE over heuristics-based algorithms (64-bit
Gaussian-distributed floating point numbers).
N INTHER
N INTHER R ANDOMIZED
20
0
−20
−40
10
25
40
55
70
85
Input size [millions of elements]
100
Figure 7. Detail: Relative speedup (slowdown if negative) of Q UICKSELECTA DAPTIVE over heuristics-based algorithms (64-bit sorted floating point numbers).
end of the scale, as expected. Fig. 10 gives only the relative speedup over heuristics-based strategies, again showing
12
2016/7/27
good improvements. Results on Gaussian-distributed data
(not shown) are similar.
8.
Summary and Future Work
We refine the definition of M EDIAN O F M EDIANS and its
variant R EPEATED S TEP to improve their speed. Measurements show that the best deterministic algorithm Q UICK SELECTA DAPTIVE outperforms state-of-the-art baselines in
the task of partitioning an array around a specified quantile.
There are several future directions we plan to follow. One
is to devise principled methods of choosing the group sizes
and layouts for each of the specialized algorithms, depending on the index searched. Currently we use three algorithms
chosen experimentally, but perhaps a single algorithm with
the proper parameterization might replace them all. We also
plan to investigate the use of fast deterministic selection algorithms for sorting.
Acknowledgments
Thanks to Timon Gehr, Ivan Kazmenko, Scott Meyers, Todd
Millstein who reviewed drafts of this document. Many participants to the D Programming Languge forum have submitted partition of 5 primitives (http://forum.dlang.org/
post/[email protected]), of which the
one by Teppo Niinimäki was used in this paper.
13
2016/7/27
BFPRTBASELINE
M EDIAN O F 3R ANDOMIZED
BFPRTI MPROVED
N INTHER R ANDOMIZED
R EPEATED S TEP BASELINE
A DAPTIVE N O S AMPLING
R EPEATED S TEP I MPROVED
R EPEATED S TEPA DAPTIVE
Time [ms]
40
30
20
10
10
20
30
40
50
60
Index sought relative to size [%]
70
80
90
Figure 8. Run times of various selection algorithms (64-bit uniformly distributed floating point numbers)
BFPRTI MPROVED
R EPEATED S TEP I MPROVED
N INTHER
N INTHER R ANDOMIZED
400
300
200
100
0
M EDIAN O F 3
M EDIAN O F 3R ANDOMIZED
Speedup of Q UICKSELECTA DAPTIVE [%]
Speedup of Q UICKSELECTA DAPTIVE [%]
BFPRTBASELINE
R EPEATED S TEP BASELINE
M EDIAN O F 3
M EDIAN O F 3R ANDOMIZED
60
40
20
0
−20
10
10
20 30 40 50 60 70 80
Index sought relative to size [%]
90
N INTHER
N INTHER R ANDOMIZED
20
30 40 50 60 70
Percentile sought [%]
80
90
Figure 10. Detail: Relative speedup of Q UICKSELECTA DAPTIVE over heuristics-based algorithms for varying order statistics.
Figure 9. Relative speedup of Q UICKSELECTA DAPTIVE
over other algorithms for varying order statistics.
14
2016/7/27
References
[17] P. Kirschenhofer and C. M. Inez. Analysis of Hoare’s find
algorithm with median–of–three partition, 1997.
[1] A. Alexandrescu. The D Programming Language. AddisonWesley Professional, 1st edition, 2010. ISBN 0321635361,
9780321635365.
[18] P. Kirschenhofer and H. Prodinger. Comparisons in Hoare’s
find algorithm.
Combinatorics, Probability and Computing, 7:111–120, 3 1998.
ISSN 1469-2163.
doi:
null. URL http://journals.cambridge.org/article_
S0963548397003325.
[2] S. Battiato, D. Cantone, D. Catalano, G. Cincotti, and
M. Hofri. An efficient algorithm for the approximate median
selection problem. In Algorithms and Complexity, pages 226–
238. Springer, 2000.
[19] D. E. Knuth. Mathematical analysis of algorithms. Technical
report, DTIC Document, 1971.
[3] J. L. Bentley and M. D. McIlroy. Engineering a sort function.
Software: Practice and Experience, 23(11):1249–1265, 1993.
[20] D. E. Knuth. The Art of Computer Programming, Volume 3:
(2Nd Ed.) Sorting and Searching. Addison Wesley Longman
Publishing Co., Inc., Redwood City, CA, USA, 1998. ISBN
0-201-89685-0.
[4] M. Blum, R. W. Floyd, V. Pratt, R. L. Rivest, and R. E. Tarjan.
Time bounds for selection. J. Comput. Syst. Sci., 7(4):448–
461, Aug. 1973. ISSN 0022-0000. doi: 10.1016/S00220000(73)80033-9.
[21] T. W. Lai and D. Wood. SWAT 88: 1st Scandinavian Workshop
on Algorithm Theory Halmstad, Sweden, July 5–8, 1988 Proceedings, chapter Implicit selection, pages 14–23. Springer
Berlin Heidelberg, Berlin, Heidelberg, 1988. ISBN 978-3540-39288-0. doi: 10.1007/3-540-19487-8 2.
[5] S. Carlsson and M. Sundström. Algorithms and Computations: 6th International Symposium, ISAAC ’95 Cairns, Australia, December 4–6, 1995 Proceedings, chapter Linear-time
in-place selection in less than 3n comparisons, pages 244–
253. Springer Berlin Heidelberg, Berlin, Heidelberg, 1995.
ISBN 978-3-540-47766-2. doi: 10.1007/BFb0015429.
[22] LDC Team. LDC—the LLVM-based D compiler, 2016.
URL https://github.com/ldc-developers/ldc. [Online; accessed 8-March-2016].
[6] K. Chen and A. Dumitrescu. Select with groups of 3 or
4. In Algorithms and Data Structures: 14th International
Symposium, WADS, 2015.
[23] H. M. Mahmoud. Distributional analysis of swaps in quick
select. Theor. Comput. Sci., 411(16-18):1763–1769, Mar.
2010. ISSN 0304-3975. doi: 10.1016/j.tcs.2010.01.029.
[7] D. Coetzee. An efficient implementation of Blum, Floyd,
Pratt, Rivest, and Tarjan’s worst-case linear selection algorithm. 2004. URL http://moonflare.com/code/
select/select.pdf.
[24] C. Martı́nez, D. Panario, and A. Viola. Mathematics and Computer Science II: Algorithms, Trees, Combinatorics and Probabilities, chapter Analysis of Quickfind with Small Subfiles,
pages 329–340. Birkhäuser Basel, Basel, 2002. ISBN 978-30348-8211-8. doi: 10.1007/978-3-0348-8211-8 20.
[8] T. H. Cormen, C. Stein, R. L. Rivest, and C. E. Leiserson.
Introduction to Algorithms. McGraw-Hill Higher Education,
2nd edition, 2001. ISBN 0070131511.
[25] C. Martı́nez, D. Panario, and A. Viola. Adaptive sampling
strategies for quickselect. ACM Trans. Algorithms, 6(3):53:1–
53:45, July 2010. ISSN 1549-6325. doi: 10.1145/1798596.
1798606.
[9] J. Daligault and C. Martı́nez. On the variance of quickselect.
In Proceedings of the Meeting on Analytic Algorithmics and
Combinatorics, pages 205–210. Society for Industrial and Applied Mathematics, 2006.
[26] M. D. McIlroy. A killer adversary for quicksort. Softw., Pract.
Exper., 29(4):341–344, 1999.
[10] K. Dinkel and A. Zizzi. Fast median finding on digital images.
In AIAA Regional Student Paper Conference, 2012.
[27] D. R. Musser. Introspective sorting and selection algorithms.
SoftwarePractice & Experience, 27(8):983–993, 1997.
[11] D. Dor and U. Zwick. Median selection requires (2 + )N
comparisons. SIAM J. Discret. Math., 14(3):312–325, Mar.
2001. ISSN 0895-4801. doi: 10.1137/S0895480199353895.
[28] H. Prodinger. Analysis of quickselect. RAPPORT DE
RECHERCHE-INSTITUT NATIONAL DE RECHERCHE EN
INFORMATIQUE ET EN AUTOMATIQUE, pages 95–98,
1996.
[12] R. W. Floyd and R. L. Rivest. Expected time bounds for
selection. Commun. ACM, 18(3):165–172, Mar. 1975. ISSN
0001-0782. doi: 10.1145/360680.360691.
[29] H. Saraogi. Median of medians algorithm. 2013. URL
http://himangi774.blogspot.com/2013/09/medianof-medians.html.
[13] M. Fouz, M. Kufleitner, B. Manthey, and N. Z. Jahromi. On
smoothed analysis of quicksort and Hoare’s find. Algorithmica, 62(3-4):879–905, 2012.
[30] U. Schning.
Mastering the master theorem.
Bulletin of the EATCS, 71:165–166, 2000.
URL
http://dblp.uni-trier.de/db/journals/eatcs/
eatcs71.html#Schoning00.
[14] R. Griffin and K. A. Redish. Remark on algorithm 347
[m1]: An efficient algorithm for sorting with minimal storage.
Commun. ACM, 13(1):54–, Jan. 1970. ISSN 0001-0782. doi:
10.1145/361953.361993.
[31] R. Sedgewick and K. Wayne. Algorithms. Pearson Education, 2011. ISBN 9780132762564. URL https://books.
google.com/books?id=idUdqdDXqnAC.
[15] C. A. R. Hoare. Algorithm 65: Find. Commun. ACM, 4(7):
321–322, July 1961. ISSN 0001-0782. doi: 10.1145/366622.
366647.
[32] R. C. Singleton. Algorithm 347: An efficient algorithm for
sorting with minimal storage [m1]. Commun. ACM, 12(3):
185–186, Mar. 1969. ISSN 0001-0782. doi: 10.1145/362875.
362901.
[16] C. A. R. Hoare. Algorithm 64: Quicksort. Commun. ACM, 4
(7):321–, July 1961. ISSN 0001-0782. doi: 10.1145/366622.
366644.
15
2016/7/27
[33] J. Tukey. The ninther, a technique for low-effort robust (resistant) location in large samples. Contributions to Survey
Sampling and Applied Statistics in Honor of HO Hartley, Academic Press, New York, pages 251–258, 1978.
[34] L. Wang and X. Wang. On the worst case data sets for order
statistics. Appl. Math, 6(2):357–362, 2012.
[35] Wikipedia. Median of medians, 2016. URL https://en.
wikipedia.org/wiki/Median_of_medians. [Online; accessed 25-Feb-2016].
[36] A. C. Yao and F. F. Yao. On the average-case complexity of
selecting the k-th best. Technical report, Stanford, CA, USA,
1979.
[37] C. K. Yap. New upper bounds for selection. Commun.
ACM, 19(9):501–508, Sept. 1976. ISSN 0001-0782. doi:
10.1145/360336.360339.
16
2016/7/27
| 8cs.DS
|
arXiv:1709.06244v1 [math.AC] 19 Sep 2017
ON ASYMPTOTIC DEPTH OF INTEGRAL CLOSURE
FILTRATION AND AN APPLICATION
TONY J. PUTHENPURAKAL
Abstract. Let (A, m) be an analytically unramified formally equidimensional
Noetherian local ring with depth A ≥ 2. Let I be an m-primary ideal and set
L
n ∗
n+1 )∗ be the
I ∗ to be the integral closure of I. Set G∗ (I) =
n≥0 (I ) /(I
associated graded ring of the integral closure filtration of I. We prove that
depth G∗ (I n ) ≥ 2 for all n ≫ 0. As an application we prove that if A is also an
excellent normal domain containing an algebraically closed field isomorphic to
A/m then there exists s0 such that for all s ≥ s0 and J is an integrally closed
ideal strictly containing (ms )∗ then we have a strict inequality µ(J) < µ((ms )∗ )
(here µ(J) is the number of minimal generators of J).
1. introduction
1.1. Setup: Throughout (A, m) is a Noetherian local ring and I is an m-primary
ideal Throughout we consider multiplicative I-stable filtration of ideals
F = {In }n≥0 , i.e., we assume
(1)
(2)
(3)
(4)
I0 = A and In+1 ⊆ In for all n ≥ 0
I1 6= A and I ⊆ I1 .
In Im ⊆ In+m for all n, m ≥ 0.
IIn = In+1 for all n ≫ 0
L
Let R(I) = n≥0 I n tn be the Rees-algebra of A with respect to I. Set R(F ) =
L
n
n≥0 In t be the Rees-algebra of F . By our assumption R(F ) is a finite extension
L
of R(I) in A[t]. Also set G(F ) = n≥0 In /In+1 the associated graded ring of F .
If G = {(In )∗ } the integral closure filtration of I then we set G∗ (I) = G(G). We
b is reduced) then the integral closure
note that if A is analytically unramified (i.e., A
filtration of I is a multiplicative I-stable filtration of ideals.
The following result is definitely known. We sketch a proof for the convenience
of the reader.
Proposition 1.2. (with hypotheses as in 1.1). Further assume that A is analytically unramified. We have In ⊆ (I n )∗ for all n ≥ 1.
Date: September 20, 2017.
1991 Mathematics Subject Classification. Primary 13A30, 13D45 ; Secondary 13H10, 13H15.
Key words and phrases. integral closure filtration, m-full ideals, asymptotic depth.
1
2
TONY J. PUTHENPURAKAL
Sketch of a proof. The integral closure of R(I) in A[t] is the Rees-algebra of the
integral closure filtration of I. By our assumption R(F ) is a finite extension of
R(I) in A[t]. The result follows.
1.3. As A is reduced we have that depth A > 0. It follows that the ideals (I n )∗
are all Ratliff-Rush. So depth G∗ (I) > 0. Recall a Noetherian local ring A is
b
b
called formally equidimensional if dim A/P
= dim A for all minimal primes of A,
the completion of A with respect to m. In the literature, the local rings with this
property are also called quasi-unmixed. Our main result is
Theorem 1.4. Let (A, m) be an analytically unramified formally equidimensional
Noetherian local ring with depth A ≥ 2. Let I be an m-primary ideal and set I ∗ to
be the integral closure of I. Then depth G∗ (I n ) ≥ 2 for all n ≫ 0.
Remark 1.5. We note that there exists n0 such that (I n )∗ is a normal ideal for
all n ≥ n0 . If A is Cohen-Macaulay then it follows from a result of Huckaba and
Huneke, [3, 3.1], that depth G(((I n )∗ )k ) ≥ 2 for all k ≫ 0. Even in this case it does
not immediately follow that depth G((I n )∗ ) ≥ 2 for all n ≫ 0.
Application: Let (A, m) be a Noetherian local ring with d = dim A ≥ 1 and
I an m-primary ideal of A. The notion of m-full ideals was introduced by D.
Rees and J. Watanabe ([9]) and they proved the “Rees property” for m-full ideals,
namely, if I is m-full ideal and J is an ideal containing I, then µ(J) ≤ µ(I), where
µ(I) = ℓA (I/mI) is the minimal number of generators of I. Also, they proved that
integrally closed ideals are m-full if A is normal.
fn , the Ratliff-Rush closure of mn , is m-full ([1,
Suppose depth A > 0. Then m
Proposition 2.2]). Thus mn is m-full for sufficiently large n.
Sometimes we need stronger property for µ(I) and we will call it “Strong Rees
property” (SRP for short).
Definition (Strong Rees Property). Let I be an m-primary ideal of A. Then
we say that I satisfies the strong Rees property if for every ideal J ) I, we have
µ(J) < µ(I).
In a recent paper the author with K. Watanabe and K. Yoshida proved the
following result, [8, 3.2]:
Theorem 1.6. Let (A, m) be a Noetherian local ring. Assume that depth A ≥ 2 and
1
HM
(G) has finite length, where G = G(m) = ⊕n≥0 mn /mn+1 . If mℓ is Ratliff-Rush
closed, then mℓ has SRP.
For integrally closed ideals we have the following weaker form of SRP.
Definition (SRP ∗ for integrally closed ideals ). Let I be an m-primary integrally closed ideal of A. Then we say that I satisfies the *-strong Rees property
for integrally closed ideals if for every ideal J ) I with J integrally closed, we have
µ(J) < µ(I).
INTEGRAL CLOSURE FILTRATION
3
Our application of Theorem 1.4 is the following:
Theorem 1.7. Let (A, m) be an excellent normal domain of dimension d ≥ 2 with
algebraically closed residue field. Then there exists s0 such that if j ≥ s0 and I is
an integrally closed ideal with I ! (mj )∗ ; then µ(I) < µ((mj )∗ ).
Technique used to prove our results:
L
The main technique for this paper is to consider LF = n≥0 A/In+1 . This is a
module over R(I), see 3.1. When F is the I-adic filtration then this technique was
developed in [6],[7]. Although this module is not finitely generated as a R-module
it has the following good properties:
Proposition 1.8. Assume A is an analytically unramified local ring. Let F be a
multiplicative I-stable filtration over an m-primary ideal I. Let M be the maximal
homogeneous ideal of R(I). Then
0
(1) HM
(LF )n = 0 for all n ≫ 0.
0
◦
◦
(2) Set HM
(LF )n = In+1
/I n+1 . Then In+1
⊆ (In+1 )∗ .
(3) If depth A ≥ 2 then we have
1
(a) HM
(LF )n = 0 for all n ≫ 0.
1
(b) ℓ(HM
(LF )n ) < ∞ for all n ∈ Z.
1
1
(c) For all n < 0 we have ℓ(HM
(LF )n−1 ) ≤ ℓ(HM
(LF )n ).
We now describe in brief the contents of this paper. In section two we describe
a few preliminaries that we need. In section three we describe our construction LF
and prove Proposition 1.8. In section four we prove Theorem 1.4. Finally in section
five we prove Theorem 1.7.
2. preliminaries
In this section we collect a few preliminaries that we need.
2.1. Some properties of multiplicative I-stable filtrations:
(1) Let F = {In }n≥0 be a multiplicative I-stable filtration and let x ∈ I. Set
B = A/(x), I = I/(x) and let F = {(In + (x))/(x)}n≥0 be the quotient
filtration of F . Then F is a multiplicative I-stable filtration.
(2) If the residue field k is infinite then we may choose x which is both I and
F -superficial.
(3) Set F <l> = {Inl }n≥0 . Then F <l> is a multiplicative I l -stable filtration.
2.2. Flat Base Change: In our paper we do many flat changes of rings. The
general set up we consider is as follows:
Let φ : (A, m) → (A′ , m′ ) be a flat local ring homomorphism with mA′ = m′ . Set
I ′ = IA′ and if N is an A-module set N ′ = N ⊗ A′ . Set k = A/m and k ′ = A′ /m′ .
Let F = {In }n≥0 be a multiplicative I-stable filtration. Then F ′ = {In′ }n≥0 is a
multiplicative I ′ -stable filtration.
4
TONY J. PUTHENPURAKAL
Properties preserved during our flat base-changes:
(1) ℓA (N ) = ℓA′ (N ′ ).
(2) dim M = dim M ′ and grade(K, M ) = grade(KA′ , M ′ ) for any ideal K of A.
(3) depth G(F ) = depth G(F ′ ).
Specific flat Base-changes:
(a) A′ = A[X]S where S = A[X] \ mA[X]. The maximal ideal of A′ is n = mA′ .
The residue field of A′ is k ′ = k(X). Notice that k ′ is infinite.
b the completion of A with respect to the maximal ideal.
(b) A′ = A
′
(c) A = A[X1 , . . . , Xn ]S where S = A[X1 , . . . , Xn ]\mA[X1 , . . . , Xn ]. The maximal
ideal of A′ is n = mA′ . The residue field of A′ is l = k(X1 , . . . , Xn ). Notice
that if I is integrally closed then I ′ is also integrally closed.
(d) When depth A ≥ 2 and A is formally equidimensional, then Ciupercă [2, Corollary 2] shows that in A′ (for n = µ(I) in (ii)) there exists a superficial element
y ∈ I ′ such that the A′ /(y) ideal J = I ′ /(y) is also integrally closed (if I is
integrally closed). We call A′ general extension of A with respect to I. Also if
Pn
I = (a1 , . . . , an ) then we can choose y to be i=1 ai Xi . We call the latter a
generic element of I ′ . We note that Ciupercă observes that y is also superficial
for the integral closure filtration of I, see [2, Section 2.5].
We will need the following result in the proof of Theorem 1.4. It is definitely
known to experts. We give a proof for the convenience of the reader
Lemma 2.3. Let (A, m) be an analytically unramified, formally equidimensional
Noetherian local ring with depth A ≥ 2. Let K be an m-primaryideal of A. Then
there exists a flat local homomorphism φ : A → A′ with mA′ = m′ and that there
exists y ∈ K ′ such that
(1) the ring A′ /(y) is analytically unramified.
(2) y is superficial for both the K ′ -adic filtration and the integral closure filtration
of K ′ .
(3) the ideal (K ′ )∗ A′ /(y) is integrally closed in A′ /(y).
(4) (K n )∗ A′ = ((K ′ )n )∗ for all n ≥ 1.
Proof. We construct A′ in two steps. First we complete A. Notice by assumption
b is equidimensional and reduced. Let K = (a1 , . . . , am ). Then K A
b is generated
A
Pm
b Set A′ = A[X
b 1 , . . . , Xm ] b
by images of ai in A.
i=1 ai Xi .
b A[X1 ,...,Xm ] and y =
m
Then as noted in 2.2(d) we get that (2), (3) hold true. Also trivially (4) holds.
b 1 , . . . , Xm ]. Let P be a minimal prime ideal of (y) in A′ .
(1) Set T = A[X
Then height P = 1. There exists a prime ideal Q of T containing (y) such that
height Q = 1 and QA′ = P . As depth A′ ≥ 2 we get that there exists ai such that
b 1 , . . . , Xm−1 ]. The
ai ∈
/ Q. Say am ∈
/ Q. Then (T /(y))Q is a localization of A[X
′
latter ring is reduced. So (T /(y))Q is reduced. As (A /(y))P is a localization of
INTEGRAL CLOSURE FILTRATION
5
(T /(y))Q it is also reduced. Thus A′ /(y) is reduced. As it is also excellent we get
that A′ /(y) is analytically unramified.
3. LF
In this section we define LF and study few of its properties. We also prove prove
Proposition 1.8 which is the most important technical result of this paper.
3.1. Let (A, m) be a Noetherian local ring and let I be an m-primary ideal of A. Let
L
F = {In }n≥0 be an I-stable multiplicative filtration of A. Set LF = n≥0 A/In+1 .
Note that we have a short exact sequence of R(I)-modules
(1)
0 → R(F ) → A[t] → LF (−1) → 0.
It follows that LF is a R(I)-module. Note that it is not finitely generated R(I)module.
Throughout we take local cohomology of R(I)-modules with respect to M =
m ⊕ R(I)+ , the maximal homogeneous ideal of R(I). Recall a graded R(I)-module
V is said to be ∗-Artinian if every descending chain of graded submodules of V
stablizes. For instance if E is a finitely generated graded R(I)-module then for all
i
i ≥ 0 the R(I)-modules HM
(E) is ∗-Artinian.
3.2. We will use the following well-known result regarding *-Artinian modules quite
often:
Let V be a *-Artinian R(I)-module. Then
(a) Vn = 0 for all n ≫ 0
(b) If ψ : V (−1) → V is a monomorphism then V = 0.
(c) If φ : V → V (−1) is a monomorphism then V = 0.
We begin with the following easy result
i
(LF ) is
Lemma 3.3. Let depth A = c ≥ 1. Then for i = 0, . . . , c − 1 we have HM
i
∗-Artinian. In particular HM
(LF )n = 0 for n ≫ 0.
Proof. We take local cohomology of short exact sequence (1) in 3.1 with respect to
M. We note that if a1 , . . . , ac is an A-regular sequence then a1 t0 , . . . , ac t0 ∈ M0 is
i
an A[t]-regular sequence. This yields that HM
(LF (−1)) to be a R(I)-sub-module
i+1
of HM
(R(F )) for i = 0, . . . , c − 1. The result follows.
As an easy consequence we get the following
Corollary 3.4. Assume A is analytically unramified. Set L∗I =
Then H 0 (L∗I ) = 0.
L
n≥0
A/(I n+1 )∗ .
0
Proof. We note that depth A > 0. So by 3.3 we get that HM
(LF ) is 8-Artinian.
n ∗
Also note that (I ) is Ratliff-Rush for all n ≥ 1. In particular we have that G∗ (I)
has positive depth, in particular H 0 (G∗ (I)) = 0.
6
TONY J. PUTHENPURAKAL
We have a short exact sequence 0 → G∗ (I) → L∗I → L∗I (−1) → 0. Taking local
0
0
cohomology with respect to M we get an inclusion HM
(L∗I ) ֒→ HM
(L∗I )(−1). The
result follows from 3.2.
We now give
Proof of Proposition 1.8. (1) This follows from 3.3.
L
(2) As A is analytically unramified the module R∗ (I)/R(F ) = n≥1 (In )∗ /In
is a finintely generated R(I)-module. We have a short exact sequence of R(I)modules
0 → W (1) → LF → L∗I → 0.
Taking local cohomology with respect to M we get
0
0
0
(L∗I )
(LF ) → HM
0 → HM
(W (1)) → HM
0
But HM
((L∗I ) = 0 by 3.4. So we have
◦
In+1
(I n+1 )∗
0
0
= HM
(LF )n = HM
(W (1))n ⊆ W (1)n =
.
In+1
I n+1
The result follows.
(3) (a) This follows from 3.3.
3(b) Let x be both F and I-superficial. Then note we have a short exact sequence
for all n ≥ 0
0→
A αn A
A
(In+1 : x)
→
→
→ 0,
−−→
In
In
In+1
(In+1 , x)
where αn (a + In ) = xa + In+1 . Set B = A/(x). Let F be the quotient filtration of
F . Thus we have a short exact sequence of R(I)-modules
xt
0 → W → LF (−1) −→ LF → LF → 0.
As x is F -superficial we get that W has finite length. So we have a short exact
sequence
(†)
0
1
1
HM
(LF )n → HM
(LF )n−1 → HM
(LF )n
1
0
Say HM
(LF )n = 0 for all n ≥ s. We note that HM
(LF )n has finite length for
1
all n ≥ 0 and is zero for n < 0. Evaluating (†) at n = s we get that HM
(LF )s−1
has finite length. Evaluating (†) at n = s − 1 yields an exact sequence
0
1
1
HM
(LF )s−1 → HM
(LF )s−2 → HM
(LF )s−1
1
1
It follows that HM
(LF )s−2 has finite length. Iterating we get that HM
(LF )n has
finite length for all n ≤ s.
0
3(c) This follows from (†) as HM
(LF )n is zero for n < 0.
INTEGRAL CLOSURE FILTRATION
7
4. Proof of Theorem 1.4
In this section we give proof of Theorem 1.4. We also prove an additional result
which will be useful in the proof of Theorem 1.7.
Proof of Theorem 1.4. Let F = {(I n )∗ } be the integral closure filtration of I. We
note that for r ≥ 1, F <r> is the integral closure of filtration of I r . Observe that
<r>
<r>
LF (−1)
= LF
(−1)
Also note that M<r> is the maximal homogenous ideal of R(I r ). As local coho1
mology commutes with the Veronese functor and as HM
(LF )j = 0 for j ≫ 0 it
1
F <r>
follows that there exists r0 such that for r ≥ r0 we have HM
(−1))j = 0
<r> (L
for all j ≥ 1.
Fix r ≥ r0 . Set K = I r . We do the construction as in 2.3. So we may assume
that there exists y ∈ K which is superficial for both the K-adic filtration and
the integral closure filtration of K. Furthermore A/(y) is analytically unramified.
Let G = {(K n )∗ } be the integral closure filtration of K and let G be it’s quotient
filtration in A/(y). Note we have an short exact sequence of R(K)-modules
yt
0 → LG (−1) −→ LG → LG → 0.
This induces a long exact sequence in cohomology. Note H 0 (LG ) = 0. Furthermore
by construction H 1 (LG )j = 0 for j ≥ 0. So for all n ∈ Z we have an exact sequence
0 → H 0 (LG )n → H 1 (LG )n−1
It follows that H 0 (LG )n = 0 for n ≥ 1. We note that as A/(y) is analytically
unramified we have by 1.8(2)
H 0 (LG )0 = U/V,
where V = K ∗ and U ⊆ (K ∗ )∗ , by 1.8(2).
But by our construction K ∗ is integrally closed in A/(y). So H 0 (LG ) = 0. We also
have a short exact sequence
0 → G(G) → LG → LG (−1) → 0.
Taking cohomology we get depth G(G) ≥ 1. By Sally descent, [4, 2.2] we get
depth G(G) ≥ 2. The result follows.
The following result is needed in the proof of Theorem 1.7.
Lemma 4.1. Let (A, m) be an analytically unramified formally equidimensional
Noetherian local ring with depth A ≥ 2. Let I be an m-primary ideal. Set F =
{(I n )∗ }n≥0 to be the integral closure filtration of I. Then H 1 (LF ) has finite length.
Proof. By Theorem 1.4 there exists r ≥ 1 such that G∗ (I r ) has depth atleast r.
Let G be the integral closure filtration of I r . So we have an exact sequence
0 → G(G) → LG → LG (−1) → 0.
8
TONY J. PUTHENPURAKAL
Taking cohomology we get an injective map H 1 (LG ) → H 1 (LG )(−1). By 3.2 we
get H 1 (LG ) = 0.
As observed before
(LF (−1))<r> = LG (−1).
As local cohomology commutes with the Veronese functor we get that
H 1 (LF )−r−1 = 0. The result now follows from 1.8(3).
5. Proof of Theorem 1.7
Our proof is quite similar in spirit to proof of Theorem 1.6. However it is different
in some places. So we are forced to give the whole proof.
b is also a
Proof of Theorem 1.7. We first note that as A is excellent and normal A
normal domain of dimension d ≥ 2, see [5, 32.2]. In particular A is analytically
unramified, formally equidimensional with depth A ≥ 2. So all the techniques
developed in the earlier sections are applicable.
The Rees algebra of the integral closure filtration of m is a finite module over
R(m). In particular there exists an s0 such that for all j ≥ s0 we have m(mj )∗ =
(mj+1 )∗ . Choose j ≥ s0 .
Step-1 We we may assume λ(I/(mj )∗ ) = 1.
By [10, 2.1] there exists a chain of integrally closed ideals
I = I0 ! I1 ! I2 ! · · · ! Is1 ! Is = (mj )∗ ,
with λ(Ii /Ii+1 ) = 1 for i = 0, . . . , s − 1. As Ij are integrally closed and A is normal
they are m-full. So we have
µ(I) = µ(I0 ) ≤ µ(I1 ) ≤ µ(I2 ) ≤ · · · ≤ µ(Is−1 ) ≤ µ((mj )∗ ).
Thus it suffices to prove µ(Is−1 ) < µ((mj )∗ ). Thus we may assume λ(I/(mj )∗ ) = 1.
Step-2 A consequence of assuming λ(I/(mj )∗ ) = 1.
Note mI ⊆ (mj )∗ . Also note that
(mj+1 )∗ = m(mj )∗ ⊆ mI.
We have an exact sequence
0→
(mj )∗
I
I
(mj )∗ ∩ mI
→
→
→
→ 0.
j+1
∗
j+1
∗
j
∗
(m )
(m )
mI
(m ) + mI
We note that (mj )∗ + mI = (mj )∗ and (mj )∗ ∩ mI = mI. Furthermore m(mj )∗ =
(mj+1 )∗ . Thus the above exact sequence yields an exact sequence
0→
mI
(mj )∗
I
I
→
→
→ j → 0.
(mj+1 )∗
m(mj )∗
mI
m
Thus to prove our result it suffices to show c = λ(mI/(mj+1 )∗ ) ≥ 2.
To show c ≥ 2 we have to consider some modules over the Rees algebra R(m) =
A[mt] of m.
INTEGRAL CLOSURE FILTRATION
9
Step 3: Some modules over the Rees algebra R(m).
We first list some finitely generated R(m)-modules that are pertinent to our result.
L
(1) R∗ = i≥0 (mi )∗ ti is a finitely generated R(m)-module.
L
(2) R∗≥j = i≥j (mi )∗ ti is an R(m)-submodule of R∗ .
L
So R∗≥j (+j) = i≥0 (mj+i )∗ ti is a finitely generated R(m)-module.
(3) The filtration
I ⊇ mI ⊇ m2 I ⊇ · · · ⊇ mi I ⊇ mi+1 I ⊇ · · ·
L
is m-stable (note that it is not multiplicative). So E = i≥0 mi Iti is a finitely
generated R(m)-module.
(4) By our assumption on j we get that R∗≥j (+j) ⊆ E. So C = E/R∗≥j (+j) is a
fintely generated R-module. Note
C=
I
mI
m2 I
mi I
⊕ j+1 ∗ t ⊕ j+2 ∗ t2 ⊕ · · · j+i ∗ ti ⊕ · · · ,
j
∗
(m )
(m )
(m )
(m )
=
mI
m2 I
mi I
I
2
⊕
t
⊕
t
⊕
·
·
·
ti ⊕ · · · .
(mj )∗
m(mj )∗
m2 (mj )∗
mi (mj )∗
We now list two NOT finitely generated modules over R(m) which is of interest to
us:
L
(1) L = i≥0 A/(mi+1 )∗ . By 3.1 we get that L is a R(m)-module.
(2) We also have an exact sequence
0 → E → A[t] → V → 0,
where V =
L
i≥0
A/mi I. It follows that V is a R-module.
Step 4: Some local cohomology computations.
Throughout we compute local cohomology with respect to M, the maximal homogeneous ideal of R(m). As depth A ≥ 2 there exists x, y ∈ m such that x, y is an
A-regular sequence. We note that xt0 , yt0 ∈ M0 is an A[t]-regular sequence. So
H 0 (A[t]) = H 1 (A[t]) = 0. It follows that H 0 (V ) = H 1 (E).
(1) Claim-1: H 0 (V ) has finite length as an A-module.
It is well-known that for all l ≥ 0 we have H l (E)n = 0 for n ≫ 0. As H 0 (V ) =
H 1 (E) we get that H 0 (V )n = 0 for n ≫ 0.
Also as H 0 (V ) ⊆ V we have H 0 (V )n = 0 for n < 0. Furthermore as H 0 (V )n ⊆
Vn it follows that H 0 (V )n has finite length for all n. So Claim-1 follows.
(2) By 3.4 we get H 0 (L) = 0. By 4.1 we get that H 1 (L) have finite length. Set
L
L(−1)≥j = i≥j A/mi ti . We have an exact sequence
0 → L(−1)≥j → L(−1) → W → 0,
where W has finite length. It follows that H 0 (L(−1)≥j ) = 0 and H 1 (L(−1)≥j ) has
finite length.
10
TONY J. PUTHENPURAKAL
Step 5: c = λ(mI/(mj+1 )∗ ) > 1.
L
i+j ∗
Set D = (L(−1)≥j )(+n) =
) . Note we have a short exact sei≥0 A/(m
quence of R(m)-modules
0 → C → D → V → 0.
Let x be A-superficial with respect to m. Then it is also superficial with respect to to
the integral closure filtration of m; see [2, section 2.5]. Notice that ((mn+1 )∗ : x) =
(mn )∗ for all n ≥ 1.
Set u = xt ∈ R1 . Let fuD : D(−1) → D be multiplication by u. Notice ker fuD =
0. It follows that ker(fuC ) = 0.
Notice ker u is D-regular. So u is C-regular.
Suppose if possible c = 1. We note that C0 = I/(mj )∗ and C1 = mI/m(mj )∗ .
u
→ C1 is an isomorphism. So we get
Note As ker(fuC ) = 0 we get the map C0 −
mI = xI + m(mj )∗ . It follows that for all i ≥ 1 we have
mi+1 I = xmi I + mi (mj )∗ .
Thus we have that
u
0 →→ C(−1) −
→ C → I/mj → 0.
(here I/mj is concentrated in degree zero). It follows that dimR C = 1.
As dimR C = 1 we get that H 1 (C) is NOT finitely generated as an A-module.
However the exact sequence 0 → C → D → V → 0 yields an exact sequence
H 0 (V ) → H 1 (C) → H 1 (D) which implies that H 1 (C) has finite length.
Thus our assumption c = 1 is not possible. So c ≥ 2 and this proves our result.
References
[1] J. Asadollahi and T. J. Puthenpurakal, An analogue of a theorem due to Levin and Vasconcelos, Commutative algebra and algebraic geometry, 9–15, Contemp. Math., 390, Amer.
Math. Soc., Providence, RI, 2005.
[2] C. Ciupercă, Integral closure and generic elements, J. Algebra 328 (2011), 122-131
[3] S. Huckaba and C. Huneke, Normal ideals in regular rings, J. Reine Angew. Math. 510 (1999),
63-82.
[4] S. Huckaba and T. Marley, Hilbert coefficients and the depths of associated graded rings, J.
London Math. Soc. 56 (2) (1997) 64-76.
[5] H. Matsumura, Commutative ring theory, Cambridge Studies in Advanced Mathematics,
vol. 8, Cambridge University Press, Cambridge, 1986.
[6] T. J. Puthenpurakal, Ratliff-Rush filtration, regularity and depth of higher associated graded
modules. I, J. Pure Appl. Algebra 208 (2007), no.1, 159–176.
[7] T. J. Puthenpurakal, Ratliff-Rush filtration, regularity and depth of higher associated graded
modules. Part II, J. Pure Appl. Algebra 221 (2017), no. 3, 611–631.
[8] T. J. Puthenpurakal, K. Watanabe and K. Yoshida, The strong Rees property of powers of
the maximal ideal and Takahashi-Dao’s question. Preprint. arXiv:1708.06090
[9] J. Watanabe, m-full ideals, Nagoya Math. J. 106 (1987), 101–111.
[10] K.-i. Watanabe, Chains of integrally closed ideals, Commutative algebra (Grenoble/Lyon,
2001), 353–358, Contemp. Math., 331, Amer. Math. Soc., Providence, RI, 2003.
INTEGRAL CLOSURE FILTRATION
11
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai
400 076, India
E-mail address: [email protected]
| 0math.AC
|
Published as a conference paper at ICLR 2018
S HAKE D ROP REGULARIZATION
Yoshihiro Yamada, Masakazu Iwamura and Koichi Kise
Graduate School of Engineering Osaka Prefecture University
1-1 Gakuen-cho, Naka-ku Sakai-shi, Osaka
[email protected],{masa,kise}@cs.osakafu-u.ac.jp
arXiv:1802.02375v1 [cs.CV] 7 Feb 2018
A BSTRACT
This paper proposes a powerful regularization method named ShakeDrop regularization. ShakeDrop is inspired by Shake-Shake regularization that decreases error
rates by disturbing learning. While Shake-Shake can be applied to only ResNeXt
which has multiple branches, ShakeDrop can be applied to not only ResNeXt but
also ResNet, Wide ResNet and PyramidNet in a memory efficient way. Important
and interesting feature of ShakeDrop is that it strongly disturbs learning by multiplying even a negative factor to the output of a convolutional layer in the forward
training pass. The effectiveness of ShakeDrop is confirmed by experiments on
CIFAR-10/100 and Tiny ImageNet datasets.
1
I NTRODUCTION
Recent advances in generic object recognition have been brought by deep neural networks. After
ResNet (He et al., 2016) opened the door to very deep CNNs of over a hundred layers by introducing the residual block, its improvements such as Wide ResNet (Zagoruyko & Komodakis, 2016),
PyramdNet (Han et al., 2017a;b) and ResNeXt (Xie et al., 2017) have broken the records of lowest error rates. On the other hand, in learning, they often suffer from problems such as vanishing
gradients. Hence, regularization methods help to learn and boost the performance of such base network architectures. Stochastic Depth (ResDrop) (Huang et al., 2016) and Shake-Shake (Gastaldi,
2017) are known to be effective regularization methods for ResNet and its improvements. Among
them, Shake-Shake applied to ResNeXt is the one achieving the lowest error rates on CIFAR-10/100
datasets (Gastaldi, 2017).
Shake-Shake, however, has following two drawbacks. (1) Shake-Shake can be applied to only multibranch architectures (i.e., ResNeXt). (2) Shake-Shake is not memory efficient. Both drawbacks
come from the same root. That is, Shake-Shake requires two branches of residual blocks to apply. If
it is true, it is not difficult to conceive its solution: a similar disturbance to Shake-Shake on a single
residual block. It is, however, not trivial to realize it.
The current paper addresses the problem of realizing a similar disturbance to Shake-Shake on a
single residual block, and proposes a powerful regularization method, named ShakeDrop regularization. While the proposed ShakeDrop is inspired by Shake-Shake, the mechanism of disturbing
learning is completely different. ShakeDrop disturbs learning more strongly by multiplying even
a negative factor to the output of a convolutional layer in the forward training pass. In addition, a
different factor from the forward pass is multiplied in the backward training pass. As a byproduct,
however, learning process gets unstable. Our solution to this problem is to stabilize the learning
process by employing ResDrop in a different usage from the usual. Based on experiments using various base network architectures, we reveal the condition that the proposed ShakeDrop successfully
works.
2
2.1
E XISTING M ETHODS R EQUIRED TO I NTRODUCE THE P ROPOSED M ETHOD
D EEP N ETWORK ARCHITECTURES
ResNet (He et al., 2016) opened the door to very deep CNNs of over a hundred layers by introducing
the residual block, given as
G(x) = x + F (x),
(1)
1
Published as a conference paper at ICLR 2018
Conv
Conv
Conv
Conv
Train
(Forward)
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
×𝜶
×𝜷
Train
(Forward)
× 𝑬(𝜶)
Test
Train
(Backward)
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
×𝜶
× (𝟏 − 𝜶)
×𝜷
× (𝟏 − 𝜷)
× 𝟎. 𝟓
× 𝟎. 𝟓
Test
Train
(Backward)
(b) Shake-Shake (ResNeXt + Shake-Shake) (Gastaldi,
2017), in which some processing layers omitted for
conciseness.
(a) ResNeXt (Xie et al., 2017), in which some processing layers omitted for conciseness.
Conv
Conv
Conv
Train
(Forward)
Test
Train
(Backward)
Conv
Conv
Conv
Conv
Conv
Conv
Conv
Conv
× (𝒃𝒍 + 𝜶 − 𝒃𝒍 𝜶)
× (𝒃𝒍 + 𝜷 − 𝒃𝒍 𝜷)
× 𝑬(𝒃𝒍 + 𝜶 − 𝒃𝒍 𝜶)
Train
(Forward)
(c) PyramidNet + “1-branch Shake.” “1-branch
Shake” is an intermediate regularization method to the
proposed ShakeDrop.
Train
(Backward)
Test
(d) PyramidNet + ShakeDrop. ShakeDrop is the proposed regularization method.
Figure 1: Network architectures.
where x and G(x) are the input and output of the residual block, respectively, and F (x) is the output
of the residual branch on the residual block.
While ResNet realized deep networks, it is imperfect. He et al. (2016) pointed out that high error
rates obtained by the 1202-layer ResNet on the CIFAR datasets (Krizhevsky, 2009) is caused by
overfitting. Veit et al. (2016) experimentally showed that some residual blocks whose channels
greatly increase cause high error rates.
PyramidNet (Han et al., 2017a;b) overcame the problem of ResNet by gradually increasing channels on each residual block. It has almost same residual block as Eqn. (1). It successfully realized
deep CNNs of up to 272 layers and achieved the lowest error rates among the vanilla residual networks on the CIFAR datasets.
2.2
N ON -D EEP NETWORK ARCHITECTURES
Wide ResNet (Zagoruyko & Komodakis, 2016) improved error rates by simply increasing channels
of ResNet. While it has almost same residual block as Eqn. (1), it has wider and shallower architecture than ResNet. Jastrzebski et al. (2017) experimentally showed that like PyramidNet, Wide
ResNet overcomes the problem of ResNet.
ResNeXt (Xie et al., 2017) achieved lower error rates than Wide ResNet on almost same number of
parameters. The basic architecture of ResNeXt is given as
G(x) = x + F1 (x) + F2 (x),
(2)
where F1 (x) and F2 (x) are the outputs of the residual branches as shown in Figure 1(a). The number
of residual branches is not limited to 2, and the number is the most important factor to control the
result.
2.3
R EGULARIZATION METHODS
Stochastic Depth (Huang et al., 2016) is a regularization method which overcame problems of
ResNet such as vanishing gradients. It makes the network apparently shallow in learning by dropping residual blocks stochastically selected. On the lth residual block from input layer, the Stochastic Depth process is given as
G(x) = x + bl F (x),
(3)
2
Published as a conference paper at ICLR 2018
where bl ∈ {0, 1} is a Bernoulli random variable with the probability of pl . Huang et al. (2016)
tested two rules for determining pl : uniform rule and linear decay rule. The uniform rule using a
constant for all pl did not work well. The linear decay rule which defines pl as
l
(1 − pL ),
L
where L is the number of all layers and pL is the initial parameter, worked well.
pl = 1 −
(4)
It was originally proposed for improving ResNet. It was also applied to PyramidNet; the combination of PyramidNet and Stochastic Depth was named PyramidDrop (Yamada et al., 2016).
Shake-Shake (Gastaldi, 2017) is a powerful regularization method for improving ResNeXt architectures. It is given as
G(x) = x + αF1 (x) + (1 − α)F2 (x),
(5)
where α is a random coefficient given as α ∈ [0, 1]. As shown in Figure 1(b), calculation of the
backward pass is disturbed by another random coefficient β ∈ [0, 1] instead of α. As a result,
Shake-Shake decreased error rates than ResNeXt.
3
3.1
P ROPOSED M ETHOD
I NTERPRETATION OF S HAKE -S HAKE REGULARIZATION AND ITS DRAWBACKS
We give an intuitive interpretation of the forward pass of Shake-Shake regularization. To the best of
our knowledge, it has not been given yet, while the phenomenon in the backward pass is experimentally investigated by Gastaldi (2017). As shown in Eqn. (5) (and also in Fig. 1(b)), in the forward
pass, Shake-Shake interpolates the outputs of two residual branches (i.e., F1 (x) and F2 (x)) with a
random variable α that controls the degree of interpolation. As DeVries & Taylor (2017a) demonstrated that interpolation of two data in the feature space can synthesize reasonable augmented data,
the interpolation of two residual blocks of Shake-Shake in the forward pass can be interpreted as
synthesizing data. Use of a random variable α generates many different augmented data. On the
other hand, in the backward pass, a different random variable β is used to disturb learning to make
the network learnable long time. Gastaldi (2017) demonstrated how the difference between α and β
affects.
The regularization mechanism of Shake-Shake relies on two or more residual branches, so that it
can be applied only to 2-branch networks architectures. In addition, 2-branch network architectures
consume more memory than 1-branch network architectures. One may think the number of learnable
parameters of ResNeXt can be kept in 1-branch and 2-branch network architectures by controlling
its cardinality and the number of channels (filters). For example, a 1-branch network (e.g., ResNeXt
1-64d) and its corresponding 2-branch network (e.g., ResNeXt 2-40d) have almost same number of
learnable parameters. However, even so, it increases memory consumption due to the overhead to
keep the inputs of residual blocks and so on. By comparing ResNeXt 1-64d and 2-40d, the latter
requires more memory than the former by 8% in theory (for one layer) and by 11% in measured
values (for 152 layers).
3.2
S IMILAR REGULARIZATION TO
ARCHITECTURES
S HAKE -S HAKE ON 1- BRANCH NETWORK
In order to realize a similar regularization to Shake-Shake on 1-branch network architectures, in the
forward pass, we need a mechanism, different from interpolation, to synthesize augmented data in
the feature space. Actually, DeVries & Taylor (2017a) demonstrated not only interpolation but also
noise addition in the feature space works well. Hence, following Shake-Shake, we apply random
perturbation, using α, to the output of a residual branch (i.e., F (x) of Eqn. (1)). In the backward
pass, we can use the same way as Shake-Shake for 1-branch network architectures.
We call the regularization method mentioned above 1-branch Shake. It is given as
G(x) = x + αF (x),
(6)
where α is a coefficient that disturbs learning in the forward pass. As shown in Figure 1(c), β is a
coefficient similar to the one of Shake-Shake and used in the backward pass. The 1-branch Shake is
3
Published as a conference paper at ICLR 2018
expected to realize powerful generalization like Shake-Shake regularization. However, by applying
it to 110-layer PyramidNet with α ∈ [0, 1] and β ∈ [0, 1] following Shake-Shake, the result on the
CIFAR-100 dataset was hopelessly bad (i.e., an error rate of 77.99%).
3.3
S TABILIZING LEARNING WITH INTRODUCTION OF MECHANISM OF R ES D ROP
Failure of 1-branch Shake is caused by too strong perturbation. However, weakening the perturbation would also weaken the effect of regularization. Thus, we need a trick to promote learning under
strong perturbation.
Our idea is to use the mechanism of ResDrop for solving the issue. As is written in Sec.2.3, in
the original form, ResDrop promotes learning by dropping some residual blocks. In other words,
it makes a network apparently shallow in learning. In our situation, however, the original usage of
ResDrop does not contribute because a shallower network to which 1-branch Shake is applied would
also suffer from strong perturbation. Thus, we use the mechanism of ResDrop as a probabilistic
switch of two network architectures: the original network (e.g., PyramidNet) and the one to which
1-branch Shake is applied (e.g., PyramidNet + 1-branch Shake). By mixing up two networks, the
following effects are expected.
1. When the original network (e.g., PyramidNet) is selected, learning is correctly promoted.
2. When the network with strong perturbation (e.g., PyramidNet + 1-branch Shake) is selected, learning is disturbed.
To achieve a good performance, two networks should be in a good balance.
The proposed ShakeDrop is given as
G(x) = x + (bl + α − bl α)F (x),
(7)
where bl is a Bernoulli random variable following the linear decay rule. We call the new method
“ShakeDrop.” According to the value of bl , Eqn. (7) is deformed as
x + F (x),
if bl = 1
G(x) =
(8)
x + αF (x), otherwise (i.e., if bl = 0).
In the case that the base network is PyramidNet, ShakeDrop is equivalent to PyramidNet if bl = 1,
and it is equivalent to “PyramidNet + 1-branch Shake” if bl = 0. We found that the linear decay rule
of ResDrop (Eqn. (4)) works well, while we also tested the uniform rule which did not work well.
As shown in Figure 1(d), similar to Shake-Shake, β is used in the backward pass instead of α. It is
noteworthy that regardless of the value of β, the weights of the network are updated. For example,
let us consider the case of β = 0. In such a case, the weights of the residual blocks selected to
perturbate (i.e., the layers with bl = 0) are not updated. However, the inputs to their succeeding
residual blocks are perturbated and the weights of the residual blocks are updated reflecting the
perturbation.
4
E XPERIMENTS
See Sec. A for implementation details.
4.1
P RELIMINARY EXPERIMENTS : SEARCH FOR PARAMETER SETTINGS OF THE PROPOSED
METHOD
The best parameter settings of the proposed method, ShapeDrop, are searched when it is applied
to PyramidNet. They include parameter ranges of α and β, and the timing these parameters are
updated. In the preliminary experiments, the CIFAR-100 dataset was used. PyramidNet had 110
layers, which consisted of a convolution layer, 54 additive pyramidal residual blocks and a fully
connected layer. The number of channels at the last residual block was 286. Table 1 shows representative parameter ranges of α and β we tested and their results. From the results, we can see how
the combinations of α and β affect. First of all, except PyramidDrop (α = 0, β = 0), only case F
(α ∈ [−1, 1], β = 0) and case H (α ∈ [−1, 1], β ∈ [0, 1]) were better than the original PyramidNet
4
Published as a conference paper at ICLR 2018
Table 1: Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with several ranges of parameters
of 4 runs at the final (300th) epoch on CIFAR-100 dataset in the “Batch” level. In some settings, it
is equivalent to PyramidNet and PyramidDrop.
A
B
C
D
E
F
G
H
α
1
0
[0, 1]
[0, 1]
[−1, 1]
[−1, 1]
[−1, 1]
[−1, 1]
β
1
0
[−1, 1]
[0, 1]
1
0
[−1, 1]
[0, 1]
Error (%)
18.01
17.74
20.61
18.27
18.68
17.28
18.26
16.22
Note
Equivalent to PyramidNet
Equivalent to PyramidDrop
Table 2: Average Top-1 errors (%) of “PyramidNet + ShakeDrop” with different levels of 4 runs at
the final (300th) epoch on CIFAR-100 dataset.
α
β
[−1, 1]
[0, 1]
Level
Batch
Image
Channel
Pixel
Error (%)
16.22
16.04
16.12
15.78
(case A: α = 1, β = 1). Among them, case H was the best. It is interesting to observe three cases E
(α ∈ [−1, 1], β = 1), F and H all of which take the same range of α. Since cases F and H were better than PyramidNet, we confirm strong perturbation is effective. However, case E was even worse
than PyramidNet. This implies that if strong perturbation is applied, the residual blocks selected to
perturbate strongly suffer from the perturbation, so that the weights of the residual blocks should not
be fully updated (i.e., β = 1 should be avoided). On the other hand, the range of β of the best case
H was in-between cases E and F. This implies that the values of β should be well balanced.
The best strategy to update the parameters (referred as scaling coefficient) among “Batch,” “Image,”
“Channel” and “Pixel” are explored. “Batch” means that the same scaling coefficients are used
for all the images in the mini-batch for each residual block. “Image” means that the same scaling
coefficients are used for each image for each residual block. “Channel” means that the same scaling
coefficients are used for the each channel for each residual block. “Pixel” means that the same
scaling coefficients are used for each element for each residual block. ShakeDrop was trained using
the best parameters found in the experiment above, i.e, ranges α ∈ [−1, 1], β ∈ [0, 1]. Table 2 shows
that “Pixel” level was the best. However, it required a lot of memory. Hence, we selected “Image”
level, which was the second best, considering memory efficiency.
Through the preliminary experiments, we found that the most effective parameters for α and β are
α ∈ [−1, 1] and β ∈ [0, 1], and the effective and memory efficient level is “Image” level. We
conduct the following experiments using these parameters.
4.2
C OMPARISON WITH REGULARIZATION METHODS
The proposed ShakeDrop is compared with vanilla (without regularization), ResDrop and 1-branch
Shake in different network architectures including ResNet, PyramidNet, Wide ResNet and ResNeXt.
Tables 3 and 4 show experimental results on CIFAR-100 dataset (Krizhevsky, 2009) and Tiny ImageNet dataset 1 , respectively.
Table 3 shows that the proposed ShakeDrop can be applied to not only PyramidNet but also ResNet,
Wide ResNet and ResNeXt. However, to successfully apply it, we found that the residual blocks
have to end with batch normalization (BN) (Ioffe & Szegedy, 2015). Since without BN, the outputs
of residual blocks are not in a certain range, sometimes factors α and β could be too big. This
seems to cause diverge in learning. With this regard, EraseReLU (Dong et al., 2017) which elimi1
https://tiny-imagenet.herokuapp.com/
5
Published as a conference paper at ICLR 2018
Table 3: Top-1 errors (%) at the final (300th) epoch of ResNet and its improvements to which
different regularization methods are applied on CIFAR-100 dataset. Method names are followed
by components of their residual blocks. For ResNet and ResNeXt, in addition to the original form,
ones without ReLU unit in the end of residual blocks following EraseReLU (Dong et al., 2017) are
also examined. For Wide ResNet, ones with bath normalization added in the end of residual blocks,
referred as “w/ BN,” are also examined. “Type A” and “Type B” of ResDrop and ShakeDrop mean
that regularization unit is inserted before and after “add” unit for residual branches, respectively.
“×” means learning did not converge. ∗ indicates the result is quoted from the literature.
(a) 1-residual-branch network architectures (ResNet, ResNeXt and PyramidNet)
Methods
ResNet-110
<Conv-BN-ReLU-Conv-BN-add(ReLU)>
ResNet-164 Bottleneck
<Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add-(ReLU)>
ResNeXt-29 8-64d Bottleneck
<Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add-(ReLU)>
PyramidNet-272 α200 Bottleneck
<BN-Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add>
Regularization
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Original (%)
28.51
24.09
24.18
×
22.00
21.96
22.20
×
20.90
20.66
22.70
×
∗
16.35
15.94
71.51
14.90
EraseReLU (%)
24.93
22.88
23.80
22.68
21.96
20.35
21.60
19.89
20.25
20.28
24.00
19.90
N/A
(b) 1-residual-branch network architectures (Wide ResNet)
Methods
Wide-ResNet-28-10k
<BN-ReLU-Conv-BN-ReLU-Conv-(BN)-add>
Regularization
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Original (%)
26.49
34.19
90.73
76.87
w/ BN (%)
24.24
26.64
58.89
19.12
(c) 2-residual-branch network architectures
Methods
ResNeXt-164 2-1-40d Bottleneck
<Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add-(ReLU)>
ResNeXt-29 2-4-64d Bottleneck
<Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add-(ReLU)>
Regularization
Vanilla
ResDrop Type-A
ResDrop Type-B
Shake-Shake
ShakeDrop Type-A
ShakeDrop Type-B
Vanilla
ResDrop Type-A
ResDrop Type-B
Shake-Shake
ShakeDrop Type-A
ShakeDrop Type-B
Original (%)
23.82
21.38
21.34
22.35
×
×
21.19
21.12
19.27
19.16
×
×
EraseReLU (%)
21.75
20.44
20.21
22.51
19.98
19.83
×
20.13
19.01
18.82
20.07
18.17
nates the last ReLU is a very convenient for us because in the EraseReLU versions of ResNet and
ResNeXt, the residual blocks ends with BN. On the other hand, in the case of Wide ResNet, we
need to intentionally add BN in the end of the residual blocks. In ResNeXt, we examined two ways,
referred as “Type A” and “Type B,” to apply ResDrop and ShakeDrop. They mean the regularization
module is inserted before and after “add” module, respectively. As far as we examined, Type B was
6
Published as a conference paper at ICLR 2018
Table 4: Top-1 errors (%) at the final (300th) epoch of ResNet and its improvements to which
different regularization methods are applied on Tiny ImageNet dataset. Method names are followed
by components of their residual blocks. For ResNet and ResNeXt, in addition to the original form,
ones without ReLU unit in the end of residual blocks following EraseReLU (Dong et al., 2017) are
also examined. For Wide ResNet, ones with bath normalization added in the end of residual blocks,
referred as “w/ BN,” are also examined. “Type A” and “Type B” of ResDrop and ShakeDrop mean
that regularization unit is inserted before and after “add” unit for residual branches, respectively.
“×” means learning did not converge. ∗ indicates the result is quoted from the literature.
(a) 1-residual-branch network architectures (ResNet, ResNeXt and PyramidNet)
Methods
ResNet-110
<Conv-BN-ReLU-Conv-BN-add(ReLU)>
ResNet-164 Bottleneck
<Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add-(ReLU)>
PyramidNet-110 α270
<BN-Conv-BN-ReLU-Conv-BN-add>
PyramidNet-200 α300 Bottleneck
<BN-Conv-BN-ReLU-Conv-BN-ReLUConv-BN-add>
Regularization
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Original (%)
42.07
43.74
45.56
×
38.20
37.17
39.29
×
36.52
33.97
85.84
32.44
32.92
32.17
78.12
31.15
EraseReLU (%)
41.24
42.50
45.16
48.92
36.52
38.09
42.10
42.80
N/A
N/A
(b) 1-residual-branch network architectures (Wide ResNet)
Methods
Wide-ResNet-28-10k
<BN-ReLU-Conv-BN-ReLU-Conv-(BN)-add>
Regularization
Vanilla
ResDrop
1-branch Shake
ShakeDrop
Original (%)
99.50
99.50
98.68
91.11
w/ BN (%)
37.88
45.80
93.62
36.39
always better than Type A. We also confirmed that ShakeDrop Type B outperformed Shake-Shake
in RexNeXt. As a conclusion, the table shows that as long as the residual blocks end with BN, the
proposed ShakeDrop achieved the lowest error rates.
Table 4 shows the results on Tiny ImageNet. To conduct the experiments, it is better to use networks
tuned for ImageNet. However, except PyramidNet, we could apply the same networks tuned for
CIFAR-100 dataset were used due to time constraint. As a result, in ResNet, the lowest error rates
were achieved by vanilla EraseReLU networks. This could be caused by use of inappropriate base
networks for the task since learning ended with relatively high training loss and no regularization
methods correctly worked. However, in PyramidNet and Wide ResNet, the proposed ShakeDrop
achieved the lowest error rates. This confirms that the proposed ShakeDrop works on Tiny ImageNet
dataset.
4.3
C OMPARISON WITH STATE - OF - THE - ART METHODS
The proposed method was compared with state-of-the-arts on the CIFAR-10/100 datasets. Stateof-the-art methods introduced some techniques that can be applied to many methods in the learning
process. One is longer learning. While most of methods related to ResNet use 300-epoch scheduling
for learning like the preliminary experiment in Sec. 4.1, Shake-Shake use 1800-epoch cosine annealing, on which the initial learning rate is annealed using a cosine function without restart (Gastaldi,
2017). Another one is image preprocessing. DeVries & Taylor (2017b) and Zhong et al. (2017)
7
Published as a conference paper at ICLR 2018
showed that accuracy is improved by data augmentation which randomly fills a part of learning
images. For fair comparison with these methods, we also applied them to the proposed method.
PyramidNet (including PyramidDrop and “PyramidDrop + ShakeDrop”) had 272 layers, which consisted of a convolution layer, 90 additive pyramidal “bottleneck” blocks (He et al., 2016) and a fully
connected layer. The number of channels at the last pyramidal residual block was 864. Table 5
shows the error rates. The proposed method, “PyramidNet + ShakeDrop,” without longer learning and image preprocessing, was 3.41% on the CIFAR-10 dataset and 14.90% on the CIFAR-100
dataset. We make following comparisons with existing methods. Note that the proposed method
used less numbers of parameters than rival methods in all conditions.
(1) Fair comparison with Shake-Shake (ResNeXt + Shake-Shake) in the same condition
“PyramidNet + ShakeDrop” combined with longer training achieved error rates of 2.67% (better by
0.19%) on the CIFAR-10 dataset and 13.99% (better by 1.86%) on the CIFAR-100 dataset.
(2) Fair comparison with Cutout (ResNeXt + Shake-Shake + Cutout) in the same condition
“PyramidNet + ShakeDrop” combined with both longer learning and image preprocessing achieved
error rates of 2.31% (better by 0.25%) and 12.19% (better by 3.01%) on the CIFAR-10/100 datasets,
respectively.
(3) Comparison with the state-of-the-arts, Cutout on CIFAR-10 and Coupled Ensemble on
CIFAR-100
“PyramidNet + ShakeDrop” combined with both longer learning and image preprocessing achieved
error rates of 2.31% (better by 0.25% than Cutout) and 12.19% (better by 2.85% than Coupled
Ensemble) on the CIFAR-10/100 datasets, respectively.
5
C ONCLUSION
We proposed a new stochastic regularization method ShakeDrop which can be successfully applied to ResNet and its improvements as long as the residual blocks ends with BN. Its effectiveness
was confirmed through the experiments on CIFAR-10/100 and Tiny ImageNet datasets. ShakeDrop
achieved the state-of-the-art performance in the CIFAR-10/100 datasets, and the best results among
existing regularization methods on Tiny ImageNet dataset except some cases using inappropriate
networks.
R EFERENCES
Terrance DeVries and Graham W. Taylor. Dataset augmentation in feature space. In Proc. ICLR
Workshop, 2017a.
Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks
with cutout. arXiv preprint arXiv:1708.04552, 2017b.
Xuanyi Dong, Guoliang Kang, Kun Zhan, and Yi Yang. EraseReLU: A simple way to ease the
training of deep convolution neural networks. arXiv preprint 1709.07634, 2017.
Anuvabh Dutt, Denis Pellerin, and Georges Qunot. Coupled ensembles of neural networks. arXiv
preprint 1709.06053v1, 2017.
Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485v2, 2017.
Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proc. CVPR,
2017a.
Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. arXiv preprint
arXiv:1610.02915v4, 2017b.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing
human-level performance on ImageNet classification. In Proc. ICCV, 2015.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016.
8
Published as a conference paper at ICLR 2018
Table 5: Top-1 errors (%) at the final epoch (300th or 1800th) on the CIFAR-10/100 datasets. Representative methods and the proposed ShakeDrop applied to PyramidNet are compared. “Reg” represents regularization methods including ResDrop (RD), Shake-Shake (SS) and proposed ShakeDrop
(SD). If “Cos” is checked, 1800-epoch cosine annealing schedule (Loshchilov & Hutter, 2016) is
used following Gastaldi (2017). Otherwise, 300-epoch multi-step learning rate decay schedule is
used following each method. If “Fil” is checked, the data augmentation used in Cutout (CO) (DeVries & Taylor, 2017b) or Random Erasing (RE) (Zhong et al., 2017), which randomly fills a part
of learning images, is used. ∗ indicates the result is quoted from the literature. + indicates the result
is quoted from Gastaldi (2017). Compared to the same condition of Cutout, the state-of-the-art, the
proposed method reduced the error rate by 0.25% on CIFAR-10 and 3.01% on CIFAR-100.
Method
Reg
Cos
Fil
Coupled Ensemble
(Dutt et al., 2017)
ResNeXt
(Xie et al., 2017)
ResNeXt + Shake-Shake
(Gastaldi, 2017)
ResNeXt + Shake-Shake + Cutout
(DeVries & Taylor, 2017b)
PyramidNet
(Han et al., 2017b)
PyramidDrop
(Yamada et al., 2016)
PyramdNet + ShakeDrop
(Proposed)
X
SS
X
SS
X
CO
X
RE
X
RE
RD
RD
SD
SD
SD
SD
RE
X
X
RE
Depth
#Param
118
106
76
64
26
29
26
29
26
29
272
272
272
272
272
272
272
272
25.7M
25.1M
24.6M
24.9M
50M
75M
100M
26.2M
34.4M
26.2M
34.4M
26.2M
34.4M
26.0M
26.0M
26.0M
26.0M
26.0M
26.0M
26.0M
26.0M
CIFAR
-10 (%)
∗
2.99
∗
2.99
∗
2.92
∗
3.13
∗
2.72
∗
2.68
∗
2.73
+
3.58
∗
2.86
∗
2.56
∗
3.31
3.42
3.83
2.91
3.41
2.89
2.67
2.31
CIFAR
-100 (%)
∗
16.18
∗
15.68
∗
15.76
∗
15.95
∗
15.13
∗
15.04
∗
15.05
+
16.34
∗
15.85
∗
15.20
∗
16.35
16.66
15.94
15.48
14.90
13.85
13.99
12.19
Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Weinberger. Deep networks with stochastic depth. arXiv preprint arXiv:1603.09382v3, 2016.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. In Francis Bach and David Blei (eds.), Proceedings of the 32nd
International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning
Research, pp. 448–456, Lille, France, 07–09 Jul 2015. PMLR. URL http://proceedings.
mlr.press/v37/ioffe15.html.
Stanisaw Jastrzebski, Devansh Arpit, Nicolas Ballas, Vikas Verma, Tong Che, and Yoshua Bengio.
Residual connections encourage iterative inference. arXiv preprint arXiv:1710.04773, 2017.
Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University
of Toronto, 2009.
Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv
preprint arXiv:1608.03983, 2016.
Yurii Nesterov. A method of solving a convex programming problem with convergence rate o(1/k 2 ).
Soviet Mathematics Doklady, 27:372–376, 1983.
David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by
back-propagating errors. Nature, 323:533–536, 1986.
Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of
relatively shallow networks. Advances in Neural Information Processing Systems 29, 2016.
9
Published as a conference paper at ICLR 2018
Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proc. CVPR, 2017.
Yoshihiro Yamada, Masakazu Iwamura, and Koichi Kise. Deep pyramidal residual networks with
separated stochastic depth. arXiv preprint arXiv:1612.01230, 2016.
Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. BMVC, 2016.
Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017.
A
I MPLEMENTATION DETAILS
All implementations used in the experiments were based on the publicly available code of
fb.resnet.torch2 . ShakeDrop was trained using back-propagation by stochastic gradient descent
with an approximate version3 of Nesterov accelerated gradient (Nesterov, 1983) and momentum
method (Rumelhart et al., 1986). 4 GPUs were used for learning acceleration; due to parallel processing, even the same layer had different values of parameters bl , α and β depending on GPUs.
CIFAR-10/100 datasets
Input images of CIFAR-10/100 datasets were processed in the following manner. An original image
of 32 × 32 pixels was color-normalized, followed by horizontally flipped with a 50% probability.
Then, it was zero-padded to be 40 × 40 pixels and randomly cropped to be an image of 32 × 32
pixels.
On PyramidNet, the initial learning late was set to 0.5 for both CIFAR-10/100 datasets following
the version 2 of Han et al. (2017b), while they use 0.1 for the CIFAR-10 dataset and 0.5 for the
CIFAR-100 dataset since version 3. Other than PyramidNet, the initial learning late was set to
0.1. The initial learning rate was decayed by a factor of 0.1 at 1/2 and 3/4 of the entire learning
process (300 epochs), respectively, following Han et al. (2017b). As the filter parameters initializer,
“MSRA” (He et al., 2015) was used. In addition, a weight decay of 0.0001, a momentum of 0.9,
and a batch size of 128 were used with 4 GPUs. On PyramidDrop and “PyramidNet + ShakeDrop,”
the linear decay parameter pL = 0.5 was used following Huang et al. (2016). 1-branch Shake used
α = [0, 1], β = [0, 1]. ShakeDrop used parameters of α = [−1, 1], β = [0, 1] and “Image” level
scaling coefficient.
Tiny ImaneNet dataset
Input images of Tiny ImageNet were processed in the following manner. An original image of
64 × 64 pixels was distorted aspect ratio and randomly cropped to be an image of 56 × 56 pixels.
Then, brightness, contrast, and saturation of the image were randomly changed. After that, the
image was color-normalized by mean and standard deviation of ImageNet, followed by horizontally
flipped with a 50% probability.
Learning settings of 110-layer PyramidNet on Tiny ImageNet were almost same as on CIFAR10/100. The initial learning late was set to 0.5. The initial learning rate was decayed by a factor
of 0.1 at 1/2 and 3/4 of the entire learning process (300 epochs), respectively, following Han et al.
(2017b). As the filter parameters initializer, “MSRA” (He et al., 2015) was used. In addition, a
weight decay of 0.0001, a momentum of 0.9, and a batch size of 128 were used with 4 GPUs.
Stride size of the first convolution layer was set to 2, and stride of average Pooling layer was set to
7. Learning settings of other than 200-layer PyramidNet on Tiny ImageNet were almost same as
110-layer PyramidNet. The initial learning late was set to 0.1.
Learning settings of 200-layer PyramidNet on Tiny ImageNet were almost same as on ImageNet.
The initial learning late was set to 0.0125. The initial learning rate was decayed by a factor of 0.1
at 1/2, 3/4 and 7/8 of the entire learning process (120 epochs), respectively, following Han et al.
2
https://github.com/facebook/fb.resnet.torch
Please be aware that the implementation in sgd.lua of Torch7 with nesterov mode is an approximation of
the original Nesterov accelerated gradient and momentum method. See https://github.com/torch/
optim/issues/27.
3
10
Published as a conference paper at ICLR 2018
Vanilla
ResDrop
ShakeDrop
Vanilla
ResDrop
ShakeDrop
(b) 1800-epoch training loss.
(a) 300-epoch training loss.
Figure 2: Change of training loss of PyramidNet with no regularization method (vanilla), ResDrop
and ShakeDrop.
(2017b). As the filter parameters initializer, “MSRA” (He et al., 2015) was used. In addition, a
weight decay of 0.0001, a momentum of 0.9, and a batch size of 32 were used with 4 GPUs. Stride
size of the first convolution layer was set to 1, and Stride of Max Pooling layer was set to 1.
B
D ISCUSSION
In order to find out why the proposed method achieved the low error rates, we investigate what kind
of difference exists with the conventional methods, particularly with PyramidNet and PyramidDrop.
B.1
T RAINING LOSS
Fig. 2 shows change of training loss. Fig. 2(a) is of the 300-epoch multi-step learning rate decay
schedule with 110-layer networks. Fig. 2(b) is of 1800-epoch cosine annealing schedule with 272layer networks. ResDrop and ShakeDrop, which are probabilistic regularization methods, decreased
the training loss more slowly than vanilla PyramidNet in both figures. Even at the final epoch, the
loss of ResDrop and ShakeDrop are much larger than 0 in contrast to vanilla PyramidNet whose loss
is close to 0.
B.2
M EAN AND VARIANCE OF GRADIENTS
In addition to the training loss, we examined the mean and variance of gradients on each iteration for
each regularization method. Under the same conditions as the preliminary experiment in Sec. 4.1,
we compared PyramidNet with no regularization method (vanilla), ResDrop and ShakeDrop in the
“Batch” level. We focused on 3 residual blocks out of 54 residual blocks and named “first,” “middle” and “final.” They mean the first, 27th and final residual blocks, respectively. For each, we
checked the 2nd convolutional layer. Figs. 3 and 4 show change of mean and variance of gradients,
respectively. Although ResDrop and ShakeDrop were similar in Fig. 2, ShakeDrop took much larger
values than ResDrop in Figs. 3 and 4. Particularly, the difference was significant in “first,” and also
large in “middle.” They would indicate that relatively intensive training lasts in ShakeDrop.
11
Published as a conference paper at ICLR 2018
0.0006
first
middle
final
0.0004
0.0006
first
middle
final
0.0004
0.0006
0.0002
0.0002
0.0002
0.0000
0.0000
0.0000
0.0002
0.0002
0.0002
0.0004
0.0004
0.0004
0.0006
0.0006
0
20000
40000
60000
80000
100000
0
20000
(a) Vanilla
40000
60000
80000
100000
first
middle
final
0.0004
0.0006
0
(b) ResDrop
20000
40000
60000
80000
100000
(c) ShakeDrop
Figure 3: The average of the gradients on each iteration in PyramidNet.
0.000010
first
middle
final
0.000008
0.000010
first
middle
final
0.000008
0.000010
0.000006
0.000006
0.000006
0.000004
0.000004
0.000004
0.000002
0.000002
0.000002
0.000000
0
20000
40000
60000
(a) Vanilla
80000
100000
0.000000
0
20000
40000
60000
(b) ResDrop
80000
100000
first
middle
final
0.000008
0.000000
0
20000
40000
60000
(c) ShakeDrop
Figure 4: The variance of the gradients on each iteration in PyramidNet.
12
80000
100000
| 1cs.CV
|
1
Retrodirective Large Antenna Energy Beamforming
in Backscatter Multi-User Networks
arXiv:1802.09035v1 [cs.IT] 25 Feb 2018
Ioannis Krikidis, Senior Member, IEEE
Abstract—In this letter, we study a new technique for energy
beamforming (EB) in multi-user networks, which combines
large antenna retrodirectivity at the transmitter side with signal
backscattering at the energy receivers. The proposed technique
has low complexity and achieves EB without any active operation
at the receivers or complicated signal processing techniques at
the transmitter. Since the average harvested energy depends on
the backscattering coefficients, we investigate different reflection
policies for various design objectives. The proposed policies are
analyzed from a system level standpoint by taking into account
spatial randomness.
Index Terms—Wireless power transfer, retrodirectivity,
backscattering, massive MIMO, stochastic geometry.
I. I NTRODUCTION
W
IRELESS power transfer (WPT) through dedicated
radio-frequency (RF) radiation has been introduced as
a promising technology for energizing low-power devices [1].
It is relevant for future mobile networks, where the majority
of billion of devices will be low-powered and autonomous. To
increase the range of WPT and boost the end-to-end power
transfer efficiency, several techniques have been proposed in
the literature that incorporate new circuit designs, network
architectures and recent advances in signal processing such
as energy beamforming (EB) [2].
EB refers to systems with multiple antennas and enables
focusing the transmitted energy to the direction of the receivers. The implementation of EB requires channel state
information (CSI) at the energy transmitter (ET), which becomes a critical assumption for WPT systems, since the energy
capabilities of the energy receivers (ERs) are limited. To tackle
this problem, most of works consider a quantized feedback
and/or exploit the channel reciprocity in association with
different channel learning methods [3], [4]. In [5], the authors
introduce an alternative EB scheme that does not require
channel estimation, by exploiting the concept of large antenna
retrodirectivity. However, its implementation requires signal
transmission from the ERs, which could be costly for WPT
scenarios. On the other hand, the work in [6] studies EB in a
backscatter communication system, where CSI is acquired by
leveraging the backscatter signal that is reflected back from the
ERs. A beneficial combination of backscatter communication
and retrodirective beamforming for EB is introduced in [7];
however, a theoretical analysis is not provided.
In this letter, we study a new EB technique for multiuser networks that combines large antenna retrodirectivity with
signal backscattering. The proposed technique overcomes the
limitations of the previous works in [5], [6], and achieves
EB without CSI estimation at the ET and without signal
I. Krikidis is with the Department of Electrical and Computer Engineering,
University of Cyprus, Nicosia 1678 (E-mail: [email protected]).
transmission at the ERs. We show that the achieved energy harvesting performance depends on the backscattering reflection
coefficients. Different reflection policies that correspond to
different complexities and energy harvesting performances are
investigated. The proposed EB technique and the associated
reflection policies are analyzed from a system level point of
view by using stochastic geometry tools.
II. S YSTEM MODEL
We consider a single-cell consisting of an ET and multiple
randomly deployed ERs, where the coverage area is modeled
as a disc D with radius ρ and an exclusion zone of radius
ξ > 1. The exclusion zone captures potential safety issues i.e.,
ERs are prohibited to be close to the ET [8]. The ET employs
a large number of antennas, denoted by M , while the ERs
are equipped with single antennas. We assume that the ERs
are located according to a homogeneous Poisson point process
(PPP) Φ = {xi } with a density λ that ensures M ≫ K, where
K is the number of ERs in D (i.e., Poisson distribution with
mean λπ(ρ2 − ξ 2 ) [9]).
The ET is connected to the power grid and transmits with a
fixed power Pt ; its antenna array has (digital) retrodirectivity
properties and thus it can re-transmit a signal back along the
spatial direction of the incoming signal without any a-priori
knowledge of its point of origin [10]. On the other hand, the
ERs do not have any local power supply and they power their
functionalities through WPT by using the signal of the ET.
Specifically, the ERs operate in a backscattering mode, where
part of the received signal is reflected back by adapting the
level of antenna impedance mismatch; let 0 ≤ βi ≤ 1 be the
reflection coefficient for the i-th ER.
The large-scale path loss is assumed to be d−α
i , where
di ≥ ξ denotes the Euclidean distance between the i-th
ER and the ET, and α > 2 is the path-loss exponent; the
exclusion zone ensures a bounded path loss region with radius
ξ. We assume a Rayleigh block fading [5], [8], where the
channel between the ET and the i-th ER is an uncorrelated
Gaussian vector with elements having zero mean and unit
variance i.e., f i = [fi,1 , . . . , fi,M ]T ∼ CN (00, I M ). We
assume channel reciprocity and thus the channels for uplink
and downlink are equivalent and constant during one block, but
vary independently from block to block. It is worth noting that
the proposed scheme becomes less interesting for scenarios
with strong line-of-sight components (e.g., for Rice fading
channel, the omnidirectional EB approaches the performance
of the perfect EB as the Rice factor increases).
III. R ETRODIRECTIVE EB WITH
BACKSCATTERING
The proposed EB technique combines a retrodirective beamforming at the ET with a passive backscattering process at the
2
ERs. The EB technique consists of two phases, as described
below.
A. Backscattering phase
The first phase enforces the ERs to notify their spatial
direction to the ET by using their backscattering capabilities. More specifically, the ET generates and
√ broadcasts an
unmodulated single-tone waveform x(t) = 2Pt cos(2πfc t)
with 0 < t < τ , where fc is the carrier frequency and τ
is the duration of the waveform1 ; the power is symmetrically
divided to M antennas. Each ER operates in the backscattering
mode and reflects back a part of the received signal without
further processing. The received signal (complex baseband
representation) at the q
ET is given by
X
βk (Pt /M )d−2α
gkf k + n (t),
(1)
y (t) =
k
xk ∈Φ
PM
where gk = m=1 fk,m denotes the equivalent channel for the
downlink (sum of M complex Gaussian random variables with
unit variance), f k is the uplink channel due to the reciprocity,
and n (t) is the additive white Gaussian noise (AWGN) vector
with n (t) ∼ CN (00, σ 2I M ). Note that the path-loss exponent
is doubled as a result of the dyadic backscatter channel.
B. EB phase
In the second phase, the ET exploits large antenna retrodirectivity and uplink/downlink channel reciprocity and effectively forms coherent beams back to the ERs. More specifically, the ET performs matching filtering on y (t) to detect
the phase of the received signal; then it transmits an energy
signal with a power Pt by employing phase conjugation. The
w+
energy transmit signal (baseband) can q
be written as x e = (w
P
−2α ∗ H
H
n) /||w
w + ñ
n||, where w =
βk (Pt /M )dk gkf k
ñ
xk ∈Φ
2
n ∼ CN (00, (σ /τ )II M ) denotes the noise at the output
and ñ
of the matched filter. Due to the channel reciprocity, the
received
signal at the i-th ER can be expressed as yi =
q
Pt d−α
x
ef i +z, where z is the AWGN component. By using
i
the energy conversion law and by ignoring energy harvesting
from AWGN, the harvested energy at the i-th ER becomes
ζPt M d−3α
βi |gi |2
i
β , di ) = ζ|yi |2 = ζPt d−α
+P
Q(β
i
−2α
| {z }
βk |gk |2 +
x ∈Φ dk
{z
| k
QOM (di )
β
QΦ
RE (β ,di )
In this section, we study the impact of reflection coefficients
on the achieved performance and we investigate different
reflection strategies.
A. Distance-inversion backscattering (DIB)
In this strategy, we enforce the average harvested energy
associated with the retrodirective component to be equal for
each ER. To satisfy this fairness requirement, we assume that
the reflection coefficient is a function of the location of each
ER i.e., βi = (di /ρ)2α ; this assumption ensures that the
reflection coefficient increases with the distance and takes the
maximum value (full reflection) at the edge of the cell. In this
case, the expression in (2) can be simplified as
#
"
M |gi |2
−α
P
for σ 2 → 0,
Q(pp, di ) ≈ ζPt di 1+
|gi |2 + xk ∈Φ,i6=k |gk |2
(3)
where p = [(d1 /ρ)2α , . . . , (dK /ρ)2α ] and the average harvested energy is given by
"
#!
2
|g
|
i
P
QDIB = ζPt EΦ [d−α
i ] 1 + ME
|gi |2 + xk ∈Φ,i6=k |gk |2
Z 2π Z ρ
= ζPt
x1−a fd (x)dθdx 1 + M E[Z]
(4a)
0
ξ
M
2ζPt (ρ2−α − ξ 2−α )
1
+
,
(4b)
=
(2 − α)(ρ2 − ξ 2 )
λπ(ρ2 − ξ 2 )
{z
}
|
Λ(ξ,ρ)
where fd (x) = 1/π(ρ2 − ξ 2 ) denotes the probability density function (PDF) of each point in the disc D, and (4b)
is based on the expected value of Beta distribution with
shape parameters M and (E[K] − 1)M , respectively, i.e.,
Z ∼ Beta(M, [λπ(ρ2 − ξ 2 ) − 1]M ).
B. Full backscattering (FB)
Mσ2
Pt τ
,
}
(2)
where ζ denotes the RF-to-DC energy conversion efficiency2.
Proof: The proof follows similar arguments with [5,
Lemma 3.1] and is based on the asymptotic massive
multiple-input multiple-output (MIMO) expressions [11] i.e.,
1
1
1 H
f 4
f 2
1, M
f k f i → 0 (with k 6= i),
M ||f i || → M +1, M ||f i || →
1
σ2
1
σ2
1 H
H
2
f
n 2
n
n
f
||f
,
and
ñ
→
0,
ñ
||
→
k
M k
M
τ
M ||ñ || → τ .
1 The
The harvested energy consists of two main components i)
an omnidirectional component QOM which only depends on
the location of the ER, and b) a retrodirective component QΦ
RE
that depends on the reflection coefficients βi of the ERs.
IV. AVERAGE HARVESTED ENERGY
proposed scheme requires the orthogonality between single-tone
waveform transmission and retrodirectivity at the ET. An appropriate design of
the parameters τ , ξ as well as a time delay between reception/reflection at the
ERs ensures that the duration of the waveform is shorter than the round-trip
delays.
2 A linear power channel is sufficient to demonstrate the proposed EB
technique [3], [4]; wireless power nonlinearities are beyond the scope of this
letter.
The FB reflection strategy refers to systems without further
intelligence, where the ERs fully reflect back their received
signals i.e., βi = 1 (∀ xi ∈ Φ). This scheme does not require
any coordination between the ERs and therefore provides
implementation simplicity. For the average harvested energy,
we state the following theorem.
Theorem 1. The average harvested energy achieved by the
FB scheme, is given by
QFB = Λ(ξ, ρ) + QFB (ξ, λ).
(5)
Proof: The first term is associated with the omnidirectional component of the harvested energy and is similar to the
first term of the DIB scheme (see (4b)), QFB (ξ, λ) refers to
the retrodirective component and is given in an integral form
in the Appendix.
For the special cases of highly dense/sparse networks with
σ 2 → 0, we introduce the following statement.
3
Statement 1. For σ 2 → 0, the average harvested energy for
highly dense (λ → ∞) and sparse networks (λ → 0) converges
to the one achieved by the non-backscattering (βi = 0) and
beamforming (perfect CSI) schemes, respectively, i.e.,
QFB → Λ(ξ, ρ)
for λ → ∞, [Non-backscattering], (6)
QFB → M Λ(ξ, ρ) for λ → 0, [Beamforming].
C. Binary backscattering schemes
(7)
In this strategy, the ERs operate either in a fullbackscattering mode or in a non-backscattering mode i.e.,
βi ∈ {0, 1}. Through this binary reflection mechanism, we
can control the harvested energy related to the retrodirective
component and provide design flexibility. The selection of the
operation mode is driven by two criteria, which are given in
the following discussion.
1) Distance-based binary backscattering (DBB): The first
criterion takes into account the spatial location of the ERs;
ERs that are located at a distance higher than ∆ from the ET,
fully reflect the received signal (βi = 1), otherwise remain
inactive (βi = 0). In this case, the average harvested energy
can be written as
QDBB = ǫΛ(ξ, ∆) + (1 − ǫ) Λ(∆, ρ) + QFB (∆, λ) , (8)
where the first and second terms refer to the location range
(ξ, ∆) and (∆, ρ), respectively, and ǫ = (ξ 2 − ∆2 )/(ρ2 − ξ 2 ).
The design parameter ∆ can be used as a means to provide an
energy harvesting balance among the ERs. More specifically,
the ERs that are close to the ET (di ≤ ∆) harvest sufficient
energy from the omnidirectional component and facilitate the
energy harvesting of the distant ERs by keeping silent. We can
design ∆ based on the max-min fairness that maximizes the
minimum average harvested energy and is given as follows
o
n
Φ(∆)
∆∗ = arg∆∈(ξ,ρ) max min QOM (∆), QOM (ρ) + Q̄RE (11, ρ)
(
Φ(∆)
Q (ξ) = QOM (ρ) + Q̄RE (11, ρ) If T is true,
= ∆ OM
Φ(∆)
QOM (∆) = QOM (ρ) + Q̄RE (11, ρ) elsewhere,
(9)
Φ(ρ)
where T = {QOM (ξ) ≤ QOM (ρ) + Q̄RE (11, ρ)}, and
Φ(∆) = {xi ∈ Φ : ||xi || ≥ ∆}. The above expression takes
into account the active/inactive ERs with the worst energy
harvesting performance and maximizes the minimum between
them; since we have closed form expressions for the harvesting
performance for both cases ((2), (14)), the optimal value ∆∗
can be found by numerically solving the equation in (9).
2) Probabilistic binary backscattering (PBB): Similar to
the DBB technique, the ERs either fully reflect the received
signal or remain inactive. In this case, the mode selection
is performed in a probabilistic way and thus the ERs with
a probability p operate in backscattering mode and with a
probability 1 − p remain silent. The average harvested energy
can be expressed as
QPBB (p) = Λ(ξ, ρ) + pQFB (ξ, λp).
(10)
Equivalently to the DBB scheme, p can be designed to boost
the energy harvesting performance of the worst case (i.e., ERs
located in the edge) such as
′
′
1, ρ) ,
p∗ = argp∈(0,1) max Q̄Φ
RE (1
(11)
where φ denotes a PPP with density pλ; due to the complexity
of the expression, p∗ does not have closed-form solution and
can be evaluated numerically.
D. Harvesting-target backscattering (HTB)
In the HTB scheme, we assume that each ER aims to
achieve a minimum harvested energy from the retrodirective
component of the received signal. Let Γi ≥ 0 denote the
energy harvesting target for the i-th ER. In this case, the
problem resembles to the conventional power control problem for time varying channels with user-specific signal-tointerference-plus-noise ratio requirements, which admits online distributed implementation (Foschini-Miljanic algorithm
[12]). Specifically, the reflection coefficient βi for the i-th ER
is given by the following iterative algorithm
Γi
βi (l) .
(12)
βi (l + 1) = min 1, Φ
β (l), di )
QRE (β
The convergence properties of the iterative scheme have
been studied by Yates [13]; it is worth noting that the channel
gains between the ET and the ERs remain constant during the
iterations (e.g., stationary users with slowly-varying channels).
For the case where the problem is feasible and thus all ERs
satisfy the minimum harvested energy constraint with βi ≤ 1,
the backscattering coefficients are hgiven in closed form i.e.,
i
Γ1 ñ1
ΓK ñK
β ∗ = (II K − F )−1u , where u = d−3α
, . . . , d−3α
,
|g |2
|g |2
1
1
and F is a K × K matrix with Fi,j =
i, j = 1, . . . , K [12].
V. N UMERICAL
K
Γi |gj |2 d−2α
j
|gi |2 d−3α
i
K
and
RESULTS
The simulation setup follows the description in Sec. II
with M = 500, λ = 0.01, fc = 900 MHz, α = 3,
σ 2 = −150 dBm, τ = 10−8 sec, ξ = 2 m, ρ = 30 m,
and ζ = 1 (unless otherwise stated). In Fig. 1, we present
the average harvested energy versus the transmit power Pt for
the proposed reflection policies. The performance of perfect
beamforming (full CSI at the ET) and non-backscattering
(omnidirectional transmission) are given as benchmarks. As
it can be seen, the FB scheme significantly outperforms the
non-backscattering scheme and provides a performance that is
close to the perfect beamforming case. On the other hand, the
DIB scheme outperforms non-backscattering, while it provides
fairness among ERs (in comparison to FB).
Fig. 2(a) plots the average harvested energy of the DBB
scheme for different ER locations versus the distance threshold
∆. As it can be seen, the parameter ∆ provides an energy
harvesting balance among ERs located in different distances
from the ET. According to the expression in (9), the value
∆∗ that ensures max − min fairness corresponds to the crossover point between the average harvested energy of the ERs
located in distance ξ and ρ, respectively. In Fig. 2(b), we deal
with the performance of the PBB scheme. By controlling the
probability p, we achieve different energy harvesting balances
among the ERs; the optimal value p∗ is selected to maximize
the energy harvesting at the edge of the cell.
Finally, Fig. 3 deals with the HTB scheme and plots the
percentage of the ERs that can achieve a common harvesting
4
10
10
10
10
10
−1
DIB (Simulation)
DIB (Bound Eq. (4))
FB (Simulation)
FB (Theory)
Non−backscattering (Simulation)
Non−backscattering (Eq. (6))
Beamforming (Simulation)
Beamforming (Eq. (7))
−2
10
10
10
10
100
2
d=ξ
d=ρ
d=15 m
d=20 m
0
90
∆*
80
−2
70
−4
8
10
12
14
16
18
20
∆ [m]
(a)
22
24
26
28
30
−3
−4
−5
−6
0
5
10
15
20
P [dBm]
25
30
35
40
10
10
0
−1
Av. harvested energy Q
−2
50
40
PBB
20
−3
10
10
−4
0
0.1
0.2
0.3
0.4
(b)
t
60
(p)
Av. harvested energy d=ρ
10
HTB, λ=0.01, Γ=−10 dBm
FB, λ=0.01, Γ=−10 dBm
HTB, λ=0.01, Γ=0 dBm
FB, λ=0.01, Γ=0 dBm
HTB, λ=0.01, Γ=10 dBm
FB, λ=0.01, Γ=10 dBm
HTB, λ=0.1, Γ=−10 dBm
FB, λ=0.1, Γ=−10 dBm
30
p*
10
Percentage of ERs [%]
Average Harvested Energy [W]
10
0
Average Harvested Energy [W]
10
1
Average Harvested Energy [W]
10
0.5
0.6
0.7
0.8
0.9
1
0
0
p
5
10
15
20
Pt [dBm]
25
30
35
40
Fig. 1. Average harvested energy versus the transmit Fig. 2. (a) DBB scheme- average harvested energy
Fig. 3. Percentage of ERs that satisfy the energy
power (DIB, FB, Non-backscattering, Beamform- versus ∆, (b) PBB scheme- average harvested energy
harvesting constraint for the HTB scheme.
ing.).
versus p; ξ = 8 m, Pt = 40 dBm
target Γi = Γ from the retrodirective component versus
the transmit power Pt . The HTB scheme is based on the
distributed implementation (see (12)) with (maximum) 100
iterations; the FB scheme is used as a benchmark. We can
observe that as the harvesting threshold and the density of
users increase, the percentage of the ERs that satisfy the energy
harvesting constraint decreases. In addition, the HTB scheme
significantly outperforms the FB scheme and the performance
gain becomes more important as the transmit power increases.
VI. C ONCLUSION
This letter has dealt with a new EB scheme for WPT multiuser networks, which jointly combines large antenna retrodirectivity with passive signal backscattering. The proposed
technique does not require CSI or active signal transmission
at the ERs, and the energy harvesting performance depends
on the reflection coefficients. Several reflection strategies
have been investigated and analyzed by taking into account
location randomness. The proposed EB technique has lower
power consumption than conventional counterparts and is a
promising solution for future machine-type communications.
An interesting extension of this work is to investigate the
proposed scheme for more complex networks with multiple
ETs.
A PPENDIX
The complementary cumulative distribution function
(CCDF) of the harvested energy (associated with the
retrodirective component of the received signal) for a given
path-loss value d is written as
1, d) > x)
F (x|d) = F (x|di = d) = P(QΦ
RE (1
2
X
M
σ
= P |gi |2 > Y (x)
d−2α
|gk |2 +
k
ζPt τ
xk ∈Φ,k6=i
X
Y (x)
d−2α
|gk |2
= exp (−U (x)) EΦ,g exp −
k
M
xk ∈Φ,k6=i
Z
Y (x)|g|2 y −2α
ydy
= exp(−U (x)) exp −λ 1−Eg exp −
M
D Z ρ
Y (x)y
= exp(−U (x)) exp −2πλ
dy ,
(13)
Y
(x)
+ y 2α
ξ
where Y (x) =
x
−1
ζPt d−α
x
(M+1−
)d−2α
−α
ζPt d
, U (x) =
Y (x)σ2
ζPt τ ,
|gi |2
follows an exponential distribution with rate parameter 1/M
and (13) follows from the probability generating functional
of a PPP [9, Sec. 4.6]. From the CCDF, we can calculate
the average harvested energy for a given path-loss value d as
follows
−α
Z
1, d) =
Q̄Φ
RE (1
(M+1)ζPt d
F (x|d)dx.
(14)
ζPt d−α
With expectation over the path-loss
Z values, we have
Φ
1, y)fd (y)ydy
Q̄Φ
QFB (ξ, λ) = E[Q̄RE (11, d)] =
RE (1
=
2
ρ2 − ξ 2
Z
ξ
ρ
Z
D
(M+1)ζPt y −α
F (x|y)ydxdy.
(15)
ζPt y −α
R EFERENCES
[1] Y. Zeng, B. Clerckx, and R. Zhang, “Communications and signal design
for wireless power transmission,” IEEE Trans. Commun., vol. 65, pp.
2264–2290, May 2017.
[2] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless
information and power,” IEEE Trans. Wireless Commun., vol. 12, pp.
1989–2001, May 2013.
[3] J. Xu and R. Zhang, “Energy beamforming with one-bit feedback,” IEEE
Trans. Signal Process., vol. 62, pp. 5370–5381, Oct. 2014.
[4] Y. Zeng and R. Zhang, “Optimized training design for wireless energy
transfer,” IEEE Trans. Commun., vol. 63, pp. 536–550, Feb. 2015.
[5] S. Lee, Y. Zeng, and R. Zhang, “Retrodirective multi-user wireless power
transfer with massive MIMO,” IEEE Wireless Commun. Lett., (accepted
for publication), 2017.
[6] G. Yang, C. K. Ho, and Y. L. Guan, “Multi-antenna wireless energy
transfer for backscatter communication systems,” IEEE J. Selec. Areas
Commun., vol 33, pp. 2974–2987, Dec. 2015.
[7] K. Huang, C. Zhong, and G. Zhu, “Some new research trends in wirelessly
powered communications,” IEEE Wireless Commun., vol. 23, pp. 19–27,
April 2016.
[8] M. Xia and S. Aissa,“On the efficiency of far-field wireless power
transfer,” IEEE Trans. Sign. Proc., vol. 63, pp. 2835–2847, June 2015.
[9] M. Haenggi, Stochastic geometry for wireless networks, Cambridge Univ.
Press, 2013.
[10] D. Masotti, A. Costanzo, M. Del Prete, and V. Rizzoli,“Retrodirective
arrays for wireless communications,” IEE Micro. Mag., vol. 3, pp. 71–79,
March 2002.
[11] Y. -G. Lim, C. -B. Chae, and G. Caire, “Performance analysis of massive
MIMO for cell-boundary users,” IEEE Trans. Wireless Commun., vol. 14,
pp. 6827–6842, Dec. 2015.
[12] G.J. Foschini, Z. Miljanic, “A simple distributed autonomous power
control algorithm and its convergence,” IEEE Trans. Veh. Tech., vol. 42,
pp. 641–646, Apr. 1993
[13] R. D. Yates, “A framework for uplink power control in cellular radio
systems,” IEEE J. Sel. Areas Commun., vol. 13,pp. 1341–1347, 1995.
| 7cs.IT
|
arXiv:1604.06392v2 [math.CO] 13 Apr 2017
Convergence criteria for FIW –algebras and polynomial
statistics on maximal tori in type B/C
Rita Jiménez Rolland∗ and Jennifer C. H. Wilson†
April 14, 2017
Abstract
A result of Lehrer describes a beautiful relationship between topological and combinatorial
data on certain families of varieties with actions of finite reflection groups. His formula relates
the cohomology of complex varieties to point counts on associated varieties over finite fields.
Church, Ellenberg, and Farb use their representation stability results on the cohomology of flag
manifolds, together with classical results on the cohomology rings, to prove asymptotic stability
for “polynomial” statistics on associated varieties over finite fields. In this paper we investigate
the underlying algebraic structure of these families’ cohomology rings that makes the formulas
convergent. We prove that asymptotic stability holds in general for subquotients of FIW –algebras
finitely generated in degree at most one, a result that is in a sense sharp. As a consequence, we
obtain convergence results for polynomial statistics on the set of maximal tori in Sp2n (Fq ) and
SO2n+1 (Fq ) that are invariant under the Frobenius morphism. Our results also give a new proof
of the stability theorem for invariant maximal tori in GLn (Fq ) due to Church–Ellenberg–Farb.
Contents
1
Introduction
2
2
Preliminaries
2.1 Representation theory of the hyperoctahedral group Bn . . . . . . . . . . . . . . . . .
2.2 The theory of FIW –modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Asymptotics and asymptotic notation . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
6
8
10
3
Asymptotic stability
3.1 Asymptotics of character polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Convergence for subquotients of FIW -algebras generated in low degree . . . . . . . .
10
10
13
4
Point counts for maximal tori in type B and C
4.1 Maximal tori in types B and C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Asymptotic results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq ) . . . . . . . .
17
18
22
24
∗ The first author is grateful for support from the Sofı́a Kovalévskaia Foundation, the Sociedad Matemática Mexicana
and PAPIIT-UNAM grant IA100816. † The second author is grateful for the support of an AMS-Simons Travel Grant.
1
R Jiménez Rolland
J C H Wilson
Introduction
1
Introduction
In this paper we study the complete complex flag varieties associated to the linear groups SO2n+1 (C)
and Sp2n (C) of type B and C, respectively. These spaces and their cohomology algebras are described in Section 4.1; their cohomology admits actions of the signed permutation group Bn . A
remarkable formula due to Lehrer [Leh92] relates the representation theory of these cohomology
groups with point-counts on varieties over Fq that parametrize maximal tori in SO2n+1 (Fq ) and
Sp2n (Fq ). One result of this paper is to exhibit a form of convergence for these formulas as the
parameter n tends to infinity.
Church, Ellenberg, and Farb prove that the cohomology groups of the flag varieties associated
to GLn (C) are representation stable with respect to the Sn –action [CEF15, Theorem 1.11]. Using this
result and a description of the cohomology due to Chevalley, they establish asymptotic stability for
“polynomial” statistics on the set of Frobenius-stable maximal tori in GLn (Fq ) [CEF14, Theorem
5.6]. In Theorem 4.3 we prove the corresponding result in type B and C, using a representation
stability result of the second author [Wil14].
To establish our results, we give general combinatorial criteria on the cohomology algebras of
the complex varieties that ensure convergence. This is one of few results in the FI–module literature that makes full use of the multiplicative structure on the FIW –algebras. Church–Ellenberg–
Farb prove convergence for statistics on maximal tori in type A using sophisticated results specific
to these cohomology algebras. This paper shows that convergence in fact follows from much simpler features of the algebras.
Table 1: Some statistics for Frobenius-stable maximal tori of Sp2n (Fq ) and SO2n+1 (Fq )
Frq -stable maximal tori statistic
for Sp2n (Fq ) and SO2n+1 (Fq )
Hyperoctahedral
character
Formula in
terms of n
Limit
as n → ∞
Total number of
Frq -stable maximal tori
1
q 2n (Steinberg)
Expected number of
X1 + Y 1
2
1+
q2
(q 2 − 1)
1
1
1
+ 4 + · · · 2n−2
q2
q
q
1-dimensional Frq -stable subtori
Expected number of split
1
2
X1
1+
1
1
1
+ 2 + · · · 2n−1
q
q
q
q
2(q − 1)
1-dimensional Frq -stable subtori
Expected value of reducible
minus irreducible Frq -stable
2-dimensional subtori
Expected value of split minus
X + Y
1
1
− (X2 + Y2 )
2
q4 −
q2 1 −
X2 − Y 2
non-split Frq -stable 2-dimensional
irreducible subtori
2
1−
1
q 2n
(q 2 −
1
q 2n
1
q 2(n−1)
1) (q 4 − 1)
2 (q 4
1−
− 1)
1
q 2(n−1)
q4
(q 2 − 1)(q 4 − 1)
q2
− 1)
2 (q 4
R Jiménez Rolland
J C H Wilson
Introduction
Examples of limiting statistics computed in Section 4.3 are presented in Table 1; the notation in
this table is defined below. The first of these result is due to Steinberg [Ste68, Corollary 14.16], and
the other results appear to be new.
A general convergence result: algebras generated in low degree
A primary goal of this paper is to describe the underlying algebraic structure on the cohomology
algebras that drives these stability results. Throughout, we will use the notation Wn to denote
either the family of symmetric groups Sn (the Weyl groups in type A) or the family of hyperoctahedral groups Bn (the Weyl groups in type B/C). We prove in Section 3.2 that convergence of
the statistics on maximal tori in type A and B/C holds because the graded pieces of the associated coinvariant algebras are, as a sequence of graded Wn –representations, subquotients of what
we call an FIW –algebra finitely generated in degree at most 1. Terminology and basic theory of FIW –
modules is summarized in Section 2.2. Speaking informally, these FIW –algebras are sequences of
graded algebras with Wn –actions that are generated by representations of W0 and W1 . Concretely,
they include the sequence of algebras Γ∗Wn which we now define.
Let k be a subfield of C. Let M (1)n denote the k–vector space with basis x1 , x2 , . . . , xn . The
symmetric group Sn acts on the variables xi by permuting the subscripts. Let M (0)n denote the
k–vector space generated by the single element y with trivial Sn –action. Then for any integers
b, c ≥ 0 let Γ∗Sn be the graded algebra generated by
M (0)n⊕b ⊕ M (1)n⊕c
with each copy of M (0)n and M (1)n assigned a positive grading. The algebra Γ∗Sn may be taken
to be commutative, anticommutative, or graded-commutative. It inherits a diagonal action of Sn
on its monomials.
Analogously, let MBC (1)n be the k–vector space with basis x1 , x1 , x2 , x2 , . . . , xn , xn . The hyperoctahedral group Bn acts on the variables xi by permuting and negating the subscripts. Again let
MBC (0)n be the k–vector space generated by a variable y with trivial Bn action. Let Γ∗Bn be the
symmetric, exterior, or graded-commutative algebra generated by
⊕c
MBC (0)⊕b
n ⊕ MBC (1)n
with each family of variables graded in positive degree.
Theorem 3.5 shows the algebras Γ∗Wn and their subquotients satisfy a convergence property
with respect to certain family of ‘polynomial’ class functions, first defined on the symmetric groups
by MacDonald [Mac79, I.7 Example 14] as follows. For any permutation σ and positive integer r,
let Xr be the function that outputs the number Xr (σ) of r–cycles in the cycle type of σ. Polynomials
in the functions Xr , called character polynomials, play a major role in the theory of representation
stability developed by Church–Ellenberg–Farb [CEF15]. In Section 2.1.2 we recall an analogous
definition for the Weyl group Bn in type B and C. These hyperoctahedral character polynomials are
polynomials
P ∈ k[X1 , Y1 , X2 , Y2 , . . .]
in the ‘signed cycle counting functions’ Xr and Yr on Bn . See Section 2.1 for the prerequisite
background on the representation theory of the groups Bn and a precise definition of these class
functions.
3
R Jiménez Rolland
J C H Wilson
Introduction
Theorem 3.5 is a stability result for the asymptotics of the inner products hPn , Adn iWn of a character polynomial P with a subquotient Adn of ΓdWn . This is precisely the stability result needed to
prove convergence of the point-counts that appear in Lehrer’s formulas.
Theorem 3.5 (Criteria for convergence). Let Wn denote one of the families Sn or Bn . Let Γ∗Wn be one
of the commutative, exterior, or graded commutative algebras defined above, and let {Adn } be any sequence
of graded Wn –representations such that Adn is a subrepresentation of ΓdWn . Then for any Wn character
polynomial P and integer q > 1 the following series converges absolutely
∞
X
limn→∞ hPn , Ad iW
n
d=0
n
qd
.
The proof (given in Section 3.2) uses an analysis of the combinatorics of coloured partitions. A
closely related result in Type A is proved in [FW] using very different methods; see Remark 3.9.
A failure of convergence
In forthcoming work [JRW17] the authors show that, in contrast to Theorem 3.5, convergence
may fail for FIW –algebras generated in degree 2 or greater. To give a concrete example, let A∗n be
the polynomial algebra k[x1,2 , x1,3 , x2,3 , x1,4 , . . . , xn−1,n ] generated by commuting variables xi,j =
xj,i , i 6= j, in graded degree 1, with an action of Wn by permuting the indices. Then the series
∞
X
limn→∞ hPn , Ad iW
n
n
qd
d=0
does not converge for any positive integer q even for the constant character polynomial P = 1. This
counterexample shows a fundamental difference in the asymptotic behaviour of FIW –algebras
generated by representations of W0 or W1 and those generated by representations of Wn in degree
n ≥ 2.
An application: polynomial statistics for coinvariant algebras and Frq -stable maximal tori
We apply Theorem 3.5 to the family {Rnd }, where Rnd denotes the dth -graded piece of the complex
coinvariant algebra Rn∗ in type A or in type B/C. A description of these algebras appears in Section
4.1.3. We obtain a convergent formula in Proposition 4.4 for coinvariant algebras of type B/C and
give a new proof of the result in type A (see [CEF14, Theorem 5.6] and Remark 4.5). Theorem 3.5
shows that the convergence of this formula follows because the complex coinvariant algebras are
quotients of the polynomial rings C[x1 , x2 , . . . , xn ]; convergence does not depend on any deeper
structural features of these algebras.
For a general connected reductive group G defined over Fq , let T Frq denote the set of maximal tori of G(Fq ) that are stable under the action of Frq , the Frobenius morphism. Frq -stable
maximal tori of reductive groups G defined over Fq are fundamental to the study of the representation theory of the finite group GFrq [DL76]. More background is given in Section 4.1. Lehrer
[Leh92] obtained remarkable formulas that relate functions that are defined on the set T Frq with
4
R Jiménez Rolland
J C H Wilson
Introduction
the character theory of the Weyl group of G. We recall one of Lehrer’s formulas in Theorem 4.2 for
the cases when G is the symplectic group Sp2n or the special orthogonal group SO2n+1 ; the formula connects functions on T Frq and the cohomology of the generalized flag varieties associated
to G(C).
In Section 4.1.2 we describe how the hyperoctahedral character polynomials of Section 2.1.2 define functions on T Frq . Specifically, for T ∈ T Frq , Xr (T ) is the number of r-dimensional Frq -stable
subtori of T irreducible over Fq that split over Fqr , and Yr (T ) is the number of r-dimensional Frq stable subtori of T irreducible over Fq that do not split over Fqr . In Section 4, we combine Lehrer’s
formula with Proposition 4.4 to establish asymptotic results for these polynomial statistics on the
set of Frq -stable maximal tori of Sp2n (Fq ) and of SO2n+1 (Fq ). We obtain Theorem 4.3 below, the
type B and C analogues to [CEF15, Theorem 5.6] for Frq -stable maximal tori for GLn (Fq ).
Theorem 4.3. (Stability of maximal tori statistics). Let q be an integral power of a prime p. For n ≥ 1,
Fr
d
denote by Tn q the set of Frq -stable maximal tori for either Sp2n (Fp ) or SO2n+1 (Fp ). Let Rm
denote the
th
∗
d -graded piece of the complex coinvariant algebra Rm in type B/C. If P ∈ C[X1 , Y1 , X2 , Y2 , . . .] is any
hyperoctahedral character polynomial, then the normalized statistic
X
2
P (T )
q −2n
Frq
T ∈Tn
converges as n → ∞. In fact,
2
lim q −2n
n→∞
X
∞
d
X
limm→∞ hPm , Rm
iBm
,
d
q
P (T ) =
Frq
d=0
T ∈Tn
and the series in the right hand converges.
At the end of Section 4, we use Theorem 4.3 and Stembridge’s formula for decomposing the
Bn –representation Rnd (see Theorem 4.7) to compute the specific asymptotic counts in Table 1.
Related work
Fulman [Ful16, Section 3] uses generating functions to recover the specific counts obtained by
Church–Ellenberg–Farb [CEF14] for Frq -stable maximal tori in GLn (Fq ). Chen uses generating
function techniques to obtain a linear recurrence on the stable twisted Betti numbers of the associated flag varieties[Che, Corollary 2 (II)]. These methods should also apply to study maximal tori
of other classical groups, including the ones considered in this paper. They offer an alternate approach to performing the computations in Section 4.3, and may shed further light on the structure
of the asymptotic formulas.
Farb and Wolfson [FW] extend the methods of Church–Ellenberg–Farb [CEF14] to establish
new algebro-geometric results. In [FW, Theorem B] they prove that configuration spaces of n
points in smooth varieties have convergent cohomology [FW, Definition 3.1] and obtain arithmetic
statistics for configuration spaces over finite fields [FW, Theorem C].
The results of this paper and the work of Church–Ellenberg–Farb [CEF14] complement the
existing literature on enumerative properties of matrix groups over finite fields; see for example
5
R Jiménez Rolland
J C H Wilson
Preliminaries
[NP95, NP98, Ful99, Wal99, NP00, Bri02, FNP05, LNP09, NP10, NPP10, NPS13, BGPW13, NPP14].
These authors seek, for example, to determine the proportion of elements in certain finite matrix
groups that are cyclic, separable, semisimple, or regular. Their methods are different from those in
this paper, and include the calculus of generating functions, complex analysis, classical Lie theory,
finite group theory, analytic number theory, and statistics. This probabilistic approach to the study
of finite groups has applications to algorithm design in group theory, random matrix theory, and
combinatorics.
Acknowledgments
We would like to thank Benson Farb for proposing this project. We thank Weiyan Chen, Tom
Church, and Jason Fulman for useful conversations. We are grateful to Ohad Feldheim for suggesting the proof of Proposition 3.6. We thank Jason Fulman for finding an error in a formula in
Table 1 in an early version of this paper. The first author would also like to thank the Centro de
Ciencias Matemáticas, UNAM-Morelia, for all its support during the writing of this paper.
2
Preliminaries
In this section we summarize some necessary background material and terminology. Section 2.1
describes the representation theory of the hyperoctahedral groups in characteristic zero. In Section
2.2 we review the foundations of FIW –modules developed by Church–Ellenberg–Farb [CEF15] in
type A and by the second author [Wil14, Wil15] in type B/C. Section 2.3 defines the asymptotic
notation that will be used in this paper.
Notation 2.1. (The inner product h− , −iG ). Throughout the paper, given a finite group G we
write h− , −iG to denote the standard inner product on the space of class functions of G. By abuse
of notation we may write a G–representation V in one or both arguments to indicate the character
of V .
2.1
Representation theory of the hyperoctahedral group Bn
Let Bn denote the Weyl group in type Bn /Cn , the wreath product
Bn ∼
= Z/2Z ≀ Sn = (Z/2Z)n ⋊ Sn ,
which we call the hyperoctahedral group or signed permutation group. By convention B0 is the trivial
group. We may view Bn as the subgroup of permutations SΩ ∼
= S2n on the set Ω = {1, 1, 2, 2, . . . , n, n}
defined by
Bn = {σ ∈ SΩ | σ(a) = σ(a) for all 1 ≤ a ≤ n }.
Here a denotes negative a; in general the bar represents the operation of negation and satisfies
a = a for a ∈ Ω.
2.1.1
Conjugacy classes of the hyperoctahedral groups
The conjugacy classes of Bn are classified by signed cycle type, defined as follows. A positive r–cycle
is a signed permutation of the form (s1 s2 · · · sr )(s1 s2 · · · sr ) ∈ SΩ , si ∈ Ω. This element reverses
6
2.1 Representation theory of the hyperoctahedral group Bn
R Jiménez Rolland
J C H Wilson
the sign of an even number of letters. A negative r–cycle is a signed permutation of the form
(s1 s2 · · · sr s1 · · · sr ) ∈ SΩ , which reverses an odd number of signs. The rth power of a positive r–
cycle is the identity, whereas the rth power of a negative r–cycle is the product of r transpositions
(si si ). Positive and negative r–cycles project to r–cycles under the natural surjection Bn → Sn . For
example, (1 3 2)(1 3 2) is a positive 3-cycle and (1 3 2 1 3 2) is a negative 3-cycle which both project
to the 3–cycle (1 3 2) ∈ Sn .
In 1930 Young [You30] proved that signed permutations factor uniquely as a product of positive
and negative cycles, and two signed permutations are conjugate if and only if they have the same
signed cycle type. We represent the signed cycle type of a signed permutation by a double partition
(λ, µ), where λ is a partition with a part of length r for each positive r–cycle, and µ a partition
encoding the negative r–cycles.
2.1.2
Character polynomials for Bn
Given a signed permutation σ, let Xr (σ) denote the number of positive r–cycles in its signed cycle
type, and let Yr (σ)`be the number of negative r–cycles. Then Xr and Yr define class functions on
the disjoint union n≥0 Bn . A character polynomial with coefficients in a ring k is a polynomial P ∈
k[X1 , Y1 , X2 , Y2 , . . .], in analogy to the character polynomials for the symmetric groups defined by
MacDonald [Mac79, I.7 Example 14]. A character polynomial P restricts to a class function on Bn
for every n ≥ 0; we denote its restriction by Pn . We define the degree of a character polynomial by
setting deg(Xr ) = deg(Yr ) = r for all r ≥ 1.
Let µ and λ be partitions, and let nr be the function on the set of partitions that takes a partition
and outputs the number of parts of size r. The space of character polynomials is spanned by
polynomials of the form
Y Xr Yr
X
Y
.
Pµ,λ =
:=
nr (µ) nr (λ)
µ
λ
r≥1
When n = |λ| + |µ| the restriction of Pµ,λ to Bn is the indicator function for signed permutations
of signed cycle type (λ, µ).
2.1.3
Classification of irreducible Bn –representations
The irreducible complex representations of Bn are in natural bijection with double partitions of n,
that is, ordered pairs of partitions (λ, µ) such that |λ| + |µ| = n. These irreducible representations
are constructed from representations of the symmetric group Sn ; this construction is described (for
example) in Geck–Pfeiffer [GP00].
From the canonical surjection Bn → Sn we can pull back representations of Sn to Bn . Let V(λ,∅)
denote the irreducible Bn –representation pulled back from the Sn –representation associated to
the partition λ of n. Let V(∅,(n)) denote the 1-dimensional representation of Bn where a signed
permutation acts by 1 or −1 depending on whether it reverses an even or odd number of signs.
Then for any partition µ of n we denote V(∅,µ) := V(µ,∅) ⊗C V(∅,(n)) . In general, for partitions λ of
k and µ of m, we define the Bk+m –representation
B
V(λ,µ) := IndBk+m
V
⊠ V(∅,µ) .
k ×Bm (λ,∅)
7
R Jiménez Rolland
J C H Wilson
2.2 The theory of FIW –modules
For each double partition (λ, µ) the representation V(λ,µ) is distinct and irreducible, and this construction gives a complete list of irreducible complex Bn –representations. As with the symmetric
group, each complex irreducible representation is defined over the rational numbers, and each
representation is self-dual [GP00, Corollary 3.2.14].
2.2
The theory of FIW –modules
Church–Ellenberg–Farb [CEF15] introduced the theory of FI–modules to study sequences of representations of the symmetric groups Sn . FI denotes the category of finite sets and injective maps;
an FI–module over a ring R is a functor from FI to the category of R–modules. Their results were
generalized by the second author to sequences of representations of the classical Weyl groups in
type B/C and D [Wil14, Wil15].
Definition 2.2. (The categories FI and FIBC ). Following Church–Ellenberg–Farb [CEF15], we let
FI denote the category of finite sets and injective maps. We write n or [n] to denote the object
{1, 2, . . . , n} and 0 := ∅. Following work of the second author [Wil14, Wil15], we let FIBC denote
the type B/C analogue of FI; we may define FIBC to be the category where the objects are finite sets
n := {1, 1, . . . , n, n}, and the morphisms are all injective maps f : m → n satisfying f (a) = f (a)
for all a in m. Notably the endomorphisms End(n) of FI are the symmetric groups Sn , and the
endomorphisms End(n) of FIBC are the hyperoctahedral groups Bn . We denote the families of
Weyl groups Sn and Bn generically by Wn , and we denote the categories FI and FIBC generically
by FIW .
The following definitions appear in Church–Ellenberg–Farb [CEF15] and Wilson [Wil14].
Definition 2.3. ((Graded) FIW –modules and FIW –algebras). Fix Wn to denote either Sn or Bn .
Let k be a ring, assumed commutative and with unit. An FIW –module V over k is a functor from
FIW to the category of k–modules; its image is sequence of Wn –representations Vn := V (n) with
actions of the FIW morphisms. A graded FIBC –module V ∗ over a ring k is a functor from FIBC to
the category of graded k–modules; a graded FIW –algebra A∗ over k is a functor from FIW to the
category of graded k–algebras. Each graded piece V d or Ad inherits an FIW –module structure. We
will refer to d as the graded–degree and n as the FIW –degree of the k–module Vnd .
Definition 2.4. (Finite generation; degree of generation; finite type). Let Wn denoted either
the family of symmetric groups or the family of signed
` permutation groups. An FIW –module V
is generated (as an FIW –module) by elements {vi } ⊆ n Vn if V is the smallest FIW –submodule
containing
elements. A graded FIW –algebra A∗ is generated (as an FIW –algebra) by the set
` those
∗
{vi } ⊆ n An if A is the smallest FIW –subalgebra containing {vi }. An FIW –module or FIW –
algebra V is finitely generated if it has a finite generating set, and V is generated in FIW –degree ≤ m
if it has a generating set vi ∈ Vmi such that mi ≤ m for all i. A graded FIW –module or algebra V ∗
has finite type if each graded piece V d is a finitely generated FIW –module.
Definition 2.5. (Weight; slope). Church–Ellenberg–Farb defined an FI–module V over a subfield k
of C to have weight ≤ m if for every n, every irreducible Sn –representation Vλ , λ = (λ1 , λ2 , . . . , λr ),
occuring in Vn satisfies (n − λ1 ) ≤ m. Analogously an FIBC –module V over a subfield of C has
weight ≤ m if for all n every irreudcible Bn –representation V(λ,µ) , λ = (λ1 , λ2 , . . . , λr ), satisfies
(n − λ1 ) ≤ m. A graded FIW –module or algebra V ∗ has finite slope M if V d has weight at most dM .
8
R Jiménez Rolland
J C H Wilson
2.2 The theory of FIW –modules
FI– and FIBC –modules over k generated in degree at most m necessarily have weight at most
m [CEF15, Proposition 2.3.5], [Wil14, Theorem 4.4].
Some of the early results on FI– and FIBC –modules concern the implications of finite generation
for the structure of an FIW –module and its characters: over characteristic zero, a finitely generated
FIW –module is uniformly representation stable in the sense of Church–Farb [CF13, Definition 2.3],
and its characters are eventually polynomial. The following result was proved by Church–Ellenberg–
Farb [CEF15] in type A and the second author [Wil14, Wil15] in types B/C and D.
Theorem 2.6. (Constraints on finitely generated FIW –modules). Let V be an FIW –module over a
subfield k of C which is finitely generated in degree ≤ m.
• (Uniform representation stability)([CEF15, Theorem 1.13], [Wil14, Theorem 4.27]). The sequence Vn is uniformly representation stable with respect to the maps induced by the FIW –morphisms
the natural inclusions In : n → (n + 1), stabilizing once n ≥ m + stabdeg(V ). In particular, in
the decomposition of Vn into irreducible Wn –representations, the multiplicities of the irreducible representations are eventually independent of n in the sense of representation stability [CF13, Definition
2.3].
• (Character polynomials)([CEF15, Proposition 3.3.3 and Theorem 3.3.4], [Wil15, Theorem
4.6]). Let χn denote the character of the Wn –representation Vn . Then there exists a unique character
polynomial FV of degree at most m such that FV (σ) = χn (σ) for all σ ∈ Wn , for all n sufficiently
large.
A major tool in the analysis of finitely generated FI– and FIBC –modules is their relationship to
represented functors and their subfunctors.
Definition 2.7. (The FIW –modules M (m)). We denote the represented FI or FIBC –module over k
by
MBC (m) := k HomFIBC (m, −)
M (m) := k HomFI (m, −)
The FI–module M (m)n ∼
= k HomFI (m, n) has k–basis
Sn /Sn−m ∼
= {ei1 ,i2 ,...,im | (i1 , i2 , . . . , im ) is an ordered m–tuple of distinct positive elements of n}
where the m–tuple (i1 , i2 , . . . , im ) encodes the image of the melements (1, 2, . . . m) under an FI–
morphism m → n. Similarly, the FIBC –module MBC (m)n ∼
= k HomFIBC (m, n) has k–basis
(i1 , i2 , . . . , im ) is an ordered m–tuple of elements of n;
∼
e
Bn /Bn−m = i1 ,i2 ,...,im
at most one of a or a appears at most once
where the m–tuple (i1 , i2 , . . . , im ) encodes the image of the m elements (1, 2, . . . m) under an FIBC –
morphism m → n. By an orbit–stabilizer argument we have isomorphisms of Wn –representations
M (m)n ∼
MBC (m)n ∼
= k Sn /Sn−m
= k Bn /Bn−m .
We denote the functors M (m) and MBC (m) generically by MW (m). An FIW –module V over k
is finitely generated in degree ≤ p if and only if it is a quotient of a finite direct sum of represented
functors MW (m) with m ≤ p; see Church–Ellenberg–Farb [CEF15, Proposition 2.3.5] and Wilson
[Wil14, Proposition 3.15].
9
R Jiménez Rolland
J C H Wilson
2.3 Asymptotics and asymptotic notation
2.3
Asymptotics and asymptotic notation
The following terminology features in the results of Section 3 on asymptotic stability.
Definition 2.8. (Asymptotic equivalence; asymptotic bounds; Big and little O notation). For
functions f (d) and g(d), d ∈ Z≥0 , we say that f is asymptotically equivalent to g and write f ∼ g if
lim
d→∞
f (d)
= 1.
g(d)
We say f is order o(g) if
f (d)
= 0.
d→∞ g(d)
lim
We say that f is order O(g) or that f is asymptotically bounded by g if
|f (d)| ≤ C|g(d)|
for some constant C and all d sufficiently large.
Note that the set of functions in O(g) are closed under linear combinations (though not products).
Definition 2.9. (Subexponential growth). We say that a function f (d) is P
subexponential if f is
order 2o(d) . Notably, if f is subexponential in d and 0 ≤ r < 1 then the series d f (d)rd converges
absolutely.
A classical example:
Proposition 2.10 (Hardy–Ramanujan [HR18], Uspensky [Usp20]). (Growth of partitions). The
number P(n) of partitons of n satisfies
P(n) ∼
√ 2n
1
√ eπ 3 .
4n 3
In particular P(n) is subexponential in n.
3
Asymptotic stability
The main goal of this section is to prove Theorem 3.5, a general convergence result for FIW -algebras
finitely generated in FIW -degrees 0 and 1. Using the terminology and notation from the previous
section, Theorem B can be restated as Theorem 3.5 below. We collect first some preliminary combinatorial results in Section 3.1 to study the asymptotic behavior of FIW –modules and their character
polynomials. We then use these results in Section 3.2 to obtain Theorem 3.5.
3.1
Asymptotics of character polynomials
Each of the results Propositions 3.1, Lemma 3.2, and Lemma 3.4 mirrors a result proven by Church–
Ellenberg–Farb [CEF14] in Type A, and their methods can generally be modified to give arguments
in Type B/C. For the sake of completeness we briefly describe these proofs in the case of the
hyperoctahedral groups.
10
R Jiménez Rolland
J C H Wilson
3.1 Asymptotics of character polynomials
Proposition 3.1. (Stability for inner products of character polynomials). Fix a family Wn to be
Sn or Bn . Let k be a subfield of C. Let P, Q be two Wn character polynomials. Then the inner product
hQn , Pn iWn is independent of n for n ≥ deg(P ) + deg(Q).
Proof. This result was proved for Sn by Church–Ellenberg–Farb [CEF14, Proposition 3.9] and
their proof adapts readily to type B/C; we briefly summarize this proof. Since hQn , Pn iBn =
h1, Qn Pn iBn it suffices to check in the case when Q = 1 and P is the element
Y Xr Yr
X
Y
Pµ,λ =
:=
µ
nr (µ) nr (λ)
λ
r≥1
associated to two partitions µ and λ. Let d = |µ| + |λ|, and notice that deg(P ) = d. Let δλ,µ denote
the indicator function for the conjugacy class (µ, λ) in Bd . Let Nµ,λ denote the size of the conjugacy
class (µ, λ) in Bd . Then for n ≥ d,
X
X
Y
1 X
,1
=
δλ,µ (σ|S∪S )
µ
λ
|Bn |
Bn
σ∈Bn S⊆[n],|S|=d
where δλ,µ (σ|S∪S ) := 0 if S ∪ S is not stabilized by σ. So
X
Y
1
n
,1
=
Nµ,λ |Bn−d |
µ
λ
|B
|
d
n
Bn
Nµ,λ
= d
2 d!
Hence for n ≥ d = deg(P ) + deg(Q) this inner product is independent of n, as claimed.
Lemma 3.2. (A convergence result for FIW –algebras of finite type and slope). Let Wn denote one
of the families Sn or Bn . Let k be a subfield of C. Suppose that A∗ is a FIW –algebra over k of finite type
and slope M . Then for each d and any Wn character polynomial P , the following limit exists:
lim hPn , Adn iWn .
n→∞
Proof. This result is proved in Type A by Church–Ellenberg–Farb; see the paragraph before [CEF14,
Corollary 3.11]. Their argument carries over to Type B/C as follows. By assumption, for each fixed
d the FIBC –module Adn is finitely generated and has weight at most dM . The second author proved
that therefore for some Dd ∈ Z, the characters of the sequence {Adn }n are given by a character
polynomial of degree at most dM for all n ≥ Dd [Wil15, Theorem 4.16]; see Theorem 2.6. By
Proposition 3.1, then, the value hPn , Adn iBn is independent of n once n = max{Dd , dM + deg(P )}.
This stable value gives the limit
lim hPn , Adn iBn = hPN , AdN iBN ,
N = max{Dd , dM + deg(P )}.
n→∞
Lemma 3.3. (A convergence result for finitely generated FIW –algebras). Let k be a subfield of C,
and let Wn represent one of the families Sn or Bn . Suppose that A∗ is an associative FIW –algebra over k
that is generated as an FIW –algebra by finitely many elements of positive graded-degree. Then for each d
and any Wn character polynomial P , the following limit exists:
lim hPn , Adn iWn .
n→∞
11
R Jiménez Rolland
J C H Wilson
3.1 Asymptotics of character polynomials
Proof. We first prove the result in type B/C. Let V be the graded FIBC –module additively generated by the generating set for the graded FIBC –algebra A∗ . By assumption on the generators V
will be supported in finitely many positive degrees. By construction V has finite type, and by Wilson [Wil14, Theorem 4.4], V has finite slope. It follows that A∗ is a quotient of the free associative
FIBC –algebra on V ,
∞
M
V ⊗j ;
khV i :=
j=0
see [Wil14, Definition 5.9]. By [Wil14, Proposition 5.2], tensor products respect finite generation,
and degree of generation is additive. Hence if we let M be the largest FIBC –degree of the generators of A∗ , the FIBC –module Ad is finitely generated in degree ≤ dM . Thus A∗ has finite type and
finite slope M ; see [Wil14, Proposition 5.10]. By Lemma 3.2, the limit limn→∞ hP, Akn iBn exists.
The proof for FI follows the same argument. The graded FI–algebra A∗ has finite type and
slope by Church–Ellenberg–Farb [CEF15, Proposition 3.2.5 and Theorem 4.2.3].
The following result will be used in the proof of Theorem 3.5.
Lemma 3.4. (On bounding growth of the graded pieces of FIW –algebras). Let Ad be the dth graded
piece of a graded FIW –module over a subfield k of C. Let g(d) be a function. The following are equivalent:
I. For each a ≥ 0 there is a function Fa (d) that independent of n and order O(g) such that
for all d and n.
dimk (Adn )Wn−a ≤ Fa (d)
II. For each Wn character polynomial P there is a function FP (d) that is independent of n and order
O(g) such that
for all d and n.
|hPn , Adn iBn | ≤ FP (d)
Proof. A special case of this result is the equivalence of the two conditions in [CEF14, Definition
3.12] for Sn ; although Church–Ellenberg–Farb only state the equivalence in Type A for the case
that g is subexponential, their proof implies the result in Type A for general functions g. Their
arguments may be adapted to the hyperoctahedral groups, and we summarize the proof in type
B/C below.
Let k denote the trivial Bn –representation. Assume (II) holds. By Frobenius reciprocity,
d
n
dimk ((Adn )Bn−a ) = hk, ResB
Bn−a An iBn−a
d
n
= hIndB
Bn−a k, An iBn
= hk[Bn /Bn−a ], Adn iBn
X1
a
, Adn iBn
= h2 a!
a
≤ F2a a!(X1 ) (d)
a
for all d and n.
So (II) implies (I). Now assume (I) holds and consider any double partition λ of a and the associated
Ba –representation Vλ with character χλ . The character of the induced representation
n
IndB
Ba ×Bn−a Vλ ⊠ k
12
3.2 Convergence for subquotients of FIW -algebras generated in low degree
is equal (for any n) to the character polynomial
Y Xr Y Yr
X
χλ (α, β)
:=
Pλ =
nr (α) r nr (β)
r
(α,β)
|α|+|β|=a
X
(α,β)
|α|+|β|=a
R Jiménez Rolland
J C H Wilson
X
Y
χλ (α, β)
.
α
β
Here χλ (α, β) denotes the value of the character χλ on a signed permutation of signed cycle type
(α, β). The character polynomials P λ are an additive basis for the space of hyperoctahedral character polynomials, and so it suffices to bound |hPnλ , Adn iBn |. Observe that
dimk ((Adn )Bn−a ) ≥ hVλ , (Adn )Bn−a iBa
d
n
= hVλ ⊠ k , ResB
Ba ×Bn−a An iBa ×Bn−a
d
n
= hIndB
Ba ×Bn−a Vλ ⊠ k , An iBn
= hPnλ , Adn iBn
which gives the desired bound.
3.2
Convergence for subquotients of FIW -algebras generated in low degree
The main result of this section is Theorem 3.5, which states that graded submodules of FIW –
algebras finitely generated in FIW –degree at most 1 satisfy our desired convergence result.
Theorem 3.5. (Criteria for convergent FIW –algebras). Let Wn denote one of the families Sn or Bn .
For nonnegative integers b, c, define a graded FIW –module V ∼
= MW (0)⊕b ⊕ MW (1)⊕c over a subfield k of
C with positive gradings. Let Γ∗ denote the free symmetric, exterior, or graded–commutative FIW –algebra
generated by V . Let A∗ be any FIW –algebra subquotient of Γ∗ . Then for any character polynomial P and
q > 1 the following sum converges absolutely
∞
X
limn→∞ hPn , Ad iW
n
d=0
qd
n
.
More generally, this sum will converge absolutely for any collection of Wn –representations {Adn | d, n ≥ 0}
such that Adn is a Wn –equivariant subquotient of Γdn .
We assume that the gradings on the algebra generators MW (0)⊕b ⊕ MW (1)⊕c are Wn –invariant.
We will use Theorem 3.5 to show convergence for coinvariant algebras in Proposition 4.4, and
to obtain Theorem 4.3 on stability for statistics on maximal tori.
The proof of Theorem 3.5 will use the following result on the asymptotics of enumerating
coloured partitions. Proposition 3.6 generalizes the result that partitions of n grow subexponentitally in n (see Proposition 2.10).
Proposition 3.6. For integers N and C, let T (N, C) denote the number of ways that N balls may be first
each coloured by one of C colours, and then partitioned into any number of multisets. Balls of the same
colour are indistinguishable. For every fixed C the sequence T (N, C) grows subexponentially in N .
For example, T (3, 2) = 14, corresponding to the 14 coloured partitions
{•, •, •}, {•, •, ◦}, {•, ◦, ◦}, {◦, ◦, ◦}, {•, •} ∪ {•}, {•, •} ∪ {◦}, {•, ◦} ∪ {•}, {•, ◦} ∪ {◦},
{◦, ◦} ∪ {•}, {◦, ◦} ∪ {◦}, {•} ∪ {•} ∪ {•}, {•} ∪ {•} ∪ {◦}, {•} ∪ {◦} ∪ {◦}, {◦} ∪ {◦} ∪ {◦}.
13
3.2 Convergence for subquotients of FIW -algebras generated in low degree
R Jiménez Rolland
J C H Wilson
Proof. Suppose throughout that C is fixed. We will create a code with a codeword recording each
partition of N balls coloured with C colours. To verify that T (N, C) is subexponential in N , it
suffices to check that the maximal length of these codewords is sublinear in N .
Consider a partition P = {P1 , P2 , . . . , Pℓ } of N coloured balls. We will create a distinct coding
scheme for (i) the “small” parts Pi containing at most s := (log N )2 balls, and (ii) the “large” parts
with more than s balls.
(i) (Small parts). There are s+C
possible sets of at most s balls coloured by C colours (this is
C
the equal to the number of monomials of C variables of degree at most s). Each of these sets
appears as a part Pi in our
coloured partition P at most N times. We can then encode these
s+C
sets by an ordered C –tuple of integers from 0 to N , each recording the number of times
the corresponding set appears in P . The number of characters needed to encode this tuple is
asymptotically bounded in N by the function
s+C
(log N )
C
< (s + C)C (log N )
= ((log N )2 + C)C (log N ).
(ii) (Large parts). There are fewer than Ns parts Pi in P with cardinality strictly greater than s.
For each of these parts, we record an ordered C–tuple of integers from 0 to N encoding the
number of balls of each given colour in that part. The total number of characters needed to
do this is asymptotically bounded by
N
CN
C
(log N ) =
s
(log N )
Combining (i) and (ii) we find that the maximal length of a codeword is asymptotically bounded
by
(log N )2 + C
C
(log N ) +
CN
.
log N
We conclude that log(T (N, C)) grows sublinearly in N , and so T (N, C) is subexponential in N as
claimed.
Having established a bound on the growth of T (N, C), we now need one final result in order
to prove Theorem 3.5.
Lemma 3.7. Let Wn denote one of the families Sn or Bn . For nonnegative integers b, c, define a graded
FIW –module V ∼
= MW (0)⊕b ⊕ MW (1)⊕c over a subfield k of C with positive gradings. Let Γ∗ denote the
free symmetric, exterior, or graded–commutative FIW –algebra generated by V . Let A∗ be any FIW –algebra
subquotient of Γ∗ . Then for each a ≥ 0 there is a function Fa (d) that is independent of n and subexponential
in d so that
dimk (Adn )Bn−a ≤ Fa (d)
for all n and d.
14
3.2 Convergence for subquotients of FIW -algebras generated in low degree
R Jiménez Rolland
J C H Wilson
It follows that, for any Wn character polynomial P , there exists a function FP (d) independent of n and
subexponential in d such that
|hPn , Adn iWn | ≤ FP (d).
More generally, if {Adn | d, n ≥ 0} is any collection of Wn –representations such that Adn is an Wn –
equivariant subquotient of Γdn , then there exists a function FP (d) as above.
Proof. The proof in Types A and B/C are extremely similar; we describe the proof in detail for the
hyperoctahedral groups and then briefly outline how to adapt the proof to the symmetric groups.
By Lemma 3.4 it suffices to show the first statement: for each a ≥ 0 there is a function Fa (d) that is
independent of n and subexponential in d so that
dimk (Adn )Bn−a ≤ Fa (d)
for all n and d.
Fix a ≥ 0. By assumption Adn is a Wn –subquotient of the Wn –representation Γdn . Since taking
Bn−a –invariants is exact in characteristic zero, the graded algebra (A∗n )Bn−a is a subquotient of
(Γ∗n )Bn−a . It suffices then to find a function Fa (d) that is independent of n and subexponential in d
so that
for all n and d.
dimk (Γdn )Bn−a ≤ Fa (d)
The monomials of Γdn are (graded– or anti–) commutative words in the (b + 2nc) variables
y (1) , y (2) , . . . , y (b)
(1)
(1)
(1)
(1)
∈ MBC (0)⊕b
n
(1)
(c)
(c)
and
(c)
(c)
x1 , x1 , x2 , x2 , . . . , x(1)
n , xn , . . . . . . , x1 , x1 , . . . , xn xn
∈ MBC (1)⊕c
n .
By assumption each variable has positive graded-degree, so the degree-d monomials must have
length at most d. The subgroup Bn−a acts on these monomials diagonally by signed permutations
on the subscript indices
{a + 1, a + 1, . . . , n, n}.
Superscripts are fixed. To bound the number of Bn−a orbits of monomials, we will in fact bound
the (larger) number of Sn−a orbits; Sn−a acts by permuting the set {a + 1, . . . , n} and simultaneously permuting the set {a + 1, . . . , n}. We can classify Sn−a –orbits of these monomials as follows:
Each monomial contains a (possibly empty) subword in the (b + 2ac) variables
y (1) , y (2) , . . . , y (b)
(1)
(1)
(1)
(1)
∈ MBC (0)⊕b
n
(c)
(1)
(c)
and
(c)
(c)
∈ MBC (1)⊕c
x1 , x1 , x2 , x2 , . . . , x(1)
n
a , xa , . . . . . . , x1 , x1 , . . . , xa xa
There are (b+2ac)+d
monomials of length at most d in the first (b + 2ac) variables. For fixed a, b, c
(b+2ac)
this formula grows polynomially in d.
(j)
Now, consider each of the remaining variables xi to be ’coloured’ by one of the (2c) symbols
(j)
(j)
x+ or x− , where the sign represents whether i is positive or negative, for j = 1, . . . , c. Next we
partition these coloured variables into multisets with one set for each subscript i. The action of
Sn−a preserves this decomposition.
Concretely, for example, if a = 2 then the Sn−a –orbit of the monomial
3
2
2
2
3
(3)
(5)
(3) (4)
(1)
(1)
(4)
(3) (4) (2)
(2) (1) (3)
x5 x5
x3 x3 x4
x2 x2 x3 x3
x6 x8 x8
y (2) y (4) y (5) x1
15
3.2 Convergence for subquotients of FIW -algebras generated in low degree
would be represented by the monomial y (2)
(2)
(4)
(4)
(3)
(4)
(1)
(1)
(1)
3
R Jiménez Rolland
J C H Wilson
2
(1)
(3) (4)
y (4) y (5) x1
x2 x2 and the coloured partition
(5)
(3)
(3)
(2)
(1)
(3)
{x+ , x+ , x+ , x− , x− } ∪ {x+ , x+ , x+ } ∪ {x+ , x− , x− } ∪ {x− } ∪ {x+ , x+ }.
Given monomials of N variables, the number of coloured partitions we may obtain in this
fashion is, in the notation of Proposition 3.6, T (N, 2c), which we proved in Proposition 3.6 grows
subexponentially in N for each fixed c.
To get an upper bound for the number of Bn−a –orbits for the monomials in Γ∗ of degree d,
we will bound the number of Bn−a –orbit of monomials in at most d variables. Some of these
monomials will have degree less than d. If some variables anticommute, then some monomials
will be zero in Γd and some monomials will vanish on passing to coinvariants, so we obtain a
strict overcount.
# of Bn−a –orbits of monomials in at most d variables
X # degree j monomials
# Bn−a -orbits of degree k
=
(c)
monomials in remaining variables
in y (1) , . . . , xa
j,k
j+k≤d
X
# degree j monomials
# Bn−a -orbits of degree d
≤
(c)
monomials in remaining variables
in y (1) , . . . , xa
j,k
j+k≤d
(b + 2ac) + d
≤
T (d, 2c)
(b + 2ac)
We thus obtain a bound that (for constant a, b, c) is independent of n and subexponential in d.
The proof in the case of the symmetric group is similar and slightly simpler: our Sn−a orbits
may be represented by a submonomial in the (b + ac) variables
y (1) , y (2) , . . . , y (b)
(1)
∈ M (0)⊕b
n
(c)
(1)
(c)
(c)
x1 , x2 , . . . , x(1)
a , . . . . . . , x1 , x2 . . . , xa
and
∈ M (1)n⊕c
and a partition that is coloured by the c colours j = 1, 2, . . . c. We obtain an asymptotic bound
(b + ac) + d
T (d, c)
(b + ac)
that is independent of n and subexponential in d.
Proof of Theorem 3.5. By Lemma 3.3, the limit in the numerator limn→∞ hPn , Adn iWn exists for every
d and moreover equals hPN , AdN iWN for N sufficiently large. By Lemma 3.7, there exist a function
FP (d) subexponential in d such that
| lim hPn , Adn iWn | ≤ FP (d).
n→∞
It follows that the sum
X limn→∞ hPn , Ad iW
n
d=0
converges absolutely.
qd
16
n
R Jiménez Rolland
J C H Wilson
Point counts for maximal tori in type B and C
op
Remark 3.8. (Convergence for graded FIW –algebras generated in low degree). Recall the FIW –
op
module V ∼
= MW (0)⊕b ⊕ MW (1)⊕c of Theorem 3.5. The module V has a natural FIW –module
structure defined (in the notation of Definition 2.7) by the maps
MW (1)n+1 −→ MW (1)n
ei , |i| = 1, . . . , n
ei 7−→
0, |i| = n + 1.
and the isomorphisms MW (0)n+1 ∼
= MW (0)n . (See Church–Ellenberg–Farb [CEF15, Section 4.1]
op
and Wilson [Wil15, Section 3] for details on the structure of simultaneous FIW and FIW –modules.)
op
These maps endow Γ∗ with the structure of an FIW –algebra. Since Wn –representations are selfdual, the representations Γ∗n are isomorphic whether viewed as representations of EndFIW (n) ∼
=
op
Wn or EndFIop (n) ∼
Wn . If A is any FIW –algebra subquotient of Γ∗ , then in particular the bigraded
=
W
piece Adn is a subquotient of Γdn , and Theorem 3.5 applies: for any character polynomial P and q > 1
the following sum converges absolutely
∞
X
limn→∞ hPn , Ad iW
n
d=0
qd
n
.
Remark 3.9. After circulating a preprint of this paper we discovered that in [FW, Theorem 3.2 and
Corollary 3.3], Farb and Wolfson prove a result which is closely related to our Theorem 3.5.
Theorem 3.10 ([FW, Theorem 3.2 and Corollary 3.3]). Let X be a connected space such that dimH ∗ (X; Q) <
∞. Then there exist constants K, L > 0 so that for each i ≥ 0, and for all n ≥ 1, the Betti numbers of the
n-fold symmetric product of X are bounded subexponentially:
√
dimH i (Symn (X); Q) < KeL i .
For each 0 ≤ a ≤ n exist constants K, L > 0 so that for each i ≥ 0, and for all n ≥ 1, the coinvariants of
the cohomology of X n are bounded subexponentially:
√
dim H i (X n ; Q)Sn−a < KeL i .
Farb–Wolfson use methods that are quite different from those of Theorem 3.5: their proof uses
a result of Macdonald on the Poincaré polynomial of symmetric products [Mac62], and an analysis
of this generating functions drawing on complex analysis.
It should be possible to deduce the subexponential growth of the Betti numbers and their coinvariants in Theorem 3.10 from Theorem 3.5 and its proof; it would take additional work to conclude their precise asymptotic formulas. Conversely, by applying Theorem 3.10 with appropriately chosen spaces X and using the Künneth Formula, it should be possible to deduce Theorem
3.5 in the case of graded-commutative algebras in type A. With some additional work it should be
possible to leverage these results to give an alternate route to proving Theorem 3.5 in all cases.
4
Point counts for maximal tori in type B and C
In this section we obtain the type B and C analogues of [CEF15, Theorem 5.6], a convergence
result for certain statistics on tori in GLn (Fq ). We combine a theorem of Lehrer with representation
17
R Jiménez Rolland
J C H Wilson
4.1 Maximal tori in types B and C
stability results of the second author and our discussion in Section 3 to compute asymptotic counts
for Frq -stable maximal tori for the symplectic and special orthogonal groups.
4.1
Maximal tori in types B and C
Let k be an algebraically closed field. We denote by G one of the following two semisimple connected linear algebraic groups defined over k: the symplectic group Sp2n (type C) or the special
orthogonal group SO2n+1 (type B). In the discussion below we identify the algebraic group G with
its set of k-rational points G(k). Recall that the symplectic group is the algebraic group defined
over k by
Sp2n (k) = {A ∈ GL2n (k) : AT JA = J},
0
In
is the matrix associated to the standard symplectic form
where J =
−In 0
ω(~u, ~v ) = u1 vn+1 + . . . + un v2n − un+1 v1 − . . . − u2n vn , for ~u, ~v ∈ k2n .
The special orthogonal group is defined as
1 0
where Q = 0 0
0 In
SO2n+1 (k) = {A ∈ SL2n+1 (k) : AT QA = Q},
0
In is the matrix associated to the quadratic form of Witt index n
0
q(u0 , u1 , . . . , u2n ) = u20 + u1 un+1 + . . . + un u2n ,
for
(u0 , u1 , . . . , u2n ) ∈ k2n+1 .
A subgroup of G is called a torus if it is k-isomorphic to a product of multiplicative groups
Gm := GL1 . For instance, we define standard maximal tori T0 to be the diagonal subgroups
×
−1
⊆ Sp2n (k)
T0 := diag(λ1 , . . . , λn , λ−1
1 , . . . , λn ) : λi ∈ k
and
×
−1
⊆ SO2n+1 (k).
T0 := diag(1, λ1 , . . . , λn , λ−1
1 , . . . , λn ) : λi ∈ k
In both cases, the maximal torus T0 has dimension n.
Let T denote the set of maximal tori of G. Since all maximal tori in G are conjugate ([Sri79,
Proposition 1.1]), the action of G on T by conjugation is transitive and T ∼
= G/N (T0 ), where
N (T0 ) is the normalizer of the torus T0 . In type B and C, the Weyl group W (T0 ) := N (T0 )/T0 is
isomorphic to the hyperoctahedral group Bn . The group Bn acts on matrices in T0 by conjugation
by permuting the n eigenvalues λ1 , . . . , λn and transposing the inverse pairs λi and λ−1
i .
4.1.1 The action of Frobenius
Let q be an integral power of a prime p and let k = Fp . The standard Frobenius morphism Frq acts
on a matrix (xij ) ∈ G = G(Fq ) by Frq : (xij ) 7→ (xqij ). The set of fixed points GFrq := {g ∈ G :
Frq (g) = g} corresponds to the Fq -points G(Fq ) of G: the finite groups Sp2n (Fq ) and SO2n+1 (Fq ).
A maximal torus of GFrq is a subgroup of G(Fq ) of the form T Frq = {g ∈ T : Frq (g) = g} for some
Frq -stable maximal torus T of G. In particular, since T0 is Frq -stable,
18
R Jiménez Rolland
J C H Wilson
4.1 Maximal tori in types B and C
Frq
T0
and
Frq
T0
−1
×
is a maximal torus of Sp2n (Fq )
= diag(λ1 , . . . , λn , λ−1
1 , . . . , λn ) : λi ∈ Fq
−1
×
is a maximal torus of SO2n+1 (Fq ).
= diag(1, λ1 , . . . , λn , λ−1
1 , . . . , λn ) : λi ∈ Fq
Other examples of Frq -stable maximal tori in Sp2 (Fp ) are given in Example 4.1 below.
An Frq -stable torus of G is defined over Fq . We say that such a torus T splits over Fq if T is
Fq -isomorphic to a product of multiplicative groups Gm . There is always a finite Galois extension
of Fq over which a given torus becomes diagonalizable, hence a split torus. The maximal torus T0
above splits over Fq (or any field) and the group G is said to be a split algebraic group. We will
investigate statistics on the set T Frq of Frq -stable maximal tori of G. For an introduction to split
reductive groups and maximal tori we refer the reader to [Mil12, Chapter I] and [Car93, Chapters
1 & 3]. See also Niemeyer–Praeger [NP10, Section 3] for a description of the maximal tori in finite
classical groups GFrq of Lie type.
4.1.2 Frq -stable maximal tori and characters of the Weyl group
The G(Fq )-conjugacy classes in T Frq correspond to conjugacy classes in the Weyl group Bn . Lehrer
observed that this implies that, in principle, functions which are defined on Frq -stable maximal tori
(for instance, functions which count rational tori) may be described in terms of the character theory
of Bn . This correspondence between G(Fq )-conjugacy classes in T Frq and conjugacy classes in Bn
is defined as follows. Consider a torus T ∈ T Frq , so T = gT0 g −1 for some g ∈ G . Since
gT0 g −1 = T = Frq (T ) = Frq (g)Frq (T0 )Frq (g)−1 = Frq (g)T0 Frq (g)−1
it follows that g −1 ·Frq (g) ∈ N (T0 ). We denote by wT the element in the Weyl group W (T0 ) ∼
= Bn
that is given by the projection of w̃T = g −1 · Frq (g) onto the quotient W (T0 ) = N (T0 )/T0 . Since
G is split, it turns out that each Frq -stable maximal torus determines a conjugacy class in the Weyl
group Bn ([Car93, Proposition 3.3.2]). Hence, given a class function χ on Bn , we can regard χ as a
function on Frq -stable maximal tori of G, by defining
χ(T ) := χ(wT ),
for T ∈ T Fq .
This correspondence between conjugacy classes of tori and conjugacy classes in the Weyl group is
illustrated in Example 4.1 for Sp2 (Fp ).
Example 4.1. (Frq -stable maximal tori of Sp2 ). For the algebraic group Sp2 (Fp ) = SL2 (Fp ), the
Frq -stable maximal torus
×
λ1
0
F
T0 =
λ
∈
p
1
0 λ1−1
1 0
is a split torus corresponding to the identity coset
T0 in the Weyl group W (T0 ), that is,
0 1
the identity element of the Weyl group B1 . On the other hand, given an element ǫ ∈ F×
q which is
not a square in Fq , consider the abelian subgroup of SL2 (Fp )
x y
Tǫ =
x, y ∈ Fp , x2 − ǫy 2 = 1 .
ǫy x
19
R Jiménez Rolland
J C H Wilson
4.1 Maximal tori in types B and C
Choose a square root
√
ǫ ∈ Fp of ǫ. If we take g =
g
−1
·
x
ǫy
y
x
·g =
1
√
2 ǫ
√1
ǫ
√
x+y ǫ
0
−1
√
ǫ
∈ SL2 (Fp ), then
0√
x−y ǫ
,
therefore g −1 Tǫ g = T0 and Tǫ is a Frq -stable maximal torus. The matrix
√
1
ǫ
1
1
−1
0
−1
−1
√
√ =
√
w̃Tǫ = g · Frq (g) = √
−1 0
2 ǫ − ǫ 1 − ǫ − ǫ
0 1
projects to the coset
T0 in the Weyl group W (T0 ). Hence wTǫ is the negative 1-cycle
1 0
(1 1) ∈ B1 .
4.1.3 The coinvariant algebra and statistics on maximal tori
There is a beautiful connection between the conjugacy-invariant functions χ(T ) on the maximal
tori of G defined over Fq , and the topological structure of certain complex varieties related to G(C).
We now describe the complex cohomology algebra of these varieties, the generalized complete
complex flag manifolds in type B and C.
Let V ∼
= Cn denote the canonical complex representation of the hyperoctahedral group Bn by
signed permutation matrices. Then the symmetric powers Sym(V ) are isomorphic to C[x1 , . . . , xn ],
and we denote by In the homogeneous ideal generated by the Bn –invariant polynomials with
constant term zero. The complex type B/C coinvariant algebra is defined as
Rn ∼
= C[x1 , . . . , xn ]/In
Let Rnd denote the dth graded piece of Rn .
Borel [Bor53] proves that Rn∗ is isomorphic as a graded C[Bn ]-algebra to the cohomology of the
generalized complete complex flag manifolds in type B and C. The cohomology groups are supported
in even cohomological degree; this isomorphism multiplies the grading by 2. We recall the definition of these varieties: in type B, let BB
n be a Borel subgroup of SO2n+1 (C). Then the associated
generalized flag manifold is
2n+1
SO2n+1 (C)/BB
| dimC Vm = m, q(Vi , V2n+1−i ) = 0},
n = {0 ⊆ V1 ⊆ . . . ⊆ V2n+1 = C
the variety of complete flags equal to their orthogonal complements. In type C, take a Borel subgroup BC
n of Sp2n (C). The associated generalized flag manifold is
2n
Sp2n (C)/BC
n = {0 ⊆ V1 ⊆ . . . ⊆ V2n = C | dimC Vm = m, ω(Vi , V2n−i ) = 0},
the variety of complete flags equal to their symplectic complements.
A result of Lehrer [Leh92, Corollary 1.10’] relates the characters of Rnd to the space of maximal
tori. In the case of the split reductive groups Sp2n and SO2n+1 defined over Fp with 2n2 roots,
Lehrer’s result specializes to the following formula.
20
R Jiménez Rolland
J C H Wilson
4.1 Maximal tori in types B and C
Theorem 4.2 (Corollary 1.10’ [Leh92]). Let q be an integral power of a prime p and let G be the linear
algebraic group Sp2n (Fp ) or SO2n+1 (Fp ). If χ is a class function on Bn , then
X
T ∈T
2
χ(T ) = q 2n
Frq
X
q −d χ, Rnd
Bn
,
(1)
d≥0
where T Frq is the set of Frq -stable maximal tori of G and Rnd is the dth -graded piece of the complex coinvariant algebra Rn in type B/C.
We remark that, in this formula, Rnd is generated in graded degree d = 1, and not (as with the
cohomology algebra) supported only in even degrees.
Formula (1) in Theorem 4.2 describes a deep relationship between the maximal tori over finite
fields and the cohomology of the topological spaces related to G(C); with these results we will (in
Sections 4.2 and 4.3) relate representation-theoretic stability results on the coinvariant algebras to
stability results for point-counts on our varieties over finite fields.
4.1.4 Character polynomials and interpreting statistics
When χ is a character polynomial, the left-hand side of Formula (1) is, in a sense, quantifying
the numbers and decompositions of maximal tori over Fq . To make this precise, we first consider
the statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq ) that correspond to the class
functions Xr and Yr .
Let T ∈ T Frq as before with T0 = g −1 T g for g ∈ G. Consider the ordered set of eigenvectors
v1 , . . . , vn , v 1 , . . . , v n of T given by the columns of g. Since T is defined over Fq , the Frobenius
2n
morphism Frq takes eigenvectors to eigenvectors. Then Frq acts on the set of lines in Fq
LT = {L1 , . . . , Ln , L1 , . . . , Ln }
where Li := spanFq (vi ) and Li := spanFq (v i ).
Moreover, since Frq (g) = g w̃T , the Frobenius morphism acts on LT by the element wT ∈ Bn that
permutes the lines Li and swaps pairs Li /Li .
Let T ∈ T Fq . For each positive or negative r-cycle of wT , we can consider its corresponding
support {Li1 , . . . , Lir , Li1 , . . . , Lir } in LT and the subtorus of T0
×
−1
.
T0r := diag(1, . . . λi1 , . . . , λir , . . . , 1, 1, . . . λ−1
i1 , . . . , λir , . . . , 1) : λij ∈ Fq
Then Tr := gT0r g −1 is an Frq -stable r-dimensional subtorus of T irreducible over Fq . A torus
defined over a field k is irreducible if it is not isomorphic over k to a product of tori. Furthermore:
• If the orbit {Li1 , . . . , Lir , Li1 , . . . , Lir } corresponds to a positive r-cycle of wT , then (Frq )r =
Frqr fixes each Lij and each Lij . This means that there is a matrix gr with entries in Fqr such
that gr−1 Tr gr = T0r and the subtorus Tr splits over Fqr .
• If the orbit {Li1 , . . . , Lir , Li1 , . . . , Lir } corresponds to a negative r-cycle of wT , then (Frq )r =
Frqr swaps the lines Lij and Lij . This implies that no matrix gr with entries in Fqr diagonalizes the subtorus Tr and hence Tr does not split over Fqr .
21
R Jiménez Rolland
J C H Wilson
4.2 Asymptotic results
Therefore, if T ∈ T Fq , for r ≥ 1 the polynomial characters Xr (T ) and Yr (T ) count the following:
Xr (T ) is the number of r-dimensional Frq -stable subtori of T irreducible over Fq that split over Fqr ,
Yr (T ) is the number of r-dimensional Frq -stable subtori of T irreducible over Fq that do not split
over Fqr .
Example 4.1 continued (Statistics for maximal tori in Sp2 ). Recall the maximual torus Tǫ ∈
Sp2 (Fp ) from Example 4.1:
x y
2
2
Tǫ =
x, y ∈ Fp , x − ǫy = 1 ,
ǫy x
1 −1
1
√
√
√
in Sp2 (Fp ), and is invariant under the
The torus Tǫ is diagonalized by the matrix 2 ǫ
ǫ
ǫ
action of Frobenius. It corresponds to the nontrivial conjugacy class wTǫ = (1 1) ∈ B1 .
The Frobenius morphism Frq acts by transposing the eigenspaces
−1
√1
√
and L1 = spanFp
.
L1 = spanFp
ǫ
ǫ
Correspondingly, no matrix that diagonalizes Tǫ can have entries in Fq . The one–dimensional torus
Tǫ is irreducible and does not split over Fq . This is consistent with its statistics
X1 (Tǫ ) = X1 ((1 1)) = 0
4.2
and
Y1 (Tǫ ) = Y1 ((1 1)) = 1.
Asymptotic results
In this section we prove Theorem 4.3, a stability result for asymptotic polynomial statistics on
Frq -stable maximal tori of the symplectic and special orthogonal groups.
Theorem 4.3 (Stability of maximal tori statistics). Let q be an integral power of a prime p. For n ≥ 1,
Fr
d
denote
denote by Tn q the set of Frq -stable maximal tori for either Sp2n (Fp ) or SO2n+1 (Fp ). Let Rm
th
∗
the d -graded piece of the complex coinvariant algebra Rm in type B/C. If P ∈ C[X1 , Y1 , X2 , Y2 , . . .] is
2 P
any hyperoctahedral character polynomial, then the normalized statistic q −2n
Fr P (T ) converges as
T ∈Tn q
n → ∞. In fact,
∞
d
X
X
2
limm→∞ hPm , Rm
iBm
lim q −2n
P (T ) =
,
d
n→∞
q
Frq
T ∈Tn
d=0
and the series in the right hand converges.
To prove this theorem we first establish a convergence result for characters of coinvariant algebras of type B/C.
Let Cn∗ denote the complex polynomial algebra C[x1 , . . . , xn ] on n variables, with generators xi
in graded degree d = 1. These polynomial rings form a complex FIBC –algebra under the natural
∗
inclusions Cn∗ ֒→ Cn+1
, which is generated as an FIBC –algebra in FIBC –degree n = 1 by C11 = hx1 i.
22
R Jiménez Rolland
J C H Wilson
4.2 Asymptotic results
Our sequence of algebras is then generated by the FIBC –module of canonical signed permutation
representations,
C•1 ∼
= MBC (∅, )• := MBC (1)• ⊗C[B1 ] V(∅, ) ∼
= hx1 , . . . , x• i.
The coinvariant algebra Rn∗ ∼
= Cn∗ /In does not admit an FIBC –module structure; the FIBC –module
structure on the polynomial rings Cn∗ does not respect the ideals In . Instead, the coinvariant algeop
bras are C–algebras over FIBC , induced by the maps
∗
Cn+1
−→ Cn∗
xi , i = 1, . . . , n
xi 7−→
0, i = n + 1.
cd of each graded piece do form finitely generated sub–
The reader may verify that the duals R
•
d
cd ∼
FIBC –modules of the dual modules C
• = C• ; see work of the second author [Wil14, Definition
5.14, Proposition 5.15, and Section 6] for details. For the following result, however, we do not
need this FIBC –module structure; we only need the observation that for each n and d, the Bn –
representation Rnd is a subquotient of the homogeneous polynomial space Cnd .
Proposition 4.4. Let Rnd the dth –graded piece of the complex type B/C coinvariant algebra Rn . Then for any
hyperoctahedral character polynomial P ∈ C[X1 , Y1 , X2 , Y2 , . . .] the following sum converges absolutely.
∞
X
limn→∞ hPn , Rd iB
n
qd
d=0
n
.
op
Proof. We have observed that Rn∗ is an FIBC –algebra quotient of the free graded-commutative
algebra C ∗ on
MBC (∅, )• := MBC (1)• ⊗C[B1 ] V(∅, )
The FIBC –module MBC (∅, )• is, in the notation of Definition 2.7, the submodule of MBC (1) on
the generators xi := ei − ei . By Remark 3.8, the result follows from Theorem 3.5.
Remark 4.5. This result was proved by Church–Ellenberg–Farb in type A [CEF14, Theorem 5.6]
using in part work of Chevalley [Che55] on the structure of the type A coinvariant algebra. Our
Theorem 3.5 allows for a combinatorial proof of [CEF14, Theorem 5.6] using only the fact that the
coinvariant algebras are quotient of the dual of the free commutative FI–algebra C[x1 , x2 , . . . , xn ]
generated by M (1)n = hx1 , x2 , . . . , xn i. This proof demonstrates that this convergence theorem
does not depend on any deeper structural features of the coinvariant algebra.
Remark 4.6. Work of the second author [Wil14, Theorem 6.1 and Corollary 6.3] observes that, for
cd dual to Rd assemble to form a finitely genereach graded–degree d, the Bn –representations R
n
n
cd , a submodule of the sequence of homogenous polynomials C
cd . The FIBC –
ated FIBC –module R
•
•
c
c
d
d
module C• , and hence R• , has weight at most d. Hence for each d there is some Dd ∈ Z such
cd are given by a character polynomial of degree ≤ d for all n ≥ Dd [Wil15,
that the characters of R
•
Theorem 4.16]. Since Bn –representations are self-dual, there is an isomorphism of representations
d
d
cd ∼
R
n = Rn , and this character polynomial gives the characters of Rn for all n ≥ Dd . It follows from
Proposition 3.1 that
d
lim hPm , Rm
iBm = hPn , Rnd iBn ,
for any n ≥ max{Dd , d + deg(P )}.
m→∞
23
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
R Jiménez Rolland
J C H Wilson
Proof of Theorem 4.3. We adapt the arguments given by Church–Ellenberg–Farb [CEF14, Theorem
3.13]. From Lemma 3.7, there exist a function FP (d) independent of n and subexponential in d
such that |hPn , Rnd iBn | ≤ FP (d) for all n. Then
d
lim hPm , Rm
iBm ≤ FP (d).
m→∞
Let ǫ > 0. The series
X FP (d)
d≥0
qd
converges absolutely, so there exist some I ∈ N such that
X FP (d)
< ǫ/2.
qd
d≥I+1
Using the notation from Remark 4.6, we choose N = max{D1 , D2 , . . . , DI , I + deg(P )}. Then
d
lim hPm , Rm
iBm = hPn , Rnd iBn
m→∞
From Proposition 4.4, the series
for d ≤ I and n≥ N .
∞
d
X
limm→∞ hPm , Rm
iBm
qd
d=0
converges to a limit L < ∞. On the other hand, by Theorem 4.2, we have that
2
q −2n
X
P (T ) =
Frq
X hPn , Rd iB
n
d≥0
T ∈Tn
qd
n
.
Therefore, if n ≥ N
2
L − q −2n
X
X limm→∞ hPm , Rd iB − hPn , Rd iB
n n
m m
qd
P (T ) =
Frq
d≥I+1
T ∈Tn
≤
d≥I+1
<ǫ
4.3
X FP (d) + FP (d)
qd
Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
In this section we compute some examples of statistics of the form given in Theorem 4.3. We use a
result on the decomposition of the coinvariant algebra Rn∗ as a Bn –representation; see Stembridge
[Ste89, Formula (2.2), Theorem 5.3], analogous to the approach taken by Church–Ellenberg–Farb
[CEF14, Theorems 5.8, 5.9, and 5.10]. These computations could also be accomplished with generating functions, using the approach of Fulman [Ful16] in type A.
24
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
R Jiménez Rolland
J C H Wilson
Decomposing the Bn –representation Rnd . The decomposition of the graded pieces Rnd into irreducible Bn –representations is described by Stembridge [Ste89] in terms of their relationships to
data called the fake degrees of Bn . Given a double partition λ = (λ+ , λ− ) of n, the multiplicity of irreducible representation Vλ in Rnd is computed by the following formula. A double standard tableau
of shape λ is a bijective labelling of the Young diagrams λ+ and λ− with the digits 1 through n,
such that the numbers are strictly increasing in each row and column. Define the flag descent set
D(T ) of a double standard tableau T as follows: draw the tableaux of shape λ+ and λ− in the
plane, with the tableau of shape λ− placed above and to the right of tableau of shape λ+ . Then
D(T ) = { j
| (j + 1) appears in a row below j
}.
and the flag major index of the double standard tableau T is the quantity
X
j + |λ− |.
f (T ) = 2
j∈D(T )
Then the multiplicity of the irreducible representation Vλ in the dth -graded piece of the coinvariant
algebra Rnd is given by the number of standard double tableaux of shape λ and flag major index d.
Theorem 4.7 (See Stembridge [Ste89] Formula (2.2), Theorem 5.3).
hRnd , Vλ iBn = # { T
|
T standard double tableau of shape λ with f (T ) = d
}.
We can use Theorem 4.7 to compute some examples of the normalized statistics for maximal
tori whose convergence is guaranteed by Theorem 4.3. A formula identifying stable sequences of
irreducible Bn –representations Vλ with a character polynomial P is given in Wilson [Wil15, Theorem 4.11]. Some stable multiplicities that we compute below, using Theorem 4.7, are summarized
in Table 3. Dots represent zero.
Our first computation is a classical result due to Steinberg (see [Ste68, Corollary 14.16]).
Theorem 4.8 (Number of Frq -stable maximal tori). The number of Frq -stable maximal tori of Sp2n (Fq )
2
and of SO2n+1 (Fq ) is q 2n .
Proof. We can compute the number of Frq -stable maximal tori using Theorem 4.2 by taking χ to be
the trivial class function on Bn . The sequence of trivial representations are given by the character
!
n
z }| {
polynomial χ = 1, and are encoded by the double partition λ =
· · · , ∅ . The single
double standard tableau of shape λ is
2
1
T =
3
···
n
, ∅
which has descent set D(T ) = ∅ and flag major index f (T ) = 0. Thus by Theorem 4.7 the trivial
representation appears with multiplicity one in degree 0 and does not occur in positive degree.
25
R Jiménez Rolland
J C H Wilson
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
Table 3: Some stable multiplicities of irreducible Bn –representations in the coinvariant alegebra
Irreducible Bn
representation
Hyperocthahedral
character polynomial
d=0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
V ((n), ∅)
1
1
·
·
·
·
·
·
·
·
·
·
·
·
·
·
·
V ((n − 1), (1))
X1 − Y1
·
1
·
1
·
1
·
1
·
1
·
1
·
1
·
1
X1 + Y1 − 1
·
·
1
·
1
·
1
·
1
·
1
·
1
·
1
·
·
·
1
·
1
·
2
·
2
·
3
·
3
·
4
·
·
·
·
·
1
·
1
·
2
·
2
·
3
·
3
·
·
·
·
·
·
·
1
·
1
·
2
·
2
·
3
·
·
·
·
1
·
2
·
3
·
4
·
5
·
6
·
7
V ((n − 1, 1), ∅)
V ((n − 2), (2))
V ((n − 2), (1, 1))
V ((n − 2, 1, 1), ∅)
V ((n − 2, 1), (1))
Y1
X1
− X1 Y1 + X2 − Y2
+
2
2
Y1
X1
− X1 Y1 − X2 + Y2
+
2
2
X1 + Y1
− X2 − Y2 − X1 − Y1 + 1
2
X12 − Y12 − 2X1 + 2Y1
Remark 4.9. Observe that the normalized formula
P
P
Fr P (T )
Fr P (T )
T ∈Tn q
T ∈Tn q
=
2
Fr
2n
q
|Tn q |
corresponds to the average of the statistic P (T ) over all maximal tori T . Using Theorem 4.3 we
obtain asymptotics of these expected values.
Proposition 4.10 (Expected number of 1-dimensional Frq -stable subtori). The expected number
of 1-dimensional Frq -stable subtori of a random Frq -stable maximal torus in Sp2n (Fq ) or in SO2n+1 (Fq )
equals
1
q 2 − q2(n−1)
1
1
1
1 + 2 + 4 + · · · + 2n−2 =
q
q
q
(q 2 − 1)
n→∞
−−−−→
q2
.
q2 − 1
Proof. To count the total number of 1-dimensional Frq -stable subtori for all maximal tori in T Frq , we
apply Formula (1) to the character polynomial P = X1 +Y1 . The goal is to show that P, Rnd B = 1
n
for d = 0, 2, 4, . . . , 2n − 2 and 0 otherwise.
The pullback of the (n − 1)–dimensional standard Sn –representation to Bn has character polynomial χ = X1 + Y1 − 1 and corresponds to the double partition
n−1
z }| {
···
, ∅ .
λ=
There are (n − 1) possible double standard tableaux of shape λ, each determined by the letter i in
26
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
the second row. Let
Ti =
1
2
···
i−1
i+1
···
i
n
,
∅
R Jiménez Rolland
J C H Wilson
for i = 2, . . . , n.
Note that we require i > 1 for Ti to form a valid standard tableau. Then D(Ti ) = {i − 1} for
i = 2, . . . , n. The flag major indices are f (Ti ) = 2i − 2 for i = 2, . . . , n, and there is one copy of
V(n−1,1),∅ in each even degree 2, 4, . . . , 2n − 2.
The character polynomial (X1 + Y1 ) is the sum (X1 + Y1 ) = (X1 + Y1 − 1) + 1, so combining
this result for multiplicities of the representation V((n−1,1),∅) and the result for V((n),∅) of Theorem
4.8 gives the desired formula.
Proposition 4.11 (Expected number of split 1-dimensional Frq -stable subtori). The expected number of split 1-dimensional Frq -stable subtori of of a random Frq -stable maximal torus in Sp2n (Fq ) or in
SO2n+1 (Fq ) equals
1
q − q2n−1
1
q
1
1
1
1
n→∞
1 + + 2 + 3 + · · · + 2n−1 =
−−−−→
.
2
q
q
q
q
2(q − 1)
2(q − 1)
Proof. We wish to evaluate the character polynomial P = X1 = 12 [(X1 + Y1 ) + (X1 − Y1 )] in
formula (1). The canonical n–dimensional representation of Bn by signed permutation matrices
has character polynomial χ = X1 − Y1 and corresponds to the double partition
n−1
z }| {
λ=
··· , .
There are n possible double standard tableaux of shape λ, each determined by the letter i in λ− .
Let
1
i
2 · · · i−1 i+1 · · · n
Ti =
for i = 1, . . . n.
,
Then D(Tn ) = ∅, and D(Ti ) = {i} for i = 1, . . . , n − 1. The flag major indices are
1,
i=n
f (Ti ) =
2i + 1, i = 1, 2, . . . , (n − 1)
and then there is a single copy of V((n−1),(1)) in each odd degree 1, 3, . . . , 2n − 1.
By combining this result with our computation for X1 + Y1 in Proposition 4.10 above, we conclude that 2X1 , Rnd Bn = 1 for d = 0, 1, 2, 3, . . . , 2n − 1 and 2X1 , Rnd Bn = 0 otherwise.
Corollary 4.12 (Expected number of eigenvectors in F2n
q ). The expected number of simultaneous
2n
eigenvectors in Fq of a random Frq -stable maximal torus in Sp2n (Fq ) equals
1+
1
1
1
1
+ 2 + 3 + · · · + 2n−1
q
q
q
q
27
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
R Jiménez Rolland
J C H Wilson
Proof. The fixed points of wT correspond to eigenvectors vi of T in F2n
q . Since the eigenvectors
come in inverse pairs (vi /v i ) for Frq -stable maximal tori in Sp(Fq ), we apply Formula (1) to the
character polynomial P = 2X1 . The result follows from the above computation.
Proposition 4.13 (Reducible v.s. Irreducible Frq -stable 2-dimensional subtori). Given a Frq -stable
torus T in Sp2n (Fq ) or in SO2n+1 (Fq ), let Rn (T ) denote the number of reducible 2-dimensional Frq -stable
subtori of T and In (T ) denote the number of irreducible 2-dimensional Frq -stable subtori of T . Then the
expected value of the function Rn − In over all Frq -stable maximal tori of Sp2n (Fq ) or SO2n+1 (Fq ) is given
by
q4 −
1
q2n
1−
1
q2(n−1)
(q 2 − 1) (q 4 − 1)
and converges to the sum
⌊ 2d−2
1
2
2
3
3
1
4 ⌋
+
+
+
+
+
+
.
.
.
+
+ ...
2
4
6
8
10
12
2d−4
q
q
q
q
q
q
q
=
q4
(q 2 − 1)(q 4 − 1)
as n tends to infinity.
Proof. To count the difference in the number of reducible and irreducible 2-dimensional Frq -stable
Fr
subtori of T ∈ Tn q , we compute
X1 (T ) + Y1 (T )
P (T ) =
− X2 (T ) + Y2 (T )
2
V2
V(n−1,1),∅ has character
and employ Formula (1). The irreducible Bn –representation
X1 + Y1
− (X2 + Y2 ) − (X1 + Y1 ) + 1
χ=
2
and associated double partition
Define
Ti,j
=
λ=
1
2
···
i−1
i+1
···
z
n−2
j−1 j+1
}|
···
···
i
j
{
, ∅
.
n
,∅
for i < j,
i, j ∈ {2, . . . , n}.
We have D(Ti,j ) = {i − 1, j − 1} for all 2 ≤ i < j ≤ n, and so f (Ti,j ) = 2(i + j) − 4 for all
2 ≤ i < j ≤ n. Stably, it follows that the multiplicity is zero in degree d if d is odd, and for d ≥ 6
28
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
R Jiménez Rolland
J C H Wilson
d
d+2
even, f (Ti,j ) = d for the ⌊ d−2
j) = (2, d2 ), (3, d2 − 1), . . . , (⌊ d+2
4 ⌋ pairs (i,
4 ⌋, 2 − ⌊ 4 ⌋ + 2). For a
X
2
given n ≥ 2, the expected value q −2n
χ(T ) of χ(T ) is given in terms of the Gaussian binomial
n
coefficient
2
Frq
T ∈Tn
by the formula
1
q2
X hχ, Rd iB
n
d≥0
qd
n
1 n
= 6
q 2
1
q2
1
1
1
1
1 − q2n
1 − q2(n−1)
1 − q2(n−1)
1 1 − q2n
=
:= 6
1
1
q
(q 2 − 1) (q 4 − 1)
1−
1−
q2
q4
As n tends to infinity, the expected value converges to
⌊ 2d−2
1
2
2
3
3
1
4 ⌋
+
+
+
+
+
+
.
.
.
+
+ ...
6
8
10
12
14
16
q
q
q
q
q
q
q 2d
=
(q 2
1
− 1)(q 4 − 1)
To obtain the result for the character polynomial
X1 + Y1
X1 + Y1
− (X2 + Y2 ) − (X1 + Y1 ) + 1 + (X1 + Y1 ) − 1
− (X2 + Y2 ) =
2
2
we combine this result with the result for irreducible representation V(n−1,1),∅ from Proposition
4.10. We find that the desired formula is
1
1
1
1 − q2n
q 2 − q2(n−1)
1 − q2(n−1)
+
−1
(q 2 − 1) (q 4 − 1)
(q 2 − 1)
1
1
1 − q2(n−1)
q 4 − q2n
=
(q 2 − 1) (q 4 − 1)
1
2
2
3
3
1
q4
n→∞
= 2 + 4 + 6 + 8 + 10 + 12 + . . .
−−−−→ 2
(q − 1) (q 4 − 1)
q
q
q
q
q
q
Proposition 4.14 (Split v.s. non-split Frq -stable irreducible 2-dimensional subtori). The expected
value of split minus non-split Frq -stable irreducible 2-dimensional subtori over all Frq -stable maximal tori
of Sp2n (Fq ) or SO2n+1 (Fq ) is given by
1
1
1 − q2(n−1)
q 2 1 − q2n
q2
n→∞
−
−
−
−
→
.
2 (q 4 − 1)
2 (q 4 − 1)
Proof. We need to consider the character polynomial X2 − Y2 in Formula (1).
The irreducible Bn –representation Λ2 V(n−1),(1) has character polynomial χ =
X1 Y1 − X2 + Y2 and corresponds to the double partition
n−2
z }| {
λ=
··· , .
29
Y1
X1
−
+
2
2
R Jiménez Rolland
J C H Wilson
4.3 Some statistics for Frq -stable maximal tori in Sp2n (Fq ) and SO2n+1 (Fq )
There are
n
possible double standard tableaux of shape λ. For i < j, with i, j ∈ {1, . . . , n}, let
2
Ti,j =
1
2
···
i−1
···
i+1
j−1 j+1
···
n
,
i
j
Then
D(Ti,n ) = {i}
D(Ti,j ) = {i, j}
f (Ti,n ) = 2i + 2
for
f (Ti,j ) = 2i + 2j + 2
for
i = 1, . . . , n − 1
1≤i<j ≤n−1
Hence for even d with 4 ≤ d ≤ 2n, f (T d−2 ,n ) = d, and for even d with 4 ≤ d ≤ 2n−2, f (Ti,j ) = d
2
d
for all i < j < n such that i + j = d2 − 1. Stably there are
copies of V(n−2),(1,1) in each even
4
degree d ≥ 4 and zero copies in odd degree. For given n ≥ 2, the expected value of the character
polynomial statistic is
1
1
1
1
−
1
−
X hχ, Rd iB
4
2n
q
q
1 n
q2(n−1)
n n
= 4
:=
1
qd
q 2 1
1−
1− 1
d≥0
q2
q2
n→∞
−−−−→
q4
2
q
(q 2 − 1) (q 4 − 1)
The irreducible Bn –representation with character polynomial χ =
and corresponds to the double partition
z
λ=
There are
Then
n−2
}| {
··· ,
Y1
X1
− X1 Y1 + X2 − Y2
+
2
2
.
n
possible double standard tableaux of shape λ. For i < j, with i, j ∈ {1, . . . , n}, let
2
1
2 · · · i−1 i+1 · · · j−1 j+1 · · · n
j
i
Ti,j =
,
D(Tn−1,n ) = ∅
f (Tn−1,n ) = 2
D(Ti,n ) = {i}
D(Ti−1,i ) = {i}
f (Ti,n ) = 2i + 2
f (Ti−1,i ) = 2i + 2
for
for
f (Ti,j+1 ) = 2i + 2j + 4
for
D(Ti,j+1 ) = {i, j + 1}
30
i = 1, . . . , n − 2
i = 2, . . . , n − 1
1≤i<j ≤n−2
R Jiménez Rolland
J C H Wilson
REFERENCES
Stably,
degrees and multiplicity
there this irreducible representation has multiplicity zero in odd
X
d
−2n2
in each even degree d ≥ 2. For n ≥ 2, the expected value q
χ(T ) is
4
Frq
T ∈Tn
X hχ, Rd iB
1 n
n n
= 2
qd
q 2
d≥0
:=
1
q2
1
q2
1
1
1
1
1 − q2n
1 − q2(n−1)
1 − q2(n−1)
q 4 1 − q2n
=
(q 2 − 1) (q 4 − 1)
1 − q12
1 − q14
n→∞
−−−−→
q4
(q 2 − 1) (q 4 − 1)
Then for the character polynomial
Y1
X1
Y1
X1
1
− X1 Y1 − X2 + Y2
+
− X1 Y1 + X2 − Y2 −
+
P = (X2 −Y2 ) =
2
2
2
2
2
X
2
P (T ) for n ≥ 2 is given by
the corresponding expected value q −2n
Frq
T ∈Tn
1
1
1
1
1
1
1 − q2(n−1)
1 − q2(n−1)
q4 1 − q2n
1 q2 1 − q2n
−
1
1
1
1
2
1 − q2
1 − q2
1 − q4
1 − q4
1
1
1 − q2(n−1)
q 2 1 − q2n
=
2 (q 4 − 1)
In the limit as n tends to infinity, this converges to
q2
.
− 1)
2 (q 4
References
[BGPW13] L. Babai, S. Guest, C. E. Praeger, and R. A. Wilson, Proportions of r-regular elements in finite classical
groups, Journal of the London Mathematical Society (2013), jdt011.
[Bor53]
A. Borel, Sur la cohomologie des espaces fibrés principaux et des espaces homogènes de groupes de Lie
compacts, Ann. of Math. (2) 57 (1953), 115–207. MR 0051508 (14,490e)
[Bri02]
J. R. Britnell, Cyclic, separable and semisimple matrices in the special linear groups over a finite field, J.
London Math. Soc. (2) 66 (2002), no. 3, 605–622. MR 1934295
[Car93]
R. W. Carter, Finite groups of Lie type, Wiley Classics Library, John Wiley & Sons, Ltd., Chichester, 1993, Conjugacy classes and complex characters, Reprint of the 1985 original, A WileyInterscience Publication.
[CEF14]
T. Church, J. Ellenberg, and B. Farb, Representation stability in cohomology and asymptotics for families
of varieties over finite fields, Algebraic topology: applications and new directions, Contemp. Math.,
vol. 620, Amer. Math. Soc., Providence, RI, 2014, pp. 1–54.
[CEF15]
, FI-modules and stability for representations of symmetric groups, Duke Math. J. 164 (2015),
no. 9, 1833–1910.
31
R Jiménez Rolland
J C H Wilson
REFERENCES
[CF13]
T. Church and B. Farb, Representation theory and homological stability, Adv. Math. 245 (2013), 250–
314.
[Che]
W. Chen, Twisted cohomology of configuration spaces and spaces of maximal tori via point-counting,
arXiv preprint:1603.03931.
[Che55]
C. Chevalley, Invariants of finite groups generated by reflections, American Journal of Mathematics
(1955), 778–782.
[DL76]
P. Deligne and G. Lusztig, Representations of reductive groups over finite fields, Ann. of Math. (2) 103
(1976), no. 1, 103–161.
[FNP05]
J. Fulman, P. M. Neumann, and C. E. Praeger, A generating function approach to the enumeration of
matrices in classical groups over finite fields, Mem. Amer. Math. Soc. 176 (2005), no. 830, vi+90.
[Ful99]
J. Fulman, Cycle indices for the finite classical groups, J. Group Theory 2 (1999), no. 3, 251–289. MR
1696313
[Ful16]
, A generating function approach to counting theorems for square-free polynomials and maximal
tori, Annals of Combinatorics 20 (2016), 587–599.
[FW]
B. Farb and J. Wolfson, Étale homological stability and arithmetic statistics, arXiv preprint: 1512.00415.
[GP00]
M. Geck and G. Pfeiffer, Characters of finite Coxeter groups and IwahoriHecke algebras, Oxford University Press, USA, 2000.
[HR18]
G. H. Hardy and S. Ramanujan, Asymptotic formulaæ in combinatory analysis, Proceedings of the
London Mathematical Society 2 (1918), no. 1, 75–115.
[JRW17]
R. Jiménez Rolland and Jennifer CH Wilson, Stability for hyperplane complements of type B/C and
statistics on squarefree polynomials over finite fields, arXiv preprint arXiv:1703.06349 (2017).
[Leh92]
G. I. Lehrer, Rational tori, semisimple orbits and the topology of hyperplane complements, Comment.
Math. Helv. 67 (1992), no. 2, 226–251.
[LNP09]
F. Lübeck, A. C. Niemeyer, and C. E. Praeger, Finding involutions in finite Lie type groups of odd
characteristic, Journal of Algebra 321 (2009), no. 11, 3397–3417.
[Mac62]
I. G. Macdonald, The Poincaré polynomial of a symmetric product, Mathematical Proceedings of the
Cambridge Philosophical Society, vol. 58, Cambridge Univ Press, 1962, pp. 563–568.
, Symmetric functions and Hall polynomials, Clarendon Press Oxford, 1979.
[Mac79]
[Mil12]
J. S. Milne, Reductive groups, 2012, Available at www.jmilne.org/math/.
[NP95]
P. M. Neumann and C. E. Praeger, Cyclic matrices over finite fields, Journal of the London Mathematical Society 52 (1995), no. 2, 263–284.
[NP98]
, Derangements and eigenvalue-free elements in finite classical groups, Journal of the London
Mathematical Society 58 (1998), no. 3, 564–586.
, Cyclic matrices in classical groups over finite fields, Journal of Algebra 234 (2000), no. 2, 367–
[NP00]
418.
[NP10]
A. C. Niemeyer and C. E. Praeger, Estimating proportions of elements in finite groups of Lie type, J.
Algebra 324 (2010), no. 1, 122–145.
[NPP10]
A. C. Niemeyer, T. Popiel, and C. E. Praeger, On proportions of pre-involutions in finite classical
groups, Journal of Algebra 324 (2010), no. 5, 1016–1043.
, Abundant p-singular elements in finite classical groups, Journal of Algebra 408 (2014), 189–
[NPP14]
204.
[NPS13]
A. C. Niemeyer, C. E. Praeger, and A. Seress, Estimation problems and randomised group algorithms,
Probabilistic Group Theory, Combinatorics, and Computing, Springer, 2013, pp. 35–82.
32
R Jiménez Rolland
J C H Wilson
REFERENCES
[Sri79]
B. Srinivasan, Representations of finite Chevalley groups, Lecture Notes in Mathematics, vol. 764,
Springer-Verlag, Berlin-New York, 1979.
[Ste68]
R. Steinberg, Endomorphisms of linear algebraic groups, Memoirs of the American Mathematical
Society, No. 80, American Mathematical Society, Providence, R.I., 1968.
[Ste89]
J. R. Stembridge, On the eigenvalues of representations of reflection groups and wreath products, Pacific
J. Math 140 (1989), no. 2, 353–396.
[Usp20]
J. V. Uspensky, Asymptotic formulae for numerical functions which occur in the theory of partitions,
Bulletin of the Russian Academy of Sciences 14 (1920), no. 6, 199–218.
[Wal99]
G. E. Wall, Counting cyclic and separable matrices over a finite field, Bull. Austral. Math. Soc. 60 (1999),
no. 2, 253–284. MR 1711918
[Wil14]
J. C. H. Wilson, F IW -modules and stability criteria for representations of classical Weyl groups, Journal
of Algebra 420 (2014), 269–332.
[Wil15]
J. C. H. Wilson, F IW –modules and constraints on classical Weyl group characters, Mathematische
Zeitschrift 281 (2015), 1–42.
[You30]
A. Young, On quantitative substitutional analysis (fifth paper), Proceedings of the London Mathematical Society 2 (1930), no. 1, 273.
Rita Jiménez Rolland
Instituto de Matemáticas, Universidad Nacional Autónoma de México
Oaxaca de Juárez, Oaxaca, México 68000
[email protected]
Jennifer C. H. Wilson
Stanford University, Department of Mathematics
Sloan Hall, 450 Serra Mall, Building 380, Stanford, CA 94305
[email protected]
33
| 4math.GR
|
NMODE — Neuro-MODule Evolution
Keyan Ghazi-Zahedi
1
∗1
Max Planck Institute for Mathematics in the Sciences, Leipzig, Germany
arXiv:1701.05121v1 [cs.NE] 18 Jan 2017
January 19, 2017
Abstract
Modularisation, repetition, and symmetry are structural features shared by almost all biological
neural networks. These features are very unlikely to be found by the means of structural evolution of
artificial neural networks. This paper introduces NMODE, which is specifically designed to operate on
neuro-modules. NMODE addresses a second problem in the context of evolutionary robotics, which
is incremental evolution of complex behaviours for complex machines, by offering a way to interface
neuro-modules. The scenario in mind is a complex walking machine, for which a locomotion module
is evolved first, that is then extended by other modules in later stages. We show that NMODE is
able to evolve a locomotion behaviour for a standard six-legged walking machine in approximately 10
generations and show how it can be used for incremental evolution of a complex walking machine.
The entire source code used in this paper is publicly available through GitHub.
1
Introduction
One of the advantages of evolutionary algorithms is that these type of algorithms can be used to find
novel solutions with a minimal amount of pre-structuring. That means that there is no inherent need to
e.g. a priori specify the structure of a neural network. In the context of embodied artificial intelligence,
this means that for an embodied agent, only the number of sensors and motors need to be specified and
everything else can be left open to evolution. In particular, artificial evolution can be used to find optimal
control structures by sequentially growing and pruning networks. We use the term optimal in the sense
of [18], i.e. in the context of structurally minimal networks that solve a given set of tasks. Furthermore,
evolutionary algorithms can be used to co-evolve the brain and body of embodied agents. Impressive
examples date back to Karl Sim’s work [26, 25], but can also be found in more recent work (e.g. [1, 19]).
A common trait of almost all experiments (early as well as contemporary) in the context of evolutionary robotics are that the morphology or the behaviour is kept simple, while the other might vary. In
cases of simple morphologies, i.e. wheel-driven robots, one can find examples of interesting behaviours
such as predator-prey scenarios [20]. If the morphology is more complex, i.e. walking machines, then
evolved behaviours, mainly solve locomotion problems (e.g. [30]), which does not mean that the results
are less impressive. Yet, it indicates that it is a challenge to evolve complex behaviours for complex
systems. We believe that these challenges result from two main problems. In the early days, genetic
algorithms [15] (GAs) were used predominantly. GAs were often applied to fully connected recurrent
neural networks, which means that the string length would increased exponentially with every additional
sensor or motor. Hence, for simple systems with limited sensors and motors, the space of policies was
small enough to search for interesting behaviours, whereas for complex morphologies, such as walking
machines, evolving a robust walking behaviour already posed a challenge. A workaround was to use
∗
[email protected]
1
neural networks of particular structures, but then the question arose, what are best suited structures to
learn a specific behaviour? An interesting alternative, which is known as HyperNEAT, will be discussed
below. The second problem, in our opinion, is that it is very difficult to add functionality to a network
in an evolutionary setting. Let us assume that we have evolved a locomotion network for a hexapod.
How can we add the functionality to add e.g. a light-seeking behaviour, while ensuring that the already
learned behaviour is preserved. One can fixate the synaptic weights of the locomotion network, but how
should the newly grown structure interface with the already existing locomotion structure? NMODE is
specifically designed to address these two problems, i.e., to reduce the search space for the evolution of
neuro-modules in a meaningful way and to provide a principled way to interface new module ,,on-top”
of already evolved structures. Before we introduce NMODE, we first present contemporary algorithms
and discuss why we did not choose to extend an existing framework instead of creating yet another
artificial evolution algorithm.
To the best of the authors knowledge, there is only one algorithm that can be considered anything
close to a standard in this context. HyperNEAT [23] is a very popular evolutionary algorithm that allows
to simultaneously evolve morphologies and neural networks. To understand how HyperNEAT works, it is
important to look at its predecessors, NEAT [27] and CPPN [24]. NEAT is a very interesting algorithm
to evolve neural networks of arbitrary structure. Two key features of NEAT are its ability to allow for
cross-over between to arbitrary structures and speciation as a method to protect innovations. CPPNs
(compositional pattern producing networks) are NEAT networks which use a set of function instead
of the standard sigmoidal transfer function. This means that one parameter of the neuron selects the
transfer function from a set of possible functions (sine wave, linear, saw tooth, . . . ). Let us assume,
that we have evolved a CPPN with two inputs and a three outputs. Let us further assume, that the two
inputs are the x and y coordinates of a two-dimensional bounded plane and the three outputs are RGB
colour values. This means that the CPPN is now able to produce a picture. The Picbreeder website1 [24]
shows impressive examples of pictures that were evolved by visitors of the website. It is striking how
CPPNs can make use of symmetries, repetitions, etc., to create complex and appealing images. The
website Endless Forms2 applies the same idea to 3D shapes [5]. The output of the CPPN is now used to
determine the boundary of geometric figures. HyperNEAT uses CPPNs to generate the synaptic weights
in a layered neural network. Each layer of a neural network is placed on a two-dimensional plane and
two layers are fully connected, which means that every neuron in the input layer is connected with every
neuron in the hidden layer and every neuron in the hidden layer is connected to every neuron in the
output layer. A CPPN is now fed with the geometric location of the pre- and post-synaptic neurons on
their layers and the output of the CPPN is the synaptic connection between the two neurons. Although
there are very interesting experiments published that use HyperNEAT (e.g. [1]), there are currently no
publications that show a successful application of HyperNEAT to complex robotic systems. For example,
HyperNEAT has not been used to evolve behaviours for non-trivial walking machines such as Svenja or
even Heaxaboard, which is discussed in Section 3.1 and Section 3.2. One reason is that HyperNEAT is
very good at finding patterns, but not very good in fine tuning parameters [4].
From our perspective, the significant reason not to use HyperNEAT is the its requirement to use a
predefined neural network architecture. The number of layers and their respective connectivity must be
specified by the experimenter before running the evolutionary algorithm. We are interested in finding
minimal and optimal neural network architectures, which is why the structure of the network (including
the number of synapses and neurons) must be be open to evolution. NEAT allows to evolve the structure
of a network together with its parameters but does not allow for modularisation. One idea that will
be followed in future work is to combine NEAT with the modular structure of NMODE. This would
allow to use cross-over operators not only on the level of modules (see Sec. 2) but also allow cross-over
between modules.
1
2
http://www.picbreeder.com
http://www.endlessforms.com
2
Besides NEAT, there aren’t many algorithms that support the co-evolution of structure and parameters of a neural network. One of the first frameworks probably is ISEE (integrated structure evolution
environment), which is described in detail in [14] and based on the ENS 3 algorithm [10]. ISEE not
only included software to evolve neural networks but also sophisticated tools to inspect the dynamics of
neural network while they were operating in the sensorimotor loop as wells as ex vitro analysis of their
dynamics. It has been applied very successfully in numerous experiments to evolve networks in simulation that were then used with minimal adaptation effort to control real robots (see e.g. [17, 29, ?, ?]).
Unfortunately, the initial implementation did not support the evolution of disjunct neuro-modules. As
an example, consider the AiboTM robot for which a behaviour was evolved in [17]. This system has a
left-right symmetry that an evolutionary algorithm can take advantage of to reduce the search space.
In ISEE, this required a few workarounds. NERD [22] is a full evolutionary environment, designed to
overcome the limitations of ISEE with a lot of additional functionality. It allows to define neuro-modules
with a rich repertoire of operators. An interesting feature is that NERD allows the definition of connection rules, e.g. to force that a specific sensor has to be connected to a specific actuator. The software is
very powerful, which comes with the cost that it is not easy to use out of the box. Additionally, NERD
is also unfortunately no longer maintained as the development has stopped.
NMODE’s development was initiated to overcome ISEE limitations by adapting NERD’s neuromodular concept in a minimalistic and easy to use way, thereby adapting interesting ideas from NEAT.
The novel idea that guided NMODE’s development is the principled way of augmenting already evolved
networks by new modules, which we see as one of the biggest problems in the context of evolutionary
robotics to achieve behaviours of higher complexity. If we look at the impressive work by Karl Sims [26,
25], then it is difficult to see a significant progress since these early experiments. Otherwise stated, in the
context of evolutionary robotics, we are still mostly working on the level of Braitenberg vehicles [3], which
means that we mostly work with purely reactive systems. Although the complexity of the morphology
has clearly increased over the decades, the behaviours that these systems show are still mainly reactive.
To the best of the authors knowledge, there are currently no systems in which a non-trivial morphology
has been evolved beyond the realms of reactive behaviours. This is where the development of NMODE
is targeted at; evolving complex behaviours for complex morphologies. This requires a method to
incrementally increase the complexity of the behaviours without loosing already gained functionality.
The method of adding new functionality in this context is called incremental evolution and is it not a
new concept [20].
The core ideas of NMODE can be summarised in the following way. A neural network is decomposed
into neuro-modules. Neural networks are composed of modules and mutations can only the change the
structure within a modules. In particular, this means that synaptic connection cannot be created between
neurones of different modules. Instead, the interaction between two modules is controlled by interface
neurones (see next Sec.). This allows to incrementally extend previously evolved modules with new
modules in a controlled manor. NMODE is not designed to be biologically plausible, but to allow to
evolve neural networks for non-trivial behaviours of complex systems. The intention of this work is to
introduce NMODE and investigate how it works on a well-known morphology (see Sec. 3.1) as well as a
more complex, biologically motivated morphology (see Sec. 3.2). Evolving non-reactive behaviours will
be investigated in future publications.
This work is organised in the following way. The next section discusses the neuro-modules and
evolutionary algorithm in detail. The source code is freely available at [12]. The following section
presents two experiments, before this work concludes with a discussion and outlook.
2
NMODE
Evolutionary algorithms can generally be divided into four functions that each operate on a population
of individuals, namely evaluation, selection, reproduction, and mutation. Let P be a population of
3
individuals, and let E : P 7→ P, S : P 7→ P, R : P 7→ P, and M : P 7→ P be the evaluation, selection,
reproduction, and mutation functions. Then, the evolution from one generation to the next can be
written as
P (n+1) = M R S E P (n)
,
(1)
where n refers to the generation index.
NMODE is designed to reduce the search space by modularisation of the neural network structure.
Therefore, we will first discuss how the modularisation is specified, before we present the algorithmic
details.
2.1
Neuro-Modules
The basic idea of NMODE is that the morphology of an embodied agent can be used to determine
how a neural network should be modularised. It is known from, e.g. stick insects, that each leg has its
own local controller and that the local leg controller are synchronised by a central nervous system (see
e.g. [8, 9, 11, 16, 29]). Hence, if we have a geometric description of an agent’s morphology (e.g. Fig. 2),
which includes the pose of the segments and the location of the sensors and actuators, we can use that
information to generate a geometric description of a neural network. By that we mean that neurons
will have 3D coordinates that refer to their position in a global Cartesian coordinate system. Currently,
this geometric description must be defined manually in NMODE, but there is no reason why this could
not be automatised in future versions. An example is the RoSiML3 robot simulation mark-up language
that is used by YARS [13] to define an experiment. It contains the pose of each segment as well as the
location of each actuator and sensor. Gazeboo [21] uses a similar file structure that could also be used
to automatically generate a neural network configuration. These files can be parsed for the names and
location of segments, sensors and actuators and also determine symmetries and repetitions.
A module in NMODE contains six different types of neurons and a list of synapses. The next
paragraph will describe each of them in detail, but we will first present the type of neuron model that
is currently implemented. We use the standard additive neuron model, which is given by the following
set of equations:
X
ai (t + 1) = θi +
wji oj (t)
(2)
j
oi (t + 1) = τ (ai (t)),
(3)
where ai (t) ∈ R is the activation of neuron i ∈ N at time t ∈ N, oi ∈ R is the output of neuron i at
time t, wji ∈ R is the synaptic strength between the pre-synaptic neuron j ∈ N and the post-synaptic
neuron i, and finally, τ : R 7→ R is the transfer function. All neurons are updated synchronously, which
means that first the activations ai (t + 1) are calculated based on the neuron outputs of the previous
time step oi (t). Once all activations are updated, the neuron outputs follow, i.e., all outputs oi (t + 1)
are calculated based on the updated actuators states ai (t + 1). As already indicated, we currently work
with time discrete neural networks. It must be noted here that the specification given in the NMODE
XML file is generic enough to allow NMODE to be extended to work with any type of neural networks
or graphical models. This is why we will sometimes refer to neurons as nodes and synapses as edges.
Currently, we support three type of transfer functions, namely τ ∈ {id, sigm, tanh}, where id(x) = x
is the identity function, sigm(x) = 1/1+e −x is the standard sigmoid function, and finally tanh is the
Hyperbolic Tangent.
Next, we will discuss the six different types of nodes, which are; sensor, actuator, hidden, input,
output, and connector. The first three node types are well-known from other neural network contexts.
3
http://keyan.ghazi-zahedi.eu/rosiml
4
Shoulder Knee
Input
Output
Leg Module
FCS
Sensor nodes
Actuator nodes
Shoulder Knee
CPG
Resulting Network
Comparable unrestricted network
Figure 1: Hexaboard NMODE configuration vs. unrestricted configuration. FCS refers to foot contact
sensors.
Sensor nodes receive input from the agent’s sensors and are typically equipped with a linear (id) transfer
function. Analogously, actuator nodes are directly connected to the actuators of the agent, and finally,
hidden nodes are only connected to other nodes within the same module. Actuator and hidden neurons
are usually equipped with one of the two sigmoidal functions presented above. Two of the new node
types are named input and output. The role of these nodes is to function as connectors between
modules. Input nodes only allow connections from itself to other neurons within the module it is
assigned to, whereas output nodes only allow connections from nodes of the corresponding module to
itself. To connect a module with another module that has input and output nodes, the connecting
module has to use a connector node. The connector node will copy the position and properties of the
node it is referring to. In particular, this means that if a connector nodes refers to an input node, it will
automatically become an output node of the connecting module.
As an example, consider the hexapod used in the first experiment presented below (see Sec. 3.1).
Its NMODE structure is shown in Figure 1. Sensor nodes are shown in red, actuator nodes in green,
input nodes in cyan, output nodes in purple, and finally, connector nodes in orange. In the experiment
discussed below, only one leg module and one CPG module are evolved. By definition, a leg module
controls a leg, which means that requires needs sensor and actuator nodes to access the leg’s state and
control its movement. For a simple locomotion movement, the CPG does not require any sensor or
actuator nodes. Instead, the CPG only consists of connector nodes to interface with the evolved leg
modules. Because all legs are morphologically identical in this example, only one leg module has to be
evolved, which is the used several times for the final controller (see Fig. 1).
We demonstrate the parameter reduction that results from this modularisation based on the hexapod
example and with the assumption that the insertion of hidden units is not allowed. This allows us to
calculate the maximal number of edge for a modularised and non-modularised neural network (see
Fig. 1), where the latter means that synapses can connect any two neurons. It must be noted, that
sensor nodes can only have outgoing connections in the current implementation, i.e., they only function
as proxy for the sensor values. Let ns,a,i,o,h be the number of sensor, actuator, input, output, and hidden
nodes. For a fair comparison, we replace all pairs of input/output nodes by a single hidden node in the
unrestricted configuration, which follows structure that was used in [17, 31]. Then the dimension of the
weight matrices for the two configurations is given by Tab. 1.
For the example of a six-legged walking machine with 2DOF per leg, we see that the unrestricted
configuration has up to 540 free parameters that the evolutionary algorithm has to take care of, compared
to the modularised network structure, which only requires the evolution of maximally 93 parameters.
In other words, the search space in the unrestricted case is about 5.8 times larger compared to the
modularised case. The next section discusses the selection, mutation, and reproduction operators on
modules.
5
NMODE configuration: Leg Module
ns · (no + na ) + ni · (ni + na ) + no · (ni + na ) + na · (no + na ) = 21
NMODE configuration: CPG Module
ni · no + no · ni = 72
NMODE configuration: Total = 93
Unrestricted configuration:
ns · (nh + na ) + nh · (nh + na ) + na · (nh + na ) = 540
Table 1: Comparison of search space dimension for the Hexapod example.
2.2
Selection
We decided to use the simplest form of selection, which is a rank based approach. After evaluation,
the individuals are ranked based on their fitness value. The top n% (user defined) are then selected for
reproduction. In the previously used ENS3 approach [10, 14], offspring were assigned to individuals with
respect to a Poisson distribution based on the individual’s fitness values. Individuals without offspring
after the reproduction process, were removed from the population. An additional parameter allowed
to save the best n individuals, which might have not been assigned an offspring despite having a high
fitness value. The idea was that, because of its stochastic nature, this selection method would provide
more diversity in the population. We believe that the speciation method introduced by [23] is more
valuable in this respect, which is why it will be included and evaluated in future versions of NMODE.
We also believe that cross-over is more useful with respect to ensuring diversity in the population, which
is why a crossover mechanism based on the exchange of modules is included (see below). In future
versions of NMODE, we will evaluate a NEAT-based cross-over technique to allow crossover between
two modules instead of a full exchange.
2.3
Reproduction
The task of the reproduction operator is to assign the right amount of offspring to each individual that
was selected after the evaluation. By right amount we mean that the number of offspring assigned
to each individual should reflect the individual’s fitness values. For this purpose, all fitness values are
normalised in the following way:
η
fi − mini {fi }
si
si =
(4)
,
ri = P ,
maxi {fi } − mini {fi }
i si
where i = 1, 2, ... , N is the index of the i-th individual, and fi is the corresponding fitness value, η is the
elitism parameter, and ri is the reproduction factor. A constant value for si is used, if max{fi } = min{fi }.
A high value of η leads to a concentration of the offspring to the individuals with the highest fitness
values. Analogously, a low value of η results in an equal distribution of offspring among the selected
individual. The resulting reproduction factors ri sum up to one, which means that the normalised ri can
directly be used to distribute the number of offspring.
One reason for the modularisation of the neural networks and the definition of the interface nodes is
that the modules can easily be exchanged between individuals. If the cross-over parameter is positive,
i.e. ζ > 0, then for each offspring (which is a copy of its mother) a father is chosen randomly from the
set of parents depending on their normalised reproduction factor ri . This means that individuals with a
high reproduction will be chosen more often as a mating partners. If two mates are found, then ζ also
determines the probability with which a module is chosen from the father instead of the mother. This can
be fatal in many cases, because the modules were not co-evolved, and hence, there is a high probability
that e.g. a leg module that works fine with a particular CPG will not work at all with another CPG.
This is accounted for, because very fit individuals will often mate with themselves (due to their high
6
fitness value), and hence, good networks will be preserved. On the other hand, less successful networks
will more likely be crossed with more successful networks, which could increase their survival chances.
As briefly mentioned above, one might think about using NEAT as a method to evolve networks within
in module. This would then allow to cross module-networks instead of completely switching them. We
will investigate this possibility in future versions of NMODE.
2.4
Mutation
NMODE is designed particularly to evolve neural networks of arbitrary structure while restricting the
search space in an meaningful way. This means that the mutations are performed on the level of modules
and not on the entire neural network. Consequently, in the following sections, E and N denote the set
of all edges and nodes within a single module.
Synapse insertion Because nodes have a three-dimensional position (see above), it makes sense to
insert edges based on the Euclidean distance between nodes. Nodes which are closer together, will
have a higher probability of being connected than those which are separated by a longer distance. This
way, sensors and actuators which e.g. read and control the same joint are more likely to be connected
early during the evolutionary process than sensors and actuators which have a larger distance. We
will investigate how a distance-based edge generation method differs from an uniformly distributed
insertion of edges with respect to the evolutionary success and the generated behaviours in an experiment
discussed below (see Sec. 3.1). Both methods are explained in the following paragraph.
At first, NMODE generates a dis-connectivity matrix. This means, that the matrix D has a positive,
non-zero entry at dij when the synapse from node j to node i does not exist. There are currently
two ways in which the non-zero entries are calculated (see Alg. 1). The first possibility is to set them
uniformly to one. This means that each new connection has the same probability of being created.
This refers to a uniformly distributed insertion of synapses. The second possibility is to set the values
proportionally to the distance between the two corresponding neurons. This means that neurons which
are closer to each other will have a higher chance of being connected by a synapse. In both cases, each
entry in the dis-connectivity matrix is then multiplied with the user-defined synapse insertion probability,
which means that the resulting matrix determines with which probability each non-existing synapse will
be created (see Alg. 1).
Synapse deletion The user specifies how many synapses can be removed during the mutation of a
module. Hence, if the user specifies p = 1, than any fraction from 0% to 100% of the synapses can
be deleted, whereas p = 0.5 means that maximally 50% of all existing synapses can be deleted. The
synapse deletion algorithm is comparably simple (see Alg. 2).
One could argue that this form of synapse insertion and deletion are too invasive, as they can both
lead to significant changes of the network structure. This is intentional for the following reason. NMODE
was not designed to be a biologically plausible evolutionary algorithm to understand how biology has
evolved complex systems. NMODE was initiated to find controller for complex systems in cases where
other approaches fail. In other words, the solution is more important than the algorithm that found it.
Therefore, if a local maximum is reached, it is desirable to increase the stochastic search area to find
other maxima. A good way of doing this is to significantly alter the structure of the neural network,
which is why we allow such large structural variations.
Synapse modification The modification of synapses is straight forward. Each synapse is modified
with a probability specified by the experimenter. The parameter δ determines the maximal variation of
the synapse and the max parameter determines the limit of the synaptic strength (see Alg. 3).
7
Algorithm 1 Synapse insertion
Require: p ∈ [0, 1]
. Probability to insert a synapse
Require: u() ∈ [0, 1]
. Function to draw a uniformly distributed random number
Require: D = (dij ) = 0, ∀i, j = 1, 2, ... , N
Require: E ∈ N2
. is the set of all existing nodes, which is empty only for the initial generation.
1: for i = 1, 2, ... , N do
2:
for j = 1, 2, ... , N do
3:
if (i, j) 6∈ E then
4:
if uniform then
5:
dij = 1
6:
else if distance then
7:
dij = 1/dist(ni ,nj )
8:
end if
9:
end if
10:
end for
11: end for
D
12: D = max
. Normalise D such that the largest entry is equal to one.
D
13: for i = 1, 2, ... , N do
14:
for j = 1, 2, ... , N do
15:
if u() < p · dij then
16:
Add synapse (i, j) to E
17:
end ifu() < psi
18:
end for
19: end for
Algorithm 2 Synapse deletion
Require: p ∈ [0, 1]
. Probability to delete a synapse
Require: u() ∈ [0, 1]
. Function to draw a uniformly distributed random number
1: N = u() · p · |E|
2: for i = 1, 2, ... , N do
3:
Delete a synapse randomly drawn from E
4: end for
Algorithm 3 Synapse modification
Require: p ∈ [0, 1]
. Probability to modify a synapse
Require: m ∈ R+
. Maximal absolute value for the synaptic strength
Require: δ ∈ R+
. Maximal absolute change of a synaptic strength
Require: u() ∈ [0, 1]
. Function to draw a uniformly distributed random number
1: for (i, u) ∈ E do
2:
if u() < p then
3:
wij = wij + δ(2u() − 1)
4:
Prune wij , such that wij ∈ [−m, m].
5:
end if
6: end for
Neuron insertion In growing a network, it is useful to insert neurons carefully, without massive structural changes [27]. In our previous implementation [14, 10], we used to add a neuron and automatically
connect it to a given percentage of the neural network. This means that every new neuron has a potentially large impact on the network structure, and hence, its function. We decided to insert neurons more
8
carefully. Hence, we adopted NEAT’s method of inserting a neuron, which means that with probability
p a synapse is chosen which is split by the new neuron. The synaptic weight of the new incoming
synapse is set to one, whereas the outgoing synaptic strength is equal to the strength of the original
synapse (see Alg. 4). The new neurons position is set to centre of the original synapse.
Algorithm 4 Neuron insertion
Require: p ∈ [0, 1]
Require: u() ∈ [0, 1]
1: if u() < p then
2:
(i, j) is drawn uniformly from E
3:
Insert new neuron k in N
4:
Remove (i, j) from E
5:
Add (i, k) with wik = 1 to E
6:
Add (k, j) with wki = wij to E
7: end if
. Probability to add a neuron
. Function to draw a uniformly distributed random number
Neuron deletion Depending on the number of connections, the deletion of a neuron can have a strong
influence on the function of the recurrent neural network. For this reason, neurons should be removed
with care, which is why in each mutation at most one neuron is removed (see Alg. 5).
Algorithm 5 neuron deletion
Require: p ∈ [0, 1]
. Probability to remove a neuron
Require: u() ∈ [0, 1]
. Function to draw a uniformly distributed random number
1: if u() < p then
2:
Remove a neuron randomly drawn from N
3: end if
Neuron modification Neurons are modified in the same way that synapses are modified. Each neuron
is modified with the specified probability. The bias value is changed with the limits specified by δ and
cropped within the limits of the specified maximal value (see Alg. 6)
Algorithm 6 neuron modification
Require: p ∈ [0, 1]
. Probability to modify a neuron
Require: m ∈ R+
. Maximal absolute value for the bias
Require: δ ∈ R+
. Maximal absolute change of the bias
Require: u() ∈ [0, 1]
. Function to draw a uniformly distributed random number
1: for n ∈ N do
2:
if u() < p then
3:
bn = bn + δ(2u() − 1)
4:
Prune bn , such that bn ∈ [−m, m].
5:
end if
6: end for
2.5
Evaluation
NMODE is primarily designed to work with YARS, although other simulators and evaluation methods
can be easily included too. YARS is a fast mobile robot simulator especially designed for evolutionary
9
robotics. It uses the bullet physics engine [6] for the physics and Ogre3D [28] for the visualisation. Robots
can be created in blender [2] and imported to YARS and blender animation data can be exported from
YARS allowing to replay and render the simulation from any camera angle. YARS and NMODE are both
available from GitHub [13, 12]. Therefore, the question of evaluation (except for a brief discussion on
fitness functions below) will not be addressed in this work but in a following publication on the current
version of YARS.
This section does not discuss fitness functions in general. There is a vast amount of literature
available on this topic (see e.g. [15, 20])). Instead, this section describes how fitness functions are
implemented in NMODE, because this is only part that requires programming by the user. Fitness
functions are dynamically loaded during runtime of the NMODE executable. The name of the fitness
function given in XML file must correspond to the name of the library. The class must inherit the Evaluate
class and implement the following functions; void updateController(), void updateFitnessFunction(),
bool abort(), void newIndividual(), void evaluationCompleted(). The first function updateController is
called after each step in the simulation, e.g. after each update of the sensor states. All sensor values
are presented in an one-dimensional array and can be used to calculate the intermediate fitness. The
function abort is then checked by NMODE to see if the evaluation should be terminated early. This
function can be used by the experimenter to indicate that some abort condition was met, e.g. if a
collision with a wall occurred. If an individual reached the full evaluation time, the evaluationCompleted
function will be called to allow post-processing. In the examples given below, the entropy [7] over joint
angles could be used as part of the fitness function. This calculation only makes sense at the end of each
evaluation for two reasons. First, the calculation of entropies can be very time consuming, and hence,
should not be performed after every simulation update. Second, such calculations are most usefull if
the full data is available, which is why the evaluationCompleted function was added. The newIndividual
function is called at the beginning of each evaluation and is meant for the initialisation and resetting of
variables.
This concludes the introduction of NMODE. The next section presents two experiments in which
NMODE was used to evolve locomotion behaviours in YARS.
3
Experiments
In the following, we will present two different experiments, namely, Hexaboard and Svenja. The reason
why we present these experiments is that the Hexaboard can be considered to be a standard hexapod
morphology with 2DOF per leg (see Fig. 2). This is a kind of benchmark experiment to show how
well NMODE performs on such tasks. The second experiment, Svenja, is a hexapod morphology that
was freely modelled after an insect. It has four degrees of freedom in each leg and the three leg pairs
are morphologically different (see Fig. 2). Svenja is particularly challenging to control because of the
arrangements of the legs and actuators. Contrary to Hexaboard, one cannot use the same network for
each leg, as the three leg pairs will have different tasks, due to their differing morphologies. The front
legs will have to pull, while the rear legs will have to push. The two centre legs have the same morphology
as Hexaboard’s legs, and hence, will have a similar task. Manually programming a behaviour for this
morphology turned out to be very challenging. This experiment was also chosen to demonstrate that
NMODE’s modularisation is well-suited for incremental evolution. Each leg-pair was evolved individually,
with its own CPG. For the final evolutionary phase, the three disjunct networks were simply merged.
3.1
Hexaboard
This experiment was chosen because it is a good representation of a standard experiment in the context
of embodied artificial intelligence. The task is to evolve a neural network for the locomotion of a sixlegged walking machine. Most walking machines are constructed with identical legs, i.e. each leg has the
same morphology and is attached to the main body in exactly the same way. This reduces the complexity
10
Figure 2: Experimental platforms. Left: Hexaboard is standard hexapod with 2DOF per leg. Right:
Svenja is a hexapod that is loosely modelled after an insect with 4DOF per leg and two light sensors.
of the control significantly, because the controller does not have to distinguish between different leg
pairs. The movements of the two front legs have the same effect on the locomotion behaviour as the
movements of the rear legs. This is in contrast to natural systems, e.g. cockroaches, for which the front
legs perform a pulling movement, whereas the rear legs can be said to push. Later in this work, we will
evolve a neural network for a hexapod with a more natural task distribution for its legs (see Sec. 3.2).
In this section we will first concentrate on the standard configuration of six identical legs, which are
mounted parallel to each other to a single segment (see Fig. 2).
For such a system, is it natural to assume that each leg should have the same local leg controller,
because each leg should show the same behavioural pattern in e.g. a tripod walking gait. Hence, to
reduce the search space, we only need to evolve a single leg controller and copy it five times for the final
controller. In addition to the leg controller, only the connecting structure (central pattern generator or
CPG) needs to be evolved, which means that the search space can be reduced significantly (see Tab. 1
an and Fig. 1). In what follows, we will first describe the morphology, followed by the user-defined
evolution parameters, before the results are presented. It must be noted here, that all experiments were
conducted as batch process, i.e., the parameters were set at the beginning of the experiments, kept
constant and the same for all experiments. This is not how NMODE was intended to be used and
other parameters might lead to even better results. The goal here is to show how fast NMODE can
evolve a good behaviour for such a standard platform. The goal is not to evolve an optimal network
structure (with respect to size and robustness against disturbances), which would require alternative
growing and pruning phases, modification of fitness function parameters, etc., and hence, a monitoring
of the evolutionary process. Interactive evolution would not allow to make statistical statements about
the evolutionary progress, which is why we chose to evolve in batch mode.
3.1.1
Morphology
We chose Hexaboard as our experimental platform for two reasons. First, it is provided in the YARS
distribution, which means that the results presented here can be easily verified, and second, it was
successfully used in previous experiments [18]. Overall, the robot was designed especially for the purpose
of learning locomotion, which means that the body part dimensions, weights, forces, angular range,
etc. were set such that the system is highly dynamic but has a very low probability of flipping over. The
exact parameters are presented in Table 1. Figure 2 visualises the arrangement of the joints and their
rotation axes.
All shapes were modelled in blender and the visual appearance of the robot is equivalent to it’s
physical form, which means that the shapes that are used for collision detection etc. are the same
shapes that are also used for the visualisation. This is also true for Svenja (see below).
11
Bounding Box (x,y,z)
Mass
Main Body 0.75 m 4.41 m 0.5 m 2.0 kg
Femur
0.23 m 0.23 m 1.17 m 0.2 kg
Tibia
0.1 m 0.09 m 0.9 m 0.15 kg
Tarsus
0.1 m 0.08 m 1.04 m 0.1 kg
αmin αmax
Fmax
Vmax
◦
◦
Shoulder -35
35
20.0 N/m 0.75 rad/s
◦
◦
Knee
-15
25
20.0 N/m 0.75 rad/s
Table 2: Hexaboard’s specification
3.1.2
NMODE parameters
To evaluate how well NMODE evolves a locomotion network for a standard hexapod platform, we ran
each experiment 10 times with the same set of parameters. From previous experience [17], we decided
to focus primarily on the evolution of the connectivity structure and omitted the insertion and deletion
of neurons in this set of experiments. Hence, the probability to add or remove a node was set to zero.
For the reproduction parameters, we chose a population size of 100 with a selection pressure of 0.1.
The elitism parameter is set to 10.0, which we determined empirically. Crossover was set to 0.1. The
mutation parameters were set to node modification of 0.01 with a maximal absolute value for the bias of
1.0 and a maximal change of the bias value to 0.01. Edge modification was set to 0.2, with a maximum
of 5.0 and a maximal step size of 0.5. Edges can be inserted with a probability of 0.05 and a maximal
absolute synaptic strength of 1.0. For the experiment in which the probability of an edge insertion is
dependent on the distance of the nodes, the minimal distance was set to 0.1, which means that neurons
with a distance smaller than 10 cm are set to 10 cm (this also applies to self-connections).
3.1.3
Fitness Function
The fitness function is the summed distance in direction of the initial orientation of Hexaboard, which is
along the positive y -axis. Let t = 1, 2, ... , T indicate the time steps, then the fitness function is defined
as
F =
T
X
y (t).
(5)
t=1
The reason for using the summed distance instead of the maximal distance is that a small progress in
distance leads to higher selection pressure, i.e. significantly more offspring are assigned to individuals
which are slightly better than their peers.
3.1.4
Results
The plot on the left-hand side of Figure 3 shows the fitness over generations for a uniform insertion
of synapses, while the plot on the right-hand side of Figure 3 shows the fitness over generations for
distance-based insertion of synapses. The standard deviations and mean values are taken with respect
to all selected individuals of all 10 experimental trials. The smaller plots in each figure show the mean
and standard deviation for only the best individuals over all 10 experiments. Hence, the large plots
show the values for 100 individuals in each generation (10 for each of the 10 experiments), whereas the
smaller plots only show the calculated values for 10 individuals (the best of each experiment). Figure 4
shows the resulting walking patterns for three different evolved individuals.
12
Best fitness averaged over 10 experiments (uniform)
Best fitness averaged over 10 experiments (distance)
20 000
15 000
Fitness
Fitness
15 000
10 000
Fitness averaged over 10 experiments
5000
10 000
Fitness averaged over 10 experiments
5000
0
0
0
10
20
30
40
50
0
Generation
10
20
30
40
50
Generation
Figure 3: Hexaboard fitness. Left: Edges are added uniformly, which means that the distance between
the connected node is not taken into account. Right: The probability of inserting and edge decreases
with the distance between the connected nodes. Both plots show the best fitness averaged over 10
experiments, with the corresponding standard deviation. The picture-in-picture plots show the fitness
of all selected individuals in each generation averaged over 10 experiments. For comparison, the best
fitness values in each generation is added to the smaller plot as blue line.
Figure 4: Examples of evolved walking patterns for Hexaboard. Left and centre: These are two walking
patterns that resulted from inserting edges dependent on the distance. Right: The classical tripod
walking behaviour that was most dominant when synapses were inserted uniformly. The tripod walking
pattern was also found in experiments in which the distance-based approach was used, but there was a
larger variety compared to the uniform approach. For the latter, the walking patterns mostly differed
in the length of the leg stance and swing phases. Videos of the walking patterns are available online.
In the same order as they are shown above: https://youtu.be/a0OtDBHpe1c, https://youtu.be/
UwFKYTIn56w, https://youtu.be/fZlmPTSmCi0.
From the results shown in Figs. 3 and 4 we can draw the following conclusions. First, both approaches
lead to a good locomotion behaviour already after approximately 10 generations. To the best of the
authors knowledge, no similar results were published for any other evolutionary algorithm. The second
conclusion is that both approaches, uniform and distance-based insertion of synapses on average lead
to similar results. In both cases, the best walking behaviours showed a tripod walking behaviour (see
Fig. 4, left-hand side). Note that the fitness function (see Eq. (5)) only favours fast walking but does
not reward a specific walking pattern. In the case of the distance-based approach, we see a higher
variance in the fitness over the different experiments. This is also reflected in the resulting walking
patterns. Two examples of the best behaviours from the ten experiments with distance-based insertion
of synapses are shown in Figure 4 (centre and right-hand side). This leads to the conclusion that
if optimality is the highest priority, uniform insertion of synapses (position-independent) seems to be
the approach of choice, whereas distance-based insertion of synapses should be used if diversity of the
solutions is desired.
3.2
Svenja
Svenja was chosen for presentation here to demonstrate how NMODE can be used for incremental
13
evolution of complex systems. Svenja’s morphology is significantly more complex compared to Hexaboard
for four main reasons. First, each leg has 4 DOF (instead of 2). Second, the leg-pairs significantly differ
in their shape and attachment to the main body. Third, the main body is segmented, with a non-even
distribution of the weight, in particular, the rear end of Svenja is twice as heavy as the head, which
means that the center of mass is not within the convex hull of the leg’s anchor points (see Fig. 2 and
Tab. 3). Fourth, due to the leg morphology and weight distribution, Svenja is very likely to flip over if
the legs are not moved properly. To further increase the difficulty of the task, we included obstacles in
the environment (see Fig. 5, right-hand side).
The approach we chose to evolve Svenja is incremental in the sense that we first evolve every leg-pair
independently and then merge the resulting networks to evolve the coordination among the legs. As for
Hexaboard, evolution was performed as a batch process, with initially fixed parameters.
Figure 5 sketches the incremental approach. On the left-hand side, the evolution of each leg pair
is shown. The leg-pairs were evolved with a single leg controller, that was mirrored for the other leg,
and a CPG that only consisted of connector nodes to the leg modules (one input and one output node
per leg module). For the final controller (see Fig. 5, left-hand side), the three leg modules were simply
appended in a single XML file. The three disjunct CPGs were appended into a single module, which was
then appended to the list of leg modules. A small Python script was enough to automate this process,
which also could have been done manually with a simple XML editor. Figure 5 (right-hand side), shows
the three different evolutionary settings for the first step of the incremental evolution. The main body
is attached to an invisible rail that only allows movements along the z- and y -axes (up/down and
forwards/backwards movements). Movements along the x-axes and rotations in general are constrained
to zero.
Front leg-pair controller
copied
Centre leg-pair controller
copied
Rear leg-pair controller
copied
First step: Evolve individual leg controller
Initial controller for the second step
Figure 5: Incremental evolution of Svenja. Left-hand side: Controller set-up. First, for each leg-pair,
individual controller are evolved (first image on the left hand side). The evolved leg controller are
then combined to control the full morphology (second image on the left-hand side). Note that the
combination of the three controller is straight forward (see text for details). Right-hand side: The
corresponding evaluation environments in YARS. From left to right and top to bottom, rear legs, centre
legs, front legs, and full morphology.
In the following paragraphs, we will first give the specifications of the morphology, then describe the
NMODE parameters and fitness function, before the results are presented.
3.2.1
Morphology
Svenja has 4DOF of which the first are located in the first joint. This means that the femur can be
rotated around two axes simultaneously (see Fig. 2). The other two DOFs allow rotations around a
single axes and connect the femur with the tibia and the tibia with the tarsus. Svenja’s main body
14
consists of three segments. This affects the centre of mass of the entire system. The head and main
segments (two which all legs are attached) have a weight of 10 kg, whereas the rear segment has a
weight of 20 kg. The full specifications of Svenja are summarised in Table 3 (see Appendix).
3.2.2
NMODE parameters
During the first phase, three independent evolutionary experiments were conducted, one for each leg
pair. In all runs (repeated five times each), we used the same NMODE parameters, which are set to; life
time is 500 iterations, generations are limited to 250, population size is set to 50, elite pressure is set to
5.0, with a cross over probability of 0.25, node mutation probability is set to 0.1, with a maximum of
1.0 and maximal step width of 0.2, nodes are added with with a probability of 0.1 (and a maximal bias
values of 1.0) and deleted with with probability of 0.1, edge mutation probability is set to 0.3, with a
maximum of 5.0 and maximal step width of 1.0, edges are added with with a probability of 0.05 (and
a synaptic strength of 1.0) and deleted with with probability of 0.05.
3.2.3
Fitness function
The main part of the fitness is equivalent to the fitness function used for Hexaboard, which means that
we used the sum over the main body’s y -coordinate in the world coordinate frame. For the evolution of
the leg pairs, an additional punishment was added for changes of the main body’s z-coordinate. This
means that in principle each evolved leg pair run could result in different walking heights as long as the
main body height changes were minimal. The fitness function read as:
F =
t
X
[y (t) − γ(z(t) − z(t − 1))],
(6)
t=1
where z(0) is the initial position of the main body before the simulation is iterated for the first time.
3.2.4
Results
The results are captured in Figure 6 and Figure 7. Figure 6 shows the fitness over generations for the
leg pairs (five runs each) and the evolution of the merged leg-pair controller. Figure 7 shows the final
behaviour for a sequence in which Svenja climbs over an obstacle. Videos of all evolved behaviours can
also be found online4 .
The first observation is that the front legs seem to be more difficult to evolve than the other two
leg pairs. It took about 100 generations to evolve a good behaviour compared to approximately 10-20
generations for the centre and rear leg pairs. The reason is that the way the front legs are attached to
the main body makes it more difficult for them to climb over the obstacles (see Fig. 5, right-hand side).
In early generations, each leg-pair’s controller would start with oscillatory movements of low amplitude
and high frequency. For the rear and centre leg-pairs small increases of the amplitude would already
increase the likelihood of climbing the obstacle, while the front legs would still be blocked by the first
obstacle. The reason is that the oscillatory leg movements of the first leg pair repel the main body more
easily from the obstacle than it is the case for the other leg pairs.
Another observation is that the combined controller very quickly learned to coordinate the leg pairs.
Remember that each leg pair was evolved independently, with no restriction of the walking height or
walking pattern. Yet, within less than 10 generations the first coordination occurs. After approximately
20 generations a good walking behaviour has been found which is shown in Figure 7.
In total, the evolution of a good walking behaviour for Svenja required only a few of hours of
CPU time. It must be noted again, that this process was done as a batch process, and hence, the
4
https://www.youtube.com/playlist?list=PLrIVgT56nVQ7zwLRxUeEEXgxpFhIJk1BY
15
Front legs
Centre legs
3000
1500
2500
Fitness
Fitness
2000
1500
1000
1000
500
500
0
0
50
100
150
200
0
250
0
50
100
150
200
250
Generation
Generation
Rear legs
Combined controller
2000
2.5 × 105
2. × 105
Fitness
Fitness
1500
1000
500
1.5 × 105
1. × 105
5. × 104
0
0
50
100
150
200
250
Generation
20
40
60
80
Generation
Figure 6: Fitness of Svenja controller over generations. For the final controller (lower right-hand side),
the best individuals for the five runs of each leg pair were chosen. For details, please read the main text.
final behaviour is not optimal with respect to size, robustness, etc. Still, this experiment shows that a
behaviour for a complex morphology such as Svenja can be learned very fast with NMODE.
4
Discussion
We presented a novel approach to artificial evolution of neural networks and demonstrated its applicability in two experiments. NMODE was not developed isolated from current research in evolutionary
robotics, rather ideas of different approaches were combined in a minimalistic way. NMODE builds
upon the main authors own experience in evolutionary robotics (ISEE), experience with a collaborator’s
software framework (NERD) and own experience and literature research on the currently most dominant
approach (HyperNEAT). We showed that NMODE is able to evolve a locomotion behaviour for two
different hexapod experiments in very few generations. To the best of the authors knowledge, no similar
results have been published so far, which shows the usefulness of NMODE in generating interesting
behaviours for complex morphologies.
This said, the development of NMODE is not completed yet. On the short-term, we will evaluate
NEAT as a method to evolve the networks within a module to allow crossover between modules instead
of completely switching modules. In addition, we will also include NEAT’s speciation method to protect
innovations. Most importantly, We will have to show that NMODE is able to evolve non-reactive
behaviours for complex morphologies, which is the topic of ongoing experiments.
16
Figure 7: This figure shows the walking behaviour of Svenja. Frames 28, 37, 46, 58, 66, and 75 are
shown here (rendered with blender), which is equivalent to approximately 2.5 s.
5
Acknowledgement
This work was partly funded by the DFG-SPP 1527 ,,Autonomous Learning”
References
[1] Auerbach, J.E., Bongard, J.C.: Environmental influence on the evolution of morphological complexity in machines. PLoS Comput Biol 10(1), 1–17 (2014)
[2] Blender Foundatation: blender 2.78a. www.blender.org (2017)
[3] Braitenberg, V.: Vehicles. MIT Press, Cambridge MA (1984)
[4] Clune, J., Beckmann, B.E., Ofria, C., Pennock, R.T.: Evolving coordinated quadruped gaits with
the hyperneat generative encoding. In: Proceedings of the Eleventh Conference on Congress on
Evolutionary Computation, CEC’09, pp. 2764–2771. IEEE Press, Piscataway, NJ, USA (2009)
[5] Clune, J., Lipson, H.: Evolving 3d objects with a generative encoding inspired by developmental
biology. SIGEVOlution 5(4), 2–12 (2011)
[6] Coumans, E.: Bullet physics simulation. In: ACM SIGGRAPH 2015 Courses, SIGGRAPH ’15.
ACM, New York, NY, USA (2015)
[7] Cover, T.M., Thomas, J.A.: Elements of Information Theory, vol. 2nd. Wiley, Hoboken, New
Jersey, USA (2006)
[8] Cruse, H.: What mechanisms coordinate leg movement in walking arthropods? Trends in Neurosciences 13(1), 15 – 21 (1990)
[9] Cruse, H., Kindermann, T., Schumm, M., Dean, J., Schmitz, J.: Walknet—a biologically inspired
network to control six-legged walking. Neural Networks 11(7–8), 1435 – 1447 (1998)
17
[10] Dieckmann, U., Pasemann, F.: Coevolution as an autonomous learning strategy for neuromodules.
In: Proceedings of the European Conference on Artificial Life (ECAL’95). Granada, Spain (1995)
[11] Ekeberg, Ö., Blümel, M., Büschges, A.: Dynamic simulation of insect walking. Arthropod Structure
& Development 33(3), 287 – 300 (2004)
[12] Ghazi-Zahedi, K.: NMODE Github Repository. https://github.com/kzahedi/NMODE (2016)
[13] Ghazi-Zahedi, K.: YARS Github Repository. https://github.com/kzahedi/YARS (2016)
[14] Ghazi-Zahedi, K.M.: Self-regulating neurons. a model for synaptic plasticity in artificial recurrent
neural networks. Ph.D. thesis, Univeristy of Osnabrück (2009)
[15] Holland, J.H.: Adaptation in Natural and Artificial Systems: An Introductory Analysis with Applications to Biology, Control and Artificial Intelligence. MIT Press, Cambridge, MA, USA (1992)
[16] Ijspeert, A.J.: Central pattern generators for locomotion control in animals and robots: A review.
Neural Networks 21(4), 642 – 653 (2008)
[17] Markelić, I., Zahedi, K.: An evolved neural network for fast quadrupedal locomotion. In: M. Xie,
S. Dubowsky (eds.) Advances in Climbing and Walking Robots, Proceedings of 10th International
Conference (CLAWAR 2007), pp. 65–72. World Scientific Publishing Company (2007)
[18] Montúfar, G., Ghazi-Zahedi, K., Ay, N.: A theory of cheap control in embodied systems. PLoS
Comput Biol 11(9), e1004,427 (2015)
[19] Nicholas Cheney, J.C., Lipson, H.: Evolved electrophysiological soft robots. In: Proceedings of the
Fourteenth International Conference on the Synthesis and Simulation of Living Systems (ALIFE
14) (2014)
[20] Nolfi, S., Floreano, D.: Evolutionary Robotics. MIT Press (2000)
[21] Open Robotics Foundation: (2016). URL http://gazebosim.org
[22] Rempis, C., Thomas, V., Bachmann, F., Pasemann, F.: NERD Neurodynamics and Evolutionary
Robotics Development Kit, pp. 121–132. Springer Berlin Heidelberg, Berlin, Heidelberg (2010)
[23] Risi, S., Stanley, K.O.: An enhanced hypercube-based encoding for evolving the placement, density,
and connectivity of neurons. Artif Life 18(4), 331–63 (2012)
[24] Secretan, J., Beato, N., D’Ambrosio, D.B., Rodriguez, A., Campbell, A., Folsom-Kovarik, J.T.,
Stanley, K.O.: Picbreeder: A case study in collaborative evolutionary exploration of design space.
Evolutionary Computation 19(3), 373–403 (2011)
[25] Sims, K.: Evolving 3d morphology and behavior by competition. Artificial Life 1(4), 353–372
(1994)
[26] Sims, K.: Evolving virtual creatures. In: SIGGRAPH ’94: Proceedings of the 21st annual conference
on Computer graphics and interactive techniques, pp. 15–22. ACM, New York, NY, USA (1994)
[27] Stanley, K.O., Miikkulainen, R.: Evolving neural networks through augmenting topologies. Evol.
Comput. 10(2), 99–127 (2002)
[28] Streeting, S., Xie, J., Castaneda, P.J., Muldowney, T., Doyle, J., O’Sullivan, J., van der Laan,
W.J.: Ogre. http://www.ogre3d.org (2006)
18
[29] Twickel, A., Büschges, A., Pasemann, F.: Deriving neural network controllers from neuro-biological
data: implementation of a single-leg stick insect controller. Biological Cybernetics 104(1-2), 95–
119 (2011)
[30] von Twickel, A., Hild, M., Siedel, T., Patel, V., Pasemann, F.: Neural control of a modular multilegged walking machine: Simulation and hardware. Robotics and Autonomous Systems 60(2), 227
– 241 (2012)
[31] Zahedi, K., Martius, G., Ay, N.: Linear combination of one-step predictive information with an
external reward in an episodic policy gradient setting: a critical analysis. Frontiers in Psychology
4(801) (2013)
A
Svenja specification
Bounding Box (x,y,z)
Head
0.614 m 0.614 m 0.812 m
Main Body
0.976 m 1.107 m 0.552 m
0.832 m 1.643 m 0.860 m
Rear
0.612 m 0.945 m 0.977 m
Front femur
Front tibia
0.204 m 0.297 m 0.901 m
Front tarsus
0.611 m 1.026 m 0.444 m
Centre femur 0.767 m 0.230 m 1.972 m
Centre tibia
0.103 m 0.094 m 0.909 m
Centre tarsus 0.973 m 0.083 m 0.436 m
Rear femur
0.804 m 1.283 m 1.004 m
Rear tibia
0.273 m 0.417 m 0.897 m
Rear tarsus
0.729 m 1.230 m 0.446 m
αmin αmax
Fmax
◦
◦
Front Ma-Fe 1
-5
20
7.5 N/m
Front Ma-Fe 2
-25 ◦
35 ◦
7.5 N/m
Front Fe-Ti
-10 ◦
10 ◦
7.5 N/m
◦
◦
Front Ti-Ta
-30
5
7.5 N/m
◦
◦
Centre Ma-Fe 1 -0.5
45
7.5 N/m
◦
◦
Centre Ma-Fe 2 -45
45
7.5 N/m
Centre Fe-Ti
-30 ◦
5◦
7.5 N/m
◦
◦
Centre Ti-Ta
-2.5
45
7.5 N/m
◦
◦
Rear Ma-Fe 1
-5
20
7.5 N/m
◦
◦
Rear Ma-Fe 2
-30
20
7.5 N/m
◦
◦
Rear Fe-Ti
-5
27.5
7.5 N/m
◦
◦
Rear Ti-Ta
-25
20
7.5 N/m
Mass
10.0 kg
10.0 kg
20.0 kg
2.0 kg
1.5 kg
1.0 kg
2.0 kg
1.5 kg
1.0 kg
2.0 kg
1.5 kg
1.0 kg
Vmax
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
0.5 rad/s
Table 3: Svenja’s specification. Please note, that the domains for the angular position are given for the
legs on the right side. Depending on how the joint is attached, the values have different meaning, i.e.,
in which direction the leg will move give an specific value. The numbers are presented here to given an
impression of the leg’s moving ranges.
19
| 9cs.NE
|
A DISCRETE ANALYSIS OF METAL-V BELT DRIVE
A. KARAM and D. PLAY
Université Libanaise
Faculté des sciences II
FANAR-MATEN B.P. 90656
Lebanon
Abstract
The metal-V belt drive includes a large number of parts
which interact between them to transmit power from the input to the
output pulleys. A compression belt composed of a great number of
struts is maintained by a tension flat belt. Power is them shared into
the two belts that moves generally in opposite directions. Due to the
particular geometry of the elements and to the great number of parts,
a numerical approach achieves the global equilibrium of the
mechanism from the elementary part equilibrium. Sliding arc on each
pulley can be thus defined both for the compression and tension
belts. Finally, power sharing can be calculated as differential motion
between the belts, is defined. The first part of the paper will present
the different steps of the quasi-static mechanical analysis and their
numerical implementations. Load distributions, speed profiles and
sliding angle values will be discussed. The second part of the paper
will deal to a systematic use of the computer software. Speed ratio,
transmitted torque, strut geometry and friction coefficients effect
will be analysed with the output parameter variations. Finally, the
effect pulley deformable flanges will be discussed.
I - Introduction
The metal V-belt of VAN-DOORNE type [1], is used in
C.V.T. mounted on cars. The pulleys are kept at a fixed axis
distance and are subjected to axial forces which can be varied in
order to adapt the radius of the belt wrapped on the pulleys so as to
obtain different speed ratios.
The belt (Fig. 1 a-b) consists of one set of thin flexible steel
bands (or flat belt) on which rides a large number of small steel
blocks (struts), whose lateral surfaces are in contact with the flanges
of the pulleys. During the motion, a strut on the driving pulley is
carried forward by the pulley via the interfacial friction between the
pulley and the strut. Thus, a strut is pressed on the preceeding strut,
generating a compressive force between them. This compressive
force increases as the struts travel from the leading edge to the
trailling edge of the driving contacting arc. On the driven pulley, this
compressive force decreases until the end of the contact arc. Power
is then transmitted from the driving pulley to the driven pulley
through the linear compressed struts zone (upper part of the figure,
Fig.1). Further, the flat belt do not only serve as a track constraints
for the travelling struts, but also works as a part of the power
transmission. The tight side of the flat belt is generally located at the
lower linear zone (Fig.1).
Fig. 1: Geometry and Modelization of C.V.T. transmission
A fine knowledge of the behaviour of each mechanical part is
obviously necessary to insure the life and the reliability of the
system. As the experimental tests become more and more expensive,
numerical models permit to give data for design considerations. Thus
final experimental tests serve only for last minor modifications.
Moreover, optimization of the car performances passes through the
interactions between the mechanical system and the microprocessored actuators. For exemple, the axial force applied on the
moveable flanges of the pulleys must be adapted in order to insure
the transmitted torque conditions with a maximum life of the
elements. Thus, the knowledge of mechanical governing parameters
becomes necessary to design efficient actuators.
The litterature on metal V-Belt behaviour appeared during
the last decade [3-4-5]. All the papers are based on the wellknown
approach of the V-belt behaviour proposed by Gerbert [6]. Their
purposes were to establish models for this highly decretized V-belt.
Various hypothesis have been made in order to get a first
understanding of the behaviour. The purpose of this paper is to make
a generalization of a model proposed in 1987 by the author [5] in
order to obtain more details on the mechanical parameter variations.
New geometries are also included and consequently the equations
are redevelopped. This paper will present:
- a theoretical mechanical analysis with the classical hypothesis of
general mechanics,
- a calculating numerical process and the resulting computing
architecture,
- results and discussion.
II - Theoretical analysis and resolution
The modelling of a mechanical system is complicated as the
number of parts is large. The interaction between parts are also
amplified if the stiffness variations gives changes of the initial
geometry. On the other hand, into a metal V-belt drive, the set of
parts is moving and the relative motions depend on the working
conditions and the whole mechanical equilibrium. It is then necessary
to describe simultaneously the equilibrium of all parts and the
kinematic conditions. Due to the large number of parts, the
resolution of the equations must be numerical. As the initial
conditions are not fixed, the numerical resolution is iterative. In this
first paragraph, the steps of calculations are summarized in a such
way to facilitate the numerical resolution. The details of the
modelization are given elsewhere [7].
II-1 Hypotheses
The following hypotheses have been made in order to
modelize the system:
- the steel struts are always in compression that means an initial
tension is considered when the metal V-belt is at rest,
- the two sets of bands are identified as one flat belt which holds the
struts in a symetrical manner, the analysis is thus bi-dimensionnal,
- the flat belt is very rigid and its length is thus, considered constant
during the motion. The strut are seeing more deformable due to
their geometry.
- the flanges of the pulleys will be considered rigid during the
analysis. Note that a discussion on the effect of the deformable
flanges will be made at the end of the paper.
The model takes also into consideration the following
assumptions:
- the coefficients of friction of flat belt/struts and struts/pulleys is
considered through the Coulomb model at this level of the
mechanical analysis of the behaviour,
- the radial and tangential displacements of struts occur relatively to
the flat belt and the flanges of the pulleys,
- full rotations of struts take place at their entry in the contacting arc,
without futher variation of angular position until the end of the
contact,
- the strains of struts follow a classical bulk material law.
After these hypotheses, the theorical analysis is detailled in three
paragraphs:- Strut equilibrium - flat belt equilibrium - global
equilibrium.
II-2 Strut equilibrium
During the study of strut equilibrium, the following simplified
hypotheses are established:
- a strut stays in a radial position, AiBi constitutes linear cord, AiA'i
and BiB'i are supposed to be transversal contacts lines between
the struts [(i+1),i] and [i,i-1],
- the action between the pulley and a strut is characterized by a
squew applied on the contact pulley strut surface at the Mpi point
(fig. 2-a),
- the inertia forces, operating only on the contacting arcs of the
pulleys are radial.
Fig. 2: Free body definition
By application of general mechanical rules, the following
equations are obtained:
-Xi-1 + Xi cos(j) + 2 XLi + 2 Xpi = 0
Xi sin(j) + 2YLi + 2YPi + YGi = 0
-2ypi XPi + 2xpi YPi + xG YGi + 2xli YLi + R(cos(j)-1)Xi-1 +
2CPi = 0
where (xpi,ypi) is a coordinate of MPi, xG is an abscissa of G, xli =
l/2, torque CPi is applied on a strut. Thus, this third equation is used
only for the determination of CPi.
II-3 Belt equilibrium
The internal forces considered into the flat belt are
longitudinal traction forces. Bending strengh and cordal effect are
neglected. The equilibrium of a flat belt element which spreads out
on a strut (fig. 2-b) gives the following equations:
Fi-1 cos(j) - Fi - 2 XLi
Fi-1 sin (j) + 2 YLi = 0
=0
The forces XLi and YLi applicated by the strut on the flat
belt element are related to the Coulomb's model and to the direction
of the relative motion between the two mechanical elements.
XLi = µL Li YLi where Li is a coefficient equal to 1 whose the
sign is defined by the relative motion.
By substitution, the preceeding equations give:
Fi = ALi Fi-1 where ALi = (cos i + µL Li sin i)
If N1, N2 are respectively the number of struts into the
contact arc 1 and 2, the global equilibrium of the flat belt is
expressed by:
FN2,2 = Fsup = F1,1
FN1,1 = Finf = F1,2
For j = 1,2 (j=1 : driven pulley, j=2 : driving pulley):
Fi+1,j = ALi,j Fi,j. Then writing all the forces:
N1
Finf = II ALi,1 Fsup;
i=1
N2
Fsup = II ALi,2 Finf
i=1
The result of the substitution is:
N1
N2
II ALi,1 II ALi,2 = 1
i=1
i=1
This equation will be used later to determinate the average velocity
VL of the flat belt.
II-4 Struts/pulleys Interaction
The contact between pulley/strut generates tangential force
TPi and a normal force NPi which are related by the Coulomb's
model: |TPi| < µp |NPi|.
TPi is situed into the contact plane (strut-pulley). It makes an angle
pi (called sliding angle) with the tangential direction (Fig. 3-a). The
components are related by:
YPi = NPi sin + TPi cos sin(
XPi = TPi cos pi )
ZPi = NPi cos - TPi sin sin(
RPi = XPi tg( pi )
pi
pi
)
)
a) Slipping of struts between the conic flanges of the pulleys.
b) Velocity profile of struts and flat belt.
fig. 3: Modelization of kinematic conditions
Each contact arc is subdivised in two zones:
- an adherence zone, where the actions between the struts X i and the
actions of pulleys on these struts stay constant:
Xi = Xi-1
ZPi = ZPi-1
- a sliding zone, where the forces decrease on the driven pulley and
increase on the driving pulley.
XPi =µpi YPi
with µpi=[µp cos( pi )]/[sin+ µp sin( pi ) cos]
The sliding angle pi is defined by the deformations of a strut from a
position (i) to an other position (i+1):
= Arctg[-GRi/GTi] where GRi and GTi are repectively the
radial and tangential motion of the strut. Note that GT and GR are
calculated from a global experimental nonlinear relation between
load and displacement (Fig. 3-a).
pi
II-5 Strut Velocities
The struts keep a constant velocity along the adherence
zones of the contact arcs. Along the sliding zones, the velocity varies
in an opposite manner to the actions Xi. Eurther, we assure that GR
and GT are independant parameters.
VTi = li/( t) = VTi-1 +( li)/( t) so:
VTi = ([1 - ( li-1)/l]/[1 - ( li)/l]) VTi-1
The velocity profile (Fig. 3-b) can be deduced. Note that at
each entry and exit of the belt circular zones on the pulleys, a
discontinuity apears because of the strut contact location change.
Thus, a parameter a/Ri must be introduced.
II-6 Flat belt velocity
The flat belt is supposed to be rigid longitudinally comparing
to the strut deformations. A constant average velocity VL is thus
considered. A presupposed known velocity profile is considered
(that means a known struts velocity) in order to calculate this
velocity (Fig. 3-b). The two struts NL1 and NL2 are considered to
have the same velocity which is the looked for one. With the global
equilibrium equation of the flat belt and the passing over all contact
arcs, determination of VL becomes possible.
II-7 Global Equilibrium of C.V.T.
The two linear strands and two contact arcs constitute the
four zones along the belt. The flat belt equilibrium in each of these
zones can be described by knowing the forces imposed on the
extremites of the considered zones (fig. 1-b). By isolating the two
parts (driving and driven) of the C.V.T., the following equations are
written:
FAXE = (Fsup + Finf - Psup - Pinf) sin(1)
C2 = (R2 + GR2) Psup - R2 Pinf - (R2 + a + GR2)Fsup+
(R2 + a)Finf
C1 = (R1 + GR1)Pinf -R1 Psup - (R1 + a + GR1)Finf+
(R1 + a)Fsup
II-8 Power and efficency calculation
These calculations are made when all the system equilibriums
are established. If the torque on the two pulleys and the rotation
velocity of each pulley are knwon, the power calculation is obvious:
PW2 = C2 W2 and PW1 = C1 W1
The efficency is given by the following expression:
µ = -PW1/PW2
In the upper and lower strands, the forces PSUP, PINF,
FSUP and FINF are transmitted from the driving pulley to the driven
pulley. Then, in these strands, no power loss is considered. The
global loss of power is obtained by PW1 + PW2 with PW1 < 0.
II-9 Method of Resolution and Computer Program
The system of equations is solved in order to obtain all
desired results. Considering the complexity of the system and the
dependence between all variables, an iterative solving method is
used. It always insures:
- a forces equilibrium:
N2
Zpi(i, 2) FAXI
i 1
- a geometrical equilibrium | LM| < where LM is the difference
between the length of the flat belt and the sum of strut
thicknesses. Note that the strut thickness variations are calculated
with the forces Xi from a non linear experimental compression
curve.
The calculation is done in three steps (Fig. 4):
Fig. 4 : Flow chart of the computer softwere
- Input data by reading data files (INPUT),
- variables initialization considering a rigid model (CINIT),
- following variables calculations are made with iterative methods
(CITER).
The iterative process of the system equilibrium is done then into
the computer module CITER. This module calculates:
- NG and NL by the equilibrium of torque (for NG) and the tension
in the belt (for NL),
- FAXI, the force imposed on the driving pulley, and FAXE a force
inter-axes pulleys (Loop 2). The numerical secant methode has
been used in order to obtain the convergence between the sum of
the calculated individual strut-pulley interactions ZPi and the
imposed input value of FAXI,
- determination of FSUP which consists in turn of the determination
of action XL, YL, PSUP, PINF, FINF, Xi, XP, YP, and the
calculation of radial and tangential slidings of struts (that means
pi angles).After the achievement of the geometric compatibility,
the angle pi along the adherent zone are calculated from the
computing module CRAD1. Finally, the velocity of struts and flat
belt, the power distribution, the efficency, the relative motion of
flat belt in relation to the struts and the distribution of powers are
calculated.
The power distribution can be visualized (Fig.5). Note that for the
same input power (15706 W), two input mechanical conditions
give two different power distributions:
- in the figure 5-a, the input power is subdivized into the strut-power
and flat belt power. As the struts along the driving contact arc
drive the flat belt, the same counter clock wise power circulation
is observed,
- in contrary in the figure 5-b, the struts along the driven contact arc
drive the flat belt. Then, in the driving zone, the strut power is the
sum of the input power and the flat belt power.
(a)- (R2R1=1, C2=60 Nm, MUL=0,01, MUP=0,12)
(b)- (R2R1=1,4, C2=60Nm, MUL=0,01, MUP=0,12)
Fig. 5 : The power distribution between different parts of system
III - Results and discussion
For one reduction ratio, a fixed inter-axe, an input torque, an
axial force imposed on the driving pulleys and an initial load F0 (An
initial load is necessary to insure the initial compression of the
struts), the computer program gives the following results (Fig. 6):
- tension in flat belt Fig. 6-a),
- normal action pulley/struts (NP Fig. 6-b),
- sliding angle ( p Fig. 6-c),
- velocity of struts (VT Fig. 6-d),
- pressure between struts (X Fig. 6-e),
- amplitude of the sliding displacement between struts and pulleys Y
[ Y GT2 + GR2 ) ] (Fig. 6-f).
- tangential action flat belt/struts (XL Fig. 6-g),
- circumferential action pulleys/struts (XP Fig. 6-h),
normal action pulleys/struts NPi
2150
tension in belt
2140
2130
2120
2110
2100
2090
2080
2070
0
20
40
60
80
0
-50
0
20
40
60
80
-100
-150
120
100
80
60
0
20
40
60
80
0
20
40
60
80
14,62
14,60
14,58
14,56
14,54
1000
800
600
400
200
0
0 10 20 30 40 50 60 70 80
sliding modulus struts/pulley Y
1200
1,6e-6
1,4e-6
1,2e-6
1,0e-6
8,0e-7
6,0e-7
4,0e-7
2,0e-7
0,0
0
circ. action pulleys/struts
pressure between struts X
tangential action belt/struts XL
140
14,52
-200
0,6
0,4
0,2
0,0
160
14,64
50
velocity of struts VT
sliding angle
100
180
0 10 20 30 40 50 60 70 80
-0,2
-0,4
-0,6
20
40
60
80
20
10
0
0
10
20
30
40
50
60
-10
-20
-30
Driving Pulley
Driven Pulley
Fig. 6 : Numerical results issued from the computer, versus the
position of struts along the contact arc (R2R1 = 1, C2 = 60 Nm,
MUL = 0,01, MUP = 0,12)
Note that, on the figure 6:
- the pressure X between struts is constant along the adherence
zone. It increases after that on the driving pulley (from the strut
NG2) and decreases on the driven pulley (from strut NG1),
- a change of the sign of the tangential action (XL) when the relative
motion flat belt /struts is reversed,
- the circonferencial action pulleys/struts (XP) depends on the sliding
direction of struts in relation to pulleys. This action is positive on
the driven pulley (the struts slides forward this pulley) and
negative on the driving pulley (the struts slides backward),
- the tension in the flat belt increases until NL1 and decreases after
that to reach the value FSUP. On the driven pulley, this
70
80
evaluation is inverted on the driving pulley : decreasing until the
strut NL2, then increasing to arrived to the FINF value. Note that
the ratio Finf/Fsup is very dependant of the coefficient of friction
µL between belt/struts [7] while it is practically independant of
the others parameters,
- the normal action pulley/struts (NP) is constant in the adherence
zone (like the pressure between the struts). On the driving pulley,
this action decreases, but it increases in the driven pulley.
Therefore, this value stays more important on the driving pulley,
- the struts velocity VT varies in the same way as the normal action.
In the adherence zones, velocity does not vary. In the figure, note
that the velocity along the linear strand is not represented,
- the sliding angle ( p) on the driven pulley is comprised between 0°
and 90°, the strut is backward pulley. On the driving pulley, it is
in the sliding zone between -90° and -180°, the strut is then
forward,
- amplitude of the sliding displacement of a strut in relation to
VT and p increases on the driven pulley (in the sliding zone)
and decreases in the driving pulley.
The computer program gives also other results as the efficency of
the mechanical system, the percentage of the transmitted power
by the flat belt and the struts, the way of distribution of power in
the different parts of the system, the relative motion between the
flat belt and the struts.
III-1 Influence of input parameters
The input parameters such as speed or reduction ratio R2R1,
input torque C2, the friction coefficients MUL and MUP influence
the results. In order to know the variation of the following results:
- efficency,
- percent of power transmitted by the flat belt,
- tension in the lower strand of the flat belt,
- pressure between the struts in the upper strand.
- axial force imposed on the flanges of driving pulley,
systematic numerical tests have been made in order to visualize
tendencies depending on these parameters. When one parameter
varies the others are kept constant and equal to : R2R1=1,
C2=80Nm, MUL=0,01, MUP=0,12.
III-2 The Efficiency
In all cases, the general efficiency is high (more than 97%
approximately). It is due to the hypothesis on power loss where
different causes are ignored (contribution of friction between
struts, hysteresis of materials, polygonal effect). Efficiency varies a
little with the input parameters. In fact, for a reduction ratio R2R1
smaller than 1, the efficiency increases, but for a mulptiplicator
C.V.T., the efficiency decreases with the ratio (Fig.7). This result is
in accordance with experimental results and with conclusions of the
litterature. For a ratio equal 1, the efficiency is pratically constant
versus the torque (fig.7) and with the struts/pulley friction coefficient
(fig.7) and increases with the flat belt/struts coefficient of friction.
Fig. 7 : Variations of C.V.T. efficiency ETA with design parameters
III-3 Percentage of transmitted power by the flat belt (%AME)
Lubrication govern obviously the C.V.T. behaviour. In a
global approach, average coefficients of friction have been
considered. The distribution of the power between the different parts
of the system is essentially related to the friction coefficients
between the flat belt/struts (MUL) and struts/pulleys (MUP).
Considering the friction coefficient between the flat belt and the
struts, the percentage of transmitted power by the flat belt increases
in relation to the coefficient (MUL). The parameter %AME
decreases for a ratio R2R1 which increases between 0,8 and 1,4 and
for a coefficient of friction struts/pulley between 0,05 and 0,12 (fig.
8).
Fig. 8 : Variations of the inner power %AME transmitted by the flat belt
III-4 The tension in the lower strand of the flat belt (FINF)
The input torque is the more influent parameter on FINF.
This tension FINF increases with this torque C2 (Fig. 9) as the
percentage of power was shown to increase with torque. It
decreases with the R2R1 ratio and the increasing of the
struts/pulleys and flat belt/struts friction coefficients.
Fig. 9 : Elongation force Finf into the lower part of the belt
III-5 The pressure between the struts in the upper strand
(PSUP)
The pressure between the more compressed struts (in the
upper strand) decreases with an increasing of reduction ratio (Fig.
10) and flat belt/struts friction coefficient (Fig. 10). This pressure
increases with the input torque (Fig. 10) and the struts/pulleys
friction coefficient(Fig. 10).
Fig. 10 : Elongation fore Psup into the upper part of the struts
III-6 The axial force imposed on the flanges of driving pulley
(FAXI)
Remember that the first part of this paper considers rigid
flanges of the pulleys. The force applied on the flanges of the pulleys
FAXI is governed by the input parameters (Fig. 11). FAXI decreases
when the ratio R2R1 increases. The same trend is observed with the
variations of the friction coefficients. FAXI increases obviously with
torque.
Fig. 11 : Variation of C.V.T. axial force FAXI with design parameters
III-7 The initial tension F0
The initial tension F0 is acheived by adding struts when the belt
have a global circular shape before mounting on the pulleys. Two
zones are isolated from the figure, the larger F0 gives the higher
efficiencies. The other parameters increases with F0 (Fig. 12). From
a global point of view, the non linear deformations of the struts ask
for a sufficient initial F0 in order to limit compressive displacement
variations. It must be also pointed out that the flat belt becomes
influent only for large F0.
Fig. 12 : Initial tension F0 influence
IV- Discussion
Incerted in the contacting arc, the strut (i) exerts an axial
force ZPi on the flange of the pulley. This forces provokes a
deformation, and concequently, the flange displaces axially of the
quantity EKF(ZPi). so, the flange has gaps with respect to the
position (i-1) of the quantity:
EKF(ZPi) - EKF(ZPi-1) where EKF is the fonction of the flange
displacement by terms of the applied force of the pulley and the
considered strut. This function has been determinated by the Finite
Element program (AMME) [8,9,10]). It is noted that this function is
not lineare. The deformation of the flange leads an additional sliding
of the strut in the pulley grouve. In effect, the position (i-1) to
another one (i), the flanges of the pulley have gaps of the quantity
ai ( ai = 2[EKF(ZPi) - EKF(ZPi-1)]). It provokes a radial sliding
of the strut with a quantity GRPi:
GRPi = - [ ai]/[2sin()]. The deformation of the struts provokes
the radial sliding at the quantity GRMi:
GRMi = -[EKL(ZPi)-EKL(ZPi-1)]/[2sin()]. So the radial sliding
GRi becomes the som: GRi = GRPi + GRMi.
In the adherent zone, we supposed that the force ZPi rests
constants because the displacements induced by the flange
deformations are small. Consequently, a radial sliding curve of the
struts (due to the pulley deformation) is obtained.
V- Conclusion
The metal V-belt is a highly discretized mechanism with a
large number of parts. The relative motion of the elements adds to
the difficulties of the mechanical behaviour modelization.
A theoretical study has been made with the equations of the
general mechanics in a quasi-static working case. The originality of
this approach lies on the consideration of the whole system
constituted by an large number of interactive components. A
computer software has been developped in order to make numerical
simulations as the cost of experimental simulations become too
expensive. The results have shown that:
- the distribution of power between the different parts of the system
is governed by the control of the friction coefficients,
- the efficiency of the system depends essentially on the rigidity of
the materials of struts. The use of less rigid materials than the
steel struts reduce the efficiency,
-the initial tension F0 has an influence on the results from a lower
limit.
Acknowledgement
A part of this work have been made under industrial contract
and the authors wish to thank Mr Douhairet and Trinquard for their
helpfull comments.
References
[1] - VAN DER VEEN S.C. "Transmatic variable drive
transmission" Hydraulic Pneumatic Mechanical Power Drives,
Transmissions and Control Vol. 23. N°268 April 77 P 133-140.
[2] - SWAIN J.C., KLAUSING T.A., WILCOX J.P. "Design
study of steel V-Belt C.V.T. for electric vehicles" NASA CR159845
June 1980, 122 P.
[3] - GERBERT, B.G. "Metal-V-Belt Mechanics", 1984, ASME
paper 84-DET-227 9 P.
[4] - SUN, D.C. "Performance Analysis of Variable Speed-Ration
Metal V-Belt Drive". J. Mech. Transm. Autom. Des., Transactions
of the ASME Vol. 110, December 1988, pp 472-481.
[5] - CARVAJAL, J.M. VILLECHAISE, B. PLAY, D. "C.V.T.
transmissions analysis, A Mechanical discrete analysis with
computer". Proceedings of the 1987 ASME International
Computers in Engineering Conference August 1987, New York,
vol. 3, pp. 461-465
[6] - GERBERT, B.G. "Force and slip Behaviour in V-belt
Drives",
Acta
polytechnica Scandinavica,
Mechanical
engineering series N° 67 Helsinki 1972, 101 P.
[7] - KARAM A. "prévision de comportement et conception de
variateur continu de vitesses. Application courroies trapézoïdales
poussantes et tirantes." PhD INSA LYON July 1991. N° 91 ISAL
0051, 222 P.
[8] - RIGAL J.F. "ANALYSE ET MODELISATION DES
SYSTEMES MECANIQUES EN CONCEPTION ASSISTEE PAR
ORDINATEUR. Application à la simulation du comportement
statique des boîtes de transmission de puissance d'hélicoptères" PhD
INSA LYON december 1989. N° 89 ISAL 0095, 243P.
[9] - RIGAL J.F. and PLAY D. "Preliminary Design and Load
Distribution of High Performance Mechanical Systems." ASME
Congress: Computers in Engineering 89, Anheim Californie, USA,
1989, Vol.2,p.93-103
[10] - RIGAL J.F. and PLAY D. "Numerical Model Definition in te
First Stages of Computer Aided Design." ASME Congress:
Computers in Engineering 90,Boston, USA, 1990, Vol. 1, p. 309315
| 5cs.CE
|
Noname manuscript No.
(will be inserted by the editor)
Twitter Sentiment Analysis: How To Hedge Your
Bets In The Stock Markets
Tushar Rao · Saket Srivastava
arXiv:1212.1107v1 [cs.CE] 5 Dec 2012
Received: 5th December 2012
Abstract Emerging interest of trading companies and hedge funds in mining
social web has created new avenues for intelligent systems that make use of
public opinion in driving investment decisions. It is well accepted that at high
frequency trading, investors are tracking memes rising up in microblogging
forums to count for the public behavior as an important feature while making
short term investment decisions. We investigate the complex relationship between tweet board literature (like bullishness, volume, agreement etc) with the
financial market instruments (like volatility, trading volume and stock prices).
We have analyzed Twitter sentiments for more than 4 million tweets between
June 2010 and July 2011 for DJIA, NASDAQ-100 and 11 other big cap technological stocks. Our results show high correlation (upto 0.88 for returns)
between stock prices and twitter sentiments. Further, using Granger’s Causality Analysis, we have validated that the movement of stock prices and indices
are greatly affected in the short term by Twitter discussions. Finally, we have
implemented Expert Model Mining System (EMMS) to demonstrate that our
forecasted returns give a high value of R-square (0.952) with low Maximum
Absolute Percentage Error (MaxAPE) of 1.76% for Dow Jones Industrial Average (DJIA). We introduce a novel way to make use of market monitoring
elements derived from public mood to retain a portfolio within limited risk
state (highly improved hedging bets) during typical market conditions.
Keywords Stock market · sentiment analysis · Twitter · microblogging ·
social network analysis
T. Rao
Netaji Subhas Institute of Technology, Delhi, India
E-mail: [email protected]
S. Srivastava
Indraprastha Institute of Technology, Delhi, India
E-mail: [email protected]
2
Tushar Rao, Saket Srivastava
1 INTRODUCTION
Financial analysis and computational finance have been an active area of research for many decades[18]. Over the years, several new tools and methodologies have been developed that aim to predict the direction as well as range of financial market instruments as accurately as possible[17]. Before the emergence
of internet, information regarding company’s stock price, direction and general
sentiments took a long time to disseminate among people. Also, the companies
and markets took a long time (weeks or months) to calm market rumors, news
or false information (memes in Twitter context). Web 3.0 is characterized with
fast pace information dissemination as well as retrieval [7]. Spreading good or
bad information regarding a particular company, product, person etc. can be
done at the click of a mouse [8], [1] or even using micro-blogging services such
as Twitter[26]. Recently scholars have made use of twitter feeds in predicting box office revenues [2], political game wagons [30], rate of flu spread [29]
and disaster news spread [12]. For short term trading decisions, short term
sentiments play a very important role in short term performance of financial
market instruments such as indexes, stocks and bonds [20].
Early works on stock market prediction can be summarized to answer the
question - Can stock prices be really predicted? There are two theories - (1)
random walk theory (2) and efficient market hypothesis (EMH)[25]. According
to EMH stock index largely reflect the already existing news in the investor
community rather than present and past prices. On the other hand, random
walk theory argues that the prediction can never be accurate since the time
instance of news is unpredictable. A research conducted by Qian et.al. compared and summarized several theories that challenge the basics of EMH as
well as the random walk model completely[27]. Based on these theories, it has
been proven that some level of prediction is possible based on various economic and commercial indicators. The widely accepted semi-strong version of
the EMH claims that prices aggregate all publicly available information and
instantly reflect new public version[21]. It is well accepted that news drive
macro-economic movement in the markets, while researches suggests that social media buzz is highly influential at micro-economic level, specially in the
big indices like DJIA [5], [22], [15] and [28]. Through earlier researches it has
been validated that market is completely driven by sentiments and bullishness
of the investor’s decisions [27]. Thus a comprehensive model that could incorporate these sentiments as a parameter is bound to give superior prediction
at micro-economic level.
Earlier work done by Bollen et. al. shows how collective mood on Twitter
(aggregate of all positive and negative tweets) is reflected in the DJIA index
movements [5]and [22]. In this work we have applied simplistic message board
approach by defining bullishness and agreement terminologies derived from
positive and negative vector ends of public sentiment w.r.t. each market security or index terms (such as returns, trading volume and volatility). Proposed
method is not only scalable but also gives more accurate measure of large
scale investor sentiment that can be potentially used for short term hedging
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
3
strategies as discussed ahead in section 6. This gives clear distinctive way for
modeling sentiments for service based companies such as Google in contrast
to product based companies such as Ebay, Amazon and Netflix. We validate
that Twitter feed for any company reflects the public mood dynamics comprising of breaking news and discussions, which is causative in nature. Therefore
it adversely affects any investment related decisions which are not limited to
stock discussions or profile of mood states of entire Twitter feed.
In section 2, we discuss the motivation of this work and related work in the
area of stock market prediction in section 3. In section 4 we explain what and
how of the techniques used in mining data and explain the terminologies used
in market and tweet board literature. In section 5 we have given prediction
methods used in this model with the forecasting results. In section 6 we discuss
how Twitter based model can be used for improving hedging decisions in a
diversified portfolio by any trader. Finally in section 7 we discuss the results
and in section 8 we present the future prospects and conclude the work.
2 MOTIVATION
”Communities of active investors and day traders who are sharing opinions
and in some case sophisticated research about stocks, bonds and other financial instruments will actually have the power to move share prices ...making
Twitter-based input as important as any other data to the stock”
-TIME (2009) [24]
High Frequency Trading (HFT) comprises of very high percentage of trading volumes in the present US market. Traders make an investment position
that is held only for very brief periods of time - even just seconds - and rapidly
trades into and out of those positions, sometimes thousands or tens of thousands of times a day. Therefore the value of an investment is as good as last
known index price. Investors will do anything that will give them an advantage in placing market bets. A large percentage of high frequency traders in
US markets, have trained AI bots to capture buzzing trends in the social media feeds without learning dynamics of the sentiment and accurate context of
the deeper information being diffused in the social networks. For example, in
February 2011 during Oscars when Anne Hathaway was trending, stock prices
of Berkshire Hathaway rose by 2.94% [3] Figure 1 highlight the incidents when
the stock price of Berkshire Hathaway jumped coinciding with an increase of
buzz on social networks/ micro-blogging websites regarding Anne Hathaway
(for example during movie releases).
The events are marked as red points in the Figure 1 , event specific news
on the pointsA: Oct. 3, 2008 - Rachel Getting Married opens: BRK.A up 0.44%
B: Jan. 5, 2009 - Bride Wars opens: BRK.A up 2.61%
C: Feb. 8, 2010 - Valentine’s Day opens: BRK.A up 1.01%
D: March 5, 2010 - Alice in Wonderland opens: BRK.A up 0.74%
4
Tushar Rao, Saket Srivastava
Fig. 1 Historical chart of Berkshire Hathaway(BRK.A) stock over the last 3 years. Highlighted points (A-F) are the days when its stock price jumped due to an increased news
volume on social networks and Twitter regarding Anne Hathaway. Courtesy Google Finance.
E: Nov. 24, 2010 - Love and Other Drugs opens: BRK.A up 1.62%
F: Nov. 29, 2010 - Anne announced as co-host of the Oscars: BRK.A up 0.25%
G: Feb. 28. 2011 - Anne hosts Oscars with James Franco: BRK.A up 2.94%
As seen in this example, large volume of tweets can create short term influential effects on stock prices. Events such as these motivate us to investigate
deeper relationship between the dynamics of social media messages and market movements [18]. This work is not directed to find a new stock prediction
technique which will counter in the effects of various other macroeconomic
factors.
The aim of this work, is to quantitatively evaluate the effects of twitter sentiment dynamics around a stocks indices/stock prices and use it in conjunction
with the standard model to improve the accuracy of prediction. Further in section 6 we investigate into how tweets can be very useful in identifying trends
in futures and options markets and to build hedging strategies to protect one’s
investment position in the shorter term.
3 RELATED WORK
There have been several works related to web mining of data (blogposts, discussion boards and news) [13], [4], [15] and to validate the significance of
assessing behavioral changes in the public mood to track movements in stock
markets. Some trivial work shows information from investor communities is
causative of speculation regarding private and forthcoming information and
commentaries[19], [31],[10] and [9]. Dewally in 2003 worked upon naive momentum strategy confirming recommended stocks through user ratings had
significant prior performance in returns [11]. But now with the pragmatic shift
in the online habits of communities around the worlds, platforms like StockTwits1 [32] and TweetTrader2 have come up and their usage is virally spreading out. Das and Chen made the initial attempts by using natural language
1
2
http://stocktwits.com/
http://tweettrader.net/
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
5
Fig. 2 Flowchart of the proposed methodology showing the various phases of sentimental
analysis beginning with Tweet collection to stock future prediction. In the final phase 4
set of results have been presented:(1) Correlation results for twitter sentiments and stock
prices for different companies (2) Granger’s casuality analysis to prove that the stock prices
are affected in the short term by Twitter sentiments (3) Using EMMS for quantitative
comparison in stock market prediction using tweet features (4) Performance of Twitter
sentiment forecasting method over different time windows
processing algorithms classifying stock messages based on human trained samples. However their result did not carried statistically significant predictive
relationships [10].
Gilbert et.al. and Zhang et.al. have used corpus from livejournal blogposts
in assessing the bloggers sentiment in dimensions of fear , anxiety and worry
making use of Monte Carlo simulation to reflect market movements in S&P
500 index [15, 33]. Similar and significantly accurate work is done by Bollen et.
al who used dimensions of Google- Profile of Mood States to reflect changes
in closing price of DJIA [5]. Sprengers et.al. analyzed individual stocks for
S&P 100 companies and tried correlating tweet features about discussions of
the stock discussions about the particular companies containing the Ticker
symbol [28]. However these approaches have been restricted to community
6
Tushar Rao, Saket Srivastava
sentiment at macro-economic level which doesn’t give explanatory dynamic
system for individual stock index for companies. Thus deriving a model that
is scalable for individual stocks/ companies and can be exploited to make
successful hedging strategies as discussed in section 6.
4 WEB MINING AND DATA PROCESSING
In this section we describe our method of Twitter and financial data collection
as shown in Figure 2. In the first phase, we mine the tweet data and after
removal of spam/noisy tweets, they are subsequently subjected to sentiment
assessment tools in phase two. In later phases feature extraction, aggregation
and analysis is done.
4.1 Tweets Collection and Processing
Out of other investor forums and discussion boards, Twitter has widest acceptance in the financial community and all the messages are accessible through
a simple search of requisite terms through an application programming interface (API)3 . Sub forums of Twitter like StockTwits and TweetTrader have
emerged recently as hottest place for investor discussion buy/sell out at voluminous rate. Efficient mining of sentiment aggregated around these tweet feeds
provides us an opportunity to trace out relationships happening around these
market sentiment terminologies. Currently more than 250 million messages are
posted on Twitter everyday (Techcrunch October 20114 ).
This study was conducted over a period of 14 months period between
June 2nd 2010 to 29th July 2011. During this period, we collected 4,025,595
(by around 1.08M users) English language tweets Each tweet record contains
(a) tweet identifier, (b) date/time of submission(in GMT), (c) language and
(d)text. Subsequently the stop words and punctuation are removed and the
tweets are grouped for each day (which is the highest time precision window in
this study since we do not group tweets further based on hours/minutes). We
have directed our focus DJIA, NASDAQ-100 and 11 major companies listed
in Table 1. These companies are some of the highly traded and discussed
technology stocks having very high tweet volumes.
As seen in Figure 3, the average message volume for the 11 companies used
to validate the working model; is more than the average discussion volume
of DJIA and NASDAQ-100. In this study we have observed that technology
stocks generally have a higher tweet volume than non-technology stocks. One
3 Twitter API is easily accessible through an easy documentation available athttps://dev.twitter.com/docs. Also Gnip - http://gnip.com/twitter, the premium platform
available for purchasing public firehose of tweets has many investors as financial customers
researching in the area.
4 http://techcrunch.com/2011/10/17/twitter-is-at-250-million-tweets-per-day/
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
7
Table 1 List of Companies
Company Name
Ticker Symbol
Amazon
Apple
AT&T
Dell
EBay
Google
Microsoft
Oracle
Samsung Electronics
SAP
Yahoo
AMZN
AAPL
T
DELL
EBAY
GOOG
MSFT
ORCL
SSNLF
SAP
YHOO
reason for this may be that all technology companies come out with new products and announcements much more frequently than companies in other sectors(say infrastructure, energy, FMCG, etc.) thereby generating greater buzz
on social media networks. However, our model may be applied to any company/indices that generate high tweet volume.
Fig. 3 Graph for average of log of daily volume over the months under study
4.2 Sentiment Classification
In order to compute sentiment for any tweet we had to classify each incoming
tweet everyday into positive or negative using nave classifier. For each day total
number of positive tweets is aggregated as P ositiveday while total number
of negative tweets as N egativeday . We have made use of JSON API from
Twittersentiment 5 , a service provided by Stanford NLP research group [16].
Online classifier has made use of Naive Bayesian classification method, which is
one of the successful and highly researched algorithms for classification giving
5
https://sites.google.com/site/twittersentimenthelp/
8
Tushar Rao, Saket Srivastava
superior performance to other methods in context of tweets. Their classification
training was done over a dataset of 1,600,000 tweets and achieved an accuracy
of about 82.7%. These methods have high replicability and few arbitrary fine
tuning elements.
In our dataset roughly 61.68% of the tweets are positive, while 38.32% of
the tweets are negative for the company stocks under study. The ratio of 3:2
indicates stock discussions to be much more balanced in terms of bullishness
than internet board messages where the ratio of positive to negative ranges
from 7:1 [11] to 5:1 [13]. Balanced distribution of stock discussion provides us
with more confidence to study information content of the positive and negative
dimensions of discussion about the stock prices on microblogs.
4.3 Tweet Feature Extraction
One of the research questions this study explores is how investment decisions for technological stocks are affected by entropy of information spread
about companies under study in the virtual space. Tweet messages are microeconomic factors that affect stock prices which is quite different type of relationship than factors like news aggregates from traditional media, chatboard
room etc. which are covered in earlier studies over a particular period [11],
[19]and [13]. Keeping this in mind we have only aggregated the tweet parameters (extracted from tweet features) over a day. In order to calculate parameters
weekly, bi-weekly, tri-weekly, monthly, 5 weekly and 6 weekly we have simply
taken average of daily twitter feeds over the requisite period of time.
Fig. 4 Tweet Sentiment and Market Features
Twitter literature in perspective of stock investment is summarized in Figure 4. We have carried forward work of Antweiler et.al. for defining bullishness
(Bt ) for each day (or time window) given equation as:
!
1 + Mt P ositive
Bt = ln
(1)
1 + Mt N egative
Where Mt P ositive and Mt N egative represent number of positive or negative
tweets on a particular day t. Logarithm of bullishness measures the share
of surplus positive signals and also gives more weight to larger number of
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
9
messages in a specific sentiment (positive or negative). Message volume for a
time interval t is simply defined as natural logarithm of total number of tweets
for a specific stock/index which is ln(Mt P ositive + Mt N egative ). The agreement
among positive and negative tweet messages is given by:
s
M P ositive − MtN egative
(2)
At = 1 − 1 − t
MtP ositive + MtN egative
If all tweet messages about a particular company are bullish or bearish,
agreement would be 1 in that case. Influence of silent tweets days in our study
(trading days when no tweeting happens about particular company) is less
than 0.1% which is significantly less than previous research [13, 28]. Carried
terminologies for all the tweet features{Positive, Negative, Bullishness, Message Volume, Agreement} remain same for each day with the lag of one day.
For example, carried bullishness for day d is given by CarriedBullishnessd−1 .
4.4 Financial Data Collection
We have downloaded financial stock prices at daily intervals from Yahoo Finance API6 for DJIA, NASDAQ-100 and the companies under study given in
Table 1. The financial features (parameters) under study are opening (Ot ) and
closing (Ct ) value of the stock/index, highest (Ht ) and lowest (Lt ) value of the
stock/index and returns. Returns are calculated as the difference of logarithm
to the base e between the closing values of the stock price of a particular day
and the previous day.
Rt = {ln Close(t) − ln Close(t−1) } × 100
(3)
Trading volume is the logarithm of number of traded shares. We estimate
daily volatility based on intra-day highs and lows using Garman and Klass
volatility measures [14] given by the formula:
r
Ht 2
Ct 2
1X1
[ln
] − [2 ln 2 − 1][ln
]
(4)
σ=
n
2
Lt
Ot
5 STATISTICAL ANALYSIS AND RESULTS
We begin our study by identifying the correlation between the Twitter feed
features and stock/index parameters which give the encouraging values of
statistically significant relationships with respect to individual stocks(indices).
To validate the causative effect of tweet feeds on stock movements we have used
econometric technique of Granger’s Casuality Analysis. Furthermore, we make
use of expert model mining system (EMMS) to propose an efficient prediction
6
http://finance.yahoo.com/
10
Tushar Rao, Saket Srivastava
model for closing price of DJIA and NASDAQ 100. Since this model does not
allow us to draw conclusion about the accuracy of prediction (which will differ
across size of the time window) subsequently discussed later in this section.
5.1 Correlation Matrix
For the stock indices DJIA and NASDAQ and 11 tech companies under study
we have come up with the correlation matrix given in Figure 10 in the appendix
between the financial market and Twitter sentiment features explained in section 4. Financial features for each stock/index (Open, Close, Return, Trade
Volume and Volatility) is correlated with Twitter features (Positive, Negative,
Bullishness, Carried Positive, Carried Negative and Carried Bullishness).The
time period under study is monthly average as it the most accurate time window that gives significant values as compared to other time windows which is
discussed later section 5.4.
Our approach shows strong correlation values between various features
(upto −0.96 for opening price of Oracle and 0.88 for returns from DJIA index etc.) and the average value of correlation between various features is
around 0.5. Comparatively highest correlation values from earlier work has
been around 0.41 [28]. As the relationships between the stock(index) parameters and Twitter features show different behavior in magnitude and sign
for different stocks(indices), a uniform standardized model would not applicable to all the stocks(indices). Therefore, building an individual model for
each stock(index) is the correct approach for finding appreciable insight into
the prediction techniques. Trading volume is mostly governed by agreement
values of tweet feeds as −0.7 for same day agreement and −0.65 for DJIA.
Returns are mostly correlated to same day bullishness by 0.61 and by lesser
magnitude 0.6 for the carried bullishness for DJIA. Volatility is again dependent on most of the Twitter features, as high as 0.77 for same day message
volume for NASDAQ-100.
One of the anomalies that we have observed is that EBay gives negative
correlation between the all the features due to heavy product based marketing
on Twitter which turns out as not a correct indicator of average growth returns
of the company itself.
5.2 Bivariate Granger Causality Analysis
The results in previous section show strong correlation between financial market parameters and Twitter sentiments. However, the results also raise a point
of discussion: Whether market movements affects Twitter sentiments or Twitter features causes changes in the markets? To verify this hypothesis we make
use of Granger Causality Analysis (GCA) to the time series averaged to weekly
time window to returns through DJIA and NASDAQ-100 with the Twitter features (positive, negative, bullishness, message volume and agreement). GCA
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
11
Table 2 Granger’s Casuality Analysis of DJIA and NASDAQ for 7 week lag Twitter sentiments. (NSDQ is short for NASDAQ)
Index
⇓
Lag PositiveNegative Bull
1
2
3
DJIA 4
5
6
7
1
2
3
NSDQ 4
5
6
7
0.614
0.033**
0.219
0.353
0.732
0.825
0.759
0.106
0.048**
0.06*
0.104
0.413
0.587
0.119
0.122
0.307
0.909
0.551
0.066
0.705
0.581
0.12
0.219
0.685
0.545
0.997
0.321
0.645
0.891
0.037**
0.718
0.657
0.651
0.928
0.809
0.044**
0.893
0.367
0.572
0.645
0.421
0.089
Msg
AgrmntVol
Carr
Positive
Carr
Carr Carr
Carr
Nega- Bull
Agrmnt Msg
tive
Vol
0.316
0.094*
0.508
0.743
0.553
0.554
0.687
0.827
0.642
0.357
0.764
0.861
0.954
0.551
0.69
0.032**
0.016**
0.116
0.334
0.961
0.867
0.02**
0.001**
0.01**
0.194
0.157
0.795
0.382
0.103
0.301**
0.845
0.221
0.045**
0.432
0.631
0.04**
0.108
0.123
0.778
0.762
0.512
0.788
0.765
0.086**
0.237
0.743
0.562
0.732
0.807
0.064*
0.022**
0.135
0.092*
0.18
0.613
0.096
0.785
0.047**
0.635
0.357
0.394
0.764
0.987
0.043**
0.828
0.401
0.649
0.485
0.898
0.196
0.759
0.265
0.357
0.999
0.987
0.261
0.865
0.704
0.255
0.008**
0.464
0.945
0.834
0.648
0.934
0.045**
0.219
0.272
0.607
0.832
0.969
0.071*
0.001**
0.131
0.343
0.028
0.591
0.544
is not used to establish causality, but as an economist tool to investigate a
statistical pattern of lagged correlation. A similar observation that cloud precede rain is widely accepted; proving cloud may may contain something that
causes rain but itself may not be actual causative of the real event.
GCA rests on the assumption that if a variable X causes Y then changes
in X will be systematically occur before the changes in Y. We realize lagged
values of X shall bear significant correlation with Y. However correlation is not
necessarily behind causation. We have made use of GCA in similar fashion as
[5, 15] This is to test if one time series is significant in predicting another time
series. Let returns Rt be reflective of fast movements in the stock market. To
verify the change in returns with the change in Twitter features we compare
the variance given by following linear models in equation 5 and equation 6 n
X
Rt = α +
βi Dt−i + t
n
n
i=1
X
X
γi Xt−i + εt
βi Dt−i +
Rt = α +
i=1
(5)
(6)
i=1
Equation 5 uses only ’n’ lagged values of Rt , i.e. (Rt−1 , ..., Rt−n ) for
prediction, while Equation 6 uses the n lagged values of both Rt and the tweet
features time series given by Xt−1 , . . . , Xt−n . We have taken weekly time
window to validate the casuality performance, hence the lag values 7 will be
calculated over the weekly intervals 1, 2, ..., 7.
From the Table 2, we can reject the null hypothesis (Ho ) that the Twitter
features do not affect returns in the financial markets i.e. β1,2,....,n 6= 0 with a
high level of confidence (high p-values). However as we see the result applies
to only specific negative and positive tweets (** for p-value < 0.05 and * for
7 lag at k for any parameter M at x week is the value of the parameter prior to x
t
t−k
week. For example, value of returns for the month of April, at the lag of one month will be
returnapril−1 which will be returnmarch
12
Tushar Rao, Saket Srivastava
p-value < 0.1 which is 95% and 99% confidence interval respectively). Other
features like agreement and message volume do not have significant casual
relationship with the returns of a stock index (low p-values).
5.3 EMMS Model for Forecasting
We have used Expert Model Mining System (EMMS) which incorporates a
set of competing methods such as Exponential Smoothing (ES), Auto Regressive Integrated Moving Average (ARIMA) and seasonal ARIMA models.
These methods are widely used in financial modeling to predict the values of
stocks/bonds/commodities/etc [23, 6]. These methods are suitable for constant
level, additive trend or multiplicative trend and with either no seasonality, additive seasonality, or multiplicative seasonality.
In this work, selection criterion for the EMMS is coefficient of determination (R squared) which is square of the value of pearson-’r’ of fit values
(from the EMMS model) and actual observed values. Mean absolute percentage error (MAPE) and maximum absolute percentage error (MaxAPE) are
mean and maximum values of error (difference between fit value and observed
value in percentage). To show the performance of tweet features in prediction
model, we have applied the EMMS twice - first with tweets features as independent predictor events and second time without them. This provides us
with a quantitative comparison of improvement in the prediction using tweet
features.
ARIMA (p,d,q) are in theory and practice, the most general class of models for forecasting a time series data, which is subsequently stationarized by
series of transformation such as differencing or logging of the series Yi . For
a non-seasonal ARIMA (p,d,q) model- p is autoregressive term, d is number
of non-seasonal differences and q is the number of lagged forecast errors in
the predictive equation. A stationary time series ∆Y differences d times has
stochastic component
∆Yi ∼ N ormal(µi , σ 2 )
(7)
2
Where µi and σ are the mean and variance of normal distribution, respectively. The systematic component is modeled as:
µi = αi ∆Yi−1 + ..... + αp ∆Yi−p + θi εi−1
(8)
+..... + θi εi−q
Where, ∆Y the lag-p observations from the stationary time series with
associated parameter vector α and i the lagged errors of order q, with associated parameter vector. The expected value is the mean of simulations from
the stochastic component,
E(Yi ) = µi = αi ∆Yi−1 + ..... + αp ∆Yi−p + θi εi−1
+..... + θi εi−q
(9)
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
13
Table 3 EMMS model fit characteristics for DJIA and NASDAQ-100
Index
Dow-30
NASDAQ-100
Predictors
Yes
No
Yes
No
Model Fit statistics
Ljung-Box Q(18)
R-squared MaxAPE Direction Statistics DF Sig.
0.95
0.92
0.68
0.65
1.76
2.37
2.69
2.94
90.8
60
82.8
55.8
11.36
9.9
23.33
16.93
18
18
18
17
0.88
0.94
0.18
0.46
Seasonal ARIMA model is of form ARIMA (p ,d ,q) (P,D,Q) where P
specifies the seasonal autoregressive order, D is the seasonal differencing order
and Q is the moving average order. Another advantage of EMMS model is
that it automatically selects the most significant predictors among all others
that are available.
In the dataset we have time series for a total of approximately 60 weeks (422
days), out of which we use approximately 75% i.e. 45 weeks for the training
both the models with and without the predictors for the time period June 2nd
2010 to April 14th 2011. Further we verify the model performance as one step
ahead forecast over the testing period of 15 weeks from April 15th to 29th July
2011 which count for wide and robust range of market conditions. Forecasting
accuracy in the testing period is compared for both the models in each case
in terms of maximum absulute percentage error (MaxAPE), mean absolute
percentage error (MAPE) and the direction accuracy. MAPE is given by the
equation 10, where yˆi is the predicted value and yi is the actual value.
M AP E =
Σ n i | yiy−iyˆi |
n
× 100
(10)
While direction accuracy is measure of how accurately market or commodity up/ down movement is predicted by the model, which is technically defined
as logical values for (yi,t̂+1 − yi,t ) × (yi,t+1 − yi,t ) > 0 respectively.
As we can see in the Table 3, there is significant reduction in MaxAPE
for DJIA(2.37 to 1.76) and NASDAQ-100 (2.96 to 2.69) when EMMS model
is used with predictors as events which in our case our all the Tweet features
(positive, negative, bullishness, message volume and agreement). Using tweet
features as part of the prediction process in the EMMS model, gives more
robust approach than the traditional forecasting methods. There is significant
decrease in the value of MAPE for DJIA which is 0.8 in our case than 1.79
for earlier approaches [5]. As we can from the values of R-square, MAPE and
MaxAPE in Table 3 for both DJIA and NASDAQ 100, our proposed model
uses Twitter sentiment analysis for a superior performance over traditional
methods. Since EMMS is a customizable and scalable technique, our proposed
model is bound to perform well in a wide range of stocks and indices.
Figures 5 shows the EMMS model fit for weekly closing values for DJIA
and NASDAQ 100. In the figure fit are model fit values, observed are values
of actual index and UCL & LCL are upper and lower confidence limits of the
prediction model.
14
Tushar Rao, Saket Srivastava
Fig. 5 Plot of Fit values (from the EMMS model) and actual observed closing values for
DJIA and NASDAQ-100. Fit are the modeled values through EMMS, observed are the actual
values of the index and UCL & LCL are the upper and the lower confidence limit.
5.4 Prediction Accuracy using OLS Regression
Our results in the previous section showed that forecasting performance of
stocks/indices using Twitter sentiments varies for different time windows.
Hence it is important to quantitatively deduce a suitable time window that
will give us most accurate prediction. Figure 6 shows the plot of R-square
metric for OLS regression for returns from stock indexes NASDAQ-100 and
DJIA from tweet board features (like number of positive, negative, bullishness,
agreement and message volume) both for carried (at 1-day lag) and same week.
The R-square metric (explained in section 5.3) is calculated as prediction
performance indicator for different time windows from daily, weekly, bi-weekly
to 6 weekly time window. From the figure 6 it can be inferred as we increase the
time window the accuracy in prediction increases but only till a certain point
that is monthly in our case beyond which value of R-square starts decreasing
again. Thus, for monthly predictions we have highest accuracy in predicting
anomalies in the returns from the tweet board features.
In the next section we will discuss the practical implementation of how
short term hedging strategies can improve efficiency by modeling mass public
opinion and behavior for a particular company or stock index through mining
of tweet sentiments.
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
15
Fig. 6 Plot of R-square values over different time windows for DJIA and NASDAQ-100.
Higher values denote greater prediction accuracy.
6 HEDGING STRATEGY USING TWITTER SENTIMENT
ANALYSIS
Portfolio protection is very important practice that is weighted as much as
portfolio appreciation. Just like a normal user purchases insurance for its
house, car or any commodity, one can also buy insurance for the investment
that is made in the stock securities. This doesn’t prevent a negative event from
happening, but if it does happen and you’re properly hedged, the impact of the
event is reduced. In a diverse portfolio hedging against investment risk means
strategically using instruments in the market to offset the risk of any adverse
price movements. Technically, to hedge investor invests in two securities with
negative correlations, which again in itself is time varying dynamic statistics.
To explain how weekly forecast based on mass tweet sentiment features
can be potentially useful for a singular investor, we will take help of a simple
example.
Let us assume that the share for a company C1 is available for $X per
share and the cost of premium for a stock option of company C1 (with strike
price $X) is $Y.
A = total amount invested in shares of a company C1 which is number of
shares (let it be N) × $X
B= total amount invested in put option of company C1 (relevant blocksize
× $Y)
And always for an effective investment (N × $X) > ( Blocksize × $Y)
An investor shall choose the value of N as per as their risk appetitive i.e.
ratio of A:B = 2:1 (assumed in our example, will vary from from investor
to investor). Which means in the rising market conditions, he would like to
16
Tushar Rao, Saket Srivastava
keep 50% of his investment to be completely guarded, while rest 50% are risky
components; whereas in the bearish market condition he would like to keep
his complete investment fully hedged by buying put options equivalent of all
the investment he has made in shares for the same security. From Figure 7, we
infer for the P/L curves consisting of shares and 2 different put options for the
company C1 purchased as different time intervals 8 ; hence the different premium price even with the same strike price of $X. Using married put strategy
makes the investment risk free but reduces the rate of return in contrast to the
case which comprises of only equity security which is completely free-fall to
the market risk. Hence the success of married put strategy depends greatly on
the accuracy of predicting whether the markets will rise of fall. Our proposed
Tweet sentiment analysis can be highly effective in this prediction to determine accurate instances when the investor should readjust his portfolio before
the actual changes happen in the market. Our proposed approach provides
an innovative technique of using dynamic Twitter sentiment analysis toexploit
the collective wisdom of the crowd for minimising the risk in a hedged portfolio. Below we summarize two different portfolio states at different market
conditions.
Table 4 Example investment breakdown in the two cases
Partially Hedged Portfolio at 50% risk
1000 shares at price of $X = 1000X
1 Block size of 500 shares put options purchased at strike price of $X with premium of
$Y each = 500Y
Total= 1000X + 500Y
Fully Hedged Portfolio at minimized risk
1000 shares at price of $X = 1000X
2 Block size of 500 shares each put options purchased at strike price of $X with premium
of $Y each = 2×500Y = 1000Y
Total = 1000X + 1000Y
To check the effectiveness of our proposed tweet based hedging strategy, we
run simulations and make portfolio adjustments in various market conditions
(bullish, bearish, volatile etc). To elaborate, we take an example of DJIA
ETF’s as the underlying security over the time period of 14th November 2010
to 30th June 2011. Approximately 76% of the time period is taken in the
training phase to tune the SVM classifier (using tweet sentiment features from
the prior week). This trained SVM classifier is then used to predict market
direction (DJIA’s index movement) in the coming week. Testing phase for the
classification model (class 1- bullish market ↑ and class 0- bearish market ↓) is
from 8th May to 30th June 2011 consisting a total of 11 weeks. SVM model is
build using KSVM classification technique with the linear (vanilladot) kernel
8 The reason behind purchase of long put options at different time intervals is because
in a fully hedged portfolio, profit arrow has lower slope as compared to partially hedged
portfolio (refer P/L graph). Thus the trade off between risk and security has to be carefully
played keeping in mind the precise market conditions.
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
17
Fig. 7 Portfolio adjustment in cases of bearish (fully hedged) and bullish (partial hedged)
market scenarios. In both the figures, strike price is the price at which a option is purchased,
Break even point (BEP) is the instance when investment starts making profit. In case of
bearish market scenario, two options at same strike price (but different premiums) are in
purchased at different instances, Option1 brought at the time of initial investment and
Option2 brought at a later stage (hence lower in premium value).
using the package ’e1071’ in R statistical language. Over the training dataset,
the tuned value of the objective function is obtained as −4.24 and the number
of support vectors is 8. Confusion matrix for the predicted over the actual
values (in percentage) is given in Table 5. Overall classifier accuracy over
the testing phase is 90.91%. Receiver operator characteristics (ROC) curve
measuring the accuracy of the classifier as true positive rate to false positive
rate is given in the figure 8. It shows the tradeoff between sensitivity i.e. true
positive rate and specificity i.e. true negative rate (any increase in sensitivity
will be accompanied by a decrease in specificity). Good statistical significance
for the classification accuracy can be inferred from the value of area under the
ROC curve (AUC) which comes out to 0.88.
Table 5 Prediction accuracy over the Testing phase (11 weeks). Values in percentage.
Confusion Matrix
Actual Direction
Market Down
Market Up
Predicted Direction
Market Down
Market Up
45
0
9
45
Figure 9 shows the DJIA index during the testing period and the arrows
mark the weeks when the adjustment is done in the portfolio based on prediction obtained from tweet sentiment analysis of prior week. At the end of the
week (on Sunday), using tweet sentiment feature we predict what shall be the
market condition in the coming week- whether the prices will go down or up.
Based on the prediction portfolio adjustment - bearish −→ bullish or bullish
−→ bearish.
18
Tushar Rao, Saket Srivastava
Fig. 8 Receiver operating characteristic (ROC curve) curve for the KSVM classifier prediction over the testing phase. ROC is graphical plot of the sensitivity or true positive rate,
vs. false positive rate (one minus the specificity or true negative rate). More the area under
curve for typical ROC, more is the performance efficiency of the machine learning algorithm.
Fig. 9 DJIA index during the testing period. In the figure green marker shows adjustment
bearish −→ bullish, while red arrow shows adjustment bullish −→ bearish. (Data courtesy
Yahoo! finance)
7 DISCUSSIONS
In section 5, we observed how the statistical behavior of market through Twitter sentiment analysis provides dynamic window to the investor behavior.
Furthermore, in the section 6 we discussed how behavioral finance can be
exploited in portfolio decisions to make highly reduced risked investment. Our
work answers the important question - If someone is talking bad/good about
a company (say Apple etc.) as singular sentiment irrespective of the overall
market movement, is it going to adversely affect the stock price? Among the
5 observed Twitter message features both at same day and lagged intervals
we realize only some are Granger causative of the returns from DJIA and
NASDAQ-100 indexes, while changes in the public sentiment is well reflected
in the return series occurring at even lags of 1, 2 and 3 weeks. Remarkably the
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
19
most significant result is obtained for returns at lag 2 (which can be inferred
as possible direction for the stock/index movements in the next week).
Table 6 given below explains the different approaches to the problem that
have been done in past by researchers [28], [5] and [15]. As can be seen from
the table, our approach is scalable, customizable and verified over a large
data set and time period as compared to other approaches. Our results are
significantly better than the previous work. Furthermore, this model can be of
effective use in formulating short-term hedging strategies (using our proposed
Twitter based prediction model).
Table 6 Comparison of Various Approaches for Modeling Markets Movements Through
Twitter
Previous Approaches Bollen et al. [5] and Sprenger et al. [28]
→
Gilbert et al. [15]
Approach
Dataset
Techniques
Results
Feedback/
backs
This Work
Mood of complete Twit- Stock Discussion with Discussion based trackter feed
ticker $ on Twitter
ing of Twitter sentiments
28th Feb 2008 to 19th 1st Jan 2010 to 30th 2nd June 2010 to 29th
Dec 2008, 9M tweets June
20100.24M July 2011- 4M tweets
sampled as 1.5% of tweets
through search API
Twitter feed
SOFNN, Grangers and OLS Regression and Corr,
GCA,
Expert
linear models
Correlation
Model Mining System
(EMMS)
* High corr values
(upto -0.96) for opening price
* Strong corr values
(upto 0.88) for re* 86.7% directional ac- * Max corr value of 0.41 turns
curacy for DJIA
for returns of S&P 100 * MaxAPE of 1.76%
for DJIA
stocks
* Directional accuracy of 90.8% for
DJIA
Draw- Individual modeling for News not taken into ac- Comprehensive
and
stocks not feasible
count, very less tweet customizable approach.
volumes
Can be used for hedging
in F&O markets
8 CONCLUSION
In this paper, we have worked upon identifying relationships between Twitter based sentiment analysis of a particular company/index and its short-term
market performance using large scale collection of tweet data. Our results show
that negative and positive dimensions of public mood carry improved power to
track movements of individual stocks/indices. We have also investigated various other features like how previous week sentiment features control the next
week’s opening, closing value of stock indexes for various tech companies and
20
Tushar Rao, Saket Srivastava
major index like DJIA and NASDAQ-100. As compared to earlier approaches
in the area which have been limited to wholesome public mood and stock ticker
constricted discussions, we verify strong performance of our alternate model
that captures mass public sentiment towards a particular index or company
in scalable fashion and hence empower a singular investor to ideate coherent
relative comparisons. Our analysis of individual company stocks gave strong
correlation values (upto 0.88 for returns) with twitter sentiment features of
that company. Further we also discuss how Twitter sentiments bring wisdom
of the crowd to use by even a singular investor in the form of simplistic married put hedging strategy. Using this technique trader can retain his portfolio
with minimum risk even during highly bullish/bearish market conditions. It is
no surprise that this approach is far more robust and gives far better results
(upto 91% directional accuracy) than any previous work. In the near future,
Twitter sentiments analysis promises to be an effective strategy for hedging
the investments in the financial markets.
References
1. Daron Acemoglu, Asuman Ozdaglar, and Ali ParandehGheibi.
Spread of
(mis)information in social networks. Games and Economic Behavior, 70(2):194 – 227,
2010.
2. Sitaram Asur and Bernardo A Huberman. Predicting the future with social media.
Computing, 25(1):492499, 2010.
3. The Atlantic. Does anne hathaway news drive berkshire hathaway’s stock?, 2011. This
is an electronic document. Date of publication: [March 18 2011]. Date retrieved: October
12, 2011. Date last modified: [Date unavailable].
4. Mark Bagnoli, Messod D. Beneish, and Susan G. Watts. Whisper forecasts of quarterly
earnings per share. Journal of Accounting and Economics, 28(1):27 – 50, 1999.
5. Johan Bollen, Huina Mao, and Xiao-Jun Zeng. Twitter mood predicts the stock market.
Computer, 1010(3003v1):1–8, 2010.
6. George Edward Pelham Box and Gwilym Jenkins. Time Series Analysis, Forecasting
and Control. Holden-Day, Incorporated, 1990.
7. danah m. boyd and Nicole B. Ellison. Social network sites: Definition, history, and
scholarship. Journal of Computer-Mediated Communication, 13(1):210–230, 2007.
8. John Seely Brown and Paul Duguid. The Social Life of Information. Harvard Business
School Press, Boston, MA, USA, 2002.
9. Zhi Da, Joseph Engelberg, and Pengjie Gao. In search of attention. Russell The Journal
Of The Bertrand Russell Archives, (919), 2010.
10. Sanjiv R. Das and Mike Y. Chen. Yahoo! for Amazon: Sentiment Parsing from Small
Talk on the Web. SSRN eLibrary, 2001.
11. Michal Dewally. Internet investment advice: Investing with a rock of salt. Financial
Analysts Journal, 59(4):65–77, 2003.
12. S. Doan, B.-K. H. Vo, and N. Collier. An analysis of Twitter messages in the 2011
Tohoku Earthquake. ArXiv e-prints, September 2011.
13. Murray Z. Frank and Werner Antweiler. Is All That Talk Just Noise? The Information
Content of Internet Stock Message Boards. SSRN eLibrary, 2001.
14. Mark B. Garman and Michael J. Klass. On the estimation of security price volatilities
from historical data. The Journal of Business, 53(1):67–78, 1980.
15. Eric Gilbert and Karrie Karahalios. Widespread worry and the stock market. Artificial
Intelligence, pages 58–65, 2010.
16. Alec Go, Richa Bhayani, and Lei Huang. Twitter Sentiment Classification using Distant
Supervision.
Twitter Sentiment Analysis: How To Hedge Your Bets In The Stock Markets
21
17. Erkam Guresen, Gulgun Kayakutlu, and Tugrul U. Daim. Using artificial neural network
models in stock market index prediction. Expert Systems with Applications, 38(8):10389
– 10397, 2011.
18. K.H. Lee and G.S. Jo. Expert system for predicting stock market timing using a candlestick chart. Expert Systems with Applications, 16(4):357 – 364, 1999.
19. Alina Lerman. Individual Investors’ Attention to Accounting Information: Message
Board Discussions. SSRN eLibrary, 2011.
20. Huaxia Rui Liangfei Qiu and Andrew Whinston. A twitter-based prediction market:
Social network approach. ICIS 2011 Proceedings. Paper 5, 2011.
21. Burton G. Malkiel. The efficient market hypothesis and its critics. Journal of Economic
Perspectives, 17(1):59–82, 2003.
22. Huina Mao, Scott Counts, and Johan Bollen. Predicting financial markets: Comparing
survey,news, twitter and search engine data. Quantitative Finance Papers 1112.1051,
arXiv.org, December 2011.
23. Garth P. McCormick. Communications to the editorexponential forecasting: Some new
variations. Management Science, 15(5):311–320, 1969.
24. Douglas McIntyre. Turning wall street on its head, 2009. This is an electronic document. Date of publication: [May 29, 09]. Date retrieved: September 24, 2011. Date last
modified: [Date unavailable].
25. Hong Miao, Sanjay Ramchander, and J. K. Zumwalt. Information Driven Price Jumps
and Trading Strategy: Evidence from Stock Index Futures. SSRN eLibrary, 2011.
26. BBC News. Twitter predicts future of stocks, 2011. This is an electronic document.
Date of publication: [April 6, 2011]. Date retrieved: October 21, 2011. Date last modified:
[Date unavailable].
27. Bo Qian and Khaled Rasheed. Stock market prediction with multiple classifiers. Applied
Intelligence, 26:25–33, February 2007.
28. Timm O. Sprenger and Isabell M. Welpe. Tweets and Trades: The Information Content
of Stock Microblogs. SSRN eLibrary, 2010.
29. Martin Szomszor, Patty Kostkova, and Ed De Quincey. swineflu : Twitter predicts swine
flu outbreak in 2009. 3rd International ICST Conference on Electronic Healthcare for
the 21st Century eHealth2010, (December), 2009.
30. Andranik Tumasjan, Timm O Sprenger, Philipp G Sandner, and Isabell M Welpe. Predicting elections with twitter: What 140 characters reveal about political sentiment.
International AAAI Conference on Weblogs and Social Media Washington DC, pages
178–185, 2010.
31. Peter Wysocki. Cheap talk on the web: The determinants of postings on stock message
boards. Working Paper, 1998.
32. Max Zeledon. Stocktwits may change how you trade, 2009. This is an electronic document. Date of publication: [2009]. Date retrieved: September 01, 2011. Date last modified: [Date unavailable].
33. Xue Zhang, Hauke Fuehres, and Peter A Gloor. Predicting stock market indicators
through twitter i hope it is not as bad as i fear. Anxiety, pages 1–8, 2009.
22
Tushar Rao, Saket Srivastava
9 APPENDIX
Correlation heatmap indicative of significant relationships between various
twitter features with the index features.
Fig. 10 Heatmap showing pearson correlation coefficients between security indices vs features from Twitter and stock features.
| 5cs.CE
|
arXiv:1802.02258v1 [cs.CE] 6 Feb 2018
A computational framework for
microstructural modelling of polycrystalline
materials with damage and failure
Ph.D. thesis submitted to the University of Palermo
by
Vincenzo Gulizzi
Tutor
Prof. Alberto Milazzo
Prof. Ivano Benedetti
Dipartimento di Ing. Civile Ambientale, Aerospaziale, dei Materiali
Università degli Studi di Palermo
Scuola Politecnica
Viale delle Scienze, Ed. 8 - 90128 Palermo
V INCENZO G ULIZZI
Palermo, December 2016
e-mail:[email protected]
Thesis of the Ph.D. course in Civil and Environmental Engineering - Materials
(Ingegneria Civile e Ambientale - Materiali)
Dipartimento di Ingegneria Civile Ambientale, Aerospaziale, dei Materiali
Università degli Studi di Palermo
Scuola Politecnica
Viale delle Scienze, Ed.8 - 90128 Palermo, ITALY
Preface
In the present thesis, a computational framework for the analysis of the deformation and damage phenomena occurring at the scale of the constituent
grains of polycrystalline materials is presented. The research falls within the
area of Computational Micro-mechanics that has been attracting remarkable
technological interest due to the capability of explaining the link between the
micro-structural details of heterogenous materials and their macroscopic response, and the possibility of fine-tuning the macroscopic properties of engineered components through the manipulation of their micro-structure. However, despite the significant developments in the field of materials characterisation and the increasing availability of High Performance Computing facilities, explicit analyses of materials micro-structures are still hindered by their
enormous cost due to the variegate multi-physics mechanisms involved.
Micro-mechanics studies are commonly performed using the Finite Element Method (FEM) for its versatility and robustness. However, finite element formulations usually lead to an extremely high number of degrees of
freedom of the considered micro-structures, thus making alternative formulations of great engineering interest. Among the others, the Boundary Element
Method (BEM) represents a viable alternative to FEM approaches as it allows
to express the problem in terms of boundary values only, thus reducing the
total number of degrees of freedom.
The computational framework developed in this thesis is based on a nonlinear multi-domain BEM approach for generally anisotropic materials and is
devoted to the analysis of three-dimensional polycrystalline microstructures.
Different theoretical and numerical aspects of the polycrystalline problem using the boundary element method are investigated: first, being the formulation based on a integral representation of the governing equations, a novel
and more compact expression of the integration kernels capable of representing the multi-field behaviour of generally anisotropic materials is presented;
iii
iv
second, the sources of the high computational cost of polycrystalline analyses are identified and suitably treated by means of different strategies including an ad-hoc grain boundary meshing technique developed to tackle the
large statistical variability of polycrystalline micro-morphologies; third, nonlinear deformation and failure mechanisms that are typical of polycrystalline
materials such as inter-granular and trans-granular cracking and generally
anisotropic crystal plasticity are studied and the numerical results presented
throughout the thesis demonstrate the potential of the developed framework.
Contents
1
2
Grain Boundary formulation for polycrystalline micro-mechanics
1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Generation of artificial polycrystalline morphologies . . . . . .
1.3 Grain boundary integral equations . . . . . . . . . . . . . . . .
1.3.1 Boundary conditions . . . . . . . . . . . . . . . . . . . .
1.3.2 Interface conditions . . . . . . . . . . . . . . . . . . . .
1.3.3 Anisotropic Kernels . . . . . . . . . . . . . . . . . . . .
1.3.4 Numerical Integration . . . . . . . . . . . . . . . . . . .
1.4 Polycrystalline implementation . . . . . . . . . . . . . . . . . .
1.4.1 System solution . . . . . . . . . . . . . . . . . . . . . . .
1.5 Key features of the formulation . . . . . . . . . . . . . . . . . .
1.6 Content of the thesis . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Fourier transform and Rayleigh expansion . . . . . . . . . . . .
2.3.1 Fundamental solutions of general anisotropic materials .
2.3.2 Convergence . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3 Pseudo-Algorithm . . . . . . . . . . . . . . . . . . . . . .
2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4.1 Isotropic materials examples . . . . . . . . . . . . . . . .
2.4.2 Anisotropic materials examples . . . . . . . . . . . . . . .
2.5 Associated Legendre polynomials . . . . . . . . . . . . . . . . .
2.6 Spherical harmonics . . . . . . . . . . . . . . . . . . . . . . . . .
2.7 Materials properties . . . . . . . . . . . . . . . . . . . . . . . . . .
v
1
5
6
11
14
16
18
19
20
21
22
23
25
25
28
30
32
35
36
37
38
41
58
59
61
vi
3
CONTENTS
An enhanced grain-boundary framework
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . .
3.2 Laguerre tessellations regularisation . . . . . . . .
3.3 Removal of cut boundary grains . . . . . . . . . .
3.4 Grain-boundary meshing . . . . . . . . . . . . . .
3.4.1 Geometrical nodes and collocation nodes .
3.4.2 Continuous/semi-discontinuous elements
3.5 Computational tests . . . . . . . . . . . . . . . . . .
3.5.1 Overall computational savings . . . . . . .
3.5.2 Computational homogenisation . . . . . .
3.5.3 Interface modelling for micro-cracking . .
3.5.4 Microcracking simulations . . . . . . . . .
3.6 Discussion . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
65
. 65
. 68
. 68
. 72
. 73
. 78
. 82
. 82
. 85
. 89
. 91
. 107
4
Modelling inter-granular and trans-granular cracking in polycrystalline materials
109
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2 Grain boundary formulation . . . . . . . . . . . . . . . . . . . . . 113
4.2.1 Boundary integral equations . . . . . . . . . . . . . . . . 113
4.2.2 Boundary conditions and interface equations . . . . . . . 114
4.2.3 Inter- and trans-granular fracture modes . . . . . . . . . 115
4.3 Numerical discretisation and solution . . . . . . . . . . . . . . . 119
4.3.1 Algorithm for inter- and trans-granular cracking . . . . . 120
4.4 Numerical tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.5 Discussion and future developments . . . . . . . . . . . . . . . . 125
5
Grain boundary formulation for crystal plasticity
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Anisotropic elasto-plastic boundary element formulation . . . .
5.2.1 Boundary integral equations . . . . . . . . . . . . . . . .
5.2.2 Numerical discretisation . . . . . . . . . . . . . . . . . . .
5.2.3 Numerical integration and singular integrals . . . . . . .
5.3 Crystal plasticity phenomenological modelling . . . . . . . . . .
5.4 The grain-boundary incremental-iterative algorithm for crystal
plasticity in polycrystalline aggregates . . . . . . . . . . . . . . .
5.4.1 Numerical discretisation of polycrystalline aggregates .
5.4.2 Polycrystalline incremental-iterative system solution . .
5.5 Numerical tests . . . . . . . . . . . . . . . . . . . . . . . . . . . .
129
130
134
134
136
138
139
141
141
143
146
CONTENTS
5.6
5.7
5.5.1 Single crystal tests . . . . . . .
5.5.2 Polycrystalline tests . . . . . .
Discussion and possible developments
Flow rules and hardening laws . . . .
5.7.1 Bassani and Wu model . . . . .
5.7.2 Méric and Cailletaud model . .
Conclusions
vii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
147
152
163
164
164
165
167
viii
CONTENTS
Chapter 1
Grain Boundary formulation for
polycrystalline micro-mechanics
Many materials of technological interest, such as metals, alloys and ceramics,
present a polycrystalline micro-structure at a scale usually ranging from nanoto micro-meters (see Figure (1.1)). A polycrystalline micro-structure is characterised by the size, shape, random crystallographic orientation and general
anisotropic behaviour of each crystal, or grain, by the possible presence of randomly distributed flaws and pores and by the physical and chemical properties of the inter-crystalline interfaces, or grain boundaries, which play a crucial
in polycrystalline micro-mechanics.
µm
cm
Figure 1.1: Schematic representation of a polycrystalline micro-structure in an
engineered structural component.
1
2
Grain Boundary formulation for polycrystalline micro-mechanics
It is nowadays widely recognised that the macroscopic features of a material are strongly influenced by the properties of and the interaction between
its micro-constituents, especially when damage initiation and evolution are
involved. With current advancements in materials science and technology,
there is interest in understanding the link between material micro-structures
and macro-properties, i.e. the structure-property relation [85, 150, 246, 143, 151],
which also attracts remarkable technological interest due to the possibility of
controlling macroscopic properties, such as strength, stiffness, fracture toughness, etc. . . , through manipulations of the micro-structure [224].
In this context, experimental measurements and numerical simulations
represent two complementary approaches that have been employed in the
literature for understanding the micro-macro relation in polycrystalline materials. Experimental techniques based for example on serial sectioning [240,
78, 77, 179] or x-ray tomography [62, 126] and usually coupled to electron
backscattering diffraction (EBSD) measurements have been used to reconstruct polycrystalline micro-morphologies and to determine the crystals’ orientation. However, although being essential for the extraction of the statistical
features of the micro-morphologies, experimental approaches may be expensive and time-consuming and generally require sophisticated equipment and
careful material preparation.
On the other hand, computational approches allow to artificially reproduce the properties of polycrystalline materials and investigate their deformation and damage mechanisms by selectively focus on particular aspects that
could be extremely challenging to reproduce in a laboratory. Furthermore,
computational micro-mechanics is enormously benefitting from the increasing affordability of High Performance Computing (HPC), which is pushing
the boundaries of multi-field micro-mechanics, and from the aforementioned
developments in micro-structural materials characterisation, which provide
micro-structural details that can be included in high-fidelity computational
models, allowing deeper understanding of materials behaviour.
In the literature, several computational approaches have been proposed
for the modelling of two- and three-dimensional polycrystalline materials. An
important concept in the field of computational micro-mechanics of heterogeneous micro-structures is the concept of the Representative Volume Element
(RVE). An RVE is usually defined as material sample sufficiently large to statistically represent the heterogeneous nature of the micro-structure, but small
enough to be considered as a infinitesimal volume at the macroscopic scale.
The resolution of the micro-macro link is known as material homogenization
3
[152] and is usually performed by computing volume and/or ensemble averages of the fields of interests over multiple RVE realisations subjected to
general loading conditions and possibly undergoing internal evolution.
To this date, the most popular computational approach is probably still
represented by the Finite Element Method (FEM). Several FEM studies have
been proposed to evaluate the effective macroscopic properties of polycrystalline materials [111, 110] or to study the local behaviour at the grain boundaries [106, 243, 101]. The concept of the representative volume element has
been investigated by Ren and Zheng [171, 172] and by Nygårds [153], who
studied the dependence of the RVE size on the micro-structural properties of
two- and three-dimensional polycrystalline aggregates and on the anisotropy
of the crystals. Kanit et al. [102] focused on the concept of the RVE from statistical and numerical points of view, highlighting the possibility of computing
the macroscopic properties by means of a sufficiently large number of small
RVEs rather than by a large representative volume. The authors also studied
the effect of different boundary conditions on the convergence of macroscopic
properties of the RVEs. Fritzen et al. [70] focused on the generation and meshing of geometrically periodic polycrystalline RVEs.
Finite element models have been also employed to study different deformation and damage mechanisms in polycrystalline materials. Due to the
anisotropic nature of the constituent crystals, one of the most important as
well as studied deformation mechanisms in polycrystalline materials is the
mechanism of crystal plasticity [178], which denotes the plastic slip over specific crystallographic planes defined by the lattice of each crystal. The phenomenon, which has been widely studied within FEM frameworks, has been
also addressed in the present thesis and more details are given in Chapter (5).
Damage and fracture mechanisms have been usually addressed by means of
the cohesive zone approach [21, 22] in which fracture evolution is represented
by a traction-separation law that encloses the complex damage phenomena
occurring at a process zone of the failing interface into a phenomenological
model. Cohesive laws have been derived either by the definition of a cohesive potential [217, 234] or by simply assuming a functional form [44, 155] and
have been used to model different fracture mechanisms [45, 193].
In the field of computational micro-mechanics, the cohesive zone approach
has been employed by Zhai and Zhou [239, 238] to model static and dynamic
micro-mechanical failure of two-phase Al2 O3 /TiB2 ceramics. Espinosa and
Zavattieri [65, 66] studied inter-granular fracture of two-dimensional polycrystalline materials subject to static and dynamic loading. The competition
4
Grain Boundary formulation for polycrystalline micro-mechanics
between the bulk deformation of grains interior and the fracture at the grain
boundaries of three-dimensional columnar polycrystalline structures was investigated by Wei and Anand [225] by coupling a crystal plasticity and an
elasto-plastic grain boundary model. Maiti et al. [128] used the cohesive zone
approach to model the fragmentation of rapidly expanding polycrystalline
ceramics. The cohesive model at the micro-structure scale was also used by
Zhou et al. [245], who investigated the effect of grain size and grain boundary
strength on the crack pattern and the fracture toughness of a ceramic polycrystalline material.
The aforementioned works mainly focused on inter-granular fracture of
polycrystalline materials, i.e. the fracture of the grain boundaries. Another
failure mechanism occurring in polycrystalline materials is the fracture of the
bulk grains, known as trans-granular fracture. Similarly to the crystal plasticity phenomenon, trans-granular fracture is a highly anisotropic fracture phenomenon mainly controlled by the orientation of the crystals within the aggregate. The problem of trans-granular fracture has been investigated by fewer
authors [207, 145, 48] within the framework of the finite element method due
to its inherent anisotropy and generally higher complexity with respect to the
inter-granular fracture mechanism. Trans-granular fracture has been also discussed in the present thesis and further details on the problem are given in
Chapter (4).
Fracture in polycrystalline materials is also strongly affected by the operating conditions of the polycrystalline component. Naturally, ductile materials
are known to present a brittle fracture behaviour in presence of an aggressive
environment, and different studies [100, 145, 200, 201] focusing on the combined effect of the applied stress and the presence of corrosive species can be
found in the literature.
In addition to the finite element approach, the polycrystalline problem has
also been addressed within the framework of the Boundary Element Method
(BEM) [17, 228, 6]. Inter-granular micro-cracking has been developed for twodimensional [190, 191] and three-dimensional [29, 28, 80, 30, 79, 33] problems
and the formulation has been usually referred to as grain boundary formulation
for polycrystalline materials.
More recently, a number of different numerical approaches have been proposed to model deformation and fracture mechanisms in polycrystalline materials. The phase-field framework has been proved to be effective for the
modelling of different deformation and fracture mechanisms [9, 137, 206, 40].
Within the polycrystalline materials framework, the phase field approach has
1.1 Notation
5
been used to model the failure of both two- and three-dimensional elastic,
elasto-plastic and ferroelectric materials [1, 2, 48, 49, 192]. A peridynamic
model for the dynamic fracture of two-dimensional polycrystalline materials was proposed by Ghajari et al. [75]. The cellular automata approach [53,
197, 58] and the non-local particle method [46] have been also investigated for
modelling inter- and trans-granular failure of polycrystalline materials.
In this Chapter, the grain boundary formulation employed for modelling
the micro-mechanics of polycrystalline aggregates is described. It is already
worth mentioning here that the salient feature of the formulation is the expression of the polycrystalline micro-structural problem in terms of inter-granular
displacements and tractions, which are the primary unknowns of the problem. In the present context, the term “grain-boundary” refers to this aspect.
The formulation has been first proposed by Sfantos and Aliabadi [190, 191]
for two-dimensional polycrystalline materials and by Benedetti and Aliabadi
[29, 28, 30] for three-dimensional polycrystalline problems. The present Chapter is intended to introduce the problem of polycrystalline micro-mechanics,
to present the 3D grain-boundary formulation and to set the stage of the subsequent developments of the present thesis.
1.1
Notation
In this Section, the notation used in the present thesis is introduced. This
work deals with three-dimensional polycrystalline materials in the 3D space
denoted by R3 . Spatial coordinates in R3 are indicated using bold lower-case
letters, i.e. x denotes the set of the three components { x1 , x2 , x3 }.
Vectorial and tensorial quantities that depend on the spatial variable x are
indicated using the indicial notation; for instance, ui (x) and σij (x) are used
to denote the set of three components of the vector {u1 (x), u2 (x), u3 (x)} and
the set of the nine components of the tensor {σ11 (x), σ12 (x), . . . , σ23 (x), σ33 (x)},
respectively. The same convention also holds for higher-order tensors. Unless
otherwise stated, the range of the subscripts is always 1, 2, 3.
By means of the indicial notation, the product of two vectorial or tensorial
quantities with repeated subscripts implies the summation over the repeated
subscripts; as an example, using the implied summation convention, the distance r between two points x and y is written as
r=
q
( x1 − y1 )2 + ( x2 − y2 )2 + ( x3 − y3 )2 =
q
( xk − yk )( xk − yk ),
6
Grain Boundary formulation for polycrystalline micro-mechanics
where the summation over the subscript k is implied. On the other hand,
N × 1 vectors and N × M matrices usually stemming from a discretisation
procedure and containing a fixed number of entries are written in bold uppercase letters. For instance, a linear system of N equations is simply written
AX = B, being X the N × 1 vector containing the unknowns, A the N × N
coefficients matrix and B the N × 1 vector representing the right-hand side.
Eventually, superscripts are used to specify a particular dependence of a
generic quantity. No summation is employed for repeated superscripts and
their range is always specified. As an example, the surface of the generic grain
g is indicated as S g .
(a)
(b)
Figure 1.2: Example of a polycrystalline aggregate generated using a Voronoibased scheme: (a) 1000-grain polycrystalline aggregate inside a cubic box and
(b) morphology of the internal grains of the aggregate.
1.2
Generation of artificial polycrystalline morphologies
In this Section, the generation of the micro-morphologies for the study of the
deformation and failure of polycrystalline materials is described. The morphology of polycrystalline materials can be reconstructed by experimental observations and measurements or by computer models that are able to reproduce the statistical features of the considered aggregate. Three-dimensional
reconstructions of polycrystalline materials have been performed by using
several techniques relying on x-ray micro-tomography [62], diffraction con-
1.2 Generation of artificial polycrystalline morphologies
7
trast tomography [104] and serial sectioning [240, 179] combined with electron backscatter diffraction (EBSD) maps [78, 77]. However, such techniques,
although necessary to measure the statistical features of polycrystalline micromorphologies, require generally expensive laboratory equipment and involve
time-consuming data processing. On the other hand, artificial morphologies,
generated on the information provided by experimental measurements, offer
a valid alternative to experimentally reconstructed aggregates. In the literature, simplified geometrical models (such as cubes or octahedra) of the grains
have been employed [244, 170, 176]. Such regular grain shapes have been
shown to produce satisfactory results in terms of homogenised properties in
linear and crystal plasticity analyses. However, the consequent regular structure of the grain interfaces is not fully appropriate to model inter-granular
failure mechanisms, which may require a more accurate representation of the
grain boundaries. Besides regular morphologies, Voronoi tessellation algorithms [219] have been extensively used to artificially generate polycrystalline
aggregates. Voronoi tessellation are analytically defined given a distribution
of points (or seeds) scattered inside a domain. In general, the bounded threedimensional Voronoi tessellation can be defined given a bounded domain
D ⊂ R3 , a set of generator seeds S = {x g ∈ D, g = 1, . . . , Ng }, being Ng
the total number of seeds, and a distance function d defined in R3 . For each
seed x g , the corresponding grain occupying the volume V g is defined as
V g = {y ∈ D | d(y, x g ) < d(y, xh ) ∀h 6= g}
(1.1)
and the resulting bounded Voronoi diagram V (D, S ) is obtained as the union
of all grains
N
V (D, S ) =
[g
V g.
(1.2)
g =1
Approximations of three-dimensional polycrystalline aggregates can be obtained using 3D Voronoi tessellations and the Euclidean distance. Voronoibased polycrystalline morphologies and their modifications [112, 67, 116], described in more details in the next part of this Section, have been widely used
within different numerical approaches to model the behaviour of polycrystalline morphologies, see for examples Refs. [111, 110, 28, 73] to cite a few. As
an example, Figure (1.2a) shows a polycrystalline aggregate generated using a
Voronoi-based scheme inside a cubic box and Figure (1.2b) shows the interior
grains of the same aggregate.
8
Grain Boundary formulation for polycrystalline micro-mechanics
3.2
0.16
α-iron
Al-2%Sn
Ti-21S β
Voronoi Tessellation
2.8
2.0
0.12
Frequency
Frequency
2.4
1.6
1.2
0.10
0.08
0.06
0.8
0.04
0.4
0.02
0
α-iron
Al-2%Sn
Ti-21S β
Voronoi Tessellation
0.14
0
0
0.4
0.8
1.2
1.6
R/hRi
(a)
2.0
2.4
2.8
3.2
0
5
10
15
20
25
30
35
40
Number of Faces
(b)
Figure 1.3: Statistics of three polycrystalline materials compared to those obtained by using Voronoi tessellations. (a) Grain size obtained as the ratio between the spherical equivalent grain radius R (i.e. the radius of a sphere with
the same volume of the grain) and the mean radius h Ri. (b) Number of grain
faces. Experimental data are taken from [240] (circles), [62] (squares) and [179]
(diamonds).
(a)
(b)
Figure 1.4: 100-grain hardcore Laguerre-Voronoi tessellation: (a) seeds and
weights (indicated by the spheres) distribution inside a cubic box and (b) resulting tessellation.
Voronoi tessellations can be generated by using open source libraries such
as TetGen (http://wias-berlin.de/software/tetgen/) [198], Voro++ (http:
1.2 Generation of artificial polycrystalline morphologies
9
//math.lbl.gov/voro++/) [180] or Neper (http://neper.sourceforge.net)
[167]. Eventually, artificial polycrystalline morphologies can also be generated by relying on grain growth model in combination with phase field [108]
or cellular automata [60] techniques.
In this thesis, the polycrystalline microstructures are generated starting
from Voronoi tessellations. Figures (1.3a) and (1.3b) show two important statistical features of three polycrystalline materials (namely α-iron [240], Al2%Sn [62] and Ti-21Sβ [179]), and the corresponding statistical features measured for a 1000-grain Poisson-Voronoi tessellation. Figure (1.3a) shows that
the artificial morphology underestimates the variability of grain size, whereas
Figure (1.3b) shows that the Voronoi tessellation slightly overestimates the
number of faces per grain.
In order to reduce such differences between artificial Voronoi morphologies and real microstructures, modified Voronoi tessellations have been used
in the literature. A first constraint prescribes a minimum distance between
the tessellation seeds, to take into account the minimum size of stable nuclei
before the solidification process initiates, which leads to the so called hardcore
Voronoi tessellations. Second, a sphere with random radius from a suitable
statistical distribution can be associated to each seed, requiring that the associated grain must contain it, which leads to the so called Laguerre tessellations
[67, 116]. Calling r g the radius associated to the seed x g , the distance d L between x g and a generic point y used to compute the Laguerre tessellations is
defined as
d L (y, x g ) = {[d E (y, x g )]2 − (r g )2 }1/2
(1.3)
where d E is the Euclidean distance. Figure (1.4) shows the Laguerre tessellation obtained combining the two modifications. The weight (radius) of each
seed is indicated by a sphere, and the hardcore condition ensures that the
spheres do not intersect. Fine-tuning the distribution of grains weights allows
to generate micro morphologies that better represent the statistical features
of real materials. Figure (1.5) compares the statistical features of a 1000-grain
Poisson-Voronoi tessellation with those of a 1000-grain hardcore Laguerre tessellation, whose sphere radii distribution was determined from the following
Weibull distribution:
b x b −1
x b
ψ( x | a, b) =
exp −
(1.4)
a a
a
where the parameters a and b have been chosen to fit the experimental data
reported in Ref. [240]. The latter clearly better approximates the observed real
10
Grain Boundary formulation for polycrystalline micro-mechanics
microstructure, in terms of both grain size and number of faces. For such
reasons, the hardcore Laguerre tessellations are adopted in the present framework.
3.2
0.16
α-iron (exp.)
Laguerre-Voronoi
Poisson-Voronoi
2.8
0.12
2.0
Frequency
Frequency
2.4
α-iron
Laguerre-Voronoi
Poisson-Voronoi
0.14
1.6
1.2
0.10
0.08
0.06
0.8
0.04
0.4
0.02
0
0
0
0.4
0.8
1.2
1.6
R/hRi
(a)
2.0
2.4
2.8
3.2
0
5
10
15
20
25
30
35
40
Number of Faces
(b)
Figure 1.5: (a) Grain size distribution and (b) number of faces statistics generated from standard Poisson-Voronoi tessellations (dashed line), experimental measurements [240] (circles) and hardcore Laguerre-Voronoi tessellations
based on experimental data (squares).
Figure 1.6: Position of the generic grain g inside the polycrystalline aggregate.
V g and S g = ∂V g denote the volume occupied by the grain and its boundary,
respectively. A local reference system Rki (x) is attached at each point x of the
boundary of the grain to facilitate the expression of the interface equations.
1.3 Grain boundary integral equations
1.3
11
Grain boundary integral equations
The polycrystalline morphology is modelled using a multi-domain boundary integral formulation in which each grain g is considered as a generally
anisotropic domain with a specific crystallographic orientation in the threedimensional space.
Let us first consider a generic three-dimensional grain g in a polycrystalline aggregate occupying the volume V g ⊂ R3 and whose boundary is
denoted by S g = ∂V g (see Figure (1.6)). In the linear elastic regime, the relag
tion between the second-order stress tensor σij (x) and the second-order strain
g
tensor ε ij (x) can be generally expressed as
g
g
g
σij (x) = cijkl ε kl (x)
(1.5)
g
where cijkl represents the components of the grain’s fourth-order stiffness tensor, which is assumed to be constant for the considered grain. It is recalled
here that, in Eq.(1.5) and throughout the present thesis, repeated subscripts
imply summation. In the formulation, the displacements field of the grain is
g
g
denoted by ui (x) and is related to the strain tensor ε ij (x) whose components
are given by
#
" g
g
∂u j
1 ∂ui
g
ε ij (x) =
(x) +
(x) ;
(1.6)
2 ∂x j
∂xi
g
the tractions field defined on a surface with normal ni (x) is denoted by ti (x)
and its components are given by
g
g
g
g
ti (x) = σij (x)n j (x) = n j (x)cijkl
∂uk
( x ).
∂xl
(1.7)
The boundary integral equations for the grain are derived by using the
reciprocity relation between the elastic solution of the grain subject to prescribed boundary conditions and the elastic solution in the infinite 3D elastic
space of a body force concentrated at the point y ∈ R3 .
For any internal point y inside the domain V g , the displacement boundary
integral equations (DBIE) in absence of body forces are written as follows [17,
228, 6]
g
u p (y) +
Z
g
Sg
g
Tpi (x, y)ui (x)dS(x) =
Z
g
Sg
g
U pi (x, y)ti (x)dS(x)
(1.8)
12
Grain Boundary formulation for polycrystalline micro-mechanics
where x is the integration point running over the boundary S g of the grain g.
y is also denoted as the collocation point and represents the point at which the
g
boundary integral equations are evaluated. The integration kernels U pi (x, y)
g
and Tpi (x, y) depend on the elastic constants of the considered grain and their
expressions are explicitly given in Section (1.3.3) in terms of the fundamental
solutions of the general anisotropic linear elastic problem.
The explicit expression of the fundamental solutions is given in Chapter
(2). However, it is worth noting here that the kernels in the boundary integral equations considered in this thesis can always bepwritten as the product
of a singular function depending on the distance r = ( xk − yk )( xk − yk ) between the collocation point y and the integration point x, and a regular function depending on the direction r̂ = (x − y)/r and the material properties, i.e.
each kernel K (x, y) can generally be written as K (x, y) = r − β K (r̂) being β > 0
an integer representing the order of the singular function.
g
For three-dimensional problems, the kernel U pi (x, y) has an order of sing
gularity β = 1, whereas the kernel Tpi (x, y) has an order of singularity β = 2.
The order of singularity is of crucial importance when the collocation point is
taken at the boundary S g [17, 228, 6].
In fact, in order to solve the elastic problem given a proper set of displacements and/or tractions boundary conditions over S g , the DBIE (1.8) must be
evaluated for points belonging to the boundary S g of the volume V g . However, for a point y ∈ S g , given the singular nature of the integration kernels,
the integrals appearing in Eq.(1.8) must be treated appropriately.
The DBIE for a boundary point are obtained by suitable limiting processes
that are well described in the BEM literature [17, 228, 6]. In order to give the
intuition of a boundary limiting process when integrating over the surface
S g , one can consider a reference system centered at y. In such a case, the beg
g
haviour of the terms dS(x), U pi (x, y) and Tpi (x, y), as x → y, are of the order
r2 , r −1 and r −2 , respectively, being r the distance between x and y. As a cong
sequence, the integral involving the U pi (x, y) is only weakly singular and the
singularity can be cancelled by the Jacobian of the integration. On the other
g
hand, the integral involving the kernel Tpi (x, y) is strongly singular and the
so-called free terms remain after the limiting process. For a detailed derivation
of the DBIE for a point belonging to the boundary, the reader is referred to
[17, 228, 6].
Following considering the limiting process, a more general form of the
1.3 Grain boundary integral equations
13
DBIE can then be written as
g
c pi (y)ui (y) +
Z
−
Sg
g
g
Tpi (x, y)ui (x)dS(x)
=
Z
g
Sg
g
U pi (x, y)ti (x)dS(x)
(1.9)
R
where the integration symbol − denotes a Cauchy principal value integral
g
and the terms c pi (y)ui (y) are the free terms that stem from the limiting process. Depending on the position of the point y, the coefficients c pi (y) take the
following values
δpi ,
c pi (y) = δpi /2,
0,
if
if
if
y inside V g
y on S g smooth ,
y outside V g
(1.10)
where δpi is the Kronecker delta function.
The strain and the stress fields at any internal point y inside the grain g
can be computed by suitably taking the derivatives of Eq.(1.8) with respect
to the coordinates yi of the collocation point y. The strain boundary integral
equations are then obtained by using the strain-displacement relationships and
are given as follows
g
ε pq (y) +
Z
ε,g
Sg
g
Tpqi (x, y)ui (x)dS(x) =
ε,g
Z
ε,g
Sg
g
U pqi (x, y)ti (x)dS(x),
(1.11)
ε,g
where the kernels U pqi (x, y) and Tpqi (x, y) are obtained in terms of the kernels
g
g
U pi (x, y) and Tpi (x, y), respectively, and their expressions are given in Section
(1.3.3). By using the grain’s constitutive relations, the stress boundary integral
equations are expressed as
g
σmn (y) +
Z
σ,g
Sg
g
Tmni (x, y)ui (x)dS(x) =
σ,g
σ,g
Z
σ,g
Sg
g
Umni (x, y)ti (x)dS(x),
(1.12)
where the kernels Umni (x, y) and Tmni (x, y) are obtained by applying the conε,g
ε,g
stitutive relation to the kernels U pqi (x, y) and Tpqi (x, y), respectively. Eqs.(1.11)
and (1.12) provide a relation to compute the strain and the stress fields, respectively, within the volume V g of the grain in terms of the values of the
displacements and the tractions at the boundary S g and are usually employed
in a post processing stage after the solution of the elastic problem is obtained
by means of Eq.(1.9).
14
Grain Boundary formulation for polycrystalline micro-mechanics
The DBIE given in Eq.(1.9) represent the starting point of the grain boundary formulation of polycrystalline mechanics. In fact, the DBIE must be written for each crystal of the aggregate and suitable boundary conditions and
interface conditions must be enforced to the grain boundaries touching the
external domain and to the internal grain boundaries, respectively. To simplify the expression of the interface boundary conditions, the displacements
and the tractions at the internal grain boundaries are usually expressed in
terms of their tangential and normal components. It is therefore convenient to
define a local reference system over each face (see Figure (1.6)) of the boundaries of the constituent grains and express displacements and tractions in that
reference system. In the local reference system, the displacements and the
tractions are written in terms of the local transformation matrix Rij (x), which
depends on the position of the considered point on the boundary. The DBIE
can then be rewritten in a more convenient form as
Z
g
e g (x, y)ueg (x)dS(x) =
cepi (y)uei (y) + − T
pi
i
Sg
Z
Sg
e g (x, y)et g (x)dS(x),
U
pi
i
(1.13)
where
cepi (y) = c pk (y) Rki (y),
e g (x, y) = U g (x, y) Rki (x),
U
pi
pk
(1.14)
e g (x, y) = T g (x, y) Rki (x)
T
pi
pk
(1.15)
and Rij (x) relate the components of the grain boundary displacements and
tractions in the global reference system to the components in the local reference system, i.e.
g
g
uk (x) = Rki (x)uei (x),
g
g
tk (x) = Rki (x)eti (x).
(1.16)
In Eqs.(1.13) to (1.16) and throughout the present thesis, the symbol e· denotes a quantity expressed in the local reference system of the grain boundary.
1.3.1
Boundary conditions
Boundary conditions (BCs) generally refer to prescribed values of displacements or tractions that are enforced at the boundary of each considered domain. However, in case of the polycrystalline problem external grain boundary and internal grain boundaries must be distinguished. Therefore, throughout the present thesis, boundary conditions will refer to prescribed values
of displacements and tractions enforced over the external boundaries of the
grains, that is those boundaries of the grains that touch the boundary of the
1.3 Grain boundary integral equations
15
Figure 1.7: Example of a grain for which the external and internal grain
boundaries are distinguished. Boundary conditions are enforced on the external boundaries (in red) whereas interface conditions are enforced on internal
boundaries (in green).
external domain as shown in Figure (1.7), in which the internal grain boundaries of a representative boundary grain are coloured in green and the external
grain boundary is coloured in red.
Different sets of boundary conditions can be assigned to polycrystalline
aggregates based on the considered applications. Kinematic or static BCs
are usually employed to model the response of a polycrystalline structure to
a particular loading condition. On the other hand, to extract homogenised
properties of polycrystalline RVEs, periodic boundary conditions (PBCs) have
been shown to provide faster convergence to the effective properties with respect to displacement or traction BCs [214]. Considering the external faces of
the aggregate, and in particular the pair of opposite master point xm and slave
point xs , PBCs are enforced as follows
ui (xs ) = ui (xm ) + Γ̄ij ( x sj − x m
j ),
(1.17a)
ti (xs ) + ti (xm ) = 0,
(1.17b)
where Γ̄ij represents the prescribed strain tensor. Even though it does not
represent a requirement, the use of periodic boundary conditions is facilitated
by the periodic structure of the aggregate and therefore by the conformity
between the meshes of coupled opposite faces. Figure (1.8) shows a periodic
polycrystalline morphology, whose generation is detailed is Section (3.3) of
16
Grain Boundary formulation for polycrystalline micro-mechanics
Figure 1.8: Example of a periodic polycrystalline morphology and detail of
the pair of opposite master point xm and slave point xs for the enforcement of
periodic boundary conditions.
Chapter (3), and the position of the master point xm and the slave point xs
on opposite faces of the aggregate for which the PBCs given in Eq.(1.17) are
enforced.
1.3.2
Interface conditions
The interface conditions (ICs) express the relationships between the displacements and the traction fields at the interface between two adjacent grains.
Consider the two grains in Figure (1.9), where the two adjacent grains g and
h share the interface I gh = V g ∩ V h that is highlighted by darker green. The
interface conditions can be generally expressed, ∀x ∈ I gh , as follows
gh
gh g
g
Ψi [uej (x), uehj (x), et j (x), ethj (x)] = 0,
(1.18a)
gh g
Θi [et j (x), ethj (x)] = 0,
(1.18b)
gh
where Θi represents grain boundary equilibrium conditions whereas Ψi
represents generic relations between the grain boundary tractions and displacements that can include from pristine to completely failed interfaces based
on the state of the grain boundaries. More details on the interface conditions
used for the polycrystalline problems addressed in this thesis are given based
on the specific considered application.
1.3 Grain boundary integral equations
17
As previously mentioned, it is more convenient to express the interface
conditions in terms of the normal and tangential components of the grain
boundary displacements and tractions and this is the reason why the functional dependence of the ICs in Eq.(1.18) is given in terms of the local fields
uei and eti . Moreover, in case of a damaged or failing interface, Eq.(1.18a) is
usually given in terms of the displacements jump at the interface. By defining
opposite local reference systems at coupled points of the interface as shown
in Figure (1.9), i.e. Rki (x ∈ V g ) = − Rki (x ∈ V h ), the local displacements jump
gh
δuei can be written as
gh
g
δuei = −(uei + ueih ).
(1.19)
Similarly, the equilibrium conditions considering opposite local reference systems can be written as
et g = etih = et gh .
i
i
(1.20)
gh
where eti is used to denote the values of the tractions at the interface I gh .
Figure 1.9: Schematic representation of two opposite reference systems at the
interface I gh between the adjacent grains g and h.
18
1.3.3
Grain Boundary formulation for polycrystalline micro-mechanics
Anisotropic Kernels
The kernels appearing in the boundary integral equations (1.8), (1.9), (1.11)
and (1.12) are computed in terms of the fundamental solutions of the generally anisotropic elastic problem. More specifically, the fundamental solutions
Φ pi (x, y) of general anisotropic elasticity are obtained as solutions of the following problem
g
∂2 Φ pk
g
cijkl
(x, y) + δpi δ(x − y) = 0,
(1.21)
∂x j ∂xl
g
where δ(x − y) is the Dirac delta function and the superscript g in Φ pk and
g
in the stiffness tensor cijkl highlights the fact that the fundamental solutions
depend on the specific considered material. The explicit solution of Eq.(1.21)
is discussed in a more general fashion in Chapter (2) wherein a novel and
unified compact expression of the fundamental solutions of a broader class of
partial differential operators, i.e. second-order homogeneous elliptic partial
differential operators, is obtained in terms of spherical harmonics expansions.
Once the fundamental solutions Φ pi (x, y) and their derivatives are computed,
the kernels in Eqs. (1.8), (1.9) are obtained as follows
g
g
U pi (x, y) = Φ pi (x, y),
(1.22a)
g
g
Σ pij (x, y)
=
g
cijkl
∂Φ pk
(x, y),
∂xl
g
g
Tpi (x, y) = Σ pij (x, y)n j (x).
(1.22b)
(1.22c)
The kernels appearing in Eq.(1.11) are computed by taking the derivatives
of the kernels in Eq.(1.22) with respect to the collocation point coordinates yi
and are written as
" g
#
g
∂Φqi
1 ∂Φ pi
g,ε
U pqi (x, y) =
(x, y) +
(x, y) ,
(1.23a)
2 ∂yq
∂y p
" g
#
g
∂Φqk
1 g ∂ ∂Φ pk
g,ε
Σ pqij (x, y) = cijkl
(x, y) +
(x, y) ,
(1.23b)
2
∂xl ∂yq
∂y p
g,ε
g,ε
Tpqi (x, y) = Σ pqij (x, y)n j (x).
(1.23c)
Finally, the kernels appearing in the stress boundary integral equations
(1.12) are obtained upon applying the constitutive relations to the kernels
1.3 Grain boundary integral equations
19
given in Eq.(1.23) and their expressions are
g
g,σ
Umni (x, y)
=
g
cmnpq
∂Φ pi
∂yq
(x, y),
(1.24a)
g
g,σ
Σmnij (x, y)
=
g,σ
g
g
cmnpq cijkl
∂2 Φ pk
∂xl ∂yq
(x, y),
(1.24b)
g,σ
Tmni (x, y) = Σmnij (x, y)n j (x).
(1.24c)
g
g,ε
In Eqs.(1.22b), (1.23b) and (1.24b) the kernels Σ pij (x, y), Σ pqij (x, y) and
g,σ
Σmnij (x, y), respectively, have been introduced as their expressions will be
used in Chapter (5).
1.3.4
Numerical Integration
To solve numerically the polycrystalline problem, the boundary integral equations (1.9) are used in the context of the boundary element method for each
grain g of the aggregate, according to the following steps:
• The boundary S g of each grain is subdivided into sets of non-overlapping
surface elements following the strategy described in Chapter (3), wherein
an enhanced meshing technique for polycrystalline aggregates is developed for computational effectiveness;
• The boundary displacements and tractions fields are then expressed in
terms of suitable shape functions N(ξ ), defined over each element in a
local 2D (surface) coordinate system ξ = {ξ 1 , ξ 2 }, and nodal values of
boundary displacements and tractions expressed in a face local reference system. After this discretisation procedure, a set of nodal points,
or simply nodes, and boundary elements are identifiable for each grain;
• The DBIE (1.9) are then written for each node of the considered grain
and it is numerically integrated, considering the explicit approximation of
the boundary fields in terms of shape functions and nodal values.
g
In this way, a set of 3Np equations, where Np is the number of nodes used in
the discretisation of the grain g, is written in terms of nodal values of displacements and tractions and the discrete version of the displacement boundary
integral equations reads
Hg Ug = Gg Tg
(1.25)
20
Grain Boundary formulation for polycrystalline micro-mechanics
g
where the vectors U g and T g are 3Np vectors collecting the nodal values of
boundary displacements and tractions, respectively, and the matrices H g and
g
g
eg
G g are 3Np × 3Np matrices stemming from the integration of the kernels T
ij
e g , respectively. In the numerical integration procedure, care must be
and U
ij
taken when integrating over the elements that contain, for a given collocation
point, the collocation point itself. Such singular elements are treated, in the
present work, using element subdivision and coordinate transformation for
e g , and rigid body conditions for the strongly singuthe weakly singular kernels U
ij
e g . The interested readers are referred to [17, 6] for such boundary
lar kernels T
ij
element specific aspects.
Enforcing displacement and traction boundary conditions on the faces of
the grains lying on the loaded external boundary of the aggregate, Eq.(1.25)
leads, for each grain, to the following system of equations
A g X g = C g Y g (λ)
(1.26)
where X g collects the unknown nodal values of grain boundary displacements
and tractions, Y g collects prescribed values of boundary displacements and
tractions, and A g and C g contain suitable combination of the columns of the
matrices H g and G g . In Eq.(1.26), the prescribed values of boundary displacements and tractions are given as a function of the load factor λ governing the
loading history.
It is clear that, at the internal grain boundaries, neither the displacements
nor the tractions are known and therefore the matrix A g is not a square matrix; that is, the system of equations (1.26) for a single grain has more unknowns than equations. To solve the polycrystalline problem, Eq.(1.26) must
be written for each grain and coupled to suitable interface conditions. The
implementation of the polycrystalline system of equations is described next.
1.4
Polycrystalline implementation
To obtain the system of equations of the entire polycrystalline aggregates, the
system of equations given in (1.26) is written for each grain of the aggregate,
that is for g = 1, . . . , Ng ; the integrity of the microstructure is retrieved in
the model by enforcing suitable inter-granular conditions, which leads to the
following system
AX − B(λ)
=0
(1.27)
M(X, λ) =
I(X)
1.4 Polycrystalline implementation
21
where
A=
A1 0 · · ·
0 A2 · · ·
..
..
..
.
.
.
0
0 ···
0
0
..
.
A Ng
, B =
C1 Y1
C2 Y2
..
.
C Ng Y Ng
, X =
X1
X2
..
.
X Ng
,
(1.28)
being X the vector collecting the unknown degrees of freedom of the polycrystalline system, i.e. the unknown displacements and tractions at the grain
boundaries. In Eq.(1.27), I(X) implements the interfaces conditions, which
are written for each pair of coupled interface nodes according to Eq.(1.18)
and therefore are generally function of the unknown nodal displacements and
tractions X.
1.4.1
System solution
The system of equations given in (1.27) must be solved at each step of the
loading history. Calling λn the load factor at the n-th load step of the loading
history, the corresponding solution Xn is obtained by solving M(Xn , λn ) = 0.
The solution is computed by employing a Newton-Raphson iterative algorithm, i.e., at the n-th load step, the (k + 1)-th approximation Xkn+1 of the system solution is obtained as
Xkn+1 = Xkn − (Jkn )−1 Mkn ,
(1.29)
where Mkn ≡ M(Xkn , λn ), the starting guess solution is chosen as the last converged solution, i.e. X0n ≡ Xn−1 , and Jkn ≡ J(Xkn , λn ) is the Jacobian of the
system defined as
∂M
J(X, λ) =
(X, λ).
(1.30)
∂X
Given the structure of the overall system of equations (1.27), it is easy
to see that the Jacobian defined in Eq.(1.30) is a highly sparse matrix, especially for a large number of grains. As a consequence, the package PARDISO
(http://www.pardiso-project.org) [114, 185, 186] is used as a sparse solver
to compute the term (Jkn )−1 Mkn in Eq.(1.29). To enhance the convergence and
the whole solution calculation, the zero and non-zero elements of the Jacobian are defined at the beginning of the loading history in such a way that its
sparsity pattern is retained throughout the analysis. Furthermore, the solver
PARDISO is used both as a direct and an iterative solver. More specifically,
22
Grain Boundary formulation for polycrystalline micro-mechanics
by considering that during subsequent load steps the changes in the Jacobian
matrix are small, the same LU factorisation at a certain load step is used as
a preconditioner of the iterative solution at the successive steps, whereas the
new LU factorisation is computed only when the iterative approach does not
converge. Such an approach drastically decreases the computational cost of
the solution. The reader is referred to the PARDISO documentation for more
details.
A direction of further computational enhancement could involve the use
of specialised high efficiency solvers, based for example on the use of the hierarchical format in conjunction with iterative solution strategies [31, 34].
1.5
Key features of the formulation
To summarise, the salient features of the method are:
• The polycrystalline microstructural problem is entirely formulated in
terms of boundary integral equations, in which the grain-boundary displacements and tractions are primary unknowns; in the present context,
the term “grain-boundary” refers to this specific aspect;
• The grain-boundary nature of the formulation induces a natural simplification of the pre-processing stage, as no volume meshing of the grains’
interior volume is required; in a context characterised by an inherent
statistical variability of the physical problem morphology, this is an important computational aspect, as it confines potential meshing issues to
the boundary of the grains;
• Since the problem is solved using only nodal points lying on the surface
of the grains, the formulation induces a reduction in the order of the
solving system, i.e. in the number of degrees of freedom; in a multiscale
perspective, this represents a computational benefit;
• The strain and stress fields at interior points, within the grains, can be
computed from grain-boundary variables using a stress boundary integral equation, in a post-processing stage;
• In the present form, non-linear constitutive behaviors within the grains
are not considered; this is not an inherent limitation of the formulation,
as such constitutive behaviors can be included introducing suitable volume terms (and volume meshes) in the boundary integral equations,
1.6 Content of the thesis
23
without introducing additional degrees of freedom, as it will be shown in
Chapter (5) when the problem of crystal plasticity is addressed.
1.6
Content of the thesis
The following Chapters of this thesis are organised as follows: Chapter (2)
presents a novel, unified and compact expression for the fundamental solutions of homogeneous second-order differential operators and their derivatives. An efficient evaluation of the fundamental solutions of partial differential equations is still of great engineering interest as they represent the key ingredient in the boundary integral equations of the boundary element method
approach employed in the present thesis, see Eqs.(1.8,1.9,1.11,1.12). The expression found in this thesis has the advantages of being general for the considered class of differential operators and simpler with respect to the formulas
found in the literature, especially when high-order derivatives of the fundamental solutions need to be computed.
In Chapter (3), an enhanced grain boundary framework for the study of
polycrystalline materials is presented. The framework is aimed at reducing
the high computational cost that hinders the development of three-dimensional
non-linear polycrystalline models using the boundary element approach. The
use of a regularised tessellation and the development of a novel ad-hoc meshing strategy to tackle the high statistical variability of polycrystalline micromorphologies are key points of the enhanced framework. The effects of the
adopted strategies on the number of degrees of freedom of polycrystalline systems and on the computational costs of micro-cracking analyses are described
by several studies.
In Chapter (4), a novel formulation for the competition between intergranular and trans-granular cracking in polycrystalline materials is developed. The two mechanisms are modelled using the cohesive zone approach.
However, unlike inter-granular cracks, which are well defined when the polycrystalline morphologies are generated, trans-granular cracks are introduced
on-the-fly during the micro-cracking analysis and the polycrystalline morphologies are remeshed accordingly.
In Chapter (5), a grain-boundary formulation for the mechanism of small
strain crystal plasticity is presented and implemented in polycrystalline materials for the first time. The boundary integral equations involving the presence of general anisotropic plastic strain in polycrystalline materials are extensively described. The iterative incremental algorithm for the solution of the
24
Grain Boundary formulation for polycrystalline micro-mechanics
crystal plasticity polycrystalline problems is discussed and numerical tests involving single crystals and polycrystalline materials are presented.
Chapter 2
Fundamental solutions for
multi-field materials based on
spherical harmonics expansions
2.1
Introduction
Fundamental solutions are essential to the solution of many boundary value
problems in engineering [143] and represent the key ingredient in boundary
integral formulations [17, 228, 6]. Simple and closed form expressions are
available only for simple cases, such as potential problems or isotropic elasticity. Therefore, the development of efficient schemes for computing the fundamental solutions of generic linear systems of partial differential equations
(PDEs) still represents a challenge of great scientific interest.
In the field of materials engineering, several works have been devoted to
finding the displacements field [69, 120] and its derivatives [23] due to a point
force in a three-dimensional anisotropic elastic medium. The formal expression of the fundamental solutions of a second order partial differential operator has been classically obtained using either Fourier or Radon transforms,
which lead to expressions in terms of a contour integral on a unit circle [208].
By using suitable variable transformations, several researchers replaced the
contour integral by an infinite line integral [220], whose solution has been obtained by means of the Cauchy’s residue theorem in terms of the Stroh eigenvalues [229, 181, 118] or eigenvectors [129, 148]. However, when the Stroh’s
eigenvalues or eigenvectors approach is used, the issue of degeneracy needs
25
26
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
to be robustly addressed, in particular when such an approach is used in a
numerical code, e.g. in boundary element implementations. Non-degenerate
cases were first studied by Dederichs and Leibfried [55] for cubic crystals.
Phan et al. [163, 164] presented a technique to compute the fundamental solutions of a 3D anisotropic elastic solid and their first derivatives in presence
of multiple roots and using the residue approach. Shiah et al. [195] used the
spherical coordinates differentiation to obtain the explicit expressions of the
derivatives of the fundamental solutions for an anisotropic elastic solid up to
the second order. Although being exact, the main disadvantage of these approaches is the necessity of using different expressions for each different case
of different roots, two coincident roots, and three coincident roots. A unified
formulation, valid for degenerate as well as non-degenerate cases, has been
first presented by Ting and Lee [215]. Their approach has been further investigated by the recent work of Xie et al. [233], in which the authors developed
a unified approach to compute the fundamental solutions of 3D anisotropic
solids valid for partially degenerate, fully degenerate and non-degenerate materials. Although not suffering from material’s degeneracy, the expressions
presented in the work of Xie et al. [233] are valid only for anisotropic elastic
materials, are given up to second order of differentiation and are rather long
and complex, in particular when the derivatives are considered. A 2D Radon
transform approach has been also used in the literature as an alternative approach to the problem [42, 232].
The aforementioned works have been mainly devoted to the derivation of
the fundamental solution of elastic anisotropic materials. Recently, multi-field
materials have been receiving an increasing interest for their application in
composite multi-functional devices [149, 161]. Closed form expressions can be
found for transversely isotropic materials showing piezo-electric [63, 64] and
magneto-electro-elastic [223, 204, 92, 61] coupling. Pan and Tonon [158] used
the Cauchy’s residue theorem to derive the fundamental solutions of nondegenerate anisotropic piezo-electric solids, and a finite difference scheme to
obtain their derivatives. Buroni and Saez [43] used the Cauchy’s residue theorem and the Stroh’s formalism to obtain the fundamental solutions of degenerate or non-degenerate anisotropic magneto-electro-elastic materials. Their
scheme was recently employed in a boundary element code for fracture analysis [142].
From a numerical perspective, the fundamental solutions represent the
essential ingredient in boundary integral formulations, such as the Boundary
Element Method (BEM) [17, 228, 6]. In practical BEM analyses of engineer-
2.1 Introduction
27
ing interest, the fundamental solutions and their derivatives are computed in
the order of million times and the availability of efficient schemes for their
evaluation is thus of great interest, especially for large 3D problems. To accelerate the computations for anisotropic elastic materials, Wilson and Cruse
[227] proposed to pre-compute the values of the fundamental solutions at regularly spaced points of a spatial grid and to use an interpolation scheme with
cubic splines to approximate their values in general points during the subsequent BEM analysis. Such an approach and similar interpolation techniques
[184] have been widely employed in the BEM literature [34, 138, 29, 28, 80, 33].
Mura and Kinoshita [144] represented the fundamental solutions of a general anisotropic elastic medium in terms of spherical harmonics expansions
and used a term-by-term differentiation to obtain the first derivative. Aubrys
and Arsenalis [16] used the spherical harmonics expansions for dislocation
dynamics in anisotropic elastic media and pointed out that line integrals and
double line integrals could be obtained analytically once the series coefficients
were computed. Recently, Shiah et al. [196] proposed an alternative scheme to
compute the fundamental solutions of 3D anisotropic elastic solids based on a
double Fourier series representation. The authors expressed the fundamental
solutions as given by Ting and Lee [215] in the spherical reference system and
then built their Fourier series representation relying on their periodic nature.
The authors underlined that the coefficients of the series were computed only
once for a given material and employed their method in a BEM code [209].
They also obtained the first and the second derivatives of the fundamental solutions whose complexity increases with the order of differentiation, despite
the use of the spherical coordinates to obtain the derivatives. The interested
reader is referred to the book by Pan and Chen [157] and to the recent paper
by Xie et al. [232] for a comprehensive overview of the available methods to
obtain the fundamental solutions.
In the present Chapter, given a generic linear system of PDEs defined by a
homogeneous elliptic partial differential operator, the fundamental solutions
and their derivatives are computed in a unified fashion in terms of spherical
harmonics expansions [81]. It is here demonstrated that the formula found
by Mura and Kinoshita [144] is in fact a particular case of a more general
representation of the fundamental solutions and their derivatives, which are
not obtained by a term-by-term differentiation and can be computed up to
the desired order. The coefficients of the series depend on the material constants and need to be computed only once, thus making the present scheme
attractive for efficient boundary integral formulations. Eventually, mathemat-
28
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
ically degenerate media do not require any specific treatment and the present
scheme can be generally employed to cases ranging from simply isotropic
to more complex general anisotropic differential operators. To the best of the
author’s knowledge, it is the first time that the fundamental solutions for generally anisotropic multi-field materials and their derivatives up to any order
are represented in such compact unified fashion.
The Chapter is organised as follows: Section(2.2) introduces the class of
systems of partial differential equations and the corresponding fundamental
solutions that will be addressed in the present study; Section(2.3) illustrates
the mathematical steps needed to obtain the expressions of the fundamental
solutions and their derivatives in terms of spherical harmonics; Section(2.4)
presents a few results from the proposed scheme: first it is shown that, in
the case of isotropic operators, the proposed representation leads to exact expressions of the fundamental solutions; then a few numerical tests covering
isotropic elastic, generally anisotropic elastic, transversely isotropic and generally anisotropic piezo-electric and magneto-electro-elastic media are presented and discussed.
2.2
Problem statement
The linear behaviour of different classes of multi-field materials, such as PiezoElectric (PE), Magneto-Electric (ME), or Magneto-Electro-Elastic (MEE) materials, can be represented by a system of generally coupled partial differential
equations (PDEs)
Lij (∂ x )φj (x) + f i (x) = 0,
(2.1)
where x =∈ R3 is the spatial independent variable, φj (x) represent the unknown functions of x, f i (x) represent the known generalized volume forces, and
i, j = 1, . . . , N where N is the number of equations as well as the number of
unknown functions. Lij (∂ x ) is supposed to be a general homogeneous elliptic partial differential operator involving a linear combination of second order
derivatives of x, i.e. Lij (∂ x ) = cijkl ∂2 (·) /∂xk ∂xl , where cijkl are the material
constants.
The system of PDEs (2.1) may be specialised to several specific problems
ranging from the classical Laplace equation up to the governing equations
for general anisotropic magneto-electro-elastic materials, as shown in Section
(2.4).
2.2 Problem statement
29
The system of PDEs in Eq.(2.1) is defined ∀x ∈ V ⊆ R3 , being V the material domain, and it is mathematically closed by enforcing a suitable set of
boundary conditions over the boundary S = ∂V of V. The Boundary Element
Method [17, 228, 6] is based on the integral representation of Eq.(2.1). In particular, using the Green’s identities, it is possible to express the values of the
functions φi (y) at any interior point y ∈ V in terms of the values of φi (x) and
their derivatives on the boundary S as follows:
φ p (y) =
Z
S
Φ pi (x, y)τi (x) − Tpi (x, y)φi (x) dS(x) +
Z
V
Φ pi (x, y) f i (x)dV (x),
(2.2)
where
τi (x) = nk (x)cijkl
∂φj
( x ),
∂xl
Tpi (x, y) = nl (x)c jikl
∂Φ pj
(x, y).
∂xk
(2.3)
Φij (x, y) are the fundamental solutions of the system of PDEs (2.1) that,
for a given material specified by the constants cijkl , depend only on the relative position between the collocation point y and the observation point x, i.e.
Φij (x, y) ≡ Φij (x − y). Defining the distance vector r ≡ x − y, Φij (r) is obtained as the solution of the adjoint problem:
Lij∗ (∂r )Φ jp (r) + δpi δ(r) = 0,
(2.4)
where Lij∗ (∂r ) = L ji (∂r ), δpi is the Kronecker delta function and δ(r) is the
Dirac delta function: Φ jp (r) is then the value of the field component function
φj (r) at the point r = x − y due to a generalised unit point force δpi at r = 0.
Upon applying the Fourier transform to Eq.(2.4), the fundamental solutions Φij (r) can be written as
Φij (r) =
where
1
8π 3
Z
R3
e ij (ξ ) exp(i ξ k rk )dξ
Φ
h
i −1
e ij (ξ ) = Lij∗ (ξ )
Φ
.
(2.5)
(2.6)
3
In Eqs.(2.5-2.6),
the Fourier transform
√ ξ ∈ R is the variable spanning
domain and i = −1 denotes the imaginary unit; Lij∗ (ξ ) = c jikl ξ k ξ l is usually
referred to as the symbol of the partial differential operator Lij∗ (∂ x ) [226]. Here
and in the following, it is assumed that Lij∗ (ξ ) can be inverted ∀ξ ∈ R3 . It
30
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
is worth noting that the symbol Lij∗ (ξ ) is homogeneous of order n = 2 and,
e ij (ξ ) is homogeneous of order n = −2.
therefore, Φ
In Eq.(2.5), Φij (r) are given in integral form and an explicit closed form expression can be obtained for particular cases only, e.g. for isotropic materials
(Kelvin solution); for general anisotropic materials, the integral in Eq.(2.5) can
be explicitly given in terms of the Stroh’s eigenvalues [215, 119, 43, 233], which
however have to be computed for each couple of collocation and observation
points, thus inducing high computational costs when employed in numerical
codes. In this work, the fundamental solutions Φij (r) and their derivatives are
obtained using the spherical harmonics expansion. The method is illustrated
in the next Section.
2.3
Fourier transform and Rayleigh expansion
Let us consider a function F (r), r ∈ R3 , and its Fourier transform representation
Z
1
F (r) =
Fe(ξ ) exp(i ξ k rk )dξ
(2.7)
8π 3 R3
where Fe(ξ ) is the Fourier transform of F (r), i.e.
Fe(ξ ) =
Z
R3
F (r) exp(−i ξ k rk )dr.
(2.8)
Following Adkins [5] and using the Rayleigh expansion (see Eq.(2.64)), in
combination with the spherical harmonic addition theorem (see Eq.(2.70)), the
quantity exp (i ξ k xk ) in Eq.(2.7) can be expressed as
∞
exp (i ξ k rk ) = 4π
`
∑ ∑
i` J` (ξ r )Y`m (r̂)Ȳ`m (ξ̂ )
(2.9)
`=0 m=−`
√
√
where r = rk rk , ξ = ξ k ξ k , r̂ = r/r, ξ̂ = ξ/ξ, J` (·) are the spherical Bessel
¯ indicates complex conjufunctions, Y`m (·) are the spherical harmonics and (·)
gation. Upon substituting Eq.(2.9) into Eq.(2.7), one obtains
F (r) =
1
2π 2
∞
`
∑ ∑
`=0 m=−`
i` Y`m (r̂)
Z
R3
Fe(ξ ) J` (ξ r )Ȳ`m (ξ̂ )dξ.
(2.10)
If we let Fe(ξ ) be homogeneous of order n, then Fe(ξ ξ̂ ) = ξ n Fe(ξ̂ ), and the
integral over R3 in Eq.(2.10) can be separated into the integral over the unit
2.3 Fourier transform and Rayleigh expansion
31
sphere and the integral over ξ ∈ [0, ∞), i.e.
1
F (r) =
2π 2
∞
`
∑ ∑
i` Y`m (r̂)
Z
`=0 m=−`
1
=
2π 2
∞
`
∑ ∑
i` Y`m (r̂)
`=0 m=−`
R3
ξ n Fe(ξ̂ ) J` (ξ r )Ȳ`m (ξ̂ )dξ =
Z ∞
ξ
0
n +2
J` (ξ r )dξ
Z
S1
Fe(ξ̂ )Ȳ`m (ξ̂ )dS(ξ̂ ), (2.11)
where S1 the surface of the unit sphere. In Eq.(2.11), the integral with respect
to ξ is evaluated as
Z ∞
0
ξ n+2 J` (ξ r )dξ =
√ 2n+1 Γ[(` + n + 3)/2]
π n +3
,
r
Γ[(` − n)/2]
(2.12)
where Γ(·) is the Gamma function. The result in Eq.(2.12) is in general valid
for −3 < n + ` < −1 + `; however, it can be extended to the range n + ` > −3,
as showed by Adkins [5]. Using the result obtained in Eq.(2.12), Eq.(2.11)
becomes
∞
`
1
F (r) = 2 n+3 ∑ a` (n) ∑ Fe`m Y`m (r̂)
(2.13)
π r
`=0
m=−`
where
a` (n) =
√
π i` 2n
Γ[(` + n + 3)/2]
=
Γ[(` − n)/2]
π i Γ[(` + n + 3)/2]
π
√ h
` + i · sin
` 2n
π cos
2
2
Γ[(` − n)/2]
and
Fe`m =
Z
S1
Fe(ξ̂ )Ȳ`m (ξ̂ )dS(ξ̂ ).
(2.14)
(2.15)
From the second equality in Eq.(2.14), it is possible to show that the coefficients a` (n) are linked to the associated Legendre polynomials P`m (t) and
Qm
` ( t ) evaluated at t = 0. Indeed, using Eq.(2.62), the following identities
hold
(−1)n/2 n+2
(−1)n/2 n+2
P` (0) + i
Q` (0), n even,
(2.16a)
4
2
(−1)(n−1)/2 n+2
(−1)(n−1)/2 n+2
a` (n) =
Q ` (0) + i π
P` (0), n odd.
(2.16b)
2
4
As an example, Figure (2.1) shows the behaviour of the coefficients a` (n) for
n = −2 and n = 0, which, as it will be demonstrated next, are related to the
fundamental solutions and their second derivatives, respectively.
a` (n) = −π
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
32
1.00
60
Re[aℓ (n)]
Im[aℓ (n)]
Re[aℓ (n)]
Im[aℓ (n)]
45
0.50
30
0.25
15
aℓ (n = 0)
aℓ (n = −2)
0.75
0.00
-0.25
0
-15
-0.50
-30
-0.75
-45
-1.00
-60
0
2
4
6
8
10 12 14 16 18 20
0
2
4
6
8
10 12 14 16 18 20
ℓ
ℓ
(a)
(b)
Figure 2.1: Behavior of the coefficients a` (n) for (a) n = −2, (b) n = 0.
2.3.1
Fundamental solutions of general anisotropic materials
h
i −1
e ij (ξ ) = L ∗ (ξ )
Starting from Eq.(2.5) and recalling that Φ
is homogeneous
ij
of order n = −2, the fundamental solutions of the system of PDEs in (2.1) can
be written using Eq.(2.13) and Eqs.(2.16). Before writing the series expression
e ij (ξ ) is a real
of the fundamental solutions, it is worth noting that, since Φ
quantity, the sum over m in Eq.(2.13) is also real and, therefore, in order to
obtain a real quantity with respect to the sum over `, only the real part of the
coefficients a` (−2) is retained. The following expression is then obtained
Φij (r) =
1
4πr
where
e ij`m =
Φ
∞
∑
`=0
` even
Z
S1
`
P` (0)
∑
e ij`m Y m (r̂)
Φ
`
(2.17)
m=−`
e ij (ξ̂ )Ȳ m (ξ̂ )dS(ξ̂ ).
Φ
`
(2.18)
In Eq.(2.17), the sum over ` ranges over even values since P` (0) = 0 for odd `.
The first derivatives of the fundamental solutions can be computed using
Eq.(2.6) and the properties of the Fourier transform of a function derivative.
In the Fourier transform domain, the first derivative of the fundamental solue ij (ξ ), which is homogeneous of
tions with respect to rα can be written as i ξ α Φ
e ij (ξ ) is purely imagorder n = −1. In this case, n is odd and the quantity i ξ α Φ
inary. Therefore, only the imaginary part of the coefficients a` (−1) is retained.
2.3 Fourier transform and Rayleigh expansion
33
The following expression is obtained
∂Φij
1
(r) =
∂rα
4πr2
where
`m
e ij,α
Φ
=
Z
S1
∞
∑
`
∑
P`1 (0)
`=1
` odd
`m m
e ij,α
Φ
Y` (r̂)
(2.19)
m=−`
e ij (ξ̂ )Ȳ m (ξ̂ )dS(ξ̂ )
ξ̂ α Φ
`
(2.20)
In Eq.(2.19), the sum over ` ranges over odd ` values as P`1 (0) = 0 for even `.
Similarly, the spherical harmonics expansion fot the second derivative with
respect to rα and r β is obtained as
∂2 Φij
1
(r) =
∂rα ∂r β
4πr3
where
`m
e ij,αβ
Φ
=
Z
S1
∞
∑
`=0
` even
P`2 (0)
`
∑
`m
e ij,αβ
Φ
Y`m (r̂)
(2.21)
m=−`
e ij (ξ̂ )Ȳ m (ξ̂ )dS(ξ̂ ).
ξ̂ α ξ̂ β Φ
`
(2.22)
The results given in Eqs.(2.17), (2.19) and (2.21) can be generalised to higherorder derivatives of the fundamental solutions and written using the following unified form
∂( I ) Φij
(i )
(i )
(i
∂r1 1 ∂r2 2 ∂r3 3
(r) =
)
where
e `m
Φ
ij,(i1 ,i2 ,i3 ) =
1
4πr I +1
Z
S1
∞
`
`∈L
m=−`
∑ P`I (0) ∑
m
e `m
Φ
ij,(i1 ,i2 ,i3 ) Y` ( r̂ ),
e ij (ξ̂ )Ȳ m (ξ̂ )dS(ξ̂ ).
(ξ̂ 1 )i1 (ξ̂ 2 )i2 (ξ̂ 3 )i3 Φ
`
(2.23)
(2.24)
In Eq.(2.23), I = i1 + i2 + i3 , L is the set of positive even (odd) integers
when I is even (odd). Mura and Kinoshita [144] pointed out that the `-th coefficients of the spherical harmonics expansion of the fundamental solutions is
the value of the Legendre polynomial of degree ` evaluated at zero. Here, it is
proved that such a relationship is in fact a particular case of the more general
expression given in Eq.(2.23). Furthermore, Eq.(2.23) shows that the fundamental solutions as well as their derivatives can be written as the product of a
regular part, expressed as a spherical harmonics expansion, by a singular part
depending only on powers of r.
34
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
e `m
In the following, the compact notation Φ
will be used to denote
ij,(i1 ,i2 ,i3 )
the coefficients of degree ` and order m of the fundamental solutions Φij (r),
whose derivative is taken i1 times with respect to r1 , i2 times with respect to
e `m
r2 and i3 times with respect to r3 . The terms Φ
will then represent the
ij,(0,0,0)
coefficients of the fundamental solutions’ expansion.
From the expression given by Eq.(2.24), two properties of the expansion
e `m
e `m
coefficients Φ
are worth mentioning. First, the coefficients Φ
ij,(i1 ,i2 ,i3 )
ij,(i1 ,i2 ,i3 )
depend only on the material properties and can be computed only once, in advance in any numerical implementation; additionally, using Eq.(2.68), it is
possible to show that
`m
ē ij,(i ,i ,i ) .
e `(−m) = (−1)m Φ
Φ
1 2 3
ij,(i ,i ,i )
(2.25)
1 2 3
Second, it is interesting to note that when the variable ξ̂ is expressed in spherical coordinates {ϑ, ϕ} as ξ̂ = {sin ϑ cos ϕ, sin ϑ sin ϕ, cos ϑ }, the following
identities hold for its components
r
ξ̂ 1 =
i
2π h −1
Y1 (ξ̂ ) − Y11 (ξ̂ ) ,
3
(2.26a)
r
ξ̂ 2 = i
i
2π h −1
Y1 (ξ̂ ) + Y11 (ξ̂ ) ,
3
r
4π 0
ξ̂ 3 =
Y (ξ̂ ).
3 1
(2.26b)
(2.26c)
Then, using the expression of the Clebsch-Gordon series in Eq.(2.69), it is
e `m
e `m
possible to compute the expansion coefficients Φ
,Φ
and
ij,(i1 +1,i2 ,i3 )
ij,(i1 ,i2 +1,i3 )
`
m
`
m
e
e
Φ
in terms of the coefficients Φ
, thus avoiding further inteij,(i1 ,i2 ,i3 +1)
ij,(i1 ,i2 ,i3 )
gration. In fact, it is sufficient to compute the coefficients of the fundamental
solutions (including the coefficients with odd values of `), and the use the
following recurrence relations for the coefficients of higher-order derivatives:
r
e `m
Φ
ij,(i1 +1,i2 ,i3 ) =
`+1
c̃`mk (−1)
`+1
c̃`mk (1)
2` + 1
e k ( m +1) + ∑ √
e k ( m −1)
− ∑ √
Φ
Φ
ij,(i1 ,i2 ,i3 )
2
2k + 1
2k + 1 ij,(i1 ,i2 ,i3 )
k =k` (−1)
k = k ` (1)
m
m
(2.27a)
2.3 Fourier transform and Rayleigh expansion
r
e `m
Φ
ij,(i1 ,i2 +1,i3 ) = i
`+1
35
`+1
c̃`mk (−1)
c̃`mk (1)
2` + 1
e k ( m +1) + ∑ √
e k ( m −1)
√
Φ
Φ
∑
ij,
(
i
,i
,i
)
2
3
1
2
2k + 1
2k + 1 ij,(i1 ,i2 ,i3 )
k =k` (−1)
k = k ` (1)
m
m
(2.27b)
e `m
Φ
ij,(i1 ,i2 ,i3 +1) =
√
`+1
2` + 1
∑
k =k`m (0)
c̃`mk (0)
√
2k + 1
e km
Φ
ij,(i1 ,i2 ,i3 )
(2.27c)
` k c 1` k
where k`m (m0 ) ≡ max{|1 − `|, |m + m0 |} and c̃`mk (m0 ) ≡ c1000
.
m0 (−m)(−m+m0 )
As a last remark on the series expansion given in Eq.(2.23), it is interesting
to note that such a representation allows to obtain some integral properties
of the fundamental solutions, which may be useful for numerical implementation. As an example, the technique proposed by Gao and Davis [72] for
the evaluation of strongly singular volume integrals, arising e.g. in isotropic
plasticity problems, required that the integral over the unit sphere S1 of the
regular part of the strain kernel, involving the second derivatives of the fundamental solution, vanished. Such a requirement has been proved by Gao and
Davis [72] for isotropic elasto-plasticity; however, it was only numerically verified by Benedetti et al. [33] for Ranisotropic crystal plasticity of copper. Using
Eq.(2.23) and considering that S1 Y`m (r̂)dS(r̂) = 0 ∀`, m 6= 0 and P0I (0) = 0
∀ I > 0, it is possible to show that, in fact, the integral over S1 of the regular
part of all the derivatives of the fundamental solutions of any second-order homogeneous elliptic operator vanishes. This constitutes a relevant byproduct
of the present unified treatment.
2.3.2
Convergence
In this Section, it is shown that the spherical harmonics representation in Eq.(2.23)
in fact coincides with the representation of the fundamental solution in terms of unit
circle integration.
Considering the fundamental solutions and using Eq.(2.23), one has
Φij (r) =
1
4πr
∞
`
`=0
m=−`
∑ P` (0) ∑
e ij`m Y m (r̂) =
Φ
`
1
8π 2 r
Z
∞
2` + 1
e ij (ξ̂ )dS(ξ̂ ) (2.28)
P` (0) P` (ξ̂ k r̂k )Φ
2
S1 `=0
∑
where the second equality is obtained using the addition theorem (2.70). Using the completeness property of the Legendre polynomials reported in Eq.(2.63),
36
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
the above expression simplifies to
1
Φij (r) =
8π 2 r
Z
S1
e ij (ξ̂ )δ(ξ̂ k r̂k )dS(ξ̂ ) =
Φ
1
8π 2 r
I
C
e ij (ξ̂ )dc
Φ
(2.29)
where the last integral is taken over the unit circle C defined over the plane
ξ̂ k r̂k = 0, which is exactly the classical unit circle integral.
Consider now the derivative with respect to rα of the fundamental solutions, which will be indicated as Φij,α ≡ ∂Φij /∂rα . Using the addition theorem
(2.70), the spherical harmonics expansion reads
1
Φij,α (r) =
8π 2 r2
Z
∞
2` + 1 1
e ij (ξ̂ )ξ̂ α dS(ξ̂ ).
P` (0) P` (ξ̂ k r̂k )Φ
2
S1 `=0
∑
(2.30)
Using now the completeness relation given in Eq.(2.63) and applying the operator (−1)(1 − t2 )1/2 ∂(·)/∂t to both sides, one obtains
∞
2` + 1 1
P` (t) P` (s) = (1 − t2 )1/2 δ0 (s − t).
2
`=0
∑
(2.31)
where the right-hand side is intended in a distributional sense. Using the
above relation with t = 0, the definition of the derivative of the Dirac delta
function and Eq.(2.30), one obtains
Φij,α (r) =
1
8π 2 r2
Z
S1
e ij (ξ̂ )ξ̂ α δ0 (ξ̂ k r̂k )dS(ξ̂ ) =
Φ
1
8π 2 r2
I
C
e ij (ξ̂ )ξ̂ α ]
∂[Φ
∂(ξˆk r̂k )
dc.
ξ̂ k r̂k =0
(2.32)
Upon considering a reference system aligned with the direction r̂, the above
expression coincides exactly with that used by Schclar [184] to derive the first
derivative of the fundamental solutions of an anisotropic elastic material.
The above technique can be generalised to higher-order derivatives and
confirms that the spherical harmonics expansion in Eq.(2.23) is in fact an alternative representation of the fundamental solutions and their derivatives.
2.3.3
Pseudo-Algorithm
To conclude the present Section, the following steps are presented to show the
compactness of representing the fundamental solutions and their derivatives
in terms of spherical harmonics, given a generic elliptic system of PDEs:
2.4 Results
37
1. Consider an elliptic system of PDEs defined by the operator Lij (∂ x ) as
given in Eq.(2.1)
2. Define the symbol Lij∗ (ξ ) of the adjoint differential operator Lij∗ (∂ x ) ≡
L ji (∂ x );
e `m
3. Compute the coefficients Φ
for the desired orders i1 , i2 , i3 using
ij,(i1 ,i2 ,i3 )
Eqs.(2.24) and (2.27);
e ij (r) and their derivatives using
4. Compute the fundamental solutions Φ
Eq.(2.23).
2.4
Results
In the present Section, the proposed technique is employed to compute the
fundamental solutions of different classes of systems of PDEs (2.1). The present
scheme is first used to exactly retrieve the fundamental solutions for two
isotropic cases, namely the classic Laplace equation and the equations of isotropic
elasticity. Then, the scheme is used to compute the fundamental solutions
and their derivatives for transversely isotropic and generally anisotropic materials, covering from the elastic to the magneto-electro-elastic cases. In such
e `m
cases, the series in Eq.(2.23) are truncated and the coefficients Φ
are
ij,(i1 ,i2 ,i3 )
computed for ` = 0, 1, . . . , L, where L will be referred to as the series truncation
number. To demonstrate the accuracy of the present scheme for any combination of source point y and observation point x, the following error is defined
over the unit sphere S1 centered at y
Φij (r̂) − ΦijR (r̂)
∆ Φij (r̂) =
.
||ΦijR (r̂)||S1
(2.33)
Similarly, the norm of the error over S1 is defined as
eS1 Φij (r̂) =
||Φij (r̂) − ΦijR (r̂)||S1
||ΦijR (r̂)||S1
.
(2.34)
R
In Eqs.(2.33) and (2.34), || · ||S1 = S1 | · |dS(r̂), Φij (r̂) represents the fundamental solutions using the present scheme, and ΦijR (r̂) represents a reference
value for the fundamental solutions computed using a reference technique.
38
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
2.4.1
Isotropic materials examples
In this Section, two classic isotropic cases are considered. In particular, the
well-known fundamental solutions of the Laplace equation and the isotropic
elasticity are retrieved showing that, in both cases, the solutions are obtained
in exact form as the series expansions involve only a finite number of terms.
Fundamental solution of the isotropic Laplace equation
In the present case, referring to Eq.(2.1), the unknown function is denoted by
φ(x) and the differential operator Lij (∂ x ) is the Laplace operator, i.e. Lij ≡
∇2 ≡ ∂2 (·) /∂xk ∂xk . The adjoint operator Lij∗ ≡ L ∗ coincides with ∇2 (selfadjoint operator) and its symbol is simply L ∗ (ξ ) ≡ ξ k ξ k ≡ ξ 2 . Then, it follows
that [L ∗ (ξ̂ )]−1 = 1. In the previous expressions, the indices i, j have been
dropped being φ(x) a scalar function.
Denoting the fundamental solution by Φ(r) and considering Eq.(2.24), the
coefficients of the spherical harmonics expansions are computed as follows
e `m
Φ
,(i1 ,i2 ,i3 ) =
Z
S1
(ξ̂ 1 )i1 (ξ̂ 2 )i2 (ξ̂ 3 )i3 Ȳ`m (ξ̂ )dS(ξ̂ ),
(2.35)
The integrals in Eqs.(2.35) can be analytically evaluated using Eqs.(2.26) and
the orthogonality property of the spherical harmonics over the unit sphere,
see Eq.(2.67); it is possible to show that the coefficients Φ,`(mi ,i2 ,i3 ) are identi1
cally zero for ` > i1 + i2 + i3 . As an example, the expressions of the fundamental solution Φ(r) and its derivatives Φ,1 (r), Φ,12 (r), Φ,112 (r) are explicitly computed in the following. The directional dependence of the fundamental solution is indicated by means of the unit vector r̂ = {r̂1 , r̂2 , r̂3 } =
{sin ϑ cos ϕ, sin ϑ sin ϕ, cos ϑ}. Moreover, only the coefficients with m ≥ 0 will
be given since the coefficients with negative m can be obtained using Eq.(2.25).
Consider the fundamental solution Φ(r) first. From Eq.(2.35), the coefficients Φ,`(m0,0,0) represent the integrals of the spherical harmonics over S1 , which
0
are
√ different from zero only for ` = m = 0. The integral of Ȳ0 (r̂) over S1 equals
2 π and, using Eq.(2.23), one has
Φ(r) =
1
.
4πr
(2.36)
Next, consider the first derivative Φ,1 (r) of the fundamental solution. Us-
2.4 Results
39
ing Eq.(2.35), the coefficients Φ,`(m1,0,0) are computed as
Φ,`(m1,0,0) =
Z
ξ̂ 1 Ȳ`m (ξ̂ )dS(ξ̂ ).
S1
(2.37)
The only non-zero coefficient is in fact
√
e 11
Φ
,(1,0,0) = − 2π/3,
(2.38)
which, using Eq.(2.23), provides
Φ,1 (r) = −
r̂1
.
4πr2
(2.39)
Then, consider the second derivative Φ,12 (r). The coefficients Φ,`(m1,1,0) are
computed as
Z
e `m
=
ξ̂ 1 ξ̂ 2 Ȳ m (ξ̂ )dS(ξ̂ ).
Φ
(2.40)
,(1,1,0)
`
S1
In this case, the only non-zero coefficient is
√
e 22
Φ
,(1,1,0) = −i 2π/15,
(2.41)
which, using Eq.(2.23), provides
Φ,12 (r) =
3r̂1 r̂2
.
4πr3
(2.42)
Eventually, the third derivative Φ,112 (r) is considered. The coefficients
Φ,`(m2,1,0) are computed as
e `m
Φ
,(2,1,0) =
Z
S1
(ξ̂ 1 )2 ξ̂ 2 Ȳ`m (ξ̂ )dS(ξ̂ )
(2.43)
and the only non-zero coefficients are
1
e 11
Φ
,(2,1,0) = i 5
√
2π/3,
1
e 31
Φ
,(2,1,0) = −i 5
√
π/21,
√
e 33
Φ
,(2,1,0) = i π/35,
(2.44)
which, using Eq.(2.23), provide
Φ,112 (r) =
3r̂2 (1 − 5r̂12 )
.
4πr4
(2.45)
40
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
Fundamental solution of isotropic elasticity
The fundamental solutions for isotropic elasticity are here retrieved using the
proposed unified formulation. In the present case, the unknown functions are
denoted by u j (x), j = 1, 2, 3 and represent the components of the displacement
field. The operator Lij (∂ x ) is
Lij = (λ + µ)
∂2
∂2
δij
+µ
∂xi ∂x j
∂xk ∂xk
(2.46)
where λ and µ are the Lamé constants. As Lij is symmetric, the adjoint operator Lij∗ coincides with Lij and its symbol is Lij∗ (ξ ) = (λ + µ)ξ i ξ j + µξ k ξ k δij .
Then, it follows that the inverse of Lij∗ (ξ̂ ) is
h
Lij∗ (ξ̂ )
i −1
=
1
1
δij −
ξ̂ i ξ̂ j
µ
2µ(1 − ν)
(2.47)
where ν is the Poisson coefficient.
Denoting the fundamental solutions of isotropic elasticity by Φij (r), the
e `m
coefficients Φ
are computed as
ij,(i1 ,i2 ,i3 )
Z
1
1
`
m
i
i
i
2
3
1
e
ξ̂ i ξ̂ j Ȳ`m (ξ̂ )dS(ξ̂ ). (2.48)
δij −
(ξ̂ 1 ) (ξ̂ 2 ) (ξ̂ 3 )
Φ
ij,(i1 ,i2 ,i3 ) =
µ
2µ(1 − ν)
S1
The integrals in Eqs.(2.48) can be once again evaluated in closed form using
Eqs.(2.26) and the orthogonality property Eq.(2.67) of the spherical harmonics
e `m
and, in particular, it is possible to show that Φ
are identically zero for
ij,(i1 ,i2 ,i3 )
` > i1 + i2 + i3 + 2. As an example, the fundamental solution Φ11 (r) and
its fourth derivative Φ11,1123 (r) will be computed. Also in this case, only the
coefficients with m ≥ 0 will be reported as the coefficients with negative m
can be obtained using Eq.(2.25).
Let us consider the fundamental solution Φ11 (r). Using Eq.(2.48), the co`m
efficients Φ11,
(0,0,0) are computed as
Z
1
1
`
m
e
Φ
δij −
ξ̂ i ξ̂ j Ȳ`m (ξ̂ )dS(ξ̂ ).
(2.49)
11,(0,0,0) =
µ
2µ(1 − ν)
S1
The only non-zero coefficients are
√
√
π (5 − 6ν)
π/5
00
20
e
e
Φ11,(0,0,0) =
, Φ11,(0,0,0) =
,
3µ(1 − ν)
3µ(1 − ν)
√
e 22
Φ
11,(0,0,0) = −
π/30
,
µ (1 − ν )
(2.50)
2.4 Results
41
which, using Eq.(2.23), provide
Φ11 (r) =
3 − 4ν + r̂12
.
16πµ(1 − ν)r
(2.51)
Consider now the fourth derivative Φ11,1123 (r) of the fundamental solution
`m
Φ11 (r). Using Eq.(2.48), the coefficients Φ11,
(2,1,1) are computed as
e `m
Φ
11,(2,1,1) =
Z
2
S1
(ξ̂ 1 ) ξ̂ 2 ξ̂ 3
1
1
ξ̂ i ξ̂ j Ȳ`m (ξ̂ )dS(ξ̂ ).
δij −
µ
2µ(1 − ν)
(2.52)
In this case, the non-zero coefficients are
√
√
π/30
(
5
−
6ν
)
π/5(8 − 11ν)
21
41
e
e
Φ
, Φ
11,(2,1,1) = i
11,(2,1,1) = −i 231µ (1 − ν ) ,
21µ(1 − ν)
√
√
π/35(8 − 11ν)
π/546
61
e
=i
, Φ
11,(2,1,1) = i 33µ (1 − ν ) ,
33µ(1 − ν)
√
√
3π/455
3π/1001
63
65
e
e
Φ
Φ
11,(2,1,1) = i 22µ (1 − ν ) ,
11,(2,1,1) = −i 6µ (1 − ν ) ,
e 43
Φ
11,(2,1,1)
(2.53)
which provide
Φ11,1123 (r) =
2.4.2
15[1 − 4ν + 14(1 + 2ν)r̂12 − 63r̂14 ]r̂2 r̂3
.
16πµ(1 − ν)r5
(2.54)
Anisotropic materials examples
In the present Section, the proposed scheme is used to compute the fundamental solutions of anisotropic elastic, piezoelectric and magneto-electro-elastic
materials. The governing equations are given for a general magneto-electroelastic material and then particularised to the elastic and piezo-electric cases.
The system of PDEs for an anisotropic MEE material can be written as follows
[156, 43, 142]:
∂2 φK
(x) + f J (x) = 0
(2.55)
ciJKl
∂xl ∂xi
where the unknown functions φK (x), K = 1, 2, . . . , 5 are given by φK (x) =
uK (x) if K ≤ 3, φK (x) = ϕ(x) if K = 4 and φK (x) = ψ(x) if K = 5, being
uk (x), k = 1, 2, 3 the components of the displacement field, ϕ(x) the electric
42
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
potential and ψ(x) the magnetic potential. The constants ciJKl are the multifield coupling coefficients of the MEE material and are defined by
ciJKl
cijkl ,
elij ,
eikl ,
qlij ,
=
qikl ,
−λil ,
−
ε il ,
−κil ,
J, K ≤ 3
J ≤ 3, K = 4
J = 4, K ≤ 3
J ≤ 3, K = 5
J = 5, K ≤ 3
J = 4, K = 5 or J = 5, K = 4
J=K=4
J=K=5
(2.56)
where cijkl , ε il and κil are the elastic stiffness tensor, the dielectric permittivity tensor and the magnetic permeability tensor, respectively, whereas elij , qlij
and λil are the piezoelectric, piezomagnetic and magneto-electric coupling
tensors. The aforementioned tensors satisfy the following symmetries
cijkl = c jikl = cklij ,
ε il = ε li ,
λil = λli ,
elij = elji ,
qkij = qkji
(2.57)
κil = κli ,
which ensure that the symbol L JK (ξ ) ≡ CiJKl ξ l ξ i of the system of PDEs (2.55)
∗ ( ξ ) are symmetric and coincident. Eventually, the genand its adjoint L JK
eralised volume forces are { f J (x)} = { f 1 (x), f 2 (x), f 3 (x), − f e (x), − f m (x)},
being f j (x), j = 1, 2, 3 the mechanical body forces, f e (x) the electric charge
density and f m (x) electric current density.
The fundamental solutions of a generally anisotropic magneto-electro-elastic
material are indicated by Φ JK (r) and are computed using the general form of
Eq.(2.23), where the coefficients of the expansions are computed using Eq.(2.24)
e JK (ξ̂ ) is defined as Φ
e JK (ξ̂ ) = [L ∗ (ξ̂ )]−1 = [CiJKl ξ̂ l ξ̂ i ]−1 .
and where Φ
JK
The fundamental solutions of elastic, piezoelectric and magneto-electroelastic materials are presented next.
2.4 Results
43
Elastic materials
First, the fundamental solutions of anisotropic elastic materials are presented.
In this case, the coupling tensors elij , qlij and λil are set to zero and only the
elastic response is considered. Three FCC crystals are considered, namely
Nickel (Ni), Gold (Au) and Copper (Cu). The non-zero elastic constants are
reported in Table (2.1) of Section (2.7). Figure (2.2) reports the error eS1 over
10 0
10
Φ11
Φ12
Φ11,1
Φ12,1
Φ11,12
Φ12,12
-2
10
eS 1
eS 1
10 -4
-3
-6
10 -5
10
-8
10
-7
10
10
-9
10
-10
0
10
20
30
Φ11
Φ12
Φ11,1
Φ12,1
Φ11,12
Φ12,12
10 -1
40
0
10
L
10
-3
Φ11
Φ12
Φ11,1
Φ12,1
Φ11,12
Φ12,12
10 -5
10
-7
10
-9
0
10
20
L
(c) Cu
40
(b) Au
eS 1
eS 1
-1
30
L
(a) Ni
10
20
30
40
10
-1
10
-3
Φ11
Φ12
Φ11,1
Φ12,1
Φ11,12
Φ12,12
10 -5
10
-7
10
-9
0
10
20
30
40
L
(d) Cu rotated
Figure 2.2: Error over the unit sphere for the fundamental solutions up to
the second derivative as a function of the series truncation number L for three
FCC crystals: (a) Nickel, (b) Gold, (c) Copper in principal reference system and
(d) Copper in a rotated reference system. The reference values are computed
using the unit circle integration.
the unit sphere, as defined in Eq.(2.34), as a function of the series truncation
44
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
number L. The unit circle integration with an accuracy of 14 digits (decimal
places) is used as reference value. The figure reports the fundamental solutions up to the second derivative and shows that the higher the order of
derivation the higher is the number of terms to be retained to obtain the same
level of accuracy.
Figure (2.2) also shows that the rate of convergence is linear in a logarithmic diagram with respect to the series truncation number. Furthermore,
it is worth underlining that rotating the material reference system does not
substantially affect the efficiency of the scheme. In particular, Figure (2.2d)
shows the error over the unit sphere for the spherical harmonics expansion
of the fundamental solutions of the FCC Copper whose reference system has
been inclined by 60 degrees with respect to the x1 -x2 plane and rotated by
30 degrees in the x1 -x2 plane. The dashed lines represent the convergence of
the spherical harmonics expansion when the material properties matrix is expressed in the principal reference system. Several combinations of rotation
angles have been tested, but the convergence of the spherical harmonics expansions remained almost unaffected.
The effect of the level of anisotropy on the accuracy of the spherical harmonics expansion is assessed considering an FCC elastic crystal with different
values of the Zener anisotropy ratio. Figure (2.3) shows the error over the unit
10 0
eS 1
10 -5
10 -10
10 -15
10 -2
10 -1
10 0
10 1
10 2
A
Figure 2.3: Error over the unit sphere for the fundamental solutions Φij (r̂)
of FCC crystals as a function of the Zener anisotropic ratio for fixed series
truncation number L = 40.
sphere for the fundamental solutions of an FCC elastic crystal whose elastic
coefficients c11 and c12 are taken from crystalline Copper, and the coefficient
2.4 Results
45
c44 is chosen as c44 = A(c11 − c12 )/2, where the Zener anisotropy ratio A takes
values from 1/50 to 50. From the figure it is possible to assess the effect of the
degree of anisotropy on the error. In particular, as the value of A moves away
from 1, i.e. the isotropic case, the error increases.
i={1,...,N }
max ∆(r̂i ) × 100 [%]
10 2
Φuc
ij
Φuc
ij,k
Φuc
ij,kl
Φsh
ij
Φsh
ij,k
Φsh
ij,kl
10 0
10 -4
10 -8
10 -12
0
0.5
1
1.5
2
2.5
computing time/N [ms]
Figure 2.4: Fundamental solutions for crystalline Cu: computing time versus
maximum error for the fundamental solutions Φij and their first and second
derivatives obtained using the unit circle integration and the spherical harmonics expansion. N denotes the number of source and observation points.
In the legend, the superscript uc stands for unit circle whereas the superscript
sh for spherical harmonics.
To show the efficiency of the spherical harmonics technique, Figure (2.4)
plots the average computing time per collocation/observation couple versus
the maximum error of the considered fundamental solutions with respect to
the selected reference value. The figure has been obtained by computing the
fundamental solutions of an FCC Cu crystal and their derivatives up to the
second order for N couples of collocation and observation points, using both
the unit circle integration and spherical harmonics expansion. The points in
the figure have been obtained by varying the number of Gauss points of the
unit circle integration and the series truncation number L. It is clear that the
higher the number of Gauss points (as well as the series truncation number),
the higher the computing time and the smaller the error. From the figure, it
is possible to see that the use of the spherical harmonics technique is more
efficient than the unit circle integration since, in order to achieve the same
accuracy in the fundamental solutions, it requires less computing time. However, a few considerations should be made:
46
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
• The figure does not include the time needed to compute the coefficients
of the spherical harmonics series since such time is fixed for any number
of evaluation points and can be predominant only if a small number of
computation points is considered. Such coefficients would be computed
ad stored in advance in any effective implementation;
• The error of the unit circle integration is linear with respect to the required computing time, whereas the error of the spherical harmonics
expansion has a quadratic behaviour with respect to the required computing time (since number of coefficients to be computed scales with
L2 ); as a consequence, there is a level of accuracy at which the unit circle
integration results more advantageous than the spherical harmonics expansion. However, reaching such a level of accuracy may be not always
necessary in a numerical code or beyond the machine precision;
• The presented results depend on the degree of anisotropy of the considered material as well as on the number of unknown functions of the
considered system of PDEs. A more exhaustive understanding of the
efficiency of the spherical harmonics expansion would definitely benefit from an a priori knowledge of the number of series coefficients required to obtain a certain level of accuracy, once the material properties
are specified. However, the task is not trivial and is left open to further
investigation.
10 0
Φ11
Φ11,1
Φ11,111
Φ11,112
Φ11,113
Φ11,122
Φ11,123
Φ11,133
Φ11,222
Φ11,223
Φ11,233
Φ11,333
eS 1
10 -2
10 -4
10 -6
10
-8
0
20
40
60
L
Figure 2.5: Fundamental solutions for crystalline Cu: error, eS1 [Φ11,ijk (r̂)], over
the unit sphere for the third derivatives of the fundamental solutions Φ11 as a
function of the series truncation number L. The error for Φ11 (r̂) and Φ11,1 (r̂)
are reported for comparison purposes.
2.4 Results
47
Figure (2.5) shows the error eS1 for the third derivatives of the fundamental
solution Φ11 of FCC Cu crystals as a function of the series truncation number
L. In this case, the fundamental solutions computed with L = 60 are used as
reference. The value L = 60 has been selected as it has been verified that the
addition of further terms would only affect after the 14th decimal place, thus
the found values would practically coincide with those provided by the unit
circle integral. For comparison purposes, the figure also reports the error eS1
for the fundamental solution Φ11 (r̂) and its first derivative Φ11,1 (r̂).
Figure (2.6) presents a few selected fundamental solutions for Cu crystals.
In order to appreciate their directional dependence, Figures (2.6a,c,e) plot the
fundamental solutions Φ11 (r̂), Φ11,1 (r̂), Φ11,12 (r̂) respectively, using a spherical representation defined as
x (ϑ, ϕ) = $(ϑ, ϕ) sin ϑ cos ϕ
y(ϑ, ϕ) = $(ϑ, ϕ) sin ϑ sin ϕ ,
(2.58)
z(ϑ, ϕ) = $(ϑ, ϕ) cos ϑ
where the amplitude $(ϑ, ϕ) is given by the absolute value of the considered
fundamental solution. Figures (2.6b,d,f) show the relative difference between
the fundamental solutions computed using the spherical harmonics expansion and the fundamental solutions computed using the unit circle integration. The difference is plotted over the unit sphere S1 .
In Figure (2.7), some selected third derivatives are plotted using the spherical representation (2.58).
48
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
(a) Φ11 (r̂) [m]
(b) ∆[Φ11 (r̂)]
(c) Φ11,1 (r̂) [-]
(d) ∆[Φ11,1 (r̂)]
(e) Φ11,12 (r̂) [m−1 ]
(f ) ∆[Φ11,12 (r̂)]
Figure 2.6: Fundamental solutions of Cu crystal: (a,c,e) spherical plots of a
few selected fundamental solutions; (b,d,f ) plot over the unit sphere of the
difference between the fundamental solutions computed using the spherical
harmonics expansions and the unit circle integration.
2.4 Results
49
(a) Φ11,111 (r̂) [m−2 ]
(b) Φ11,112 (r̂) [m−2 ]
(c) Φ11,113 (r̂) [m−2 ]
(d) Φ11,122 (r̂) [m−2 ]
(e) Φ11,123 (r̂) [m−2 ]
(f ) Φ11,222 (r̂) [m−2 ]
Figure 2.7: Fundamental solutions for crystalline Cu: spherical plots of some
selected third derivatives of the fundamental solution Φ11 .
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
50
Piezo-electric materials
10
-2
10
-6
10
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[2, 2]
[2, 3]
[2, 4]
[3, 3]
[3, 4]
[4, 4]
-10
10 -14
0
10
20
L
(a)
30
10
-2
10
-6
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[2, 2]
[2, 3]
[2, 4]
[3, 3]
[3, 4]
[4, 4]
eS 1
R
|Φij,2 − ΦR
ij,2 |/|Φij,2 |
Next, piezo-electric materials are studied. In this case, the coupling tensors
qlij and λil are set to zero, whereas the piezoelectric coupling tensor elij is
retained. Two piezoelectric materials are considered, namely a transversely
isotropic lead zirconate titanate (PZT-4) ceramic and an orthotropic piezoelectric polyvinylidene fluoride (PVDF), whose non-zero elastic, piezoelectric and
dielectric constants are reported in Table (2.2) and Table (2.3), respectively. The
plane of isotropy of PZT-4 is the x1 -x2 plane.
Figure (2.8a) shows the convergence of the spherical harmonics expansions of the first derivative Φij,2 (r) of PZT-4 computed at r = {1, 1, 1}. As
reference, the results reported in Ref. [43] and computed using the explicit expressions of the fundamental solutions are used. Figure (2.8b) shows the error
eS1 for the same derivative.
10 -10
10
-14
0
10
20
30
L
(b) eS1 [Φij,2 (r̂)]
Figure 2.8: Fundamental solutions for transversely isotropic PZT-4: a) error
for the first derivative Φij,2 (r) computed at r = {1, 1, 1} as a function of the
series truncation number L (the reference values are computed using the explicit expressions of the fundamental solutions as in Ref. [43]); b) error over
the unit sphere for the first derivatives Φij,2 (r̂) for the fundamental solutions
as a function of the series truncation number L (the reference values are computed using the unit circle integration).
Figure (2.9) reports the error eS1 of the fundamental solutions and their
derivatives up to the second order for the orthotropic piezoelectric material
PVDF. The unit circle integration is used as reference value. Figure (2.10)
presents a few selected fundamental solutions for PVDF. Figures (2.10a,c,e)
2.4 Results
51
plot the fundamental solutions Φ34 (r̂), Φ34,1 (r̂), Φ34,12 (r̂) respectively using
the spherical representation (2.58), whereas Figures (2.10b,d,f) show the error
with respect to the unit circle integration.
-6
eS 1
10
10 -10
10
10
10
20
-6
10 -10
-14
0
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[2, 2]
[2, 3]
[2, 4]
[3, 3]
[3, 4]
[4, 4]
10 -2
eS 1
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[2, 2]
[2, 3]
[2, 4]
[3, 3]
[3, 4]
[4, 4]
10 -2
10
30
-14
0
10
L
(a) eS1 [Φij (r̂)]
0
10
-3
10
-6
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[2, 2]
[2, 3]
[2, 4]
[3, 3]
[3, 4]
[4, 4]
10 -9
10
-12
0
10
30
(b) eS1 [Φij,1 (r̂)]
20
L
(c) eS1 [Φij,11 (r̂)]
30
10
eS 1
eS 1
10
20
L
0
10
-3
10
-6
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[2, 2]
[2, 3]
[2, 4]
[3, 3]
[3, 4]
[4, 4]
10 -9
10
-12
0
10
20
30
L
(d) eS1 [Φij,12 (r̂)]
Figure 2.9: Fundamental solutions for orthotropic piezoelectric PVDF: error
for a) the fundamental solutions Φij (r̂), b) the first derivative Φij,1 (r̂), c) the
second derivative Φij,11 (r̂), and d) the second derivative Φij,12 (r̂) as a function
of the series truncation number L. The reference values are computed using
the unit circle integration.
52
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
(a) Φ34 (r̂) [m]
(b) ∆S1 [Φ34 (r̂)]
(c) Φ34,1 (r̂) [-]
(d) ∆S1 [Φ34,1 (r̂)]
(e) Φ34,12 (r̂) [m−1 ]
(f ) ∆S1 [Φ34,12 (r̂)]
Figure 2.10: Fundamental solutions for orthotropic piezoelectric PVDF: (a,c,e)
spherical plots of a few selected fundamental solutions; (b,d,f ) plot over the
unit sphere of the difference between the fundamental solutions computed
using the spherical harmonics expansions and the unit circle integration.
2.4 Results
53
Magneto-electro-elastic materials
The fundamental solutions of magneto-electro-elastic materials are finally computed. Two MEE materials are considered: a transversely isotropic MEE material indicated by M1 and an orthotropic MEE material indicated by M2 , whose
properties are reported in Tables (2.4) and (2.5), respectively.
10
-3
10
-6
10
-9
10
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[1, 5]
[3, 3]
[3, 4]
[3, 5]
[4, 4]
[4, 5]
[5, 5]
-12
0
10
20
L
(a)
30
10 0
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[1, 5]
[3, 3]
[3, 4]
[3, 5]
[4, 4]
[4, 5]
[5, 5]
10 -3
eS 1
R
|Φij − ΦR
ij |/|Φij |
10 0
10
-6
10 -9
10
-12
0
10
20
30
L
(b) eS1 [Φij (r̂)]
Figure 2.11: Fundamental solutions for transversely isotropic magnetoelectro-elastic material M1 : (a) error for the fundamental solutions Φij (r) computed at r = {1, 1, −1} as a function of the series truncation number L (the reference values are computed using the explicit expressions of the fundamental
solutions as in Ref. [156]); (b) error over the unit sphere for the fundamental
solution Φij (r̂) as a function of the series truncation number L (the reference
values are computed using the unit circle integration).
Figure (2.11a) shows the convergence of the spherical harmonics expansions of the fundamental solutions Φij (r) of the material M1 computed at
r = {1, 1, −1}. As reference, the results reported in Ref. [156] and computed
using the explicit expressions of the fundamental solutions are used. Figure
(2.11b) shows the error eS1 for the same fundamental solutions.
Figure (2.12) shows the error eS1 for the fundamental solutions and their
derivatives up to the second order of the orthotropic MEE material M2 . The
unit circle integration is used as reference value. Figure (2.14) displays a few
selected fundamental solutions for the considered material: Figures (2.14a,c,e)
plot the fundamental solutions Φ11 (r̂), Φ11,1 (r̂), Φ11,12 (r̂) respectively using
the spherical representation (2.58), whereas Figures (2.14b,d,f) show the error
with respect to the unit circle integration.
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
54
Figure (2.13) shows the error eS1 for the fourth derivatives of the fundamental solution Φ11 of the MEE material M2 as a function of the series truncation number L. In this case, the fundamental solutions computed with L = 40
are used as reference. For comparison purposes, the figure also reports the
error eS1 for the fundamental solution Φ11 (r̂) and its first, second and third
derivative Φ11,1 (r̂), Φ11,11 (r̂) and Φ11,111 (r̂). In Figure (2.15) some selected
fourth derivatives are plotted using the spherical representation (2.58).
0
10
-4
10
-8
10 -12
10
-16
0
10
20
30
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[1, 5]
[2, 2]
[2, 3]
[2, 4]
[2, 5]
[3, 3]
[3, 4]
[3, 5]
[4, 4]
[4, 5]
[5, 5]
10
eS 1
eS 1
10
0
10
-4
10
-8
10 -12
10
-16
0
10
L
(b) eS1 [Φij,1 (r̂)]
eS 1
-6
10 -10
0
10
20
L
(c) eS1 [Φij,11 (r̂)]
30
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[1, 5]
[2, 2]
[2, 3]
[2, 4]
[2, 5]
[3, 3]
[3, 4]
[3, 5]
[4, 4]
[4, 5]
[5, 5]
10 -2
10
-6
eS 1
10 -2
10 -14
30
L
(a) eS1 [Φij (r̂)]
10
20
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[1, 5]
[2, 2]
[2, 3]
[2, 4]
[2, 5]
[3, 3]
[3, 4]
[3, 5]
[4, 4]
[4, 5]
[5, 5]
10 -10
10 -14
0
10
20
30
[1, 1]
[1, 2]
[1, 3]
[1, 4]
[1, 5]
[2, 2]
[2, 3]
[2, 4]
[2, 5]
[3, 3]
[3, 4]
[3, 5]
[4, 4]
[4, 5]
[5, 5]
L
(d) eS1 [Φij,12 (r̂)]
Figure 2.12: Fundamental solutions for orthotropic magneto-electro-elastic
material M2 : error for (a) the fundamental solutions Φij (r̂), (b) the first derivative Φij,1 (r̂), (c) the second derivative Φij,11 (r̂), and (d) the second derivative
Φij,12 (r̂) as a function of the series truncation number L. The reference values
are computed using the unit circle integration.
2.4 Results
55
10
-2
eS 1
10 -6
10 -10
10 -14
0
10
20
L
30
40
Φ11
Φ11,1
Φ11,11
Φ11,111
Φ11,1111
Φ11,1112
Φ11,1113
Φ11,1122
Φ11,1123
Φ11,1133
Φ11,1222
Φ11,1223
Φ11,1233
Φ11,1333
Φ11,2222
Φ11,2223
Φ11,2233
Φ11,2333
Φ11,3333
Figure 2.13: Fundamental solutions for orthotropic magneto-electro-elastic
material M2 : error, eS1 [Φ11,ijkl (r̂)], over the unit sphere for the fourth derivatives of the fundamental solutions Φ11 as a function of the series truncation
number L. The reference values are fundamental solutions computed for
L = 40. The error for Φ11 (r̂), Φ11,11 (r̂), and Φ11,111 (r̂) are reported for comparison purposes.
56
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
(a) Φ11 (r̂) [m]
(b) ∆S1 [Φ11 (r̂)]
(c) Φ11,1 (r̂) [-]
(d) ∆S1 [Φ11,1 (r̂)]
(e) Φ11,12 (r̂) [m−1 ]
(f ) ∆S1 [Φ11,12 (r̂)]
Figure 2.14: Fundamental solutions for orthotropic magneto-electro-elastic
material M2 : (a,c,e) spherical plots of a few selected fundamental solutions;
(b,d,f ) plot over the unit sphere of the difference between the fundamental
solutions computed using the spherical harmonics expansions and the unit
circle integration.
2.4 Results
57
(a) Φ11,1111 (r̂) [m−3 ]
(b) Φ11,1122 (r̂) [m−3 ]
(c) Φ11,1133 (r̂) [m−3 ]
(d) Φ11,1223 (r̂) [m−3 ]
(e) Φ11,1233 (r̂) [m−3 ]
(f ) Φ11,3333 (r̂) [m−3 ]
Figure 2.15: Fundamental solutions for orthotropic magneto-electro-elastic
material M2 : spherical plots of a few selected fourth derivatives of the fundamental solution Φ11 .
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
58
2.5
Associated Legendre polynomials
This appendix summarises some of the properties of the associated Legendre
polynomials used in this Chapter. Further details about orthogonal polynomials and more comprehensive discussions on their properties can be found
in Ref. [4]. The associated Legendre polynomials are the particularisation to
integer values of ` of the associated Legendre functions, which are solutions
of the associated Legendre differential equation
d
m2
2 d m
ym (t) = 0.
(1 − t ) y` (t) + `(` + 1) −
dt
dt
1 − t2 `
(2.59)
m
m
The general solution of Eq.(2.59) is written as ym
` ( t ) = αP` ( t ) + βQ` ( t ),
m
m
where α and β are constants and P` (t) and Q` (t) are referred to as the associated Legendre functions of the first and second kind, respectively. Since
the associated Legendre polynomials P`m (t) are mainly employed for obtaining the fundamental solutions in terms of spherical harmonics expansions,
some of their properties are listed in the present appendix.
The associated Legendre polynomials of degree ` and order m, P`m (t), can
be expressed in terms of the Legendre polynomials as
`+m
dm
(−1)m
2 m/2 d
P
(
t
)
=
)
(
1
−
t
( t2 − 1) `
`
dtm
2` ` !
dt`+m
(2.60)
where (·)! denotes the factorial. From their definition, the associated Legendre polynomials with negative m, P`−m , can be written as
P`m (t) = (−1)m (1 − t2 )m/2
P`−m (t) = (−1)m
(` − m)! m
P ( t ).
(` + m)! `
(2.61)
The associated Legendre polynomials satisfy several identities and recurrence relations that can be found in handbooks [4]. One of the identities,
which is of interest for the developments presented here, is the value of P`m (t)
and Qm
` ( t ) at t = 0:
hπ
i Γ[(` + m + 1)/2]
2m
P`m (0) = √ cos
(` + m)
;
2
Γ[(` − m + 2)/2]
π
hπ
i Γ[(` + m + 1)/2]
√
m −1
Qm
(
0
)
=
−
2
π
sin
(`
+
m
)
.
`
2
Γ[(` − m + 2)/2]
(2.62a)
(2.62b)
2.6 Spherical harmonics
59
For m = 0, the associated Legendre polynomials coincide with the Legendre polynomials that satisfy the following completeness relation
∞
2` + 1
P` (t) P` (s) = δ(t − s)
2
`=0
∑
(2.63)
where δ(·) is the Dirac delta function and t, s ∈ (−1, 1). Furthermore, it is
worth recalling here the Rayleigh expansion, also known as plane wave expansion, which expresses a plane wave as the infinite sum of spherical waves [12]
and is written as follows
∞
exp(iξ k rk ) =
∑ (2` + 1)i` J` (ξ r) P` (ξ̂ k r̂k )
(2.64)
`=0
where J` (·) are the spherical Bessel functions, r = {rk }√and ξ = {ξ k }, k =
√
1, 2, 3, are three-dimensional vectors and r = rk rk , ξ = ξ k ξ k , r̂k = rk /r and
ξ̂ k = ξ k /ξ.
2.6
Spherical harmonics
This appendix summarises some of the properties of the spherical harmonics
Y`m (ξ̂ ) used in the present Chapter. Further details about the spherical harmonics and more comprehensive discussions on their properties can be found
in Ref. [12]. Let ξ̂ = ξ̂ (ϑ, ϕ), ϑ ∈ [0, π ], ϕ ∈ [0, 2π ], a unit vector spanning the
unit sphere S1 in R3 , the spherical harmonic Y`m (ξ̂ ) ≡ Y`m (ϑ, ϕ) of degree `
and order m is defined as follows:
m
Y`m (ξ̂ ) = Y`m (ϑ, ϕ) = cm
` P` (cos ϑ ) exp(imϕ ),
with
s
cm
`
=
2` + 1 (` − m)!
,
4π (` + m)!
(2.65)
(2.66)
√
and where i = −1 is the imaginary unit, (·)! denotes the factorial and P`m (·)
are the associate Legendre polynomials. The normalisation coefficient cm
` is
m1
m2
chosen such that two spherical harmonics Y`1 (ϑ, ϕ) and Y`2 (ϑ, ϕ) are orthogonal over S1 as follows
Z
S1
Y`m1 1 (ϑ, ϕ)Ȳ`m2 2 (ϑ, ϕ)dS(ϑ, ϕ) = δm1 m2 δ`1 `2 ,
(2.67)
60
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
¯ denotes the complex conjugate and δij is the Kronecker delta funcwhere (·)
tion. The spherical harmonic with negative m, Y`−m , can be written as
Y`−m (ϑ, ϕ) = (−1)m Ȳ`m (ϑ, ϕ).
(2.68)
The product of two spherical harmonics can be obtained using the ClebschGordon series
s
kf
(2`1 + 1)(2`2 + 1) `1 `2 k `1 `2 k
c000 cm m (m +m ) Ykm1 +m2 (ϑ, ϕ)
Y`m1 1 (ϑ, ϕ)Y`m2 2 (ϑ, ϕ) = ∑
2
1 2
1
4π
(
2k
+
1
)
k=ki
(2.69)
`1 `2 `
where k i = max{|`1 − `2 |, |m1 + m2 |}, k f = `1 + `2 and cm1 m2 m denotes the
Clebsch-Gordon coefficients, that is c`m11`m2 `2 m ≡ h`1 `2 m1 m2 |`1 `2 ` mi.
The spherical harmonics are related to the Legendre polynomials through
the following addition theorem
`
2` + 1
P` (ξ̂ k r̂k ) = 2π ∑ Y`m (ξ̂ )Y`m (r̂)
2
m=−`
where r̂ and ξ̂ are two unit vectors, and ξ̂ k r̂k = ξ̂ 1 r̂1 + ξ̂ 2 r̂2 + ξ̂ 3 r̂3 .
(2.70)
2.7 Materials properties
2.7
61
Materials properties
Table 2.1: Elastic properties for Copper (Cu), Gold (Au) and Nickel (Ni) from
Ref. [29].
material
Cu
property
elastic constants [109 N/m2 ]
Au
elastic constants [109 N/m2 ]
Ni
elastic constants [109 N/m2 ]
component
c1111 , c2222 , c3333
c1122 , c1133 , c2233
c1212 , c1313 , c2323
c1111 , c2222 , c3333
c1122 , c1133 , c2233
c1212 , c1313 , c2323
c1111 , c2222 , c3333
c1122 , c1133 , c2233
c1212 , c1313 , c2323
value
168
121
75
185
158
39.7
251
150
124
Table 2.2: Piezoelectric properties for transversely isotropic lead zirconate titanate (PZT-4) ceramic from Ref. [158].
property
elastic constants [109 N/m2 ]
piezoelectric constants [C/m2 ]
dielectric permeability
constants [10−9 C/(V · m)]
component
c1111 , c2222
c3333
c1122
c1133 , c2233
c2323 , c1313
c1212
e113 , e223
e333
e322 , e311
ε 11 , ε 22
ε 33
value
139
115
77.8
74.3
25.6
(c1111 − c1122 )/2
12.7
15.1
-5.2
6.461
5.620
62
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
Table 2.3: Piezoelectric properties for orthotropic piezoelectric polyvinylidene
fluoride (PVDF) from Ref. [142].
property
elastic constants [109 N/m2 ]
piezoelectric constants [C/m2 ]
dielectric permeability
constants [10−9 C/(V · m)]
component
c1111
c1122
c1133
c2222
c2233
c3333
c2323
c1313
c1212
e113
e223
e333
e311
e322
ε 11
ε 22
ε 33
value
3.61
1.61
1.42
3.13
1.31
1.63
0.55
0.59
0.69
-0.016
-0.013
-0.021
0.032
-0.004
0.054
0.066
0.059
2.7 Materials properties
63
Table 2.4: Magneto-electro-elastic properties for the transversely isotropic material M1 from Ref. [156].
property
elastic constants [109 N/m2 ]
piezoelectric constants [C/m2 ]
piezomagnetic constants [N/(A · m)]
magneto-electric
coefficients [N · s/(A · m)]
dielectric permeability
coefficients [10−9 C/(V · m)]
magnetic permeability
coefficients [10−6 N · s2 /C2 ]
component
c1111 , c2222
c3333
c1122
c1133 , c2233
c1313 , c2323
c1212
e113 , e223
e333
e322 , e311
q113 , q223
q333
q322 , q311
λ11 , λ22
λ33
e11 , e22
e33
κ11 , κ22
κ33
value
166
162
77
78
43
(c1111 − c1122 )/2
11.6
18.6
-4.4
550.0
699.7
580.3
0
0
11.2
12.6
5
10
64
Fundamental solutions for multi-field materials based on spherical
harmonics expansions
Table 2.5: Magneto-electro-elastic properties for the orthotropic material M2
from Ref. [142].
property
elastic constants [109 N/m2 ]
piezoelectric constants [C/m2 ]
piezomagnetic constants [N/(A · m)]
magneto-electric
coefficients [N · s/(A · m)]
dielectric permeability
constants [10−9 C/(V · m)]
magnetic permeability
coefficients [10−6 N · s2 /C2 ]
component
c1111
c1122
c1133
c2222
c2233
c3333
c2323
c1313
c1212
e113
e223
e333
e311
e322
q113
q223
q333
q311
q322
λ11
λ22
λ33
ε 11
ε 22
ε 33
κ11
κ22
κ33
value
3.61
1.61
1.42
3.13
1.31
1.63
0.55
0.59
0.69
-0.016
-0.013
-0.021
0.032
-0.004
550.0
570.0
699.7
580.3
590.0
0.6
0.8
1.0
0.054
0.066
0.059
5.0
7.0
10.0
Chapter 3
An enhanced grain-boundary
framework
In this Chapter, an enhanced three-dimensional framework for computational
homogenisation and intergranular cracking of polycrystalline materials is presented. The framework is aimed at reducing the computational cost of polycrystalline micro simulations. The scheme is based on the grain-boundary formulation described in Chapter (1). A regularisation scheme is used to avoid
excessive mesh refinements often induced by the presence of small edges and
surfaces in mathematically exact 3D Voronoi or Laguerre morphologies. For
homogenisation purposes, periodic boundary conditions are enforced on nonprismatic periodic micro Representative Volume Elements (µRVEs), eliminating pathological grains generally induced by the procedures used to generate
prismatic periodic µRVEs. An original meshing strategy is adopted to retain
mesh effectiveness without inducing numerical complexities at grain edges
and vertices. The proposed methodology offers remarkable computational savings and high robustness, both highly desirable in a multiscale perspective. The
determination of the effective properties of several polycrystalline materials
demonstrate the accuracy of the technique. Several microcracking simulations
complete the study and confirm the performance of the method.
3.1
Introduction
In Chapter (1), it was underlined how the high and less expensive affordability of HPC facilities has been pushing the boundaries of more and more
65
66
An enhanced grain-boundary framework
realistic micro-mechanical models taking into account multiple aspects of the
micro-structures of interest. However, despite such remarkable advancements,
any analysis of macro components or structures including all the details of
the micro-structure, although conceivable in principle, is practically unfeasible,
due to its enormous computational cost. As a consequence, there is interest in developing techniques able to extract information from the finer scales
that can be profitably used at the scale relevant to the technological application. Upon the premise of spatially separated length-scales, multi-scale modelling, responds to such demand, being in fact an effective way to exchange
information between the scales, avoiding the excessive cost of fine-scale modelling of the whole problem domain. In this framework, computational homogenisation plays a key role: a µRVE is associated to each relevant point of
a macro-continuum and provides, through computational homogenisation,
the macro-continuum constitutive behaviour. The collection of µRVEs identify the micro-scale underlying the application macro-scale.
Although multiscale modelling should in principle provide huge savings
with respect to fine scale analyses of the whole domain, the need of resolving
many micro Boundary Value Problems (µBVPs), nested within the macroscale,
still generates huge computational costs, especially when 3D modelling is
pursued, when fine details are included in the µRVEs morphologies, and
when highly non-linear phenomena are investigated, e.g. initiation and propagation of damage. For such reasons, in recent years, remarkable efforts have
been focused on the development of techniques for reducing the computational costs of µRVEs computations [235, 115].
Due to the increased affordability of HPC, three-dimensional modelling
of polycrystalline RVEs has recently received much attention [145, 100, 201,
28, 30]. However, the computational cost of 3D µRVEs simulations is consistently reported as a limiting factor for the scalability of the proposed methods
and, as a consequence, only few 3D multiscale frameworks have been developed. A two-scale finite element approach, successfully demonstrated on two
3D applications, has been developed by Han and Dawson [83]. Nakamachi
et al. [147] proposed a multi-scale finite element procedure employing realistic 3D micro polycrystalline structures obtained by scanning electron microscopy and electron backscattering diffraction (SEM-EBSD) measurements.
The technique was further refined for formability tests [113] and accelerated
with HPC and domain partitioning techniques. Benedetti and Aliabadi [30]
have recently proposed a 3D, homogenisation based, grain-boundary twoscale formulation for degradation and failure in polycrystalline materials: al-
3.1 Introduction
67
though the potential of the technique has been demonstrated, the two-scale
3D non-linear simulations have been hindered by the computational costs of
the µRVEs simulations, pointing out the strong need for model order reduction.
In this Chapter, the polycrystalline problem is tackled adopting a comprehensive strategy aimed at reducing the order of the µRVEs simulations. The
methodology is based on the synergistic use of state-of-the-art techniques for
the analysis of polycrystalline microstructures. First, regularised Laguerre tessellations [167] are employed to remove the majority of those mathematically
exact, but physically inessential, small entities (edges and faces) that generally
challenge the meshing algorithms, inducing unduly large mesh refinements.
The regularised tessellations are then employed within the grain-boundary
formulation developed in Chapter (1), which employs only grain-boundary
variables, thus avoiding the grains internal meshing.
The grain-boundary meshing is enhanced employing continuous and semidiscontinuous elements for the faces of the grains, allowing a smoother and
computationally more effective representation of the intergranular fields with
respect to the basic discontinuous scheme. Additionally, a modified version
for the polycrystalline unit cell is adopted: usually, Voronoi or Laguerre unit
cells are obtained by cutting the unbounded tessellation, generated by some
initial seeds, through some cutting walls used to produce a prismatic RVE.
The cutting operation, however, often re-introduces unnecessary mesh refinements, probably due to the impossibility of avoiding small or pathological
residual entities produced by the cutting operation itself. This can be simply avoided by retaining, in the final morphology, only those grains generated by the original seeds. This simple strategy produces polycrystalline
RVEs with more regular and affordable overall meshes, at the expense of irregular external shapes and a slightly more complex implementation of the
RVEs’ boundary conditions. Although this is a relatively simple expedient,
it has not been used often in the literature, see e.g. [154, 91, 73]. The above
enhancements synergistically contribute to outstanding savings in terms of
overall DoFs count. Their practical implementation and effectiveness is challenged by the high statistical variability of Voronoi or Laguerre aggregates, in
terms of topology and morphology, so that many different cases can be met.
These aspects are thoroughly discussed in the paper, which is organised as
follows. The generation and regularisation of polycrystalline morphologies
is discussed in Sections (3.2) and (3.3), while the adopted meshing strategy is
described in Section (3.4). The tests in Section (3.5) complete the study.
68
3.2
An enhanced grain-boundary framework
Laguerre tessellations regularisation
The adoption of the hardcore and Laguerre constraints, described in Chapter
(1), produces more regular morphologies with respect to the basic PoissonVoronoi tessellation, namely it reduces the occurrence of excessively small
grains edges and faces; however, it does not ensure a minimum smallest edge
length in the tessellation. In other words, very small entities are still mathematically possible in the tessellation and, if they occur, they usually induce
pathological meshing behaviours, i.e. excessive mesh refinements.
Quey et al. [167] have recently developed a regularisation scheme that iteratively removes the small edges whose length is below a predetermined
threshold value. The removal is accomplished by merging the two vertices
defining the small edges. If the algorithm encounters a face formed by three
edges and one of them must be removed, the whole face is deleted. The technique naturally leads to topological and morphological modifications of the
tessellation that, strictly speaking, it is not a Voronoi or a Laguerre tessellation
anymore: the resulting tessellation may have more than four grains intersecting at one vertex and more than three faces intersecting at one edge. Moreover, while the grain faces lying on the external boundary of the aggregate
remain flat, a regularised intergranular interface may result not planar, then
inducing a loss of convexity for the two grains in contact through the considered interface itself. However, the regularised morphology results more
affordable in terms of generated mesh, as the number of small geometrical
entities (edges and faces) is drastically reduced. Further information on the
effects of the regularization on the statistical features of the artificial polycrystalline topology and morphology can be found in Ref. [167].
In the present work, regularised tessellations, produced in accordance with
the referred procedure, are used for the first time in conjunction with the
grain-boundary formulation presented in Chapter (1). This is expected to
produce noticeable computational advantages, which will be tested in the
next sections. The regularisation is performed using the open source software
Neper (http://neper.sourceforge.net/) [167].
3.3
Removal of cut boundary grains
Besides the adoption of regularised tessellations, another modification is proposed here for the generation of computationally effective polycrystalline µRVEs.
In a material homogenisation framework, different boundary conditions
3.3 Removal of cut boundary grains
69
can be enforced on the µRVEs to retrieve the macro-scale properties [102]. Periodic boundary conditions (PBCs) have been shown to provide faster convergence to the overall macro properties with respect to displacement or traction
boundary conditions [214]. Even though it does not represent a requirement,
the use of PBCs is facilitated by the conformity between the meshes of coupled
opposite faces, and therefore by a periodic structure of the µRVE.
Fritzen et al. [70] suggested the following procedure to obtain periodic
microstructures:
1. given the initial seeds distribution inside the original domain, the Ng
seeds are copied into the 26 boxes surrounding the original one;
2. the obtained extended bounded domain, containing 27 × Ng seeds, is partitioned into convex polyhedra by means of a suitable tessellation algorithm;
3. the periodic RVE is extracted cutting out the original domain from the
extended one.
Although this procedure leads to an apparently simple cubic periodic microstructure, the use of the cutting planes introduces additional small geometrical entities, e.g. very small grains, which introduces additional sources of
undesired mesh refinement.
Periodic microstructures may also be obtained by repeating steps 1 and 2
above (see Figure (3.1)) and, instead of cutting out the extended tessellation,
by considering only the cells originated by original seeds, that is, those seeds that have
been copied. Given, the same initial distribution of seeds within a unit cube,
two periodic morphologies generated using the above mentioned procedures
are shown in Figure (3.2).
Two couples of opposite boundary faces of the cut morphology are highlighted in Figure (3.3). The adopted construction algorithm enforces opposite
faces of the domain to be of the same shape (they can be perfectly superimposed through simple translation). This feature facilitates the generation of
conforming meshes between opposite faces of the domain external boundary,
and thus the application of PBCs on the µRVEs. Although this procedure has
been successfully employed in homogenisation and multi-scale analyses, see
e.g. [30], the cutting process generally leads to the generation of very small
grains. These small entities cannot be avoided during the analysis of the RVE
and cannot be generally regularised, using for example Neper, and as a consequence they represent a source of overrefinement in the µRVE’s mesh.
70
An enhanced grain-boundary framework
Extended domain
Original domain
Figure 3.1: Schematic generation of periodic morphologies: original domain
containing Ng scattered seeds and extended volume surrounding the original
box (marked in red) and containing additional 26 × Ng points.
Figure (3.4) shows instead some of the coupled faces from the periodic
microstructure, obtained by considering only the grains corresponding to the
initial seeds, those contained within the original domain. In this case, the presence of small cells depends only on the seeds distribution and weights. Nevertheless, the microstructure retains the consistency between opposite faces
of the tessellation. In fact, it can be shown that for each external grain face,
i.e. each face that is not an intergranular interface, there exists an opposite
periodic face of the same shape: PBCs can then be applied as in the case of
cubic unit cells, the only difference resulting from a slightly more involved
detection of coupled opposite periodic faces.
The effects of the regularisation and of the non-prismatic periodic morphology on the number of degrees of freedom for polycrystalline RVEs are
investigated and discussed in Section (3.5.1).
3.3 Removal of cut boundary grains
(a)
71
(b)
Figure 3.2: (a) periodic morphology as obtained after the cutting process; (b)
periodic morphology obtained considering only the cells whose generator
seeds fall within the original domain (the original box is sketched in black
lines).
(a)
(b)
Figure 3.3: (a-b) Two examples of opposite coupled faces obtained after the
cutting process.
72
An enhanced grain-boundary framework
(a)
(b)
Figure 3.4: (a-b) Two examples of opposite coupled faces from the nonprismatic periodic microstructure.
3.4
Grain-boundary meshing
The generation of appropriate meshes for general polycrystalline microstructures is a demanding task, due to the need of preserving accuracy while maintaining affordable system order. Apart the case of regular morphologies (cubes,
octahedra, dodecahedra, etc.), for which the inherent regularity enables regular meshes, the difficulty of meshing randomly generated morphologies originates from their high statistical variability.
In the literature either structured or unstructured meshes have been employed for polycrystalline aggregates [70, 38, 20]. Structured meshes use elements of regular shape, such as cubes, to discretise the entire domain and
attribute to each element the crystallographic properties of the grain they belong to. Such meshes are easy to generate and present high regularity.
However, the interfaces between the grains are poorly resolved, even if
in case of very refined meshes, so that they do not appear adequate for the
modelling of intergranular fields. Unstructured meshes, on the other hand,
are built on the morphology itself and are thus naturally suitable to represent
grains boundaries. Unstructured meshes are usually generated starting from
1D meshes of the tessellation edges. The nodes of this 1D mesh are then used
as input to mesh the grains faces (2D mesh), which in turn provide the nodes
for meshing the grains interior volumes (3D mesh). However, the presence of
small entities causes either the presence of very low quality element or overly
3.4 Grain-boundary meshing
73
resolved regions.
Several meshing algorithms have been proposed in the literature to generate high quality polycrystalline meshes [194, 189, 74, 167]. In this study, the
algorithm proposed by Persson and Strang [162] has been suitably modified
for the generation of regular meshes of the grain faces. The difference with
respect to the original scheme is that the mesh nodes of the grains edges were
used as input for the grains faces meshing, to ensure consistency between the
mesh of contiguous grain faces. Mesh homogeneity was enforced as in Ref.
[29], and a mesh density parameter dm to control mesh refinement. Figures
(3.5a), (3.5b) and (3.5c) show the surface mesh of a Laguerre cell for three different values of dm , respectively dm = 1.0, dm = 2.0, and dm = 3.0.
(a)
(b)
(c)
Figure 3.5: Surface mesh of a Laguerre cell for different values of the discretisation parameter dm : (a) dm = 1.0; (b) dm = 2.0; (c) dm = 3.0.
3.4.1
Geometrical nodes and collocation nodes
As described in Section (1.3) of Chapter (1), the displacement boundary integral equations (1.13), which are reported here for the sake of readability
Z
g
e g (x, y)uei (x)dS(x) =
cepi (y)uei (y) + − T
pi
Sg
Z
Sg
e g (x, y)eti (x)dS(x),
U
pi
(3.1)
are numerically evaluated at the collocation points y of the grain g in order
to obtain the discretised version H g U g = G g T g (see Eq.(1.25)). The procedure is straightforward if such points lie on a smooth surface, as the normal
unit outward vectors are uniquely defined; however, if the collocation points
belong to edges or vertices of the tessellation, the unit normal vector is not
unique. This is a well known issue in the boundary element literature: corner
points, as they are usually referred to, induce more unknowns than equations
74
An enhanced grain-boundary framework
when they are external points with prescribed displacements and when they
are interface points [76]. In a polycrystalline µRVE, such issue is encountered
at each grain edge and vertex, so that the approach adopted to tackle it has a
noticeable effect on the overall effectiveness of the methodology.
In [29, 28], for simplicity of implementation, discontinuous triangular elements were employed, so to ensure that the collocation points would always
lie on smooth surfaces. However, this has been identified as a source of computational ineffectiveness, leading to an unnecessary increase of number of
DoFs, particularly critical in the case of intergranular degradation modelling.
In polycrystalline modelling, and in general in the framework of multi-scale
modelling, the average number of DoFs used in the µBVP is of crucial importance when memory and solution time performances are of concern. For
this reason and in this sense, model reduction with accuracy preservation are
undoubtedly of great interest.
Different strategies can be used to address corner points in boundary elements [76, 56]. In this work, to reduce the computational inefficiency of discontinuous meshing while keeping a relatively easy implementation, continuous and semi-discontinuous elements are simultaneously used for the representation of the grain boundary fields, while a continuous representation of
the geometry is naturally retained.
Consider a sample face as shown in Figure (3.6a), where the edge nodes
are indicated with big dots. The 1D mesh of the edges of the entire tessellation
is performed first and then used as input for the 2D mesh of the tessellation
faces. The face is subdivided into non-overlapping triangles, Figure (3.6b) by
means of the procedure described in Section (3.4). The use of discontinuous
elements for the approximation of the field variables leads to 3 × Nt collocation nodes, where Nt denotes the number of discretisation triangles. Figure
(3.6c) shows the position of the geometrical nodes, which are indicated with
big dots, and the position of the collocation points, indicated with crosses.
The use of continuous and semi-discontinuous elements considerably reduces the number of collocation points. Two different distributions of the
collocation nodes can be obtained, depending on the algorithm selected. The
first distribution, as shown in Figure (3.6d), is obtained by moving the nodes
lying on a face’s edge inside the triangles. In this case each edge node generates as many collocation points as the number of triangles sharing that specific
edge node. Internal geometrical nodes must be added to such edge collocation
nodes to provide the overall number of collocation points associated with a
given grain face.
3.4 Grain-boundary meshing
75
(a)
(b)
(c)
(d)
(e)
Figure 3.6: (a) Sample face with marked edge nodes; (b) 2D triangulation of
the face; (c) Position of the collocation nodes when a discontinuous mesh is
adopted; (d,e) Position of the collocation nodes when two continuous/semidiscontinuous approximations are adopted. In the figures, the small circles
represent the geometrical nodes whereas the small crosses represent the collocation nodes.
76
An enhanced grain-boundary framework
The second distribution, as shown in Figure (3.6e), is instead obtained by
moving the edge nodes along the edges of the mesh triangles. In this case,
each edge node generates as many collocation points as the number of triangles’ edges meeting at the considered node. The overall number of collocation nodes is obtained by adding to these the number of internal geometrical
nodes. As it can be inferred from Figures (3.6d) and (3.6e), the former distribution produces a higher number of collocation nodes with respect to the latter.
However, each triangle may always contain only three collocation nodes. The
latter distribution, on the contrary, produces a smaller number of collocation
nodes, but the resulting triangles may contain from three to six collocation
nodes. The latter distribution therefore requires a suitable specialisation of the shape
functions for each possible case, i.e. number of collocation nodes per triangle.
In the last part of this Section, it is shown that a surface mesh combining
triangular and quadrangular elements leads to a further reduction of collocation nodes with respect to the approach shown in Figure (3.6e), while retaining simplicity in the shape functions representation. Figure (3.7a) shows
some triangles containing more than three collocation nodes after the semidiscontinuous approximation described in Figure (3.6e). In particular, the
four dark-grey triangles contain five collocation nodes, whereas the mid-grey
triangle contains four collocation nodes. By looking at the neighbouring triangles, it can be noted that, by appropriately merging two triangles, a quadrangle with four collocation node can be obtained. This is shown in Figure (3.7b),
where the mid-grey triangles represent the couples of triangles that, after the
merging procedure, would produce a quadrangle with four collocation nodes.
Figure (3.7b) also shows two triangles (in dark-grey) whose merging would
not produce a quadrangle with four collocation nodes. To avoid the generation of this unwanted element, an additional internal node is added to the
face, which is triangulated again as shown in Figure (3.7c). Additional internal nodes are added until the presence of either triangles with three nodes or
quadrangles with four nodes is ensured (usually very few additional nodes
are needed). Triangles with four collocation nodes are allowed, as they are
treated as collapsed quadrangles. Finally, the merging algorithm is employed
and a continuous/semi-discontinuous mesh is obtained as shown in Figure
(3.7d). The resulting mesh has the advantage of drastically reducing the number of collocation nodes with respect to a completely discontinuous approach,
as the one previously adopted in [29, 28]. Furthermore, as it combines triangular and quadrangular elements, the shape functions are straightforward to
deal with.
3.4 Grain-boundary meshing
77
(a)
(b)
(c)
(d)
Figure 3.7: (a) Sample grain face showing semi-discontinuous triangles containing more than three collocation nodes; (b) Couples of matched triangles whose merging produces quadrangular elements with four (mid-grey) or
five (dark-grey) collocation nodes; (c) New triangulation obtained by adding
supplementary internal nodes to avoid generation of quadrangular elements
with more than four collocation points; (d) Resulting mixed triangularquadrangular continuous-semidiscontinuous grain face’s mesh. In the figures, the small circles represent the geometrical nodes whereas the small
crosses represent the collocation nodes.
Figure (3.8) shows the surface mesh of the same Laguerre cell as in Figure (3.5) when the described merging algorithm is employed. It is worth
noting that the internal structure of the mesh is retained as the triangles are
very well-shaped thanks to the meshing algorithm of Section (3.4). More-
78
An enhanced grain-boundary framework
over, merging two internal triangles into a quadrangle would not reduce the
number of collocation nodes. On the other hand, on the faces edges, merging
triangles into quadrangles leads to a reduction of the DoFs associated to the
Laguerre cell. Figs (3.8a), (3.8b) and (3.8c) show the surface mesh for the three
different values of dm , respectively dm = 1.0, dm = 2.0, and dm = 3.0, produced by the optimized meshing algorithm. In this thesis, the last described
meshing algorithm, that is the one combining triangular and quadrangular
semi-discontinuous elements, is adopted.
(a)
(b)
(c)
Figure 3.8: Surface mesh of a Laguerre cell obtained by using the optimised
meshing algorithm with triangular and quadrangular elements, for different
values of the discretisation parameter dm : (a) dm = 1.0; (b) dm = 2.0; (c) dm =
3.0. The small dots represent the collocation points of the mesh.
3.4.2
Continuous/semi-discontinuous elements
In this Section, the triangular and quadrangular elements used for the grain
boundary mesh are discussed with more details.
Similarly to a finite element discretisation, the surface of a grain is discretised into non-overlapping elements. For each element, the geometrical
position of a point x = { x1 , x2 , x3 } in the global coordinate system can be
represented in terms of the coordinates ξ = {ξ 1 , ξ 2 } of the element’s local coordinate system by means of a transformation x = x(ξ ) that, for the triangular
and quadrangular elements used in this thesis, can be expressed as follows
µ
x=
∑ M (α) ( ξ ) x(α) ,
(3.2)
α =1
where M(α) (ξ ) represents the α-th shape function, x(α) represents the coordinates of the α-th geometrical node of the element and µ represent the num-
3.4 Grain-boundary meshing
79
ber of geometrical nodes of the elements. In this thesis, linear triangular and
quadrangular boundary elements have been employed. The two types of elements are reported in Figure (3.9) where the black dots indicate the geometrical nodes.
Unlike the geometry that is represented by continuous shape functions
M(α) (ξ ), a generic unknown boundary field φ is approximated by means of
continuous/semi-discontinuous shape functions that are suitably specialised
to avoid corner problems discussed in Section (3.4.1). The field φ is approximated as follows
ν
φ=
∑ N (β) (ξ, d)φ(β) ,
(3.3)
β =1
where N ( β) (ξ, d) represents the β-th shape function, d is a set of discontinuity
parameters that can be adjusted to obtain continuous to completely discontinuous elements, φ( β) is the value of the field at the collocation node y( β) of the
element and ν is the number of nodal values of the field as well as the number of collocation nodes of the element. Figures (3.9a) and (3.9b) show the
position of the collocation nodes of a discontinuous triangular and a discontinuous quadrangular element, respectively, indicated by small crosses. The
position of the collocation nodes and the expressions of the shape functions
in terms of the discontinuity parameters d are discussed next.
x (3)
x (3)
x (4)
y (4)
y (3)
ξ2
ξ2
y (1)
x
y (3)
ξ1
(1)
y (2)
ξ1
y (2)
y (1)
x (2)
(a)
x (2)
x (1)
(b)
Figure 3.9: (a) discontinuous triangle; (b) discontinuous quadrangle. In the
figures, x(α) , α = 1, . . . , µ are the geometrical nodes whereas y( β) , β = 1, . . . , ν
are the collocation nodes.
80
An enhanced grain-boundary framework
Continuous/semi-discontinuous triangles
For continuous/semi-discontinuous triangular elements, the number of geometrical nodes µ as well as the number of collocation nodes ν is 3; the classical local triangular reference system is centred at x(1) , as shown in Figure (3.9a). The position of the collocation nodes y( β) and the expressions of
the shape functions N ( β) (ξ, d) are given by two discontinuity parameters, i.e.
d ≡ {λ1 , λ2 }. Referring to Figure (3.9a), the local coordinates ξ ( β) of the collocation nodes are given by
ξ (1) = { λ 2 , λ 2 }
(2)
(3.4a)
= {1 − λ1 − λ2 , λ2 }
(3.4b)
ξ (3) = { λ 2 , 1 − λ 1 − λ 2 } .
(3.4c)
ξ
The geometrical position of the collocation node in the global reference system
is then obtained as y( β) = ∑α M(α) [ξ ( β) ]x(α) . From Eqs. (3.4), it is clear that, if
λ1 = λ2 = 0, then the collocation nodes coincide with the geometrical nodes,
i.e. y( β) ≡ x( β) , β = 1, 2, 3, and the element is continuous.
The shape functions N ( β) (ξ, d) of a continuous/semi-discontinuous triangular element can be expressed as follows
1
(1 − λ1 − ξ 1 − ξ 2 )
∆
1
N (2) (ξ, d) = (ξ 1 − λ2 )
∆
1
(3)
N (ξ, d) = (ξ 2 − λ2 )
∆
N (1) (ξ, d) =
(3.5a)
(3.5b)
(3.5c)
where ∆ = 1 − λ1 − 2λ2 . It is straightforward to see that the shape functions
defined in Eqs. (3.5) satisfy the condition that N ( β) (ξ, d) has unit value at ξ =
ξ ( β) and that ∑ β N ( β) (ξ, d) = 1.
Continuous/semi-discontinuous quadrangles
For continuous/semi-discontinuous quadrangular elements, the number of
geometrical nodes µ as well as the number of collocation nodes ν is 4; the local
reference system is centred at ξ = {0, 0}, as shown in Figure (3.9b). The position of the collocation nodes y( β) and the expressions of the shape functions
N ( β) (ξ, d) are given by four discontinuity parameters, i.e. d ≡ {λ1 , λ2 , κ1 , κ2 }.
3.4 Grain-boundary meshing
81
Referring to Figure (3.9b), the local coordinates ξ ( β) of the collocation nodes
are given by
ξ (1) = {−1 + λ1 , −1 + κ1 }
(3.6a)
(2)
= {+1 − λ2 , −1 + κ1 }
(3.6b)
(3)
= {+1 − λ2 , +1 − κ2 }
(3.6c)
ξ (4) = {−1 + λ1 , +1 − κ2 }
(3.6d)
ξ
ξ
The geometrical position of the collocation node in the global reference system is then obtained as y( β) = ∑α M(α) [ξ ( β) ]x(α) . From Eqs. (3.6), it is clear
that, if λ1 = λ2 = κ1 = κ2 = 0, then the collocation nodes coincide with the
geometrical nodes, i.e. y( β) ≡ x( β) , β = 1, 2, 3, 4, and the element is continuous.
The shape functions N ( β) (ξ, d) of a continuous/semi-discontinuous quadrangular element can be expressed as follows
1
(1 − λ2 − ξ 1 )(1 − κ2 − ξ 2 )
∆
1
N (2) (ξ, d) = (1 − λ1 + ξ 1 )(1 − κ2 − ξ 2 )
∆
1
N (3) (ξ, d) = (1 − λ1 + ξ 1 )(1 − κ1 + ξ 2 )
∆
1
(4)
N (ξ, d) = (1 − λ2 − ξ 1 )(1 − κ1 + ξ 2 )
∆
N (1) (ξ, d) =
(3.7a)
(3.7b)
(3.7c)
(3.7d)
where ∆ = (2 − λ1 − λ2 )(2 − κ1 − κ2 ). Similarly to the triangular element, the
shape functions defined in Eqs. (3.7) satisfy the condition that N ( β) (ξ, d) has
unit value at ξ = ξ ( β) and that ∑ β N ( β) (ξ, d) = 1.
82
3.5
An enhanced grain-boundary framework
Computational tests
In this Section, the performance and accuracy of the developed framework
are assessed: the overall computational savings in terms of DoFs count, the
computational homogenisation accuracy and the performance of the microcracking algorithm are considered. The homogenisation and micro-cracking
computations have been performed on single 8-core processors of the Galileo
super-computing facility of the CINECA Super Computing Applications and
Innovation Department in Bologna, Italy (http://www.hpc.cineca.it/hardware/
galileo).
3.5.1
Overall computational savings
The computational savings obtained with the proposed methodology are presented and discussed in terms of overall number of DoFs. First, the effect of
the adoption of the continuous and semi-discontinuous mesh elements is discussed, as it is valid for both non-periodic and periodic morphologies. Then,
the number of DoFs of the non-prismatic periodic morphologies and cubic
periodic morphologies are compared.
Figure (3.10a) shows the number of DoFs, obtained using the different
meshing approaches discussed in Section (3.4), as a function of the number
of tessellation grains Ng . The represented data is the average value over 100
different realisations for each selected number of grains. The dashed lines
with smaller markers refer to non-regularised tessellations, whereas the solid
lines with bigger markers refer to regularised tessellations. It can be noted
that, in all cases, the overall number of DoFs varies linearly with respect to the
number of grains, as the DoFs of each grain add linearly over the DoFs of the
aggregate, but the adoption of the continuous/semi-discontinuous approach,
with simultaneous use of quadrangular and triangular elements, guarantees
the lowest rate after that of the totally continuous approach, which would
however introduce complexities in terms of corner nodes in the formulation.
Considering, as an example, tessellations with Ng = 200, it might be of
interest to observe that the use of only discontinuous elements increases the
number of DoFs of approximately 210% with respect to the use of continuous elements only, whereas the use of the continuous/semi-discontinuous
mesh increases the number of DoFs of only 60% with respect to the continuous mesh, see Table (3.1). The same savings ratios among different meshing
schemes, for the considered number of grains, are maintained if regularised
3.5 Computational tests
83
tessellations are considered, as shown in the second row of Table (3.1). On
the other hand, Table (3.2) exemplifies the effect of the morphology regularisation on the number of DoFs for the different meshing approaches, showing
that the regularisation algorithm provides in average an additional 10% reduction in the number of DoFs with respect to a non-regularised morphology
for tessellations with Ng = 200.
Figure (3.10b) shows the average number of DoFs as a function of the mesh
density parameter dm , for morphologies with Ng = 200 over 100 realisations.
For discontinuous meshes, the number of DoFs grows faster with dm with
respect to semi-discontinuous and continuous meshes.
Figure (3.11) shows the number of DoFs as a function of the number of
grains Ng , for the periodic morphologies discussed in Section (3.4). The nonprismatic periodic morphology allows to further reduce the number of DoFs
in average from 45.3% for Ng = 10 to 27.8% for Ng = 200 with respect to the
cubic periodic morphology. The fact that the higher the number of grains, the
smaller the reduction is consistent with the fact less grains are affected by the
finite size of the cube when Ng increases.
Finally, as an example, considering Ng = 200, it can be interesting to note
that by combining the continuous/semi-discontinuous mesh, the regularisation and the non-prismatic morphology, the overall reduction in terms of DoFs
is around 70%.
Table 3.1: Average increase in the number of DoFs with respect to the (ideal)
continuous mesh for the different analysed meshing schemes, for tessellations
with Ng = 200. The second row refers to regularised morphologies.
DoF−DoFcont
[%]
DoFcont
DoF−DoFcont
DoFcont
reg
[%]
disc.
semidisc. A
semidisc. B
cont./semidisc.
216.8%
201.8%
111.2%
59.9%
211.1%
200.6%
110.7%
58.7%
Table 3.2: Average effect of the morphology regularisation on the number of
DoFs for tessellations with Ng = 200.
DoF−DoFreg
[%]
DoF
disc.
semidisc. A
semidisc. B
cont./semidisc.
11.8%
10.5%
10.4%
9.7%
84
An enhanced grain-boundary framework
300k
DoF
200k
1200k
900k
discontinuous
semidiscontinuous A
semidiscontinuous B
cont./semidiscont.
continuous
DoF
250k
1500k
discontinuous
semidiscontinuous A
semidiscontinuous B
cont./semidiscont.
continuous
150k
600k
100k
300k
50k
0
1020
50
100
200
1.0
1.5
2.0
3.0
dm
Ng
(a)
(b)
Figure 3.10: (a) Comparison of the number of DoFs obtained using different
meshing approaches. The dashed lines with smaller markers refer to nonregularised morphologies; the solid lines with bigger markers refer to regularised morphologies. Data collected on 100 realisations using dm = 1.0; (b)
Number of DoFs as a function of the mesh density parameter dm . The dashed
lines with smaller markers refer to non-regularised morphologies; the solid
lines with bigger markers refer to regularised morphologies. Data collected
on 100 realisations with Ng = 200.
200k
prismatic morphology
non-prismatic morphology
DoF
150k
100k
50k
0
1020
50
100
200
Ng
Figure 3.11: Number of DoFs vs Ng for non-prismatic and prismatic periodic morphologies using the continuous/semi-discontinuous mesh. Data collected over 100 realisations.
3.5 Computational tests
3.5.2
85
Computational homogenisation
In this Section, the developed model is employed to determine the elastic
properties of polycrystalline aggregates, using the non-prismatic periodic morphologies presented in Section (3.2) and the periodic BCs introduced in Section (1.3.1) of Chapter (1). For these analyses, the grain boundaries are considered as perfect and displacements continuity and tractions equilibrium are
enforced, i.e. the interface equations (1.18) of Section (1.3.2) are particularised
as follows
gh
g
gh
g
gh
Ψi = uei (x) + ueih (x) = −δuei (x) = 0,
Θi = eti (x) − etih (x) = 0,
(3.8a)
(3.8b)
where it is recalled that the symbol e· denotes a quantity expressed in the
boundary local reference system, ui and ti denote the grain boundary displacements and tractions, respectively, δui denotes the displacements jumps
and g and h are the two generic adjacent grains sharing an interface. Similarly,
in order to account for the general orientation of the faces in the non-prismatic
periodic morphology, the periodic boundary conditions given in Eq.(1.17) are
expressed in the local reference system of the periodic faces, i.e.
uei (xs ) + uei (xm ) = Rki (xs )Γ̄ki ( x sj − x m
j ),
(3.9a)
eti (xs ) − eti (xm ) = 0,
(3.9b)
where the change of signs is obtained by choosing the rotation matrix Rki (xs )
at the slave point as opposite to the rotation matrix Rki (xm ) at the master point.
The macro-stress and macro-strain averages are calculated for the µRVE
using the following boundary integrals
Σij =
1
V
Z
S
ti (x) x j dS(x),
Γij =
1
2V
Z
S
(ui (x)n j (x) + u j (x)ni (x))dS(x),
which eventually allow to compute the apparent properties Cijkl entering the
constitutive relationships Σij = Cijkl Γkl , see e.g. [29].
As first test, the isotropy of the macroscopic elastic properties of an aggregate of Copper (Cu) crystals is assessed. Similar tests, for the basic grainboundary formulation, have been carried out in [29]. However, it necessary
to assess that the new regularisation and meshing strategies, and the nonprismatic morphology, do not introduce any artefact in the homogenisation
procedure, for example altering the isotropy of the apparent constants Cijkl .
86
An enhanced grain-boundary framework
For this purpose, the complete apparent
notation) of a single 200-grain aggregate
obtaining:
200.1 105.9 102.1
105.1 199.3 105.2
102.9 105.0 201.1
C=
−0.3 0.5
0.6
−1.9 1.0
0.4
−1.5 2.8 −1.4
stiffness matrix C, (6 × 6, in Voigt
of Cu crystals has been computed
−0.6 −4.9 0.5
0.6
1.7
2.4
−0.5 0.3 −0.8
49.1 0.0
0.3
−1.4 46.4 −0.3
1.0
0.7 49.6
.
The matrix has been obtained from six linearly independent mechanical tests,
conducted in displacement control (three independent tensile and three independent shear tests). As it appears, the aggregate, containing 200 randomly
oriented crystals, tends to behave macroscopically as an isotropic material. It is
underlined that only a single microstructural realisation has been considered,
and increased accuracy is to be expected if ensemble averages over several
microstructural realisations are carried out. However, the previous result confirms an almost isotropic behaviour and is completely satisfactory. For further
details about the effect of the number of grains and number of realisations on the
macroscopic averages, the interested reader is referred to Ref. [29].
After the previous assessment, other homogenisation tests are carried out
and the results are presented in terms of effective Young’s and shear moduli,
computed as ensemble averages of the apparent properties over 100 realisations, for aggregates with Ng = {10, 20, 50, 100, 200} grains. In this study,
crystalline Cu, crystalline SiC and a two-phase steel have been considered
as test materials. Figure (3.12) reports the apparent Young’s and shear moduli versus Ng for copper crystal aggregates. For comparison, also values obtained with kinematic boundary conditions [29] are displayed. As expected, it
is apparent that when periodic boundary conditions are enforced, the apparent properties approach the effective ones much faster with respect to those
obtained using displacement boundary conditions. The error bars represent
the standard deviation of the homogenisation results. The dashed line, corresponding to the reference effective property, has been taken as the mean value
of the fifth order bounds reported in Ref. [70]. In Figure (3.12), the theoretical
bounds for the elastic properties of copper polycrystalline aggregates are reported [35]. More specifically, the light grey region corresponds to the upper
and lower 1st-order bounds, which are the Voigt and Reuss bounds, respectively; the dark grey region corresponds to the Hashin and Shtrikman bounds
3.5 Computational tests
87
[86, 87].
Figure (3.13) shows the apparent Young’s and shear moduli for polycrystalline SiC. The error bars represent the standard deviation of the homogenisation results, whereas the reference value of the effective macroscopic properties has been taken from Ref. [28]. Also in this case, the theoretical bounds
are reported. Again, the light grey region corresponds to the upper and lower
1st-order bounds, whereas the dark grey region corresponds to the Hashin
and Shtrikman bounds.
Finally, Figure (3.14) shows the apparent Young’s and shear moduli for a
two-phase steel alloy, namely two-phase stainless steel SAF 2507. The properties of the micro-constituents have been taken from Ref. [97], and consist of
58% austenite (γ) and 42% ferrite (α) in volume. The figures show the apparent properties of ferrite aggregates, of austenite aggregates, and of the steel
alloy. The theoretical bounds for the alloy constituents are also reported. In
this case, only the Hashin and Shtrikman bounds are reported. It can be noted
that the alloy’s apparent properties approach the known values for SAF 2507.
The slight misalignment of the curves has been intentionally added for the
sake of clarity in the figure.
150
60
Ref. [29]
present
Ref. [70]
1st-order bounds
H-S bounds
140
130
120
110
100
1st-order bounds
H-S bounds
Shear modulus [GPa]
Young’s modulus [GPa]
160
Ref. [29]
present
Ref. [70]
55
50
45
40
35
1020
50
100
200
1020
50
100
Ng
Ng
(a)
(b)
200
Figure 3.12: Apparent (a) Young’s and (b) shear moduli for polycrystalline
copper. Data collected on 100 realisations. The light grey region represents
the 1st order bounds, whereas the dark grey region represents the Hashin and
Shtrikman bounds.
88
An enhanced grain-boundary framework
200
1st-order bounds
H-S bounds
1st-order bounds
H-S bounds
present
Ref. [28]
Shear modulus [GPa]
Young’s modulus [GPa]
470
460
450
440
present
Ref. [28]
195
190
185
1020
50
100
200
1020
50
100
Ng
Ng
(a)
(b)
200
Figure 3.13: Apparent (a) Young’s and (b) shear moduli for polycrystalline
SiC. Data collected on 100 realisations. The light grey region represents the
1st order bounds, whereas the dark grey region represents the Hashin and
Shtrikman bounds.
250
105
230
α-iron
γ-iron
SAF
210
190
170
150
α-iron H-S bounds
γ-iron H-S bounds
Shear modulus [GPa]
Young’s modulus [GPa]
α-iron H-S bounds
γ-iron H-S bounds
95
α-iron
γ-iron
SAF
85
75
65
55
1020
50
100
200
1020
50
100
Ng
Ng
(a)
(b)
200
Figure 3.14: Apparent (a) Young’s and (b) shear moduli for polycrystalline
aggregates of a two-phase steel. Data collected on 100 realisations. The grey
regions represent the Hashin and Shtrikman bounds.
3.5 Computational tests
3.5.3
89
Interface modelling for micro-cracking
Prior to presenting the results of micro-cracking simulations of polycrystalline
aggregates, the explicit expression of the interface equations introduced in
Section (1.3.2) and herein used for modelling polycrystalline micro-cracking is
recalled for completeness. The ICs given in Eqs.(1.18) are particularised to be
the cohesive laws used in Ref. [28] for modeling polycrystalline inter-granular
micro-cracking.
As detailed in [28], during the loading history the generic grain boundary interface I gh , shared by the grains g and h, undergoes different states,
which are subsequently modelled using the appropriate interface equations.
At the beginning of the analysis all the interfaces are assumed to be in a pristine state and, likewise the homogenisation analyses, displacements continuity
and tractions equilibrium are enforced as in Eqs.(3.8).
The pristine state of I gh holds as long as the following inter-granular
threshold condition
τe < Tmax
(3.10)
is satisfied. In Eq.(3.10), Tmax is the interface cohesive strength and τe is an effecgh
tive traction expressed as a function of the local grain boundary traction eti as
follows
s
2
β
2
τe = hτn i +
τs ,
(3.11)
α
where τn and τs are
qthe normal and tangential tractions components defined as
gh
gh
gh
τn = et3 and τs = (et1 )2 + (et2 )2 , respectively, α and β are suitable constants
weighing the sliding and opening failure modes of the interface and h·i are the
gh
Macaulay brackets, defined as hτ i = max(0, τ ). It is recalled that eti denotes
the tractions at the interface I gh , which, due to the equilibrium conditions
g
at the interface (see Eq.(1.20)), is equal to the traction eti or the traction etih , i.e.
et gh = et g = eth .
i
i
i
As soon as the effective traction overcomes the interface strength, the interface enters the damage state and micro-cracking initiation and evolution is
modelled by replacing the displacement continuity equations (3.8a) with a cohesive law embodying suitably defined parameters representing the initiation
and evolution of damage. The cohesive law can be derived from the definition of a potential [155] or its expression can assumed a priori. The reader is
referred to Ref.[28] for the derivation of the cohesive law used in the present
90
An enhanced grain-boundary framework
thesis. Here, the following irreversible extrinsic cohesive traction-separation law
is introduced in place of Eq.(3.8a)
gh
gh
gh
gh
)δue3
gh
− et3
Ψi = Cs (d∗ )δuei − eti = 0,
(3.12a)
gh
Ψ3
(3.12b)
= Cn (d
∗
= 0,
where Eq.(3.12a) is valid for i = 1, 2. The term irreversible refers to the fact
the damage of the interface is allowed only to accumulate during the loading
history and is represented by the monotonically increasing parameter d∗ ; the
term extrinsic refers to the fact that such cohesive law is introduced only after the threshold condition ceases to hold; finally, the term traction-separation
refers to the fact the the cohesive law expresses a relation between the tracgh
gh
tions eti and the displacements jump δuei at the interface between the grain g
and h.
In Eqs.(3.12), Cs and Cn are the constitutive constants of the cohesive law
given as
αTmax 1 − d∗
,
δucr
d∗
s
Tmax 1 − d∗
,
Cn (d∗ ) =
δucr
d∗
n
Cs (d∗ ) =
(3.13a)
(3.13b)
cr
where δucr
s and δun denote the critical values of the normal and sliding displacement jumps at which the interface fails in the case of pure opening mode
or pure sliding mode, respectively. In Eqs.(3.12) and (3.13), d∗ denotes the irreversible damage parameter, which is given by the maximum value reached
by an effective opening d during the loading history Hd , i.e.
d∗ = max d ∈ [0, 1],
(3.14)
Hd
gh
where d is defined in terms of the local displacements jumps δuei at the interface I gh as
s
δun 2
δus 2
d=
+ β cr
(3.15)
δucr
δus
n
q
gh
gh
gh
being δus = (δue1 )2 + (δue2 )2 and δun = δue3 the tangential and normal
components of the displacements jump, respectively.
It is worth noting that Eqs.(3.12), in the case of very low levels of damage
d∗ , still provide consistent interface conditions; in fact, it is straightforward
3.5 Computational tests
91
gh
to see that, by multiplying Ψi by d∗ ≈ 0, Eqs.(3.12) corresponds to enforce
gh
δuei = 0, i.e. pristine interface conditions. Furthermore, in case the two grains
g and h come in contact, the continuity of the displacements along the local
gh
normal direction is enforced, i.e. Eq.(3.12b) is replaced by δue3 = 0.
Given the definition of τs and τn , it is easy to see that Eqs.(3.12) can also be
written in terms of the tangential and normal components as
τs = Cs (d∗ )δus ,
∗
τn = Cn (d )δun ,
(3.16a)
(3.16b)
and, therefore, it is possible to show that the work of separation G I in case of
pure normal mode and the work of separation G I I in pure sliding mode can
be written as
Z δucr
n
1
Tmax δucr
n,
2
0
Z δucr
s
α
GI I =
τs dδus = 2 Tmax δucr
s ,
2β
0
GI =
τn dδun =
(3.17a)
(3.17b)
respectively. Eqs.(3.17) will be used to compute the critical displacement δucr
s
in sliding mode in terms of the interface strength Tmax , the critical displacement δucr
n in normal mode, the parameters α and β and the ratio G I I /G I .
Eventually, the interface enters the failed state when d∗ reaches the critical
values d∗ = 1. In such a case, the cohesive law given in Eqs.(3.12) is replaced
gh
by the condition of zero tractions, i.e. by the equations eti = 0, i = 1, 2, 3.
gh
However, also in this case, if the two grains come in contact, δue3 = 0 is
enforced.
For a clearer representation of cohesive law used in this thesis, Figures
(3.15a) and (3.15b) show the behaviour of the tangential and normal tractionseparation law given in Eqs.(3.16a) and (3.16b), respectively. In the figures, the
cr
shaded region of the plane ( βδus /δucr
s )-( δun /δun ) corresponds to the region
d ≥ 1, i.e. the region where the interface is in the failed stated. For further
details about the derivation of the cohesive law described in this Section and
used in this thesis, the reader is referred to Ref.[28] and the references therein.
3.5.4
Microcracking simulations
In this Section, the results of some micro-cracking simulations for SiC aggregates are reported. The SiC grains bulk elastic constants are taken from Ref.
92
An enhanced grain-boundary framework
(a)
(b)
Figure 3.15: Schematic representation of the (a) tangential component τs and
(b) normal component τn of the tractions of the cohesive law used in this thesis
given in Eqs.(3.16a) and (3.16b), respectively.
[28]. The work of separation G I is
√estimated by the macroscopic values of the
fracture toughness K IC = 3 MPa m in mode I and in the case of plain strain,
of the Young modulus E = 448 GPa and of the Poisson’s ratio ν = 0.168 of
SiC by means of the relation
GI =
1 − ν2 2
K IC .
E
(3.18)
The inter-granular cohesive
√ properties are assumed to be: Tmax =cr500 MPa;
G I /G I I = 1; α = 1; β = 2. The critical displacements δucr
n and δus are then
computed using Eqs.(3.17a) and (3.17b), respectively.
Differently from what it has been done in the homogenisation tests, regularised prismatic and non-periodic microstructures are used in the cracking
computations. This is due to the circumstance that the simple implementation of Eqs.(3.9), as periodic boundary conditions for a non-prismatic unit cell,
without suitable modifications to take into account the presence of possible
displacement jumps at the RVE’s external boundary, would over-constrain
the micro-cracking, preventing the cracks from reaching the external walls of
the unit cell. To avoid this, for simplicity, prismatic unit cells subjected to
non-periodic kinematical boundary conditions are used here. This is not a
limitation of the formulation: there is no problem in using periodic boundary
conditions in a micro-cracking simulation, as shown in Ref. [30]. The only
difficulty arises if non-prismatic RVEs are used without suitably modifying
3.5 Computational tests
93
Eqs.(3.9). On the other hand, general kinematic or static boundary conditions
can be enforced on the non-prismatic cells. The results presented here serve
thus the purpose of testing the use of regularised morphologies in conjunction
with optimised meshes within the scope of microcracking simulations: it has
been actually found that the introduction of such enhancements has a remarkable effect in cutting down the microcracking simulation time to several hours,
for morphologies with few hundred grains, which represents an outstanding
gain with respect to the computation times previously reported in Ref. [28],
where similar computations required several days on similar machines.
Another points of interest concerns the mesh size of the grain-boundary
elements. For the basis formulation, this point has been carefully discussed
in Ref. [28] where it has been highlighted that the characteristic length of the
grain-boundary elements le should be fine enough to resolve the strain and
stress distributions inside the cohesive zone [173, 66, 216]. This condition,
which ensures the reproducibility of results, is always respected in the present
formulation and implementation that, as a matter of fact, smooths the representation of the inter-granular fields without altering the reference mesh size.
In the following, prismatic microstructures whose edges are aligned with
the global reference system are tested. Figure (3.16) schematises the boundary
conditions that will be used during the numerical tests: for each face, the
boundary conditions are given as prescribed displacements along the normal
of the face and written as uk nk = ūn where nk is the outward unit normal of
the face and ūn is the prescribed value of displacement. Along the remaining
directions, the external faces of the aggregate are traction-free.
A first set of tests reports the micro-cracking behaviour of aggregates with
Ng = 100 and different ASTM G values, namely ASTM G = 10, 12, 14 [205]
under tensile loading in displacement control. The boundary conditions consist of prescribed displacements over the top and bottom faces of the aggregates given by ui ni = λ/2, whereas the lateral faces of the aggregates are subject ui ni = 0. The remaining loading directions are traction-free. Figure (3.17a)
shows three 100-grain morphologies with ASTM G values G = 10, 12, 14 to
scale, whereas Figure (3.17b) shows the mesh of the ASTM G = 10 aggregate. Figure (3.18a) reports the value of the macro-stress Σ33 corresponding to
different values of enforced nominal strain Γ33 computed as Γ33 = λ/H being H the height of the aggregate. It is worth noting as different ASTM grain
sizes induce different macro-behaviours in terms of brittleness: aggregates
with bigger grains (ASTM G = 10) are generally more brittle with respect to
morphologies with smaller grains, if the inter-granular cohesive strength and
94
An enhanced grain-boundary framework
Figure 3.16: Scheme of the boundary conditions used to investigate the microcracking of polycrystalline aggregates. The boundary conditions are given as
prescribed displacement along the direction defined by the normal of each
face.
toughness are kept constant; this is consistent with the model physics. Figures
(3.18b), (3.18c) and (3.18d) show the microcracking patterns and the damage
distribution for the above three morphologies with Ng = 100 and average
ASTM grain size G = 10, 12, 14, respectively, at the last computed snapshot
reported in Figure (3.18a).
The second sets of tests investigates the effect of the number of grains on
the micro-cracking response of the aggregates. Similarly to the previous set of
tests, the top and bottom faces of the aggregates are subject to ui ni = λ/2, the
lateral faces are subject to ui ni = 0 and the remaining directions are tractionfree. Figure (3.19a) shows three morphologies with average ASTM grain size
G = 12 and number of grains Ng = 50, 100, 200 and Figure (3.19b) shows
the boundary mesh of the internal grains of the 200-grain aggregate, i.e. those
grains that non are affected by the presence of the prismatic box. The microcracking stress-strain behaviour is reported in Figure (3.20a) in which it is
interesting to note that the morphologies with Ng = 100 and Ng = 200 have
similar response in terms of averaged stress component, so that is appears
reasonable to assume a representative behaviour. Figures (3.20b), (3.20c) and
(3.20d) show the microcracking patterns and the damage distribution for the
morphologies with Ng = 50, Ng = 100 and Ng = 200, respectively, at the last
computed snapshot reported in Figure (3.20a).
A third set of tests investigates the micro-cracking behaviour of a polycrys-
3.5 Computational tests
95
(a)
(b)
Figure 3.17: (a) Relative size of three different morphologies with Ng = 100
and different ASTM grain size G = 10, 12, 14. (b) Mesh of the ASTM G = 10
aggregate and particular of the mesh of a grain.
talline aggregate subjected to loads in different directions, so to investigate the
isotropic behavior of the macro-aggregate also in micro-cracking simulations.
The investigated cubic 200-grain aggregate, with ASTM G = 12 grain-size,
and the corresponding mesh are shown in Figures (3.21a) and (3.21b), respectively. Five different sets of boundary conditions are prescribed on the aggregate, which is subject to uniaxial tensile loading along the directions x1 , x2 and
x3 , to biaxial tensile loading on the plane x1 -x2 and to triaxial tensile loading.
Similarly to the two previous sets of tests, the uniaxial boundary conditions along the direction xk are enforced as follows
(
ui ni =
λ/2,
0,
if nk = ±e pk
,
otherwise
(3.19)
where ni is the unit normal of the aggregate’s faces and e pk is the unit vector of
the k-th loading direction. It is worth noting that when the aggregate is subject to uniaxial boundary conditions along the xk direction, the corresponding
prescribed strain is given by Γ̄ = Γkk = λ/H being H the side of the aggregate.
Figures (3.22a) and (3.23a) shows the values of the macro-stress components
Σijkl versus the load factor λ. In the notation of Σijkl , the first two subscripts
ij indicate the considered homogenized stress component, whereas the subscript k indicates the loaded face, i.e. the face with e pk as unit normal, and the
96
An enhanced grain-boundary framework
600
500
ASTM G = 14
ASTM G = 12
ASTM G = 10
Σ33 [MPa]
400
300
600
200
560
100
0
0.00
520
0.12
0.04
0.08
0.14
0.16
0.12
0.16
0.20
Strain Γ33 [%]
(b) ASTM G = 10
(a)
(c) ASTM G = 12
(d) ASTM G = 14
(e)
Figure 3.18: (a) Volume average stress component Σ33 versus nominal applied
strain Γ33 . (b-d) Microcracking patterns and damage distribution for three
morphologies with Ng = 100 and average ASTM grain size (b) G = 10, (c)
G = 12 and (d) G = 14. (e) Colormap of the damage.
3.5 Computational tests
(a)
97
(b)
Figure 3.19: (a) Relative size of three different morphologies with ASTM grain
size G = 12 and different number of grains Ng = 200, 100, 50. (b) Grain
boundary mesh of those grain not affected by the presence of the prismatic
box (represented by black lines) of the 200-grain morphology.
subscript l indicates the loading direction. In particular, Figure (3.22a) shows
the values of the direct macro-stress components Σijkl , i.e. the components of
the macro-stress Σii when a prescribed displacement in the direction i is applied on the face i, which corresponds to a prescribed macro-strain component
Γii . Similarly, Figure (3.23a) shows the values of the cross macro-stress components Σijkl , that is the components of the macro-stress Σii when a prescribed
displacement in the direction j is imposed on the face j, with j 6= i. The insets
in Figures (3.22a) and (3.23a) are close-ups of the curves when the stresses
reach their maximum values. Figures (3.22b-d) show the cracked morphology when subject to the three uniaxial loading conditions, whereas Figures
(3.23b-d) show the corresponding damage distribution. It is worth noting the
correspondence between the crack pattern,the distribution of damage and the
homogenised curves. More specifically, when the morphology is loaded along
the direction x2 and x3 , it shows an almost well-defined unique crack perpendicular to the direction of loading and the corresponding homogenised curves
overlap reasonably well. On the other hand, when the morphology is loaded
along the x1 direction, it shows a more tortuous crack, which corresponds to
a slightly higher values of the homogenised stresses. However, in all three
tests the isotropic behaviour of the macro direct and cross macro components
is verified. Unlike Figures (3.18b-d) and (3.20b-d) that show all the damaged
98
An enhanced grain-boundary framework
600
500
Ng = 50
Ng = 100
Ng = 200
Σ33 [MPa]
400
300
200
100
0
0.00
0.04
0.08
0.12
0.16
0.20
Strain Γ33 [%]
(b) Ng = 50
(a)
(c) Ng = 100
(d) Ng = 200
(e)
Figure 3.20: (a) Volume average stress component Σ33 versus nominal applied strain Γ33 . (b-d) Microcracking patterns and damage distribution for
three morphologies with ASTM grain size G = 12 and number of grains (b)
Ng = 50, (c) Ng = 100 and (d) Ng = 200. (e) Colormap of the damage.
3.5 Computational tests
99
interfaces in the aggregates, in order to have a clearer representation of the
cracks, Figures (3.23b-d) display the interface whose damage is higher than
5%.
Then, a tensile biaxial and a tensile triaxial test have been performed on
the morphology shown in Figure (3.21). The biaxial has been simulated enforcing the following boundary conditions: ui ni = λ/2 on the lateral faces
of the morphology and ui ni = 0 on the top and bottom faces. On the other
hand, the triaxial has been simulated enforcing ui ni = λ/2 on all the faces
of the morphology. Figure (3.24a) shows the macro-stress components Σ11 ,
Σ22 and Σ33 versus λ, when the morphology is subject to a biaxial tensile load.
For comparison, also the macro-stress components Σ1111 and Σ1122 , taken from
corresponding the uniaxial tests are reported in the figure. Similarly, Figure
(3.24b) shows the macro-stress components Σ11 , Σ22 and Σ33 versus λ, when
the morphology of is subject to a triaxial tensile load. The macro-stress components Σ1111 and Σ1122 taken from the corresponding uniaxial test and the
macro-stress Σ11 taken from the biaxial test are reported in the figure. The
crack patterns for the biaxial and triaxial loading conditions, at the last computed values of λ, are also reported in Figures (3.24c) and (3.24d) and the corresponding damage distribution are reported in Figures (3.24e) and (3.24f). It
is interesting to note that for these last two tests, and in particular for the triaxial case, multiple cracks show during the loading history. Again, in Figures
(3.24e,f) the interfaces whose damage is higher than 5% are reported.
Finally, the microcracking response of a 1000-grain cubic polycrystalline
aggregates with ASTM grain size G = 12 is considered. The 1000-grain
morphology is generated in a cubic box by the distribution of the weighted
seeds shown in Figure (3.25a). The resulting tessellation is reported in Figure (3.25b). The aggregate is subject to uniaxial tensile boundary conditions
along the x3 , i.e. ui ni = λ/2 on the top and bottom faces and ui ni = 0 on the
lateral faces. Figure (3.26a) show the volume stress averages Σ11 , Σ22 and Σ33
as a function of the load factor λ. The crack pattern at the last computed step
of the loading history is shown in Figure (3.26b) where it is possible to see
the presence of two fully developed micro-cracks, one originating from the
bottom-right corner and the second one originating from the center of leftmost edge of the aggregate. The cracks configuration will be clearer by means
of the following figures. Figures (3.27a,c,e) show the evolution of the microcrack pattern at the different values of the load factor λ. At the same values of
λ, Figures (3.27b,d,f) show the evolutions of the damage distribution within
the aggregate. In this case, all the damaged interface are displayed and it is in-
100
An enhanced grain-boundary framework
(a)
(b)
Figure 3.21: (a) 200-grain morphology with ASTM grain size G = 12 and (b)
its boundary mesh.
teresting to note how at λ = 0.077, corresponding the load factor at which the
softening initiates, the damage is almost uniform throughout the morphology.
During the next load steps, the aggregate undergoes softening and the cracks
localize. In order to better show the shape and configuration of the cracks
within the aggregate, Figure (3.28) shows from two different views the interfaces of the aggregate whose damage is higher than 5% at the three load steps
λ = 0.077, λ = 0.079 and λ = 0.086. By looking at the figures, it is possible
to recognise the coalescence of failing interfaces into the two fully developed
cracks, one close to the bottom face and the second one at half the height of
the aggregate. Also, it is clear that the cracks initiate at the external faces and
tend to grow towards the interior of the aggregate.
3.5 Computational tests
101
600
Σijkl [MPa]
500
570
Σ1111
Σ2222
Σ3333
550
530
400
0.040
0.050
300
200
100
0
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Load factor λ
(a)
(b) Γ̄ = Γ11
(c) Γ̄ = Γ22
(d) Γ̄ = Γ33
Figure 3.22: (a) Volume average of the direct stress component Σijkl versus
applied load factor λ for the three uniaxial loading configurations. (b-d) Microcracking patterns for the three uniaxial loading configurations denoted by
Γ̄ = Γkk where k is the loading direction.
102
An enhanced grain-boundary framework
140
130
125
120
120
Σijkl [MPa]
100
0.04
0.05
80
Σ2211
Σ3311
Σ1122
Σ3322
Σ1133
Σ2233
60
40
20
0
0.00
0.02
0.04
0.06
0.08
0.10
0.12
Load factor λ
(a)
(b) Γ̄ = Γ11
(c) Γ̄ = Γ22
(d) Γ̄ = Γ33
(e)
Figure 3.23: (a) Volume average of the cross stress component Σijkl versus applied load factor λ for the three uniaxial loading configurations. (b-d) Damage
distribution for the three uniaxial loading configurations denoted by Γ̄ = Γkk
where k is the loading direction. (e) Colormap of the damage.
3.5 Computational tests
103
600
600
Σijkl [MPa]
500
400
300
200
100
0
0.00
Σ1111
Σ1122
Σ11biax
Σ11triax
Σ22triax
Σ33triax
500
Σijkl [MPa]
Σ1111
Σ1122
Σ11biax
Σ22biax
Σ33biax
400
300
200
100
0.02
0.04
0.06
0.08
0.10
0
0.00
0.12
Load factor λ
0.02
0.04
0.06
0.08
0.10
0.12
Load factor λ
(a)
(b)
(c)
(d)
(e)
(f )
(g)
Figure 3.24: Micro-cracking results of (a,c,e) biaxial test and (b,d,f ) triaxial test:
(a,b) Volume averages of the stress components Σ11 , Σ22 and Σ33 versus applied load factor λ; (c,d) micro-crack pattern; (e,f ) Damage distribution. (g)
Colormap of the damage.
104
An enhanced grain-boundary framework
(a)
(b)
Figure 3.25: 1000-grain morphology with ASTM grain size G = 12: (a) seeds
and weights distribution inside the cubic box; (b) resulting polycrystalline aggregate.
600
500
Σ33
Σ11
Σ22
Σij [MPa]
400
300
200
100
0
0.00
0.02
0.04
0.06
0.08
0.10
Load factor λ
(a)
(b)
Figure 3.26: (a) Volume stress averages Σ11 , Σ22 and Σ33 as a function of the
load factor λ. (b) Micro-crack pattern of the 1000-grain morphology subject to
tensile load at the last computer step.
3.5 Computational tests
105
(a) λ = 0.077
(b) λ = 0.077
(c) λ = 0.079
(d) λ = 0.079
(e) λ = 0.086
(f ) λ = 0.086
(g)
Figure 3.27: Microcracking snapshots at the loading steps (a,b) λ = 0.077,
(c,d) λ = 0.079 and (e,f ) λ = 0.086. (a,c,e) Cracked aggregate. (b,d,f ) Damage
distribution. (g) Colormap of the damage.
106
An enhanced grain-boundary framework
(a) λ = 0.077
(b) λ = 0.077
(c) λ = 0.079
(d) λ = 0.079
(e) λ = 0.086
(f ) λ = 0.086
(g)
Figure 3.28: Two different views of the microcracks pattern at the loading
steps (a,b) λ = 0.077, (c,d) λ = 0.079 and (e,f ) λ = 0.086. (g) Colormap of the
damage.
3.6 Discussion
3.6
107
Discussion
In this Chapter of the present thesis, an effective computational tool for the microstructural analysis of polycrystalline materials, with applications in computational homogenisation and inter-granular micro-cracking simulations, has
been presented. The key contribution of the present work relies in the use
of the grain-boundary formulation in conjunction with regularized Laguerre
morphologies and a specific grain-boundary meshing strategy. In particular:
• The employment of the tessellation regularization [167] has allowed to
remove the sources of unnecessary mesh refinements in the original
morphologies, thus reducing the overall number of degrees of freedom
necessary to model the crystal aggregate. The effect of this regularization is particularly appreciated in microcracking simulations: in the
original scheme [28], the occurrence of many small grain-boundary elements hindered the computational performance of the technique and
could represent an issue for numerical convergence, due to the concentration of collocation nodes on excessively small grain faces;
• The employment of a specific grain boundary meshing strategy, using
continuous and semi-discontinuous boundary elements, has simultaneously allowed: a) to smooth the representation of the inter-granular displacement and traction fields, with advantages in terms of numerical
convergence of the non-linear solution scheme; b) to reduce the number
of inter-granular unknowns, i.e. degrees of freedom; c) to avoid the typical complexities arising, within a boundary element framework, when
grid nodes are collocated over edges or corners, where a unit normal
cannot be uniquely defined [76].
Additionally, for computational homogenization applications, non-prismatic
RVEs have been adopted, with the subsequent further reduction of meshing
complexities potentially induced by the cutting operations usually performed
to obtain periodic prismatic morphologies. The above enhancements have induced a remarkable acceleration in the numerical solution strategy, retaining
the accuracy of the original formulation [28], which is particularly welcome if
the technique is to be used in multiscale perspective [30].
However, some numerical/computational aspects deserve further study.
First of all, although the reduction of the system order, in terms of DoFs, undoubtedly represents a benefit in any numerical technique, it should be noted
that it is not the only factor affecting the computational effectiveness. In this
108
An enhanced grain-boundary framework
sense, the features of the resolving system, e.g. the sparsity pattern, the symmetry and the definiteness of the coefficient matrices, play a crucial role. The
coefficient matrix produced by the proposed formulation is not symmetric.
For such reasons, the general purpose parallel direct sparse solver PARDISO
has been used in the present study, as described in Chapter (1). However,
the development of specific iterative Krylov solvers for sparse systems having
the structure of the system described in Section (1.4.1), [11, 10], in conjunction
with hierarchical matrices [31, 34], would probably produce computational
savings, both in terms of storage memory and solution time. To improve the
numerical convergence under general loading conditions, the implementation
of arc-length procedures [50, 131] could be of interest.
Chapter 4
Modelling inter-granular and
trans-granular cracking in
polycrystalline materials
In this Chapter, a novel grain boundary formulation for inter-granular and
trans-granular cracking of three-dimensional polycrystalline materials is presented. The formulation is based on the displacement and stress boundary
integral equations introduced in Chapter (1) and still retains the advantage of
expressing the considered polycrystalline problem in terms of grain boundary variables only. Trans-granular cracking is assumed to occur along specific
cleavage planes, whose direction depends on the crystallographic orientation
of the grains of the aggregate. Both inter- and trans-granular cracking are
modelled using the cohesive zone approach. However, unlike grain boundary cracking for which the failure interface is well defined, trans-granular
cracks are introduced only if a suitably defined effective stress over the potential cleavage plane overcomes the threshold strength for that plane. During the load history, whenever a trans-granular crack needs to be introduced,
the polycrystalline morphology is remeshed accordingly and a cohesive law,
whose parameters characterise the behaviour of trans-granular failure, is associated to the newly introduced interface.
Some preliminary results show that the method is able to capture the competition between inter-granular and trans-granular cracking.
109
Modelling inter-granular and trans-granular cracking in polycrystalline
110
materials
4.1
Introduction
The present Chapter presents a model for studying the competition between
inter-granular and trans-granular cracking modes, which represent two of
the main mechanisms of failure in brittle polycrystalline materials. Intergranular cracking denotes the failure of the interfaces between adjacent grains
whereas trans-granular cracking denotes the failure, often along specific crystallographic planes, of the bulk grains. The occurrence of the two cracking modes is influenced by several factors such as crystallographic lattice,
temperature and aggressive environment [8, 51]. The crystallographic lattice
plays a key role in determining to which deformation mechanism a polycrystalline materials is susceptible. As an example, body-centered cubic (BCC)
and hexagonal close packed (HCP) crystals usually show ductile-to-brittle
transition [203, 94, 95] with decreasing temperature and inter-granular and
trans-granular cracking phenomena are the result of the limited number of
slip systems in these crystal lattices at low temperature. On the other hand,
although face-centered cubic (FCC) lattices generally favour ductile deformation over a wide range of temperatures as a consequence of the large number of slip systems, aggressive environments are well-known to induce grain
boundary embrittlement and therefore favour inter-granular and trans-granular
fracture in naturally ductile materials [133, 139, 169]. At room temperature,
HPC ceramics such as 6H silicon carbide (SiC) also exhibit inter- and transgranular brittle fracture [71].
Several studies, both experimental and numerical, have been devoted to
understanding the complex interaction between inter- and trans-granular cracking and their relationship with the morphological, chemical and physical properties of the micro-structure of polycrystalline materials. Sukumar et al. [207]
presented a extended finite element (XFEM) approach to model the interand trans-granular behaviour of 2D brittle polycrystalline microstructures. In
their work, the authors assumed a linear elastic isotropic behaviour of the
crystals and studied the micro-cracking response of the polycrystalline morphologies and the corresponding crack patterns as a function of the ratio between the fracture toughness of the grain boundaries and the fracture toughness of the grain interiors. Within the XFEM framework, Wang et al. [221] investigated the effect of the average grain diameter, of a secondary phase and
of the grain boundary fraction on the the fracture toughness of 2D polycrystalline aggregates, and recently, Prakash et al. [166] studied the effect of the
grain boundary strength on the crack propagation in polycrystalline tungsten
4.1 Introduction
111
(W) using XFEM.
Verhoosel and Gutiérrez [218] modelled the inter-granular and trans-granular
crack propagation in 2D piezoelectric polycrystals. The model was developed in a finite element framework and crack propagation was obtained using interface elements and two different cohesive laws for modelling the grain
boundary inter-granular failure and the bulk crystals trans-granular cracking.
Musienko and Cailletaud [145] used a rate-dependent finite element framework to model inter- and trans-granular stress corrosion cracking of 2D and
3D Zircaloy polycrystalline morphologies in an iodine environment. However, trans-granular crack propagation was allowed only for 2D morphologies. For inter-granular crack propagation, the authors used grain boundary finite elements with a specific constitutive relations involving a damage
parameter whose evolution was governed by a rate-dependent power law
flow rule. Trans-granular cracking was allowed for a specific crystallographic
plane of the Zircaloy crystal, namely the basal plane, which failed to cleavage
when the normal stress reached a threshold value.
Kraft and Molinari [107] developed a 2D finite element model to study
the effect of grain boundaries distribution on the inter- and trans-granular
fracture of polycrystalline aluminum oxide (Al2 O3 ). The authors employed a
cohesive zone approach to model tensile and shear separation of grain boundaries as well as cleavage cracks. Unlike the aforementioned studies, transgranular fracture was considered in the model by inserting cohesive interfaces along specific crystallographic planes within the crystals and by remeshing the morphology accordingly. New trans-granular planes were introduced
during the analyses as soon as the values of the resolved normal and shear
stress over those plane overcame their corresponding critical values, which
were estimated from the fracture toughness of the trans-grnular interface.
Inter- and trans-granular dynamic cracking in 2D polycrystalline materials
were recently investigated by Mousavi et al. [140, 141] and by Lin et al. [123]
using the cohesive zone approach within the finite element framework.
More recently, models based on different approaches such as peridynamics or nonlocal lattice particle method have been presented. Ghajari et al. [75]
proposed a peridynamic model to study the fracture propagation in 2D Al2 O3
polycrystalline aggregates accounting for the random orientation and therefore anisotropy of the microstructure. They reported the crack patterns as
a function of the ratio between the fracture toughness of the grains boundaries and the grain interiors showing a transition from inter-granular to transgranular behaviour corresponding to weaker to tougher grain boundaries.
Modelling inter-granular and trans-granular cracking in polycrystalline
112
materials
Chen et al. [46] studied the fracture behaviour of 2D polycrystalline morphologies using a nonlocal lattice particle framework. In their model, all the
crystallites belonging to a grain were rotated to match the crystallographic
orientation of the grain and grain boundaries were naturally generated at the
location where two grains meet. Crack propagation was obtained by considering a spring-based model among the particles and inter-granular to transgranular cracking mode was realised by changing the strength of the grain
boundary springs parameters.
While several two-dimensional models accounting for the interplay or
competition between inter- and trans-granular polycrystalline cracking mechanisms are present in the literature, few 3D models have been developed, due
to increased geometrical and mechanical complexity and much higher computational requirements. However, the combined effect of inter- and transgranular cracking is a naturally three-dimensional phenomenon due to the
crucial role of the 3D crystallographic orientation and therefore of the potential cleavage planes within polycrystalline aggregates. Furthermore, as
pointed out by some researchers [203, 94, 95], 2D model may not be able to
fully capture the role of grain boundary fracture in accommodating transgranular crack propagation through misaligned adjacent grains.
Three-dimensional polycrystalline morphologies were studied by Clayton
and Knap [48] by means of a phase field approach. The authors studied silicon
carbide (SiC) and zinc (Zn) polycrystalline aggregates taking into account the
anisotropic linear elastic behaviour of the crystals as well as the fracture behaviour of the trans-granular failure, which occurred along the basal planes of
the considered HCP crystals. By varying the ratio between the elastic stiffness
and the surface energy of the grains with respect to the grain boundaries, they
studied inter-granular and trans-granular crack patterns within the considered morphologies. Recently [49], they also incorporated deformation twinning in their phase field model. Abdollahi and Arias [1, 2] employed the phase
field model to study the fracture and toughening mechanisms in ferroelectric
polycrystalline microstructures. They studied 2D [1] and 3D [2] morphologies and modelled the inter-granular and trans-granular fracture behaviour
by considering a specific ratio of the grain interior and grain boundaries fracture toughnesses. Within the phase-field framework, Shanthraj et al. [192] developed an elasto-viscoplastic phase field model to capture anisotropic cleavage failure in 2D polycrystalline aggregates. Three-dimensional modelling of
cleavage in polycrystals has been also proposed by Shterenlikht and Margetts
[197] and Di Caprio et al. [58] using a cellular automata approach.
4.2 Grain boundary formulation
113
In this Chapter, inter- and trans-granular crack evolution in three-dimensional
polycrystalline morphologies is modelled using a boundary element and a
cohesive zone approach. The Chapter is organised as follows: Section (4.2)
briefly recalls the grain boundary formulation of Chapter (1) focusing on the
displacements and stress boundary integral equations and the cohesive zone
modelling of inter- and trans-granular failure; Section (4.3) describes the numerical discretisation of the displacement boundary integral equations as well
as the stress boundary integral equations and presents the numerical algorithm accounting for the two considered failure mechanisms and involving
the on-the-fly remeshing of the polycrystalline morphology; Section (4.4) shows
some preliminary results of the developed formulation.
4.2
Grain boundary formulation
In this Section, the grain boundary formulation described in Chapter (1) is
briefly recalled.
4.2.1
Boundary integral equations
To model inter-granular and trans-granular cracking in 3D polycrystalline
materials, the displacement boundary integral equations (1.13) and the stress
boundary integral equations (1.12) are simultaneously used during the analysis. They are recalled here for the sake of completeness. The displacement
boundary integral equations are used to express the behaviour of the generic
g
grain g in terms of the grain boundary values of displacements uei (x) and
g
tractions eti (x) and read
Z
g
e g (x, y)ueg (x)dS(x) =
cepi (y)uei (y) + − T
pi
i
Z
Sg
Sg
e g (x, y)et g (x)dS(x)
U
pi
i
(4.1)
where x and y are the integration and collocation points, respectively, S g is the
g
e g (x, y) and T
e g (x, y)
boundary of the grain g, cepi (y)uei (y) are the free-terms, U
pi
pi
are boundary equations kernels and it is recalled that the symbol e· denotes a
quantity expressed in the boundary local reference system.
On the other hand, the stress boundary integral equations, as given in
g
Eq.(1.12), allow to compute the stress tensor σmn (y) at any internal point y ∈
V g of the grain g in terms of the grain boundary displacements and tractions
as follows
g
σmn (y) +
Z
σ,g
Sg
g
Tmni (x, y)ui (x)dS(x) =
Z
σ,g
Sg
g
Umni (x, y)ti (x)dS(x),
(4.2)
Modelling inter-granular and trans-granular cracking in polycrystalline
114
materials
σ,g
σ,g
where once again x is the integration point and Umni (x, y) and Tmni (x, y) are
the kernels of the integral equations. In this case, there is no need to express
the equation in the local boundary reference system since the stress boundary
integral equations are used in a post-processing stage, that is after the solution
has been obtained. The reader is referred to Section (1.3) or Chapter (1) for a
more detailed description of terms in Eqs.(4.1) and (4.2).
4.2.2
Boundary conditions and interface equations
To close the polycrystalline problem, a proper set of boundary conditions and
interface conditions must be enforced. The boundary conditions are enforced
on the external faces of the aggregate whereas the interface conditions express the relation between the inter-granular displacements and tractions as
described in Sections (1.3.1) and (1.3.2).
In Chapter (3), inter-granular cracking was modelled using a suitably defined traction-separation cohesive law, which accounted for the degradation
and failure of the grain boundaries. In this work, although in a different way
with respect to inter-granular failure that will be described in Section (4.2.3),
trans-granular cracking is also modelled using the cohesive-zone approach.
First, the cohesive zone model is recalled.
As discussed in Section (3.5.3), damage arises over certain interfaces, which
for the purpose of this study may represent inter-granular regions as well as
cleavage planes, when a suitably defined effective traction τe overcomes the interface strength Tmax . When such criterion is fulfilled for a point of the generic
couple of adjacent grains g and h sharing the interface I gh , an extrinsic cohesive law of the form
gh
gh
gh
)δue3
gh
− et3
Cs (d∗ )δuei − eti = 0,
Cn (d
∗
= 0,
(i = 1, 2)
(4.3a)
(4.3b)
gh
is introduced to link the boundary traction components eti with the boundary
gh
displacements jump δuei . It is worth noting that if the interface I gh represent
a trans-granular crack, the grains g and h denote those grains originating from
the trans-granular failure of a single grain. The cohesive law (4.3) is given in
terms of the constitutive constants Cs (d∗ ) and Cn (d∗ ), which are function of
the irreversible damage parameter d∗ defined as maximum value of the effective
opening displacement d during the loading history Hd , see Eqs.(3.13), (3.14) and
(3.15).
4.2 Grain boundary formulation
115
Subsequently, upon assuming the work of separation G I , the interface
strength Tmax , and the relative contribution between opening and sliding failure modes represented by the parameters α, β and the ratio G I I /G I , it is possible to completely determine the cohesive law.
To account for inter-granular and trans-granular failures, two different cohesive laws are employed in the present study and their parameters are set
to represent the different properties of the two types of interface as well as
the competition between the two mechanisms. Although inter- and transgranular processes are modelled using the same cohesive approach, it is worth
underlining their different nature and the consequence in terms of numerical
implementation: inter-granular failures occur along the interfaces defined by
the grain boundaries, which are natural sites of crack initiation and are welldefined when the polycrystalline morphology is generated; on the other hand,
at a generic point within the bulk crystallographic lattice of a crystal, transgranular failure may potentially occur along different planes, which are not a
priori known and may or may not activate depending on the fulfilment of a
threshold condition involving the comparison between an actual local stress
acting over the plane and the corresponding strength. As an example, Figure
(4.1a) shows a grain boundary interface between two adjacent grains, whereas
Figure (4.1b) shows an example of a potential trans-granular interface identified by a crystallographic plane described by the unit normal vector mi and
passing through the point y within a bulk grain. In terms of numerical implementation, the trans-granular failure is addressed by considering the grain
as a bulk homogeneous solid and introducing a cohesive interface along the
activated cleavage plane whenever the trans-granular threshold condition is
met.
4.2.3
Inter- and trans-granular fracture modes
In this Section, the parameters entering the cohesive laws employed to describe the two failure mechanisms are discussed. As regards the notation of
this Section, the grain boundaries quantities referring to the interface between
the adjacent grains g and h are denoted by the superscript gh since they in fact
characterise the behaviour of the interface I gh between the grains g and h and
can in general vary from interface to interface. On the other hand, the quantities characterising the behaviour of the cleavage planes of the grain g are
denoted by the superscript g only since they refer to the properties of the bulk
grain g.
Modelling inter-granular and trans-granular cracking in polycrystalline
116
materials
(a)
(b)
Figure 4.1: (a) Grain boundary (in darker green) between two adjacent grains
and (b) potential cleavage plane (in darker green) defined by the normal vector mi and passing through the point y.
In Section (3.5.3), it was shown that, at the interfaces between adjacent
grains, the effective traction of the grain boundaries at the interface I gh , degh
noted by τe , is computed in terms of the local grain boundary tractions as
s
2
D E2 β gh
gh
gh
gh
τe =
τs
,
(4.4)
τn
+
α gh
where β gh and α gh are the values of the cohesive law coefficients introduced in
Section (3.5.3) and characterising the cohesive behaviour of the grain boundgh
ary. τsq denotes the traction along the sliding direction that is computed as
gh
gh
gh
(et1 )2 + (et2 )2 ; τn denotes the traction along the normal direction and
gh
gh
g
is simply τn = et3 . It is recalled that the tractions eti = eti = etih are obtained as
solution of the displacements BIE (4.1) written for each grain of the aggregate
and coupled to the interface equations, whose numerical implementation is
described in Section (4.3).
In this case, it is highlighted that the interfaces between different grains
are known when the polycrystalline tessellation is generated and are retained
during the loading history even if the grain boundaries do not undergo intergranular cracking. In fact, if the grain boundary between the generic grains
gh
g and h is in a pristine state, which holds as long as effective traction τe
τs =
4.2 Grain boundary formulation
117
gh
is smaller the grain boundary strength Tmax , the interface equations simply
gh
enforce a zero displacements jump, i.e. δuei = 0.
g
Similarly to the inter-granular fracture mode, an effective traction τe of the
trans-granular fracture mode can be defined as follows
s
g 2
β g
g
g 2
τe =
τn +
τs .
(4.5)
αg
However, unlike the inter-granular case where the interface is already wellg
defined, given a point within the considered grain, the sliding traction τs and
g
g
the normal traction τn must be computed as a function of the stress tensor σij
and all the potential cleavage planes of the grain g at that point. It is clear that
the crystallographic lattice of the grains plays a crucial role in the determination of the orientation of the trans-granular cracks. Using the stress BIE (4.2),
g
the stress tensor σij at any internal point of the grain g can be computed as a
function of the boundary displacements and tractions. Then, given the unit
g
normal mi representing the potential cleavage plane of the grain g and two
g
g
g
mutually orthogonal directions pi and qi on the plane, the normal traction τn
g
g g g
is computed as τn = mi σij m j , whereas the sliding traction τs is computed as
q
g
g
g
g
g g g
g
g g g
τs = (τp )2 + (τq )2 being τp = pi σij m j and τq = qi σij m j . During the loadg
ing history, the values of the effective stress τe is computed for each potential
g
g
cleavage plane. If τe overcomes the cleavage plane strength Tmax of the considered plane, a trans-granular cohesive interface is introduced into the grain as
a flat surface extending up to the boundaries of the grain. This assumption
is justified by experimental observations of trans-granular cracking in several
classes of metallics [175, 109, 95, 98] and ceramics materials [96].
Inter- and trans-granular fracture competition
The competition between the inter- and trans-granular modes of failure in
polycrystalline materials is modelled by considering suitable sets of parameters entering the corresponding cohesive laws. Here, it is assumed that the
coefficients α, β and G I I /G I do not differ between the two mechanisms, i.e.
gh
gh
g
g
α gh = α g = α, β gh = β g = β and G I I /G I = G I I /G I ∀ g = 1, . . . , Ng and
∀ gh = 1, . . . , Ni , being Ng the number of grains and Ni the number of grain
boundary interfaces. On the other hand, different ratios between the fracgh
g
ture energy G I in mode I of the grain boundaries and the fracture energy G I
Modelling inter-granular and trans-granular cracking in polycrystalline
118
materials
(a)
(b)
Figure 4.2: Schematic representation of (a) the tangential and (b) normal
components of the two cohesive laws used in this work model inter- and
trans-granular cracking. The red surfaces represent a reference cohesive law
whereas the blu surfaces represent a scaled cohesive law with γG < 1.
in mode I of the cleavage planes are considered. More specifically, considering the relation G I = Tmax δucr
n /2 (see Eq.(3.17a)), it is easy to see that the
g
gb
ratio γG ≡ G I /G I between the two aforementioned fracture energies can be
modified by changing Tmax and/or δucr
n . In this thesis, to scale the fracture eng
g
gh
ergy G I of the factor γG , i.e. G I = γG G I , it is assumed that both the interface
strength Tmax and the critical displacement δucr
n in mode I are scaled by a same
√
√
g
gh
cr,g
cr,gh
quantity, i.e. Tmax = γG Tmax and δun = γG δun . By the same fashion,
it is easy to see that the the critical displacement δucr
s in mode II is scaled by
√
√
cr,g
cr,gh
γG , i.e. δus = γG δus .
The effect of the parameter γG on the cohesive law is represented in Figures (4.2). Figure (4.2a) represents the tangential component τs as a function of
cr
( βδus /δucr
s ) and ( δun /δun ) according to Eq.(3.16) whereas Figure (4.2b) repcr
resents the normal component τn as a function of ( βδus /δucr
s ) and ( δun /δun )
according to Eq.(3.16). In the figures, the red surfaces represent a reference cohesive law, e.g. associated to the behaviour of the grain boundaries, whereas
the blue surfaces represent a scaled cohesive law with γG < 1.
4.3 Numerical discretisation and solution
4.3
119
Numerical discretisation and solution
The solution of the polycrystalline problems of inter- and trans-granular microcracking is obtained by numerically discretising Eqs.(4.1) and (4.2). The
boundary S g of the generic grain g is divided into non-overlapping elements
according to the meshing strategy developed in Chapter (3), where combined
continuous and semi-discontinuous triangular and quadrangular elements
were used to reduce the computational cost of the polycrystalline problems.
The mesh size of the surface mesh is chosen such that the average element
length le is much smaller than characteristic length Lcz of the cohesive zone,
which can be estimated in terms of the material fracture toughness and the
interface strength [173, 66, 216].
The discretised version of the displacement boundary integral equations
(4.1) is computed using the discretisation technique and the meshing strategy
combining triangular and quadrangular elements described in Section (3.4).
For the generic grain g, a linear system of equations of the form H g U g = G g T g
is obtained. The equations of the entire polycrystalline aggregate are then
retrieved by writing such equations for each grain of the aggregate and by
coupling them with the interface equations as discussed in Section (1.4). A
system of equations of the form M(X, λ) = 0 is obtained, where X collects the
unknowns of the system and λ is the load factor.
Unlike the displacement BIE, which are evaluated on the boundary of the
grains, the stress boundary integral equations (4.2) are evaluated at selected
control points within the grain. The discretised version of Eq.(4.2) is then obtained as follows
Σ g = Gσ,g T g − Hσ,g U g ,
(4.6)
where Σ g contain the components of the stress tensor evaluated at the internal
control points and the matrices Hσ,g and Gσ,g stem from the integration of the
σ,g
σ,g
kernels Tmni (x, y) and Umni (x, y), respectively, over the boundary S g . It is
g
worth noting that, if nv is the number of control points for the grain g and
the stress tensor is represented in Voigt notation, the matrices Hσ,g and Gσ,g
g
g
g
have 6nc rows and 3ns columns, being ns the number of collocation points of
grain g. As already mentioned, Eq.(4.6) are used in a post-processing stage
and therefore the reordering due the boundary conditions enforced on the
aggregate is not introduced.
In the next Section, the use of the stress equations (4.6) in combination
with the solution of the polycrystalline system of equations is represented by
a flow chart describing the algorithm for inter- and trans-granular cracking.
Modelling inter-granular and trans-granular cracking in polycrystalline
120
materials
4.3.1
Algorithm for inter- and trans-granular cracking
Figure (4.4) schematises the flow chart of the algorithm used in the present
work to address inter- and trans-granular cracking in polycrystalline materials. At the beginning of the analysis, the load factor λ is initialised to 0, the
load step counter n is initialised to 1 and the remesh boolean variable is initialised to .false.. The polycrystalline morphology is then generated and
the properties of the constituent grains and grain boundaries are loaded. The
algorithm then enters the incremental loop that is governed by the counter n
and can be described by the following steps:
1. At the beginning of the load step n, the load factor is incremented by
the load factor increment ∆λn chosen based on the convergence speed
at the previous load step [28].
2. If the analysis has just started, i.e. n = 1, or the remesh is .true., the
aggregate is discretised, the boundary element matrices are computed,
the overall element system is assembled and the remesh variable is set
to .false..
3. At the n-th load step, the system of equation M(Xn , λn ) = 0 is to be
solved. The equilibrium solution Xn is obtained by employing the NewtonRaphson algorithm and the solver PARDISO [114, 185, 186] as described
in Section (1.4). During the Newton-Raphson search for the solution,
the consistency of the grains interfaces is checked according to the procedure described in Ref. [28]
g
4. The values of the stress tensor σmn at the control points of each grain of
the aggregate is computed by means of Eq.(4.6) and the possible occurrence of trans-granular cracking is then checked by comparing the local
g
g
effective cleavage stress τe with the local threshold Tmax .
5. If the cleavage threshold is overcome, new cohesive interfaces are introduced along the cleavage planes of those grains undergoing transgranular cracking. The cohesive properties of the newly introduced interfaces are those corresponding to trans-granular cracks and are in general different from the cohesive properties of the grain boundaries. At
this point, the morphology has been modified and the boolean variable
remesh is set to .true.. As an example, Figure (4.3a) shows the boundary mesh of a grain whose threshold condition for trans-granular cracking is fulfilled, Figure (4.3b) shows the cleavage plane (in darker green)
4.3 Numerical discretisation and solution
(a)
(b)
121
(c)
Figure 4.3: (a) Boundary mesh of a grain at the point of undergoing cleavage
cracking; (b) cleavage plane (in darker green) that needs to be introduced into
the grain; (c) mesh of the two child grains originated from the grain in figure
(a).
that needs to be introduced into the grain and Figure (4.3c) shows the
remesh of the two child grains originated from the cut of the grain in
Figure (4.3a). The flow goes back to step 2 in order to find the equilibrium solution of the new aggregate for the same load factor λn .
6. If no trans-granular cracks are introduced, the current load factor λn
is compared to the final load factor λ f and either the next load step is
considered or the analysis is ended.
Modelling inter-granular and trans-granular cracking in polycrystalline
122
materials
Start
λ0 ← 0, n ← 1, remesh ← .false.
Initial morphology:
• load polycrystalline morphology;
• load grains and grain boundaries properties;
Update load factor: λn ← λn−1 + ∆λn
yes
(n = 1) ∨ (remesh)
no
• perform grains cut;
• remesh ← .true.;
• compute the mesh;
• compute the boundary element matrices;
• enforce BCs and ICs;
• assemble the overall system;
• remesh ← .false.;
Solve system: M(Xn , λn ) = 0
Compute internal stress: Eq.(4.6)
yes
trans-granular cracks?
no
n ← n+1
no
λn ≥ λ f
yes
End
Figure 4.4: Flow chart of the algorithm for inter- and trans-granular cracking.
4.4 Numerical tests
4.4
123
Numerical tests
In this tests, polycrystalline SiC aggregates with hexagonal crystal lattice are
considered. For hexagonal 6H SiC polytypes, the preferred cleavage plane is
the basal plane [48] identified by (0001) Miller indices. The elastic properties
of the crystals and the cohesive properties of the grain boundaries and the
cleavage planes are listed in Table (4.1).
In this thesis, only preliminary results of the formulation are presented.
Figure (4.5a) shows a 10-grain polycrystalline aggregate with ASTM grain
size G = 12. In Figure (4.5a), the small reference systems inside each grain
represents the orientation of the crystallographic lattice where the normal to
the basal plane is indicated by the longer arrow. For simplicity, in the following tests, the threshold condition for activating the trans-granular cracking is
checked at the centroid of each grain of the aggregate.
Consistently with the notation of Section (3.5.4), the aggregate is loaded
along the x3 direction by means of uniaxial boundary conditions prescribed
as ui ni = λ/2 over the top and bottom faces and ui ni = 0 over the lateral faces,
whereas the remaining loading directions are kept traction-free. Two values
gb
g
of the ratio γG = G I /G I are considered for these analyses, namely γG = 1
and γG = 1/4. Figure (4.5b) shows the macro-stress response Σ33 as a function
of the load factor λ corresponding to the two values of γG . As expected, the
response corresponding to a lower fracture energy of the cleavage planes, that
is γG = 1/4, is characterised by a smaller value of the maximum macro-stress
reached during the loading history.
The effect of the ratio γG on the micro-cracking pattern and on the damage
distribution is reported in Figures (4.6) and (4.7), respectively. Figures (4.6a),
(4.6c) and (4.6e) report the micro-cracking pattern at the load steps λ = 0.009,
λ = 0.011 and λ = 0.025, respectively, corresponding to γG = 1; in this case,
since the two mechanisms have the same fracture energy and the cleavage
planes are not favourably oriented with respect to the loading direction, the
crack path occur along the grain boundaries of the aggregate. On the other
hand, as shown in Figures (4.6b), (4.6d) and (4.6f ), by reducing the fracture
energy of the cleavage planes, that is by choosing a value γG = 1/4, the cleavage planes start to fail in correspondence to a lower value of the load factor,
for which the stress on the grain boundaries is not high enough to initiate the
damage. However, the three-dimensional random orientation of the grains
does not allow a perfect propagation of the trans-granular crack within the
aggregate and, as a consequence, partial grain boundaries failure is necessary
Modelling inter-granular and trans-granular cracking in polycrystalline
124
materials
to accommodate the crack propagation and the final failure of the aggregate.
Similarly, Figures (4.7a), (4.7c) and (4.7e) show the inter-granular damage distribution at the load steps λ = 0.009, λ = 0.011 and λ = 0.025, respectively,
corresponding to γG = 1. Figures (4.7b), (4.7d) and (4.7f ) show the damage
distribution corresponding to γG = 1/4 at the same load steps. By looking
at Figures (4.7a) and (4.7b), it is clear that the damage over a cleavage plane
initiates in correspondence of a much lower value of the load factor, and, due
to the local softening of failing cleavage plane, the corresponding stress redistribution and the three-dimensional random orientation of the grains, the
crack path for this choice of γG is characterised by inter- and trans-granular
failure.
Table 4.1: Elastic and cohesive properties of SiC polycrystalline aggregates.
domain
Bulk crystals
property
elastic constants
[109 N/m2 ]
Grain
boundaries
interface strength [MPa]
Cleavage
planes
cohesive law constants [-]
critical displacements
jumps [µm]
interface strength [MPa]
cohesive law constants [-]
critical displacements
jumps [µm]
component
c1111 , c2222
c3333
c1122
c1133 , c2233
c2323 , c1313
c1212
gh
Tmax
α gh
β gh
cr,gh
δun
cr,gh
δus
g
Tmax
αg
βg
cr,g
δun
cr,g
δus
value
502
565
95
96
169
(c1111 − c1122 )/2
500
1√
2
7.8089 · 10−2
1.5618 · 10−1
√
gh
γG · Tmax
α gh
β gh
√
cr,gh
γG · δun
√
cr,gh
γG · δus
4.5 Discussion and future developments
125
600
γG = 1
γG = 1/4
500
Σ33 [MPa]
400
300
200
100
0
0
0.005
0.010
0.015
0.020
0.025
Load factor, λ
(a)
(b)
Figure 4.5: (a) 10-grain polycrystalline morphology with ASTM grain size G =
12; (b) Volume stress averages Σ33 as a function of the load factor λ for the two
g
gh
g
values of the considered ratio γG = G I /G I between the fracture energy G I
gh
of the cleavage planes and the fracture energy G I of the grain boundaries.
4.5
Discussion and future developments
In this Chapter, a numerical scheme for addressing the competition between
inter- and trans-granular crack propagation has been developed and tested.
Although the numerical tests involved a polycrystalline aggregate with a rather
small number of constituent grains and subject to one type of boundary conditions only, the results showed that the developed formulation is able to capture the behaviour of polycrystalline aggregates when the two mechanisms
are active. The next step of this formulation is to consider polycrystalline
morphologies with a much larger number of grains in order to avoid the
boundary effects of the finiteness of the container and to investigate the coupling between the statistical variability of polycrystalline micro-structure and
the competition between the two mechanisms. Furthermore, different sets of
boundary conditions can also be tested to study the polycrystalline response
along different directions and eventually how the existence of a representative volume element can be affected by the simultaneous presence of the two
mechanisms.
Modelling inter-granular and trans-granular cracking in polycrystalline
126
materials
(a) λ = 0.009
(b) λ = 0.009
(c) λ = 0.011
(d) λ = 0.011
(e) λ = 0.025
(f ) λ = 0.025
Figure 4.6: Micro-cracking snapshots at the loading steps (a,b) λ = 0.009, (c,d)
λ = 0.011 and (e,f ) λ = 0.025. (a,c,e) γG = 1. (b,d,f ) γG = 1/4.
4.5 Discussion and future developments
127
(a) λ = 0.009
(b) λ = 0.009
(c) λ = 0.011
(d) λ = 0.011
(e) λ = 0.025
(f ) λ = 0.025
(g)
Figure 4.7: Micro-cracking snapshots of the damage distribution at the loading steps (a,b) λ = 0.009, (c,d) λ = 0.011 and (e,f ) λ = 0.025. (a,c,e) γG = 1.
(b,d,f ) γG = 1/4. (g) Colormap of the damage.
Modelling inter-granular and trans-granular cracking in polycrystalline
128
materials
Chapter 5
Grain boundary formulation for
crystal plasticity
In this Chapter, a three-dimensional grain-boundary formulation for smallstrain crystal plasticity in polycrystalline aggregates is presented. The method
is developed starting from single grain crystal plasticity equations and implemented in polycrystalline aggregates [33, 32, 82]. It is based on the use
of a suitable set of boundary integral equations for modelling the individual grains, which are represented as anisotropic elasto-plastic domains. In
the boundary integral framework, crystal plasticity is modelled resorting to
an initial strains approach, and specific aspects related to the integration of
strongly singular volume integrals in the anisotropic elasto-plastic grain-boundary
equations are discussed and suitably addressed. In the polycrystalline case,
Laguerre-type micro-morphologies are discretised using robust non-structured
boundary and volume meshes. A general grain-boundary incremental/iterative
algorithm, embedding rate-dependent flow and hardening rules for crystal
plasticity, is developed and discussed. The method has been assessed through
several numerical simulations, for both single and polycrystalline aggregates,
which confirm its robustness and accuracy and suggest directions for further
developments. The formulation retains the key feature of the grain boundary formulation described in Chapter (1), that is the expression of the micromechanical problem in terms of grain-boundary variables only, namely intergranular displacements and tractions, which results in a reduction of the total
number of degrees of freedom and may be appealing in a multi-scale framework.
129
130
5.1
Grain boundary formulation for crystal plasticity
Introduction
Crystal plasticity represents one of the main deformation mechanisms active
in grains of a polycrystal and consists of plastic slip over well defined crystallographic planes. The phenomenon is characterised by its inherent anisotropy,
induced by the structure of the crystalline lattice itself. Several numerical and
computational methods have been proposed in the literature for the analysis
of plasticity in single crystals and crystal aggregates. A detailed comparison
among three different theoretical frameworks, i.e. conventional continuum
plasticity, discrete dislocation plasticity and nonlocal plasticity, is reported by
Needleman [150], who strongly underlines the importance of computational
studies. The importance of computational approaches addressing multiple
scales and involving the explicit representation of the micro-structural details
characterising the crystal plasticity phenomenon is highlighted in the work
of Dawson [54]. A very complete review of Crystal Plasticity Finite Element
Methods (CPFEMs) can be found in [178] where a wide range of continuumbased techniques and contributions, with applications in fields as diverse as
texture evolution, creep, nano indentation, forming and others, is discussed.
Single crystal plastic slip studies can be traced back to the works of Taylor
et al. [211, 210, 212]. Classical contributions in the field are those by Mandel [132], Hill [89], Rice [174], Hill and Rice [90], Asaro and Rice [15], Asaro
[13] (also addressing the mechanics of polycrystals), Peirce et al. [159, 160]
and Bassani et al. [231, 24]. In Ref. [52], Cuitino and Ortiz reviewed some
computational models for single crystal plasticity and proposed a statistical
mechanical model based on dislocation mechanics. Both rate-dependent (RD)
[174, 160, 124] and rate-independent (RI) [7] models have been proposed and
compared [135]. Much research has also been focused on the development
of crystal plasticity models for polycrystalline aggregates [178]. Several 3D
CPFEMs were initially developed to assess texture development in polycrystals. In several studies, the aggregate was represented with simple morphologies and coarse meshes, with individual grains often modelled by a single
finite element. Despite such approximations, the models were quite accurate
in predicting texture evolution. In [230], a Taylor-type polycrystal plasticity
model was tested in large strain reverse torsion tests, with the different slip
hardening laws presented by Asaro and Needleman [14], by Harren et al. [84],
by Bassani and Wu [24] and by Kalidindi et al. [99].
Both 2D [105] and fully 3D [183] polycrystalline models, with the explicit
representation of individual grains, have been developed, using a rate-dependent
5.1 Introduction
131
CPFEM formulations. Quilici and Cailletaud [168] developed a CPFEM framework that was subsequently used for the analysis of stress/strain fields within
3D polycrystalline aggregates in the small strains regime [18, 19]. Voronoi tessellations have been used to represent the microstructure and a visco-plastic
crystallographic constitutive model, with isotropic and kinematic hardening,
was employed. Aggregates up to 200 grains were analysed and the FETI (Finite Element Tearing and Interconnecting) method [68] was employed, in conjunction with High Performance Computing, to tackle the high computational
requirements of 3D computations. The same strategy was later employed
by Diard et al. [59], who analysed micro-plasticity in hexagonal close-packed
materials, with as realistic as possible microstructures. Sarma and Radhakrishnan [182] applied CPFEM for studying the hot deformation of grains in
polycrystalline aluminum. They report that, due to the time-consuming nature of such simulations, it was possible to consider only a limited number
of specially constructed microstructures. Zeghadi et al. [236] performed, in
a CPFEM framework, a statistical analysis on several different 3D morphologies generated starting from the same 2D surface. Aggregates with around
40 grains were considered and the plastic strain fields over the reference 2D
surface, induced by different internal 3D morphologies of the assessed specimen, were investigated. It was found that large fluctuations in the equivalent
plastic slip level are induced by different internal 3D morphologies, under
the same external loading conditions, thus advocating the need of fully 3D
studies over simplified pseudo 3D (or 2D columnar) studies.
Similar studies, focused on the comparison between experimental and
computational results, with the aim of identifying possible sources of error
and better calibrating microstructural constitutive parameters were performed
by Héripré et al. [88], Musienko et al. [146], Zhang et al. [241] and Pinna et al.
[165]. Musienko et al. [146] analyzed an experimentally reconstructed copper
specimen with approximately 100 grains. CPFEM was used and the results
of fully 3D and columnar 2D (pseudo-3D) simulations were compared to assess the bias introduced by the simplified morphologies. It was found that
the computational results of the fully 3D mesh were in good agreement with
the experimental data, while only some traces of the experimental behaviour
were retained in the extended 2D simulations. Zhang et al. [241] compared, in
terms of stress-strain curves and lattice reorientation, the experimental results
obtained from a tensile test of a Ti-5Al-2.5Sn specimen with those obtained
from numerical CPFEM simulations using 2D columnar and 3D microstructures. The stress-strain curve was used to calibrate the parameters of the
132
Grain boundary formulation for crystal plasticity
crystal plasticity constitutive model. Using the lattice reorientation maps, the
authors showed that fully 3D microstructures better reproduced the experimental observations with respect to the pseudo 3D microstructures. Pinna
et al. [165] simulated the deformation of a FCC aluminium alloy AA5052
at high temperatures using a structured mesh where each cubic volume element had its own crystallographic orientation, and a Fe-30wt%Ni alloy using
a two dimensional mesh with plane strain assumption. Besides the qualitative agreement between the numerical results and the experimental measurements in terms of stress-strain curves, the authors reported the limitations
of simplified CPFEM models to capture the intra-granular strain distribution
and suggested the use of fully three-dimensional models. Barbe and Quey
[20] presented a CPFEM model for 3D polycrystalline microstructures, including polycrystal-to-polycrystal diffusive transformations. Schneider et al.
[188, 187] investigated the mechanical behaviour of αFe-Cu composites under
large plastic deformations in simple tension and compression, using 3D FEM
simulations (ABAQUS) with an elasto-visco-plastic material model. Crack
initiation and evolution in polycrystalline aggregates is addressed in [202],
where the cohesive-zone type contact is implemented in a CPFEM framework employing Bassani-Wu hardening laws. In [177], a comparison between the self-consistent, finite element and spectral methods is carried out in
the context of homogenisation of polycrystalline aggregates with an elasticviscoplastic behaviour. Zhang et al. [242] studied the anisotropic mechanical response of a rolled AA1050 aluminium plate with respect to the loading direction using five different crystal plasticity models, showing the ability
of CPFEM to qualitatively reproduce the experimental findings. Zeng et al.
[237] developed a smoothed crystal plasticity finite element framework for
modelling two-dimensional single crystals and polycrystals. Kim and Yoon
[103] simulated the necking behaviour of AA 6022-T4 sheet using a CPFEM
formulation involving four different continuum damage mechanics models,
discussing the performance of such damage models when compared to experimental tests.
More advanced crystal plasticity phenomenological models, able to capture different deformation mechanisms besides slip, such as twinning in HCP
crystals or non-Schmid effects in BCC crystals, are still under investigation.As
a few recent examples, Abdolvand et al. [3] implemented in a finite element
formulation a crystal plasticity model comprising twinning as well as slip in
HCP materials, in order to simulate the deformation of a rolled Zircaloy-2 element. Wang et al. [222] proposed a physically-based crystal plasticity model
5.1 Introduction
133
including also de-twinning for HCP crystals and used it in a self-consistent
model to simulate cyclic loading in a magnesium alloy. Lim et al. [122] developed a BCC crystal plasticity finite element model and employed it to simulate the deformation of tantalum oligo-crystals [121]. The authors compared
numerical and experimental results finding good agreement in terms of both
in-plane and out-of-plane strain fields and lattice rotations. An interesting
point raised by the authors was the lower agreement between predicted and
measured fields at the grain boundaries, which prompted the authors to recommend more detailed grain boundary models in future studies.
The short review given above confirms the maturity of the CPFEM for the
investigation of plasticity in crystalline solids and also the current engineering interest toward the development of new or more enriched models. On the
other hand, while several boundary element formulations for elasto-plastic
isotropic problems are present in the literature, see e.g. Refs. [41, 17, 6, 39,
47, 26, 131, 130], the present Chapter represents the first boundary element
application to anisotropic or crystal plasticity in polycrystalline materials. A
boundary element approach may in fact result in accurate stress/strain predictions at reduced discretisation effort.
In the present Chapter, a novel three-dimensional grain-boundary formulation for small strains crystal plasticity in single crystals and polycrystalline
materials is presented for the first time. The boundary integral equations
given in Chapter (1) are suitably modified to account for general anisotropic
elasto-plasticity and are given in Section (5.2). The presence of plastic flow
is taken into account through an initial strains approach, which is commonly
adopted in boundary element formulations for addressing elasto-plastic problems [17, 6, 41]. The free terms involved in the formulation are explicitly
given, Section (5.2.1), and the strategy for numerical integration of the strongly
singular volume integrals is discussed in Section (5.2.3). The adopted phenomenological continuum-based crystal plasticity description is discussed in
Section (5.3).
The grain-boundary polycrystalline implementation for crystal plasticity,
which is based on the formulation given in Chapter (1), is given in Section (5.4)
and the incremental/iterative grain-boundary algorithm for crystal plasticity
is presented in a rate-dependent framework in Section (5.4.2). The salient
feature of the formulation is the expression of the problem in terms of intergranular variables only, namely grain-boundary displacements and tractions,
also in presence of an interior plastic flow, which allows a remarkable reduction in the order of the numerical system, especially for polycrystalline analy-
134
Grain boundary formulation for crystal plasticity
sis. In Section (5.5), numerical tests are performed both for single crystals and
crystal aggregates, and they confirm the potential of the technique. Directions
of further investigation are discussed in Section (5.6).
5.2
Anisotropic elasto-plastic boundary element formulation
In the present Section, the different ingredients of the boundary integral formulation of the crystal plasticity problem are presented. To maintain a unified
notation, the basic equations are presented in rate form. This is a common procedure for time-dependent problems, such as creep or visco-plasticity. Rateindependent crystal plasticity problems, on the other hand, could be formulated in purely incremental terms, due to the lack of time-dependent effects.
However, also rate-independent problems can be generally associated to a
time-like parameter that plays the role of an ordering parameter, such as the
load factor λ. Keeping this in mind, it is possible to retain a unique rate notation for both classes of problems.
Considered a generic domain V g ⊂ R3 representing a single grain with
specific symmetry undergoing elasto-plastic deformations, the proposed formulation is based on the use of the set of boundary integral equations representing the displacements and stresses of the points of the domain in terms
of boundary integrals of the displacements and tractions of points on the surface S g of the domain itself, see Section (1.3) of Chapter (1). When a region
V p,g ⊆ V g undergoes plastic slip, such equations are modified by the presence of some volume integral terms. The salient feature of the method is that
only the boundary points carry degrees of freedom, irrespective of the presence of internal plastic regions, allowing a considerable reduction in the size
of the solved system, which may be appealing in computational terms.
In the following Sections, the previous outline is further developed and
attention is devoted to the fundamental boundary integral equations, to the
expressions of the integral kernels appearing in the equations and their numerical integration, and to the form of the discrete algebraic systems.
5.2.1
Boundary integral equations
Let us consider the generic grain V g and its boundary S g = ∂V g . When plastic
deformations are considered within the context of the small strains theory, the
5.2 Anisotropic elasto-plastic boundary element formulation
135
g
total strain rate ε̇ ij (x) can be expressed as the sum of the elastic strain rate
e,g
p,g
ε̇ ij (x) and plastic strain rate ε̇ ij (x), i.e.
" g
#
g
∂u̇ j
1 ∂u̇i
g
e,g
p,g
(x) +
(x) = ε̇ ij (x) + ε̇ ij (x),
ε̇ ij (x) =
2 ∂x j
∂xi
(5.1)
where it is recalled that ui denotes the displacements components. The presence of plastic strains can be addressed in boundary element formulations by
adopting an initial strain or initial stress approach [17, 6, 41]: in the present
work, the initial strain formulation is used, as it appears particularly suitable
to the form of the crystal plasticity constitutive equations, see Section (5.3).
p,g
When some initial plastic strain ε ij is present in V p,g ⊆ V g , the displacement
boundary integral equations (1.9) are modified and rewritten, in rate form, as
g
c pi (y)u̇i (y) +
Z
g
g
− Tpi (x, y)u̇i (x)dS(x) =
Sg
Z
g
Sg
g
U pi (x, y)ṫi (x)dS(x) +
Z
g
V p,g
p,g
Σ pij (x, y)ε̇ ij (x)dV (x) (5.2)
where, consistently with Eq.(1.9), y denote the collocation point, x denotes the
boundary and volume integration point and ui (x) and ti (x) are the components of boundary displacements and tractions, respectively.
The kernels apR
pearing in Eq.(5.2) are given in Eq.(1.22) of Chapter (1). − denotes the Cauchy
principal value integral and c pi (y)u̇i (y) are free terms. The volume integral
in Eq.(5.2) is performed over the interior region V p,g ⊆ V g undergoing plasp,g
tic deformation, i.e. where ε̇ ij 6= 0. Similarly to the integral containing the
g
kernel U pi (x, y), the volume integral, although containing a singular kernel of
order r −2 , is a weakly singular integral. In fact, upon considering a reference
system centered at y and taking the limit x → y, the order of the volume eleg
ment dV (x) is r2 dr whereas the kernel Σ pij (x, y) is of the order r −2 , being r is
the distance between the point x and y.
By considering an interior point y ∈ V g , the strain field inside the grain
g is computed by suitably taking the derivatives of Eq.(5.2) with respect to
the coordinates yi of the collocation point y. In rate form, the strain boundary
integral equations (1.11) are modified as
g
ε̇ pq (y) +
Z
Sg
Z
ε,g
Sg
g
Tpqi (x, y)u̇i (x)dS(x) =
ε,g
g
U pqi (x, y)ṫi (x)dS(x) +
Z
−
V p,g
p,g
p,g
ε
Σ pqij (x, y)ε̇ ij (x)dV (x) + f pqij
ε̇ ij (y), (5.3)
ε,g
136
Grain boundary formulation for crystal plasticity
g
where the kernels U pqi (x, y), Σ pqij (x, y) and Tpqi (x, y) are given in Eq.(1.23). In
this case, unlike the DBIE (5.2), the volume integral must be evaluated in the
Cauchy principal value sense since, as x → y, the order of the volume element
ε,g
dV (x) is r2 drdS and the order of the kernel Σ pqij (x, y) is r −3 . Consequently,
ε,g
ε,g
g,p
ε ε̇ ( y ) remain and the coefficients
after the limiting process, the free terms f pqij
ij
ε
f pqij are computed as
ε
f pqij
1
=−
2
Z
S1
[Σ pij (r̂)nq (r̂) + Σqij n p (r̂)]dS(r̂),
(5.4)
where S1 is the three dimensional unit sphere and r̂ is the unit vector centered
at y. The numerical treatment of the volume integrals in Eq.(5.3) is addressed
in Section 5.2.3.
The boundary integral representation for the stress rate σ̇mn (y) at the generic
interior point y ∈ V g is obtained from Eq.(5.3) using the constitutive relatione,g
g
p,g
ships σ̇mn (y) = cmnpq ε̇ pq (y) = cmnpq [ε̇ pq (y) − ε̇ pq (y)]. In rate form, the stress
boundary integral representation reads
g
σ̇mn (y) +
Z
Sg
Z
g
σ,g
Sg
Tmni (x, y)u̇i (x)dS(x) =
σ,g
g
Umni (x, y)ṫi (x)dS(x) +
Z
−
p,g
p,g
σ
Σmnij (x, y)ε̇ ij (x)dV (x) + f mnij
ε̇ ij (y), (5.5)
σ,g
V p,g
where the kernels Umni (x, y), Tmni (x, y) and Σmnij (x, y) are given in Eq.(1.24)
and the coefficients of the free terms are
Z
σ
f mnij = −cmnpq
Σ pij (r̂)nq (r̂)dS(r̂) + δpi δqj .
(5.6)
σ,g
σ,g
σ,g
S1
5.2.2
Numerical discretisation
In the initial strains approach [41, 6], the crystal plasticity problem is closed
considering the displacement and stress boundary integral equations, Eq.(5.2)
and Eq.(5.5), together with the crystal plasticity constitutive equations given
in Section (5.3). In order to write the discrete counterpart of the boundary
integral equations, the following steps are followed:
• The boundary S g and the volume V g of each grain domain are subdivided into sets of non-overlapping surface and volume elements:
5.2 Anisotropic elasto-plastic boundary element formulation
137
– the surface is discretised using the strategy developed in Chapter (3), where a mesh combining continuous/semi-discontinuous
triangular and quadrangular two-dimensional elements is implemented in order to tackle and reduce the high computational costs
of the polycrystalline problem
– the volume of the grain is divided into linear tetrahedral elements
and the plastic strain field is assumed constant and equal to the
value at the centroid of each element. Such elements are usually
referred to as sub-parametric elements1 ;
• The unknown fields, namely displacements and tractions on the surface
and plastic strains in the volume, are expressed in terms of nodal values
and suitable shape functions;
• Eq.(5.2) is written for every boundary discretisation node and it is then
numerically integrated;
• Eq.(5.5) is written at the centroids of each tetrahedral volume element
and it is then numerically integrated.
The discrete version of the DBIE written in rate form coincide with Eq.(1.25)
plus an additional term accounting for the plastic deformation of the volume
as follows
p,g
H g U̇ g = G g Ṫ g + D g Ė ,
(5.7)
where U̇ g and Ṫ g contain the components of the nodal values of the boundary
displacements and tractions, the matrices H g and G g stem from the integrae g (x, y) and T
e g (x, y) in the local boundary refertion over S g of the kernels U
pi
pi
p,g
ence system (see Eq.(1.13) in Section(1.3)), respectively, Ė contain the components of the values of plastic strains at the tetrahedral elements’ centroids
g
and the matrix D g is generated from the numerical integration of Σ pij (x, y)
over the interior volume of the grain g undergoing plastic flow.
To solve the crystal plasticity problem, Eq.(5.7) has to be used with the
discretised version of the stress integral equations (5.5) that, written for each
internal point, read
g
Σ̇ = Gσ,g Ṫ g − Hσ,g U̇ g + Dσ,g Ė
1 In
p,g
,
(5.8)
sub-parametric elements the order of interpolation of the unknown fields is lower than
the order of interpolation of the geometry.
138
Grain boundary formulation for crystal plasticity
g
where Σ̇ contain the components of the nodal values of stress at the centroids
of the tetrahedra, the matrices Hσ,g and Gσ,g stem from the integration of the
σ,g
σ,g
kernels Tmni (x, y) and Umni (x, y), respectively, over the boundary S g and Dσ,g
σ,g
stems from the numerical integration of Σmnij (x, y) over the interior volume
of the grain g undergoing plastic flow.
Eqs.(5.7) and (5.8) must be used in conjunction with the suitable boundary conditions and the constitutive crystal plasticity equations for the solution
of the grain-boundary elasto-plastic problem, as will be described in Section
(5.4). The interested reader is referred to references [17, 6, 41] for further details about the boundary element method.
5.2.3
Numerical integration and singular integrals
To obtain the discrete equations introduced above, it is necessary to accurately evaluate the singular integrals appearing in Eqs.(5.2) and (5.5). While
the singular integrals appearing in the displacement equations, Eq.(5.2), have
been widely considered in the literature, the integration of the singular kerε,g
σ,g
nels Σ pqij (x, y) and Σmnij (x, y) for the anisotropic case was recently presented in
[30]. In this Section, the procedures used for the evaluation of the singular
integrals are briefly recalled, and the strategy used for treating the integrals
σ,g
involving Σmnij (x, y) is described. The same approach applies to the integrals
involving Σ pqij (x, y).
Considering Eq.(5.2), the weak singularity induced on the boundary by
g
Uij (x, y) is treated subdividing the singular elements into triangles and representing each sub-triangle as a quadrangle collapsed into the singular point;
g
the strong singularity induced by Tij (x, y), treated as a Cauchy singular value,
is tackled using rigid body considerations [17, 6]; the volume terms appearing in Eq.(5.2) may give rise to weak singularities only when the plastic region
reaches the boundary: in this case, the interested volume cells are subdivided
into sub-tetrahedra, which afterwards are represented as 8-node cubes collapsed into the singular point. In this way, the Jacobian of the integration
cancels the singularity.
In the strain and stress integral equations, Eqs.(5.3) and (5.5), since constant volume cells are used in the implementation, the collocation points y
always fall in the interior domain, being the centroids of the volume cells. For
such a reason, the boundary integrals may become at most nearly singular,
and this happens when the plastic region reaches the boundary. Nearly sinε,g
5.3 Crystal plasticity phenomenological modelling
139
gular integrals are simply evaluated through boundary element subdivision.
The volume integrals, on the contrary, must be treated in the Cauchy singular
ε,g
σ,g
value sense, being the kernels Σ pqij (x, y) and Σmnij (x, y) strongly singular in
the volume.
In the present work, the starting point for the evaluation of the volume
Cauchy singular values is the work by Gao and Davies [72], who, for isotropic
σ,g
materials, have shown that the integral of Σmnij (x, y) over the singular volume
cell V c can be expressed as
Z
−
Vc
σ,g
Σmnij (x, y)dV (x)
=
Z
Σmnij (x, y)( x p − y p )n p (x) log(r )dS(x)
σ,g
Sc
(5.9)
as long as the condition
Z
Σmnij (r̂)dS(r̂) = 0,
σ,g
S1
(5.10)
where S1 is the unit sphere centered at y, is satisfied. In Eq.(5.9), Sc = ∂V c
is the external boundary of the singular volume cell. Gao and Davies [72]
proved Eq.(5.10) for isotropic materials, taking into account the explicit analytical form of the isotropic fundamental solutions. Using the spherical harmonics expansion developed in Chapter (2), theR fact that the integral of the
spherical harmonics Y`m (r̂) over the unit sphere S1 Ym` (r̂)dS(r̂) = 0 for m, ` 6=
0 and that P02 (0) = 0 (P`m (0) is the associated Legendre function), it is possible
to demonstrate that Eq.(5.10) also holds for general anisotropic materials.
In the next Section, the crystal plasticity constitutive model linking Eqs.(5.7)
and (5.8) and employed in this thesis is discussed.
5.3
Crystal plasticity phenomenological modelling
In crystallography, a slip system family describes a set of crystallographically
equivalent slip planes associated with a family of crystallographically equivalent slip directions for which dislocation motion can occur, producing plastic
deformation. Different slip systems are associated with different classes of
crystal lattice. Generally, slip planes are those with the highest surface density
of atoms, i.e. close-packed planes, while slip directions are those corresponding to highest linear density of atoms, i.e. close-packed directions. Once a
specific slip plane and a specific slip direction over it are identified, a certain
resolved shear stress is needed to trigger the slip, which is an important deformation mechanism in crystals.
140
Grain boundary formulation for crystal plasticity
Crystal plasticity can be generally modeled through phenomenological or
physically-based approaches [178]. Moreover, the crystallographic slip within
single crystals can be idealised as either rate-dependent or rate-independent. In
the present work, a phenomenological rate-dependent description is adopted.
In the present Section, an introductory functional description of crystal slip is
briefly recalled.
The slip over a specific slip plane α identified by the unit normal vector
α
ni , and along a specific direction identified by the unit vector siα , is activated
by the Schmid resolved shear stress
τ α = siα σij nαj ,
(5.11)
which triggers the corresponding slip or shear rate
γ̇α = Ψα (τ α , τcα ) ,
(5.12)
where Ψα expresses a specific flow rule and τcα is a certain critical resolved shear
stress, which plays the role of a state variable
for the slip system α. τcα generally
R
depends on the accumulated slips γ β = t |γ̇ β |dt and on the slip rates γ̇ β on all
the Ns slip systems and its evolution can be expressed as
τ̇cα = Ξα γ β , γ̇ β
(5.13)
where Ξα expresses some specific hardening law and β = 1, . . . , Ng , meaning
that the evolution of τcα of the generic slip system α is a function of slip rate
and accumulated slip of all the slip systems. To avoid possible confusion, the
sum of the accumulated slips γα over all the Ns slip systems, which is defined
as γ ≡ ∑α γα , will be referred to as the overall cumulative slip. Eqs.(5.11-5.13)
are valid ∀α = 1, · · · , Ns .
Once the slip rates are known, the plastic strains rate at a point in the
crystal can be expressed as
p
ε̇ ij
1 Ns α α α
α α
= ∑ γ̇ ni s j + si n j .
2 α =1
(5.14)
The developed formulation can be coupled with different types of flow and
hardening rules. The flow and hardening rules used in this Chapter are reported in Section (5.7).
5.4 The grain-boundary incremental-iterative algorithm for crystal
plasticity in polycrystalline aggregates
5.4
141
The grain-boundary incremental-iterative algorithm
for crystal plasticity in polycrystalline aggregates
Following Refs. [213, 41], in the framework of the initial strain formulation the
elasto-plastic grain-boundary problem can be expressed in terms of accumulated values of displacements U g , tractions T g and plastic strains ε p,g . Eq.(5.7)
can then be rewritten as
H g U g = G g T g + D g E p,g
BCs
−−→
A g X g = C g Y g (λ) + D g E p,g
(5.15)
where the column reordering induced by the enforcement of the boundary
conditions (BCs) is indicated. In Eq.(5.15), the matrix A g contains a selection of columns from H g and Gg corresponding to unknown components of
boundary displacements and tractions, collected in the unknowns vector X g ;
the matrix C g is a linear combinations of the columns from H g and G g multiplying the known components of boundary displacements and tractions, collected in the vector Y g , and whose value is governed by the load factor λ.
Similarly, Eq.(5.8) can be rewritten, in terms of accumulated values, as
Σ g = Aσ,g X g − Cσ,g Y g (λ) + Dσ,g E p,g ,
(5.16)
where, similarly to the matrices A g and C g , the matrices Aσ,g and Cσ,g contains a suitable combination of columns from Hσ,g and Gσ,g corresponding to
unknown and known, respectively, components of boundary displacements
and tractions. However, it is recalled that these equations are generally used
in a post-processing stage and the distinction between known and unknown
values of boundary fields is adopted for the sake of notation.
5.4.1
Numerical discretisation of polycrystalline aggregates
Similarly to the assemblage procedure of the polycrystalline aggregate described in Section (1.4) of Chapter (1), Eq.(5.15) is written for each grain of the
aggregate, i.e. for g = 1, . . . , Ng , and coupled to an appropriate set of interface
equations. The system of equations of the whole aggregate is then written as
AX − B(λ) − DE p
p
M(X, λ, E ) =
=0
(5.17)
IX
where the matrix A and the vector B are given in Eq.(1.28). Unlike the system
of the equations (1.27), in which non-linear behaviours within the grains are
142
Grain boundary formulation for crystal plasticity
not considered, in Eq.(5.17), the presence of plastic slip within the polycrystalline aggregate is represented by the term DE p where
D=
D1 0 · · ·
0 D2 · · ·
..
..
..
.
.
.
0
0 ···
0
0
..
.
D Ng
and
Ep =
E p,1
E p,2
..
.
E p,Ng
.
(5.18)
In the system of equations (5.17), the interface equations implement perfect
bonding between contiguous grains and are given by the relation IX = 0,
where the matrix I contains just zeros and ±1’s enforcing the interface condigh
g
tions δuei = 0 and eti = etih for pairs of interface points belonging to the generic
adjacent grains g and h.
In the solution algorithm described in Section (5.4.2), the system of equations (5.17) is used in conjunction with the stress equations (5.16) which are
used to evaluate the stress at each interior test point within each grain. It
is worth noting that, being Eq.(5.16) used after the solution of Eq.(5.17), in a
post-process stage within each iteration, it can be used for each grain separately, as U g and T g are known. However, for the sake of notation, the computation of the stress at the interior points of each grain of the polycrystalline
aggregate is indicated as
Σ(X, λ, E p ) = Aσ X − Bσ (λ) + Dσ E p ,
(5.19)
where
Σ=
Bσ =
Σ1
Σ2
..
.
Σ Ng
, Aσ X =
Cσ,1 Y1
Cσ,2 Y2
..
.
Cσ,Ng Y Ng
Aσ,1 X1
Aσ,2 X2
..
.
Aσ,Ng X Ng
, Dσ E p =
,
Dσ,1 E p,1
Dσ,2 E p,2
..
.
Dσ,Ng E p,Ng
.
The incremental-iterative algorithm for the solution of the crystal plasticity problem is described in the next Section.
5.4 The grain-boundary incremental-iterative algorithm for crystal
plasticity in polycrystalline aggregates
5.4.2
143
Polycrystalline incremental-iterative system solution
The incremental-iterative algorithm used to solve the crystal plasticity problem is described in Algorithm (1), which is detailed in the following. In the
Algorithm, the subscript n refers to the n-th load step whereas the superscript
k refers to the k-th iterative step of the considered load step. Moreover, the
superscript g refers to the g-th grain of the aggregate, ye is the position of the
centroid of e-th volume cell and α denotes the α-th slip system of the grain.
The Algorithm starts at Step (1) at which the load factor is initialised to 0
and the index n is initialised to 1. The Algorithm runs until the load factor λ
reaches the given final load factor λ f . Once the solution at the load step n − 1
corresponding to the load factor λn−1 has been found, the load increment λ̇∆tn
is introduced, being λ̇ the loading rate and ∆tn the time step, and the new load
factor λn is computed. At the beginning of the analysis, the plastic strain
p
Ė n=0 throughout the whole polycrystalline aggregate as well as the plastic
g
slip γnα=0 (ye ) for each slip system of the volume cells of each grain are set to
p
g
zero, i.e. Ė n=0 = 0 and γnα=0 (ye ) = 0. On the other hand, the initial value of
g
α
the critical resolved shear stress τc,n
=0 ( ye ) is set to an initial value depending
on the considered material.
At Step (6), the iterative cycle starts and the solution of the system of equations (5.17) and (5.19) coupled to the constitutive relations (5.11), (5.12) and
(5.13) is computed. At Step (7), Eq.(5.17) is solved for the current (k + 1)-th
iteration as a function of λn and of the value of the plastic strain at the k-th
p
iteration, which is obtained as the sum of the last converged value E n−1 plus
the increment at the k-th step given by the product of the time step ∆tn and
p,k
the last computed value of the plastic strain rate Ė n . It is worth noting that
p
in the Algorithm, E n−1 collects the converged plastic strain components up to
the load level λn−1 , but not including the last increment λ̇∆tn .
The updated values of boundary displacements and tractions, given by
k
+
Xn 1 and by the prescribed values governed by λn , are then used in Step
(8) to compute the internal stress components at the volume cells’ centroids
collected in Σkn+1 , which in turn are used to obtain the resolved shear stress
g
τnα,k+1 (ye ) for each slip system α of the centroids of all the volume cells of
each grain (Step (9)).
g
At Step (10), Eq.(5.12) is used to compute the slips rate γ̇nα,k+1 (ye ) and
g
the accumulated slips γnα,k+1 (ye ) for each slip system of the volume cells’
centroids of each grain. Similarly, at Step (11) the critical shear stress rate
α,k +1 g
(ye ) is computed using Eq.(5.13) and the previously computed values
τ̇c,n
144
Grain boundary formulation for crystal plasticity
of the slips rate and the accumulated slips. At Step (12), the rate of the plastic
p,k+1 g
strain components ε̇ ij,n (ye ) is computed using Eq.(5.14).
g
At this point, the overall cumulative slip γnk+1 (ye ) is computed for the
volume cells of each grain and the convergence condition is checked. The
convergence conditions is given in terms of the difference between two cong
secutive cumulative slip between γnk (ye ) compared to a tolerance value econv .
If the convergence is met then the accumulated slips, the critical shear stresses
and the accumulated plastic strains are updated and stored and another load
increment can be introduced and computed if needed. If the convergence is
not met, the iteration counter k is updated and another iteration is started.
5.4 The grain-boundary incremental-iterative algorithm for crystal
plasticity in polycrystalline aggregates
145
Algorithm 1 CPBEM algorithm
1:
2:
3:
4:
5:
6:
7:
8:
START
λ0 ← 0, n ← 1;
while (λn < λ f ) do
λn ← λn−1 + λ̇∆tn ;
k ← 0;
converged ← .false.;
while (.not.converged) do
Solve:
p,k
p
M(Xkn+1 , λn , E n−1 + Ė n ∆tn ) = 0;
Compute the internal stress field:
p
9:
10:
. Eq.(5.17)
p,k
Σkn+1 = Σ(Xkn+1 , λn , E n−1 + Ė n ∆tn );
. Eq.(5.19)
∀ g, e, α compute the resolved shear stress:
g
g
g,k +1 g
g
τnα,k+1 (ye ) = siα (ye )σij,n (ye )nαj (ye );
. Eq.(5.11)
∀ g, e, α compute the slips rate and the accumulated slips:
g
g
g
α,k g
α
γ̇nα,k+1 (ye ) = Ψα [τnα,k+1 (ye ), τc,n
. Eq.(5.12)
−1 ( ye ) + τ̇c,n ( ye ) ∆tn ];
g
g
g
γnα,k+1 (ye ) = γnα−1 (ye ) + |γ̇nα,k+1 (ye )|∆tn ;
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
∀ g, e, α compute the critical shear rate:
β,k +1 g
β,k +1 g
α,k +1 g
τ̇c,n
( y e ) = Ξ α [ γn
(ye ), γ̇n
(ye )];
. Eq.(5.13)
∀ g, e compute the plastic strain rate:
. Eq.(5.14)
g α g
g α g
p,k +1 g
α,k+1 g
1
α
α
(ye )[ni (ye )s j (ye ) + n j (ye )si (ye )]
ε̇ ij,n (ye ) = 2 ∑α γ̇n
g
g
∀ g, e compute the cumulative slip γnk+1 (ye ) = ∑α γnα,k+1 (ye );
g
g
if (∀ g, e γnk+1 (ye ) − γnk (ye ) < econv ) then
converged ← .true.
else
k ← k+1
end if
end while
α g
g
α,k +1 g
α
(ye )|∆tn
γn (ye ) ← γn−1 (ye ) + |γ̇n
g
g
α,k +1 g
α
α
∀ g, e, α update and store τc,n (ye ) ← τc,n−1 (ye ) + τ̇c,n (ye )∆tn
p
p,k +1
p
E n ← E n−1 + Ė n
∆tn
n ← n + 1;
end while
EXIT
146
5.5
Grain boundary formulation for crystal plasticity
Numerical tests
In this Section, the developed and implemented formulation is numerically
and computationally assessed. The phenomenological framework by Bassani
and Wu [24] is used in the first series of tests performed in the present work,
considering FCC Copper (Cu) as benchmark material. The explicit expressions of the flow rule and the hardening law of the Bassani and Wu crystal
plasticity framework are given in Section (5.7). The single crystal properties
as well as the slip systems numbering used in the present study coincide with
those in Refs. [231, 24]. The elastic constants and the crystal plasticity parameters for FCC copper single crystals are summarised in Table (5.1).
Table 5.1: Elastic and crystal plasticity properties for FCC Copper from [24].
To interpret the symbols, the reader is referred to Section (5.7.1).
property
elastic constants [109 N/m2 ]
initial critical resolved
shear stress [106 N/m2 ]
reference shear rate [1/s]
rate sensitivity [-]
stage I reference stress [106 N/m2 ]
initial hardening modulus [106 N/m2 ]
hardening during easy glide [106 N/m2 ]
interaction peak slip [-]
latent hardening factor [-]
cross-hardening moduli f αβ [-]
component
c1111 , c2222 , c3333
c1122 , c1133 , c2233
c1212 , c1313 , c2323
value
170
124
64.5
τ0α ≡ τ0 , α = 1, · · · , Ns
17
γ̇0α , α = 1, · · · , Ns
n
τs
h0
hs
γp
q
no junction
Hirth lock
Coplanar interaction
Glissile junction
Sessile junction
0.001
50
1.3 · τ0
90 · τ0
1.5 · τ0
0.001
0
8.0
8.0
8.0
15.0
20.0
5.5 Numerical tests
5.5.1
147
Single crystal tests
To assess the accuracy and numerical performance of the formulation, tests on
a single crystal are performed first. The analysed domain is a cubical crystal,
with the edges aligned with the reference system and subjected to uniaxial
stress acting along the x3 direction.
Single crystal test: algorithm test
In Figure (5.1), the behaviour of the implemented scheme is tested with respect to the value of the rate sensitivity n appearing in Eq.(5.23), for the three
different crystallographic loading directions [001], [111] and [632], chosen as
in Ref. [24] for comparative purposes. In all tests, the loading rate λ̇ is always
set as λ̇ = 10−3 s−1 . The curves plot the stress component σ33 at the centroid
of the cube versus the strain component ε 33 at the same point. It is worth noting that nominal components of strains are considered, being the developed
formulation valid for small strains. However, strains up to 10% are plotted
to investigate when the small strains assumption breaks down: it is interesting to note that, up to this value of strains, the converged curves match
reasonably well with those reported in [24], which have been replicated in the
figure using a simple rate-independent scheme and represented with dashed
lines. It is confirmed that, as n rises, the rate-dependent scheme simulates
a rate-independent process, although the convergence is not the same for all
directions, being noticeably slower (with respect to n) for the direction [632].
Also, it is interesting to note that for the [001] and [111] loading directions, the
rate-dependent scheme predicts a lower bound for the stress, whereas for the
[632] loading direction it predicts an upper bound.
Single crystal test: mesh convergence
In the following series of tests, the single crystal response is tested against different domain volume meshes, in order to assess the robustness of the computational homogenisation scheme that will be used in the polycrystalline tests.
Unlike the curves reported in Figs.(5.1), which are obtained considering the
values of stress and strain at the centroid of the cubic single crystal domain,
the stress-strain curves in the subsequent part of this work are obtained as
averaged results of the stress and the strain fields over the crystalline or poly-
148
Grain boundary formulation for crystal plasticity
60
100
50
80
40
RD
RD
RD
RD
RD
RI
20
10
-
n = 10
n = 20
n = 30
n = 40
n = 50
σ33 /τ0
σ33 /τ0
60
30
RD
RD
RD
RD
RD
RI
40
20
0
n = 10
n = 20
n = 30
n = 40
n = 50
0
0
0.02
0.04
0.06
0.08
0.1
0
0.02
ε33
0.04
0.06
0.08
0.1
ε33
(a) Loading direction [001]
(b) Loading direction [111]
60
25
40
30
-
n = 10
n = 20
n = 30
n = 40
n = 50
n = 100
n = 200
RD
RD
RD
RD
RD
RD
RD
RI
20
15
σ33 /τ0
RD
RD
RD
RD
RD
RD
RD
RI
50
σ33 /τ0
-
-
n = 10
n = 20
n = 30
n = 40
n = 50
n = 100
n = 200
10
20
5
10
0
0
0
0.02
0.04
0.06
0.08
ε33
(c) Loading direction [632]
0.1
0
0.005
0.010
0.015
0.020
ε33
(d) Loading direction [632]
Figure 5.1: Effect of the rate sensitivity parameter n for three different loading
crystallographic directions. Plot (d) shows a zoom from plot (c). It is apparent as the rate-dependent results converge to the rate-independent ones with
increasing values of n.
5.5 Numerical tests
149
crystalline domain. Strain and stress volume averages are given by
Γij =
1
V
Z
V
γij (x)dV (x) =
Σij =
1
V
Z
V
1
2V
Z
S
[ui (x)n j (x) + u j (x)ni (x)]dS(x)
σij (x)dV (x) =
1
V
Z
S
xi (x)t j (x)dS(x)
(5.20)
(5.21)
where S and V represents the boundary and the volume of the domain, respectively. The terms involving the surface integrals allow to write the volume averages in terms of boundary displacements and tractions, which are
the primary variables of the model. The terms involving the volume integrals, on the other hand, are obtained as the weighed sum
R of the strains and
stresses computed at the centroids of the tetrahedra, i.e. V (·)dV = ∑e (·)e Ve
being Ve the volume of the e-th volume element. The comparison between the
volume averages computed using the volume and the surface integrals can
be used as a tool to assess the accuracy in the integration of the kernels in the
displacement and stress boundary integral equations.
Figure (5.2) shows four surface and volume meshes of a cubic domain.
Following the procedure described in Chapter (3) and used in Refs. [29, 28, 80],
the surface and volume meshes of the crystalline domain are built considering
an as uniform as possible mesh size, which is computed as the average length
of the edges of the considered crystals divided by a mesh density parameter
dm . In the considered case, the average edge length is clearly the cube’s edge.
Figures (5.2a,e), (5.2b,f), (5.2c,g), (5.2e,h) correspond to dm = 2, 3, 4, 5, respectively. Although they may not represent the best meshing choice for a cubic domain, the combination of triangular and quadrangular elements in the
surface meshes and the tetrahedra in the volume meshes is the same used to
adapt to the geometrical variability of general Laguerre-based polycrystalline
morphologies. Some statistics about the meshing algorithm are summarised
in Table (5.2), which reports the number of degrees of freedom and the number of surface and volume elements of the single crystal meshes involved in
the present test, as well as those for the polycrystalline meshes considered in
the next Section.
The cube in Figure (5.2) is loaded in uniaxial stress, with uniform tractions
acting over the top and bottom faces, whereas the lateral faces are tractionfree. To avoid the rigid body motion, the technique proposed by Lutz et al.
[127] is used. Figure (5.3a) shows the stress-strain curves of a single crystal cubic domain loaded along the [632] direction for the four considered meshes.
150
Grain boundary formulation for crystal plasticity
(a)
(b)
(c)
(d)
(e)
(f )
(g)
(h)
Figure 5.2: Single crystal surface (a-d) and volume (e-h) meshes: (a,e) dm = 2,
(b,f ) dm = 3, (c,g) dm = 4, (d,h) dm = 5. The small circles in the surface meshes
represent the collocation points.
The curves are plotted in the strain range 0-2% considering a slip rate sensitivity n = 50 and are compared to that obtained for the single crystal centroid.
The four meshes replicate the results obtained in Figure (5.1d) as uniform
strain and stress fields are expected for this configuration. In Figure (5.3a),
the close-up view shows a comparison between the curves obtained using the
surface and the volume integrals of Eqs.(5.20) and (5.21). A good match is
found among the curves showing a satisfactory accuracy in the integration of
the kernels of the boundary integral equations. Finally, Figure (5.3b) shows
the same response in a smaller strain range, i.e. 0-0.2%.
5.5 Numerical tests
151
Table 5.2: Statistics about the meshes of the single crystal and polycrystalline
domains: dm is the mesh density parameter; Nes is the number of surface elements; DoFs is the number of degrees of freedom for the system in Eq.(5.17);
Nev is the number of volume elements. The last column reports the number
of volume elements per grain, which is considered as reference to capture the
volume field distribution.
dm
2
3
4
5
1
2
3
4
5
2.5
2.5
morphology
single crystal
polycrystalline, 10 grains
polycrystalline, 100 grains (I)
polycrystalline, 100 grains (II)
8
5.0
7
4.9
Nes
36
72
150
255
430
895
1811
3291
5242
19902
18075
DoFs
162
288
504
729
3594
6285
10131
15669
22143
132369
122166
Nev
46
161
436
836
562
1576
4097
8791
16119
45704
40181
Nev /Ng
46
161
436
836
56
158
410
897
1612
457
402
3.0
4.7
1.50
5
1.55
Macro stress Σ33 /τ0
Macro stress Σ33 /τ0
2.5
4.8
6
1.60
4
dm = 5
dm = 4
dm = 3
dm = 2
Single crystal
3
2
1
2.0
1.5
dm = 5
dm = 4
dm = 3
dm = 2
Single crystal
1.0
0.5
0
0
0
0.5
1.0
1.5
2.0
0
0.05
0.10
Macro strain Γ33 [%]
Macro strain Γ33 [%]
(a)
(b)
0.15
0.20
Figure 5.3: Volume stress average Σ33 plotted versus volume strain average
Γ33 for the four considered meshes of the single crystal loaded along the [632]
direction. (a) stress-strain response in the strain range 0-2.0%; the close-up
shows the values obtained using the surface integral (solid lines) compared
to those obtained using the volume integral (dashed lines). (b) stress-strain
response in the range strain 0-0.2%.
152
Grain boundary formulation for crystal plasticity
(a)
(b)
(c)
(d)
(e)
(f )
(g)
(h)
(i)
(j)
Figure 5.4: Surface (a-e) and volume (f -j) meshes of a 10-grain aggregate: (a,f )
dm = 1, (b,g) dm = 2, (c,h) dm = 3, (d,i) dm = 4, (e,j) dm = 5. The small circles
in the surface meshes represent the collocation points.
5.5.2
Polycrystalline tests
In the present Section, the proposed model is used to simulate the plastic
behaviour of Cu polycrystalline aggregates. To generate the polycrystalline
morphologies, the Neper package [167] is used for its capability of eliminating pathologically small geometrical entities typical of domains subdivided
by the Voronoi or Laguerre algorithm, which may induce unnecessarily heavy
meshes.
10-grain aggregate
First, a convergence analysis is performed to obtain a proper mesh size ensuring appropriate grains boundary and volume fields resolution and mesh
independency of the results. Figure (5.4) shows five different meshes (dm =
1, 2, ..., 5) of a 10-grain aggregate. The statistics about the number of DoFs and
the number of surface and volume elements are summarised in Table (5.2).
The aggregate is loaded in uniaxial tension, with uniform tractions acting over
the top and bottom faces, while the lateral faces are traction-free. In this case,
to avoid rigid body motion, the elimination of rigid body modes technique is
applied to one grain of the aggregate.
5.5 Numerical tests
153
8
dm
dm
dsm
dm
dm
dm
400
300
=1
=2
= 2, dvm = 3
=3
=4
=5
80
200
60
40
100
e(Σ133 , Σ233 )
e(Σ233 , Σ333 )
e(Σ333 , Σ433 )
e(Σ433 , Σ533 )
7
Macro stress Error [%]
Macro stress Σ33 [MPa]
500
6
5
4
3
2
20
1
0
0
0
0
0.5
0.1
1.5
1.0
0.2
2.0
0
0
0.5
1.0
Macro strain Γ33 [%]
Macro strain Γ33 [%]
(a)
(b)
1.5
2.0
Figure 5.5: Volume stress average Σ33 versus volume strain average Γ33 for
the considered meshes of the 10-grain aggregate. (a) Stress-strain response in
the strain range 0-2.0%; the close-up shows the values obtained in the strain
range 0-0.2%. (b) Error between two consecutive mesh refinements.
Figure (5.5a) shows the convergence of the macro stress-strain response
of the Cu 10-grain aggregate with respect to the considered meshes. As the
model allows to build the surface and the volume meshes separately, Figure
(5.5a) shows also the macro response obtained combining a surface mesh with
a mesh density parameter dm = 2 and a volume mesh with a mesh density
parameter dm = 3, which corresponds to the curve labelled with dsm = 2,
dvm = 3. Figure (5.5b) reports the macro-stress error during the loading history
between two consecutive mesh refinements defined as:
i
i +1
e(Σ33
, Σ33
) = 100 ·
i − Σ i +1 |
|Σ33
33
Σ533
(5.22)
i corresponds to the macro stress computed for d = i.
where Σ33
m
Figure (5.6) shows the deformed configurations of the aggregate for the
meshes of Figure (5.4) at the last computed step of the load history, i.e. for
Γ33 = 2%. A good convergence is obtained in terms of the grain boundary
displacements starting from dm = 2. Figure (5.7) and Figure (5.8) report the
contour plots of the Von Mises stress and cumulative slip fields, respectively.
Also in this case, the convergence in terms of the overall behaviour and stress
and slip concentration is satisfactorily reached through mesh refinement. It is
however noted that a dm = 3 mesh is required to appreciate the fields distribution, likely due to the use of constant volume elements. Considering Table
154
Grain boundary formulation for crystal plasticity
(5.2), an estimate of the number of volume elements per grain to obtain satisfactory fields distributions within the aggregate appears to be around 400
elements per grain, at least for the present load condition.
(a)
(b)
(c)
(d)
(e)
Figure 5.6: Deformed shape for the five considered meshes of the 10-grain
polycrystalline copper aggregate subjected to tensile load at the last computed
step (Γ33 = 2%). Figs.(a) to (e) correspond to the (a) to (e) surface meshes and
(f ) to (j) volume meshes of Figure (5.4), respectively.
(a)
(f )
(b)
(c)
(d)
(e)
(g)
(h)
(i)
(j)
Von Mises stress [MPa]
0
100
200
300
400
500
600
700
800
(k)
Figure 5.7: Von Mises stress contour plots for the 10-grain polycrystalline copper aggregate subjected to tensile load at the last computed step (Γ33 = 2%).
Figs.(a-e) and Figs.(h-j) correspond to two different views of the aggregate. (k)
Colormap.
5.5 Numerical tests
155
(a)
(b)
(c)
(d)
(e)
(f )
(g)
(h)
(i)
(j)
Cumulative slip
0
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10
(k)
Figure 5.8: Cumulative slip contour plots of the 10-grain polycrystalline copper aggregate subjected to tensile load at the last computed step (Γ33 = 2%).
Figs.(a-e) and Figs.(h-j) correspond to two different views of the aggregate. (k)
Colormap.
600
600
100-grain aggregate (I)
100-grain aggregate (II)
linear elastic response
BZ model
500
Macro Stress Σg33 [MPa]
Macro Stress Σ33 [MPa]
500
400
300
80
60
200
40
100
20
400
300
80
60
200
40
100
20
0
0
0
0
0
0.5
1.0
Macro Strain Γ33 [%]
(a)
0.1
1.5
0.2
2.0
0
0
0
0.5
1.0
1.5
2.0
Macro Strain
2.5
Γg33
0.1
3.0
3.5
0.2
4.0
[%]
(b)
Figure 5.9: (a) Volume stress average Σ33 versus volume strain average Γ33
for the two 100-grain aggregates; (b) Averaged stress-strain behaviour for each
constituent grain of the aggregate (I).
156
Grain boundary formulation for crystal plasticity
100-grain aggregate
The same tensile test is performed on two 100-grain Cu aggregates, in order
to evaluate the plastic response of more complex and representative Cu microstructures. Figure (5.9a) shows the macro stress-strain behaviour of the
two aggregates, which are labelled as aggregate (I) and (II) and reported in
Figures (5.10a) and (5.10d), respectively. Figures (5.10b,e) show the corresponding surface meshes, whereas Figure (5.10c,f) show the volume meshes.
The meshes are built using a mesh density parameter that ensures an average
number of 400 volume elements per grain, see Table (5.2), which represents a
good tradeoff between solution accuracy and number of DoFs.
The macro stress-strain behaviour reported in Figure (5.9a) is qualitatively
similar to that obtained for the 10-grain aggregate. The insert of the figure
reports a close-up view of the macro response in the range 0-0.2%, showing
the transition from the linear to the plastic slip behaviour. By comparing the
response of the two microstructures, it can be noted that an aggregate with
100 grains is satisfactorily representative for polycrystalline Cu undergoing
small strains crystal plasticity. The solid line in Figure (5.9a) represents the
linear response of the microstructure whose averaged stiffness is calculated
using the Hashin-Shtrikman bounds [86, 87], whereas the dashed line reports
the response of the aggregate using the Berveiller and Zaoui (BZ) model [36].
It can be noted that the BZ model and the present model agree very well with
respect to the elastic to plastic behavior transition. On the other hand, the
BZ model provides values of stress higher than the present model, but this
appears consistent with Barbe et al. [18], who reported an overestimate with
respect to FEM as well. Figure (5.9b) shows the averaged stress-strain behavior of each of the 100 grains of the aggregates, obtained using Eqs.(5.20) and
(5.21) where the integration is performed over the boundary of each grain.
Figures (5.11a,c,e) show the contour plot of the Von Mises stress over the
external boundary of the aggregate (I) at the macro-strain steps Γ33 = 0.3%,
1.0% and 2.0%, respectively, whereas Figures (5.11b,d,f) report the same contour plot at the same macro-strain steps for a section of the aggregate giving an insight of the internal volume fields distribution. Similarly, Figures
(5.12a,c,e) show the contour plot of the cumulative slip at the macro-strain
steps Γ33 = 0.3%, 1.0% and 2.0%, respectively, and Figures (5.12b,d,f) report
the same contour plot at the same macro-strain steps for a section of the aggregate.
5.5 Numerical tests
157
(a)
(b)
(c)
(d)
(e)
(f )
Figure 5.10: (a,b) 100-grain polycrystalline microstructure (I) and (II), respectively; corresponding (c,d) surface and (e,f ) volume meshes.
158
Grain boundary formulation for crystal plasticity
(a) Γ33 = 0.3%
(b) Γ33 = 0.3%
(c) Γ33 = 1.0%
(d) Γ33 = 1.0%
(e) Γ33 = 2.0%
(f ) Γ33 = 2.0%
Von Mises stress [MPa]
0
100
200
300
400
500
600
700
800
900
(g)
Figure 5.11: Von Mises stress contour plots for the 100-grain polycrystalline
copper aggregate (I) at the macro-strain steps (a,b) Γ33 = 0.3%, (c,d) 1.0% and
(e,f ) 2.0%; (g) Colormap.
5.5 Numerical tests
159
(a) Γ33 = 0.3%
(b) Γ33 = 0.3%
(c) Γ33 = 1.0%
(d) Γ33 = 1.0%
(e) Γ33 = 2.0%
(f ) Γ33 = 2.0%
Cumulative slip
0
0.02
0.04
0.06
0.08
0.10
0.12 0.13
(g)
Figure 5.12: Cumulative slip contour plots for the 100-grain polycrystalline
copper aggregate (I) at the macro-strain steps (a,b) Γ33 = 0.3%, (c,d) 1.0% and
(e,f ) 2.0%; (g) Colormap.
160
Grain boundary formulation for crystal plasticity
In conclusion, a comparison with the homogenisation results obtained using the crystal plasticity finite element method is presented. The reference is
the work by Barbe et al. [18], where the response of polycrystalline aggregates
at small strains is studied using FEM. In their work, the authors employ the
model proposed by Méric and Cailletaud [134], which is briefly recalled in
Section (5.7) for the sake of completeness.
In [18], the FEM aggregate consists of 200 grains with random crystallographic orientations, discretized using a structured mesh and subject to isovolumic boundary conditions. The BEM results are instead obtained with the
polycrystalline aggregate (I) whose morphology, surface and volumes meshes
are reported in Figures (5.10a,c,e), respectively. As in Ref. [18], the BEM polycrystalline aggregate is subjected to iso-volumic boundary conditions and is
loaded along the x3 direction, so that the axial stress and strain correspond to
the averaged macro-stress Σ33 and macro-strain Γ33 , respectively. The lateral
stresses correspond to the averaged macro-stresses Σ11 and Σ22 . The elastic,
slip rate and hardening constants are taken from [18], whereas the interaction hardening moduli from the work of Gérard et al. [73], who adopted such
coefficients based on the work on dislocation dynamic simulations for FCC
crystals of Devincre et al. [57]. All the constants are reported in Table (5.3). It
is worth noting that in the elastic regime, the grains are considered isotropic,
as in [18], and this is why their elastic behaviour is given in terms of Young’s
modulus E and Poisson’s ratio ν.
Figure (5.13) shows the comparison between the results obtained using the
FEM model in [18] and those obtained from the present formulation. In Figure
(5.13), the FEM points are denoted by small circles whereas the BEM points
are denoted by small squares. As shown in the figures, for the strain range
0-0.6%, it is possible to assess the capabilities of the present formulation, as
the plastic behaviour of the polycrystalline aggregate is well developed. The
presented results highlight a satisfactory agreement between the results of the
present formulation and those reported in the literature.
Eventually, Figures (5.14a,c,e) show the contour plot of the Von Mises
stress over the external boundary of the aggregate (I) at the macro-strain steps
Γ33 = 0.14%, 0.30% and 0.59%, respectively, using the present grain boundary formulation and the Méric and Cailletaud model [134]. Similarly, Figures
(5.14b,d,f) show the contour plot of the cumulative slip at the same macrostrain steps.
5.5 Numerical tests
161
Table 5.3: Material parameters for FEM-BEM comparison. The elastic, slip rate
and hardening constants are taken from [18], whereas the interaction hardening moduli from [73]. To interpret the symbols, the reader is referred to
Section (5.7.2).
property
Young’s modulus [109 N/m2 ]
Poisson’s ratio [-]
slip rate constant [106 N · s1/n ]
rate sensitivity [-]
initial critical resolved
shear stress [106 N/m2 ]
hardening constant [106 N/m2 ]
hardening constant [-]
hardening moduli [-]
250
value
169
0.3
10
25
τ0
111
Q
b
self hardening
Hirth lock
Coplanar interaction
Glissile junction
Collinear interaction
Lomer lock
35
7
8.0
0.2
1.0
3.0
90.0
2.5
0
FEM - Σ11
FEM - Σ22
BEM - Σ11
BEM - Σ22
-20
Lateral Stress [MPa]
200
Axial Stress [MPa]
component
E
ν
K
n
150
100
50
FEM
BEM
0
-40
-60
-80
-100
-120
0
0.1
0.2
0.3
0.4
Axial Strain [%]
(a)
0.5
0.6
0
0.1
0.2
0.3
0.4
0.5
0.6
Axial Strain [%]
(b)
Figure 5.13: Comparison between FEM results [18] and the present BEM formulation. (a) Macro stress-strain behaviour of a polycrystalline aggregate subjected to iso-volumic boundary conditions; (b) Lateral macro-stresses versus
axial macro-strain.
162
Grain boundary formulation for crystal plasticity
(a) Γ33 = 0.14%
(b) Γ33 = 0.14%
(c) Γ33 = 0.30%
(d) Γ33 = 0.30%
(e) Γ33 = 0.59%
(f ) Γ33 = 0.59%
Cumulative slip
Von Mises stress [MPa]
0
150
300
(g)
450
600 0
0.01
0.02
0.03
0.04
(h)
Figure 5.14: (a,c,e) Von Mises stress contour plots for the polycrystalline aggregate at the macro-strain steps Γ33 = 0.14%, 0.30% and 0.59%, respectively.
(b,d,f ) Cumulative slip contour plots for the polycrystalline aggregate at the
macro-strain steps Γ33 = 0.14%, 0.30% and 0.59%, respectively. (g,h) Colormaps of (g) Von Mises stress and (h) cumulative slip.
5.6 Discussion and possible developments
5.6
163
Discussion and possible developments
The presented formulation addresses crystal plasticity in a boundary element
framework, both for single crystals and crystal aggregates. From this point of
view, it may be interesting to note that, while the possibility of using boundary elements for crystal plasticity is mentioned in the very exhaustive review
[178], no reference is provided to crystal plasticity boundary elements papers.
The salient feature of the method is its expression in terms of grain-boundary
variables only, namely in terms of inter-granular displacements and tractions.
In previous applications, see e.g. Chapters (1), (3) and (4) and [29, 28, 80],
this aspect ensured simplification in data preparation, as only meshing of the
inter-granular interfaces was required, and a reduction in the number of DoFs
for the analysis, particularly appreciated in polycrystalline problems. In crystal plasticity applications, the data preparation simplification with respect to
other techniques is lost, due to the need of meshing the grains interior volume; however, the system order reduction is maintained, as the volume mesh
does not carry additional degrees of freedom. The volume mesh is in fact only
used to compute the plastic contribution on the right-hand side of Eq.(5.19),
while the number of degrees of freedom with respect to the purely elastic case remains
unchanged. Additionally, it is worth noting that the plastic accumulation just
modifies the right-hand side of Eq.(5.19): such circumstance is particularly
advantageous in terms of numerical solution, as the factorisation of the coefficient matrix in the left-hand side of Eq.(5.19) is performed just once and used
throughout the analysis, thus ensuring numerical effectiveness.
Although the proposed scheme allows a considerable reduction in the
number of DoFs with respect to other volume discretisation schemes, the solution of system (5.19) still represents a formidable computational task, especially in the polycrystalline case, due to the fact that the numerical blocks
associated to each grain in the boundary integral scheme are fully populated.
For such reason, at least in the present implementation, the typical computation time is longer than the typical times with CPFEMs or Fast Fourier Transform based formulations, see e.g. Ref. [117]. An interesting direction of further investigation, from this point of view, could be related to the use of fast
multipoles [125] or hierarchical matrices [25, 31, 34, 27, 138] in conjunction
with iterative solvers to reduce the computational time and further compress
the storage memory requirements of the proposed framework. On the other
hand, the implementation of effective rate-independent schemes [7, 136, 37],
could speed up the iterative convergence of the crystal plastic analysis, avoid-
164
Grain boundary formulation for crystal plasticity
ing the numerical stiffness, and the associated computational costs, associated
to high values of the rate sensitivity n.
In the developed model, a phenomenological description of crystal plasticity has been adopted. However, the model offers flexibility and could be
coupled with more sophisticated crystal plasticity laws, included physically
based approaches.
Another natural extension could be the introduction of different models
of grain interfaces, which at the moment are modelled as perfect and remain
intact during the loading history. Sliding, separation and contact could be
included within the formulation by suitably redefining the interface equations
in (5.19).
The presented formulation has also some limitations, which provide input for further investigation. First of all, the formulation is confined within
the framework of small strains, so that finite lattice rotations are not taken into
account. For this reason, applications to texture evolution problems are not
currently possible. In fact, the present formulation is based on an elasto-plastic
analysis, while crystal plastic finite element formulations for texture evolution
are often based on rigid-plastic approaches, where the initial elastic behaviour
is neglected [24, 93]. The extension of the developed grain-boundary framework to finite strains and the inclusion of finite rotations is not a trivial task
and poses considerable research challenges that are left as future task.
5.7
Flow rules and hardening laws
In the present work two different crystal plasticity phenomenological frameworks have been used in the performed numerical tests: the model by Bassani
and Wu [24] and the model by Méric and Cailletaud [134].
5.7.1
Bassani and Wu model
Bassani and Wu used as flow rule a power law of the form
γ̇α = γ̇0α
τα
τcα
n
sgn (τ α ) ,
∀α = 1, · · · , Ns ,
(5.23)
where γ̇0α and n are material parameters known as reference shear rate and rate
sensitivity of slip respectively, while τcα is the evolving critical stress. Eq.(5.23)
describes, in general, a rate-dependent phenomenon in which all the slip systems are always active with a specific slip rate γ̇α , as long as the corresponding
5.7 Flow rules and hardening laws
165
resolved shear stress is non-zero, so that no distinction is made between active and inactive slip systems. It can be demonstrated that when n → ∞ the
formulation becomes rate-independent.
They also expressed the hardening laws, i.e. the evolution of τcα with respect
to the material state and evolution, as
Ns
τ̇cα =
∑ hαβ
γ̇ β ,
∀α = 1, · · · , Ns ,
(5.24)
β =1
where the self hardening moduli hαα and the latent hardening moduli hαβ were
expressed as
"
β #
h
−
h
γ
s
0
hαα = hs + (h0 − hs ) sech2
· γα
· 1 + ∑ f αβ tanh
α
τs − τ0
γp
β6=α
hαβ = q · hαα ,
β 6= α,
(5.25a)
(5.25b)
where: h0 is the initial hardening modulus; hs is the hardening modulus during easy glide within stage I hardening; τs is the stage I reference stress, where
large plastic flow initiates; τ0α R= τcα (0) is the initial critical resolved shear stress
on the α-th slip system; γα = t |γ̇α |dt; γ p is the amount of slip after which the
interaction between slip systems reaches the peak strength; f αβ is a specific
slip interaction strength; q is a constant latent hardening factor. The model
by Bassani and Wu has been implemented in several FEM formulations, also
for polycrystalline micromechanics [199, 201], often using the implementation
developed by Huang [93].
5.7.2
Méric and Cailletaud model
On the other hand, neglecting the kinematic hardening, Méric and Cailletaud
represented the slip rate γ̇α of the α-th slip system using the following flow
rule
α
|τ | − τcα n
α
γ̇ =
sgn (τ α ) , ∀α = 1, · · · , Ns ,
(5.26)
K
where h·i = max(0, ·) and K and n are material parameters. In addition, in
their model, the hardening law, which comprises self and latent hardening
through the hardening moduli hαβ , has the form
Ns
τcα = τ0 + Q
∑ hαβ
β =1
h
i
1 − exp(−b|γ β |) ,
∀α = 1, · · · , Ns
(5.27)
166
Grain boundary formulation for crystal plasticity
where τ0 is the initial critical resolved shear stress; Q is a material parameter
that, multiplied by the hardening modulus hαβ , represents the maximal increase of the critical shear stress on the α-th slip system due to the slip in the
β-th system; τ0 + Q ∑ β hαβ represents the maximum, or saturated, value that
the critical shear stress on the α-th slip system can reach; eventually, b is a
material parameter that governs the exponential saturation law.
Conclusions
In this thesis, a computational framework for micro-structural modelling of
polycrystalline materials with damage and failure has been developed and
presented. The framework is based on a non-linear multi-domain grain boundary formulation for generally anisotropic materials and has been employed
for modelling the deformation, damage and failure mechanisms of polycrystalline aggregates at the scale of the constituent grains.
In the thesis, several aspects of the polycrystalline problem within a boundary element formulation have been investigated and addressed.
In Chapter (2), a novel unified compact scheme to compute the fundamental solutions of generally anisotropic multi-field materials, as well as their
derivatives up to the desired order, has been presented. The scheme is based
on the relationship between the Rayleigh expansion and the three-dimensional
Fourier representation of a second-order homogenous partial differential operator, which allows to obtain a convergent series representation of the fundamental solutions and their derivatives in terms of spherical harmonics. The
presented expression is easy to implement as it does not require any termby-term differentiation to obtain the derivatives of the fundamental solution.
Furthermore, the coefficients of the series depend on the material properties
only and it has been shown that the technique does not suffer from material
degeneracies, thus making it very attractive for efficient boundary element
implementations. Two isotropic operators, namely Laplace and isotropic elasticity operators, have been analysed as particular cases and it has been shown
that, in such cases, the exact solutions are easily retrieved as the series contain a finite number of terms. Numerical tests have been presented to assess the accuracy, convergence and robustness of the proposed scheme when
isotropic elastic, generally anisotropic elastic, transversely isotropic and generally anisotropic piezo-electric and magneto-electro-elastic materials are considered.
167
168
Grain boundary formulation for crystal plasticity
In Chapter (3), a computational effective framework for micro-structural
analysis of polycrystalline materials, with micro damage and cracks, has been
presented. The framework, based on the grain-boundary formulation of Chapter (1), has been developed for computational effectiveness, with an aim towards affordable three-dimensional modelling. The sources of high computational costs in the original strategy have been identified and suitably treated.
First, the original Voronoi or Laguerre morphologies have been regularised, to
remove the small entities usually inducing excessively refined meshes. Then,
special periodic non-prismatic unit cells have been considered, to remove the
pathological grains often induced by the operations used to generate prismatic periodic RVEs. Eventually, a specific grain meshing strategy has been
adopted, to conjugate mesh effectiveness with formulation simplicity. The
combined effects of such enhancements has been demonstrated to lead to outstanding reductions in terms of overall number of degrees of freedom, with
no compromise over accuracy for both homogenisation and cracking analyses. In turn, this is reflected in remarkable savings in terms of solution time,
especially for micro-damage and micro-cracking simulations. In conclusion,
micro-cracking simulations involving few hundred grains, typically requiring
several days with the original scheme, are now performed in several hours.
In Chapter (4), a numerical formulation for inter- and trans-granular cracking in three-dimensional polycrystalline materials has been presented. The
formulation still retains the advantages of expressing the polycrystalline problem involving trans-granular cracking in terms of grain boundary variables
thus reducing the computational cost of polycrystalline simulations. Transgranular cracking has been modelled by computing the stress in the interior
of the grains, by introducing cohesive interfaces in correspondence of the failing cleavage planes according to a threshold condition and by remeshing the
polycrystalline morphology accordingly. The competition between inter- and
trans-granular cracking has been investigated by changing the parameters of
the cohesive laws used to model the two mechanisms. Some preliminary results of the formulation have been presented showing that, by changing the
ratio between the fracture energies of the two mechanisms, the failure behaviour switch from inter-granular to trans-granular cracking. Future improvements of the presented formulation will include analyses of polycrystalline morphologies with a higher number of grains and different loading
conditions to investigate the effect of the statistical variability of polycrystalline morphologies under different boundary conditions on the competition
between the two considered mechanisms.
5.7 Flow rules and hardening laws
169
In Chapter (5), a boundary element formulation for small strains crystal plasticity in polycrystalline aggregates has been presented. The method,
based on the use of grain-boundary integral equations for the anisotropic
elasto-plastic problem, is formulated in terms of boundary displacements and
tractions, which play the role of primary variables, thus allowing a considerable reduction in the number of DoFs with respect to other volume discretisation techniques. The scheme has been implemented both for single crystals
and polycrystalline aggregates and a general rate-dependent scheme has been
used in the proposed incremental/iterative grain-boundary crystal plasticity
algorithm. Several numerical tests have been performed to assess the accuracy and robustness of the method and the results confirm the potential of the
developed tool. Although the still high computational costs associated with
the analysis of general plastic polycrystalline problems call for further developments and refinements, the presented framework provides an alternative
tool for the analysis of this class of problems.
To conclude, it is worth highlighting that, since the developed framework
is capable of dealing with non-linear behaviours of bulk anisotropic domains
(as in the crystal plasticity study), with non-linear behaviours of the interdomain interfaces (as in the study of inter-granular failure), with on-the-fly
remeshing (as in the trans-granular cracking study), it can be easily extended
to investigate the micro-mechanisms of damage and failure of other classes of
materials whose macroscopic behaviour is strongly influenced by their microstructure, such as fibre-reinforced composites. Furthermore, it was shown
that a broader class of physical behaviours can be modelled using the unified representation of the fundamental solutions obtained in Chapter (2) and
therefore the same framework can be also employed to model the damage and
failure mechanisms of materials showing multi-field coupling.
170
Grain boundary formulation for crystal plasticity
Bibliography
[1] A. Abdollahi and I. Arias. Numerical simulation of intergranular and
transgranular crack propagation in ferroelectric polycrystals. International journal of fracture, 174(1):3–15, 2012.
[2] A. Abdollahi and I. Arias. Three-dimensional simulation of crack propagation in ferroelectric polycrystals: Effect of combined toughening
mechanisms. Acta materialia, 65:106–117, 2014.
[3] H. Abdolvand, M. R. Daymond, and C. Mareau. Incorporation of twinning into a crystal plasticity finite element model: evolution of lattice strains and texture in zircaloy-2. International Journal of Plasticity,
27(11):1721–1738, 2011.
[4] M. Abramowitz and I. A. Stegun. Handbook of mathematical functions:
with formulas, graphs, and mathematical tables, volume 55. Courier Corporation, 1964.
[5] G. S. Adkins. Three-dimensional fourier transforms, integrals of spherical bessel functions, and novel delta function identities. arXiv preprint
arXiv:1302.1830, 2013.
[6] M. H. Aliabadi. The boundary element method: applications in solids and
structures., volume 2. John Wiley & Sons Ltd, England, 2002.
[7] L. Anand and M. Kothari. A computational procedure for rateindependent crystal plasticity. Journal of the Mechanics and Physics of
Solids, 44(4):525–558, 1996.
[8] T. L. Anderson and T. Anderson. Fracture mechanics: fundamentals and
applications. CRC press, 2005.
171
172
BIBLIOGRAPHY
[9] I. Aranson, V. Kalatsky, and V. Vinokur. Continuum field description of
crack propagation. Physical review letters, 85(1):118, 2000.
[10] F. Araújo, E. F. d’Azevedo, and L. J. Gray. Constructing efficient
substructure-based preconditioners for bem systems of equations. Engineering Analysis with Boundary Elements, 35(3):517–526, 2011.
[11] F. C. d. Araújo, E. F. dAzevedo, and L. J. Gray. Boundary-element
parallel-computing algorithm for the microstructural analysis of general composites. Computers & Structures, 88(11):773–784, 2010.
[12] G. B. Arfken, H. J. Weber, and F. E. Harris. Mathematical methods for
physicists: A comprehensive guide. Academic press, 2011.
[13] R. J. Asaro. Micromechanics of crystals and polycrystals. Advances in
applied mechanics, 23:1–115, 1983.
[14] R. J. Asaro and A. Needleman. Overview no. 42 texture development
and strain hardening in rate dependent polycrystals. Acta metallurgica,
33(6):923–953, 1985.
[15] R. J. Asaro and J. Rice. Strain localization in ductile single crystals. Journal of the Mechanics and Physics of Solids, 25(5):309–338, 1977.
[16] S. Aubry and A. Arsenlis. Use of spherical harmonics for dislocation
dynamics in anisotropic elastic media. Modelling and Simulation in Materials Science and Engineering, 21(6):065013, 2013.
[17] P. K. Banerjee and R. Butterfield. Boundary element methods in engineering
science, volume 17. McGraw-Hill London, 1981.
[18] F. Barbe, L. Decker, D. Jeulin, and G. Cailletaud. Intergranular and intragranular behavior of polycrystalline aggregates. part 1: Fe model.
International journal of plasticity, 17(4):513–536, 2001.
[19] F. Barbe, S. Forest, and G. Cailletaud. Intergranular and intragranular behavior of polycrystalline aggregates. part 2: Results. International
Journal of Plasticity, 17(4):537–563, 2001.
[20] F. Barbe and R. Quey. A numerical modelling of 3d polycrystal-topolycrystal diffusive phase transformations involving crystal plasticity.
International Journal of Plasticity, 27(6):823–840, 2011.
BIBLIOGRAPHY
173
[21] G. I. Barenblatt. The formation of equilibrium cracks during brittle fracture. general ideas and hypotheses. axially-symmetric cracks. Journal of
Applied Mathematics and Mechanics, 23(3):622–636, 1959.
[22] G. I. Barenblatt. The mathematical theory of equilibrium cracks in brittle
fracture. Advances in applied mechanics, 7:55–129, 1962.
[23] D. Barnett. The precise evaluation of derivatives of the anisotropic elastic green’s functions. Physica status solidi (b), 49(2):741–748, 1972.
[24] J. L. Bassani and T.-Y. Wu. Latent hardening in single crystals ii. analytical characterization and predictions. Proceedings of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences, 435(1893):21–
41, 1991.
[25] M. Bebendorf. Hierarchical matrices. Springer, 2008.
[26] A. Benallal, C. A. Fudoli, and W. S. Venturini. An implicit bem formulation for gradient plasticity and localization phenomena. International
Journal for Numerical Methods in Engineering, 53(8):1853–1869, 2002.
[27] I. Benedetti and M. Aliabadi. A fast hierarchical dual boundary element
method for three-dimensional elastodynamic crack problems. International journal for numerical methods in engineering, 84(9):1038–1067, 2010.
[28] I. Benedetti and M. Aliabadi. A three-dimensional cohesive-frictional
grain-boundary micromechanical model for intergranular degradation
and failure in polycrystalline materials. Computer Methods in Applied
Mechanics and Engineering, 265:36–62, 2013.
[29] I. Benedetti and M. Aliabadi. A three-dimensional grain boundary
formulation for microstructural modeling of polycrystalline materials.
Computational Materials Science, 67:249–260, 2013.
[30] I. Benedetti and M. Aliabadi. Multiscale modeling of polycrystalline
materials: A boundary element approach to material degradation
and fracture. Computer Methods in Applied Mechanics and Engineering,
289:429–453, 2015.
[31] I. Benedetti, M. Aliabadi, and G. Davi. A fast 3d dual boundary element
method based on hierarchical matrices. International journal of solids and
structures, 45(7):2355–2376, 2008.
174
BIBLIOGRAPHY
[32] I. Benedetti and V. Gulizzi. A novel numerical formulation for crystal plasticity. In Key Engineering Materials, volume 713, pages 317–320.
Trans Tech Publications, 2016.
[33] I. Benedetti, V. Gulizzi, and V. Mallardo. A grain boundary formulation
for crystal plasticity. International Journal of Plasticity, 83:202–224, 2016.
[34] I. Benedetti, A. Milazzo, and M. Aliabadi. A fast dual boundary element method for 3d anisotropic crack problems. International journal for
numerical methods in engineering, 80(10):1356–1378, 2009.
[35] J. G. Berryman. Bounds and self-consistent estimates for elastic constants of random polycrystals with hexagonal, trigonal, and tetragonal
symmetries. Journal of the Mechanics and Physics of Solids, 53(10):2141–
2173, 2005.
[36] M. Berveiller and A. Zaoui. An extension of the self-consistent scheme
to plastically-flowing polycrystals. Journal of the Mechanics and Physics
of Solids, 26(5):325–344, 1978.
[37] M. B. Bettaieb, O. Débordes, A. Dogui, L. Duchêne, and C. Keller. On
the numerical integration of rate independent single crystal behavior at
large strain. International Journal of Plasticity, 32:184–217, 2012.
[38] T. Böhlke, K. Jöchen, O. Kraft, D. Löhe, and V. Schulze. Elastic properties
of polycrystalline microcomponents. Mechanics of Materials, 42(1):11–23,
2010.
[39] M. Bonnet and S. Mukherjee. Implicit bem formulations for usual and
sensitivity problems in elasto-plasticity using the consistent tangent operator concept. International Journal of Solids and Structures, 33(30):4461–
4480, 1996.
[40] M. J. Borden, C. V. Verhoosel, M. A. Scott, T. J. Hughes, and C. M. Landis. A phase-field description of dynamic brittle fracture. Computer
Methods in Applied Mechanics and Engineering, 217:77–95, 2012.
[41] C. A. Brebbia, J. C. F. Telles, and L. Wrobel. Boundary element techniques:
theory and applications in engineering. Springer Science & Business Media,
2012.
BIBLIOGRAPHY
175
[42] F. C. Buroni and M. Denda. Radon–stroh formalism for 3d theory of
anisotropic elasticity. In Advances in Boundary Element Techniques, pages
295–300. EC, Ltd, UK, 2014.
[43] F. C. Buroni and A. Sáez. Three-dimensional green’s function and
its derivative for materials with general anisotropic magneto-electroelastic coupling. Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 466(2114):515–537, 2010.
[44] G. T. Camacho and M. Ortiz. Computational modelling of impact damage in brittle materials. International Journal of solids and structures,
33(20):2899–2938, 1996.
[45] N. Chandra, H. Li, C. Shet, and H. Ghonem. Some issues in the application of cohesive zone models for metal–ceramic interfaces. International
Journal of Solids and Structures, 39(10):2827–2855, 2002.
[46] H. Chen, Y. Jiao, and Y. Liu. Investigating the microstructural effect
on elastic and fracture behavior of polycrystals using a nonlocal lattice
particle model. Materials Science and Engineering: A, 631:173–180, 2015.
[47] A. Cisilino, M. Aliabadi, and J. Otegui. A three-dimensional boundary element formulation for the elastoplastic analysis of cracked bodies.
International journal for numerical methods in engineering, 42(2):237–256,
1998.
[48] J. Clayton and J. Knap. Phase field modeling of directional fracture
in anisotropic polycrystals. Computational Materials Science, 98:158–169,
2015.
[49] J. Clayton and J. Knap. Phase field modeling and simulation of coupled fracture and twinning in single crystals and polycrystals. Computer
Methods in Applied Mechanics and Engineering, 2016.
[50] M. A. Crisfield. A fast incremental/iterative solution procedure that
handles snap-through. Computers & Structures, 13(1-3):55–62, 1981.
[51] A. Crocker, P. Flewitt, and G. Smith. Computational modelling of fracture in polycrystalline materials. International materials reviews, 50(2):99–
125, 2005.
176
BIBLIOGRAPHY
[52] A. M. Cuitino and M. Ortiz. Computational modelling of single crystals.
Modelling and Simulation in Materials Science and Engineering, 1(3):225,
1993.
[53] S. Das, A. Shterenlikht, I. C. Howard, and E. J. Palmiere. A general
method for coupling microstructural response with structural performance. Proceedings of the Royal Society of London A: Mathematical, Physical
and Engineering Sciences, 462(2071):2085–2096, 2006.
[54] P. R. Dawson. Computational crystal plasticity. International Journal of
Solids and Structures, 37(1):115–130, 2000.
[55] P. Dederichs and G. Leibfried. Elastic green’s function for anisotropic
cubic crystals. Physical Review, 188(3):1175, 1969.
[56] Q. Deng, C. G. Li, S. L. Wang, H. Tang, and H. Zheng. A new method to
the treatment of corners in the bem. Engineering Analysis with Boundary
Elements, 37(1):182–186, 2013.
[57] B. Devincre, L. Kubin, and T. Hoc. Physical analyses of crystal plasticity
by dd simulations. Scripta Materialia, 54(5):741–746, 2006.
[58] D. di Caprio, J. Stafiej, G. Luciano, and L. Arurault. 3d cellular automata simulations of intra and intergranular corrosion. Corrosion Science, 112:438–450, 2016.
[59] O. Diard, S. Leclercq, G. Rousselier, and G. Cailletaud. Evaluation of
finite element based analysis of 3d multicrystalline aggregates plasticity: Application to crystal plasticity model identification and the study
of stress and strain fields near grain boundaries. International Journal of
Plasticity, 21(4):691–722, 2005.
[60] H. Ding, Y. He, L. Liu, and W. Ding. Cellular automata simulation of
grain growth in three dimensions based on the lowest-energy principle.
Journal of Crystal Growth, 293(2):489–497, 2006.
[61] H. Ding, A. Jiang, P. Hou, and W. Chen. Green’s functions for twophase transversely isotropic magneto-electro-elastic media. Engineering
Analysis with Boundary Elements, 29(6):551–561, 2005.
[62] K. Döbrich, C. Rau, and C. Krill III. Quantitative characterization of the
three-dimensional microstructure of polycrystalline al-sn using x-ray
BIBLIOGRAPHY
177
microtomography. Metallurgical and Materials Transactions A, 35(7):1953–
1961, 2004.
[63] M. L. Dunn. Electroelastic green’s functions for transversely isotropic
piezoelectric media and their application to the solution of inclusion
and inhomogeneity problems. International Journal of Engineering Science, 32(1):119–131, 1994.
[64] M. L. Dunn and H. Wienecke. Green’s functions for transversely
isotropic piezoelectric solids. International Journal of Solids and Structures, 33(30):4571–4581, 1996.
[65] H. D. Espinosa and P. D. Zavattieri. A grain level model for the study
of failure initiation and evolution in polycrystalline brittle materials.
part i: Theory and numerical implementation. Mechanics of Materials,
35(3):333–364, 2003.
[66] H. D. Espinosa and P. D. Zavattieri. A grain level model for the study of
failure initiation and evolution in polycrystalline brittle materials. part
ii: numerical examples. Mechanics of Materials, 35(3):365–394, 2003.
[67] Z. Fan, Y. Wu, X. Zhao, and Y. Lu. Simulation of polycrystalline structure with voronoi diagram in laguerre geometry based on random
closed packing of spheres. Computational materials science, 29(3):301–308,
2004.
[68] C. Farhat. Implicit parallel processing in structural mechanics. Compt.
Mech. Adv., 2:1–124, 1994.
[69] I. Fredholm. Sur les équations de léquilibre dun corps solide élastique.
Acta Mathematica, 23(1):1–42, 1900.
[70] F. Fritzen, T. Böhlke, and E. Schnack. Periodic three-dimensional mesh
generation for crystalline aggregates based on voronoi tessellations.
Computational Mechanics, 43(5):701–713, 2009.
[71] C. Gandhi and M. Ashby. Overview no. 5: Fracture-mechanism maps
for materials which cleave: Fcc, bcc and hcp metals and ceramics. Acta
Metallurgica, 27(10):1565–1602, 1979.
[72] X.-W. Gao and T. G. Davies. An effective boundary element algorithm
for 2d and 3d elastoplastic problems. International Journal of Solids and
Structures, 37(36):4987–5008, 2000.
178
BIBLIOGRAPHY
[73] C. Gérard, G. Cailletaud, and B. Bacroix. Modeling of latent hardening
produced by complex loading paths in fcc alloys. International Journal of
Plasticity, 42:194–212, 2013.
[74] C. Geuzaine and J.-F. Remacle. Gmsh: A 3-d finite element mesh generator with built-in pre-and post-processing facilities. International Journal
for Numerical Methods in Engineering, 79(11):1309–1331, 2009.
[75] M. Ghajari, L. Iannucci, and P. Curtis. A peridynamic material model for
the analysis of dynamic crack propagation in orthotropic media. Computer Methods in Applied Mechanics and Engineering, 276:431–452, 2014.
[76] L. Gray and E. Lutz. On the treatment of corners in the boundary element method. Journal of Computational and Applied Mathematics,
32(3):369–386, 1990.
[77] M. Groeber, S. Ghosh, M. D. Uchic, and D. M. Dimiduk. A framework
for automated analysis and simulation of 3d polycrystalline microstructures.: Part 1: Statistical characterization. Acta Materialia, 56(6):1257–
1273, 2008.
[78] M. Groeber, B. Haley, M. Uchic, D. Dimiduk, and S. Ghosh. 3d reconstruction and characterization of polycrystalline microstructures using
a fib–sem system. Materials Characterization, 57(4):259–273, 2006.
[79] V. Gulizzi and I. Benedetti. Micro-cracking of brittle polycrystalline materials with initial damage. European Journal of Computational Mechanics,
25(1-2):38–53, 2016.
[80] V. Gulizzi, A. Milazzo, and I. Benedetti. An enhanced grain-boundary
framework for computational homogenization and micro-cracking
simulations of polycrystalline materials. Computational Mechanics,
56(4):631–651, 2015.
[81] V. Gulizzi, A. Milazzo, and I. Benedetti. Fundamental solutions for general anisotropic multi-field materials based on spherical harmonics expansions. International Journal of Solids and Structures, 100:169–186, 2016.
[82] V. Gulizzi, A. Milazzo, and I. Benedetti. A micro-mechanical model for
grain-boundary cavitation in polycrystalline materials. In Key Engineering Materials, volume 665, pages 65–68. Trans Tech Publications, 2016.
BIBLIOGRAPHY
179
[83] T.-S. Han and P. R. Dawson. A two-scale deformation model for polycrystalline solids using a strongly-coupled finite element methodology.
Computer methods in applied mechanics and engineering, 196(13):2029–2043,
2007.
[84] S. Harren, T. C. Lowe, R. Asaro, and A. Needleman. Analysis of largestrain shear in rate-dependent face-centred cubic polycrystals: correlation of micro-and macromechanics. Philosophical Transactions of the
Royal Society of London A: Mathematical, Physical and Engineering Sciences,
328(1600):443–500, 1989.
[85] Z. Hashin. Analysis of Composite Materials - A survey. ASME Journal
of Applied Mechanics, 50:481–505, 1983.
[86] Z. Hashin and S. Shtrikman. On some variational principles in
anisotropic and nonhomogeneous elasticity. Journal of the Mechanics and
Physics of Solids, 10(4):335–342, 1962.
[87] Z. Hashin and S. Shtrikman. A variational approach to the theory of the
elastic behaviour of polycrystals. Journal of the Mechanics and Physics of
Solids, 10(4):343–352, 1962.
[88] E. Heripre, M. Dexet, J. Crepin, L. Gélebart, A. Roos, M. Bornert, and
D. Caldemaison. Coupling between experimental measurements and
polycrystal finite element calculations for micromechanical study of
metallic materials. International Journal of Plasticity, 23(9):1512–1539,
2007.
[89] R. Hill. Generalized constitutive relations for incremental deformation
of metal crystals by multislip. Journal of the Mechanics and Physics of
Solids, 14(2):95–102, 1966.
[90] R. Hill and J. Rice. Constitutive analysis of elastic-plastic crystals at
arbitrary strain. Journal of the Mechanics and Physics of Solids, 20(6):401–
413, 1972.
[91] A. Hlilou, I. B. Naceur, K. Sai, C. Gérard, S. Forest, and G. Cailletaud.
Generalization of the polycrystalline β-model: Finite element assessment and application to softening material behavior. Computational Materials Science, 45(4):1104–1112, 2009.
180
BIBLIOGRAPHY
[92] P.-F. Hou, H.-J. Ding, and J.-Y. Chen. Green’s functions for transversely
isotropic magnetoelectroelastic media. International Journal of Engineering Science, 43(10):826–858, 2005.
[93] Y. Huang. A User-material Subroutine Incroporating Single Crystal Plasticity in the ABAQUS Finite Element Program. Harvard Univ., 1991.
[94] G. Hughes, G. Smith, A. Crocker, and P. Flewitt. An examination of
the linkage of cleavage cracks at grain boundaries. Materials science and
technology, 21(11):1268–1274, 2005.
[95] G. M. Hughes, G. E. Smith, P. E. Flewitt, and A. Crocker. The brittle
fracture of polycrystalline zinc. Proceedings of the Royal Society of London
A: Mathematical, Physical and Engineering Sciences, 463(2085):2129–2151,
2007.
[96] S. Ii, C. Iwamoto, K. Matsunaga, T. Yamamoto, and Y. Ikuhara. Tem
in situ observation of fracture behavior in ceramic materials. Applied
surface science, 241(1):68–74, 2005.
[97] N. Jia, R. L. Peng, Y. Wang, S. Johansson, and P. Liaw. Micromechanical
behavior and texture evolution of duplex stainless steel studied by neutron diffraction and self-consistent modeling. Acta Materialia, 56(4):782–
793, 2008.
[98] M. Joo, D.-W. Suh, J. Bae, and H. Bhadeshia. Role of delamination and
crystallography on anisotropy of charpy toughness in api-x80 steel. Materials Science and Engineering: A, 546:314–322, 2012.
[99] S. R. Kalidindi, C. A. Bronkhorst, and L. Anand. Crystallographic texture evolution in bulk deformation processing of fcc metals. Journal of
the Mechanics and Physics of Solids, 40(3):537–569, 1992.
[100] M. Kamaya and M. Itakura. Simulation for intergranular stress corrosion cracking based on a three-dimensional polycrystalline model. Engineering Fracture Mechanics, 76(3):386–401, 2009.
[101] M. Kamaya, Y. Kawamura, and T. Kitamura. Three-dimensional local
stress analysis on grain boundaries in polycrystalline material. International Journal of Solids and Structures, 44(10):3267–3277, 2007.
BIBLIOGRAPHY
181
[102] T. Kanit, S. Forest, I. Galliet, V. Mounoury, and D. Jeulin. Determination
of the size of the representative volume element for random composites: statistical and numerical approach. International Journal of solids
and structures, 40(13):3647–3679, 2003.
[103] J.-B. Kim and J. W. Yoon. Necking behavior of aa 6022-t4 based on the
crystal plasticity and damage models. International Journal of Plasticity,
73:3–23, 2015.
[104] A. King, G. Johnson, D. Engelberg, W. Ludwig, and J. Marrow. Observations of intergranular stress corrosion cracking in a grain-mapped
polycrystal. Science, 321(5887):382–385, 2008.
[105] M. Kovač and L. Cizelj. Modeling elasto-plastic behavior of polycrystalline grain structure of steels at mesoscopic level. Nuclear engineering
and design, 235(17):1939–1950, 2005.
[106] K. Kozaczek, B. Petrovic, C. Ruud, S. Kurtz, and A. McIlree. Microstructural modelling of grain-boundary stresses in alloy 600. Journal of materials science, 30(9):2390–2400, 1995.
[107] R. Kraft and J. Molinari. A statistical investigation of the effects of
grain boundary properties on transgranular fracture. Acta Materialia,
56(17):4739–4749, 2008.
[108] C. Krill Iii and L.-Q. Chen. Computer simulation of 3-d grain growth
using a phase-field model. Acta materialia, 50(12):3059–3075, 2002.
[109] S. Kumar and W. A. Curtin. Crack interaction with microstructure. Materials today, 10(9):34–44, 2007.
[110] S. Kumar, S. Kurtz, and V. Agarwala. Micro-stress distribution within
polycrystalline aggregate. Acta mechanica, 114(1-4):203–216, 1996.
[111] S. Kumar and S. K. Kurtz. Simulation of material microstructure using a 3d voronoi tesselation: Calculation of effective thermal expansion
coefficient of polycrystalline materials. Acta metallurgica et materialia,
42(12):3917–3927, 1994.
[112] S. Kumar, S. K. Kurtz, J. R. Banavar, and M. Sharma. Properties of
a three-dimensional poisson-voronoi tesselation: A monte carlo study.
Journal of Statistical Physics, 67(3-4):523–551, 1992.
182
BIBLIOGRAPHY
[113] H. Kuramae, Y. Ikeya, H. Sakamoto, H. Morimoto, and E. Nakamachi.
Multi-scale parallel finite element analyses of ldh sheet formability tests
based on crystallographic homogenization method. International Journal
of Mechanical Sciences, 52(2):183–197, 2010.
[114] A. Kuzmin, M. Luisier, and O. Schenk. Fast methods for computing
selected elements of the greens function in massively parallel nanoelectronic device simulations. In F. Wolf, B. Mohr, and D. Mey, editors,
Euro-Par 2013 Parallel Processing, volume 8097 of Lecture Notes in Computer Science, pages 533–544. Springer Berlin Heidelberg, 2013.
[115] H. Lamari, A. Ammar, P. Cartraud, G. Legrain, F. Chinesta, and
F. Jacquemin. Routes for efficient computational homogenization
of nonlinear materials using the proper generalized decompositions.
Archives of Computational methods in Engineering, 17(4):373–391, 2010.
[116] C. Lautensack. Fitting three-dimensional laguerre tessellations to foam
structures. Journal of Applied Statistics, 35(9):985–995, 2008.
[117] R. A. Lebensohn, A. K. Kanjarla, and P. Eisenlohr. An elasto-viscoplastic
formulation based on fast fourier transforms for the prediction of micromechanical fields in polycrystalline materials. International Journal of
Plasticity, 32:59–69, 2012.
[118] V.-G. Lee. Explicit expression of derivatives of elastic green’s functions
for general anisotropic materials. Mechanics Research Communications,
30(3):241–249, 2003.
[119] V.-G. Lee. Derivatives of the three-dimensional green’s function for
anisotropic materials. International Journal of Solids and Structures,
46(18):3471–3479, 2009.
[120] I. Lifshitz and L. Rozentsveig. On the construction of the green’s tensor
for the basic equation of the theory of elasticity of an anisotropic infinite
medium. Zh. Eksp. Teor. Fiz., 17(9):783–791, 1947.
[121] H. Lim, J. Carroll, C. Battaile, T. Buchheit, B. Boyce, and C. Weinberger.
Grain-scale experimental validation of crystal plasticity finite element
simulations of tantalum oligocrystals. International Journal of Plasticity,
60:1–18, 2014.
BIBLIOGRAPHY
183
[122] H. Lim, C. R. Weinberger, C. C. Battaile, and T. E. Buchheit. Application
of generalized non-schmid yield law to low-temperature plasticity in
bcc transition metals. Modelling and Simulation in Materials Science and
Engineering, 21(4):045015, 2013.
[123] L. Lin, X. Wang, and X. Zeng. The role of cohesive zone properties on intergranular to transgranular fracture transition in polycrystalline solids.
International Journal of Damage Mechanics, page 1056789515618732, 2015.
[124] X. Ling, M. Horstemeyer, and G. Potirniche. On the numerical implementation of 3d rate-dependent single crystal plasticity formulations.
International Journal for Numerical Methods in Engineering, 63(4):548–568,
2005.
[125] Y. Liu. Fast multipole boundary element method: theory and applications in
engineering. Cambridge university press, 2009.
[126] W. Ludwig, A. King, P. Reischig, M. Herbig, E. M. Lauridsen,
S. Schmidt, H. Proudhon, S. Forest, P. Cloetens, S. R. Du Roscoat, et al.
New opportunities for 3d materials science of polycrystalline materials
at the micrometre lengthscale by combined use of x-ray diffraction and
x-ray imaging. Materials Science and Engineering: A, 524(1):69–76, 2009.
[127] E. Lutz, W. Ye, and S. Mukherjee. Elimination of rigid body modes from
discretized boundary integral equations. International Journal of Solids
and Structures, 35(33):4427–4436, 1998.
[128] S. Maiti, K. Rangaswamy, and P. H. Geubelle. Mesoscale analysis of
dynamic fragmentation of ceramics under tension. Acta Materialia,
53(3):823–834, 2005.
[129] K. Malen. A unified six-dimensional treatment of elastic green’s functions and dislocations. physica status solidi (b), 44(2):661–672, 1971.
[130] V. Mallardo. Integral equations and nonlocal damage theory: a numerical implementation using the bdem. International journal of fracture,
157(1-2):13–32, 2009.
[131] V. Mallardo and C. Alessandri. Arc-length procedures with bem in
physically nonlinear problems. Engineering Analysis with Boundary Elements, 28(6):547 – 559, 2004.
184
BIBLIOGRAPHY
[132] J. Mandel. Généralisation de la théorie de plasticité de wt koiter. International Journal of Solids and structures, 1(3):273–295, 1965.
[133] C. McMahon. Hydrogen-induced intergranular fracture of steels. Engineering Fracture Mechanics, 68(6):773–788, 2001.
[134] L. Méric and G. Cailletaud. Single crystal modeling for structural calculations: Part 2finite element implementation. Journal of Engineering
Materials and Technology, 113(1):171–182, 1991.
[135] C. Miehe and J. Schröder. A comparative study of stress update algorithms for rate-independent and rate-dependent crystal plasticity. International Journal for Numerical Methods in Engineering, 50(2):273–298, 2001.
[136] C. Miehe, J. Schröder, and J. Schotte. Computational homogenization
analysis in finite plasticity simulation of texture development in polycrystalline materials. Computer methods in applied mechanics and engineering, 171(3):387–418, 1999.
[137] C. Miehe, F. Welschinger, and M. Hofacker. Thermodynamically consistent phase-field models of fracture: Variational principles and multifield fe implementations. International Journal for Numerical Methods in
Engineering, 83(10):1273–1311, 2010.
[138] A. Milazzo, I. Benedetti, and M. Aliabadi. Hierarchical fast bem for
anisotropic time-harmonic 3-d elastodynamics. Computers & Structures,
96:9–24, 2012.
[139] G. Mori, D. Scherer, S. Schwentenwein, and P. Warbichler. Intergranular
stress corrosion cracking of copper in nitrite solutions. Corrosion science,
47(9):2099–2124, 2005.
[140] S. M. T. Mousavi, B. Hosseinkhani, C. Vieillard, M. Chambart, P. J. J.
Kok, and J.-F. Molinari. On the influence of transgranular and intergranular failure mechanisms during dynamic loading of silicon nitride.
Acta Materialia, 67:239–251, 2014.
[141] S. T. Mousavi, N. Richart, C. Wolff, and J.-F. Molinari. Dynamic crack
propagation in a heterogeneous ceramic microstructure, insights from a
cohesive model. Acta Materialia, 88:136–146, 2015.
BIBLIOGRAPHY
185
[142] M. Muñoz-Reja, F. Buroni, A. Sáez, and F. Garcı́a-Sánchez. 3d explicitbem fracture analysis for materials with anisotropic multifield coupling.
Applied Mathematical Modelling, 40(4):2897–2912, 2016.
[143] T. Mura. Micromechanics of defects in solids. Springer Science & Business
Media, 2013.
[144] T. Mura and N. Kinoshita. Green’s functions for anisotropic elasticity.
physica status solidi (b), 47(2):607–618, 1971.
[145] A. Musienko and G. Cailletaud. Simulation of inter-and transgranular
crack propagation in polycrystalline aggregates due to stress corrosion
cracking. Acta materialia, 57(13):3840–3855, 2009.
[146] A. Musienko, A. Tatschl, K. Schmidegg, O. Kolednik, R. Pippan, and
G. Cailletaud. Three-dimensional finite element simulation of a polycrystalline copper specimen. Acta materialia, 55(12):4121–4136, 2007.
[147] E. Nakamachi, N. Tam, and H. Morimoto. Multi-scale finite element
analyses of sheet metals by using sem-ebsd measured crystallographic
rve models. International Journal of Plasticity, 23(3):450–489, 2007.
[148] G. Nakamura and K. Tanuma. A formula for the fundamental solution
of anisotropic elasticity. The Quarterly Journal of Mechanics and Applied
Mathematics, 50(2):179–194, 1997.
[149] C.-W. Nan, M. Bichurin, S. Dong, D. Viehland, and G. Srinivasan. Multiferroic magnetoelectric composites: historical perspective, status, and
future directions. Journal of Applied Physics, 103(3):031101, 2008.
[150] A. Needleman. Computational mechanics at the mesoscale. Acta Materialia, 48(1):105–124, 2000.
[151] S. Nemat-Nasser and M. Hori. Micromechanics: overall properties of heterogeneous materials, volume 37. Elsevier, 2013.
[152] V. P. Nguyen, M. Stroeven, and L. J. Sluys. Multiscale continuous and
discontinuous modeling of heterogeneous materials: a review on recent
developments. Journal of Multiscale Modelling, 3(04):229–270, 2011.
[153] M. Nygårds. Number of grains necessary to homogenize elastic materials with cubic symmetry. Mechanics of Materials, 35(11):1049–1057,
2003.
186
BIBLIOGRAPHY
[154] M. Nygårds and P. Gudmundson. Three-dimensional periodic voronoi
grain models and micromechanical fe-simulations of a two-phase steel.
Computational Materials Science, 24(4):513–519, 2002.
[155] M. Ortiz and A. Pandolfi. Finite-deformation irreversible cohesive elements for three-dimensional crack-propagation analysis. International
Journal for Numerical Methods in Engineering, 44(9):1267–1282, 1999.
[156] E. Pan. Three-dimensional green’s functions in anisotropic magnetoelectro-elastic bimaterials. Zeitschrift für angewandte Mathematik und
Physik ZAMP, 53(5):815–838, 2002.
[157] E. Pan and W. Chen. Static Green’s Functions in Anisotropic Media. Cambridge University Press, 2015.
[158] E. Pan and F. Tonon. Three-dimensional greens functions in anisotropic
piezoelectric solids.
International Journal of Solids and Structures,
37(6):943–958, 2000.
[159] D. Peirce, R. Asaro, and A. Needleman. An analysis of nonuniform
and localized deformation in ductile single crystals. Acta metallurgica,
30(6):1087–1119, 1982.
[160] D. Peirce, R. J. Asaro, and A. Needleman. Material rate dependence and localized deformation in crystalline solids. Acta metallurgica,
31(12):1951–1976, 1983.
[161] M. Peng, Y. Zhang, Y. Liu, M. Song, J. Zhai, and Z. L. Wang. Magneticmechanical-electrical-optical coupling effects in gan-based led/rareearth terfenol-d structures. Advanced Materials, 26(39):6767–6772, 2014.
[162] P.-O. Persson and G. Strang. A simple mesh generator in matlab. SIAM
review, 46(2):329–345, 2004.
[163] A.-V. Phan, L. Gray, and T. Kaplan. On the residue calculus evaluation
of the 3-d anisotropic elastic green’s function. Communications in numerical methods in engineering, 20(5):335–341, 2004.
[164] A.-V. Phan, L. Gray, and T. Kaplan. Residue approach for evaluating
the 3d anisotropic elastic green’s function: multiple roots. Engineering
analysis with boundary elements, 29(6):570–576, 2005.
BIBLIOGRAPHY
187
[165] C. Pinna, Y. Lan, M. Kiu, P. Efthymiadis, M. Lopez-Pedrosa, and D. Farrugia. Assessment of crystal plasticity finite element simulations of the
hot deformation of metals from local strain and orientation measurements. International Journal of Plasticity, 73:24–38, 2015.
[166] C. Prakash, H. Lee, M. Alucozai, and V. Tomar. An analysis of the influence of grain boundary strength on microstructure dependent fracture
in polycrystalline tungsten. International Journal of Fracture, 199(1):1–20,
2016.
[167] R. Quey, P. Dawson, and F. Barbe. Large-scale 3d random polycrystals for the finite element method: Generation, meshing and remeshing.
Computer Methods in Applied Mechanics and Engineering, 200(17):1729–
1745, 2011.
[168] S. Quilici and G. Cailletaud. Fe simulation of macro-, meso-and
micro-scales in polycrystalline plasticity. Computational materials science,
16(1):383–390, 1999.
[169] V. Raja and T. Shoji. Stress corrosion cracking: theory and practice. Elsevier,
2011.
[170] S. I. Ranganathan and M. Ostoja-Starzewski.
Scaling function,
anisotropy and the size of rve in elastic random polycrystals. Journal
of the Mechanics and Physics of Solids, 56(9):2773–2791, 2008.
[171] Z.-Y. Ren and Q.-S. Zheng. A quantitative study of minimum sizes of
representative volume elements of cubic polycrystalsnumerical experiments. Journal of the Mechanics and Physics of Solids, 50(4):881–893, 2002.
[172] Z.-Y. Ren and Q.-S. Zheng. Effects of grain sizes, shapes, and distribution on minimum sizes of representative volume elements of cubic
polycrystals. Mechanics of Materials, 36(12):1217–1229, 2004.
[173] J. R. Rice. Mathematical analysis in the mechanics of fracture. Fracture:
an advanced treatise, 2:191–311, 1968.
[174] J. R. Rice. Inelastic constitutive relations for solids: an internal-variable
theory and its application to metal plasticity. Journal of the Mechanics and
Physics of Solids, 19(6):433–455, 1971.
188
BIBLIOGRAPHY
[175] J. Riedle, P. Gumbsch, and H. F. Fischmeister. Cleavage anisotropy in
tungsten single crystals. Physical review letters, 76(19):3594, 1996.
[176] H. Ritz and P. Dawson. Sensitivity to grain discretization of the simulated crystal stress distributions in fcc polycrystals. Modelling and Simulation in Materials Science and Engineering, 17(1):015001, 2008.
[177] C. Robert and C. Mareau. A comparison between different numerical
methods for the modeling of polycrystalline materials with an elastic–
viscoplastic behavior. Computational Materials Science, 103:134–144, 2015.
[178] F. Roters, P. Eisenlohr, L. Hantcherli, D. D. Tjahjanto, T. R. Bieler, and
D. Raabe. Overview of constitutive laws, kinematics, homogenization and multiscale methods in crystal plasticity finite-element modeling: Theory, experiments, applications. Acta Materialia, 58(4):1152–1211,
2010.
[179] D. Rowenhorst, A. Lewis, and G. Spanos. Three-dimensional analysis
of grain topology and interface curvature in a β-titanium alloy. Acta
Materialia, 58(16):5511–5519, 2010.
[180] C. H. Rycroft. Voro++: A three-dimensional Voronoi cell library in C++.
Chaos, 19:041111, 2009.
[181] M. A. Sales and L. Gray. Evaluation of the anisotropic green’s function
and its derivatives. Computers & structures, 69(2):247–254, 1998.
[182] G. Sarma and B. Radhakrishnan. Modeling microstructural effects on
the evolution of cube texture during hot deformation of aluminum. Materials Science and Engineering: A, 385(1):91–104, 2004.
[183] G. Sarma, B. Radhakrishnan, and T. Zacharia. Finite element simulations of cold deformation at the mesoscale. Computational Materials Science, 12(2):105–123, 1998.
[184] N. Schclar. Anisotropic analysis using boundary elements, volume 20. Computational mechanics, 1994.
[185] O. Schenk, M. Bollhöfer, and R. A. Römer. On large-scale diagonalization techniques for the anderson model of localization. SIAM Rev.,
50(1):91–112, Feb. 2008.
BIBLIOGRAPHY
189
[186] O. Schenk, A. Wchter, and M. Hagemann. Matching-based preprocessing algorithms to the solution of saddle-point problems in large-scale
nonconvex interior-point optimization. Computational Optimization and
Applications, 36(2-3):321–341, 2007.
[187] Y. Schneider, A. Bertram, and T. Böhlke. Three-dimensional simulation
of local and global behaviour of αfe-cu composites under large plastic
deformation. Technische Mechanik, 33(1):34–51, 2013.
[188] Y. Schneider, A. Bertram, T. Böhlke, and C. Hartig. Plastic deformation
behaviour of fe–cu composites predicted by 3d finite element simulations. Computational Materials Science, 48(3):456–465, 2010.
[189] J. Schöberl. Netgen an advancing front 2d/3d-mesh generator based on
abstract rules. Computing and visualization in science, 1(1):41–52, 1997.
[190] G. Sfantos and M. Aliabadi. A boundary cohesive grain element formulation for modelling intergranular microfracture in polycrystalline
brittle materials. International journal for numerical methods in engineering, 69(8):1590–1626, 2007.
[191] G. Sfantos and M. Aliabadi. Multi-scale boundary element modelling
of material degradation and fracture. Computer Methods in Applied Mechanics and Engineering, 196(7):1310–1329, 2007.
[192] P. Shanthraj, B. Svendsen, L. Sharma, F. Roters, and D. Raabe. Elastoviscoplastic phase field modelling of anisotropic cleavage fracture. Journal of the Mechanics and Physics of Solids, 2016.
[193] C. Shet and N. Chandra. Analysis of energy balance when using cohesive zone models to simulate fracture processes. Journal of engineering
materials and technology, 124(4):440–450, 2002.
[194] J. R. Shewchuk. Triangle: Engineering a 2D Quality Mesh Generator and
Delaunay Triangulator. In M. Lin and D. Manocha, editors, Applied Computational Geometry: Towards Geometric Engineering, volume 1148 of Lecture Notes in Computer Science, pages 203–222. Springer-Verlag, Berlin,
1996. (From the First ACM Workshop on Applied Computational Geometry.).
190
BIBLIOGRAPHY
[195] Y. Shiah, C. Tan, and V. Lee. Evaluation of explicit-form fundamental
solutions for displacements and stresses in 3d anisotropic elastic solids.
CMES: Computer Modeling in Engineering & Sciences, 34(3):205–226, 2008.
[196] Y. Shiah, C. Tan, and C. Wang. Efficient computation of the green’s function and its derivatives for three-dimensional anisotropic elasticity in
bem analysis. Engineering Analysis with Boundary Elements, 36(12):1746–
1755, 2012.
[197] A. Shterenlikht and L. Margetts. Three-dimensional cellular automata
modelling of cleavage propagation across crystal boundaries in polycrystalline microstructures. Proceedings of the Royal Society of London
A: Mathematical, Physical and Engineering Sciences, 471(2177):20150039,
2015.
[198] H. Si. Tetgen, a delaunay-based quality tetrahedral mesh generator.
ACM Transactions on Mathematical Software (TOMS), 41(2):11, 2015.
[199] I. Simonovski and L. Cizelj. Automatic parallel generation of finite element meshes for complex spatial structures. Computational Materials
Science, 50(5):1606–1618, 2011.
[200] I. Simonovski and L. Cizelj. Computational multiscale modeling of intergranular cracking. Journal of Nuclear Materials, 414(2):243–250, 2011.
[201] I. Simonovski and L. Cizelj. Towards modeling intergranular stress
corrosion cracks on grain size scales. Nuclear Engineering and Design,
246:107–114, 2012.
[202] I. Simonovski and L. Cizelj. Cohesive zone modeling of intergranular
cracking in polycrystalline aggregates. Nuclear Engineering and Design,
283:139–147, 2015.
[203] G. Smith, A. Crocker, R. Moskovic, and P. Flewitt. Models to describe
brittle and ductile fracture in ferritic steels. Philosophical Magazine A,
82(17-18):3443–3453, 2002.
[204] A. K. Soh, J.-x. Liu, and K. H. Hoon. Three-dimensional green’s functions for transversely isotropic magnetoelectroelastic solids. International Journal of Nonlinear Sciences and Numerical Simulation, 4(2):139–148,
2003.
BIBLIOGRAPHY
191
[205] A. Standard. E112: Standard test methods for determining average
grain size. West Conshocken, 1996.
[206] I. Steinbach. Phase-field models in materials science. Modelling and Simulation in Materials Science and Engineering, 17(7):073001, 2009.
[207] N. Sukumar, D. Srolovitz, T. Baker, and J.-H. Prévost. Brittle fracture in polycrystalline microstructures with the extended finite element method. International Journal for Numerical Methods in Engineering,
56(14):2015–2037, 2003.
[208] J. L. Synge. The hypercircle in mathematical physics. University Press Cambridge, 1957.
[209] C. Tan, Y. Shiah, and C. Wang. Boundary element elastic stress analysis
of 3d generally anisotropic solids using fundamental solutions based on
fourier series. International Journal of Solids and Structures, 50(16):2701–
2711, 2013.
[210] G. Taylor and C. Elam. The plastic extension and fracture of aluminium
crystals. Proceedings of the Royal Society of London. Series A, Containing
Papers of a Mathematical and Physical Character, 108(745):28–51, 1925.
[211] G. I. Taylor and C. F. Elam. Bakerian lecture. the distortion of an aluminium crystal during a tensile test. Proceedings of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences, 102(719):643–
667, 1923.
[212] G. I. Taylor and H. Quinney. The plastic distortion of metals. Philosophical Transactions of the Royal Society of London. Series A, Containing Papers
of a Mathematical or Physical Character, 230:323–362, 1932.
[213] J. Telles and C. Brebbia. The boundary element method in plasticity.
Applied Mathematical Modelling, 5(4):275–281, 1981.
[214] K. Terada, M. Hori, T. Kyoya, and N. Kikuchi. Simulation of the multiscale convergence in computational homogenization approaches. International Journal of Solids and Structures, 37(16):2285–2311, 2000.
[215] T. Ting and V.-G. Lee. The three-dimensional elastostatic green’s function for general anisotropic linear elastic solids. The Quarterly Journal of
Mechanics and Applied Mathematics, 50(3):407–426, 1997.
192
BIBLIOGRAPHY
[216] V. Tomar, J. Zhai, and M. Zhou. Bounds for element size in a variable
stiffness cohesive finite element model. International journal for numerical
methods in engineering, 61(11):1894–1920, 2004.
[217] V. Tvergaard. Effect of fibre debonding in a whisker-reinforced metal.
Materials science and engineering: A, 125(2):203–213, 1990.
[218] C. Verhoosel and M. Gutiérrez. Modelling inter-and transgranular
fracture in piezoelectric polycrystals. Engineering Fracture Mechanics,
76(6):742–760, 2009.
[219] G. Voronoı̈. Nouvelles applications des paramètres continus à la théorie
des formes quadratiques. deuxième mémoire. recherches sur les parallélloèdres primitifs. Journal für die reine und angewandte Mathematik,
134:198–287, 1908.
[220] C. Wang. Elastic fields produced by a point source in solids of general
anisotropy. Journal of Engineering Mathematics, 32(1):41–52, 1997.
[221] D. Wang, J. Zhao, Y. Zhou, X. Chen, A. Li, and Z. Gong. Extended
finite element modeling of crack propagation in ceramic tool materials
by considering the microstructural features. Computational Materials Science, 77:236–244, 2013.
[222] H. Wang, P. Wu, J. Wang, and C. Tomé. A crystal plasticity model
for hexagonal close packed (hcp) crystals including twinning and detwinning mechanisms. International Journal of Plasticity, 49:36–52, 2013.
[223] X. Wang and Y.-p. Shen. The general solution of three-dimensional
problems in magnetoelectroelastic media. International Journal of Engineering Science, 40(10):1069–1080, 2002.
[224] T. Watanabe and S. Tsurekawa. The control of brittleness and development of desirable mechanical properties in polycrystalline systems by
grain boundary engineering. Acta Materialia, 47(15):4171–4185, 1999.
[225] Y. Wei and L. Anand. Grain-boundary sliding and separation in polycrystalline metals: application to nanocrystalline fcc metals. Journal of
the Mechanics and Physics of Solids, 52(11):2587–2616, 2004.
[226] R. O. Wells. Differential analysis on complex manifolds, volume 65.
Springer Science & Business Media, 2007.
BIBLIOGRAPHY
193
[227] R. Wilson and T. Cruse. Efficient implementation of anisotropic three
dimensional boundary-integral equation stress analysis. International
Journal for Numerical Methods in Engineering, 12(9):1383–1397, 1978.
[228] L. C. Wrobel. The boundary element method, applications in thermo-fluids
and acoustics, volume 1. John Wiley & Sons, 2002.
[229] K.-C. Wu. Generalization of the stroh formalism to 3-dimensional
anisotropic elasticity. Journal of elasticity, 51(3):213–225, 1998.
[230] P. Wu, K. Neale, and E. Van der Giessen. Simulation of the behaviour of
fcc polycrystals during reversed torsion. International Journal of Plasticity, 12(9):1199–1219, 1996.
[231] T.-Y. Wu, J. L. Bassani, and C. Laird. Latent hardening in single crystals
i. theory and experiments. Proceedings of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, 435(1893):1–19, 1991.
[232] L. Xie, C. Hwu, and C. Zhang. Advanced methods for calculating
green’s function and its derivatives for three-dimensional anisotropic
elastic solids. International Journal of Solids and Structures, 80:261–273,
2016.
[233] L. Xie, C. Zhang, J. Sladek, and V. Sladek. Unified analytical expressions of the three-dimensional fundamental solutions and their
derivatives for linear elastic anisotropic materials. Proceedings of the
Royal Society of London A: Mathematical, Physical and Engineering Sciences,
472(2186):20150272, 2016.
[234] X.-P. Xu and A. Needleman. Void nucleation by inclusion debonding in
a crystal matrix. Modelling and Simulation in Materials Science and Engineering, 1(2):111, 1993.
[235] J. Yvonnet and Q.-C. He. The reduced model multiscale method
(r3m) for the non-linear homogenization of hyperelastic media at finite
strains. Journal of Computational Physics, 223(1):341–368, 2007.
[236] A. Zeghadi, S. Forest, A.-F. Gourgues, and O. Bouaziz. Ensemble averaging stress–strain fields in polycrystalline aggregates with a constrained surface microstructure–part 2: Crystal plasticity. Philosophical
Magazine, 87(8-9):1425–1446, 2007.
194
BIBLIOGRAPHY
[237] W. Zeng, J. Larsen, and G. Liu. Smoothing technique based crystal plasticity finite element modeling of crystalline materials. International Journal of Plasticity, 65:250–268, 2015.
[238] J. Zhai, V. Tomar, and M. Zhou. Micromechanical simulation of dynamic
fracture using the cohesive finite element method. Journal of engineering
materials and technology, 126(2):179–191, 2004.
[239] J. Zhai and M. Zhou. Finite element analysis of micromechanical failure
modes in a heterogeneous ceramic material system. International Journal
of fracture, 101(1-2):161–180, 2000.
[240] C. Zhang, M. Enomoto, A. Suzuki, and T. Ishimaru. Characterization of
three-dimensional grain structure in polycrystalline iron by serial sectioning. Metallurgical and Materials Transactions A, 35(7):1927–1933, 2004.
[241] C. Zhang, H. Li, P. Eisenlohr, W. Liu, C. Boehlert, M. Crimp, and
T. Bieler. Effect of realistic 3d microstructure in crystal plasticity finite
element analysis of polycrystalline ti-5al-2.5 sn. International Journal of
Plasticity, 69:21–35, 2015.
[242] K. Zhang, B. Holmedal, O. S. Hopperstad, S. Dumoulin, J. Gawad,
A. Van Bael, and P. Van Houtte. Multi-level modelling of mechanical
anisotropy of commercial pure aluminium plate: crystal plasticity models, advanced yield functions and parameter identification. International
Journal of Plasticity, 66:3–30, 2015.
[243] Y. Zhao and R. Tryon. Automatic 3-d simulation and micro-stress distribution of polycrystalline metallic materials. Computer Methods in Applied
Mechanics and Engineering, 193(36):3919–3934, 2004.
[244] Z. Zhao, S. Kuchnicki, R. Radovitzky, and A. Cuitino. Influence of ingrain mesh resolution on the prediction of deformation textures in fcc
polycrystals by crystal plasticity fem. Acta materialia, 55(7):2361–2373,
2007.
[245] T. Zhou, C. Huang, H. Liu, J. Wang, B. Zou, and H. Zhu. Crack propagation simulation in microstructure of ceramic tool materials. Computational Materials Science, 54:150–156, 2012.
[246] T. I. Zohdi and P. Wriggers. An introduction to computational micromechanics. Springer Science & Business Media, 2008.
| 5cs.CE
|
INsight: A Neuromorphic Computing System for
Evaluation of Large Neural Networks
Taehwan Shin
Department of Electronic Engineering
Incheon National University
Incheon, Korea 406-772
arXiv:1508.01008v1 [cs.NE] 5 Aug 2015
Jaeyong Chung
Department of Electronic Engineering
Incheon National University
Incheon, Korea 406-772
[email protected]
Yongshin Kang
Department of Electronic Engineering
Incheon National University
Incheon, Korea 406-772
Abstract
Deep neural networks have been demonstrated impressive results in various cognitive tasks such as object detection and image classification. In order to execute
large networks, Von Neumann computers store the large number of weight parameters in external memories, and processing elements are timed-shared, which
leads to power-hungry I/O operations and processing bottlenecks. This paper describes a neuromorphic computing system that is designed from the ground up
for the energy-efficient evaluation of large-scale neural networks. The computing
system consists of a non-conventional compiler, a neuromorphic architecture, and
a space-efficient microarchitecture that leverages existing integrated circuit design methodologies. The compiler factorizes a trained, feedforward network into
a sparsely connected network, compresses the weights linearly, and generates a
time delay neural network reducing the number of connections. The connections
and units in the simplified network are mapped to silicon synapses and neurons.
We demonstrate an implementation of the neuromorphic computing system based
on a field-programmable gate array that performs the MNIST hand-written digit
classification with 97.64% accuracy.
1
Introduction
Large-scale neural networks such as deep convolutional neural networks have been shown state-ofthe-art results on various tasks in computer vision, automatic speech recognition, natural language
processing, and audio recognition, and on certain tasks, their performance has become comparable
to humans [1]. However, these networks contain a huge number of parameters (e.g., 108 [2]), and the
evaluation of such models is computationally expensive. This makes it problematic to deploy these
models to embedded platforms, where computing power, memory, storage, and energy are limited.
It is also problematic to evaluate the large models at the server side. For example, processing images
and videos uploaded by millions of users require massive amounts of computation, and running a
data center that supports such computation carries enormous costs including cooling expenses and
electricity bills.
The evaluation of large neural networks can be accelerated using general-purpose graphical processing units (GPGPU), and several custom accelerators have been proposed as well [3, 4]. However,
all these processing units are based on Von Neumann architecture, and the large amount of param1
eters are stored in high-density external memories. The traditional architecture requires inefficient
data transfer between parameter-storing memories and processing units, leading to high-energy consumption and processing bottleneck [5].
As neural networks are models inspired by brains, it is natural to execute them in the processing units that mimic brains. Neuromorphic architectures emulate biological neural networks, and
they locate memory and processing elements in a close proximity. This type of architectures has
potential for eliminating the issues in the today’s computers [6]. However, mixed analog-digital
neuromorphic systems such as BrainScaleS [7] and Neurogrid [8] are not designed to run real-world
applications, but aim to simulate the brain for study on computational neuroscience. They usually
adopt biologically plausible neuron models. The neuron models at a low-level of abstractions are
considered expressive and versatile as biological neurons do, but it comes at an implementation cost
such as complexity, space, and energy. To run specific applications, certain details are not necessary.
TrueNorth [9, 10, 11, 12], the neuromorphic architecture from IBM, aims to balance the computational capability and the cost judiciously, and adopts digital spiking neurons based on the leaky
integrate-and-fire model. A recently unveiled chip [6] based on TrueNorth contains 4096 customdesigned cores and employs asynchronous circuits for low-power operations. The chip is used for
both the brain simulation and cognitive applications [13].
This paper presents a novel neuromorphic computing system that is designed solely for the execution
(i.e., inference or testing) of deep neural networks. To the best of our knowledge, this is the first
neuromorphic computing system that stores all the weight parameters on-chip by compressing neural
networks. Designing a new computer architecture is a complex task and has many trade-offs. Our
design philosophy to deal with it is summarized in the next section.
2
Design Considerations
The performance of neural networks such as classification accuracy is extremely valuable. As will
be shown later, in some way, even 1% degradation of the performance is translated into a huge
saving of the implementation cost, and we do not want to sacrifice the performance in other ways.
The performance of learning algorithms for spiking neurons such as spike timing dependent plasticity (STDP) [14] is not close to the state of the art yet [15]. While non-spiking neural networks
trained using state-of-the-art algorithms can be converted into spiking neural networks, constraints
arisen from the hardware architecture and inefficiency in information coding could degrade the performance [9].
We desire to build an “artificial” “brain”. How much should we learn from brains and what should
we throw away from the architecture of existing processors? This may be the central question in
designing neuromorphic architectures. We take a moderate approach to this problem. As much
as we want to learn from the nature, we also desire to take advantage of the results of humans’
effort over several decades. Since we still make the artificial brain into a silicon chip, we need to
continue to use techniques optimized for structures of silicon chips. Until today, the man-made
information processing device at the lowest power consumption is application-specific integrated
circuits (ASICs) and, for general purpose, field programmable gate arrays (FPGAs). Thus, we
leverage the ASIC/FPGA design methodology, and balance the ASIC/FPGA architecture and the
structure of brains judiciously.
Our neuromorphic computing system aims at the evaluation (i.e., the forward-execution) of neural
network models and is not intended to train the models. Training a neural network is computationally more expensive than the forward-execution, and there have been research efforts to accelerate
the training using FPGAs [16, 17]. Our main objective is low-power signal processing rather than
high-performance computing, although it can provide high performance in power-limited environments including data centers. For various applications such as object recognition, a model can be
trained using a cluster of CPU or GPU-based machines, and copies of the model can provide desired
functionality without modification. We also do not consider on-line learning. STDP facilitates efficient on-line learning, but as above-mentioned, it does not provide the state-of-the-art performance.
Our neuromorphic system is digital. Digital systems use a high level of abstraction on signals,
which allows us to deal with complexity and makes it easier to ensure the correctness of man-made
2
ag
a
af
bg
bf
be
a
a
ae
Factorization
a
b
b
cg
cf
ce
b
c
g
f
e
Rolling,
Memorization
e
b
f
g
c
c
c
Figure 1: A fully connected feedforward network of a single layer is transformed into a sparsely
connected network with two layers by factorization. The number of parameters is reduced from 9
to 6 by factorization, but the number of connections rather increases from 9 to 12. However, when
it is converted into a TDNN, the number of connections with parameters becomes 6. The reduction
of the connections is critical in neuromorphic architectures, where each connection uses a dedicated
computational resource called the synapse.
systems. While analog silicon neurons can provide a higher integration density than that of digital
neurons, we make it digital for those benefits.
The rest of the paper is organized as follows. Section 3 describes the motivation of our work.
Section 4, 5, and 6 propose the compiler, the neuromorphic architecture, and the microarchitecture of
our computing system, respectively. Section 7 shows experimental results, and Section 8 concludes
the paper.
3
Motivation
Each unit (node) in feedforward neural networks mainly performs weighted sum operations, and the
functionality can be easily implemented in a hardware module using logic multipliers, logic adders,
and weight-storing registers. For each unit in the model, we create an instance of the module and
these instances are interconnected following the model, yielding a circuit design. However, this type
of systems has not built until today for the following reasons. 1) Neural networks for interesting
applications are huge. Recent deep neural network architectures contain millions and billions of
parameters. Each parameter at least requires a logic multiplier and a register to store the parameter, and we cannot integrate such a high number of multipliers and registers in a chip. For this
reason, existing accelerators of neural networks store parameters in high-density external memories
and equip with a small number of processing elements. The external memory access draws much
power, and the limited memory I/O bandwidth becomes a processing bottleneck. Also, the processing elements are time-shared, which necessitates complex scheduling to maximize the utilization
and creates another processing bottleneck. 2) Neural networks are densely connected unlike logic
(circuit) networks. Densely interconnected designs cause routing congestion issues in the conventional ASIC/FPGA design methodology. For this reason, existing neuromorphic architectures [18]
rely on custom-designed crossbars, where inputs and outputs are fully connected. 3) Silicon chips
have limited I/O bandwidth and all the inputs of neural networks are usually not available at the
same time. For example, traditional cameras transfer each pixel at a time, but neural network models for processing images take all the pixels of a whole image as input in parallel. Thus, the direct
realization of neural network models results in systems with very low utilization.
Fortunately, weight parameters in neural networks are heavily redundant [19], i.e., these networks
are highly over-parameterized. This is necessary to obtain a high quality model during training by
providing much flexibility to learning algorithms, but for the forward-execution, such a high number
of parameters are not necessary. Our neuromorphic architecture is mainly motivated by the fact that
the high redundancy in weight parameters can be exploited to minimize the implementation cost.
However, it is not enough just to compress the weight parameters linearly. We also want the result
of the compression to be represented in a neural network again for the execution in neuromorphic
architectures. In this sense, we simplify neural networks. Surprisingly, simplifying neural networks
using factorization for the first issue handles the remaining two issues as well. A motivational
example is depicted in Fig. 1. The feedforward network can be considered a 2D filter. Since the
3
2D kernel is separable, it can be decomposed into a vertical filter and a horizontal filter. Using
the decomposition, the single-layer network is factorized into a deeper, sparse network, providing
a solution to the second issue. Besides, the weight-shared network can be “rolled” to a network
with fewer units and connections. This process often require adding delay units for memorizing a
previous state of a unit, resulting in a time delay neural network (TDNN), which is a special case
of biologically plausible recurrent networks. This TDNN takes input serially and allows us to timeshare resources, dealing with the last issue. Any 2D filter can be decomposed into a set of separable
filters using singular value decomposition. We extend this to the bank of 3D filters for typical deep
neural networks.
4
Neural Network Compiler
In a convolutional neural network for computer vision tasks, the weights in each convolutional layer
are typically represented in a 4th-order tensor. The weights in the first fully-connected layer can also
be represented in a 4th-order tensor. Let χ ∈ RC×Kh ×Kw ×F be the weight tensor of a convolutional
layer (a fully-connected layer), where C is the number of input channels, Kh and Kw are the height
and width of the kernel (the input map), respectively, and F is the number of output feature maps.
Applying a tensor decomposition method to the weight tensor, a layer can be factorized into 5 sublayers. We denote the weight tensors of the sub-layers by χc , χv , χh , χg , and χi , respectively, and
we represent the weights of each sub-layer in the same format as those of the original layer. The
sub-layers are:
• Channel filter layer: Each unit in this layer is connected to the input units at the same
vertical and horizontal position across channels. χc ∈ RC×1×1×Rc .
• Vertical filter layer: Each unit in this layer belongs to a channel and is connected to the
inputs at the same vertical position. χv ∈ RRc ×Kh ×1×Rv .
• Horizontal filter layer: As in the vertical filter layer, each unit acts on a channel and is
connected to the inputs at the same horizontal position. χh ∈ RRv ×1×Kw ×Rv .
• Code generation layer: As in the first layer, each unit in this layer acts on the same
position across channels. For each position, a code for the original output is generated.
χg ∈ RRv ×1×1×Rf .
• Inverse transform layer: As in the first layer, each unit in this layer acts on the same
position across channels. For each position, the code is converted into the original output.
χi ∈ RRf ×1×1×F .
If the original layer is convolutional, the sublayers also become convolutional. This factorization
makes the network deeper and less dense. We use our own tensor decomposition method, but one can
use CP decomposition referring to a recent preprint [20]. In this case, the code generation layer is not
needed, and the horizontal filter layer generates the codes directly, and the rank of CP decomposition
equals Rc = Rv = Rf . After the decomposition, we perform low-rank approximation, greatly
simplifying the network. In the case of CP decomposition, a simplified network can be obtained
directly. Even for a very high degree of approximation (e.g., 90% error), the performance of the
network is degraded marginally [20]. However, for greater simplification, we fine-tune the factored
network after an aggressive approximation, recovering the original performance.
Then, we construct a TDNN corresponding to a factored layer directly. Let I ∈ RC×H×W be
an input signal where C is the number of input maps, and H and W are the height and width of
the maps, respectively. We first represent the input signal I as a set of C sequences sz (t), z =
0, ..., C − 1, such that
sz (t) = I(z, y, x), t = yW + x,
(1)
for y = 0, ..., H − 1, x = 0, ..., W − 1. We require W (N − 1) delay units to obtain N contiguous
data points along the vertical axis from an input sequence, and N − 1 delay units in the case of the
horizontal axis. In the TDNN, the total number of delay units is W (Kh − 1)Rc + (Kw − 1)Rv .
The delay for the first valid output is W (Kh − 1) + Kw − 1. An example of the TDNNs and its
approximation are shown in Figure 2.
4
(a) 0% synaptic connection reduction with
0% error
(b) 61.11% synaptic connection reduction
with 50.55% error
Figure 2: We randomly generate a single layer network. This network can be considered a feature
detector that takes 2 input maps of size 2 × 3 and produces 3 feature maps. The positive-weight
connections are colored with red. The negative-weight connections are colored with blue. The
darker the color is, the larger the absolute value of the weight is. The connections with weight of 1
are colored with black.
5
Neuromorphic Architecture
We describe our neuromorphic architecture briefly. Since it is neuromorphic, the architecture may be
familiar to readers already. The proposed architecture uses the second-generation neuron model [21]
and consists of synapses, neurons, and delay elements, a special type of short-term memories.
Synapse and neurons are processing elements. Each synapse stores a real-valued weight and performs the synaptic operation that multiplies the incoming input value by the weight. Each neuron
produces the sum of the output values of the connected synapses, and it can add a bias and evaluate
an activation function such as the sigmoid and rectified linear functions. We call the connections
with the weight not equal to one the synaptic connection. Unlike the traditional computers, a processing element is not timed-shared for the operations for different synaptic connections; a synapse
is dedicated to the operations for a synaptic connection. Thus, for the execution, each synaptic
connection is mapped, not scheduled, to a synapse, and synaptic connections and synapses have a
one-to-one correspondence. Each delay unit is also mapped into a delay element.
6
Microarchitecture
From neuron
at previous sublayer
n cycles
Neuron
i0
Synapse
Delay element
i1
pp
w
n
1
pp
w
n
1
To
neuron
at next
sublayer
Pipe
Registers
Delay element
ik-1
n cycles
2n cycles
clk
input
m fractional
ø0
phase
synapse
output
neuron_out
@sublayer 0
synapse_in
@sublayer 1
sublayer 0
valid
sublayer 1
valid
2m fractional part
2m fractional part
m fractional
ø1
ø0
ø1
2m fractional part
2m fractional part
2m fractional part
(a) Block diagram
(b) Timing diagram
Figure 3: A k-input neuron at the horizontal sublayer takes the current input and the last k − 1 inputs
using delay elements. The bias and activation circuits are omitted. The k-input adder has the zero
latency, and the 2n-bit result is aligned by the pipeline registers, which truncates the least significant
m bits.
The proposed architecture can be realized in various ways. We can continue to be bio-inspired to implement a microarchitecture, or can use conventional digital signal processing (DSP) circuits. At one
extreme, one can employ analog spiking neurons to be biologically plausible. At the other extreme,
numbers are represented in n-bit binary, and conventional logic multiplier, adders, and flip-flops can
5
be used. Especially, the decision on the coding of numbers leads to various trade-offs in circuit area,
performance (e.g., classification accuracy), energy-efficiency, fault-tolerance, etc. Various neural
codes are introduced in [22]. One can use the rate coding, and a number can be represented as the
number of ones in a bitstream, which also enables cost-effective stochastic computing [23]. This is
an interesting research avenue, but it may require a long bitstream if even 1% performance degradation is considered critical, and the high robustness provided by the rate coding may not be necessary
in digital systems. We take a moderate approach, and use the conventional binary coding. However,
we employ the fixed-point, bit-serial representation, which provides the following strengths. 1) Unlike the rate coding, the precision increases exponentially in the number of bits, so we can easily
avoid the performance degradation due to the quantization error. If the performance degradation is
permitted, it saves more implementation cost to perform a higher degree of low-rank approximation
to the network. 2) We can use the cost-effective, bit-serial arithmetic [24]. Only a handful of logic
gates and sequential elements are needed to implement bit-serial adders and bit-serial multipliers. 3)
We can leverage the existing integrated circuit design tools fully because they understand the binary
coding well.
In our implementation, a number is represented as an n-bit stream where the least significant bit
(LSB) comes first. To represent signed real numbers in the n-bit, we use the fixed-point format and
let m denote the fractional bit-width. Figure 3(a) depicts a k-input neuron. The k-input neuron
comprises k synapses, a k-input bit-serial adder to add up the outputs of the k synapses, a bias
circuit and an activation circuit to evaluate an activation function. A synapse is implemented by an
n-bit register to store the weight and a bit-serial multiplier, which mainly consists of n full adders
and several registers to store intermediate results. The output of the k-input adder can go to the
bias circuit followed by the activation circuit, or directly go to the output of the neuron. A delay
element is realized by a 1-bit n stage shift register. Figure 3(b) depicts a timing diagram of our
implementation. It takes 2n cycles to compute the product of two n-bit numbers. To put the product
back to the n-bit format, we truncate the least significant m-bits, and the first valid output becomes
available m cycles later. Thus, the synapse comes to have an inherent m cycle delay. To simplify
the design, we pad a 2n − m cycle delay at the output of the neuron, and the registers for this
delay can be moved forward to improve timing (i.e., they are used as pipeline registers and can be
re-timed). Then, the outputs of a sublayer become available 2n cycles after the inputs arrive, and
all the synapses in the system start a synaptic operation at the same time at every 2n cycle. Delay
elements perform shift only for the first half of the 2n cycle period.
7
Experiments
We implement the neural network compiler using pylearn2 [25]. We use our own tensor approximation method, but CP-decomposition can also perform similarly. Neural network models are trained
using Nvidia GTX780 GPU and are fed into the compiler. We chose Artix7 XC7A100TCSG324-1
field-programmable gate array (FPGA) as the target platform. Note that this FPGA has relatively
small capacity and 40 times larger FPGAs are available in the market. Our experiments can be
considered in two ways. First, FPGAs are considered a general-purpose, not-highly-optimized neuromorphic processor, and the experiments are regarded as making software for the processor in
part using existing hardware synthesis tools. FPGAs already have plastic connections, and processing elements and memories are mixed in space. The SRAM-based look-up tables (LUTs) serve as
memories as well as processing elements. Second, our experiments can also be considered to build
a prototype of an application-specific neuromorphic processor. While the neural network complier
generates a hardware model in verilog in an application-specific fashion, we can easily extend it into
a general-purpose hardware model if programmable interconnects are available. We elaborate the
generated hardware models for the target platform using Xilinx Vivado. The shift registers in delay
elements are refined into an SRAM-based shift register, not a chain of flip-flops, and even long shift
registers are implemented efficiently.
To demonstrate our approach, we use a softmax regression for hand-written digit classification.
The neural network for the regression does not require the hardware for activation functions and
pooling, which allows us to focus on the reduction of the implementation cost by the low-rank
approximation. The original neural network has 7840 weight parameters. We train the network
using the MNIST data set and achieve 92.34% accuracy. For various degrees of approximation,
the neural network compiler generates TDNNs and the corresponding hardware models. Using
6
Table 1: The resulting TDNNs and their implementations for varying degrees of approximation.
Approx.
error
10%
16%
31%
40%
51%
61%
82%
(ConvNet2)
Original
0.9234
0.9234
0.9234
0.9234
0.9234
0.9234
0.9234
0.9913
Classification Accuracy
Post-approx.
Post-tuning
0.9131
0.9226
0.908
0.9217
0.8287
0.9154
0.7684
0.9055
0.5403
0.8943
0.5135
0.8377
0.3586
0.6115
0.1010
0.9764
Post-imp.
0.9155
0.9146
0.9109
0.9008
0.8943
0.8368
0.5989
0.9764
Params
3226
2274
1200
854
564
376
132
1288
TDNN
Delay units
8316
7857
6588
5670
4779
3186
1566
8844
Total LUTs
NA
53467
31223
23810
16954
11343
4754
45843
Implementation
Mem. LUTs
NA
4189
4065
2853
3007
1581
779
6809
Power (W)
NA
2.548
1.483
1.222
0.869
0.669
0.363
1.120
Table 2: The results of the neural network compiler for a convolutional neural network.
Layer 1
Layer 2
Layer 3
Original
Params
1600
102400
5760
Approx. error
12%
88%
40%
ConvNet1
Params
Param. reduction
588
2.72×
2474
41.39×
1786
3.23×
Approx. error
50%
94%
83%
ConvNet2
Params
Param. reduction
168
9.52×
829
123.52×
291
19.79×
Vivado, the hardware model is synthesized, and the power consumption and the hardware cost in
terms of LUTs are obtained. All the designs operate at 160MHz. Table 1 compares the results.
The classification accuracy is evaluated after the approximation, fine-tuning, and implementation.
The post-implementation accuracy is the performance of the actual system and reflects the finiteprecision effects. We use 7 bits for the fractional part. For TDNNs, the number of parameters and the
number of delay units indicate the size of the models. The number of parameters equals the number
of synapses, and LUTs in the FPGA are mainly used to implement synapses. A large number of
delay units are realized by a small number of memory LUTs, and a memory LUT occupies a similar
space to a LUT in the chip. Thus, the hardware cost is mainly determined by the number of synapses
(parameters). The original feedforward network with 7840 parameters cannot be implemented in the
small FPGA in a neuromorphic fashion, but the TDNN conversion and the approximation make it
possible. The TDNN with 3226 parameters is slightly larger to be fitted to the FPGA, but the smaller
TDNNs are easily contained in the target system. In neuromorphic architectures, the reduction of
parameters leads to the reduction of synapses, which is directly translated into the reduction of
hardware cost and power/energy consumption. In this experiment, the parameter reduction rate with
negligible accuracy loss is not that high possibly because it is a single-layer network. For recent deep
architectures, more than 100× parameter reduction is possible with marginal loss of accuracy [20].
To show that our approach can be extended to multi-layer neural networks, we use a convolutional
neural network (CNN) for the same task. The network has 2 convolutional layers with max-pooling
and the rectified linear function as the activation function, and a final softmax layer. The baseline
model has 109,760 parameters and achieves 99.13% classification accuracy. From the baseline,
the compiler creates two approximated networks with reduced parameters. The first network (ConvNet1) maintains the classification accuracy after re-training, and the second network (ConvNet2)
is simplified until the resulting hardware fits to our target platform. The results are summarized in
Table 1 and 2. ConvNet2 is simplified too much to be fitted to the small FPGA sacrificing some accuracy, but ConvNet1 will provide the original accuracy on an industrial FPGA. Our neuromorphic
implementation of ConvNet2 provides the exactly same accuracy as the model satisfying our design
goal that is not to sacrifice the accuracy except the low-rank approximation. For this task, to our best
knowledge, the best accuracy achieved on neuromorphic architectures is 91.94% [10].
A TDNN model is refined to a network of silicon neurons and delay elements preserving the topology, and the IC design tool places the neurons and the delay elements in a space and connects them
automatically. This provides interesting visualization of neural networks. ConvNet2 on the FPGA
is shown in Fig. 4. The network on the FPGA is stimulated by images in the test set of MNIST,
which are transferred via UART, and the output of the network is visualized in a screen as shown
in Figure 5. The neuromorphic system successfully classifies images in real-time, and the results
match with the simulation exactly.
7
(a)
(b)
(c)
Figure 4: Neural network on a chip: (a) The neurons and the delay elements in ConvNet2 are
highlighted in blue and red, respectively. (b) The three layers of ConvNet2 are highlighted in yellow,
green, and magenta, respectively. (c) The connectivity of ConvNet2 is shown.
Figure 5: Our neuromorphic system based on a FPGA is executing a neural network originally
with 100K weight parameters for the hand-written digit classification task. Now both software and
hardware are bio-inspired. The classification accuracy of the system is 97.64% even though a small
FPGA is used, and it processes an input image in 156.8µs. The internal block RAMs are used as the
frame buffer, and any external memory is not used in this system, which shows that it is a non-Von
Neumann architecture.
8
Discussion
In this paper we have presented a neuromorphic computing system that is newly designed from the
microarchitecture to the compiler in order to forward-execute neural networks with minimum energy
consumption. We have also demonstrated that the parameter reduction via low-rank approximation
is much more valuable than considered in the community. This neuromorphic system can scale by
simply using a larger FPGA or by fabricating a custom chip. Since more than 40 times larger FPGAs
than the target platform are available in the market, with our approach, it now becomes easy to build
neuromorphic computing system that can execute neural networks with more than 4 million real8
valued parameters, fully leveraging existing integrated circuit design techniques. Our neuromorphic
system has just built and is not fully optimized. Through all the levels of the computing stack,
there is huge room for optimization. Our future work may include the circuit-level optimization and
neural network transformations to minimize the implementation cost.
References
[1] Dan Ciresan, Ueli Meier, and Jürgen Schmidhuber. Multi-column deep neural networks for
image classification. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 3642–3649, 2012.
[2] Pierre Sermanet, David Eigen, Xiang Zhang, Michaël Mathieu, Rob Fergus, and Yann LeCun. Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint arXiv:1312.6229, 2013.
[3] Clément Farabet, Berin Martini, Polina Akselrod, Selçuk Talay, Yann LeCun, and Eugenio
Culurciello. Hardware accelerated convolutional neural networks for synthetic vision systems.
In Circuits and Systems (ISCAS), Proceedings of 2010 IEEE International Symposium on,
pages 257–260, 2010.
[4] Srimat Chakradhar, Murugan Sankaradas, Venkata Jakkula, and Srihari Cadambi. A dynamically configurable coprocessor for convolutional neural networks. In ACM SIGARCH Computer Architecture News, volume 38, pages 247–257, 2010.
[5] John Backus. Can programming be liberated from the von neumann style?: a functional style
and its algebra of programs. Communications of the ACM, 21(8):613–641, 1978.
[6] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million
spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668–673, 2014.
[7] Johannes Schemmel, D Bruderle, A Grubl, Matthias Hock, Karlheinz Meier, and Sebastian
Millner. A wafer-scale neuromorphic hardware system for large-scale neural modeling. In
Circuits and Systems (ISCAS), Proceedings of 2010 IEEE International Symposium on, pages
1947–1950, 2010.
[8] Ben Varkey Benjamin, Peiran Gao, Emmett McQuinn, Swadesh Choudhary, Anand R Chandrasekaran, J Bussat, Rodrigo Alvarez-Icaza, John V Arthur, PA Merolla, and Kwabena Boahen. Neurogrid: A mixed-analog-digital multichip system for large-scale neural simulations.
Proceedings of the IEEE, 102(5):699–716, 2014.
[9] John V Arthur, Paul A Merolla, Filipp Akopyan, Rodrigo Alvarez, Andrew Cassidy, Shyamal
Chandra, Steven K Esser, Nabil Imam, William Risk, Daniel Ben Dayan Rubin, et al. Building
block of a programmable neuromorphic substrate: A digital neurosynaptic core. In Neural
Networks (IJCNN), The 2012 International Joint Conference on, pages 1–8, 2012.
[10] Steven K Esser, Alexander Andreopoulos, Rathinakumar Appuswamy, Pallab Datta, Davis
Barch, Arnon Amir, John Arthur, Andrew Cassidy, Myron Flickner, Paul Merolla, et al. Cognitive computing systems: Algorithms and applications for networks of neurosynaptic cores.
In Neural Networks (IJCNN), The 2013 International Joint Conference on, pages 1–10, 2013.
[11] Arnon Amir, Pallab Datta, William P Risk, Andrew S Cassidy, Jeffrey A Kusnitz, Steven K
Esser, Alexander Andreopoulos, Theodore M Wong, Myron Flickner, Rodrigo Alvarez-Icaza,
et al. Cognitive computing programming paradigm: a corelet language for composing networks
of neurosynaptic cores. In Neural Networks (IJCNN), The 2013 International Joint Conference
on, pages 1–10, 2013.
[12] Andrew S. Cassidy, Paul Merolla, John V. Arthur, Steve K. Esser, Bryan Jackson, Rodrigo
Alvarez-Icaza, Pallab Datta, Jun Sawada, Theodore M. Wong, Vitaly Feldman, Arnon Amir,
Daniel Ben-Dayan Rubin, Filipp Akopyan, Emmett McQuinn, William P. Risk, and Dharmendra S. Modha. Cognitive computing building block: A versatile and efficient digital neuron
model for neurosynaptic cores. In Neural Networks (IJCNN), The 2013 International Joint
Conference on, pages 1–10, 2013.
9
[13] Andrew S Cassidy, Rodrigo Alvarez-Icaza, Filipp Akopyan, Jun Sawada, John V Arthur,
Paul A Merolla, Pallab Datta, Marc Gonzalez Tallada, Brian Taba, Alexander Andreopoulos,
et al. Real-time scalable cortical computing at 46 giga-synaptic ops/watt with ∼100x speed up
in time-to-solution and ∼100,000x reduction in energy-to-solution. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis,
pages 27–38, 2014.
[14] Jae-sun Seo, Bernard Brezzo, Yong Liu, Benjamin D Parker, Steven K Esser, Robert K Montoye, Bipin Rajendran, José A Tierno, Leland Chang, Dharmendra S Modha, et al. A 45nm
cmos neuromorphic chip with a scalable architecture for learning in networks of spiking neurons. In Custom Integrated Circuits Conference (CICC), 2011 IEEE, pages 1–4, 2011.
[15] Joseph M Brader, Walter Senn, and Stefano Fusi. Learning real-world stimuli in a neural
network with spike-driven synaptic dynamics. Neural computation, 19(11):2881–2912, 2007.
[16] Daniel L Ly and Paul Chow. A high-performance fpga architecture for restricted boltzmann
machines. In Proceedings of the ACM/SIGDA international symposium on Field programmable
gate arrays, pages 73–82, 2009.
[17] Lok-Won Kim, Sameh Asaad, and Ralph Linsker. A fully pipelined fpga architecture of a
factored restricted boltzmann machine artificial neural network. ACM Transactions on Reconfigurable Technology and Systems (TRETS), 7(1):5, 2014.
[18] G. Snider. Molecular-junction-nanowire-crossbar-based neural network, April 15 2008. US
Patent 7,359,888.
[19] Misha Denil, Babak Shakibi, Laurent Dinh, Nando de Freitas, et al. Predicting parameters
in deep learning. In Advances in Neural Information Processing Systems, pages 2148–2156,
2013.
[20] Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan Oseledets, and Victor Lempitsky.
Speeding-up convolutional neural networks using fine-tuned cp-decomposition. arXiv preprint
arXiv:1412.6553, 2014.
[21] Jilles Vreeken. Spiking neural networks, an introduction. Technical Report UU-CS, (2003008):1–5, 2003.
[22] Wulfram Gerstner and Werner M Kistler. Spiking neuron models: Single neurons, populations,
plasticity. Cambridge university press, 2002.
[23] Armin Alaghi and John P Hayes. Survey of stochastic computing. ACM Transactions on
Embedded computing systems (TECS), 12(2s):92, 2013.
[24] Peter B Denyer and David Renshaw. VLSI signal processing: a bit-serial approach, volume 1.
Addison-Wesley Reading, MA, 1985.
[25] Ian J. Goodfellow, David Warde-Farley, Pascal Lamblin, Vincent Dumoulin, Mehdi Mirza,
Razvan Pascanu, James Bergstra, Frédéric Bastien, and Yoshua Bengio. Pylearn2: a machine
learning research library. arXiv preprint arXiv:1308.4214, 2013.
10
| 9cs.NE
|
UDC 512.624.3 + 512.552
MSC: 12E12, 16U99, 16P10
BALANCED FACTORISATIONS IN SOME ALGEBRAS
Anton A. Klyachko Andrey M. Mazhuga Anastasia N. Ponfilenko
Faculty of Mechanics and Mathematics, Moscow State University
Moscow 119991, Leninskie gory, MSU
[email protected]
[email protected]
[email protected]
arXiv:1607.01957v1 [math.RA] 7 Jul 2016
We prove that, in any field of characteristic not two and not three except F5 , each element decomposes
into a product of four factors whose sum vanishes. We also find all k, n, q such that every n×n matrix
over the q -element field decomposes into a product of k commuting matrices whose sum vanishes.
0. Introduction
Suppose that an element a of a ring is factored into a product: a = a1 a2 . . . ak ; we call this factorisation balanced if
P
ai = 0. It is easy to show (see [KV16]) that,
in any field of characteristic not two, each element admits a balanced decomposition into a
product of k factors for each k > 5. (While, for each k < 5, this assertion is no longer valid.)
We prove the following theorem answering thereby two question from [KV16] (one of which had been known earlier [Iva13]).
Theorem 1. In any field of characteristic not two and not three except the five-element field F5 , each element admits
a balanced decomposition into a product of four factors; if the field is infinite, then each element admits infinitely
many such decompositions.
We are unsure whether this result was known earlier, because
- A high-school teacher of the second author, Dmitrii Vitalievich Andreev, in 2003 suggested such a problem on
classes (for the field of rationals); the problem was not solved, but the teacher’s hints make the second author
suspect (now, when he already knows a solution) that Dmitrii Vitalievich was able to solve this problem;
- in 2016 (already knowing a solution), the second author asked such a question (again for the field of rationals)
on the forum http://math.stackexchange.com/ and shortly someone posts a solution, different from ours, but
probably also valid; now this question is, unfortunately, removed and we cannot give an exact references.
The following theorem from [KV16] describes, for each k, all finite fields, in which each element admits a balanced
decomposition into a product of k factors.
Theorem on balanced factorisations in finite fields [KV16]. Suppose that k > 2 is an integer and F is a finite
field. Then any element of F has a balanced factorisation into a product of k factors if and only if
either |F | = 2 and k is even,
or |F | = 4 and k 6= 3,
or |F | is a power of two but neither two nor four (and k is arbitrary),
or |F | ∈ {3, 5} and k ∈
/ {2, 4},
or |F | = 7 and k ∈
/ {2, 3},
or |F | is neither a power of two nor three nor five nor seven and k 6= 2.
In other words, the situation in finite fields is the following:
F2
F3
F4
F5
F7
F8 , F16 , F32 , F64 , . . .
F9 , F11 , F13 , F17 , . . .
k=2
yes
no
yes
no
no
yes
no
k=3
no
yes
no
yes
no
yes
yes
k=4
yes
no
yes
no
yes
yes
yes
k = 5, 7, 9, . . .
no
yes
yes
yes
yes
yes
yes
k = 6, 8, 10, . . .
yes
yes
yes
yes
yes
yes
yes
Table 1
The following theorem complements this result from [KV16].
Theorem 2. Let k, n > 2 be integers and F be a finite field. Any n × n matrix over F decomposes into a product of
k commuting matrices whose sum vanishes if and only if
either k = 3 and |F | = 5, or k = 3 and |F | > 8, or k = 4 and |F | = 4, or k = 4 and |F | > 7, or k > 5 and |F | > 3.
In other words, the situation in the matrix algebras over finite fields is the following (ignore the superscripts for now):
The work of the first two authors was supported by the Russian Foundation for Basic Research, project no.
15-01-05823.
1
F2
F3
F4 , F7
F5
F8 , F9 , F11 F13 F16 , . . .
k=2
no1
no1
no1
no1
no1
k=3
no0
no8
no0
yes5
yes5,6
k=4
no2
no0
yes5,7
no0
yes5,7
k = 5, 6, 7, 8, . . .
no2
yes3,4
yes3,4,5,7
yes3,4
yes3,4,5,7
Table 2
Note that the size of matrices does not affect the answer, provided this size is at least two.
In Section 1, we prove Theorem 1 and also another theorem on finite fields (Theorem 3), which plays the key role
in the proof of Theorem 2. All argument in Section 1 is elementary, except that the proof of Theorem 3 is substantially
based on results of [KV16] (that in turn are based on the theory of elliptic curves). In Section 2, we prove Theorem 2.
Notation, which we use are standard. Note only that the symbol Fq denotes the q-element field and the letter E
always denotes the identity matrix.
1. Fields
Proof of Theorem 1. Look at this identity
x=
2(1 − 4x)2 −(1 + 8x) −(1 + 8x)
18x
·
·
·
.
3(1 + 8x)
6
2(1 − 4x) (1 − 4x)(1 + 8x)
We can only suggest readers to verify this identity and the fact that the sum of factors is zero. In the exceptional cases,
/ { 41 , − 81 , 0} (this
when denominators vanishes, i.e. for x ∈ { 14 , − 81 }, we can multiply x by y 4 choosing y such that xy 4 ∈
is possible in any field of characteristic not two and not three except F5 ), write a similar decomposition for xy 4 and
then divide each factor by y:
x=
2(1 − 4xy 4 )2 −(1 + 8xy 4 ) −(1 + 8xy 4 )
18xy 4
·
·
·
.
3y(1 + 8xy 4 )
6y
2y(1 − 4xy 4 ) y(1 − 4xy 4 )(1 + 8xy 4 )
Thus we obtain a balanced decomposition of any element into a product of four factors. Moreover, the last identity
implies that there are infinitely many such decompositions if the field is infinite, because of the following elementary
fact (whose proof is left for readers as an easy exercise):
any non-constant rational fraction over an infinite field takes infinitely many values.
(Note that, for each x, the second factor in the last identity is a non-constant rational fraction f (y).) This completes
the proof. The five-element field is indeed an exception, see Table 1.
Similarly, we can obtain infinitely many balanced decompositions into products of any larger number of factors,
i.e.
in any infinite field characteristic not two each element admits infinitely
many balanced decompositions into products of k factors for each k > 5.
For example, the following identity is obtained by a slight modification of an identity from [KV16]; this gives infinitely
many balanced decompositions of any nonzero element of an infinite field of characteristic not two into products of
2017 factors:
1006
1006
xy 2016 xy 2016
2
2
1
1
2016
x=
· 2018 · − 2018 ·
·
· −xy
· −
.
2
2
xy
xy
y
y
A decomposition x = x1 x2 . . . xk is called power if all factors are equal: x1 = . . . = xk [KV16].
Theorem 3. Let k > 2 be an integer and let F be a finite field. Any element of F admits a non-power balanced
decomposition into a product of k factors if and only if
either k = 3 and |F | = 5, or k = 3 and |F | > 8, or k = 4 and |F | = 4, or k = 4 and |F | > 7, or k > 5 and |F | > 3.
In other words, the answer here is the same as in Theorem 2 (Table 2).
Proof. Superscripts of the particles yes and no in Table 2 indicate references to the cases below.
Case 0: no, because in these cases for some elements there are no balanced decompositions (by Theorem on balanced
factorisations in finite fields).
Case 1: k = 2 — no. For any element a, consider its balanced decomposition a = xy, x + y = 0. If the characteristic
is 2, then the decomposition a = (−x)x is power; if the characteristic is not two, then not any element is square and,
therefore, not any element admits a balanced decomposition into a product of two factors.
Case 2: |F | = 2 — no. The decomposition of 1 can comprises only 1s and, therefore, it is power.
2
Case 3: k = 5 + 2n, where n > 0 and char F 6= 2 — yes. Let us use the universal formula for balanced decompositions
into a product of 5 + 2n factors from [KV16]:
±a = (−a) ·
a a 2 −2 n
· · ·
· 1 · (−1)n
2 2 a a
(for a 6= 0).
char F 6= 2, therefore, a2 6= − a2 and, hence, this decomposition is non-power. The zero element has an obvious balanced
non-power decomposition: 0 = (−1) · 1 · 0k−2 .
Case 4: k = 6 + 2n, where n > 0 and char F 6= 2 — yes. Let us use a formula for balanced decompositions into a
product of 6 + 2n factors from [KV16]: Consider c ∈ F such that c2 6= a (such c exists, except for the case where
2
F = F3 and b = 1; but in this case everything is obvious). Put b = c c−a . Then
±a = (−c) · (c − b) ·
b b 2 −2 n
· · ·
· 1 · (−1)n .
2 2 b b
Since char F 6= 2, we have 2b 6= − 2b and, therefore, the decomposition is non-power.
Case 5: In these cases, balanced decompositions exist by Theorem on balanced factorisations in finite fields; these
decompositions cannot be power, because the number of factors is not divisible by the characteristic.
Case 6: |F | > 9, char F 6= 2, and k = 3 — yes.
Consider two cases. If char F 6= 3, then the Theorem on balanced factorisations in finite fields provides us with a
balanced decomposition for any element; since char F 6= 3, this decomposition cannot be power.
To prove the assertion for characteristic three, we need a lemma.
Lemma. In the field F3n , where n > 2, there exists a nonzero square u such that u + 1 is also a nonzero square.
Proof. If 2 is a square, then u = 1 is the required element. Otherwise, suppose that, for any u ∈
/ {0, 1, 2}, the following
holds:
if u is a square, then u + 1 is not a square.
Then, any set of the form {u, u + 1, u + 2} contains at most one square. Therefore, the number of squares is at
n
most 3n−1 + 1. On the other hand, in the field of characteristic 3, there are precisely 3 2+1 squares. This implies the
n
inequality 3 2+1 6 3n−1 + 1 that holds only for n = 1. This contradiction completes the proof of the lemma.
Let us resume the proof of Theorem 3. We have to show that, in a finite field of characteristic three and cardinality
at least nine, each element has a balanced non-power decomposition into a product of three factors.
Note that, in such a field, any element is a cube. We want to find a balanced decomposition of an element
a = b3 6= 0:
a = xyz, x + y + z = 0. Eliminating z, we obtain yx2 + y 2 x + a = 0.
(∗)
Let us solve this equation with respect to x. By lemma, there exists τ 2 6= 0 such that τ 2 + 1 6= 0 is also a square:
τ 2 + 1 = π 2 6= 0. Take y = b+bπ
2 . Note that y 6= b, because the equality y = b would mean that π = 1 and τ = 0.
Therefore, the discriminant of the quadratic equation (∗) is a square:
bπ + b
D = y − 4ay = y(y − b ) = y(y − b) =
·
2
4
3
3
3
bπ − b
2
3
= (b2 (1 + τ 2 ) − b2 )(bπ − b)2 = b2 τ 2 (bπ − b)2
and equation (∗) has a solution. The obtained decomposition is not power, because y 3 6= a (since y 6= b).
It remains to find a non-power balanced decomposition of zero, but this is an easy task: 0 = (−1) · 1 · 0.
Case 7: |F | = 2n and k = 4 + 2m, where m > 0. Since char F = 2, any element is a square: a = b · b. If a 6= 1, then
2
b 6= 1 and a = b2 · 12m+2 is a required decomposition. If a = 1, then a = 1 = c2 · 1c · 12m is a required decomposition,
where c is any element different from 0 and 1.
Case 8: |F | = 3 and k = 3 — no. The decomposition of 1 cannot can contain 0 and, therefore, must contain 1 and
−1, to make decomposition non-power. Then, the third factor must be zero, because the decomposition is balanced.
This contradiction completes the proof of Theorem 3.
2. Matrices
Proof of Theorem 2. First note that, for k = 2, Theorem 2 is valid:
the Jordan cell J with eigenvalue zero and size n × n is not a square in the ring of n × n matrices if n > 2
(we leave the proof of this fact to readers as an exercise) and, therefore, the matrix −J has no balanced decompositions
into a product of two factors.
In the case k > 3, Theorem 3 implies that it suffices to prove the following statement.
3
Theorem 2′ . Let n > 2 and k > 3 are integers and let F be a field (not necessarily finite). Then the following
conditions are equivalent:
a) each matrix n × n over F has a balanced decomposition into a product of k commuting factors;
b) each element of F has a non-power balanced decomposition into a product of k factors.
Proof.
The implication b) =⇒ a) follows immediately from the following fact proven in [KV16]:
Let F be a field and let k be a positive integer larger than two. If, in all finite extensions of the
field F , each element has a non-power balanced decomposition into a product k elements, then
the same is true for each element of each finite-dimensional associative algebra with unit over F .
The implication a) =⇒ b). First, note that 0 ∈ F has a balanced non-power decomposition into a product of k
factors for any k > 3: 0 = 0k−2 · 1 · (−1). To obtain a decomposition of a nonzero element a ∈ F , we need the
following simple fact from linear algebra (the proof is left to readers as an exercise):
The centraliser of the nilpotent Jordan cell J of size n × n in the algebra of n × n
matrices consists of polynomials in J, i.e. C(J) = {a0 E+a1 J +. . .+an−1 J n−1 | ai ∈ F }.
Thus, if the Jordan cell aE + J has a balanced decomposition aE + J = X1 . . . Xk into a product of commuting
matrices, then all matrices Xi lie in the centraliser of J and, by virtue of the fact mentioned above, we obtain a
balanced decomposition of a in F :
a = x1 . . . xk ,
where xi is the (unique) eigenvalue of Xi .
It remains to note that this decomposition cannot be power for a 6= 0. Indeed, assuming the contrary, we would obtain
a balanced decomposition in the ring of matrices:
aE + J = (xE + J1 ) . . . (xE + Jk ),
where Ji are nilpotent commuting matrices.
P
The balancedness of this decomposition means that k is divisible by char F and
Ji = 0. But then, multiplying out
brackets, we obtain
aE + J = (xE + J1 ) . . . (xE + Jk ) = aE + f (J1 , . . . , Jk ),
where the polynomial f has no terms degree 1, i.e. the right-hand side of this equality is a matrix of the form aE + J ′ ,
where (J ′ )n−1 = 0. This contradiction completes the proof of Theorems 2′ and 2.
Generally (without the commutativity condition), the question on balanced decompositions of matrices over finite
fields remains open.
Question. For which q, k, and n, it is true that any n×n matrix over the q-element field has a balanced decomposition
into a product of k matrices?
We can say only the following.
1. In some cases the decompositions exist by Theorem 2.
2. For k = 2 6 n, the decomposition does not exist, e.g., because factors such balanced decomposition must commute.
Moreover, computer experiments show the following facts.
1 1
0 1
3. Over the two-element field, all 2 × 2 matrices, except
and (a similar matrix)
, admit balanced
1 0
1 1
decompositions
intoa product of three factors, while these two matrices have no such decompositions.
1 0
4. The matrix
and two similar matrices over the two-element field have no balanced decompositions into a
1 1
product of four
factors,while all other 2 × 2 matrices over F2 have such decompositions.
1 0 0
5. The matrix 1 1 0 and similar matrices over the two-element field have no balanced decompositions into a
0 0 1
product of three factors, while the remaining 3 × 3 matrices over F2 have such decompositions.
6. All 3 × 3 matrices over F2 have balanced decompositions into products of four factors.
7. All 2 × 2 matrices over F3 , F4 , F5 , and F7 have balanced decompositions into products of three and four factors.
This implies that the same is true for any larger number of factors, because we can increase the number of factors
by two multiplying decompositions by E and −E.
REFERENCES
[Iva13] Ivanishchuk A. V. The experience of learning and research activity of students in Lyceum 1511 (MEPhI)
(in Russian) // published in the book Sgibnev A. I. Research problems for beginners. Moscow: MCCME, 2013.
(Freely available at http://www.mccme.ru/free-books/)
[KV16] Klyachko A. A., Vassiliyev A. N. Balanced factorisations // arXiv:1506.01571.
4
| 0math.AC
|
arXiv:1706.02897v1 [cs.AI] 7 Jun 2017
Bandit Models of Human Behavior:
Reward Processing in Mental Disorders
Djallel Bouneffouf, Irina Rish, Guillermo A. Cecchi
IBM Thomas J. Watson Research Center
Yorktown Heights, NY USA
Abstract. Drawing an inspiration from behavioral studies of human decision making, we
propose here a general parametric framework for multi-armed bandit problem, which extends the standard Thompson Sampling approach to incorporate reward processing biases
associated with several neurological and psychiatric conditions, including Parkinson’s and
Alzheimer’s diseases, attention-deficit/hyperactivity disorder (ADHD), addiction, and chronic
pain. We demonstrate empirically that the proposed parametric approach can often outperform the baseline Thompson Sampling on a variety of datasets. Moreover, from the behavioral modeling perspective, our parametric framework can be viewed as a first step towards
a unifying computational model capturing reward processing abnormalities across multiple
mental conditions.
1 Introduction
In daily-life decision making, from choosing a meal at a restaurant to deciding on a place to
visit during a vacation, and so on, people often face the classical exploration versus exploitation
dilemma, requiring them to choose between following a good action chosen previously (exploitation) and obtaining more information about the environment which can possibly lead to better
actions in the future, but may also turn out to be a bad choice (exploration).
The exploration-exploitation trade-off is typically modeled as the multi-armed bandit (MAB)
problem, stated as follows: given N possible actions (“arms”), each associated with a fixed, unknown and independent reward probability distribution [1,2], an agent selects an action at each
time point and receives a reward, drawn from the corresponding distribution, independently of
the previous actions.
In order to better understand and model human decision-making behavior, scientists usually
investigate reward processing mechanisms in healthy subjects [3]. However, neurogenerative and
psychiatric disorders, often associated with reward processing disruptions, can provide an additional resource for deeper understanding of human decision making mechanisms. Furthermore,
from the perspective of evolutionary psychiatry, various mental disorders, including depression,
anxiety, ADHD, addiction and even schizophrenia can be considered as “extreme points” in a
continuous spectrum of behaviors and traits developed for various purposes during evolution, and
somewhat less extreme versions of those traits can be actually beneficial in specific environments
(e.g., ADHD-like fast-switching attention can be life-saving in certain environments, etc.). Thus,
modeling decision-making biases and traits associated with various disorders may actually enrich the existing computational decision-making models, leading to potentially more flexible and
better-performing algorithms.
Herein, we focus on reward-processing biases associated with several mental disorders, including Parkinson’s and Alzheimer disease, ADHD, addiction and chronic pain. Our questions
are: is it possible to extend standard stochastic bandit algorithms to mimic human behavior in
such disorders? Can such generalized approaches outperform standard bandit algorithms on specific tasks?
We show that both questions can be answered positively. We build upon the Thompson Sampling, a state-of-art approach to multi-arm bandit problem, and extend it to a parametric version
which allows to incorporate various reward-processing biases known to be associated with particular disorders. For example, it was shown that (unmedicated) patients with Parkinson’s disease
appear to learn better from negative rather than from positive rewards [4]; another example is
addictive behaviors which may be associated with an inability to forget strong stimulus-response
associations from the past, i.e. to properly discount past rewards [5], and so on. More specifically,
we propose a parameteric model which introduces weights on incoming positive and negative
rewards, and on reward histories, extending the standard parameter update rules in Bernoulli
Thompson Sampling; tuning the parameter settings allows us to better capture specific rewardprocessing biases.
Our empirical results demonstrate that the proposed approach outperforms the baseline Thompson Sampling on a variety of UCI benchmarks Furthermore, we show how parameter-tuning in
the proposed model allows to mimic certain aspects of the behavior associated with mental disorders mentioned above, and thus may provide a valuable tool for improving our understanding of
such disorders.
The rest of this paper is organized as follows. Section 2 reviews related work. Section 3
describes the MAB model and the proposed algorithm. The experimental evaluation for different
setting is presented in Section 5. The last section concludes the paper and identifies directions for
future works.
2 Related Work
2.1 Reward Processing in Mental Disorders
The literature on the reward processing abnormalities in particular neurological and psychiatric
disorders is quite extensive; below we summarize some of the recent developments in this fastgrowing field.
Parkinson’s disease (PD). It is well-known that the neuromodulator dopamine plays a key
role in reinforcement learning processes. PD patients, who have depleted dopamine in the basal
ganglia, tend to have impaired performance on tasks that require learning from trial and error. For
example, [4] demonstrate that off-medication PD patients are better at learning to avoid choices
that lead to negative outcomes than they are at learning from positive outcomes, while dopamine
medication typically used to treat PD symptoms reverses this bias.
Alzheimer’s disease (AD). This is the most common cause of dementia in the elderly and,
besides memory impairment, it is associated with a variable degree of executive function impairment and visuospatial impairment. As discussed in [3], AD patients have decreased pursuit of
rewarding behaviors, including loss of appetite; these changes are often secondary to apathy, associated with diminished reward system activity. Furthermore, poor performance on certain tasks
is correlated with memory impairments.
Frontotemporal dementia, behavioral variant (bvFTD). Frontotemporal dementia (bvFTD)
typically involves a progressive change in personality and behavior including disinhibition, apathy, eating changes, repetitive or compulsive behaviors, and loss of empathy [3], and it is hypothesized that those changes are associated with abnormalities in reward processing. For example,
changes in eating habits with a preference for sweet, carbohydrate rich foods and overeating in
bvFTD patients can be associated with abnormally increased reward representation for food, or
impairment in the negative (punishment) signal associated with fullness.
Attention-deficit/hyperactivity disorder (ADHD). Authors in [6] suggest that the strength
of the association between a stimulus and the corresponding response is more susceptible to
degradation in ADHD patients, which suggests problems with storing the stimulus-response associations. Among other functions, storing the associations requires working memory capacity,
which is often impaired in ADHD patients.
Addiction. In [5], it is demonstrated that patients suffering from addictive behavior are not
able to forget the stimulus-response associations, which causes them to constantly seek the stimulus which generated such association.
Chronic pain. In [7], it is suggested that chronic pain results in a hypodopaminergic (low
dopamine) state that impairs motivated behavior, resulting into a reduced drive in chronic pain
patients to pursue the rewards. Decreased reward response may underlie a key system mediating
the anhedonia and depression common in chronic pain.
A variety of computational models was proposed for studying the disorders of reward processing in specific disorders, including, among others [4,8,9,10,5,11].
However, none of the above studies is proposing a unifying model that can represent a wide
range of reward processing disorders; moreover, none of the above studies used the multi-arm
bandit model simulating human online decision-making.
2.2 Multi-Armed Bandit (MAB)
The multi-armed bandit (MAB) problem models a sequential decision-making process, where at
each time point a player selects an action from a given finite set of possible actions, attempting to
maximize the cumulative reward over time.
MAB is frequently used in reinforcement learning to study the exploration/exploitation tradeoff, and is an active area of research since the 1950s. Optimal solutions have been provided
using a stochastic formulation [1,2], or using an adversarial formulation [12,13,14]. Recently,
there has been a surge of interest in a Bayesian formulation [15], involving the algorithm known
as Thompson sampling [16]. Theoretical analysis in [17] shows that Thompson sampling for
Bernoulli bandits asymptotically achieves the optimal performance limit. Empirical analysis of
Thompson sampling, including problems more complex than the Bernoulli bandit, demonstrates
that its performance is highly competitive with other approaches [15,18].
Psychological study done in [19] shows that, instead of maximizing output by a deliberate
mean-variance trade-off, participants approach dynamic decision-making problems by utilizing a
probability matching heuristic. Thus, their behavior is better described by the Thompson sampling
choice rule than by the Upper Confidence Bound (UCB) approach [2]. However, none of the above
studies bandit models of the behavior of patients with mental disorders and impaired reward
processing.
To the best of our knowledge, this work is the first one to propose a generalized version of
Thompson Sampling algorithm which incorporates a range of reward processing biases associated
with various mental disorders and shows how different parameter settings of the proposed model
lead to behavior mimicking a wide range of impairments in multiple neurological and psychiatric
disorders. Most importantly, our bandit algorithm based on generalization of Thompson sampling
outperforms the baseline method on multiple datsasets.
3 Background and Definitions
The Stochastic Multi-Armed Bandit. Given a slot machine with N arms representing potential
actions, the player must chose one of the arms to play at each time step t = 1, 2, 3, ..., T . Choosing
an arm i yields a random real-valued reward according to some fixed (unknown) distribution with
support in [0, 1]. The reward is observed immediately after playing the arm. The MAB algorithm
must decide which arm to play at each time step t, based on the outcomes during the previous
t − 1 steps.
Let µi denote the (unknown) expected
for arm i. The goal is to maximize the expected
Preward
T
total reward during T iterations, i.e., E[ t=1 µi(t) ], where i(t) is the arm played in step t, and
the expectation is over the random choices of i(t) made by the algorithm. We could also use
the equivalent performance measure known as the expected total regret, i.e. the amount of total
reward lost because of playing according to a specific algorithm rather than choosing the optimal
arm in each step.
The expected total regret is formally defined as:
E[R(T )] = E[
T
X
t=1
(µ∗ − µi(t) )] =
X
∆i E[ki (T )].
(1)
i
where µ∗ := maxi µi , ∆i := µ∗ − µi , and ki (t) denote the number of times arm i has been
played up to step t.
Thompson Sampling. Thompson sampling (TS) [20], also known as Basyesian posterior sampling, is a classical approach to multi-arm bandit problem, where the reward ri (t) for choosing an
arm i at time t is assumed to follow a distribution P r(rt |µ̃) with the parameter µ̃. Given a prior
P r(µ̃) on these parameters, their posterior distribution is given by the Bayes rule, P r(µ̃|rt ) ∝
P r(rt |µ̃)P r(µ̃) [17].
A particular case of the Thompson Sampling approach, presented in Algorithm 1, assumes a
Bernoulli bandit problem, with rewards being 0 or 1, and the parameters following the Beta prior.
TS initially assumes arm i to have prior Beta(1, 1) on µi (the probability of success). At time
t, having observed Si (t) successes (reward = 1) and Fi (t) failures (reward = 0), the algorithm
updates the distribution on µi as Beta(Si (t), Fi (t)). The algorithm then generates independent
samples θi (t) from these posterior distributions of the µi , and selects the arm with the largest
sample value.
4 Proposed Approach: Human-Based Thompson Sampling
We will now introduce a more general rule for updating the parameters of Beta distribution in
steps 10 and 11 of the Algorithm 1; this parameteric rule incorporates weights on the prior and the
current number of successes and failures, which will allow to model a wide range of reward processing biases associated with various disorders. More specifically, the proposed Human-Based
Algorithm 1: Thompson Sampling
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
Foreach arm i = 1, ..., K
set Si (t) = 1, Fi (t) = 1
End for
Foreach t = 1, 2, ..., T do
Foreach i = 1, 2, ..., K do
Sample θi (t) from Beta(Si (t), Fi (t))
End do
Play arm it = argmaxiθi (t), obtain reward r(t)
if r(t) = 1, then
Si (t) = Si (t) + 1
else Fi (t) = Fi (t) + 1
End do
Thompson Sampling (HBTS), outlined in Algorithm 2, replaces binary incremental updates in
lines 10 and 11 of TS (Algorithm 1) with their corresponding weighted version (lines 10 and
11 in Algorithm 2), using the four weight parameters: τ and φ are the weights of the previously
accumulated positive and negative rewards, respectively, while α and β represent the weights on
the positive and negative rewards at the current iteration.
Algorithm 2: Human-Based Thompson Sampling (HBTS)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
Foreach arm i = 1, ..., K
set Si (t) = 1, Fi (t) = 1
End for
Foreach t = 1, 2, ..., T do
Foreach i = 1, 2, ..., K do
Sample θi (t) from Beta(Si (t), Fi (t))
End do
Play arm it = argmaxiθi (t), obtain reward r(t)
if ri (t) = 1, then
Si (t) = τ Si (t) + αri (t)
else Fi (t) = φFi (t) + β(1 − ri (t))
End do
4.1 Reward Processing Models with Different Biases
In this section we describe how specific constraints on the model parameters in the proposed
algorithm can yield different reward processing biases discussed earlier, and introduce several
instances of the HBTS model, with parameter settings reflecting particular biases. The parameter settings are summarized in Table 2, where we use list our models associated with specific
disorders.
It is important to underscore that the above models should be viewed as only a first step towards a unifying approach to reward processing disruptions, which requires further extensions, as
Table 1: Algorithms parameters
τ
α
φ
β
UCI Datasets
AD (addiction)
1 ± 0.1
1 ± 0.1 0.5 ± 0.1 1 ± 0.1
0.2 ± 0.1 1 ± 0.1 0.2 ± 0.1 1 ± 0.1
ADHD
AZ (Altzheimer’s) 0.1 ± 0.1 1 ± 0.1 0.1 ± 0.1 1 ± 0.1
CP (chronic pain) 0.5 ± 0.1 0.5 ± 0.1 1 ± 0.1
1 ± 0.1
bvFTD
0.5 ± 0.1 100 ± 10 0.5 ± 0.1 1 ± 0.1
PD (Parkinson’s) 0.5 ± 0.1 1 ± 0.1 0.5 ± 0.1 100 ± 10
M (“moderate”) 0.5 ± 0.1 1 ± 0.1 0.5 ± 0.1 1 ± 0.1
TS
1
1
1
1
well as tuning and validation on human subjects. Our main goal is to demonstrate the promise of
our parametric approach at capturing certain decision-making biases, as well as its computational
advantages over the standard TS, due to the increased generality and flexibility facilitated by
multi-parametric formulation. Note that the standard Thompson sampling (TS) approach correspond to setting the four (hyper)parameters used in our model to 1. Next, we introduce the model
which incorporates some mild forgetting of the past rewards or losses, using 0.5 weights, just
as an example, and calibrating the other models with respect to this one; we refer to this model
as M for “moderate” forgetting, which serves here as a proxy for somewhat “normal” reward
processing, without extreme reward-processing biases associated with disorders. We will use the
subscript M to denote the parameters of this model.
We will now introduced several models inspired by certain reward-processing biases in a
range of mental disorders. It is important to note that, despite using disorder names for these
models, we are not claiming that they provide accurate models of the corresponding disorders,
but rather disorder-inspired versions of our general parametric family of models.
Parkinson’s disease (PD). Recall that PD patients are typically better at learning to avoid
negative outcomes than at learning to achieve positive outcomes [4]; one way to model this is to
over-emphasize negative rewards, by placing a high weight on them, as compared to the reward
processing in healthy individuals. Specifically, we will assume the parameter β for PD patients
to be much higher than normal βM (e.g., we use β = 100 here), while the rest of the parameters
will be in the same range for both healthy and PD individuals.
Frontotemporal Dementia (bvFTD). Patients with bvFTD are prone to overeating which
may represent increased reward representation. To model this impairment in bvFTD patients, the
parameter of the model could be modified as follow: αM << α (e.g., α = 100 as shown in Table
2), where α is the parameter of the bvFTD model has, and the rest of these parameters are equal
to the normal one.
Alzheimer’s disease (AD). To model apathy in patients with Altzheimer’s, including downplaying rewards and losses, we will assume that the parameters φ and τ are somewhat smaller
than normal, φ < φM and τ < τM (e.g, set to 0.1 in Table 2), which models the tendency to
forget both positive and negative rewards.
ADHD. Recall that ADHD may be involve impairments in storing stimulus-response associations. In our ADHD model, the parameters φ and τ are smaller than normal, φM > φ and
τM > τ , which models forgetting of both positive and negative rewards. Note that while this
model appears similar to Altzheimer’s model described above, the forgetting factor will be less
Table 2: Datasets
UCI Datasets
Instances Classes
Covertype
581 012
7
CNAE-9
1080
9
Internet Advertisements 3279
2
1 025 010
9
Poker Hand
pronounced, i.e. the φ and τ parameters are larger than those of the Altzheimer’s model (e.g., 0.2
instead of 0.1, as shown in Table 2).
Addiction. As mentioned earlier, addiction is associated with inability to properly forget (positive) stimulus-response associations; we model this by setting the weight on previously accumulated positive reward (“memory” ) higher than normal, τ > τM , e.g. τ = 1, while τM = 0.5.
Chronic Pain. We model the reduced responsiveness to rewards in chronic pain by setting
α < αM so there is a decrease in the reward representation, and φ > φM so the negative rewards
are not forgotten (see table 2).
Of course, the above models should be treated only as first approximations of the reward
processing biases in mental disorders, since the actual changes in reward processing are much
more complicated, and the parameteric setting must be learned from actual patient data, which
is a nontrivial direction for future work. Herein, we simply consider those models as specific
variations of our general method, inspired by certain aspects of the corresponding diseases, and
focus primarily on the computational aspects of our algorithm, demonstrating that the proposed
parametric extension of TS can learn better than the baseline TS due to added flexibility.
5 Empirical Evaluation
In order to evaluate the proposed framework empirically and compare its performance with the
standard Thompson Sampling, we used the following four classification datasets from the UCI
Machine Learning Repository1 : Covertype, CNAE-9, Internet Advertisements and Poker Hand.
A brief summary of the datasets is listed in Table 2.
In order to simulate an infinite data stream, we draw samples randomly without replacement,
from each dataset, restarting the process each time we draw the last sample. In each round, the
algorithm receives the reward 1 if the instance is classified correctly, and 0 otherwise. We compute
the total number of classification errors as a performance metric. Note that we do not use the
features (context) here, as we try to simulate the classical multi-arm bandit environment (rather
than contextual bandit), and use the class labels only. As the result, we obtain a non-stationary
environment, since even if P (reward|context) is fixed, switching from a sample to a sample
(i.e., from a context to a context) results into different P (reward) at each time point.
In order to test the ability of our models to reflect decision-making biases in various disorders,
as well as to evaluate the advantages of our model in comparison with the baseline TS, under
different test conditions, we consider the following settings:
- Positive reward environment: we modify the reward function so that the agent receives only
positive rewards (the lines 11 is not executed). This environment allows us to evaluate how our
models deal with positive reward.
1
https://archive.ics.uci.edu/ml/datasets.html
Table 3: Average Results
Addiction ADHD Alzheimer’s Chronic Pain bvFTD Parkinson M
Datasets
Positive Environment
51.46
Negative Environment
62.83
Normal Reward Environment 52.81
52.35 52.53
55.06 55.54
53.68 51.48
52.88
55.48
53.11
59.16 56.23
56.03 56.21
49.55 58.01
TS
52.64 55.62
52.74 61.21
50.92 56.95
Table 4: Positive-reward Environment
Datasets
Internet Advertisements
CNAE-9
Covertype
Poker Hand
Addiction
ADHD
Alzheimer’s Chronic Pain bvFTD
Parkinson
M
TS
34.06 ± 0.34
40.25 ± 0.85
65.04 ± 0.52
66.5 ± 0.24
31.85 ± 3.51
39.89 ± 2.70
66.5 ± 0.75
71.18 ± 0.12
32.40 ± 1.91
40.08 ± 3.69
66.75 ± 1.52
70.19 ± 1.87
43.61 ± 1.51
40.28 ± 2.27
68.05 ± 1.72
73 ± 1.87
37.69 ± 1.88
40.16 ± 1.99
65.01 ± 1.75
67.73 ± 1.87
38.34 ± 1.77
40.06 ± 1.66
67.08 ± 1.23
77.03 ± 1.87
Addiction
ADHD
Alzheimer’s Chronic Pain bvFTD
Parkinson
M
TS
41.346 ± 0.21
40.248 ± 0.35
73.26 ± 0.30
96.51 ± 0.35
37.833 ± 1.20
39.97 ± 0.20
71.28 ± 0.32
71.18 ± 0.22
40.76 ± 1.93
39.89 ± 3.49
71.35 ± 1.75
70.19 ± 2.77
41.4 ± 1.17
39.95 ± 1.73
70.5 ± 1.8
73 ± 1.87
33.22 ± 1.7
39.96 ± 1.33
70.05 ± 1.87
67.73 ± 1.51
38.19 ± 1.6
40.02 ± 1.11
69.93 ± 0.83
96.71 ± 1.16
Addiction
ADHD
Alzheimer’s Chronic pain bvFTD
Parkinson
M
TS
32.28 ± 0.20
40.16 ± 0.38
73.54 ± 0.31
65.29 ± 0.33
36.8 ± 1.28
40.09 ± 0.31
64.27 ± 0.30
73.57 ± 0.33
36.56 ± 1.63
39.99 ± 3.01
63.54 ± 1.30
65.83 ± 2.68
44.69 ± 1.85
40.25 ± 1.71
72.67 ± 1.82
74.44 ± 1.07
33.65 ± 1.81
39.86 ± 1.10
64.61 ± 0.8
65.58 ± 1.62
37.71 ± 0.66
39.78 ± 0.80
64.63 ± 1.87
85.71 ± 1.09
32.96 ± 1.66
39.94 ± 0.73
69.49 ± 1.75
69.14 ± 2.57
55.67 ± 1.68
40.14 ± 2.33
70.62 ± 1.73
70.26 ± 0.81
Table 5: Negative-reward Environment
Datasets
Internet Advertisements
CNAE-9
Covertype
Poker Hand
41.08 ± 1.64
40.27 ± 0.23
71.45 ± 1.87
69.14 ± 0.88
42.633 ± 1.23
39.89 ± 1.33
71.34 ± 1.87
70.26 ± 1.19
Table 6: Normal Reward Environment
Datasets
Internet Advertisements
CNAE-9
Covertype
Poker Hand
35.53 ± 1.43
39.75 ± 0.28
68.69 ± 1.84
68.49 ± 0.92
28.59 ± 1.76
40.13 ± 1.81
63.69 ± 1.85
65.69 ± 1.01
- Negative reward environment: we modify the reward function so that the agent receives only
negative rewards (the lines 10 is not executed). This environment helps to evaluate the negativereward processing by our models.
- Normal environment: the agent can see both negative and positive rewards.
The average error rate results on the UCI datasets, for each type of the environment, and
over 10 runs of each algorithm, are shown in Table 3. We compute the error rate by dividing
the total accumulated regret by the number of iterations. The best results for each dataset are
shown in bold. Note that our parametric approach always outperforms the standatrd TS method:
AD (addiction) model is best in positive reward environment, M (moderate) version is best in
negative environment, and bvFTD happens to outperform other models in regular (positive and
negative) reward environment. While further modeling and validation on human subjects may
be required to validate neuroscientific value of the proposed models, they clearly demonstrate
computational advantages over the classical TS approach for the bandit problem.
We now present the detailed results for all algorithms and for each of the three environments,
in Tables 4, 5, and 6. Lowest errors for each dataset (across each row) are again shown in bold.
Note that, in all three environments, and for each of the four datasets, the baseline Thompson
Sampling was always inferior to the proposed parametric family of methods, for each specific
settings, different versions of our HBTS framework were performing best.
Positive Reward Environment. Table 4 summarizes the results for positive reward setting. Note
that most versions of the proposed approach frequently outperform the standard Thompson sampling. ADHD model yields best results on two datasets out of four, while AD (addiction) and M
(moderate) models are best at one of each remaining datasets, respectively.
Note that PD (Parkinson’s) and bvFTD (behavioral-variant fronto-teporal dementia) yield the
worst results on most datasets. The behavior of PD model is therefore consistent with the literature
on Parkinson’s disease, which suggests, as mentioned earlier, that Parkinson’s patients do not
learn as well from positive rewards as they do from negative ones.
Ranking the algorithms with respect to their mean error rate, we note that the top three performing algorithms were AD (addiction), ADHD and AZ (Alzheimer’s), in that order. One can
hypothesise that these observations are consistent with the fact that those disorders did not demonstrate such clear impairment in learning from positive rewards as, for example, Parkinson’s.
Negative Reward Environment. As shown in Table 5, for negative reward environment, we
again observe that the proposed algorithms alwyas work better than the state of the art Thompson
sampling.
Overall, M (moderate) model performs best in this environment, on three out of four datasets.
Note that PD (Parkinson’s) and CP (chronic pain) models outperform many other models, performing much better with negative rewards than they did woith the positive ones, which is consistent with the literature discussed before. AD (addiction) is the worst-performing out of HBTS
algorithms, which may relate to its bias towards positive-reward driving learning, but impaired
ability to learn from negative rewards.
Ranking the algorithms with respect to their mean error rate, we note that the two bestperforming algorithms were ADHD and AZ (Alzheimer’s), in that order.
Normal Reward Environment. Similarly to the other two environments, the baseline Thompson
Sampling is always inferior to the proposed algorithms, as shown in Table 6). Interestingly, model
M was never a winner, either, and different disorder models performed best for different data sets.
PD and CP show worst performance, suggesting that negative-reward driven learning is impairing.
6 Conclusions
This paper proposes a novel parametric family of algorithms for multi-arm bandit problem,
extending the classical Thompson Sampling approach to model a wide range of potential reward processing biases. Our approach draws an inspiration from extensive literature on decisionmaking behavior in neurological and psychiatric disorders stemming from disturbances of the
reward processing system. The proposed model is shown to consistently outperform the baseline
Thompson Sampling method, on all data and experiment settings we explored, demonstrating better adaptation to each domain due to high flexibility of our multi-parameter model which allows
to tune the weights on incoming positive and negative rewards, as well as the weights on memories about the prior reward history. Our empirical results support multiple prior observations
about reward processing biases in a range of mental disorders, thus indicating the potential of
the proposed model and its future extensions to capture reward-processing aspects across various
neurological and psychiatric conditions. Our future work directions include extending our model
to the more realistic contextual bandit setting, as well as testing the model on human decision
making data.
References
1. Lai, T.L., Robbins, H.: Asymptotically efficient adaptive allocation rules. Advances in Applied Mathematics 6(1) (1985) 4–22
2. Auer, P., Cesa-Bianchi, N., Fischer, P.: Finite-time analysis of the multiarmed bandit problem. Machine
Learning 47(2-3) (2002) 235–256
3. Perry, D.C., Kramer, J.H.: Reward processing in neurodegenerative disease. Neurocase 21(1) (2015)
120–133
4. Frank, M.J., Seeberger, L.C., O’reilly, R.C.: By carrot or by stick: cognitive reinforcement learning in
parkinsonism. Science 306(5703) (2004) 1940–1943
5. Redish, A.D., Jensen, S., Johnson, A., Kurth-Nelson, Z.: Reconciling reinforcement learning models
with behavioral extinction and renewal: implications for addiction, relapse, and problem gambling.
Psychological review 114(3) (2007) 784
6. Luman, M., Van Meel, C.S., Oosterlaan, J., Sergeant, J.A., Geurts, H.M.: Does reward frequency or
magnitude drive reinforcement-learning in attention-deficit/hyperactivity disorder? Psychiatry research
168(3) (2009) 222–229
7. Taylor, A.M., Becker, S., Schweinhardt, P., Cahill, C.: Mesolimbic dopamine signaling in acute and
chronic pain: implications for motivation, analgesia, and addiction. Pain 157(6) (2016) 1194
8. Seeley, W.W., Zhou, J., Kim, E.J.: Frontotemporal dementia: what can the behavioral variant teach us
about human brain organization? The Neuroscientist 18(4) (2012) 373–385
9. Hauser, T.U., Fiore, V.G., Moutoussis, M., Dolan, R.J.: Computational psychiatry of adhd: neural gain
impairments across marrian levels of analysis. Trends in neurosciences 39(2) (2016) 63–73
10. Dezfouli, A., Piray, P., Keramati, M.M., Ekhtiari, H., Lucas, C., Mokri, A.: A neurocomputational
model for cocaine addiction. Neural computation 21(10) (2009) 2869–2893
11. Hess, L.E., Haimovici, A., Muñoz, M.A., Montoya, P.: Beyond pain: modeling decision-making deficits
in chronic pain. Frontiers in behavioral neuroscience 8 (2014)
12. Auer, P., Cesa-Bianchi, N.: On-line learning with malicious noise and the closure algorithm. Ann.
Math. Artif. Intell. 23(1-2) (1998) 83–99
13. Auer, P., Cesa-Bianchi, N., Freund, Y., Schapire, R.E.: The nonstochastic multiarmed bandit problem.
SIAM J. Comput. 32(1) (2002) 48–77
14. Bouneffouf, D., Féraud, R.: Multi-armed bandit problem with known trend. Neurocomputing 205
(2016) 16–21
15. Chapelle, O., Li, L.: An empirical evaluation of thompson sampling. In: Advances in neural information
processing systems. (2011) 2249–2257
16. Thompson, W.: On the likelihood that one unknown probability exceeds another in view of the evidence
of two samples. Biometrika 25 (1933) 285–294
17. Agrawal, S., Goyal, N.: Analysis of thompson sampling for the multi-armed bandit problem. In: COLT
2012 - The 25th Annual Conference on Learning Theory, June 25-27, 2012, Edinburgh, Scotland. (2012)
39.1–39.26
18. Bouneffouf, D.: Freshness-aware thompson sampling. In: Neural Information Processing - 21st International Conference, ICONIP 2014, Kuching, Malaysia, November 3-6, 2014. Proceedings, Part III.
(2014) 373–380
19. Schulz, E., Konstantinidis, E., Speekenbrink, M.: Learning and decisions in contextual multi-armed
bandit tasks. In: In Proceedings of the 37th annual conference of the cognitive science society. (2015)
2204–2212
20. Thompson, W.R.: On the likelihood that one unknown probability exceeds another in view of the
evidence of two samples. Biometrika 25(3/4) (1933) 285–294
| 2cs.AI
|
A short note on Simulation and Abstraction
Chris Hankin
Institute for Security Science and Technology
Imperial College London, UK
[email protected]
This short note is written in celebration of David Schmidt’s sixtieth birthday. He has now been active
in the program analysis research community for over thirty years and we have enjoyed many interactions with him. His work on characterising simulations between Kripke structures using Galois
connections was particularly influential in our own work on using probabilistic abstract interpretation to study Larsen and Skou’s notion of probabilistic bisimulation. We briefly review this work and
discuss some recent applications of these ideas in a variety of different application areas.
1
Introduction
Since his earliest contributions on state transition machines for lambda calculus expressions [13], David
Schmidt has been at the forefront of research in programming language theory, particularly program
analysis.
His work on program analysis started through his collaboration with Neil Jones. No doubt partly
inspired by Patrick and Radhia Cousot’s work on abstract interpretation [2], he has made a study of
various aspects of Galois connections. An early contribution was [7]. A later example, which was
influential in our own work, was [15] where he shows how to characterise simulation relations using
Galois Connections.
The early work of our group was also inspired by the Cousots [1]. Over the last fourteen years,
we have been working on the analysis of probabilistic and quantitative programming languages and
systems [12]. This has led to a framework called probabilistic abstract interpretation (PAI). Rather than
using lattices and Galois Connections, PAI uses Hilbert Spaces and Moore-Penrose Pseudo Inverses.
Inspired by [15], we have used PAI to characterise probabilistic bisimulation [8, 10]; we demonstrated
the probabilistc analogue of Dave’s earlier results which amounted to characterising Larsen and Skou’s
probabilistic bisimulation [5] using Moore-Penrose Pseudo Inverses. A major feature of our approach is
that it becomes natural to introduce a notion of approximate bisimulation – this has proved to be very
useful in studies of language-based security [9].
The author first met David Schmidt nearly thirty years ago. He visited Imperial College whilst
developing the work in [7], his visit coincided with our own work on the ideas in [1]. We found a lot
to talk about and it was the start of many further interactions. In addition to his scientific contributions,
he has always given time to developing text books; his [14] was used to educate several generations of
Imperial students in the principles of programming language design. The author has also worked with
Dave on many programme committees; perhaps the high point was when they were co-General Chairs
of POPL in 2001. Dave has a had a long and successful career and we wish him many more years. The
author would like to add his congratulations to Dave on this important milestone.
A. Banerjee, O. Danvy, K.-G. Doh, J. Hatcliff (Eds):
David A. Schmidt’s 60th Birthday Festschrift
EPTCS 129, 2013, pp. 337–340, doi:10.4204/EPTCS.129.20
Simulation and Abstraction
338
2
Simulation and Galois Connections
In [15], Dave addresses, inter alia, the question of characterising simulation relations between Kripke
structures using Galois connections.
Recall that we define (L, α , γ , M) to be a Galois connection between the complete lattices (L, ⊑) and
(M, ⊑) if and only if
α : L → M and γ : M → L are monotone functions
that satisfy:
γ ◦ α ⊒ idL
α ◦ γ ⊑ idM
For Kripke structures C = hΣC , →C , IC i and A = hΣA , →A , IA i, a binary relation R ⊆ ΣC × ΣA is a
simulation of C by A (C ⊳R A), if for every c ∈ ΣC , a ∈ ΣA :
if c R a and c →C c′ then ∃a′ ∈ ΣA [a →A a′ and c′ R a′ ]
In the framework studied in [15], the concrete Kripke structures are often infinite state structures
representing programs and the abstract Kripke structures are some program analysis. In this setting,
we should abstract sets of states of the concrete structure to a single state in the abstract structure.
Given a Galois connection, (α : P(ΣC ) → ΣA , γ : ΣA → P(ΣC )), we can construct the relation, R(α ,γ ) ⊆
P(ΣC ) × ΣA :
S R(α ,γ ) a if and only if α (S) ⊑A a
This can be shown to be the suitable basis for a simulation relation.
Whilst [15] achieves much more than described here, it was these ideas that inspired our own work
on probabilistic bisimulation (originally introduced by Larsen and Skou [5]) for reactive systems (also
called fully probabilistic systems):
Definition 1 A probabilistic bisimulation is an equivalence relation ∼b on states of a probabilistic transition system satisfying for all actions a ∈ A:
p ∼b q and p →a π ⇔ q →a ρ and π ∼b ρ .
where π and ρ are distributions of states.
.
3
Probabilistic Bisimulation and Moore-Penrose Pseudo Inverses
In [10, 8] we introduced an approximate version of bisimulation and confinement where the approximation can be used as a measure ε for the information leakage of the system under analysis. We represented
probabilistic transition systems by linear operators, i.e. by their transition matrices M. In the case
of probabilistic programs and systems these matrices M are the usual well known stochastic matrices
which are the generators of the corresponding Markov chains (for details see [10, 8]).
We then showed that two systems M1 and M2 are bisimilar if there exist simplified, or abstracted,
versions of M1 and M2 , represented by matrices M#1 and M#2 , such that M#1 = M#2 . In the probabilistic
Chris Hankin
339
abstract interpretation setting that we use, a bounded linear operator and its Moore-Penrose Pseudo
Inverse are the analogue of the adjoined pair of monotonic functions in a Galois insertion. The abstract
systems are obtained by lumping states, i.e. by identifying each concrete state si with a class C j of states
which are all behavioural equivalent to each other.
Concretely, we compute this via n × m matrices K (where n is the number of concrete states and m
the number of abstract classes) with Kij = 1 iff si ∈ C j and 0 otherwise. We refer to such matrices which
have exactly one entry 1 in each row while all other entries are 0 as classification matrices, and denote
the set of all classification matrices by K . The abstract systems are then given by M#i = K†i Mi Ki with
Ki some classification matrix and † constructing the so called Moore-Penrose pseudo-inverse – in the
case of classification matrices K† can be constructed as the row-normalised transpose of K.
The problem of showing that two systems M1 and M2 are behaviourally equivalent, i.e. are (probabilistically) bisimilar, is now translated into finding two classification matrices Ki ∈ K such that
M#1 = K†1 M1 K1 = K†2 M2 K2 = M#2 .
In case that two systems are not bisimilar we can still define a quantity ε which describes how
(non-)bisimilar the two systems are. This ε is formally defined in terms of the norm of a linear operator
representing the partition induced by the ‘minimal’ bisimulation on the set of the states of a given system,
i.e. the one minimising the observational difference between the system’s components (see again [10]
for further details, in particular regarding labeled probabilistic transition systems):
Definition 2 Let M1 and M2 be the matrix representations of two probabilistic transition systems. We
say that M1 and M2 are ε -bisimilar, denoted by M1 ∼εb M2 , iff
inf
K1 ,K2 ∈K
kK†1 M1 K1 − K†2 M2 K2 k = ε
where k.k denotes an appropriate norm, e.g. the supremum norm k.k∞ .
In [10] we show that, when ε = 0 this gives the standard notion of probabilistic bisimulation.
4
Conclusion
This short note has sketched some early work by David which forms part of his deep study of the use
of Galois connections and relations in program analysis. Our own work on characterising probabilistic
bisimulation using probabilistic abstract interpretation has found a number of applications, including:
• the detection and removal of timing channels in probabilistic transition systems [11] – we study a
concept called probabilistic time bisimilarity and use it to detect timing channels;
• the detection of sub-communities in social media [6] – we evaluate a number of algorithms including one using the notion of stability from [3] which effectively lumps nodes together if their
mutual interactions are “stronger” than interactions outside the group; and
• the abstraction of stochastic and Bayesian games to provide decision support in cyber security [4] –
where we hope to apply probabilistic abstract interpretation directly to the underlying probabilistic
transition systems in the games, thereby developing a principled way of reducing the state spaces
to achieve tractability of game solutions.
We look forward to discussing some of this work with David in the future but, in the meantime, reiterate
our best wishes on this important anniversary.
Simulation and Abstraction
340
5
Acknowledgements
Much of the work discussed above was done in collaboration with Alessandra Di Pierro and Herbert
Wiklicky. More recently, I have enjoyed working on the application of these ideas to other areas with
Erwan Le Martelot and Pasquale Malacaria.
References
[1] Geoffrey L. Burn, Chris Hankin & Samson Abramsky (1986): Strictness Analysis for Higher-Order Functions. Sci. Comput. Program. 7(3), pp. 249–278, doi:10.1016/0167-6423(86)90010-9.
[2] Patrick Cousot & Radhia Cousot (1977): Abstract Interpretation: A Unified Lattice Model for Static
Analysis of Programs by Construction or Approximation of Fixpoints. In: POPL, ACM, pp. 238–252,
doi:10.1145/512950.512973.
[3] Jean-Charles Delvenne, Sophia Yaliraki & Mauricio Barahona (2010): Stability of graph communities across
time scales. Proc. Nat. Acad. Sci. 107(29), pp. 12755–12760, doi:10.1073/pnas.0903215107.
[4] Chris Hankin & Pasquale Malacaria (2013): Payoffs, Intensionality and Abstraction in Games. In: Computation, Logic, Games, and Quantum Foundations - The Many Facets of Samson Abramsky, Lecture Notes in
Computer Science 7860, Springer, doi:10.1007/978-3-642-38164-5-6.
[5] Kim Guldstrand Larsen & Arne Skou (1989): Bisimulation Through Probabilistic Testing. In: POPL, ACM,
pp. 344–352, doi:10.1145/75277.75307.
[6] Erwan Le Martelot & Chris Hankin (2013): Fast Multi-Scale Detection of Relevant Communities in LargeScale Networks. Computer Journal, doi:10.1093/comjnl/bxt002.
[7] Austin Melton, David A. Schmidt & George E. Strecker (1986): Galois Connections and Computer
Science Applications. In: CTCS, Lecture Notes in Computer Science 240, Springer, pp. 299–312,
doi:10.1007/3-540-17162-2-130.
[8] Alessandra Di Pierro, Chris Hankin & Herbert Wiklicky (2003): Quantitative Relations and Approximate
Process Equivalences. In: CONCUR, Lecture Notes in Computer Science 2761, Springer, pp. 498–512,
doi:10.1007/978-3-540-45187-7-33.
[9] Alessandra Di Pierro, Chris Hankin & Herbert Wiklicky (2004): Approximate Non-interference. Journal of
Computer Security 12(1), pp. 37–82.
[10] Alessandra Di Pierro, Chris Hankin & Herbert Wiklicky (2005): Measuring the confinement of probabilistic
systems. Theor. Comput. Sci. 340(1), pp. 3–56, doi:10.1016/j.tcs.2005.03.002.
[11] Alessandra Di Pierro, Chris Hankin & Herbert Wiklicky (2011): Probabilistic timing covert channels: to
close or not to close? Int. J. Inf. Sec. 10(2), pp. 83–106, doi:10.1007/s10207-010-0107-0.
[12] Alessandra Di Pierro & Herbert Wiklicky (2000): Concurrent constraint programming: towards probabilistic
abstract interpretation. In: PPDP, ACM, pp. 127–138, doi:10.1145/351268.351284.
[13] David A. Schmidt (1980): State transition machines for lambda calculus expressions. In: SemanticsDirected Compiler Generation, Lecture Notes in Computer Science 94, Springer, pp. 415–440,
doi:10.1007/3-540-10250-7-32.
[14] David A. Schmidt (1986): Denotational semantics: a methodology for language development. Allyn and
Bacon.
[15] David A. Schmidt (1999): Binary relations for abstraction and refinement. In: Workshop on Refinement and
Abstraction.
| 6cs.PL
|
Vulnerability of Deep Learning
Richard Kenway∗
arXiv:1803.06111v1 [stat.ML] 16 Mar 2018
School of Physics and Astronomy, University of Edinburgh,
James Clerk Maxwell Building, Peter Guthrie Tait Road, Edinburgh EH9 3FD, UK
(Dated: March 19, 2018)
The Renormalisation Group (RG) provides a framework in which it is possible to assess whether
a deep-learning network is sensitive to small changes in the input data and hence prone to error,
or susceptible to adversarial attack. Distinct classification outputs are associated with different RG
fixed points and sensitivity to small changes in the input data is due to the presence of relevant
operators at a fixed point. A numerical scheme, based on Monte Carlo RG ideas, is proposed for
identifying the existence of relevant operators and the corresponding directions of greatest sensitivity
in the input data. Thus, a trained deep-learning network may be tested for its robustness and, if it
is vulnerable to attack, dangerous perturbations of the input data identified.
I.
A.
INTRODUCTION
The Threat to Deep Learning
Despite many successful applications, deep learning [1–
3] is poorly understood. The absence of an underlying
theory means that we cannot be certain under what circumstances a given trained network will operate correctly
and this uncertainty calls into question the use of such
networks in safety-critical applications. There is also the
possibility that slightly perturbed input data can be constructed that would fool the network, but not a human,
making the network susceptible to adversarial attack, or
simply cause it to misclassify and thereby conceal information in the data [4].
It is important to distinguish between ambiguous data,
which any system, machine or human, would find difficult
to classify and data that has been imperceptibly altered
to cause the deep network to misclassify it. This paper
introduces a theoretical framework, based on the Renormalisation Group (RG), which specifies the conditions
under which the latter may occur and proposes a calculational method for testing a trained network for this
vulnerability. Further, it determines the directions in the
space of input data along which small perturbations are
amplified by such a vulnerable deep network, causing it
to generate a qualitatively incorrect output.
The theoretical framework captures the salient features
of deep learning: the multi-layer structure, trained one
layer at a time. It is demonstrated for the specific case
in which each layer is a Restricted Boltzmann Machine
(RBM) [3] and the network is trained to classify a data
set {x} in terms of outputs {y}, whose exact conditional
probability distribution is p(y|x). However, the approach
may be extended straightforwardly to other deep networks and to data generation as well as classification.
∗
Email address: [email protected]
B.
Analogy with the Renormalisation Group
The formal analogy between deep learning and the RG
has been noted before, although with some controversy
about whether the context should be unsupervised, or
supervised learning [5–7]. There are more fundamental
issues.
The first is that the RG applied to critical phenomena describes how features of the microscopic degrees of
freedom of a system (analogous to the input data {x})
depend on length scale, successively eliminating shortdistance fluctuations and generating a sequence of effective Hamiltonians, which describe the surviving features
on larger and larger length scales. A critical point is associated with scale-invariant fluctuations and convergence
of the sequence of effective Hamiltonians to a fixed-point
Hamiltonian which describes them. Although many applications of deep learning do involve data which exhibit
structure on different length scales, e.g., data drawn from
the physical world, and this appears to be captured by
deep learning so that it is “cheap” [6], this need not be
the case.
The second is that the RG transformation effected by
each layer of the network may be different, reflecting
the emergence of different types of features in the data,
whereas in critical phenomena each application of the
RG transformation changes the length scale by a fixed
amount.
We will assume that training of the deep network converges to a conditional probability distribution that is a
good approximation to p(y|x), provided there is sufficient
number of layers. Thus, the sequence of RG transformations corresponding to applying the RBM in each layer to
the output from the hidden nodes in the previous layer,
starting with the input data, converges.
Presumably, some more generalised form of scaling
takes place in deep learning and we will assume that this
has the effect of successively scaling to smaller values all
but a finite number of eigenvalues of the Fisher Information Matrix (FIM) of the trained deep network, so that
its spectrum is sloppy [8]. We will say more about this
in Section IV.
In Section II we introduce the RG formalism in the
2
context of a deep RBM network used for classification
and, in particular, define the sequence of effective Hamiltonians whose fixed points and universality classes define
distinct outputs. In Section III we develop the computational scheme, based on Monte Carlo RG [9, 10], for
estimating the stability of these fixed points with respect
to the depth of the network. Section IV relates the instability of a particular fixed point to direction(s) in the
input data space {x}, along which small changes in the
data cause major changes in the output probability distribution and, hence, are sources of vulnerability.
II.
RENORMALISATION GROUP FOR DEEP
LEARNING
Consider a layered network of N RBMs with input
nodes h0 , N −1 layers of hidden nodes hk , k = 1, . . . , N −
1, and output nodes hN . For example, the hidden nodes
may be binary vectors of the same dimension as the input
data {x} and the outputs {y}, and the joint probability
distribution for the k th layer is of the form [3]
T
tk (hk , hk−1 ) =
T
T
ehk Wk hk−1 +ak hk +bk hk−1
,
zk
T
T
T
zk = Trhk ,hk−1 ehk Wk hk−1 +ak hk +bk hk−1 , (1)
where Wk , ak and bk are weights determined by the
layerwise training. We will not require the precise form
of tk .
The trained deep RBM network generates the probability distribution for the output y given input data x,
qN (y|x), iteratively for k = 1, . . . , N through
qk (hk |x) = Trhk−1 tk (hk |hk−1 )qk−1 (hk−1 |x),
q0 (h0 |x) = δh0 ,x ,
qN (y|x) ≈ p(y|x),
(2)
At this point, it is worth noting that the input data x
determines the couplings in the space of effective Hamiltonians via the first layer,
(1)
e−Hx
(1)
Zx
tk (hk , hk−1 )
.
Trhk tk (hk , hk−1 )
(3)
We define the effective Hamiltonian for the k
by
(k)
qk (h|x) =
e−Hx
th
layer
(h)
,
(k)
Zx
(k)
Zx(k) = Trh e−Hx
(h)
(4)
.
Then, the sequence of Hamiltonians is generated by the
RBMs in each layer via
(k+1)
e−Hx
(h)
(k+1)
Zx
(k)
= Trh′ tk+1 (h|h′ )
e−Hx
(h′ )
(k)
Zx
,
(5)
such that each Hamiltonian depends only on the previous
one in the sequence. This has the same form as an RG
transformation with kernel tk+1 [5].
= t1 (h|x),
(6)
and the couplings in each subsequent layer are functions
solely of the couplings in the previous layer.
We parametrise the space of effective Hamiltonians in
terms of a complete set of operators Oα (h), which, in the
case of binary variables, consist of all possible products
of the components of h, with couplings gα . Thus, the
effective Hamiltonian for layer k is
X
Hx(k) (h) =
gα(k) Oα (h).
(7)
α
Then the effect of the RG transformation in eq. (5) is
to define {g (k+1) } solely in terms of {g (k) } and, thereby,
to generate a flow in the coupling-constant space of the
effective Hamiltonians.
The key assumption we make is that the training of
the deep RBM network converges, so that, for N large
enough, the sequence of effective Hamiltonians converges
to a fixed point:
X
Hx∗ (h) =
(8)
gα∗ Oα (h),
α
such that
∗
e−Hx (y)
= qN (y|x) ≈ p(y|x).
Zx∗
(9)
The analogue of the RG universality class of the fixed
point Hx∗ is the subset of the input data, Cx , which
are associated with the same output distribution, p(y|x).
Thus, assuming the training correctly classifies the data,
where, according to Bayes Theorem,
tk (hk |hk−1 ) =
(h)
x, x′ ∈ Cx ⇒ Hx∗ = Hx∗′
(10)
p(y|x) 6= p(y|x′ ) ⇒ Hx∗ 6= Hx∗′ ,
(11)
and
so that there must be more than one fixed point. In
any non-trivial classification problem, there is a distinct
fixed point of the correctly trained deep network for every
distinct output.
Next, we consider the flow in coupling-constant space
close to a fixed point and, in particular, the stability
properties of the fixed point as the number of layers is
increased. We assume that the deeper the network the
more accurately it is able to learn the classification problem. Close to the fixed point {g ∗ }, the relationship between the couplings in adjacent layers is, to leading order,
gα(k+1) ({g (k) }) = gα(k+1) ({g ∗ })
(k+1)
X (k)
∂gα
+
(gβ − gβ∗ )
(k)
∂gβ
β
. (12)
{g∗ }
3
(k+1)
Since gα
({g ∗ }) = gα∗ , this becomes
X
(k)
∗
gα(k+1) − gα∗ =
Tαβ
(gβ − gβ∗ ),
(13)
β
where
(k+1)
∗
Tαβ
=
∂gα
(14)
(k)
∂gβ
{g∗ }
is the stability matrix of the fixed point. The stability
properties of a fixed point determine whether it becomes
sensitive to small changes in the input data as the number
of layers is increased.
Define the eigenvalues and eigenvectors of the stability
matrix by
X
(µ)
∗
(15)
Tαβ
φβ = Λµ φ(µ)
α
the hidden nodes and output nodes for a given input,
but don’t enable us to construct the sequence of effective Hamiltonians, or the fixed-point Hamiltonian. In
the next section, we will explain how Monte Carlo RG
methods may be used to estimate the stability matrix
∗
Tαβ
, defined in eq. (14), and how its eigenvalues behave
in deep networks close to the fixed point. This will be
sufficient to determine whether a particular trained deep
network has any fixed points with relevant operators.
It is worth noting that the existence of multiple fixed
points (associated with a non-trivial classification problem) does not imply the existence of relevant operators
at any of the fixed points. In the robust situation, where
no fixed point has a relevant direction, data close to the
boundary between two universality classes are simply ambiguous and do not represent a source of adversarial attack (because a human would be equally likely to misclassify them).
β
and express the couplings, gα , and the corresponding operators, Oα (h), in the basis of eigenvectors of the stability
matrix as
X
gα =
(16)
φ(µ)
α g̃µ ,
µ
Õµ (h) =
X
φ(µ)
α Oα (h).
(17)
α
Then the effective Hamiltonian in this basis is
X
Hx(k) =
g̃µ(k) Õµ (h)
(18)
µ
and, close to the fixed point, from eqs (13) and (15),
X
Hx(k+1) − Hx∗ =
(g̃µ(k+1) − g̃µ∗ )Õµ (h)
III.
STABILITY OF DEEP LEARNING
Having trained a deep network of RBMs on a classification problem whose exact solution is p(y|x), we have
available a sequence of RBMs given, for example, by
eq. (1), which enables us to use the conditional probability distributions in eq. (2) to sample the hidden nodes
in each layer and the output layer. Using this capability,
we apply methods developed for Monte Carlo RG [9, 10]
to estimate the stability matrix for each fixed point in
the space of effective Hamiltonians.
Denote by hOγ i(k) the expectation value of the operator Oγ with respect to the effective Hamiltonian in layer
k, eq. (4), i.e.,
hOγ i(k) = Trh Oγ (h)qk (h|x)
µ
=
X
Λµ (g̃µ(k) − g̃µ∗ )Õµ (h).
(k)
=
(19)
Trh Oγ (h)e−Hx
.
(k)
Zx
µ
In the language of the RG, if Λµ > 1, Õµ (h) is relevant
and, if Λµ < 1, it is irrelevant. In the context of deep
learning, the existence of a relevant operator at a fixed
point creates a sensitivity to input data which generate a
non-zero coupling to this operator. Such data will lead to
a probability distribution for the output which increasingly deviates from the fixed-point distribution as the
network is made deeper (more layers are added). This
is the source of vulnerability of deep networks to tiny
changes in the input data which excite a relevant operator, leading to the sequence of effective Hamiltonians
converging to the wrong fixed point, i.e., misclassifying
the data.
As it stands, this provides an explanation of the potential vulnerability of deep networks, but doesn’t help
to diagnose whether a particular network is susceptible.
This is because, in practice, we only have access to the
RBM weights, e.g., eq. (1), which enable us to sample
(h)
(20)
We start from the chain rule applied to hOγ i(k+1) close
to a fixed point (i.e., k ≈ N for N large),
∂hOγ i(k+1)
(k)
∂gβ
=
X ∂gα(k+1) ∂hOγ i(k+1)
(k)
α
=
X
α
∂gβ
(k+1)
∂gα
(k+1)
(k+1) ∂hOγ i
.
(k+1)
∂gα
Tαβ
(21)
(k+1)
Here Tαβ
is the approximation to the stability matrix
∗
Tαβ
in eq. (14) obtained from the k th layer, which should
∗
converge to Tαβ
for large k. Differentiating eq. (20) and
using eq. (7), we obtain for the term on the RHS of
eq. (21)
∂hOγ i(k+1)
(k+1)
∂gα
= hOγ i(k+1) hOα i(k+1) − hOγ Oα i(k+1) . (22)
4
To obtain a similar expression for the term on the LHS of
eq. (21), we need to use tk+1 , the RBM for layer k + 1, to
(k+1)
(k)
relate Hx
to Hx , as in eq. (5). In Monte Carlo RG,
this is where the RG blocking step enters. We obtain
∂hOγ i(k+1)
(k)
∂gβ
= hOγ i(k+1) hOβ i(k)
−Trhh′ Oγ (h)tk+1 (h|h′ )Oβ (h′ )qk (h′ |x)
= hOγ i(k+1) hOβ i(k) − hOγ tk+1 Oβ i(k) .
(23)
Thus, with a suitable choice of operators, given a specific input x, by sampling the hidden nodes in the upper
layers of the network (large k) and computing a set of
expectation values of these operators, we can construct
a set of linear equations for the elements of the stability
(k+1)
∗
matrix, Tαβ
≈ Tαβ
,
X
α
(k+1)
Tαβ
h
i
hOγ i(k+1) hOα i(k+1) − hOγ Oα i(k+1)
= hOγ i(k+1) hOβ i(k) − hOγ tk+1 Oβ i(k) .
(24)
Hence, we may obtain a set of successive approximations
(µ)
for the eigenvalues Λµ and eigenvectors φα in eq. (15).
By varying the subset of operators used in eq. (24), we
can test the convergence of the estimates of the largest
eigenvalues of the stability matrix associated with a given
layer of the network and, by computing the eigenvalues
from successive layers, we can investigate whether increasing depth of the network is associated with scaling
behaviour of the corresponding operators, as in eq. (19).
If this scaling behaviour is due to an eigenvalue of magnitude greater than one, then even a tiny component of
the initial data, which creates a non-zero coupling to the
corresponding eigenvector, will cause instability of the
fixed point if the network is deep. In the next section,
we will explain how to identify the specific perturbation
of the data which does this.
admixture of this component in the data will lead to erroneous classification if the network is too deep.
The FIM is defined in terms of the Kullback-Liebler
divergence, DKL , which is itself a measure of how distinguishable two probability distributions p1 and p2 are,
using data sampled from p1 :
p1 (x)
DKL (p1 ||p2 ) = Trx p1 (x) log
.
(25)
p2 (x)
This enables us to measure the dependence of qN (y|x) on
x by considering two nearby data points, x and x′ , and
expanding the Kullback-Liebler divergence as a Taylor
series in their difference [8],
qN (y|x)
′
DKL (qN (y|x)||qN (y|x )) = Try qN (y|x) log
qN (y|x′ )
X
1
=
(x′i − xi )Fij(N ) (x′j − xj )
2 ij
+...,
(26)
where, using eqs (4), (7) and (20),
∂2
DKL (qN (y|x)||qN (y|x′ ))
∂x′i ∂x′j
x′ =x
(N )
(N )
(N )
(N )
(N )
(N )
∂Hx ∂Hx
∂Hx(N )
∂Hx
=
−
∂xi ∂xj
∂xi
∂xj
X ∂g (N ) ∂g (N′ )
α
α
[hOα Oα′ i(N ) − hOα i(N ) hOα′ i(N ) ]
=
∂xi ∂xj
′
Fij(N ) =
αα
(27)
is the FIM.
Having chosen a subspace of operators, {Oα }, in which
to express the effective Hamiltonians in each layer of the
network (via eq. (7)), such that sufficiently precise estimates of the largest eigenvalues of the stability matrix
(k+1)
associated with each layer, Tαβ , are obtained by the
method in Section III, we can compute the FIM using
the chain rule:
(1)
IV.
(N )
i ∂g
Xh
∂gα
β
=
.
T (N ) . . . T (2)
∂xi
αβ ∂xi
GENERATION OF ADVERSARIAL DATA
(28)
β
We consider the exact conditional probability distribution, p(y|x), and hence the trained network, qN (y|x), to
be parametrised by the input data, x, via the fixed-point
Hamiltonian in eqs (8) and (9). The Fisher Information
Matrix (FIM) [8, 11] is a metric which measures how a
probability distribution changes along different directions
in parameter space. If the fixed-point Hamiltonian has
an unstable direction due to a relevant operator, then
this will correspond to the direction in parameter space
along which the probability distribution changes fastest,
i.e., the stiffest direction. Hence, to generate adversarial data we need to compute the eigenvector associated
with the largest eigenvalue of the Fisher information matrix for the fixed-point probability distribution. A tiny
The last derivative may be computed from eq. (6) and
the explicit form of the RBM for the first layer, e.g.,
T
eq. (1), by expressing hT
1 W1 x + a1 h in terms of {Oα }
(1)
to determine gβ (x). Alternatively, we may solve the
system of linear equations in eq. (24) for the first layer:
X
α
(1)
i
∂gα h
hOγ i(1) hOα i(1) − hOγ Oα i(1)
∂xi
= Trh Oγ (h)
∂
t1 (h|x).
∂xi
(29)
Thus, computing the FIM involves translating the input
data into the chosen coupling-constant subspace via the
5
RBM in the first layer, taking the product of the stability matrices associated with each layer in eq. (28), and
computing a set of expectation values of operators in this
subspace using the full network. The fact that the FIM
is built from the product of stability matrices connects
the scaling behaviour in eqs (13) and (19) with the sloppiness of the FIM spectrum observed in [8], i.e., all but a
fixed number of eigenvalues of the FIM are scaled to very
small values, resulting in a fixed number of stiff (relevant)
directions and many sloppy (irrelevant) directions.
We may then obtain the eigenvector corresponding to
the largest eigenvalue of the FIM. For deep networks
which have an unstable fixed point, small perturbations
of the corresponding input data in the direction of this
eigenvector of the FIM are likely to result in the data
being misclassified.
V.
CONCLUSIONS
In this paper we have extended the formal analogy between deep learning and the RG. This enabled us to interpret the classification problem in terms of a sequence of
effective Hamiltonians associated with the layers of the
network, whose couplings are determined by the input
data, and the convergence of this sequence to a distinct
fixed point for each distinct output of the classification
problem. Input data with the same classification should
all result in a flow in the coupling-constant space of these
effective Hamiltonians which converges to the same fixed
[1] G.E. Hinton, S. Osindero and Y-W. Teh, Neural Comput., 18, 1527 (2006).
[2] Y. LeCun, Y. Bengio and G. Hinton, Nature, 521, 436
(2015).
[3] R. Salakhutdinov, Annu. Rev. Stat. Appl., 2, 361 (2015).
[4] X. Yuan, P. He, Q. Zhu, R.R. Bhat and X. Li, Adversarial Examples: Attacks and Defenses for Deep Learning,
arXiv:1712.07107.
[5] P. Mehta and D.J. Schwab, An exact mapping between the
Variational Renormalization Group and Deep Learning,
arXiv:1410.3831.
point.
This exposed the possibility that a fixed point might
have an unstable direction, associated with a relevant operator in the language of the RG. Small perturbations of
the input data that create a non-zero overlap with such
a relevant operator tend to be amplified exponentially by
successive layers of the network, so that the flow diverges
from the correct fixed point and the data may be misclassified. Fixed points with no relevant operators are
not prone to this sort of error and will only misclassify
data that is ambiguous even to a human.
Convergence of the sequence of effective Hamiltonians
to a fixed point for a given subset of the input data is
guaranteed by the training (assuming this converges to a
good approximation to the exact conditional probability
distribution). The existence of a relevant direction at the
fixed point is associated with sensitivity to perturbations
which take the data outside of this subset. The size of
the amplification factor for the perturbation is greater
the greater the depth of the network. Hence, the vulnerability to misclassification is directly due to the depth of
a network.
The key result of this paper is a method for diagnosing
whether a given trained network is vulnerable to adversarial attack, or simply misclassification due to noise in
the data, based only on computing expectation values of
operators in the hidden and output layers. The computations are fairly demanding, but may be justified for a
network that is intended for use in safety-critical applications.
[6] H.W. Lin, M. Tegmark and D. Rolnick, J. Stat. Phys.,
168, 1223 (2017).
[7] D.J. Schwab and P. Mehta, Comment on “Why does deep
and cheap learning work so well?”, arXiv:1609.03541.
[8] B.B. Machta, R. Chachra, M. Transtrum and J.P.
Sethna, Science 342, 604 (2013).
[9] S-K. Ma, Phys. Rev. Lett., 37, 461 (1976).
[10] R.H. Swendsen, Phys. Rev. Lett., 42, 859 (1979).
[11] A. Raju, B.B. Machta and J.P. Sethna, Information geometry and the renormalization group, arXiv:1710.05787.
| 2cs.AI
|
Data-flow Analysis of Programs with Associative Arrays∗
David Hauzar, Jan Kofroň, and Pavel Baštecký
Department of Distributed and Dependable Systems
Faculty of Mathematics and Physics, Charles University in Prague
Czech Republic
{hauzar, kofron}@d3s.mff.cuni.cz, [email protected]
Dynamic programming languages, such as PHP, JavaScript, and Python, provide built-in data structures including associative arrays and objects with similar semantics—object properties can be created at run-time and accessed via arbitrary expressions. While a high level of security and safety of
applications written in these languages can be of a particular importance (consider a web application storing sensitive data and providing its functionality worldwide), dynamic data structures pose
significant challenges for data-flow analysis making traditional static verification methods both unsound and imprecise. In this paper, we propose a sound and precise approach for value and points-to
analysis of programs with associative arrays-like data structures, upon which data-flow analyses can
be built. We implemented our approach in a web-application domain—in an analyzer of PHP code.
1
Introduction
Dynamic languages are often used for the development of security-critical applications, in particular
web applications. To assure a reasonable level of reliability, various static analyses such as security
analysis and bug finding are applied. In these cases, data-flow analysis is a necessary prerequisite.
Unfortunately, dynamic features pose major challenges here. For instance, any interprocedural data-flow
analysis needs to track types of variables to determine targets of virtual method calls. This becomes
even more important in the case of languages with dynamic type systems, where types of variables can
be completely unspecified. Moreover, names of methods to be called or files to be included can be
computed at run-time. Next, all these data can be manipulated using built-in dynamic data structures,
such as multi-dimensional associative arrays and objects with similar semantics—object properties can
be created at run-time and accessed via arbitrary expressions, e.g., variables. This happens relatively
often, e.g., in web applications, which manipulate a lot of input.
Tracking values of variables and analysing the shape of these data structures is thus essential for
data-flow analysis of dynamic languages. In this paper we present our approach to these challenges. Our
contribution includes: (1) value and points-to analysis of associative arrays with an arbitrary depth and
accessible using statically unknown values and arbitrary expressions, (2) modeling explicit aliases, and
(3) a prototype implementation.
2
Motivation and Overview
In this section, we show some dynamic features that impact the data-flow analysis and present an
overview of our approach. We use PHP as the representative of a dynamic language; However, other
languages, especially those connected with the development of web applications, provide built-in associative arrays-like data structures. These languages includes Java-script, Python, Ruby, etc. Moreover,
libraries of “ordinary” programming languages emulate some of these features and offer the developer
API behaving in a similar way. Hence, our approach is not limited to PHP. For the illustration of our
concepts, we use the code in Fig. 1 as a running example.
∗ This work was partially supported by the Grant Agency of the Czech Republic project 14-11384S and by Charles University
Foundation grant 431011 and by Charles University institutional funding SVV-2014-260100.
J. Pang and Y. Liu (Eds.): 3rd International Workshop
on Engineering Safety and Security Systems 2014 (ESSS 2014)
EPTCS 150, 2014, pp. 56–70, doi:10.4204/EPTCS.150.6
c D. Hauzar, J. Kofroň, P. Baštecký
This work is licensed under the
Creative Commons Attribution License.
D. Hauzar, J. Kofroň, P. Baštecký
1
2
3
4
5
6
7
8
9
10
57
$any = $ GET[’user input’]; // an arbitrary user input
$alias = 1; $alias2 = 1; $alias3 = 1;
if ($any) {
$arr[$any] = &$alias;
$t = $arr[1]; // t can be either undefined or can
have value 1
$t[2] = 2; // can update also $alias[2] and e.g.
$arr[1][2]
$arr[1][2] = 3;
$arr[1][3] = 4;
$arr[2][3] = 5;
} else {
$arr[$any][2] = 6;
$arr[1][$any] = 7; // can update also some of
variables involved by the previous update
11
12
13
}
14
15
16
17
18
19
20
$arr[2][1] = &$alias2; // $arr[2][1] and $alias2 can be
aliased also with $alias[1]
$arr[2] = &$alias3;
$arr2 = $arr; // deep−copies $arr, including aliases
$arr2[2] = 8; // updates also $arr[2] and $alias3
$arr2[3] = 9; // can update also $arr[3] and $alias
$arr[$any] = arr2;
Figure 1: Running example.
2.1
Variables, Arrays, and Objects
Variables as well as indices and object properties need not be declared. If a specified index exists in an
array, it is overwritten; if not, it is created. At line 7 in Fig. 1, a new array is created in $arr and index 2
is added to this array. Next, at line 8, index 3 is added to this array.
Arrays can have an arbitrary depth. Unfortunately, updates of such structures cannot be decomposed.
That is, splitting the update at line 7 into two updates at lines 5-6 results in different semantics. The first
reason is that the array assignment statement deep-copies the operand. The update at line 6 thus does not
update the array stored at $arr[1], but its copy. The second reason is that while updates create indices
if they do not exist, read accesses do not; while the update at line 7 creates an index containing an array
in $arr[1] in the case it does not exist, the read access at line 5 returns null in this case and the update
at line 6 fails.
The semantics of the PHP object model is similar to the semantics of associative arrays. Objects’
properties need not to be declared. If a non-existing property is written, it is created. As well as indices,
properties can be accessed via arbitrary expressions. Objects can also have an arbitrary depth in the sense
of reference chains. In the following, we describe associative arrays, however, the same principles apply
to objects as well. We write associative arrays-like data structures to emphasize this fact.
2.2
Dynamic Accesses
In dynamic languages, variables, indices of arrays, and properties of objects can be accessed with arbitrary expressions. At line 4 in Fig. 1 the $arr array with an index determined by the variable $any
is assigned; if a given index exists in $arr, it is overwritten; if not, it is created. Therefore, the set of
variables, array indices and object fields is not evident from the code.
An update can involve more than one element and can be statically unknown. The update at line 4 is
statically unknown and thus may or may not influence accesses at lines 5, 7, 8, 9, 15, and 20. Similarly,
line 11 can access index 2 in any index at the first level. In particular, it can access also index 1 at the first
level, which is updated at the following line. That is, reading $arr[1][2] can return either of values 6,
7, and undefined, reading $arr[1][1] can return 7 and undefined, reading $arr[2][2] can return
6 and undefined, and reading $arr[2][1] always returns undefined. Next, after two branches of the
if statement are merged at line 13, reading of $arr[1][2] can return values 6, 7, 3, and undefined.
2.3
Explicit Aliasing
PHP makes it possible for a variable, index of an array, and property of an object to be an alias of another
variable, index, or property. After an update of an element, all its aliases are also updated. Aliasing in
PHP is thus similar to references in C++ in many aspects.
Unlike C++, in PHP each variable, index, and property can be aliased and later un-aliased from its
previous aliases and become an alias of a new element. As an example, the statement at line 16 un-aliases
58
Data-flow Analysis of Programs with Associative Arrays
$arr[2] from its previous aliases. Moreover, a variable can be an alias of another variable only at some
paths to a given program point, e.g., if it is made an alias in a single branch of an if statement.
The statement at line 4 makes variable $alias an alias of a statically unknown index of array
$arr. Hence, the statement at line 7 accesses $arr[1][2] and may also access $alias[2]. Similarly, the statement at line 15 makes $alias2 an alias of $arr[2][1] and may also make it an alias
of $alias[1]. If an array is assigned, it is deep-copied. However, if an index in the source array has
aliases, the set of aliases in the corresponding index in a target array consists of these aliases and the
source index. Consequently, the statement at line 18 updates also $arr[2] and its alias $alias3. Similarly, the statement at line 19 may update also $arr[3] and $alias, because the statement at line 3 may
make these aliases of each other.
2.4
Overview of the Approach
Our approach consists of the following key parts: (1) definition of analysis state, (2) definition of read
accesses to associative arrays, (3) definition of write accesses to associative arrays (i.e. the transfer
function), and (4) definition of merging associative arrays (i.e. the join operator).
The fundamental part of analysis state consists of represenation of associative arrays. Each associative array contains a set of indices including a special index called unknown field. This index stores information that has been written to statically unknown indices of the array. All indices (including unknown
fields) can contain a set of values and also can point to another associative array—the next dimension.
Next, unknown fields always contain value undefined. Note that (multi-dimensional) unknown fields
allow for dealing with statically unknown accesses, however, they pose a challenge both to definition of
new indices and definition of merging associative arrays.
Both indices that are read by a read access and indices that are updated by a write access to a multidimensional associative array are specified using a list of expressions. Each expression specifies an
access to one dimension of the array. At each level, indices corresponding to values of an expression
corresponding to the level are followed. If the expression yields a statically unknown value, all indices
(including the unknown field) are followed. The read access differs from the write access in the way
it handles the case when there is no corresponding index defined. While the read access follows the
unknown field, the write access defines the index. Note that in the latter case, all data that could have
been assigned to the new index using statically unknown updates are copied to this index. The reason is
that these data are stored in unknown fields and unknown fields are not followed by statically known read
accesses. Next, unlike a read access, a write access distinguishes indices that certainly must be updated
and indices that only may be updated. Former indices are strongly updated—original data are replaced
with new data, the latter indices are weakly updated—original data are joined with new data.
Example 1: At line 5, value 1 is used to read-access the first dimension of an associative array.
Because the array has no index corresponding to this value, the unknown field is followed. It contains
value 1 from the update at line 4 and also value undefined.
Example 2: At line 12, value 1 is used to write-access the first dimension of an array $arr. The
array has no index corresponding to this value and the index $arr[1] is thus created. Because statically
unknown assignment at line 11 could involve also the index $arr[1], the data from this assignment are
copied to a new index and thus also the index $arr[1][2] (with values 6 and undefined) is defined.
The principal challenge of merging multi-dimensional associative arrays with unknown fields is to
determine the set of indices of the resulting array. That is, the resulting array may contain indices that
are not present in any array being merged. The reason again stems from the fact that unknown fields are
not followed by statically known read accesses.
Example 3: As an example, see the join point at line 13. The array $arr contains all indices that
are defined in either of merged branches plus the index $arr[2][2] (with values 6 and undefined).
D. Hauzar, J. Kofroň, P. Baštecký
s∈Σ
m ∈ Map
i ∈ Index
a ∈ Aliases
amust ∈ Aliasesmust
amay ∈ Aliasesmay
59
= Var × Map × Index × Aliases
= Var → P(Val)
= (Var ×Val) → Var
= Aliasesmust × Aliasesmay
= Var ×Var
= Var ×Var
Table 1: Data-flow analysis state-space.
undefVar ∈ Var
∗ ∈ Val
undefined ∈ Val
• ∈ Val
is a variable representing an undefined variable.
is the statically-unknown value.
is the undefined value.
is the value representing index-name of unknown field.
Table 2: Special variables and values.
The reason of creating new index is that while in the second branch the read access to this index in the
first level follows the unknown field and then reaches the value assigned at line 11, in the merged array
the read access follows the index $arr[2].
3
Formalization
We formalized our data-flow analysis using data-flow equations for the forward data-flow analysis [7].
The formalization includes handling associative arrays of unlimited depth and accesses with arbitrary
expressions to such structures. It does not explicitly include handling of objects. While objects are
treated analogously to arrays in our implementation, there are subtle differences. Therefore, we excluded
handling of objects from our formalization to make it more clear.
3.1
Analysis State Space
Tab. 1 presents elements of the state space of our data-flow analysis. Every state contains a variable,
which represents the symbol-table. Because top-level variables can be accessed dynamically ($$var is
a variable whose name is given by a value of variable $var), we model top-level variables as indices
of the symbol-table variable1 . Function Map maps a variable to a set of its possible values. Function
Index maps a variable and an index name to a variable containing an array which has the first variable
on the index with this name. In the following, we say that variable v is an index of variable p identified
by value ind if ((v, ind), p) ∈ indexOf (i.e., v is p[”ind”]). A pair of relations Aliases relates variables
that are must and may aliases. Tab. 2 presents special variables and values. The value • identifies the
unknown field of a given variable. The unknown field of a variable is used to access statically-unknown
indices of the variable. In Tab. 3, there are several helper functions (projections) defined; we use them in
the subsequent definitions.
3.2
Data-flow equations
For propagating states through the nodes of the control-flow graph (CFG), we use a modification of
standard data-flow equations for the forward problem. Each node k of CFG has six states associated:
INk , GENk0 , INk0 , GENk , KILLk and OUTk . INk represents the data coming to k; it is created by merging
the states going out from all predecessors of k. In the case that the node has more predecessors, we call
this state join point and the operation mergeStates merges information from different states. If the node
has only one predecessor, the operation mergeStates only copies the information. The state GENk0 defines
1 Consequently, the notions of index and variable refer to the same abstraction and we use them interchangeably. That is, an
index of an associative array in an arbitrary depth is a first class name the same way as a variable.
60
Data-flow Analysis of Programs with Associative Arrays
variables that are newly defined by the update and data of these variables (note that these data can come
only from unknown fields). The state KILLk defines data that is removed from variables by the update
while GENk defines data that is added from the right-hand side of the update statement. Finally, the state
OUTk represents updated data, i.e., the data going out from this node. The predicate pred(k) returns the
set of n output states associated with the predecessors of k.
The data-flow equations are:
INk = mergeStates({OUTp }), p ∈ pred(k)
INk0 = INk ∪ GENk0
OUTk = GENk ∪ (INk0 − KILLk )
For the the initial node i, the output state is as follows:
OUTi = (root, {(root, 0),
/ (unk, {undefined})}, {((unk, •), root)}, (0,
/ 0))
/
That is, the state contains variable root representing a symbol table and a variable unk which is its
unknown field—it represents statically-unknown variables.
#
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
Expression
Let x = (root, map, i, (amay , amust )) be a state.
r(x)
= root
values(x, v ∈ Var)
= {vals, (v → vals) ∈ map}
values(x,V ∈ P(Var))
= {values(x, v), v ∈ V }
valuesunde f (x,V ∈ P(Var))
= values(x,V )
if V 6= 0/
= {undefined}
if V = 0/
values(x, undefVar)
= {undefined}
indexOf(x)
=i
indices(x,V ∈ P(Val))
= {iv, ∃v∈V ∃n∈Val ((iv, n), v) ∈ i}
indices(x,V ∈ P(Var), I ∈ P(Val)) = {iv, ∃v∈V ∃ind∈I ((iv, ind), v) ∈ i}
aliasesmust/may (x)
= {(v1 , v2 ), (v1 , v2 ) ∈ amust/may ∨ (v2 , v1 ) ∈ amust/may }
aliasesmust/may (x, v ∈ Val)
= {a, (v, a) ∈ aliasesmust/may (x)}
aliasesmust/may (x,V ∈ P(Val))
= {(a, v), a ∈ aliasesmust/may (x, v), v ∈ V }
aliases(x, v ∈ Val)
= aliasesmay (x, v) ∪ aliasesmust (x, v)
Table 3: List of helper functions and projections.
3.3
Access Paths
We describe expressions for accessing variables and associative arrays of an arbitrary depth using access
paths. An access path consists of a single value or a sequence of access paths:
AP ::= a, a ∈ Val
::= []([AP])∗
Each access path from the sequence represents the expression for accessing the level of an associative
array given by the position of the access path in the sequence. This makes it possible to perform accesses
to any associative array of an arbitrary depth where at each level of the array the set of values used for
indexing is specified with an arbitrary read-access. As we will see later, a read access using an access
path returns a set of values, which can include the undefined and the ∗ values.
D. Hauzar, J. Kofroň, P. Baštecký
61
# Expression
(13) Eval(x, AP)
= {a}
if AP = a, a ∈ Val
= {values(x, v), v ∈ Vars(x, AP)} if AP = [][AP]∗
(14) Vars(x, AP) = Vars(x, r(x), AP)
(15) Vars(x, v ∈ Var, AP) = {undefVar}
if AP = a, a ∈ Val
= {undefVar}
if AP = [][AP1 ][AP2 ]...[APn ] ∧Varsn (x, v, AP) = 0/
= Vars0 (x, v, AP) if AP = []
= Varsn (x, v, AP) if AP = [][AP1 ][AP2 ]...[APn ] ∧Varsn (x, v, AP) 6= 0/
(a)
(b)
(c)
(d)
(16) Vars0 (x, v, AP) = {v}
(17) ∀i∈1,...,n :
Varsi (x, v, AP) = indices(x,Varsi−1 (x, v, AP))
if ∗ ∈ Eval(x, APi )
= indicesr (x,Varsi−1 (x, v, AP), Eval(x, APi ))
if ∗ ∈
/ Eval(x, APi )
(18) indicesr (x,V ∈ P(Var), I ∈ P(Val)) = indices(x,V, I)∪
S
v∈V,ind∈I {u, ((u, •), v) ∈ indexOf(x)∧ 6 ∃w∈Vars ((w, ind), v) ∈ indexOf(x)}
(a)
(b)
(a)
(b)
Table 4: Definition of read accesses.
An access path describes an access from a given index (variable). In the following, [v]([AP])∗ denotes an access path []([AP])∗ from variable v. Access paths express any PHP expression describing
data-access without loss of information. For example, consider the following PHP expressions and the
corresponding access paths in the state s: $a[$b]–[r(s)][a][[r(s)][b]], $$a–[r(s)][[r(s)][a]], $a[$b[$c]][2]–
[r(s)][a][[r(s)][b][[r(s)][c]]][2].
3.4
Read Accesses
Tab. 4 defines the read access. Eval defines the set of values accessible via an access path from a symboltable variable at a given state (13). Vars defines the set of variables accessible via an access path from
the symbol-table variable (14) or from an arbitrary variable (15).
If the access does not identify any variable, the set consists of the undefined variable undefVar (15a)–(15-b). Otherwise the set of variables is obtained by a traversal of the indexOf relation. The traversal
starts from the specified variable (16). At each level, the access can be performed either with a staticallyunknown value (17-a) or with a set of statically-known values (17-b). In the first case, the set of variables
at the next level involves all indices of all variables at the current level. Note that these indices include
unknown fields. If the access is performed with a set of statically-known values, the set of variables at
the next level includes indices of all variables at the current level that are identified by the values (18-a).
Moreover, if the accessed index is not yet defined for a variable at the current level, the unknown field
of the variable is added (18-b). Note that it is not necessary to follow aliases. The reason is that a write
access copies the data to all possible targets, including all possible aliases.
If a set of values at each level of an access path contains exactly one value, the Var function yields a
single variable. We use the notation [var][v1 ]..[vn ] in state x to denote the variable Vars(x, var, [][v1 ]...[vn ]).
If the state and the variable from which the access is performed is clear from the context, we write only
[][v1 ]...[vn ].
Example 4: Assume the read-access at line 5 in Fig 1. The access is performed from the root
variable of the state using the access path [][arr][1]. The variable [][arr] is defined at line 4, while the
index [][arr][1] of this variable is not defined. Thus, the first level consists of variable [][arr], while the
second level consists of its unknown field—[][arr][•].
62
Data-flow Analysis of Programs with Associative Arrays
3.5
Write Accesses
Tab. 6 defines the GEN 0 set, which contains variables that are created by the assignment and alias
statements—the variables statically mentioned for the first time in the left-hand side of the statement2 . It
also defines the variables that are updated by these statements. Tab. 7 defines KILL and GEN sets, which
contains data that are removed and added to these variables. Tab. 5 defines the deep copy of a variable,
which is used when a new variable is created and when an existing variable is assigned to another one.
Collecting Variables
Tab. 6 defines four sets of variables. Must and may (22) are variables that either must or may, respectively, be updated by the assignment statement, must 0 and may0 (23) are variables for the alias statement.
If a variable must be updated, a strong update is performed—new information replaces current information. If a variable only may be updated, a weak update is performed—new information is added to the
information already present at the variable.
Example 5: An example of a weak update is the update at line 4 in Fig. 1—it is not statically known
which index of the variable [][arr] is updated. Weak updates are also performed, e.g., at line 19 in Fig. 1.
While variable [][arr2][3] is strongly-updated, variables [][arr][3] and [][alias] that may be aliases of
[][arr2][3] are only weakly-updated.
# Expression
(19) deepcopyassign (S ∈ Σ ×Var, T ∈ Σ ×Var) :
(xs , vs )
∈ S∧
(xt , vt )
∈ T∧
values(xt , vt )
⊇ values(xs , vs )∧
(a)
∀vsi ∈Var ∀isi ∈Val ((vsi , isi ), vs ) ∈ indexOf(xs ) =⇒ (
(b)
vti = createindex(xt , vt , isi )∧
deepcopy((xs , vsi ), (xt , vti ))∧
(6 ∃vunkn ∈Var ((vunkn , •), vs ) ∈ indexOf(x)∧
v0unkn
= createindex(xt , vt , •)∧
0
values(xt , vunkn ) ⊇ {undefined}))
(20) deepcopy(S ∈ Σ ×Var, T ∈ Σ ×Var) :
deepcopyassign (S, T )∧
(xs , vs )
∈ S∧
(xt , vt )
∈ T∧
aliasesmust (xt , vt ) ⊇ {a, a ∈ aliasesmust (xs , vs )}∧
aliasesmay (xt , vt ) ⊇ {a, a ∈ aliasesmay (xs , vs )}∧
(21) createindex(x, parent ∈ Var, ind ∈ Val) = {var, var = newvar(x)∧
indexOf(x)
⊇ {((var, ind), parent)}∧
values(x, var)
⊇ {undefined}∧
aliasesmust (x, var) ⊇ {var}}
Table 5: Definition of deep copy of the index.
Similarly to read accesses, the variables which are updated by a statement are defined by a traversal
of the indexOf relation starting in the root variable of the state. The traversal uses the access path of the
left-hand side (LHSAP). However, the traversal differs from that of a read access. The first difference is
2 In
PHP, variables are defined also when they are mentioned for the first time in the right-hand side of the alias statement,
i.e., the statement $a = &$b defines the variable $b if it is not defined. While we model this behavior in our implementation,
we ommit it from the formalization to make the presentation of our approach more clear.
D. Hauzar, J. Kofroň, P. Baštecký
#
(22)
(23)
(24)
(25)
(26)
(27)
(28)
63
Expression
Assignemnt / Alias: LHSAP = RHSAP/LHSAP = &RHSAP
LHSAP ∼ [][AP1 ][AP2 ]...[APn ]
∀ j=1,2,...,n I j = Eval(APj )
must = mustn ∧ may = mayn
must 0 = mustn0 ∧ may0 = may0n
indicesw (V ∈ P(Var), I ∈ P(Val))
= indices(IN,V, I)∪
S
0
v∈V,ind∈I {defindex(v, ind), 6 ∃w∈Var ((w, ind), v) 6∈ indexOf(IN ∪ GEN )}
defindex(v ∈ Var, ind ∈ Val)
= {i, ((u, •), v) ∈ indexOf(IN ∪ GEN 0 )∧
i = createindex(GEN 0 , v, ind)∧
deepcopy((IN ∪ GEN 0 , u), (GEN 0 , i))}
must0 = {r(IN)} ∧ may0 = 0/
∀ j∈1,2,...,n
((|I j | = 1 ∧ I j 6= {∗}) ∧ (
(a)
must j = aliasesmust (indicesw (must j−1 , I j ))∧
may j = aliases(INk0 , indicesw (may j−1 , I j )) ∪ aliasesmay (IN, indicesw (must j−1 , I j ))))∨
((|I j | > 1 ∧ ∗ ∈
/ I j) ∧ (
(b)
must j = 0∧
/
may j = aliases(INk0 , indicesw (may j−1 ∪ must j−1 , I j ))))∨
(∗ ∈ I j ∧ (
(c)
must j = 0∧
/
may j = aliases(INk0 , indices(INk0 , may j−1 ∪ must j−1 ))))
((In | = 1 ∧ In 6= {∗}) ∧ (
mustn0 = indiceswj (must j−1 ))∧
may0n = indiceswj (may j−1 )))∨
((|I j | > 1 ∧ ∗ ∈
/ I j) ∧ (
0
mustn = 0∧
/
may0n = indiceswj (may j−1 ∪ must j−1 )))∨
(∗ ∈ In ∧ (
mustn0 = 0∧
/
may0n = indices(INk0 , mayn−1 ∪ mustn−1 )))
Table 6: Definition of collecting variables for an update.
that for a write access also the corresponding aliases are followed. That is, all aliases whose data can be
possibly changed are updated by the write access. That is why it is not necessary to follow the aliases
during read accesses. The second difference is that if a write access to an index identified by a statically
known value is performed and this index does not exist, it is created.
Creating new indices when traversing the indexOf relation is handled by the definition of the indicesw
set (24). Note that write accesses to unknown fields in preceding program points could update also the
newly defined index and its sub-indices. Thus the new index contains a deep copy of the unkwnown field
(19). That is, it contains all values and aliases from the unknown field (20)–(21-a) and also a deep copy
of all indices of the unknown field (19-b).
Example 6: The statement at line 12 in Fig. 1 creates a new variable [][arr][1]. The write access to
the unknown field [][arr][•] at line 11 could update also this new variable and the data from this unknown
64
Data-flow Analysis of Programs with Associative Arrays
field is thus copied to a new variable. Thus the sub-index [][arr][1][2] of the variable [][arr][1] is defined.
The traversal begins with the must0 set initialized with the variable r(INk ), which corresponds to the
symbol table and the may0 set initialized with the empty set (26). The statement (27) describes a single
step of the traversal at the level j. The set I j of values used to access the j-th level of an associative array
can have (27-a) a single statically-known value, (27-b) several statically-known values, or (27-c) it can
contain a statically-unknown value.
Example 7: As an example of case (27-a), see the statement at line 9 in Fig. 1. The must2 set consists
of variable [r(GEN 0 )][arr][2], the may2 set of variable [r(IN)][alias], and the set of values I3 consists of
value 3. Thus, the must3 set contains must aliases of the index [r(GEN 0 )][arr][2][3]. The only must-alias
of this index is the index itself. The may3 set contains all aliases of index [r(GEN 0 )][alias][3], which is
again only the index itself and all may-aliases of index [r(GEN 0 )][arr][2][3], which is the empty set.
In (27-b) and (27-c), the must j set is empty—it is not known which index is accessed. The may j set
consists of all aliases of the indices of variables in must j−1 and may j−1 . In case of (27-c), the variables
do not need to be identified by values of I j and no new variables are created.
Example 8: As an example, see the statement at line 12 in Fig. 1. At the second level, must2 set
consists of the variable [r(GEN 0 )][arr][1] and the may2 set is empty. The must3 set is empty, while the
may3 set consists of variables [r(GEN 0 )][arr][1][2] and [r(GEN 0 )][arr][1][•].
The difference between computing variables that will be updated by the assignment and the alias
statement (28) is caused by the fact that the assignment statement updates the variable and all its aliases
with new values while the alias statement un-aliases the variable from all its original aliases while keeping
their values unaffected. However, the alias statement respects the aliases at previous levels the same way
as the assignment statement. In other words, the expressions for obtaining indices to be updated for the
assignment and the alias statements treat differently only the last level.
Example 9: In the case of the alias statement at line 15 in Fig. 1, both variables [r(GEN 0 )][arr][2][1]
and [r(GEN 0 )][alias][1] will be updated, since ([r(GEN 0 )][arr][2] is a may-alias of [r(GEN 0 )][alias]).
In the case of the alias statement at line 16, only the variable corresponding to [r(IN)][arr][2] will be
updated—the variable r(IN)][arr] has no alias.
Performing Update
Tab. 7 defines how the collected variables are updated by the assignment and alias statements. Expressions (29)–(31) describe the data that is removed from the variables. Both the alias and assignment
statements remove all the values and indices of all updated variables that were present in these variables before the update including the data added in collecting phase (29)–(30). The alias statement also
removes aliasing data (31).
Expressions (32)–(35) describe the data that is added to the variables. In the case of a strong update
(32), both the statements add just the data that results from merging variables obtained by the read
accesses of the right-hand-side access path (RHSAP). In the case of a weak update, the original variable
is merged too, so the original data is preserved (33). Technically, the update is described as first merging
the data to a temporary fresh variable and then copying it from this variable to the variable being updated
(34)–(35). The difference between the assignment and the alias statements is that while the former one
does not copy the alias data at the first level (34), the latter one does (35). Note that for the other levels,
the alias data are copied also in the case of the assignment statement (20-b).
Example 10: As an example, see the update at line 20 in Fig 1. The must set is empty, the may
set consists of variables [][arr][1], [][arr][2], [][arr][•], [][arr2][1], [][arr2][2], [][arr2][3], and [][arr2][•].
Consider the update of variable [][arr][1]. Because the update is weak, new data results from merging
the result of read access of RHSAP, which is [][arr2], with the variable that is updated, which is [][arr][1].
D. Hauzar, J. Kofroň, P. Baštecký
#
(29)
(30)
(31)
(32)
(33)
(34)
(35)
65
Expression
Assignemnt / Alias: LHSAP = RHSAP/LHSAP = &RHSAP
∀v∈must∪may∪must 0 ∪may0 values(KILL, v) = values(INk0 , v)
∀v∈must∪may∪must 0 ∪may0 indexOf(KILL, v)= indexOf(INk0 , v)
∀v∈must 0
aliases(KILL, v) = aliases(INk0 , v)
∀vt ∈must∪must 0 src = {(IN, v), v ∈ Vars(IN, RHSAP)}
∀vt ∈may∪may0 src = {(IN, v), v ∈ Vars(IN, RHSAP)} ∪ {(IN 0 , vt )}
∀vt ∈must∪may deepcopyassign ((GEN 0 , v = fresh(GEN 0 )), (GEN, vt )) ∧ mergeVars((GEN 0 , v), src)
∀vt ∈must 0 ∪may0 deepcopy((GEN 0 , v = fresh(GEN 0 )), (GEN, vt )) ∧ mergeVars((GEN 0 , v), src)
Table 7: Definition of updates for assignment and alias statement and new object expression.
Consequently, after the update, the variable []arr][1] contains indices [][arr][1][1], [][arr][1][2], and
[][arr][1][3]. E.g., index [][arr][1][1] contains the data merged from indices [][arr] [1][1] and [][arr2][1].
Example 11: Now consider the update at line 17. The must set consists of [] [arr2], the may set is
empty. The read-access of the RHSAP results in reading [][arr]. Because the update is strong, [][arr] is
the only variable that is merged and thus it is only deep-copied. Note that in the case of the assignment
statement, the alias data is copied for all the levels except for the first one. Thus, because of the alias
statement at line 16, [][arr2][2] is a must-alias of [][alias3] and [][arr][2] and due to the alias statement
at line 4, e.g., [][arr2][•] is a may-alias of [][alias] and [][arr2][•]. Consequently, the statement at line 18
strongly updates not only [][arr2][2], but also [][arr][2] and [][alias3]. Similarly, the statement at line 19
strongly updates [][arr2][3] and weakly updates [][arr][•] and [][alias]. Thus, the subsequent read access
using access path [][arr][3] would read also value 9.
3.6
Merge
Tab. 8 defines the merge operation. Expression (36) defines the operation mergeStates, which is used
in the first data-flow equation to define the IN state of a node. It merges the root variables of the OUT
states of all predecessors of the node to the root variable in the IN state. Note that if the node has only
one predecessor, the merge actually corresponds to a deep copy.
Expression (37) defines how variables in given states are merged into the resulting state. Note that
this operation is used also when an update is performed (34)–(35). In (37-a), for each variable being
merged and a state in which it is defined, the access paths of all sub-indices of the variable are collected.
The empty access path [], which corresponds to the variables being merged, is added to these access paths
(37-a).
Example 12: For merging at the join point at line 13 in Fig. 1 and for the symbol-table variable
of the first branch, the following access paths are collected: [][alias], [][alias][2], [][alias][3], [][arr],
[][arr][•], [][t], [][t][2], [][arr][1], [][arr][1][2], [][arr][1][3], [][arr][2], [][arr][2][3].
Sub-expression (37-b) further extends the set of access paths. After the extension, it contains an
access path for each variable that will be defined in the resulting state. For each access path that contains
the value • (corresponding to the unknown field) at a certain level it adds the access paths that are
created from this access path by replacing the value • with all the values that are in the input access
paths at this level. Note that this adds new access paths to the resulting set only if there were performed
corresponding statically-known write accesses from different variables being merged. This is analogous
to copying indices of the unknown field when there is a write access with a given value for the first
time and a new variable is thus defined. While write-accesses to unknown fields in preceding program
points could also create sub-indices of a newly defined variable, in the case of merge there could be
write-accesses to unknown fields that could create sub-indices of variables created elsewhere. Both these
66
Data-flow Analysis of Programs with Associative Arrays
# Expression
(36) mergeStates(OUT ∈ P(Σ)) = {IN, mergeVars((IN, r(IN)), {(o, r(o)), o ∈ OUT })}
(37) mergeVars(R ∈ Σ ×Var, M ∈ P(Σ ×Var)) :
S
APs = (x,vr )∈M (accessPaths(x, vr , [])) ∪ {[]}∧
(a)
ResAPs = extend(APs)∧
(b)
∀AP∈ResAPs (mergeAP(R, M, AP))
(c)
(38) mergeAP(R = (xR ∈ Σ, vR ∈ Var), M ∈ P(Σ ×Var), AP) :
resVar
= createVar(R, AP)∧
(a)
S
mergedVars
= (xM ,vM )∈M {(xM ,Var(xM , vM , AP))}∧
(b)
S
values(o, resVar)
= (xm ,vm )∈mergedVars {valuesunde f (xm , vm )}∧
(c)
T
aliasesmust (xR , resVar) = (xm ,vm )∈mergedVars {aliasesmust (xm , vm )}∧
(d)
S
aliasesmay (xR , resVar) = (xm ,vm )∈mergedVars {aliasesmust (xm , vm ) ∪ aliasesmay (xm , vm )}− (e)
aliasesmust (xR , resVar)∧
S
indices(xR , resVar)
= (xm ,vm )∈mergedVars {createVar((xR , resVar), [][n]),
(f)
n ∈ indicesNames(xm , vm )}
S
(39) accessPaths(x ∈ Σ, vR ∈ Var, AP)
= ((i,iname ),vR )∈indexOf(x) (
{AP[iname ]} ∪ accessPaths(x, i, AP[iname ]))
S
(40) extend(APs) = APs ∪ l∈levels(APs)∧(AP∈APs∧level(AP)>l) {newAP(AP, l, values(APs, l))}
(41) newAP([][v1 ]...[vn ], l ∈ Int,V ∈ P(Val)) = {[][u1 ]...[un ], ∀i=1,...,n∧i6=l (ui = vi ) ∧ ul ∈ V } if vl = •
= 0/
if vl 6= •
(42) level([][v1 ]..[vn ])
=n
(43) levels(APs)
= {l, l = level(AP) ∧ AP ∈ APs}
(44) value([][v1 ]..[vn ], l ∈ Int)
= vl
(45) values(APs)
= {v, v = value(AP) ∧ AP ∈ APs}
(46) indicesNames(x, var) = {n, ∃iv∈Var ∃n∈Val ((iv, n), var) ∈ indexOf(x)}
Table 8: Definition of merge.
operations are thus necessary to preserve the invariant that if there could be a write-access to an index
using a statically known value at a given level, all the data that could be possibly written to this index are
stored there.
Example 13: When the merge at line 13 in Fig. 1 is performed, the set of access paths is extended
with [][arr][2][2], which then causes the corresponding variable in the resulting state to be created. The
reason is that while in the else branch [][arr][2] is not defined, there is a write access to the unknown
field at line 11 that could create a sub-index of this variable. The read access using [][arr][2][2] will
follow [][arr][•] in the second level and will finally [][arr][•][2], which contains values 6 and undefined.
In the then branch, [][arr][2] is created and it will be thus added to resulting state of the merge. The read
access follows this variable at the second level. Thus, to access value 6 with access path [][arr][2][2],
there must be a variable [][arr][2][2] which contains this value in the resulting state. This is analogous
to copying data from the unknown field to a variable that is statically stated for the first time when the
update is performed, e.g, a new variable [][arr][1][2] containing values 6 and undefined is created during
the update at line 12.
Sub-expression (37-c) merges variables corresponding to an access path in the merged states to the
variable which corresponds to this access path in the resulting state. First the variable in the resulting
state using the access path and the output variable is created (38-a). Then, the corresponding variables in
merged states are obtained (38-b). Finally, the data of these variables are merged to the resulting variable
(38-c)–(38-f). Note that while in the case of (38-a), the access path is used to create the variable which
D. Hauzar, J. Kofroň, P. Baštecký
67
directly corresponds to the access path, in the case of (38-b) the access path is used to get variables in
merged states by the read access (15). That is, in the case of (38-b), the variables can be accessible by
unknown fields even at levels where the access path contains a static value. Note that both (38-a) and
(38-f) contain the expression createVar, however, resulting variables are created only once—if expression
createVar is used the second time with the same arguments, it returns the existing variable.
Example 14: As an example, see the merge corresponding to the join point at line 13 in Fig. 1. For
the access path [][arr][1][3], variable [][arr][1][3] is be created in the resulting state. The merged variable
in the first branch is [][arr][1][3], however, for the second branch, the data corresponding to this access
path is located in [][arr][1][•]. For the access path [][arr][2][2], a variable is created in the resulting state,
the merged variables are undefVar in the first branch and [][arr][•][2] in the second branch.
3.7
Termination and Soundness
Termination: The values in our model are represented either by constants present in the program or
values ∗, undefined, and •. Thus the number of values is finite. From this it follows that the number
of defined indices in a single dimension of arrays is finite. However, due to the presence of loops and
recursion, the infinite number of dimension may be generated. To ensure the termination, the number of
dimensions must be limited. This can be done either explicitly [5] or implicitly by using, e.g, allocationsite abstraction for creating new dimensions of arrays. Then, the total number of indices is finite and
alias and indexOf relations are finite as well. The transfer functions defined in Tab. 6, Tab. 7, and Tab. 8
are monotonic so the fixpoint computation terminates.
Our approach is implemented as a part of a static analyzer with the support of operators such as +.
Thus, potentially an infinite number of values can be generated in the program due to presence of loops
and recursion. To ensure termination of fixpoint computation in this case, it is necessary to limit the size
of value sets of each variable by a constant—larger value sets would be represented either by value ∗ or
by a finite abstract domain.
Soundness: We use the following soundness argument:
If a value can be written to a given variable (index) by a write-access at the node n1 of CFG,
it is read from this variable by a read access in node n2 of CFG that follows n1 if and only if
there is a path from n1 to n2 in CFG where the variable is not strongly-updated by different
value. Moreover, if there is a path from the initial node to a given node such that a variable
was not strongly-updated, the set of the variable values returned by a read access always
includes the undefined value.
Note that a value can be written to a given index also if the write access is statically unknown at any level.
Also note that there can be an arbitrary number of join points between n1 and n2 . We do not provide
the proof of the argument, however, its validity follows from definitions of read accesses, write accesses,
and merge.
4
Implementation and Evaluation
We implemented our approach in the context of the Weverca [2] static analyzer. Besides associative
arrays, the implementation supports also objects. Weverca makes it possible to perform static taint
analysis [11] as well as other analyses that can be used for finding bugs and evaluation the system
safety. It constructs CFGs on the fly. If the processed program point is a virtual method call, it uses a
read-access to get the method’s receiver object and uses its type to determine the method to be called.
It also uses a read-access to get values of variables that specify the name of a method and the include
target in case of a dynamic method call or dynamic include. Weverca also implements several options of
context-sensitivity for function and method calls.
68
Data-flow Analysis of Programs with Associative Arrays
n CFG nodes
Variables
(mCODE n/CODE n)
1
2
3
4
235 / 117
463 / 231
919 / 459
1831 / 915
Analysis Time (s)
(mCODE n/CODE n)
107
211
419
835
0.4 / 0.3
1.3 / 0.7
4.9 / 2.5
22.8 / 10.3
Table 9: The evaluation results.
The novelty of our approach is that it is sound and precise even if statically-unknown data from the
input are used to access associative arrays-like structures at an arbitrary level. Other approaches, such
as [10] are more limited, e.g., they model only associative arrays of the depth 1. Since it is not an issue
to implement a precise analysis but a precise scalable analysis, a question we want to answer is whether
our approach scales well. In particular, we want to know how the merge function scales with respect to
the number of variables being merged in the presence of dynamic write accesses to multiple levels of
associative arrays.
To evaluate our approach we used the code CODE n that was generated from the code fragment at
lines (2)–(19) in Fig. 1 replicated 2n times with all the variables except for the variable $any with the
prefix unique for each replica. This code contains non-trivial dynamic accesses to multiple levels of
associative arrays; the number of variables in this code grows exponentially with n. However, the number
of variables that are being merged at each join point is constant. To evaluate the complexity of the merge
operation, we use the code mCODE n (Fig 2) which is defined using CODE n. In mCODE n, all variables
defined in the code are involved in the top-level merge operation.
$any = $ GET[’user input’];
if ($any) { CODE n }
else { CODE n }
Figure 2: The code used for the evaluation
Tab. 9 shows the results of analysis of mCODE n and CODE n for different values of n. The table
shows the number of nodes of generated CFG, the number of variables defined in our representation in
the OUT state of the program end point, the running time of the analysis. The comparison of running
times of mCODE n and CODE n shows that the merge operation is efficient even if all the variables all
involved—mCODE n analyzes two copies of CODE n and additionally merges all the variables. Thus the
increase of time consumption is caused mostly by the increase of the number of variables. This is caused
mainly by the fact that the amount of data stored in a state grows with the number of variables defined
in the state and that in the implementation there is no sharing of data between the states. We believe that
by optimizing the implementation, the scalability can be highly improved and the approach can scale up
to thousands of variables and tens of thousands nodes of CFG.
5
Related work
In this section, we discuss tools and techniques related to the area of our interest.
Pixy [4] is an open-source tool for detection of taint-style vulnerabilities in PHP 4. It involves a flowsensitive, interprocedural, and context-sensitive data flow analysis along with literal and alias analysis
to achieve precise results. The main limitations of Pixy include limited support for statically-unknown
updates to associative arrays, ignoring classes and the eval command, and limited support for aliasing
and handling file inclusion, which all represent principle differences from programming languages such
as Java and C. Alias analysis introduced in Pixy incorrectly models aliasing between arrays and array
indices. Web applications use associative arrays and objects extensively, thus we believe that this is an
D. Hauzar, J. Kofroň, P. Baštecký
69
essential limitation. Importantly, Pixy does not perform type inference, which also limits its precision
and soundness.
Stranger [14] is an automata-based string analysis tool for PHP, which is built upon Pixy. It adds a
more precise string manipulation techniques that enable the tool to prove that an application is free from
attack patterns specified as regular expressions.
Jang [3] presents flow-insensitive points-to analysis for JavaScript. Their work is closely related to
our approach, since JavaScript provides dynamic features similar to those in PHP. The same way as our
approach, they model variables, arrays, and objects using associative arrays. However, they precisely
model only assignments to constant indices and they use special unknown field for all other assignments.
Moreover, the same as Sridharan [10], they limit the accesses to associative arrays to depth one. Next,
PHP supports creating aliases between variables which JavaScript does not. Therefore, we must maintain
alias information to perform updates of points-to information correctly.
Phantm [5] is a PHP 5 static analyzer for type mismatch based on data-flow analysis; it aims at
detection of type errors. It combines run-time information from the bootstrapping phase of an application
and static analysis when instrumentation using this information is used. To obtain precise results, Phantm
is flow-sensitive, i.e., it is able to handle situations when a single variable can be of different types.
However, they omit updates of associative arrays and objects with statically-unknown values and aliasing,
which can lead to both missing errors and reporting false positives.
F4F [9] focuses on static taint analysis of web applications that use frameworks. This work uses
a semi-automatically generated specification of framework-related behaviors to reduce the amount of
statically-unknown information, which arises, e.g., from reflective calls.
Andromeda static taint analyzer [12] fights the problem of scalability of taint analysis by computing
data-flow propagations on demand. It uses forward data-analysis to propagate tainted data and ignores
propagation of other data. If tainted data are propagated to the heap, it uses backward analysis to compute
all targets to which the data should be propagated. Andromeda analyzes Java, .NET, and JavaScript
applications. The drawback of the approach is that it propagates only taint information. The control-flow
of the application can depend on other information which are then not available.
Sridharan [10] et. al. present static flow-insensitive points-to analysis for JavaScript. They model
objects in JavaScript using associative arrays that can be accessed by arbitrary expressions. However,
they limit the accesses to depth one. They show that in this setting, the complexity of flow-insensitive
points-to analysis becomes O(N 4 ), where N is the program size, in contrast to the O(N 3 ), which is the
case when the accesses are constant. To enhance the precision and scalability of the analysis, they identify
correlations between dynamic property read and write accesses. If the updated location and stored value
can be accessed by the same first class entity (variable), it is extracted to a function parametrized by
this entity; this function is then analyzed context-sensitively with the context be the variable. Thus, the
correlation between the update and store is preserved.
Schafer [8] et. al. present a dynamic analysis for identifying variables and expressions that always
have the same value at a given program point and finding this value. Such values can be used, e.g, to
make a constrained dynamic language constructs static and thus enhance the scalability and precision of
static analysis.
Wei [13] et. al. reduce the number of statically-unknown information in static analysis by collecting
such information at run-time.
Livshits [6] et. al. propose a method of fully automatic placement of security sanitizers and declassifiers. They place sanitizers statically whenever possible and they try to minimize the amount of run-time
tracking. The input of their analysis is a data-flow graph generated by a static analyzer. The quality of
sanitization placement—the reduction of the amount of run-time tracking—depends on the quality of
70
Data-flow Analysis of Programs with Associative Arrays
this data-flow graph.
6
Conclusion and future work
Dynamic languages such as PHP contain features that pose significant challenges for static analysis. In
our previous position paper [1] we introduced our approach to static analysis of PHP and described particular parts of the Weverca analyzer [2]. In this paper we focused on the data modeling part. We described
dynamic accesses to associative array-like structures, which make it hard to apply value and points-to
analysis, and presented our approach to this challenge. We focused on soundness and precision—we do
not want to overwhelm the user with too many false alarms, which is often the case of related tools.
The evaluation shows that the prototype implementation of our approach scales to hundreds of variables and thousands of nodes in the control-flow graph. We believe that the scalability of the tool can be
further enhanced. Therefore, as future work, we plan to enhance the scalability by implementing optimizations, in particular, sharing the data between nodes of the control-flow graph, and then to evaluate
the scalability on real-life applications.
References
[1] David Hauzar & Jan Kofroň (2012): On Security Analysis of PHP Web Applications. In: STPSA 2012, IEEE,
pp. 577–582, doi:10.1109/COMPSACW.2012.106.
[2] David Hauzar & Jan Kofroň (2014): W EVERCA. http://d3s.mff.cuni.cz/projects/formal_
methods/weverca/.
[3] Dongseok Jang & Kwang-Moo Choe (2009): Points-to analysis for JavaScript. SAC ’09, ACM, New York,
NY, USA, pp. 1930–1937, doi:10.1145/1529282.1529711.
[4] N. Jovanovic, C. Kruegel & E. Kirda (2006): Pixy: a static analysis tool for detecting Web application
vulnerabilities. In: S&P’06, IEEE, doi:10.1109/SP.2006.29.
[5] Etienne Kneuss, Philippe Suter & Viktor Kuncak (2010): Runtime Instrumentation for Precise Flow-Sensitive
Type Analysis. In: RV, pp. 300–314, doi:10.1007/978-3-642-16612-9 23.
[6] Benjamin Livshits & Stephen Chong (2013): Towards Fully Automatic Placement of Security Sanitizers and
Declassifiers. POPL ’13, ACM, New York, NY, USA, pp. 385–398, doi:10.1145/2429069.2429115.
[7] Flemming Nielson, Hanne R. Nielson & Chris Hankin (1999): Principles of Program Analysis. SpringerVerlag New York, Inc., Secaucus, NJ, USA, doi:10.1007/978-3-662-03811-6.
[8] Max Schäfer, Manu Sridharan, Julian Dolby & Frank Tip (2013): Dynamic Determinacy Analysis. PLDI ’13,
ACM, New York, NY, USA, pp. 165–174, doi:10.1145/2499370.2462168.
[9] Manu Sridharan et al. (2011): F4F: Taint Analysis of Framework-based Web Applications. OOPSLA ’11,
ACM, New York, NY, USA, pp. 1053–1068, doi:10.1145/2048066.2048145.
[10] Manu Sridharan et al. (2012): Correlation Tracking for Points-to Analysis of Javascript. ECOOP’12,
Springer-Verlag, Berlin, Heidelberg, pp. 435–458, doi:10.1007/978-3-642-31057-7 20.
[11] Omer Tripp et al. (2009): TAJ: Effective Taint Analysis of Web Applications. PLDI ’09, ACM, New York,
NY, USA, pp. 87–97, doi:10.1145/1542476.1542486.
[12] Omer Tripp et al. (2013): ANDROMEDA: Accurate and Scalable Security Analysis of Web Applications.
FASE’13, Springer-Verlag, Berlin, Heidelberg, pp. 210–225, doi:10.1007/978-3-642-37057-1 15.
[13] Shiyi Wei & Barbara G. Ryder (2013): Practical Blended Taint Analysis for JavaScript. ISSTA 2013, ACM,
New York, NY, USA, pp. 336–346, doi:10.1145/2483760.2483788.
[14] Fang Yu, Muath Alkhalaf & Tevfik Bultan (2010): Stranger: An automata-based string analysis tool for
PHP. TACAS’10, doi:10.1007/978-3-642-12002-2 13.
| 6cs.PL
|
Property Testing in High Dimensional Ising models
arXiv:1709.06688v1 [math.ST] 20 Sep 2017
Matey Neykov∗
Han Liu†
Abstract
This paper explores the information-theoretic limitations of graph property testing in zerofield Ising models. Instead of learning the entire graph structure, sometimes testing a basic
graph property such as connectivity, cycle presence or maximum clique size is a more relevant
and attainable objective. Since property testing is more fundamental than graph recovery, any
necessary conditions for property testing imply corresponding conditions for graph recovery,
while custom property tests can be statistically and/or computationally more efficient than
graph recovery based algorithms. Understanding the statistical complexity of property testing requires the distinction of ferromagnetic (i.e., positive interactions only) and general Ising
models. Using combinatorial constructs such as graph packing and strong monotonicity, we
characterize how target properties affect the corresponding minimax upper and lower bounds
within the realm of ferromagnets. On the other hand, by studying the detection of an antiferromagnetic (i.e., negative interactions only) Curie-Weiss model buried in Rademacher noise, we
show that property testing is strictly more challenging over general Ising models. In terms of
methodological development, we propose two types of correlation based tests: computationally
efficient screening for ferromagnets, and score type tests for general models, including a fast
cycle presence test. Our correlation screening tests match the information-theoretic bounds for
property testing in ferromagnets.
Keywords: Ising models; Minimax hypothesis testing; Dobrushin’s comparison theorem; Antiferromagnetic Curie-Weiss detection; Two-point function bounds; Correlation tests.
1
Introduction
Undirected graphical models are a powerful framework for summarizing and visualizing complicated
multivariate distributions, where conditional independencies among the variables are represented
via a graph. These models have been applied in a variety of fields such as image processing (Besag,
1993), computational biology (Durbin et al., 1998) and social network analysis (Wasserman and
Faust, 1994; Grabowski and Kosiński, 2006). A central topic of interest in graphical models research
is estimating the structure (also known as structure learning) of, or inferring questions about,
the underlying graph based on a sample of observations. Substantial progress has been made
∗
Department of Statistics & Data Science, Carnegie Mellon University, Pittsburgh PA 15213, USA; e-mail:
[email protected]
†
Department of Operations Research and Financial Engineering, Princeton University, Princeton NJ 08542, USA;
e-mail: [email protected]
1
towards understanding structure learning. Popular procedures developed for high dimensional
graph estimation include `1 -regularization methods (Yuan and Lin, 2007; Rothman et al., 2008;
Liu et al., 2009; Ravikumar et al., 2011; Cai et al., 2011), neighborhood selection (Meinshausen
and Bühlmann, 2006; Bresler et al., 2008) and thresholding (Montanari and Pereira, 2009).
Instead of focusing on learning the structure of the entire graph, we study the weaker inferential
problem of property testing, i.e., testing whether the graph structure obeys certain properties based
on a sample of n observations. The questions we are interested in are inspired by Neykov et al.
(2016), but unlike Neykov et al. (2016) who work exclusively with Gaussian models, here we focus
on the zero-field Ising model. The Ising model is a pairwise binary model introduced by statistical
physicists as a model for spin systems with the goal of understanding spontaneous magnetization
and phase transitions (Ising, 1925). More recently the model has found applications in diverse areas
such as image analysis (Geman and Geman, 1984), bioinformatics and social networks (Ahmed and
Xing, 2009). Formally, a zero-field Ising model is a collection of d binary ±1 valued random variables
X = (X1 , X2 , . . . Xd ), hereto referred to as spins, which are distributed according to the law
X
P(X = x) ∝ exp θ
ξuv xu xv ,
(1.1)
1≤u<v≤d
where θ ≥ 0. It is easily seen that the vector X is Markov to the graph G whose edges connect
vertices u and v if and only if ξuv 6= 0. Notice that model (1.1) is overparametrized. However,
when ξuv are viewed as fixed constants, this specification allows one to study the behavior of X
for different values of θ. In the statistical physics literature the parameter θ = T1 where T stands
for temperature, and is often referred to as the inverse temperature of the system. To give some
intuition behind the importance of the inverse temperature it is worth discussing the phenomenon
of phase transitions.
A phase transition occurs when the temperature drops below a critical threshold causing the
distribution of the spins to change from “balanced” to a distribution in which the majority of spins
align similarly. In ferromagnetic Ising models, that is models whose spin-spin interactions ξuv are
non-negative, the main culprit responsible for phase transitions is the graph structure. At low
temperatures neighboring spins are more “rigid” tending to align well with one another, while at
high temperatures they act erratically as the model approaches a collection of independent spins.
This is not unlike how water molecules behave at different temperatures causing water to change
phase. In systems with both positive and negative interactions ξuv , phase transitions are more
challenging to characterize, as the signs of the interactions on the graph also exert influence on
the transition. The complicated behavior of the Ising model at different temperatures suggests
that the difficulty of property testing is related to the temperature. The main focus of this paper
is uncovering necessary and sufficient conditions on the temperature, sample size, dimensionality
and graph properties, allowing one to conduct property tests even when the data is sampled from
the most challenging models. Understanding such limitations is practically useful, since necessary
conditions can provide a benchmark for algorithm comparisons, while mismatches between sufficient
and necessary conditions can prompt to searching for better algorithms.
To elaborate on the type of problems we study, let Vd = {1, . . . , d} be a vertex set of cardinality
d and let G 1 be the set of all graphs over the vertex set Vd . A binary graph property P is a map
P : G 7→ {0, 1}. Given a sample of n observations from a zero-field Ising model with an underlying
1
G depends on d but we omit this dependence to simplify the notation.
2
graph G, the goal of property testing is to test the hypotheses
H0 : P(G) = 0 versus H1 : P(G) = 1.
(1.2)
Below we give three specific instances of property tests, which we study in detail in the later sections
of the paper.
Connectivity. A graph is connected if and only if each pair of its vertices is connected via a path.
Define P as P(G) = 0 if G is disconnected and P(G) = 1 otherwise. Testing for connectivity is
equivalent to testing whether the variables can be partitioned into two independent sets.
Cycle Presence. If a graph is a forest, i.e., a graph containing no cycles, its structure can
be estimated efficiently using a graph selection procedure based on a maximum spanning tree
construction proposed by Chow and Liu (1968). It is therefore sometimes of interest to test whether
the underlying graph is a forest. In this example P satisfies P(G) = 0 if G is a forest and P(G) = 1
otherwise.
Clique Size. A relevant question is to test whether the size of a maximum clique (i.e., a maximum
complete subgraph) contained in the graph is less than or equal to some integer m − 1 versus the
alternative that a maximum clique is of size at least m. In this example set P(G) = 0 if G contains
no m-clique, and P(G) = 1 otherwise.
By definition, property testing is a statistically simpler task compared to learning the entire graph
structure, since if a graph estimate is available, property testing can be done via a deterministic
procedure (although possibly a computationally challenging one). An important implication of
this observation is that any quantification on how hard testing a particular graph property is,
immediately implies a “lower bound” on how hard estimating the entire graph is. Conversely, any
algorithm capable of learning the graph structure with high confidence can be applied to test any
property while preserving the same confidence. Importantly however, there could exist tests geared
towards particular graph properties which can statistically and/or computationally outperform
generic graph learning methods.
Under the assumption that the maximum degree2 of G is at most s, foundational results on
the limitations of structure learning of Ising models were given by Santhanam and Wainwright
(2012). Specifically, Santhanam and Wainwright (2012) gave two conditions on the triplet (n, d, s):
under the first condition it was argued that no algorithm can learn the graph structure with high
confidence; under the second condition a correlation decoding algorithm was shown to estimate the
graph with high probability. In view of the relationship between property tests and structure learning described earlier, our work can be seen as a generalization of necessary conditions for structure
learning. Our results also help to paint a more complete picture of the statistical complexity of
testing in Ising models. Unlike in structure learning, understanding property testing requires the
distinction of ferromagnetic and general Ising models, of which the latter exhibit strictly stronger
limitations. In terms of methodological development, we formalize correlation based property tests
which can be customized to target any graph property. We now outline the three major contributions of this work.
2
The largest number of neighbors of any vertex of G.
3
1.1
Summary of Contributions
Our first contribution is to provide necessary conditions for property testing in ferromagnets. We
give a generic lower bound on the inverse temperature (Theorem 2.4), demonstrating that property
testing is difficult in high temperature regimes. A key role in the proof is played by Dobrushin’s
comparison theorem (Föllmer, 1988), which a is powerful tool for comparing discrepancies between
Gibbs measures based on their local specifications. We further formalize the class of strongly
monotone graph properties, and show that when the temperature drops below a certain property
dependent threshold, testing strongly monotone properties becomes challenging (Theorem 2.6).
Our second contribution is to study necessary conditions for general Ising models, i.e., models including both ferromagnetic and antiferromagnetic 3 interactions. Specifically we argue that
testing strongly monotone properties over general models requires more stringent conditions than
performing the same tests over ferromagnets (Theorem 4.1 and Proposition 4.3). In order to prove
this result we demonstrate that it is very difficult to detect the presence of an antiferromagnetic
Curie-Weiss4 (e.g., see Kochmański et al., 2013) model buried in Rademacher noise, which to the
best of our knowledge is the first attempt to analyze this problem.
Our third contribution is to design several correlation based tests and understand their limitations. First, we formalize and study a generic correlation screening algorithm for ferromagnets.
We show that this algorithm works well at high temperature regimes (Corollary 3.7), and could be
successful even beyond this regime for some properties (Section 3.1). To analyze the algorithms at
low temperature regimes we develop a novel “no-edge” two-point function (i.e., correlation) bound
for graphs of bounded degree, which may be of independent interest. Second, we adapt the correlation decoders of Santhanam and Wainwright (2012) to property testing for general Ising models,
and we develop a computationally tractable cycle test (Section 5.1).
1.2
Related Work
Pseudo-likelihood parameter estimation and inference of the inverse temperature for Ising models
of given structures was studied by Bhattacharya and Mukherjee (2017). Property testing is a
fundamentally different problem, and our work is in part inspired by Neykov et al. (2016). We
show that the graph packing constructions introduced by Neykov et al. (2016) for Gaussian models,
can also be used to give upper bounds on the temperature for property testing in Ising models
(Theorem 2.4). Unlike Neykov et al. (2016) however, we do not restrict our study to graphs of
bounded degree, and we give a more complete picture of the complicated landscape of property
testing in Ising models, by distinguishing ferromagnetic from general models (see Theorems 2.6 and
4.1, and Proposition 4.3).
Structure learning is very relevant to property testing. The first minimal sample size requirements for graph selection in Ising models were established by Santhanam and Wainwright (2012).
In a follow up work, restricted to the class of ferromagnetic models, Shanmugam et al. (2014) related structural conditions of the graph with information-theoretic bounds. Similarly to these works
we exploit graph properties to obtain information-theoretic lower bounds for testing. However, as
we discussed previously, property testing is a more fundamental problem than graph estimation,
and hence our results imply bounds for structure learning and not vice-versa. One of the first
3
Inspired by statistical physics jargon, throughout the paper we use the terms ferromagnetic and antiferromagnetic
to refer to positive and negative interactions respectively.
4
i.e., an antiferromagnetic model with a complete graph.
4
algorithms for structure learning of high dimensional Ising models based on `1 -regularization was
given by Ravikumar et al. (2010), but it was later shown by Montanari and Pereira (2009) to work
well only at high temperature regimes. Santhanam and Wainwright (2012) suggested correlation
decoders, which are computationally inefficient but to the best of our knowledge have the smallest sample size requirements for general models. Anandkumar et al. (2012) gave a polynomial
time neighborhood selection method for models whose graphs satisfy special properties. The first
polynomial time algorithm which works for general Ising models was given by Bresler (2015) and
was motivated by earlier works on structure recovery (Bresler et al., 2008, 2014). Inspired by the
simplicity of the correlation algorithms studied by Montanari and Pereira (2009); Santhanam and
Wainwright (2012) we use similar ideas to develop property tests, and demonstrate that for some
properties our tests work in vastly different regimes compared to graph recovery.
Finally, we remark that property testing is not unrelated to the more basic detection problems
considered by Addario-Berry et al. (2010); Arias-Castro et al. (2012, 2015b,a). In detection, the
null hypothesis is typically a model with an empty graph, while the alternative consists of models
with graphs containing a certain type of subgraph. This is a major difference with property testing,
where both the null and alternative hypothesis may, and usually do contain models with non-empty
graphs. Nevertheless, to show minimax bounds of property testing over the class of general Ising
models we do consider a detection problem (see Sections 4 and 6 for details).
1.3
Notation
For convenience of the reader we summarize
the notation used throughout the paper. For a vector
Pd
q 1/q
T
d
v = (v1 , . . . , vd ) ∈ R , let kvkq = ( i=1 vi ) , 1 ≤ q < ∞ with the usual extension for q = ∞:
kvk∞ = maxi |vi |. Moreover, for a matrix A ∈ Rd×d we denote kAkp = maxkvkp =1 kAvkp for
p ≥ 1. For any n ∈ N we use the shorthand [n] = {1, . . . , n}, and we reserve the special notation
Vd to denote the vertex set Vd = [d]. We denote N0 = N ∪ {0}. For a set N ⊂ N we define
N
2 = {(u, v) | u < v, u, v ∈ N } to be the set of ordered pairs of numbers in N . For a graph
G = (V, G) we use V (G) = V , E(G) = E, maxdeg(G) to refer to the vertex set, edge set and
maximum degree of G respectively. For two graphs G, G0 we use G0 E G if G0 is a spanning
subgraph of G, i.e., V (G0 ) = V (G) and E(G0 ) ⊆ E(G); we use G0 ⊆ G if G0 is a subgraph of G but
not necessarily a spanning one, i.e., V (G0 ) ⊆ V (G) and E(G0 ) ⊆ E(G). For a probability measure
P, the notation P⊗n means the product measure of n independent and identically distributed (i.i.d.)
samples from P.
d
Any vector θ ∈ R(2) naturally induces an undirected graph G(θ) = (Vd , E(θ)), where the edge
set E(θ) = {(u, v) | θuv 6= 0, u < v}. If E is an arbitrary edge set for e := (j, k) ∈ E we use the
notation θe = θjk interchangeably to denote the element e of the vector θ. Given two sequences
{an }, {bn } we write an = O(bn ) if there exists a constant C < ∞ such that an ≤ Cbn ; an = Ω(bn )
if there exists a positive constant c > 0 such that an ≥ cbn ; an = o(bn ) if an /bn → 0, and an bn if
there exists positive constants c and C such that c < an /bn < C; an & bn if there exists an absolute
constant C > 0 so that an ≥ Cbn . Finally we use ∧ and ∨ for min and max of two numbers
respectively.
1.4
Organization
The remainder of the paper is structured as follows. Minimax bounds for ferromagnetic models
are given in Section 2. Section 3 is dedicated to correlation screening algorithms for testing in
5
ferromagnets. Section 4 provides minimax bounds for general models. Section 5 studies correlation
based algorithms for general models. The proof of the main result of Section 4 is given in Section 6.
Discussion is postponed to the final Section 7. All remaining proofs are relegated to the appendices.
2
Bounds for Ferromagnets
This section discusses lower and upper bounds on the temperature for ferromagnetic models. We
begin by formally introducing the simple zero-field ferromagnetic Ising models. Denote the set of
d
all vectors θ ∈ R(2) whose entries are either equal to 0 or θ > 0 by
d
S(θ) = {θ ∈ R(2) | θe = 0 if e 6∈ E(θ), θe = θ if e ∈ E(θ)}.
(2.1)
Given a θ ∈ S(θ), the simple zero-field ferromagnetic Ising model with signal θ > 0 is given by
Pθ (X = x) =
1
exp θ
Zθ
xu xv ,
X
(2.2)
(u,v)∈E(θ)
where the vector of spins x ∈ {±1}d and
X
Zθ =
exp θ
x∈{±1}d
X
xu xv ,
(u,v)∈E(θ)
denotes the normalizing constant, also known as partition function. Model (2.2) is equivalent to
(1.1), where the spin-spin interactions ξuv are either equal to 1 or 0 hence the term “simple”. The
term “zero-field” refers to the fact that we have set all “main-effects” parameters of the spins xu
to zero, and “ferromagnetic” refers to the fact that all spin-spin interactions are non-negative. As
discussed in the introduction, the parameter θ is the inverse temperature but will also be referred
to as signal strength interchangeably.
A key concept allowing us to quantify the difficulty of testing a graph property P under the
worst possible scenario is the minimax risk. Formally, given data generated from model (2.2) and
a property P, testing (1.2) is equivalent to testing H0 : θ ∈ S0 (θ) versus H1 : θ ∈ S1 (θ) where
S1 (θ) := {θ ∈ S(θ) | P(G(θ)) = 1}.5
S0 (θ) := {θ ∈ S(θ) | P(G(θ)) = 0},
The minimax risk of testing P is defined as
⊗n
Rn (S0 (θ), S1 (θ)) := inf sup P⊗n
θ (ψ = 1) + sup Pθ (ψ = 0) ,
ψ
θ∈S0 (θ)
(2.3)
(2.4)
θ∈S1 (θ)
where the infimum is taken over all measurable binary valued test functions ψ, and recall the
notation ⊗n for a product measure of n i.i.d. observations. Criteria (2.4) evaluates the sum of
the worst possible type I and type II errors under the best possible test function ψ. One can
always generate ψ ∼ Ber( 12 ) independently of the data, which yields a minimax risk equal to 1.
In the remainder of this section we derive upper and lower bounds on the temperature beyond
which Rn (S0 (θ), S1 (θ)) asymptotically equals 1, which implies that asymptotically the best test
5
The sets S0 (θ) and S1 (θ) depend on P, but we suppress this dependence to ease the notation.
6
of P would be as good as a random guess. Importantly, here and throughout the manuscript we
implicitly assume the high dimensional regime d := d(n), so that asymptotically d → ∞ as n → ∞.
To formalize our general signal strength bound for combinatorial properties in Ising models,
we need several definitions. Similar definitions were previously used by Neykov et al. (2016) to
understand the limitations of combinatorial inference in Gaussian graphical models. The first
definition allows us to measure a graph based pre-distance between edges.
Definition 2.1 (Edge Geodesic Pre-distance). Let G be a graph and {e, e0 } be a pair of edges
which need not belong to G. The edge geodesic pre-distance is given by
dG (e, e0 ) := min 0 dG (u, v),
u∈e,v∈e
where dG (u, v) denotes the geodesic distance6 between vertices u and v on G. If such a path does
not exist dG (e, e0 ) = ∞.
Here we use the term pre-distance since dG (e, e0 ) does not obey the triangle inequality. Having
defined a pre-distance we can define edge packing sets and packing numbers.
Definition 2.2 (Packing Number). Given a graph G = (V, E) and a collection of edges C with
vertices in V , an r-packing of C is any subset of edges S, i.e., S ⊆ C such that each pair of edges
e, e0 ∈ S satisfy dG (e, e0 ) ≥ r. We define the r-packing number:
N (C, dG , r) = max{|S| | S ⊆ C, S is r-packing}, i.e.,
N (C, dG , r) is the maximum cardinality of an r-packing set.
A large r-packing number implies that the set C has a large collection of edges that are far
away from each other. Hence the packing number can be understood as a complexity measure of
an edge set. The final definition before we state our first result formalizes constructions of graphs
belonging to the null and alternative hypothesis and differing in a single edge.
Definition 2.3 (Null-Alternative Divider). For a binary graph property P, let G0 = (Vd , E0 ) be a
graph satisfying P(G0 ) = 0. We refer to an edge set
C = {e1 , . . . , em },
as a null-alternative divider (or simply divider for short) with a null base G0 if for any e ∈ C the
graphs Ge := (Vd , E0 ∪ {e}) satisfy P(Ge ) = 1.
Intuitively, a large divider set C implies that testing P is difficult since there exist multiple
graphs Ge with which one can confuse the graph G0 . We make this intuition precise in
Theorem 2.4 (Signal Strength Necessary Conditions). Given a binary graph property P, let G0
satisfy P(G0 ) = 0, and the set C be a divider set with a null base G0 . Suppose that |C| → ∞
asymptotically. If for some c < 1 and 0 < ε < 1 we have
r
log N (C, dG0 , ((− log c)−1 + ε) log log |C|)
θ ≤ (1 − ε)
∧ atanh(c(maxdeg(G0 ) + 1)−1 ),
n
then liminf Rn (S0 (θ), S1 (θ)) = 1.
n→∞
6
The geodesic distance between u and v is the number of edges on the shortest path connecting u and v.
7
Theorem 2.4 gives a strategy for obtaining lower bounds on θ using purely combinatorial constructions. Its proof utilizes Dobrushin’s comparison theorem (Föllmer, 1988) to bound the χ2
divergence between Ising measures deferring in a single edge. The second inequality on θ is required to ensure that the system is in a “high temperature regime” which is where Dobrushin’s
theorem holds. In all of our examples one can select a graph G0 of constant maximum degree
and hence the real obstruction on θ is given by the entropy term. Theorem 2.4 is reminiscent of
Theorem 2.1 of Neykov et al. (2016); remarkably, similar constructions can be used to give lower
bounds on the signal strength in both the Gaussian and Ising models. Even though the statements
of the two results are related, their proofs are vastly different. The proof in the Gaussian case
heavily relies on the fact that the partition functions can be evaluated in closed form, which is
generally impossible in Ising models. We demonstrate the usefulness of Theorem 2.4 in Section 2.1
where we apply it to several examples.
We complement Theorem 2.4 by an upper bound on the inverse temperature θ above which the
minimax risk cannot be controlled. The need for such bounds arises due to identifiability issues
in Ising models at low temperatures. In such regimes the model develops long range correlations,
i.e., even spins which are not neighbors on the graph can become highly correlated. A simple
implication of this fact for instance is that it is challenging to tell apart a triangle graph from a
vertex with its two disconnected neighbors at low temperatures (see Figure 3). To formalize the
statement we first define a class of graph properties. To this end recall the distinction between the
spanning subgraph and subgraph inclusions E, ⊆ introduced in Section 1.3.
Definition 2.5 (Monotone and Strongly Monotone Properties). A binary graph property P : G 7→
{0, 1}, is called monotone if for any two graphs G0 E G we have P(G0 ) ≤ P(G). A binary property
P is called strongly monotone if for any two graphs G0 ⊆ G we have P(G0 ) ≤ P(G).
By definition any strongly monotone property is monotone, however the converse is not true. An
example of a strongly monotone property is the size of the largest clique in a graph. On the other
hand, an example of a monotone property which is not strongly monotone is graph connectivity.
We now state our result giving an upper bound on θ when testing strongly monotone properties.
Recall definition (2.3) of the sets S0 (θ), S1 (θ). We have
Theorem 2.6 (Strongly Monotone Properties Necessary Conditions). Let P be a strongly monotone property, and H0 be a graph satisfying P(H0 ) = 1. Assume there exists an l × r biclique7 B
with r ≥ 2 such that B E H0 . Suppose there are two vertices u, v belonging to the right side of B,
3
so that adding (u, v) to H0 gives a graph H1 satisfying P(H1 ) = 1. Let θ satisfy θ ≥ 2l and θ ≥ r−2
when r > 2 or θ ≥ log 2 for r = 2. Then if for some κ > 1 we have
θ≥
2κnr
log logbd/(l+r)c
l
,
(2.5)
it holds that liminf Rn (S0 (θ), S1 (θ)) = 1.
n→∞
Theorem 2.6 shows how to prove upper bounds on θ using graph constructions. One needs
to find a graph H0 containing a large biclique B, so that adding edges to H0 transfers it to an
alternative graph. The number of “left” vertices l of B appears in (2.5), and therefore the larger B
is the harder it is to test P in the worst case. The intuition behind this is as follows. The existence
7
A complete bipartite graph.
8
of the biclique B is a measure of how dense H0 is. The denser H0 is the harder it is to tell it
apart from H1 when θ is large. On the other hand the strong monotonicity of P ensures that if a
subgraph H1 of G satisfies P(H1 ) = 1 then P(G) = 1. Therefore if G contains H0 as a subgraph
it becomes hard to test for P when the value of θ is large. Examples 2.8 and 2.9 of the following
section illustrate how to apply Theorem 2.6 in practice.
2.1
Examples
In this section we apply Theorems 2.4 and 2.6 to establish necessary conditions on θ for the three
examples discussed in the introduction. When we apply Theorem 2.4 we will use ε = 12 and c = e−2
so that (− log c)−1 + ε = 1.
Example 2.7 (Connectivity). In this example we derive a lower bound on θ for graph connectivity
testing. Define “graph connectivity” P as P(G) = 0 if G is disconnected and P(G) = 1 otherwise.
Construct a base graph G0 := (Vd , E0 ) where
bd/2c−1
E0 := {(j, j + 1)j=1
, (bd/2c, 1), (j, j + 1)dj=bd/2c+1 , (bd/2c + 1, d)},
and let
bd/2c
C := {(j, bd/2c + j)j=1 } (see Figure 1).
Adding any edge from C to G0 results in a connected graph, so C is a divider with a null base G0 . To
construct a packing set of C, we collect all edges (j, bd/2c + j) for j ≤ bd/2c − dlog log |C|e satisfying
dlog log |C|e divides j. jThis procedure
results in a packing set with radius at least dlog log |C|e and
k
|C|
cardinality of at least dlog log |C|e − 1. Therefore
log N (C, dG0 , log log |C|) ≥ log
hj
k
i
|C|
− 1 log |C| log d.
dlog log |C|e
By Theorem 2.4 the asymptotic risk of connectivity testing is 1 if for some absolute constant κ > 0
p
θ < κ log d/n ∧ atanh(1/(3e2 )),
(2.6)
where in the last bound we used maxdeg(G0 ) = 2.
e0
e
Figure 1: The graph G0 with two edges e, e0 ∈ C : dG0 (e, e0 ) = 2, d = 10.
Example 2.8 (Cycle Presence). Consider testing the property P “cycle presence”, i.e. P(G) = 0
if G is a forest and P(G) = 1 otherwise. Define the null base graph G0 = (Vd , E0 )
E0 := {(j, j + 1)d−1
j=1 },
9
and let the edge set
C := {(j, j + 2)dj=1 },
where the addition is taken modulo d (see Figure 2 for a visualization). By construction we have
P(G0 ) = 0 and |C| = d. Adding any edge from C to G0 yields a graph with a cycle, and hence the
edge set C is a divider with a null base G0 . Moreover, there exists a log log |C|-packing set of C of
cardinality at least dlog |C|−2
log |C|e+2 . To see this collect the edges (j, j + 2) for j = k(log log |C| + 2) + 1
for k = 0, 1, . . . and j ≤ d − 2. The last observation implies that
log N (C, dG0 , log log |C|) log
|C| − 2
log d.
dlog log |C|e + 2
Hence by Theorem 2.4, using maxdeg(G0 ) = 2, we conclude that the asymptotic risk of cycle testing
is 1 when for some absolute constant κ > 0
p
θ < κ log d/n ∧ atanh(1/(3e2 )).
(2.7)
e e0
Figure 2: The graph G0 with two (dashed) edges e, e0 ∈ C such that dG0 (e, e0 ) = 2, d = 7.
By definition cycle presence is a strongly monotone property. Figure 3 shows an example of a
graph H0 satisfying the conditions of Theorem 2.6. Concretely, H0 is a 1×2 biclique which contains
no cycle and has the property that adding one edge on its right side gives a graph with a cycle.
Applying Theorem 2.6 we obtain an upper bound on θ. If there exists a constant κ > 1 such that
θ ≥ 2 ∨ log
4κn
,
logbd/3c
(2.8)
the minimax risk of cycle testing is asymptotically 1.
Figure 3: The graph H0 in the left panel is a triangle with a missing edge. H0 contains a biclique
with no cycles. On the other hand, if we add the dashed edge we obtain a triangle graph H1 , which
has a cycle. In terms of the notation of Theorem 2.6 we have l = 1 and r = 2.
Example 2.9 (Clique Size). In our final example we consider testing the “maximum clique size”
property P, where P is such that P(G) = 0 if G has no m-clique and P(G) = 1 otherwise.
10
eM := M \ {(i1 , i2 )} denote the m-clique with vertices
For a subset M = {i1 , . . . , im } ⊂ Vd , let C
2
in M excluding the edge (i1 , i2 ). Define the null base graph G0 = (Vd , E0 ) where
bd/mc−1
E0 :=
[
e{jm+1,...,(j+1)m} ).
E(C
j=0
Let the edge set
bd/mc−1
C :=
[
j=0
{(jm + 1, jm + 2)},
be the union of missing edges in the incomplete m-cliques in E0 . We refer to Figure 4 for a
visualization of this construction. By design have P(G0 ) = 0 and |C| = bd/mc. Adding any edge
from C to G0 results in a graph with an m-clique, and hence the edge set C is a divider with a
null base G0 . Moreover, note that taking any two edges e, e0 ∈ C we have that dG0 (e, e0 ) = ∞ (see
Figure 4) and hence C is an ∞-packing. Consequently we have
log N (C, dG0 , log log |C|) = log(|C|) = logbd/mc.
Since maxdeg(G0 ) = m − 1 by Theorem 2.4 we have that if
p
θ ≤ κ logbd/mc/n ∧ atanh(1/(me2 )),
(2.9)
the minimax risk of testing is asymptotically 1.
e0
e
Figure 4: A graph G0 in solid edges, consisting of incomplete 4-cliques. The set C consists of all
dashed edges. Adding any single one of the dashed edges transfers G0 to a graph in the alternative.
Since P is a strongly monotone property, we can apply Theorem 2.6 to upper bound θ. To obtain
more precise bounds, let us further restrict the underlying true graph G so that maxdeg(G) ≤ s,
where s is a known integer such that m ≤ s+1. The following construction of the graph H0 from the
statement of Theorem 2.6 is inspired by Turán’s Theorem (e.g., Bollobás, 2004). We build H0 by
s−1
taking b m−2
c(m − 1) vertices, splitting them in m − 1 equally sized groups and connecting any two
vertices belonging to different groups; see Figure 5 for a visualization of H0 . It is simple to check that
H0 does not contain an m-clique, and adding an edge to H0 , gives a graph containing an m-clique
s−1
s−1
m−1
with maximum degree bounded by s. Furthermore, H0 contains a b m−2
cd m−1
2 e × b m−2 cb 2 c
biclique; to see this split the m − 1 vertex groups into 2 vertex groups: one with all vertices in the
first d m−1
2 e groups, and the other with all remaining vertices.
11
Figure 5: In the left panel we show an example of the graph H0 . The concrete values of s and m
used are s = 7 and m = 4. H0 contains no 4-clique, and its maximum degree is smaller than 7. On
the other hand adding any edge (such as the dashed edge in the figure on the right) to H0 results
in a graph H1 which contains a 4-clique, and whose maximum degree remains bounded by 7. The
graph H0 contains a 6 × 3 biclique, whose right side consists of taking the three rightmost vertices.
We are now in a position to apply Theorem 2.6. To render bound (2.5) in a reader friendly
s−1
s−1
m−1
form, we use that the terms b m−2
cb m−1
2 c, b m−2 cd 2 e and s/2 are equal up to constants for large
enough values of s and m. We have that the minimax risk is asymptotically 1 if for any κ > 1
κns
log logbd/sc
3
θ&
∨
.
s/2 − 2
s/2
(2.10)
Notably, the maximum degree s of the graph appears in (2.10) unlike in the previous two examples.
The bigger the maximum degree is allowed to be, the smaller the signal θ has to be in order for
meaningful clique size tests to exist.
3
Correlation Screening for Ferromagnets
In this section we formulate and study the limitations of a greedy correlation screening algorithm on
monotone property testing problems. We pay special attention to the examples discussed in Section
2.1. Unlike correlation based decoders, such as the ones studied by Santhanam and Wainwright
(2012), this algorithm is designed to directly target the graph property of interest, and also has
polynomial runtime for many instances. Moreover, for different properties, the regimes in which
the algorithm works differ vastly from graph recovery algorithms. For generality we expand model
class (2.2) to include all zero-field ferromagnetic models such that:
X
Pθ (X = x) ∝ exp
θuv xu xv ,
(3.1)
(u,v)∈E(θ)
where x ∈ {±1}d and for (u, v) ∈ E(θ): θuv > 0. In contrast to (2.2), (3.1) allows for the
interactions to have different magnitude.
Akin to Section 3 in Neykov et al. (2016), we define a class of graphs “witnessing” the alternative.
For a monotone property P, and a graph G satisfying P(G) = 1, define the collection of graphs
WP (G) = {G0 E G | P(G0 ) = 1},
where recall that the notation G0 E G means that G0 is a spanning subgraph of G. The set WP (G)
is therefore the collection of all spanning
subgraphs of G satisfying P. We refer to graphs in WP (G)
Vd
as witnesses of P. If Gd = Vd , 2 denotes the complete graph, the set WP (Gd ) is the collection
of all possible witnesses of P. It is therefore clear that by the monotonicity of P for two graphs G
12
and G0 such that P(G) = 0 and G0 ∈ WP (Gd ), the set E(G0 ) \ E(G) 6= ∅. Define the parameter
sets
d
S0 := {θ ∈ R(2) | P(G(θ)) = 0},
d
S1 (θ) := θ ∈ R(2) P(G(θ)) = 1,
max
min
G0 ∈WP (Gd ) (u,v)∈E(G0 )
θuv ≥ θ > 0 ,
where we suppressed the dependence of S0 and S1 (θ) on the property P to simplify the notation.
The set S0 imposes no signal strength restrictions, while S1 (θ) requires the existence of at least
one witness of P, each edge of which corresponds to an interaction with magnitude at least θ. In
Section 2.1 we saw that some property tests, such as cycle testing and clique size testing, necessitate
further restrictions on their parameters (see (2.8) and (2.10)). Let R be an appropriately chosen
for the property P restriction set on θ. For instance, an appropriate set R for cycle testing could
be R = {θ | kθk∞ ≤ Θ} for some Θ ≥ θ > 0.
Before we formulate the correlation screening algorithm, it is useful to first define the extremal
correlation
T := T (θ, R, P) =
min
min
θ∈S1 (θ)∩R (u,v)∈E(θ)
Eθ Xu Xv .
(3.2)
T is the smallest possible correlation between neighboring vertices under any model from the
alternative. Provided that we have sufficiently many samples, and the data is generated under an
alternative model, all empirical correlations between neighboring vertices should be approximately
at least T . To formally define the empirical correlations, let X (1) , X (2) , . . . , X (n) ∼ Pθ be n i.i.d.
b so that for any
samples from the ferromagneticPIsing model (3.1). Define the empirical measure P,
n
(i)
d
−1
b
b To this end,
b
∈ A). Put E for the expectation under P.
Borel set A ⊂ R : P(A) = n
i=1 1(X
for a given δ > 0 define the universal threshold
r
4 log d + log δ −1
τ := τ (n, d, δ) =
,
(3.3)
n
and consider the following correlation screening Meta-Algorithm 1 for monotone property testing
in ferromagnetic Ising models.
Algorithm 1 Correlation Screening Test
Input: {X (i) }i∈[n] , θ, R, P
Set ψ = 0
b u Xv }u,v∈V
Calculate the matrix M := {EX
d
Solve
b = argmax
G
min
0
G0 ∈WP (Gd ) (u,v)∈E(G )
Muv
Set ψ = 1 if mine∈Gb Me > T 8 − τ .
return ψ
8
It is assumed that either T or an upper bound of T are available in closed form.
13
(3.4)
The only potentially computationally intensive task in Algorithm 1 is optimization (3.4), which
aims to find a witness whose smallest empirical correlation is the largest. However, for many
properties solving (3.4) can be done in polynomial time via greedy procedures. We remark that step
(3.4) treats Muv as a surrogate of θuv . Instead, one could opt to substitute Muv with an estimate
of the parameter θuv , which can be obtained via a procedure such as `1 -regularized vertex-wise
logistic regressions (Ravikumar et al., 2010), e.g. Here we prefer to focus on correlation screening
due to its simplicity, while we recognize that the estimate Muv may not be a good proxy of θuv in
models at low temperature regimes, which are known to develop long range correlations. We have
the following result regarding the performance of Algorithm 1.
Theorem 3.1. Suppose that
T − max
max
min
θ∈S0 ∩R G0 ∈WP (Gd ) (u,v)∈E(G0 )
Eθ Xu Xv > 2τ.
(3.5)
Then Algorithm 1 satisfies
sup Pθ (ψ = 1) ≤ δ
and
θ∈S0 ∩R
sup
θ∈S1 (θ)∩R
Pθ (ψ = 0) ≤ δ.
(3.6)
Condition (3.5) ensures that the gap between the minimal correlations in models under the
null and alternative hypothesis is sufficiently large even in worst case situations. Specifically, the
term T selects a parameter θ and an edge in E(G(θ)) corresponding the least possible correlation
between neighboring vertices among all alternative models. The second term on the left hand side
(LHS) of (3.5), selects a parameter θ under the null and a witness G0 , yielding the largest possible
minimal correlation on any of the edges of G0 .
Note that any witness G0 ∈ WP (Gd ) contains at least one edge (u, v) ∈ E(G0 ) corresponding
to a parameter θuv = 0 when θ ∈ S0 , while there exists at least one G0 ∈ WP (G(θ)) such that
θuv ≥ θ for all (u, v) ∈ E(G0 ) if θ ∈ S1 (θ). Hence, if the pairwise correlations are good proxies for
the parameters θ the strategy of Algorithm 1 is viable, since T would be sufficiently large while
the second term of (3.5) would be sufficiently small. Indeed, we show this formally below. Given a
sparsity level s and a real number Θ define the ratio
R(s, Θ) :=
cosh(2sΘ) + 2se−2(s−1)Θ (cosh(2(s − 1)Θ)
.
2se−2(s−1)Θ cosh(2Θ) + 1
The following holds
Corollary 3.2. Assume that the parameter θ ∈ R, where the restriction set R is
d
R = {θ ∈ R(2) | maxdeg(G(θ)) ≤ s, kθk∞ ≤ Θ}.
i. Then if s ≥ 3
9
and
R(s, Θ) − 1
> 2τ,
R(s, Θ) + 1
(3.7)
s tanh2 (Θ)
> 2τ,
1 − (s − 1) tanh(Θ)
(3.8)
T −
ii. and/or if (s − 1) tanh(Θ) < 1, and
T −
9
Similar condition to (3.5) holds for the case s = 2. For details see the proof of Corollary 3.2.
14
Algorithm 1 satisfies (3.6).
Remark 3.3. Corollary 3.2 ii. ensures that Algorithm 1 satisfies (3.6) in the high temperature
regime s tanh(Θ) 1 when the entries of θ are approximately equal. To see this, note that the
FKG inequality (see Theorem 6.2.1 Alon and Spencer (2004)) guarantees
T ≥ tanh(θ),
and therefore if θ Θ, s tanh(Θ) is sufficiently small, and θ ≥ Ω(τ ), (3.8) implies that Algorithm
1 controls the type I and type II errors. This fact, coupled with the results of Section 2 suggests
that correlation screening is optimal up to scalars in the high temperature regime.
Below we study three specific instances of Algorithm 1 to obtain better understanding of its
limitations. Importantly, we observe that the correlation screening test can be constant optimal
beyond the high temperature regime for some properties.
3.1
Examples
We now revisit the three examples of Section 2.1.
Example 3.4 (Connectivity). Here we implement the correlation screening algorithm for connectivity testing (see Algorithm 2), and we take the opportunity to contrast property testing to graph
recovery. We will argue that correlation screening can test graph connectivity even in graphs of
unbounded degree. In contrast, correlation based algorithms fail to learn the structure even in
unconnected graphs when the signal strength θ ≥ Ω( 1s ), where s denotes the maximum degree of
the graph, as argued by Montanari and Pereira (2009).
Algorithm 2 Connectivity Test
Input: {X (i) }i∈[n]
Set ψ = 0
b u Xv }u,v∈V
Calculate the matrix M := {EX
d
Estimate Tb the maximum spanning tree (MST) on M10 // Equivalent to solving (3.4)
Set ψ = 1 if mine∈Tb Me > tanh(θ) − τ
return ψ
It is simple to see that the parameter set definitions reduce to
S0 := {θ | G(θ) is disconnected}, S1 (θ) := θ
max
min
T EG(θ) (u,v)∈E(T )
θuv
tree
≥θ>0 ,
d
and there are no further parameter restrictions, i.e., R ≡ R(2) . We have
Corollary 3.5. Assume that tanh(θ) > 2τ . Then Algorithm 2 satisfies (3.6).
Corollary 3.5 underscores the difference between property testing and structure learning. Montanari and Pereira (2009) and Santhanam and Wainwright (2012), showed that one cannot recover
the graph structure in a ferromagnetic model when the parameter θ exceeds a critical threshold.
We also note that the condition tanh(θ) ≥ 2τ matches the lower bound prediction (2.6) up to
constant terms.
10
Finding an MST can be done efficiently.
15
Example 3.6 (Cycle Presence). Here we revisit cycle testing. The parameter sets reduce to
S0 := {θ | G(θ) is a forest}, S1 (θ) := θ
max
min θuv ≥ θ > 0 .
C ⊆G(θ) (u,v)∈E(C)
cycle
Motivated by (2.8) we take the restriction set as R = {θ | kθk∞ ≤ Θ}. The correlation screening
algorithm for cycle testing is given in Algorithm 3. We have the following corollary of Theorem
3.1.
Algorithm 3 Cycle Test
Input: {X (i) }i∈[n] , θ
Set ψ = 0
b u Xv }u,v∈V
Calculate the matrix M := {EX
d
b emerges11 // i.e., solve (3.4)
Add edges with weights from the matrix M until a cycle C
Set ψ = 1 if mine∈Cb Me > tanh(θ) − τ
return ψ
Corollary 3.7. Assume that tanh(θ) − tanh2 (Θ) > 2τ . Then Algorithm 3 satisfies (3.6).
Remark 3.8. For the purpose of this remark, let us restrict to model class (2.2), i.e., assume θ ≡ Θ.
In this setting, the condition of Corollary 3.7 reduces to the quadratic inequality t − t2 > 2τ , where
we put t := tanh(θ) for brevity. Equivalently, the feasible values of θ should satisfy
√
√
1 − 1 − 8τ
1 + 1 − 8τ
≤t≤
.
2
2
To make the calculation more accessible we will now use the notation
f (x) ≈ g(x)
in the sense that
√
√
f (x)
1+ 1−8τ
1− 1−8τ
≈ 2τ and
≈ 1 − 2τ . It
limx↓0 g(x) = 1. When τ = o(1), it is simple to check that
2
2
therefore follows that Algorithm 3 is successful when
2τ ≈ atanh(2τ ) ≤ θ ≤ atanh(1 − 2τ ) ≈ log(1/τ )/2.
Up to scalars the above inequalities agree with bounds (2.7) and (2.8) given in Section 2.1. An
alternative correlation based cycle test which works for general models is given in Section 5.1.
Example 3.9 (Clique Size). Finally we revisit clique size testing. The parameter sets reduce to
S0 (Θ) := {θ | G(θ) has no m-clique}, S1 (θ, Θ) := θ
max
min θuv ≥ θ > 0 ,
C
m-clique
⊆G(θ) (u,v)∈E(C)
and let the restriction set R = {θ | kθk∞ ≤ Θ, maxdeg(G(θ)) ≤ s}, where 2 ≤ m ≤ s + 1. We
summarize the correlation screening implementation for clique size testing in Algorithm 4. To this
end for a Z ∼ N (0, 1) define
√
e2θ E coshm−2 ( θZ + 2θ)
√
r(m, θ) :=
.
E coshm−2 ( θZ)
16
Algorithm 4 Clique Size Test
Input: {X (i) }i∈[n] , θ
Set ψ = 0
b u Xv }u,v∈V
Calculate the matrix M := {EX
d
b emerges12 // i.e., solve (3.4)
Add edges with weights from the matrix M until an m-clique C
r(m,θ)−1
Set ψ = 1 if mine∈Cb Me > r(m,θ)+1
−τ
return ψ
The following holds
Corollary 3.10. For θ ∈ S1 (θ) ∩ R we have T = r(m,θ)−1
r(m,θ)+1 . Hence if (s, m, d, n, θ, Θ) are such that
either of (3.7) or (3.8) hold then Algorithm 4 satisfies (3.6).
Remark 3.11. We now comment on the implications of Corollary 3.10 over model class (2.2), i.e.,
we assume θ ≡ Θ. In Remark 3.3 we already argued that if s tanh(θ) < ν for some sufficiently small
ν > 0, Algorithm 4 controls the type I and type II errors as long as θ ≥ Ω(τ ). Since tanh(θ) ≈ θ
for small θ, in the high temperature regime θ < ν/s Algorithm 4 matches the lower bound (2.9).
Unfortunately, condition (3.7) is not strong enough to ensure that Algorithm 4 works in the
regime θ ≥ Ω(1/s) for all values of m. However, (3.7) suffices to show that it works when m = s+1.
In Remark C.4 of Appendix C, we formally argue that when m = s + 1, θ = Θ, esθ s, provided
that
θ≤
log (2/τ )
,
4(s − 1)
(3.9)
Algorithm 4 satisfies (3.6). Note that (3.9) matches bound (2.10) up to scalars. Hence the special
case of clique size testing when m = s + 1 is yet another example confirming that correlation
screening can be useful for property testing even at low temperatures.
4
Bounds for General Models
So far we studied model classes admitting only ferromagnetic interactions. The situation drastically
changes if one considers the general class of zero-field Ising models, which includes models with
antiferromagnetic, i.e., negative interactions. The main result of this section is an impossibility
theorem, which shows that testing strongly monotone properties over this general class of models
requires boundedness of a certain maximum functional of the property. We also argue more specifically, that unlike in the ferromagnetic case, connectivity testing is not feasible at low temperatures
over the general model class unless the degree of the graph is bounded. Both of these results sharply
contrast what we have seen in the previous sections of the paper.
11
b takes at most d steps, and can be done efficiently.
Finding the cycle C
Fast algorithms for m-clique presence exist when either m or s − m are bounded. For example, when a new edge
ds
(u, v) is added, walk over common neighbors of both
u and v, and check for an m-clique. There are at most
2 steps
s−1
2 s−1
and at each step we have to check at most m−2 m-cliques, giving a runtime bound of O(ds s + m m−2 ) .
12
17
Concretely, we will work with the simple zero-field models specified by the parameters θ ∈ S(θ)
d
(recall the definition of S(θ) (2.1)) and ξ ∈ {±1}(2) as
X
ξuv xu xv .
(4.1)
Pθ,ξ (X = x) ∝ exp θ
(u,v)∈E(θ)
Expression (4.1) has more degrees of freedom compared to (2.2) since the spin-spin interactions
in (4.1) are allowed to be negative. Intuitively, interactions corresponding to ξuv = −1 have a
“repelling” effect on the corresponding spins u and v, whereas interactions with ξuv = 1 have an
“attracting” effect.
Given a monotone property P, recall definition (2.3) of the parameter sets S0 (θ), S1 (θ). Below
we redefine the minimax risk to reflect the model class expansion. Let
⊗n
⊗n
sup
Rn (S0 (θ), S1 (θ)) := inf
sup Pθ,ξ (ψ = 1) + sup
sup Pθ,ξ (ψ = 0) ,
(4.2)
ψ
d θ∈S0 (θ)
d θ∈S1 (θ)
(
)
(
)
2
2
ξ∈{±1}
ξ∈{±1}
where P⊗n
θ,ξ denotes the product measure of n i.i.d. observations of (4.1). Armed with this new
definition we have
d
(2)
Theorem
√ 4.1. Assume θ belongs to the restriction set R = {θ ∈ R | maxdeg(G(θ)) ≤ s}, where
s = o( d). Suppose that the strongly monotone property P satisfies P(∅) = 0 and P(Cs ) = 1,
where Cs denotes an s-clique graph. Then if for some small ε > 0,
s log d/s2
s log d/(2s)
√
> 2 + ε and
≥ 1 + ε,
n
n log 2s
(4.3)
we have liminf inf Rn (S0 (θ) ∩ R, S1 (θ) ∩ R) = 1.
n→∞ θ≥0
Remark 4.2. At first glance it might appear that the statement of the above result is vacuous,
d
= o(1) is often assumed. This assumption
since in the high dimensional literature the scaling s log
n
q
d
is related to the fact that the quantity s log
corresponds to optimal rates of estimation for the `2
n
norm (e.g., Lemma 3 of Ravikumar et al., 2010). We would like to emphasize that the conclusion of
Theorem 4.1 has implications regarding testing and detection problems, and therefore while related
to, should not be confused with minimax optimal estimation rates.
We would further like to contrast our result with similar known bounds such Theorem 1 of
Santhanam and Wainwright (2012) and Theorem 1 of Bresler et al. (2008). The key differences
between our result and these known bounds are that, first (4.3) is valid for property testing and
even more generally for a certain detection problem (see the proof in Section 6 for more details),
while previous results are valid only for structure recovery; and therefore second — the worst
cases are very different. In fact both previously known bounds remain valid in the smaller class of
ferromagnetic models, while as we saw in Section 3, some strongly monotone property tests such
as cycle presence do not exhibit such limitations.
Note that for any non-constant strongly monotone P one has P(∅) = 0. Further, the requirement that P holds true on Cs is mild, since for any non-zero strongly monotone P one can always
find a sufficiently large s for which P is satisfied. The only true √
restriction of Theorem 4.1 on P is
thus that one has to be able to find s in the sparse regime s d.
18
−θ
θ
Figure 6: Testing connectivity on a model with a connected graph as above is difficult. All solid
edges correspond to positive interactions with magnitude θ, while the dashed edge corresponds to
a negative interaction with magnitude −θ. The cliques are chosen of maximum size s so that the
total degree remains at most than s. When the value of θ is large, the majority of the spins in each
clique will tend to have the same sign. Hence the two interactions with the node in the path graph
are “likely” to cancel out, which will make it hard to tell this graph from the disconnected graph
consisting of the connected cliques and the path graph. We exploit this construction in the proof
of Proposition 4.3.
Loosely speaking Theorem 4.1 shows that when the quantity s lognd/s (up to a log factor) is
large, strongly monotone
property testing over the model class (4.1) is very difficult in the sparse
√
regime when s d. What is more, this statement remains valid regardless how large or small
the signal strength parameter θ ≥ 0 is allowed to be. This contrasts sharply with our results in the
ferromagnetic case, where we have already seen an example which did not require such a condition.
Take the cycle testing example in Section 3.6. In this example if s denotes the maximum degree of
the graph, we can always take an s-clique Cs (which certainly contains a cycle), and hence s has to
satisfy (4.3) in order for tests with reasonable minimax risk (4.2) to exist. In contrast Corollary 3.7
shows that controlling the minimax risk is possible without requirements on the maximum degree
of the graph. Theorem 4.1 shows that this is no longer the case over the broader model class (4.1).
The proof of Theorem 4.1 is involved, and in fact shows that a more basic problem than property
testing is impossible when (4.3) holds. Namely, the proof argues that testing the null hypothesis
of a model with an empty graph versus the alternative of a model containing an antiferromagnetic
s-clique is very difficult. Clearly, one would be able to tell those two scenarios apart if there exists
a test of whether a property P satisfying the requirements of Theorem 4.1 holds. The details of
the proof can be found in Section 6.
Theorem 4.1 sheds some light on the complexity involved in testing within model class (4.1).
However, it also leaves something to be desired, namely it does not quantify the effect θ has, and it
does not address specific properties which may potentially exhibit different complexity. Moreover,
it only applies to strongly monotone properties and not to all monotone properties, and thus in
particular it does not apply to connectivity testing. Below we give an explicit upper bound on the
parameter θ for connectivity testing within the model class (4.1). We show a particularly hard case
for connectivity testing in Figure 6 and include a brief explanation in its caption.
Proposition 4.3. Let P be graph connectivity. Assume θ belongs to the restriction set R = {θ ∈
d
3
R(2) | maxdeg(G(θ)) ≤ s}. Let s, n, d be sufficiently large, and suppose θ ≥ 2bs/4c−2
and there
19
exists a κ > 1 so that
θ>
κsn
2 log log(ds)
s − 16
.
(4.4)
Then the minimax risk (4.2) satisfies liminf Rn (S0 (θ) ∩ R, S1 (θ) ∩ R) = 1, where S0 (θ), S1 (θ) are
n→∞
defined in (2.3).
Importantly, condition (4.4) implies that the maximum degree cannot be too large with respect
to the other parameters if we hope for a connectivity test with a good control over both type I and
type II errors to exist. Recall that no such conditions were needed in Corollary 3.5 when testing
connectivity in ferromagnets.
5
Correlation Testing for General Models
Section 4 made it apparent that property testing is more challenging in the enlarged model space
(4.1). In this section we work with an even larger class of zero-field models compared to (4.1) which
are specified as:
X
Pθ (X = x) ∝ exp
θuv xu xv ,
(5.1)
(u,v)∈E(θ)
where θ ≤ |θuv | ≤ Θ, (u, v) ∈ E(θ) are unknown parameters and x ∈ {±1}d . For a monotone
property P define the parameter sets
d
S0 (θ, Θ) := {θ ∈ R(2) | θ ≤ |θuv | ≤ Θ, P(G(θ)) = 0},
d
S1 (θ, Θ) := {θ ∈ R(2) | θ ≤ |θuv | ≤ Θ, P(G(θ)) = 1}.
Different from Section 3, here we impose signal strength restrictions even in the null set S0 (θ, Θ)
and leave the more general setting for future work. We note that one can no longer rely on the
screening algorithms of Section 3 to perform a property test for data generated by (5.1). The success
of correlation screening hinges on the fact that in ferromagnets deleting edges reduces correlations,
which no longer holds in the model class (5.1). One alternative would be to perform exact structure
recovery, and check whether the graph property in question holds on the desired graph. Possibilities
of exact graph recovery include methods developed in (Bresler et al., 2008; Ravikumar et al., 2011;
Santhanam and Wainwright, 2012; Anandkumar et al., 2012; Bresler, 2015).
Below we take a different route, and modify the correlation decoders of Santhanam and Wainwright (2012) by specializing them to property testing. Specifically we consider a score test type
of approach, which only involves model fitting assuming the null hypothesis holds. Suppose there
exists an algorithm A mapping the data input as
e
A({X (i) }i∈[n] , θ, Θ, P) 7→ θ,
(5.2)
so that the output θe ∈ S0 (θ, Θ) and in addition if the true underlying graph G(θ) satisfies
P(G(θ)) = 0 then
b u Xv − E eXu Xv | ≤ ε(δ), 13
max |EX
θ
u,v∈Vd
20
(5.3)
holds with probability at least 1 − δ. Define the test
e := 1(max |EX
b u Xv − E eXu Xv | ≥ ρ).
ψρ ({X (i) }i∈[n] , θ)
θ
u,v
(5.4)
Recall the definition of the threshold τ (3.3) and let
T := T (θ, Θ, s) =
sinh2 (θ/4)
.
2sΘ(3 exp(2sΘ) + 1)
(5.5)
The following holds
Proposition 5.1. Suppose that A is an algorithm satisfying (5.3), and (s, n, d, θ, Θ) are such that
T ≥ τ + ε(δ),
(5.6)
for a small δ > 0. Then the test ψε(δ) given in (5.4) satisfies
sup
θ∈S0 (θ,Θ)
P(ψε(δ) = 1) ≤ δ
and
sup
θ∈S1 (θ,Θ)
P(ψε(δ) = 0) ≤ δ.
(5.7)
A key component for the existence of a successful test (5.4) is the algorithm A satisfying condition (5.3). One example how to construct such an algorithm, is to solve the following optimization
problem
b u Xv − Eθ Xu Xv |.
A({X (i) }i∈[n] , θ, Θ, P) = argmin max |EX
θ∈S0 (θ,Θ) u,v∈Vd
(5.8)
In the following lemma we show that (5.3) indeed holds.
Lemma 5.2. If the algorithm A is defined by (5.8), then (5.3) holds with ε(δ) = τ.
By combining the statements of Proposition 5.1 and Lemma 5.2, we arrive at an abstract generic
property test summarized in Algorithm 5.
Algorithm 5 Generic Property Test
Input: {X (i) }i∈[n] , θ, Θ, P
b u Xv }u,v∈V
Calculate the matrix {EX
d
b u Xv − Eθ Xu Xv |
Solve θe = argminθ∈S0 (θ,Θ) maxu,v∈Vd |EX
e
Output ψτ ({X (i) }i∈[n] , θ)
A sufficient condition for Algorithm 5 to satisfy (5.7) is T ≥ 2τ , where T is defined in (5.5).
One potential problem with Algorithm 5 is that solving (5.8) in general requires combinatorial
optimization, which will likely result in non-polynomial runtime complexity for most properties.
However, unlike the structure learning procedure of Santhanam and Wainwright (2012) which
also requires combinatorial optimization, Algorithm 5 has the advantage that it does not need to
optimize over the entire set of graphs G but only over the smaller set {G | P(G) = 0}. We conclude
this section by proposing a custom variant of this algorithm specialized to cycle testing which uses
a different algorithm A and can be ran in polynomial time.
13
b is the empirical expectation defined in Section 3, while E e is the expectation with respect to the measure
Here E
θ
Pθe from (5.1).
21
5.1
A Computationally Efficient Cycle Test
In this section we propose an efficient algorithm A satisfying (5.3) for cycle testing. Having computationally efficient algorithms for cycle testing is beneficial in practice, since if we have enough
evidence that graph is a forest, we can recover its structure efficiently (Chow and Liu, 1968). We
summarize the algorithm A called “cycle test map” in Algorithm 6.
Algorithm 6 Cycle Test Map
Input: {X (i) }i∈[n] , θ, Θ
b u Xv }u,v∈V
Calculate the matrix M := {EX
d
e
Find a MST T with edge weights |Muv |
for 1 ≤ u < v ≤ d do
if (u, v) 6∈ E(Te) or |Muv | < tanh(θ) − τ then
θeuv ← 0; E(Te) ← E(Te) \ {(u, v)}
else
θeuv ← sign(Muv )((atanh(|Muv |) ∧ Θ) ∨ θ)
end if
end for
e
return (Te, θ)
e of Algorithm 6, evaluating the expectations E eXu Xv needed in (5.4)
Given the output (Te, θ)
θ
can be done in polynomial time via the simple formula
Y
Y
EθeXu Xv =
EθeXk X` =
tanh(θek` ), 14
e
T
(k,`)∈Pu→v
e
T
(k,`)∈Pu→v
Te
e Next, we show the
where Pu→v
denotes the path between vertices u and v in the forest Te = G(θ).
validity of the test in (5.4).
Proposition 5.3. Suppose that tanh(θ)(1 − tanh(Θ)) > 2τ . Then the output of Algorithm 6
satisfies (5.3) with
ε(δ) = τ
2 − tanh(Θ)
.
1 − tanh(Θ)
(5.9)
tanh(Θ)
By combining Propositions 5.1 and 5.3 we immediately conclude that if T ≥ τ 3−2
1−tanh(Θ) , and
e of Algorithm 6 with the test ψε(δ)
the constraints of Proposition 5.3 hold, using the output (Te, θ)
of (5.4) with ε(δ) as in (5.9), satisfy (5.7).
6
Proof of Theorem 4.1
This section is dedicated to the proof of the main result of Section 4. To ease the presentation we
define the following notation. For a set S ⊂ R, a set V ⊂ N by x ∈ S V we understand x = (xk )k∈V
14
The validity of this formula follows by Proposition A.3 and Lemma A.4 which can be found in the supplement.
22
and xk ∈ S. For a vector x ∈ RW , and a set V where V ⊆ W ⊂ N let
X
X
SxV :=
xk , and SSxV =
xk x` ,
k∈V
(k,`)∈(V2 )
(6.1)
where we remind the reader that V2 = {(k, `) | k < `, k, ` ∈ V }. To relieve the sub-indeces, we
further introduce a slight abuse of notation. For a fixed θ > 0 and a multigraph G (i.e., a graph
allowed to have more than one edge joining two vertices) we denote with PG the Ising measure
X
xu xv ,
PG (X = x) ∝ exp − θ
(u,v)∈E(G)
for θ ≥ 0. In this way, when G is a simple graph, this notation is consistent with Pθ,−1 where
θ ∈ S(θ), G = G(θ) and −1 is the vector whose entries are all equal to −1. For the convenience of
the reader below is a definition of χ2 -divergence which we use in the proof.
Definition 6.1 (χ2 -divergence). For two measures P and Q satisfying P Q the χ2 -divergence is
defined by
2
dP
Dχ2 (P, Q) = EQ
−1 .
dQ
We proceed with the
Proof of Theorem 4.1. The proof is concerned with a more fundamental problem than testing the
strongly monotone property P. Namely we consider the problem of detecting an antiferromagnetic
s-clique. To elaborate in detail, for a set V with cardinality |V | = s, let Cs (V ) denote the s-clique
with vertices in the set V . For a vertex set V ⊂ Vd , put GV := (Vd , E(Cs (V ))). The pdf of the
Ising model PGV is given by
PGV (X = x) ∝ exp − θSSxV ,
where x ∈ Rd and θ ≥ 0. Let P∅ denote a d-dimensional Rademacher vector corresponding to the
null hypothesis, i.e., the hypothesis with an empty graph, or equivalently the null hypothesis when
θ = 0. Given a sample of n observations from the measure P, the antiferromagnetic clique detection
problem aims to test
H0 : P = P∅ vs H1 : P = PGV for some V ⊂ Vd , |V | = s.
(6.2)
It is clear that any test of P, can test (6.2) with the same type I and type II error control. Let
π denote the distribution uniformly sampling the vertex set V ⊂ Vd of an s-clique. Define the
mixture measure
P
⊗n
(X1 = x1 , . . . , Xn = xn ) := EV ∼π P⊗n
GV (X1 = x1 , . . . , Xn = xn ).
Using Le Cam’s Lemma (Yu, 1997), we have
Rn (S0 (θ) ∩ R, S1 (θ) ∩ R) ≥ 1 −
Therefore it suffices to control the divergence Dχ2 (P
23
⊗n
1
2
q
Dχ2 (P
⊗n
, P⊗n
∅ ).
, P⊗n
∅ ) (recall Definition 6.1).
Figure 7: The multigraph Cs (V ) ⊕ Cs (V 0 ) is the graph whose adjacency matrix is the sum of
the adjacency matrices of Cs (V ) and Cs (V 0 ), i.e., ACs (V )⊕Cs (V 0 ) = ACs (V ) + ACs (V 0 ) . Above is a
depiction of Cs (V ) ⊕ Cs (V 0 ), where s = 5, d = 10 and |V ∩ V 0 | = 3; Cs (V ) is plotted with solid
edges, while Cs (V 0 ) is plotted with dashed edges.
⊗n
To write the expression Dχ2 (P , P⊗n
∅ ) in a convenient form, we need to introduce several
quantities. To this end take the s-clique graph Cs (V ) and for a θ > 0 define the partition function
X
ZCs (V ) (θ) :=
exp − θSSxV .
x∈{±1}V
Taking two s-clique graphs Cs (V ) and Cs (V 0 ) define the partition function based on the multigraph
Cs (V ) ⊕ Cs (V 0 ) (see Figure 7):
X
0
ZCs (V )⊕Cs (V 0 ) (θ) :=
exp − θ SSxV + SSxV .
x∈{±1}V ∪V 0
Define the standardized version of the partition function above by
T (θ, V ∩ V 0 ) :=
Note that when V ∩ V 0 = ∅, T (θ, ∅) =
write
Dχ2 (P
⊗n
ZCs (V )⊕Cs (V 0 ) (θ)
22s−|V ∩V 0 |
ZCs (V ) (θ)ZCs (V 0 ) (θ)
.
22s
, P⊗n
∅ ) = EV,V 0 ∼π
.
A bit of algebra shows that one can
T n (θ, V ∩ V 0 )
− 1,
T n (θ, ∅)
(6.3)
where V, V 0 ∼ π indicates a two sample i.i.d. draw from π. We now state the key result enabling
us to prove Theorem 4.1, and defer its proof to after we complete this proof.
Theorem 6.2. For any pair of vertex sets V, V 0 of cardinality s, the ratio
ZCs (V )⊕Cs (V 0 ) (θ)
,
ZCs (V ) (θ)ZCs (V 0 ) (θ)
is non-decreasing in θ for θ ≥ 0.
Although the statement of Theorem 6.2 is elementary, its proof is involved. The main difficulty
stems from the fact that we are considering an antiferromagnetic model, which is known to exhibit
24
much more complicated phase transitions compared with the ferromagnetic case and also does not
in general obey correlation inequalities.
Fix two vertex sets V, V 0 of cardinality s, and let I = V ∩ V 0 for brevity. By virtue of Theorem
(θ,I)
6.2, we have that TT (θ,∅)
is increasing in θ. Hence to upper bound this ratio it suffices to understand
V \I
its behavior in the limit θ → ∞. Define U := Sx
T (θ, I) = esθ
|I|
X
I=−|I|, 2|[I+|I|]
|I|
[I+|I|]/2
2|I|
and I := SxI , and rewrite T (θ, I) as
s−|I|
X
U =−(s−|I|), 2|[U +(s−|I|)]
s−|I|
2
[U +s−|I|]/2 − θ (U +I)2
e 2
.
2s−|I|
After a bit of algebra, depending on the parity of s we have the following identity
|I| P
2|I|
|I| s−|I| 2
2
, if s is even,
2
s
j=0,2|j
j/2
(s−j)/2
T (θ, I)
T (θ, I)
(s/2)
sup
= lim
=
P2|I|
|I|
|I|
s−|I|
s−|I| 2
θ→∞ T (θ, ∅)
θ≥0 T (θ, ∅)
2 s 2 j=0,2|j j/2 (s−1−j)/2 + (s+1−j)/2 , if s is odd.
4((s−1)/2)
We now distinguish two cases.
Case I. First consider the case when s is even. We have
2|I|
2|I|
X
X
|I|
s − |I| 2
|I|
s − |I|
s − |I|
s
s − |I|
≤
max
=
,
j
j/2 (s − j)/2
j/2 (s − j)/2
(s − j)/2
s/2 s/2 − b|I|/2c
j=0,2|j
j=0,2|j
where in the last equality we used Vandermonde’s identity. We conclude that
2|I|
T (θ, I)
≤
sup
θ≥0 T (θ, ∅)
s−|I|
s/2−b|I|/2c
s
s/2
=: F (|I|).
(6.4)
(|I|)
A bit of algebra shows that the ratio FF(|I|+2)
≤ 1 for all values of |I|. Hence the maximum is
achieved at |I| ∈ {s, s − 1} (and a direct check shows that it is achieved at |I| = s). Therefore
sup
θ≥0
√
2s
T (θ, I)
≤ s ≤ 2s,
T (θ, ∅)
s/2
where the last inequality follows by Lemma 17.5.1 of Cover and Thomas (2012). Moreover, by
expanding the binomials on the right hand side (RHS) of (6.4) and some algebra we have
T (θ, I)
≤
sup
T
θ≥0 (θ, ∅)
1
1+
s − 2b|I|/2c + 1
b|I|/2c
.
Case II. Now consider the case when s is odd. Using the same ideas as in case I. shows that the
following bound holds
T (θ, I)
sup
≤
θ≥0 T (θ, ∅)
1
1+
s − 2d|I|/2e + 2
25
d|I|/2e
∧
√
2s.
Compiling all inequalities yields
sup
θ≥0
nd|I|/2e
√
√
nd|I|/2e
1
∧ ( 2s)n
1+
∧ ( 2s)n ≤ exp
s − 2b|I|/2c + 1
s − 2b|I|/2c + 1
√
2n|I|
+ 1(|I| ≥ s/2)( 2s)n .
≤ exp
s
T n (θ, I)
≤
T n (θ, ∅)
The final step of our proof is to control (6.3). Taking expectation with respect to V, V 0 ∼ π and
subtracting 1 from the preceding display yields:
√
2n|V ∩ V 0 |
⊗n
⊗n
− 1 + EV,V 0 ∼π 1(|V ∩ V 0 | ≥ s/2)( 2s)n
Dχ2 (P , P0 ) ≤ EV,V 0 ∼π exp
s
0
√
2n|V ∩ V |
V = [s] − 1 + EV 0 ∼π 1(|V ∩ V 0 | ≥ s/2) V = [s] ( 2s)n ,
= EV 0 ∼π exp
s
(6.5)
where in the final inequality we used the symmetry of the problem to condition on the event
V = [s]. In order to handle (6.5) we first recognize that given V = [s], the distribution of |V ∩ V 0 |
is hypergeometric: the number of red balls out of s balls drawn without replacement from an urn
consisting of s red balls and d − s blue balls. Using a standard bound of the tail probability of a
hypergeometric distribution (Chvátal, 1979) we have
s s d − s s
s s
0
EV 0 ∼π 1(|V ∩ V | ≥ s/2) V = [s] ≤ 2s
≤ 2s
.
d
d
d
P
To deal with the first term of the RHS of (6.5), we use the representation |V ∩V 0 | = v∈V 0 1(v ∈ V ),
and the fact that {1(v ∈ V )}v∈V 0 are negatively associated (Joag-Dev and Proschan, 1983, see,
e.g., for a proper definition). We have
Y
2n|V ∩ V 0 |
EV 0 ∼π exp
V = [s] − 1 ≤
P(v ∈ V )e2n/s + 1 − P(v ∈ V )
s
0
v∈V
≤ 1 + e2n/s
s s
s2 2n/s
≤ exp
e
.
d
d
Hence, continuing the bounds in (6.5) we finally have
Dχ2 (P
⊗n
, P⊗n
∅ ) ≤ exp
s2 2n/s
2s s √ n
e
− 1+
( 2s) .
d
d
It can be checked that under the sufficient conditions (4.3) the above expression goes to 0 asymptotically. With this our proof is complete.
Proof of Theorem 6.2. For brevity we will use the shorthand notation G := Cs (V ) and G0 :=
Cs (V 0 ). Recall the definitions of SxV and SSxV in (6.1). The statement is equivalent to showing that
V
V 0
V
d
ZG⊕G0 (θ)
V0
log
= EG SSX
+ EG0 SSX
− EG⊕G0 SSX
+ SSX
≥ 0.
dθ
ZG (θ)ZG0 (θ)
26
The latter is implied if we show that
V
V
EG SSX
≥ EG⊕G0 SSX
.
(6.6)
To this end, for any 0 ≤ t ≤ 1 define the model
0
PG⊕tG0 (X = x) ∝ exp − θ SSxV + tSSxV .
Clearly, PG⊕tG0 interpolates between the measures PG and PG⊕G0 , and (6.6) will be implied if we
showed that
d
EG⊕tG0 SSxV ≤ 0.
dt
A direct calculation yields
V
V
V
V 0
d
V0
EG⊕tG0 SSX
.
EG⊕tG0 SSX
= −θ EG⊕tG0 SSX
SSX
− EG⊕tG0 SSX
dt
Thus it suffices to show the following positive correlation inequality, for all 0 ≤ t ≤ 1.
V
V
V 0
V0
EG⊕tG0 SSX
SSX
≥ EG⊕tG0 SSX
EG⊕tG0 SSX
.
Notice that for any V and x ∈ {±1}V we have the identity (SxV )2 = 2SSxV + |V |. Hence the above
is precisely equivalent to
V 2 V 0 2
V 2
V0 2
EG⊕tG0 SX
SX
≥ EG⊕tG0 SX
EG⊕tG0 SX
.
(6.7)
If we were in the ferromagnetic case (6.7) would have followed by Griffiths inequality (Griffiths,
1967), which would have completed the proof. However, in antiferromagnetic cases the situation
is more challenging, as such positive correlation inequalities are not expected to hold in general.
Therefore below we utilize the special structure of our problem.
Due to the definition of PG⊕tG0 , the random vectors XV \(V ∩V 0 ) = (Xi )i∈V \(V ∩V 0 ) and XV 0 \(V ∩V 0 ) =
V ∩V 0 . Let X
f be an independent copy of X.
(Xi )i∈V 0 \(V ∩V 0 ) are conditionally independent given SX
Put for brevity
f := EG⊕tG0 S V 2 S V ∩V 0 − EG⊕tG0 S V 2 S V ∩V 0
HG (X, X)
X
X
f
f
X
X
f := EG⊕tG0 S V 0 2 S V ∩V 0 − EG⊕tG0 S V 0 2 S V ∩V 0 ,
HG0 (X, X)
X
X
f
f
X
X
and note that (6.7) can be equivalently expressed as
f G0 (X, X)
f ≥ 0.
EG⊕tG0 HG (X, X)H
The above will be implied if for any feasible h, e
h we showed that the following two numbers
V 2 V ∩V 0
V 2 V ∩V 0
EG⊕tG0 SX SX
= h − EG⊕tG0 SX
SX
=e
h ,
f
f
V 0 2 V ∩V 0
V 0 2 V ∩V 0
EG⊕tG0 SX
SX
= h − EG⊕tG0 SX
SX
=e
h ,
f
f
have the same sign. On the other hand, due to sign symmetry in X, the above is implied if the
following functions are increasing in h on the set N0 ∩ {−|V ∩ V 0 |, −|V ∩ V 0 | + 2, . . . , |V ∩ V 0 | −
2, |V ∩ V 0 |},
V 2 V ∩V 0
V 0 2 V ∩V 0
SX
=h .
EG⊕tG0 SX
SX
= h , EG⊕tG0 SX
27
0
By a simple calculation one can check that for any x ∈ {±1}|V \(V ∪V )| ,
2
θ V \(V ∩V 0 )
V ∩V 0
,
+h
PG⊕tG0 XV \(V ∩V 0 ) = x|SX
= h ∝ exp − Sx
2
and similarly for any x ∈ {±1}|V
0 \(V
PG⊕tG0 XV 0 \(V ∩V 0 ) =
∪V 0 )|
V ∩V 0
x|SX
= h ∝ exp
2
θt V 0 \(V ∩V 0 )
−
Sx
+h
.
2
Hence, due to this symmetry and the form of the distribution, it is sufficient to show that for any
k-clique graph H with |V (H)| = k, E = V (H)
, and any h ∈ N0
2
2
2
V (H)
V (H)
EH,h+2 SX
+ h + 2 − EH,h SX
+ h ≥ 0,
(6.8)
where
PH,h (X = x) ∝ exp
2
θ V (H)
.
+h
− Sx
2
If ZH,h (θ) denotes the partition function of PH,h , one can verify that the sign of the RHS of (6.8)
coincides with the sign of the derivative
ZH,h (θ)
d
log
,
dθ
ZH,h+2 (θ)
and therefore if suffices to show that for all θ ≥ θ0 ≥ 0 and h ∈ N0 we have
ZH,h (2θ0 )
ZH,h (2θ)
≤
,
0
ZH,h+2 (2θ )
ZH,h+2 (2θ)
(6.9)
where we scaled the parameters θ and θ0 by 2 for convenience. For two vectors X and Y put for
P
P
V (H)
V (H)
= i∈V (H) Yi , and let V (H) = {v1 , . . . , vk }.
brevity SX = SX
= i∈V (H) Xi and SY = SY
(6.9) is equivalent to
E∅ exp − θ(SY + h + 2)2 − θ0 (SX + h)2 ≤ E∅ exp − θ0 (SY + h + 2)2 − θ(SX + h)2 ,
where E∅ denotes the expectation with respect to uniformly drawing (Xv1 , . . . , Xvk , Yv1 , . . . , Yvk )
from the set {±1}2k . The random variables inside the expectations are discrete and non-negative
and therefore the inequality from the preceding display can be written as
Z ∞
2
0
2
P∅ exp − θ(SY + h + 2) − θ (SX + h) > t dt
0
Z ∞
0
2
2
≤
P∅ exp − θ (SY + h + 2) − θ(SX + h) > t dt,
0
where P∅ is the uniform measure on {±1}2k . Hence to prove (6.8) it suffices to show that for all
θ ≥ θ0 ≥ 0 the following stochastic dominance holds
2
0
2
0
2
2
P∅ θ(SY + h + 2) + θ (SX + h) < t ≤ P∅ θ (SY + h + 2) + θ(SX + h) < t
(6.10)
where t ≥ 0. We show this in the following
28
Proposition 6.3. Let {Xi }i∈[k] and {Yi }i∈[k] are i.i.d. Rademacher random variables, h ∈ N0 and
θ ≥ θ0 ≥ 0. Then (6.10) holds.
The proof of Proposition 6.3 is technical and we defer it to Appendix D. With this our proof is
complete.
7
Discussion
In this manuscript we formalized necessary and sufficient conditions for property testing in Ising
models. Specifically, we showed lower and upper information-theoretic bounds on the temperature
for ferromagnetic models. Furthermore, we argued that greedy correlation screening works well at
high temperature regimes, and can also be useful in low temperature regimes for certain properties.
We also demonstrated that testing strongly monotone properties over the class of general Ising
models is strictly more difficult than testing in ferromagnets. We discussed generic property tests
based on correlation decoding, and developed a computationally efficient cycle test.
Important problems that we plan to investigate in future work include — searching for more
sophisticated algorithms than correlation screening which will work at low temperature regimes for
testing any property in ferromagnets; relating the temperature of the system to the informationtheoretic limits of strongly monotone property testing in models with antiferromagnetic interactions; utilizing computationally efficient structure recovery algorithms, such as those of Bresler
(2015), to obtain tractable algorithms for property testing in general models.
Acknowledgments
The authors would like to express their gratitude to Professors Michael Aizenman, Ramon van
Handel and Allan Sly, for thought provoking discussions, their insights and encouragements. Special
thanks are due to Ramon, who suggested the interpolating measure PG⊕tG0 towards the proof of
Theorem 6.2.
References
Addario-Berry, L., Broutin, N., Devroye, L. and Lugosi, G. (2010). On combinatorial
testing problems. The Annals of Statistics 38 pp. 3063–3092.
Ahmed, A. and Xing, E. P. (2009). Recovering time-varying networks of dependencies in social
and biological studies. Proceedings of the National Academy of Sciences 106 11878–11883.
Alon, N. and Spencer, J. H. (2004). The probabilistic method. John Wiley & Sons Inc.
Anandkumar, A., Tan, V. Y., Huang, F. and Willsky, A. S. (2012). High-dimensional
structure estimation in ising models: Local separation criterion. The Annals of Statistics 1346–
1375.
Arias-Castro, E., Bubeck, S., Lugosi, G. and Verzelen, N. (2015a). Detecting markov
random fields hidden in white noise. arXiv preprint arXiv:1504.06984 .
Arias-Castro, E., Bubeck, S., Lugosi, G. et al. (2012). Detection of correlations. The
Annals of Statistics 40 412–435.
Arias-Castro, E., Bubeck, S., Lugosi, G. et al. (2015b). Detecting positive correlations in
a multivariate sample. Bernoulli 21 209–241.
29
Besag, J. (1993). Statistical analysis of dirty pictures. Journal of applied statistics 20 63–87.
Bhattacharya, B. B. and Mukherjee, S. (2017). Inference in ising models. Bernoulli .
Bollobás, B. (2004). Extremal graph theory. Courier Corporation.
Boucheron, S., Lugosi, G. and Massart, P. (2013). Concentration inequalities: A nonasymptotic theory of independence. OUP Oxford.
Bresler, G. (2015). Efficiently learning ising models on arbitrary graphs. In Proceedings of the
Forty-Seventh Annual ACM on Symposium on Theory of Computing. ACM.
Bresler, G., Gamarnik, D. and Shah, D. (2014). Structure learning of antiferromagnetic ising
models. In Advances in Neural Information Processing Systems.
Bresler, G., Mossel, E. and Sly, A. (2008). Reconstruction of markov random fields from samples: Some observations and algorithms. In Approximation, Randomization and Combinatorial
Optimization. Algorithms and Techniques. Springer, 343–356.
Cai, T. T., Liu, W. and Luo, X. (2011). A constrained `1 minimization approach to sparse
precision matrix estimation. Journal of the American Statistical Association 106 594–607.
Chow, C. and Liu, C. (1968). Approximating discrete probability distributions with dependence
trees. IEEE Trans. Inf. Theory 14 462–467.
Chvátal, V. (1979). The tail of the hypergeometric distribution. Discrete Mathematics 25 285–
287.
Cover, T. M. and Thomas, J. A. (2012). Elements of information theory. John Wiley & Sons
Inc.
Durbin, R., Eddy, S. R., Krogh, A. and Mitchison, G. (1998). Biological sequence analysis:
probabilistic models of proteins and nucleic acids. Cambridge university press.
Fisher, M. E. (1967). Critical temperatures of anisotropic ising lattices. ii. general upper bounds.
Physical Review 162 480.
Föllmer, H. (1988). Random fields and diffusion processes. École d’Été de Probabilités de SaintFlour XV–XVII, 1985–87 101–203.
Geman, S. and Geman, D. (1984). Stochastic relaxation, gibbs distributions, and the bayesian
restoration of images. IEEE Transactions on pattern analysis and machine intelligence 721–741.
Golub, G. H. and Van Loan, C. F. (2012). Matrix computations, vol. 3. JHU Press.
Grabowski, A. and Kosiński, R. (2006). Ising-based model of opinion formation in a complex
network of interpersonal interactions. Physica A: Statistical Mechanics and its Applications 361
651–664.
Griffiths, R. B. (1967). Correlations in ising ferromagnets. i. Journal of Mathematical Physics
8 478–483.
Ising, E. (1925). Beitrag zur theorie des ferromagnetismus. Zeitrschrift für Physik A Hadrons and
Nuclei 31 253–258.
Joag-Dev, K. and Proschan, F. (1983). Negative association of random variables with applications. The Annals of Statistics 286–295.
Kochmański, M., Paszkiewicz, T. and Wolski, S. (2013). Curie–weiss magnet—a simple
model of phase transition. European Journal of Physics 34 1555.
Liu, H., Lafferty, J. D. and Wasserman, L. A. (2009). The nonparanormal: Semiparametric
estimation of high dimensional undirected graphs. Journal of Machine Learning Research 10
2295–2328.
Meinshausen, N. and Bühlmann, P. (2006). High dimensional graphs and variable selection
with the lasso. The Annals of Statistics 34 1436–1462.
30
Montanari, A. and Pereira, J. A. (2009). Which graphical models are difficult to learn? In
Advances in Neural Information Processing Systems.
Neykov, M., Lu, J. and Liu, H. (2016). Combinatorial inference for graphical models. arXiv
preprint arXiv:1608.03045 .
Ravikumar, P., Wainwright, M. J. and Lafferty, J. D. (2010). High-dimensional ising
model selection using `1 -regularized logistic regression. The Annals of Statistics 38 1287–1319.
Ravikumar, P., Wainwright, M. J., Raskutti, G. and Yu, B. (2011). High-dimensional
covariance estimation by minimizing `1 -penalized log-determinant divergence. Electronic Journal
of Statistics 5 935–980.
Rothman, A. J., Bickel, P. J., Levina, E. and Zhu, J. (2008). Sparse permutation invariant
covariance estimation. Electronic Journal of Statistics 2 494–515.
Santhanam, N. P. and Wainwright, M. J. (2012). Information-theoretic limits of selecting
binary graphical models in high dimensions. Information Theory, IEEE Transactions on 58
4117–4134.
Shanmugam, K., Tandon, R., Ravikumar, P. K. and Dimakis, A. G. (2014). On the information theoretic limits of learning ising models. In Advances in Neural Information Processing
Systems.
Wasserman, S. and Faust, K. (1994). Social network analysis: Methods and applications, vol. 8.
Cambridge university press.
Yu, B. (1997). Assouad, fano, and le cam. In Festschrift for Lucien Le Cam. Springer, 423–435.
Yuan, M. and Lin, Y. (2007). Model selection and estimation in the gaussian graphical model.
Biometrika 19–35.
31
A
Auxiliary Results
In this section we spell out the details of some auxiliary results.
For the next result assume we have m distinct graphs Gj for j ∈ [m] which differ from a graph
G0 in a single edge, i.e. Gj = (V, Ej ), where |Ej \ E0 | = 1. For each j, denote the only edge in
the set Ej \ E0 by ej = (uj , vj ). Define for brevity Pj = Pθj for θj such that G(θj ) = Gj and
P0 = Pθ0 for a θ0 with G(θ0 ) = G where the measures Pθ0 , Pθj are defined in (2.2). Denote the
corresponding expectations with Ej and E0 .
Lemma A.1 (Edge Addition). The following bound holds
E0
Pj Pk
≤ 1 + tanh(θ)[Ej Xuk Xvk − E0 Xuk Xvk ].15
P0 P0
Remark A.2. By the FKG inequality we always have Ej Xuk Xvk − E0 Xuk Xvk ≥ 0, and therefore
P
E0 P0j PPk0 ≥ 1. Furthermore, if we are interested in product measures over n i.i.d. observations, i.e,
P⊗n
and P⊗n
0 Lemma A.1 immediately to yields the following bound
j
⊗n
P⊗n
Pj Pk n
j Pk
EP⊗n ⊗n ⊗n = E0
≤ (1 + tanh(θ)[Ej Xuk Xvk − E0 Xuk Xvk ])n .
0 P
P
P
P
0
0
0
0
Proof of Lemma A.1. Note that by definition we have:
Z θ0
Pk (x)
=
exp(θxuk xvk ).
P0 (x)
Z θk
Hence
Z θk
=
Z θ0
X
x∈{±1}d
exp(θxu xvk )
X )
k vk
Hence P0 (x) E0 exp(θXku
Z θk
Pk (x) =
Z θ0
X
exp(θxuk xvk )P0 (x) = E0 exp(θXuk Xvk ).
x∈{±1}d
= Pk (x). This shows that:
E0
Pj Pk
Ej exp(θXuk Xvk )
Pk
= Ej
=
.
P0 P0
P0
E0 exp(θXuk Xvk )
Using that for any c ∈ R and x ∈ {±1} we have the identity
exp(cx) = cosh(c)(1 + x tanh(c)).
(A.1)
implies that for Q = P0 , Pj
EQ exp(θXuk Xvk ) = sinh(θ)EQ Xuk Xvk + cosh(θ),
Consequently
E0
Ej exp(θXuk Xvk )
sinh(θ)[Ej Xuk Xvk − E0 Xuk Xvk ]
Pj Pk
=
=1+
P0 P0
E0 exp(θXuk Xvk )
sinh(θ)E0 Xuk Xvk + cosh(θ)
Ej Xuk Xvk − E0 Xuk Xvk
=1+
≤ 1 + tanh(θ)[Ej Xuk Xvk − E0 Xuk Xvk ]
E0 Xuk Xvk + coth(θ)
where for the last inequality we note that by the FKG inequality E0 Xuk Xvk ≥ 0. With this the
proof is complete.
15
We omit writing the argument X in the ratio
Pj (X) Pk (X)
P0 (X) P0 (X)
32
for brevity.
For a given graph G = (V, E) and vertex set W let G|W = (W, E|W ), where E|W = {(u, v) ∈
E | u, v ∈ W } is the restriction of G on the vertex set W . The next proposition states that when
studying correlations it suffices to consider a potentially smaller graph.
Proposition A.3. Let G = (V, E) be any graph, and Pθ be the measure of an Ising model with
G
graph G(θ) = G as in (5.1). Let u, v ∈ V be two fixed vertices. Denote by Pu→v
the set of all
G
e = G|V
simple paths on G connecting u and v, and let VG be the set of all vertices in Pu→v . Let G
G
e
and PθGe be the measure of the Ising model with θ restricted to G. Then
Eθ Xu Xv = EθGe Xu Xv .
e c = (V, E \ E|V ). Take any two distinct vertices
Proof of Proposition A.3. Construct the graph G
G
e c . To see
v1 , v2 ∈ VG . Note that v1 and v2 belong to two distinct connected components in G
c
e
this, assume the contrary. This implies that there exists a path on G connecting v1 and v2 which
G . For each vertex ` ∈ V
is a contradiction since such a path does not belong to the set Pu→v
G
c
e
let C` denote the connected component of G containing `, and let V (C` ) be its vertex set. Let
V c = V \ ∪`∈VG (V (C` ) \ {`}) denote the vertex set which is completely disconnected from the graph
e In terms of this notation we have the decomposition:
G.
V = ∪`∈VG (V (C` ) \ {`}) ∪ VG ∪ V c .
Let (X` )`∈VG and (Y` )`∈VG be two fixed configurations of spins in the vertex set VG . Take any
configuration of spins (x` )`∈V \VG , and consider the following two states
SX = ((X` )`∈VG , (X`
SY = ((Y` )`∈VG , (Y`
where by
(xw )w∈V (C` )\{`} )`∈VG , (x` )`∈V c )
(xw )w∈V (C` )\{`} )`∈VG , (x` )`∈V c ),
we mean element-wise multiplication. Due to the construction we conclude:
Pθ (SX )
= exp
Pθ (SY )
X
(w,`)∈E|VG
Pθ ((X` )`∈V )
G
e
.
θw` (Xw X` − Yw Y` ) = G
PθGe ((Y` )`∈VG )
(A.2)
Note that as we vary the spins (x` )`∈V \VG while holding (X` )`∈VG and (Y` )`∈VG fixed, the values (X` (xw )w∈V (C` )\{`} )`∈VG , (x` )`∈V c ) and (Y` (xw )w∈V (C` )\{`} )`∈VG , (x` )`∈V c ) take all possible
values. The last observation and (A.2) complete the proof.
Lemma A.4 (Path Graph Correlations). Let G = ([`], {(j, j + 1)}j∈[`−1] ) be a path graph, and Pθ
be the measure of an Ising model with the graph G(θ) = G as in (5.1). For any k ≤ j ∈ [`] we
have:
j−1
Y
Eθ Xk+1 Xj =
tanh(θu,u+1 ).
u=k+1
Proof of Lemma A.4. The proof follows by a standard induction argument, and is omitted.
33
B
Bounds with Ferromagnets
Since in this section all entries of the vector θ are of magnitude equal to θ there is a one to one
correspondence between a vector θ ∈ S(θ) and a graph G ∈ G. Hence it is more convenient to
denote the measure Pθ with PG where G = G(θ), and suppress the dependency on the signal
strenght θ whenever that does not cause confusion.
Proof of Theorem 2.4. Let S ⊆ C denote the r-packing set of C with cardinality m := |S| =
N (C, dG0 , r), where we have set for brevity r = ([− log c]−1 + ε) log log |C|. To ease of notation we
will use P0 for the measure PG0 and for an edge ej ∈ S we let Pj = PGj where Gj = G0 ∪ {ej }.
⊗n
⊗n
1 P
Recall the notation ⊗n for a product measure of n i.i.d. observations. Let P = m
j∈[m] Pj
denote the mixture density of the graphs from the alternative. By Definition 6.1 we have
⊗n
Dχ2 (P
, P⊗n
0 ) = EP⊗n
0
P⊗n
P⊗n
0
2
P⊗n
P⊗n
1 X
j
i
EP⊗n ⊗n
−1 = 2
⊗n − 1.
0 P
m
P
0
0
i,j∈[m]
⊗n
P⊗n P
We now handle the terms EP⊗n Pi⊗n Pj⊗n one by one, and we distinguish two cases.
0
0
0
Case I. First assume that i 6= j. We will show that the following bound holds.
Lemma B.1 (High Temperature Bound). Let G = (V, E) and G0 = (V, E 0 ) be two graphs such that
E\E 0 = {e}, where e = (u, v). Then if AG denotes the adjacency matrix of G, and kAG k1 tanh(θ) ≤
c < 1 for some small c > 0 the following holds
EG Xk X` − EG0 Xk X` ≤ 2
∞
X
l+1
tanh(θ)
[AlG ]uk + [AlG ]vk + [AlG ]u` + [AlG ]v` ,
l=0
where the expectations EG , EG0 are taken with respect to distributions as specified by (2.2).
Let Λ := maxdeg(G0 ) = kAG0 k1 , where AG0 is the adjacency matrix of G0 . If kAG0 k2 denotes
the operator norm of AG0 , recall that kAG0 k2 ≤ Λ by Gershgorin’s circle Theorem (Golub and
Van Loan, 2012). By the triangle inequality kAGj k2 ≤ kAG0 k2 + 1 ≤ Λ + 1 for all j. It immediately
follows by our assumption that
kAGj k1 tanh(θ) ≤ c < 1,
and hence the condition of Lemma B.1 is satisfied.
Recall that for any two vertices r and t in Vd , [AlGj ]rt equals the number of paths of length l
between vertex r and vertex t. Since [AlGj ]rt ≤ kAGj kl2 ≤ (Λ + 1)l , it follows that (Λ + 1)l is an
upper bound on the number of paths on the graph Gj between vertices r and t of length l. Take
e := (Λ + 1). Let dG (ei , ej ) = lij . Applying
the two distinct edges ei = (k, `), ej = (u, v) ∈ S. Put Λ
0
Lemma B.1 we have:
Ej Xk X` − E0 Xk X` ≤ 2
≤
∞
X
l+1
tanh(θ)
[AlGj ]uk + [AlGj ]vk + [AlGj ]u` + [AlGj ]v`
l=0
X
l≥lij
e l tanhl+1 (θ) ≤
8Λ
34
e tanh(θ))lij
8 tanh(θ)(Λ
e tanh(θ)
1−Λ
Case II. When i = j, Lemma 4 of Shanmugam et al. (2014) gives the following bound
Ej Xk X` − E0 Xk X` ≤ tanh(θ).
Applying Lemma A.1
⊗n
Dχ2 (P
, P⊗n
0 )+1≤
1 X
1 X
(1 + tanh(θ)[Ej Xk X` − E0 Xk X` ])n + 2
(1 + tanh(θ)2 )n
2
m
m
ei ,ej ∈S
i6=j
ej ∈S
e tanh(θ))lij n (1 + tanh(θ)2 )n
8 tanh2 (θ)(Λ
1 X
1+
≤ 2
+
e tanh(θ)
m
m
1−Λ
ei ,ej ∈S
i6=j
n
1 X
8
(1 + tanh(θ)2 )n
l
+2
ij
e tanh(θ))
≤ 2
1+
(Λ
+
e2
m
m
(1 − c)Λ
ei ,ej ∈S
i6=j
1
≤ 2
m
exp n
X
ei ,ej ∈S
i6=j
8
e2
(1 − c)Λ
e tanh(θ))lij +2
(Λ
+
exp(n tanh(θ)2 )
,
m
e −1 ) and c < 1, and the last one uses x + 1 ≤
where the next to last inequality uses θ ≤ atanh(cΛ
exp(x). Since S is anqr-packing we have that lij + 2 ≥ r + 2 > 2. Under the assumption that
tanh(θ) ≤ θ ≤ (1 − ε)
⊗n
Dχ2 (P
, P⊗n
0 )
log m
n ,
and
log d
n
= o(1) we have
exp(n tanh(θ)2 )
m−1
8
r+2
e
+
+1≤
(Λ tanh(θ))
exp n
e2
m
m
(1 − c)Λ
m−1
8
e tanh(θ))r+2 + o(1)
(Λ
≤
1 + (e − 1)n
e2
m
(1 − c)Λ
8
e tanh(θ))r+2 + o(1),
≤ 1 + (e − 1)n
(Λ
2
e
(1 − c)Λ
8n
r+2 ≤ 1 since by convexity ex ≤
e
e 2 (Λ tanh(θ))
(1−c)Λ
n
r+2 = o(1), hence
e
argue that in fact
e 2 (Λ tanh(θ))
(1−c)Λ
where the next to last inequality holds when
1 + (e − 1)x when x ≤ 1. In what follows we
the inequality is always true in an asymptotic sense. We have
n
e2
(1 − c)Λ
e tanh(θ))r+2 =
(Λ
2
(1 − ε)2
e tanh(θ))r ≤ (1 − ε) log(|C|)(Λ
e tanh(θ))r
log(m)(Λ
1−c
1−c
Taking a log results in
e tanh(θ)) → −∞,
log log |C| + log((1 − ε)2 /(1 − c)) + r log(Λ
e tanh(θ)) ≤ r log(c) ≤
where the limit holds since r = ([− log c]−1 + ε) log log(|C|) and hence r log(Λ
− log log |C| + ε log c log log |C|, and ε log c < 0 while |C| → ∞ asymptotically. We have established
that under the condition of the theorem
⊗n
lim sup Dχ2 (P
n→∞
35
, P⊗n
0 ) = 0.
Recall that by Le Cam’s lemma we have the bound
Rn (S0 (θ), S1 (θ)) ≥ inf
h
ψ
P⊗n
0 (ψ
= 1) + P
⊗n
q
i
1
⊗n
(ψ = 0) ≥ 1 −
Dχ2 (P , P⊗n
0 ).
2
(B.1)
Hence since the LHS of the inequality above is a lower bound on the risk, the proof is completed.
Proof of Lemma B.1. To prove this result we make usage of Dorbushin’s comparison technique (see
Theorem (2.8) Föllmer, 1988) which a is powerful tool for comparing discrepancies between Gibbs
measures based on their local specifications.
We first estimate Dobrishin’s interaction matrix C which measures the influence of a vertex r
on vertex m, and is defined elementwise as
Crm =
sup {DTV (PG,m (·|X−m = x), PG,m (·|X−m = y))}.
x=y
off
r
In the above, x and y are d − 1 dimensional vectors, x = y off r means that x and y coincide
except the entries xr and yr , PG,m denotes the conditional distribution of Xm given the values of
all remaining spins X−m , and we have denoted the total variation between two measures by DTV 16 .
For a vertex m let Nm denote the set neighbors of m in the graph G. We have
X
PG,m (X = x|X−m = x) = g θx
xj
j∈Nm
for x ∈ {±1} and g(z) = ez P
/(e−z + ez ). It therefore
immediately follows that Crm = 0 if r 6∈ Nm .
P
Suppose r ∈ Nm . Let Sr = j∈Nm \{r} xj = j∈Nm \{r} yj . We have
Crm = sup |g θ(Sr + 1) − g θ(Sr − 1) | = sup
Sr
≤
Sr
e2θ
2+
−
e2θ
e−2θ
+ e−2θ
e2θ − e−2θ
e−2Sr θ + e2Sr θ + e2θ + e−2θ
= tanh(θ).
We conclude that Dobrushin’s interaction matrix is C = AG tanh(θ), where AG is the adjacency
matrix of the graph G. Notice that the matrix C satisfies Dobrushin’s condition
lim kCl k1 = 0,
l→∞
since kCk1 = kAG k1 tanh(θ) < 1, and therefore liml→∞ kCl k1 ≤ liml→∞ kCkl1 = 0. Next we will
find the discrepancy between the averaged marginal conditional measures PG and PG0 . We define
X
bm :=
DTV (PG,m (·|X−m = x), PG0 ,m (·|X−m = x))PG0 (X−m = x).
x∈{±1}d−1
16
Recall that for two probability measures P and Q on a probability space (Ω, Σ), we have DTV (P, Q) :=
supA∈Σ |P(A) − Q(A)|.
36
Since G and G0 differ
P only on the edge (u, v) we have that bm = 0 unless m = u or m = v. Let
m = u. Put S = j∈Nm xj . We have
|PG,u (X = x|X−u = x) − PG0 ,u (X = x|X−u = x)| = |g θxS − g θx(S − xv ) |
eθ − e−θ
+ e−2θS+θxv + e−θ + eθ
tanh(θ)
≤
≤ tanh(θ).
1/ cosh(θ) + 1
=
e2θS−θxv
Therefore we conclude that bu ≤ tanh(θ) and hence by symmetry the same inequality also holds
for bv . Since we are interested in bounded the correlation of Xk and X` under the two measures
we define the comparison function f : {±1}d 7→ R
f (X) := Xk X` .
and the oscillation of f at site r:
δr (f ) := sup{|f (x) − f (y)| : x = y off r}.
Note that δr (f ) = 0 if r 6∈ {k, `} and δk (f ) = δ` (f ) = 2. Define the matrix
D :=
∞
X
l
tanh(θ)AG ,
l=0
and let b be the vector with entries bm . According to Dobrushin’s comparison theorem we have
EG Xk X` − EG0 Xk X` ≤ 2([b> D]k + [b> D]` ).
(B.2)
Hence (B.2) is equivalent to
EG Xk X` − EG0 Xk X` ≤ 2
∞
X
l+1
tanh(θ)
[AlG ]uk + [AlG ]vk + [AlG ]u` + [AlG ]v` ,
l=0
which is what we wanted to show.
Lemma B.2 (Low Temperature Bound). Let G = (V, E) be a graph, and k, ` ∈ V be vertices such
that (k, `) 6∈ E. Let r ≥ 2 and l are integers such that there exists an l × r biclique containing k
3
and ` on its right, i.e., r side. Then for values of θ ≥ 2l , θ ≥ r−2
for r > 2 and θ ≥ log 2 when r = 2
we have:
2(r − 1)
EG Xk X` ≥ 1 −
.
exp(θl) + r − 1
Proof of Lemma B.2. By the FKG inequality (Alon and Spencer, 2004) pruning edges reduces the
correlations. Hence we may assume without loss of generality that G is an l × r-biclique. We first
show the case when r ≥ 3. By a direct calculation we have
Pr−2 Pl
r−2 l
PG (Xk X` = 1)
m=0
j=0 m
j exp[θ(r − 2m)(l − 2j)]
= Pr−2 Pl
.
r−2 l
PG (Xk X` = −1)
exp[θ(r − 2m − 2)(l − 2j)]
m=0
j=0
m
37
j
To this end note that
r−2
r−2 X
l
X
X
r−2
r−2
l
l
cosh(θ(r − 2m))l .
exp[θ(r − 2m)(l − 2j)] = 2
m
m
j
m=0
m=0 j=0
Similarly,
r−2 X
l
X
r−2
l
m=0 j=0
m
j
exp[θ(r − 2m − 2)(l − 2j)] = 2
l
r−2
X
r−2
m=0
m
cosh(θ(r − 2m − 2))l .
Hence we obtain the identity
Pr−2 r−2
l
PG (Xk X` = 1)
m=0 m cosh(θ(r − 2m))
= Pr−2
.
r−2
l
PG (Xk X` = −1)
m=0 m cosh(θ(r − 2m − 2))
(B.3)
To obtain a bound on (B.3) we will search for the argmax of the denominator. We will first argue
that if m∗ is the argmax of the denominator them m∗ satisfies 0 ≤ m∗ ≤ r−2
e . Recall that cosh is
an even function and hence by symmetry (m ↔ r − 2 − m):
r−2
r−2
=
, cosh(θ(r − 2m − 2)) = cosh(θ(r − 2(r − 2 − m) − 2)).
m
r−2−m
Therefore, to find a maximum in the denominator we only need to focus on terms satisfying m ≤
b(r − 2)/2c. To show that 0 ≤ m∗ ≤ r−2
e , we begin by comparing the term at m = 0 with
any term at m satisfying (r − 2)/e < m ≤ d(r − 2)/2e, i.e., we will compare cosh(θ(r − 2))l to
r−2
l
m cosh(θ(r − 2m − 2)) . First we record a well known bound on the binomial coefficients
r−2
(r − 2)e m
.
≤
m
m
We will now argue that
cosh(θ(r − 2))
≥ exp(mθ).
cosh(θ(r − 2m − 2))
(B.4)
Direct calculation shows that the above is equivalent to
exp(θ(2r − 4 − 4m)) ≥
1 − exp(−3mθ)
.
exp(mθ) − 1
(B.5)
−3x
−3x (2ex + e2x + 3) < 0), and hence
Note that the function 1−e
ex −1 is decreasing (its derivative is −e
reaches its maximum at small values of θm. We have:
θm >
3 r−2
3
= .
r−2 e
e
−9/e
It is simple to check that 1−e
< 1. Meanwhile since 2r − 4 − 4m ≥ 0 the left hand side of (B.5)
e3/e −1
is at least 1 and consequently (B.4) holds. Therefore when
(r − 2)e m
m
< exp(θml),
38
i.e., when m >
r−2
e
(using θl > 2) we have
r−2
cosh(θ(r − 2)) ≥
cosh(θ(r − 2m − 2))l .
m
l
Hence the maximum is reached at m∗ such that 0 ≤ m∗ ≤ r−2
e . We therefore have the bound:
r−2
∗ l
cosh(θ(r − 2m∗ ))l
PG (Xk X` = 1)
m∗ cosh(θ(r − 2m ))
=
.
≥
∗
l
PG (Xk X` = −1)
(r − 1) cosh(θ(r − 2m∗ − 2))l
(r − 1) r−2
m∗ cosh(θ(r − 2m − 2))
Finally we will argue that
cosh(θ(r − 2m∗ ))
≥ exp(θ).
cosh(θ(r − 2m∗ − 2))
Similarly to before we need to verify:
exp(θ(2r − 4 − 4m∗ )) ≥
1 − exp(−3θ)
,
exp(θ) − 1
(B.6)
1−exp(−3x)
∗
We have 1−exp(−3θ)
exp(θ)−1 ≤ limx→0 exp(x)−1 = 3, and thus (B.6) is implied when exp(θ(2r −4−4m )) ≥
3. Using the fact that m∗ ≤ r−2
e we have:
exp(θ(2r − 4 − 4m∗ )) ≥ exp(θ(2 − 4/e)(r − 2)) ≥ exp(6 − 12/e) > 3,
where we used the fact that θ(r − 2) ≥ 3. Compiling these results yields:
PG (Xk X` = 1)
exp(θl)
≥
PG (Xk X` = −1)
r−1
Hence
EG Xk X` = PG (Xk X` = 1) − PG (Xk X` = −1) ≥ 1 −
(B.7)
2(r − 1)
,
exp(θl) + r − 1
as we claimed. This completes the proof when r > 2. In the case when r = 2, a direct calculation
shows
PG (Xk X` = 1)
= cosh(2θ)l .
PG (Xk X` = −1)
Clearly then when θ ≥ log 2, we have cosh(2θ) ≥ exp(θ), and the proof can proceed in the same
way as before. With this the result is complete.
log
2κnr
logbd/(l+r)c
Proof of Theorem 2.6. First we point out a simple implication of (2.5). If θ ≥
, then
l
certainly
h
i
2κn(r−1)
log logbd/(l+r)c
− (r − 1)
θ≥
.
l
We will now argue that even if the above holds the asymptotic minimax risk is still 1. Note that
since B E H0 we have |V (H0 )| = |V (B)| = l + r. Consider a graph G0 based on the union of
m = bd/(l + r)c disconnected copies of the graph H0 . Let ek = (uk , vk ) is the local copy of the
edge e = (u, v) in the k th copy of H0 . Since the property P can be represented as a maximum
over subgraphs of G0 , by the assumption of the theorem P(G0 ) = 0, while adding the edge ek
39
for any k ∈ [m] to the k th copy of H0 transfers G0 to a graph Gk satisfying P(Gk ) = 1. For
each graph in i ∈ {0, 1, . . . , m} let Pi be the measure corresponding to Ising models with graphs
{G0 , G1 , . . . , Gm }, and Ek be the corresponding expectation under Pk . Define the mixture measure
⊗n
⊗n
1 P
P =m
k∈[m] Pk . Using Lemma A.1
Dχ2 (P
⊗n
, P⊗n
0 )+1≤
1 X
(1 + tanh(θ)[Ej Xuk Xvk − E0 Xuk Xvk ])n .
m2
j,k∈[m]
By Proposition A.3, for i 6= j we have
Ej Xuk Xvk − E0 Xuk Xvk = 0,
while by Lemma B.2 if i ≡ j
Ej Xuk Xvk − E0 Xuk Xvk ≤ 1 − E0 Xuk Xvk ≤
2(r − 1)
.
exp(θl) + r − 1
We conclude that
Dχ2 (P
⊗n
n
m−1
1
2(r − 1)
+
1 + tanh(θ)
m
m
exp(θl) + r − 1
n
m−1
1
2(r − 1)
≤
+
1+
m
m
exp(θl) + r − 1
2n(r−1)
exp
exp(θl)+r−1
m−1
+
.
≤
m
m
, P⊗n
0 )+1≤
The proof is completed via an application of Le Cam’s argument (B.1).
Remark B.3. Theorem 2.6 demonstrates that some strongly monotone properties have upper
bound signal strength limitations. In fact it is evident from the proof that the scaling on θ in (2.5)
can be slightly improved. It turns out that under the same conditions, it suffices that if
2 tanh(θ)(r − 1)
d
= o log
,
exp(θl) + r − 1
|l + r|
holds instead of (2.5) we still have that the minimax risk goes to 1 asymptotically.
C
Correlation Screening for Ferromagnets
Lemma C.1. Given n samples from an Ising model satisfying (5.1). We have that
b u Xv − Eθ Xu Xv | ≥ ε ≤ 2 d e−nε2 /2 .
Pθ max |EX
u,v∈Vd
2
(C.1)
Note that since (3.1) is a special case of (5.1), the same conclusion is also valid for ferromagnets
satisfying (3.1).
Proof of Lemma C.1. This is a direct consequence of Hoeffding’s inequality (Boucheron et al., 2013)
and the union bound.
40
Proof of Theorem 3.1. Recall the definitions of T (3.2) and τ (3.3). Using Lemma C.1 we can
guarantee that if θ ∈ S1 (θ) ∩ R
min Me ≥ T − τ,
b
e∈G
with probability at least 1 − δ. This immediately shows that
sup
θ∈S1 (θ)∩R
Pθ (ψ = 0) ≤ δ.
b ∈ WP (Gd ), we have that if P(G) = 0
Next since G
min Me ≤ max
b
e∈G
max
min
θ∈S0 ∩R G∈WP (Gd ) (u,v)∈E(G)
Eθ Xu Xv + τ,
on an event of probability at least 1 − δ. When (3.5) holds we have
min Me < T − τ,
b
e∈G
and thus Algorithm 1 will return ψ = 0. This completes the proof.
Lemma C.2 (Generic No Edge Correlation Upper Bound). Assume we have a ferromagnetic Ising
model as specified by (3.1) where all coefficients θuv satisfy 0 < θuv ≤ Θ. Then if the maximum
degree of G(θ) is bounded by s ≥ 3, for any two vertices k and ` such that (k, `) 6∈ E(θ) we have
Eθ Xk X` ≤
cosh(2sΘ) + 2se−2(s−1)Θ (cosh(2(s − 1)Θ) − cosh(2Θ)) − 1
R(s, Θ) − 1
=
.
−2(s−1)Θ
R(s, Θ) + 1
cosh(2sΘ) + 2se
(cosh(2(s − 1)Θ) + cosh(2Θ)) + 1
If s = 2 we have
Eθ Xk X` ≤
(C.2)
cosh(4Θ) − 1
.
cosh(4Θ) + 3
Finally if s = 1, Eθ Xk X` = 0.
Proof of Lemma C.2. First, by the FKG inequality (see Theorem 6.2.1 Alon and Spencer (2004))
it follows that since all 0 ≤ θuv ≤ Θ the correlation Eθ Xk X` will only increase if we were to set all
θuv = Θ. Hence we may assume that all θuv = Θ for all u, v.
Let Nk and N` denote the neighbors of vertices k and ` respectively. Again, due to the FKG
inequality adding edges increases correlations so we can assume that |Nk | = |N` | = s. We have
X
X
X
X
X
exp ξΘ
xu +
xu + Θ
xu xv
d
u∈N
u∈N
(u,v)∈E(θ),k,`6∈{u,v}
Pθ (Xk X` = 1)
x∈{±1} ξ∈{±1}
k
`
X
.
= X
X
X
X
Pθ (Xk X` = −1)
exp ξΘ
xu −
xu + Θ
xu xv
x∈{±1}d ξ∈{±1}
u∈Nk
u∈N`
(u,v)∈E(θ),k,`6∈{u,v}
(C.3)
Let N = Nk ∩ N` and suppose |N | = m for 0 ≤ m ≤ s. Put N k := Nk \ N and similarly let
N ` := N` \ N . For a vector x ∈ Rd and a set S ⊆ Vd let xS = (xi )i∈S . For ξ ∈ {±1} define the sets
X
X
S1 := x ∈ {±1}d
xu +
xu = 2s ,
u∈N`
u∈Nk
S2 := x ∈ {±1}d
X
xu +
u∈Nk
41
X
u∈N`
xu ≤ 2(s − 2) ,
and let S3 := {±1}d \ (S1 ∪ S2 ). We now record several identities for the three sets. For any x ∈ S1
we have
X
X
X
X
X
X
exp ξΘ
xu −
xu = 2.
exp ξΘ
xu +
xu = 2 cosh(2sΘ),
u∈Nk
ξ∈{±1}
u∈N`
u∈Nk
ξ∈{±1}
u∈N`
The
P first identity
P holds by the definition of S1 , while the second holds since if x ∈ S1 it follows that
x
−
u
u∈Nk
u∈N` xu = 0. For any x ∈ S2 :
X
X
X
X
X
X
xu ≥ 2.
xu −
exp ξΘ
xu ≤ 2 cosh(2(s − 2)Θ),
xu +
exp ξΘ
u∈Nk
ξ∈{±1}
u∈N`
u∈Nk
ξ∈{±1}
u∈N`
The first inequality follows by the definition of S2 , and the second is true since since for any x ∈ R
ex + e−x ≥ 2. Finally for any x ∈ S3 :
X
X
X
X
X
X
exp ξΘ
xu −
xu ≥ 2 cosh(2Θ).
exp ξΘ
xu +
xu = 2 cosh(2(s−1)Θ),
u∈Nk
ξ∈{±1}
u∈N`
u∈Nk
ξ∈{±1}
P
u∈N`
P
To see why the first identity holds, note that the sum u∈Nk xu + u∈N` xu contains 2s odd terms
P
P
(equal to ±1), and is therefore even. Hence since 2(s − 2) <
u∈Nk xu +
u∈N` xu < 2s we must
P
P
have
x
+
x
=
2(s
−
1).
To
show
why
the
inequality
holds
suppose the contrary,
u
u
u∈Nk P
u∈N` P
P
P
P
P
i.e., suppose
xu − u∈N ` xu
u∈N` xu ≤ 1. Since
u∈Nk xu −
u∈N` xu =
u∈Nk xu −
u∈N
k
P
P
and |N k ∪ N ` | = 2(s − m), the sum u∈N k xu − u∈N ` xu is even. Hence the only possibility
P
P
P
is
Nk xu = 2(s − 1). The latter is impossible since
u∈N ` xu , and therefore 2
P u∈N k xu =
x
≡
s
(mod
2).
Nk u
We first show the result when s ≥ 3. Let 1 ∈ Rd denote a vector of 1’s. Observe the following
inclusions
x ∈ {±1}d | ∃ξ ∈ {±1}, xNk ∪N` = ξ1Nk ∪N` = S1 ,
S 2 := x ∈ {±1}d | ∃ξ ∈ {±1}, ∃u ∈ N : xNk ∪N` \{u} = ξ1Nk ∪N` \{u} , xu = −ξ ⊆ S2 ,
S 3 := x ∈ {±1}d | ∃ξ ∈ {±1}, ∃u ∈ N k ∪ N ` : xNk ∪N` \{u} = ξ1Nk ∪N` \{u} , xu = −ξ ⊆ S3 .
Notice that whens ≥ 3: |S 2 | ≥ m|S1 | and |S
3 | ≥ 2(s − m)|S1 |. For i ∈ [3] define the three sums
P
P
Σi := x∈Si exp Θ (u,v)∈E(θ),k,`6∈{u,v} xu xv . Using (C.3) and the identities we recorded above
we conlcude
Pθ (Xk X` = 1)
cosh(2sΘ)Σ1 + cosh(2(s − 2)Θ)Σ2 + cosh(2(s − 1)Θ)Σ3
≤
.
Pθ (Xk X` = −1)
Σ1 + Σ2 + cosh(2Θ)Σ3
(C.4)
We now use the following elementary inequality which can be checked via cross multiplication: for
positive real numbers a, b, c, d, x, y > 0 such that ac ≥ db and x ≥ y it holds that
a + by
a + bx
≤
.
c + dx
c + dy
Taking x = ∞, y = 1 in (C.5) implies that
cosh(2sΘ)Σ1 + cosh(2(s − 1)Θ)Σ3
cosh(2(s − 1)Θ)
≥ cosh(2sΘ) ∧
≥ cosh(2(s − 2)Θ),
Σ1 + cosh(2Θ)Σ3
cosh(2Θ)
42
(C.5)
where the last inequality follows by simple algebra and the fact that cosh(x) is increasing for x ≥ 0.
We now record the inequality
P
P
(u,v)∈E(θ),k,`6∈{u,v} xu xv
x∈S 2 exp Θ
Σ2
≥ me−2(s−2)Θ ,
≥
P
P
Σ1
x∈S1 exp Θ
(u,v)∈E(θ),k,`6∈{u,v} xu xv
which follows by the fact that |S 2 | ≥ m|S1 | and that all vertices u ∈ N are connected to at most
s − 2 vertices in the set Vd \ {k, `}. Combining the last two inequalities with (C.4) and (C.5) we
obtain
3
cosh(2sΘ) + cosh(2(s − 2)Θ)m exp(−2(s − 2)Θ) + cosh(2(s − 1)Θ) Σ
Pθ (Xk X` = 1)
Σ1
. (C.6)
≤
3
Pθ (Xk X` = −1)
1 + m exp(−2(s − 2)Θ) + cosh(2Θ) Σ
Σ
1
We now consider two cases.
Case I. Suppose
cosh(2sΘ) + cosh(2(s − 2)Θ)me−2(s−2)Θ
cosh(2(s − 1)Θ)
≤
.
cosh(2Θ)
1 + me−2(s−2)Θ
(C.7)
Taking y = 0 in (C.5), and using (C.6) yields
cosh(2(s − 1)Θ)
Pθ (Xk X` = 1)
≤
.
Pθ (Xk X` = −1)
cosh(2Θ)
(C.8)
Case II. Assume that (C.7) holds in the opposite direction. We have
P
P
(u,v)∈E(θ),k,`6∈{u,v} xu xv
x∈S 3 exp Θ
Σ3
≥ 2(s − m)e−2(s−1)Θ ,
≥
P
P
Σ1
x∈S1 exp Θ
(u,v)∈E(θ),k,`6∈{u,v} xu xv
where we used that |S 3 | ≥ 2(s−m)|S1 | and that for any x ∈ S 3 the vertex u ∈ N k ∪N ` is connected
to at most s − 1 vertices in the sed Vd \ {k, `}. (C.5) and (C.6) imply
Pθ (Xk X` = 1)
cosh(2sΘ) + cosh(2(s − 2)Θ)me−2(s−2)Θ + cosh(2(s − 1)Θ)2(s − m)e−2(s−1)Θ
≤
Pθ (Xk X` = −1)
1 + me−2(s−2)Θ + cosh(2Θ)2(s − m)e−2(s−1)Θ
Taking the supremum over 0 ≤ m ≤ s on the RHS shows that the maximum is reached at m ≡ 0.
The latter can be verified via comparing consecutive values of m and arguing that the RHS is
decreasing in m; we omit this lengthly calculation. Finally a simple comparison between the RHS
of the last bound evaluated at m = 0 and the RHS of (C.8) shows that the last bound is larger.
Putting everything together we have concluded that for s ≥ 3
Pθ (Xk X` = 1)
cosh(2sΘ) + 2se−2(s−1)Θ cosh(2(s − 1)Θ)
≤
.
Pθ (Xk X` = −1)
1 + 2se−2(s−1)Θ cosh(2Θ)
43
For the special case s = 2, the major difference is that when m = 2, |S 2 | ≥ |S1 | and not
|S 2 | ≥ 2|S1 | as before. Using the same ideas as in the proof of the case s ≥ 3 one can show that
when s ≡ 2:
Pθ (Xk X` = 1)
cosh(4Θ) + 1
≤
.
Pθ (Xk X` = −1)
2
The last two inequalities combined with the fact Eθ Xk X` = 1 − 2Pθ (Xk X` = −1) complete the
proof.
The following simple lemma gives a sharper correlation bound when no edge is present, in a
regime where the maximum parameter Θ is sufficiently small.
Lemma C.3 (High Temperature No Edge Correlation Upper Bound). Let the assumptions of
Lemma C.2 hold, and let additionally (s − 1) tanh(Θ) < 1. Then we have
Eθ Xk X` ≤
s tanh2 (Θ)
.
1 − (s − 1) tanh(Θ)
Proof of Lemma C.3. As in the proof of Lemma C.2, we may assume that all θuv = Θ. Next, notice
that each path P connecting k and ` has length m ≥ 2. Let θP be the restriction of the parameter
θ to the path graph P . By Lemma A.4 we know that EθP Xi Xj = tanhm (θ), and furthermore there
are at most s(s − 1)m−2 such paths, due to the fact that each vertex has at most s neighbors. The
conclusion now follows from a result by Fisher (1967) which bounds the correlation between two
nodes with a sum over the correlations of all path graphs between the two vertices, and the formula
for summing converging geometric progressions.
Proof of Corollary 3.2. Note that if θ ∈ S0 ∩ R for any G ∈ WP (θ) there exists an edge e ∈ E(G)
with a zero entry in θ. Fact i. now follows since by Lemma C.2 and the definition of R(s, Θ) and
we have for s ≥ 3
R(s, Θ) − 1
max
max
min Eθ Xu Xv ≤
,
θ∈S0 ∩R G∈WP (Gd ) (u,v)∈E(G)
R(s, Θ) + 1
and fact ii. follows by noting that when (s − 1) tanh(Θ) < 1 by Lemma C.3
max
max
min
θ∈S0 ∩R G∈WP (Gd ) (u,v)∈E(G)
Eθ Xu Xv ≤
s tanh2 (Θ)
.
1 − (s − 1) tanh(Θ)
Proof of Corollary 3.5. We will now argue that if G is under the alternative, i.e., if G is connected,
then for any spanning tree T ⊆ G(θ) we have
min Eθ Xu Xv ≥ tanh(θ).
(u,v)∈T
(C.9)
By the FKG inequality (Alon and Spencer, 2004) we have min(u,v)∈T Eθ Xu Xv ≥ min(u,v)∈T EθT Xu Xv ,
where θT is the restriction of θ on T , i.e., setting all entries of θ outside of T to 0 while retaining
the values of all entries belonging to T . Using Proposition A.3, Lemma A.4 and the fact that on a
44
tree T the only path between vertices u and v connected by an edge (u, v) is the edge we conclude
that
EθT Xu Xv = tanh(θuv ).
Hence
min Eθ Xu Xv ≥ min EθT Xu Xv = min tanh(θuv ) ≥ tanh(θ),
(u,v)∈T
(u,v)∈T
(u,v)∈T
as we claimed. On the other hand, when G(θ) is under the null for any tree T , we have
min Eθ Xu Xv = 0.
(u,v)∈T
Hence an application of Theorem 3.1 completes the proof.
Proof of Corollary 3.7. Similarly to the proof of Proposition 3.5 we can convince ourselves with
the help of the FKG inequality, Proposition A.3 and Lemma A.4 that under the alternative there
exists at least one cycle C ⊆ G(θ) so that
min Eθ Xu Xv ≥ min EθC Xu Xv ≥ min tanh(θuv ) ≥ tanh(θ),
(u,v)∈C
(u,v)∈C
(u,v)∈C
where θC is the restriction of θ on C. Next, note that under the null hypothesis for any given cycle
C we will have at least one edge (u, v) ∈ C such that θuv = 0. Since under the null the graph is a
forest by Proposition A.3 and Lemma A.4 we know that
Y
min Eθ Xu Xv ≤
tanh(θkl ) ≤ tanh2 (Θ),
(u,v)∈C
G(θ)
(k,l)∈Pu→v
G
where Pu→v
is the direct path that connects u to v. Applying Theorem 3.1 completes the proof.
Proof of Corollary 3.10. Take any model with a graph G containing an m-clique. By the FKG
inequality we know that the correlation for any edge (u, v) ∈ E(θ): Eθ Xu Xv will only decrease if
we set all parameters to their lower bound θ, and if we remove edges so that we reduce the graph G to
an m-clique. We recognize that under such manipulation we obtain a Curie-Weiss model17 (albeit
under non-standard parametrization) for m of the variables, and remaining d − m variables are
independent and have Rademacher distributions. Hence we only need to calculate Pθ (Xk X` = ξ)
for ξ ∈ {±1} in an m-dimensional Curie-Weiss model. This is a standard calculation which we
nevertheless include for completeness. We have
Pθ (Xk X` = 1) =
eθ−θ(m−2)/2
S(2θ, θ/2),
Zm (θ)
Pθ (Xk X` = −1) =
e−θ−θ(m−2)/2
S(0, θ/2),
Zm (θ)
(C.10)
where Zm (θ) is the partition constant of an m-dimensional Curie-Weiss magnet with inverse temperature θ and no external magnetic field, i.e.,
X
X
Zm (θ) =
exp θ
xu xv ,
(C.11)
x∈{±1}m
17
i.e., a fully connected m-clique with inverse temperature θ.
45
u<v
and the function S(h, µ) = S(−h, µ) given by
S(h, µ) :=
2
X
X
.
xu + µ
xu
exp h
X
x∈{±1}m−2
u∈[m−2]
u∈[m−2]
Clearly since P(Xk X` = 1) + P(Xk X` = −1) = 1, one can express Zm (θ) in terms of S(2θ, θ/2) and
S(0, θ/2). We set out to find a closed form expression of the more general quantity S(h, µ). Using
the identity
Z ∞
√
1
2
x2
e =√
e−y /2+ 2xy dy,
2π −∞
we can rewrite S(h, µ) as
1
S(h, µ) = √
2π
X
x∈{±1}m−2
Z
∞
e−y
2 /2+(y
√
2µ+h)
P
u∈[m−2]
xu
dy
−∞
p
= 2m−2 EZ coshm−2 ( 2µZ + h).
(C.12)
Putting (C.10) and (C.12) together we conclude
Eθ Xk X` = Pθ (Xk X` = 1) − Pθ (Xk X` = −1) =
r(m, θ) − 1
.
r(m, θ) + 1
This implies that under the alternative
min Eθ Xu Xv ≥
(u,v)∈C
r(m, θ) − 1
,
r(m, θ) + 1
for any m-clique C ⊆ G, which is what we wanted to show.
Remark C.4. In this remark we will first show a low temperature expansion of
the equivalent formulation
r(m, θ) − 1
d log Zm (θ)/dθ
=
,
m
r(m, θ) + 1
2
r(m,θ)−1
r(m,θ)+1 .
Note
where Zm (θ) is defined in (C.11). Now observe that
m
m −4(m−2)θ
θ
−2(m−1)θ
(
)
2
Zm (θ) = 2e
1 + me
+
e
+ O(m3 e−6(m−3)θ ) ,
2
where the first term corresponds to all spins being 1 or −1, the second term corresponds to the m
terms where one spin is −1 and all remaining spins are 1 and vice versa, and so on. Hence
m
m −4(m−2)θ
−2(m−1)θ
log Zm (θ) = log 2 +
θ + log 1 + me
+
e
+ O(m3 e−6(m−3)θ )
2
2
m
m −4(m−2)θ
= log 2 +
θ + me−2(m−1)θ +
e
+ O(m3 e−6(m−3)θ )
2
2
46
Under the assumption m2 e−2mθ = o(1), we have the approximation
r(m, θ) − 1
≈ 1 − 4e−2(m−1)θ − 4(m − 2)e−4(m−2)θ .
r(m, θ) + 1
Now recalling bound (C.2), we can equivalently rewrite its RHS as
1−
4se−2(s−2)Θ + 4se−2sΘ + 4
.
1 + e2sΘ + e−2sΘ + s(1 + e−4(s−1)Θ + e−2(s−2)Θ + e−2sΘ )
Under the assumption e2sΘ s the above expression is asymptotically equivalent to
1 − 4e−2sΘ − 4se−4(s−1)Θ + o(1).
Therefore, when m = s + 1, and θ = Θ the asymptotic difference is of magnitude 4e−4(s−1)θ . Hence
the LHS of (3.7) is asymptotically 4e−4(s−1)θ and thus if
r
4 log d + log δ −1
−4(s−1)θ
,
2e
≥
n
the test is successful. This is equivalent to θ ≤
scalars.
D
log
n
log d+log δ −1 /4
8(s−1)
, which matches bound (2.10) up to
Bounds for General Models
Proof of Proposition 6.3. Define the two ellipses
E1 = {(x, y) ∈ R2 : θ(x+h)2 +θ0 (y+h+2)2 < t}, and E2 = {(x, y) ∈ R2 : θ0 (x+h)2 +θ(y+h+2)2 < t}.
Clearly both ellipses are centered at the point (−h, −(h + 2)). Let us begin by calculating the
intersection points of the contours of the two ellipses in the case when θ 6= θ0 (the case θ = θ0 is
trivial, as (6.10) holds with an equality). The contours intersect when
θ(x + h)2 + θ0 (y + h + 2)2 = θ0 (x + h)2 + θ(y + h + 2)2 = t.
Hence at the intersection, we necessarily have (x − y − 2)(x + y + 2h + 2) = 0. Note that the
equations x − y − 2 = 0, and x + y + 2h + 2 = 0 define two lines, which pass through the center
(−h, −(h + 2)), and partition the x, y-plane into four regions. Denote those regions as
E = {(x, y) ∈ R2 : x − y − 2 ≥ 0, x + y + 2h + 2 ≥ 0},
N = {(x, y) ∈ R2 : x − y − 2 ≤ 0, x + y + 2h + 2 ≥ 0},
W = {(x, y) ∈ R2 : x − y − 2 ≤ 0, x + y + 2h + 2 ≤ 0},
S = {(x, y) ∈ R2 : x − y − 2 ≥ 0, x + y + 2h + 2 ≤ 0}.
To facilitate the readability of the proof, we plot two examples of E1 and E2 , along with the two
lines and four regions in Figure 8. By the above definitions we know that E1 \ E2 ⊂ N ∪ S, and
47
30
20
30
20
10
N
10
N
W
E
-20
-20
-10
0
E
-10
0
W
-30
-20
-10
0
S
-30
-30
S
10
20
30
-30
-20
-10
0
10
20
30
Figure 8: An example of the two ellipses E1 and E2 with the same values of θ but two different
values of θ0 . The points inside the ellipses represent points from the set S. The point of intersection
of the two lines is the center (−h, −h−2), where for this specific example we have set h = 8, k = 26.
Points “o” form the set E1 ∩ S, and points “x” form the set E2 ∩ S. The four regions E, N , W, S
are also depicted. Points of warmer colors are more likely to occur than points with cooler colors.
E2 \ E1 ⊂ E ∪ W. Note that the pairs (SX , SY ) can only take integer values, and in such a way so
that −k ≤ SX ≤ k, −k ≤ SY ≤ k, and SX ≡ SY ≡ k (mod 2). Define
S = {(x, y) ∈ (N0 )2 | − k ≤ x ≤ k, −k ≤ y ≤ k, x ≡ y ≡ k (mod 2)}
In terms of this notation (6.10) can be restated as
P∅ ((SX , SY ) ∈ E2 ∩ S) ≤ P∅ ((SX , SY ) ∈ E1 ∩ S),
or equivalently
P∅ ((SX , SY ) ∈ (E2 \ E1 ) ∩ S) ≤ P∅ ((SX , SY ) ∈ (E1 \ E2 ) ∩ S).
(D.1)
In the following, our proof plan is to show (D.1) by taking pairs of points from the set (E1 \E2 )∩S
and matching them in a 1-1 manner with pairs of points from the set (E2 \ E1 ) ∩ S, while ensuring
that the former pairs are more likely to occur than the latter. Sometimes one or more of the points
in the two pairs will lie outside of the set S (see, e.g., the right panel of Figure 8), but we will make
sure that this does not alter the conclusion.
Take a point (i, j) ∈ (E1 \E2 )∩S∩N . Define the transformation (bi, b
j) = (−j−2h−2, −i−2h−2),
b
b
b
b
and consider the four points (i, j), (j + 2, i − 2), (j + 2, i − 2), (i, j). Since (i, j) ∈ (E1 \ E2 ) ∩ N it is
simple to check that (j +2, i−2) ∈ (E2 \E1 )∩E, (bi, b
j) ∈ (E2 \E1 )∩W and (b
j +2, bi−2) ∈ (E1 \E2 )∩S.
We will now argue that
P∅ (SX , SY ) ∈ {(i, j), (b
j + 2, bi − 2)} ≥ P∅ (SX , SY ) ∈ {(j + 2, i − 2), (bi, b
j)} .
(D.2)
48
Note that for any point (m, `) ∈ S
P∅ ((SX , SY ) = (m, `)) =
k
k+m
2
−2k
.
k+` 2
k
2
To this end we consider several cases.
i. First assume that all of the above four points belong to S. Therefore, we need to show that
k
k
k
k
k
k
k
k
−
−
≥
k+j
k+j+2
k+b
j
k+b
j+2 .
k+i
k+i−2
k+bi−2
k+bi
2
2
2
2
2
2
2
2
A direct calculation yields that the above is equivalent to
b
j−i
j−bi
(k
+
1)
+
(k
+
1)
2
2 (k + 1) + (k + 1)
.
k−i
k+j
k−j ≥
b
b
k+i
bi k−bi
k+
!
+ 1 ! k+j + 1 ! k−j !
2 ! 2 +1 ! 2 +1 ! 2 !
2
2
2
2
j−i
We note that b
j − bi = j − i, and furthermore the numerator 2 (k + 1) + (k + 1) ≥ 0, since
j − i ≥ −2 by the fact that (i, j) ∈ N . Rearaging and cancelling terms shows that the above
is equivalent to showing that
k+j
k + j + 2h + 2
k+i
k + i + 2h + 2
+ 2 ...
+1
+ 1 ...
≥
2
2
2
2
k − i − 2h − 2
k−i
k − j − 2h − 2
k−j
+ 2 ...
+1
+ 1 ...
.
2
2
2
2
This inequality holds true since (i, j) ∈ N and hence i + j + 2h + 2 ≥ 0.
ii. In the case when (b
j +2, bi−2) ∈ S but either or both of the points (bi, b
j) 6∈ S or (j +2, i−2) 6∈ S
then (D.2) continues to hold by i.
iii. Assume (b
j + 2, bi − 2) 6∈ S but (bi, b
j) ∈ S and (j + 2, i − 2) ∈ S. By the fact that (b
j + 2, bi − 2) 6∈ S
but (bi, b
j) ∈ S we have that either b
j = k or bi = −k. The former is impossible since otherwise
i = −(k + 2h + 2) < −k. Hence the only viable option is to have bi = −k, in which case
j = k − (2h + 2). We need to show that
k! j−i
1
2 (k + 1) + (k + 1)
k−i
k+j
k−j ≥
k−bj .
b
k+i
k+
j
2 ! 2 +1 ! 2 +1 ! 2 !
2 ! 2 !
Since (i, j) ∈ E1 \ E2 we cannot have j = i − 2 (which can only happen in the intersection of
the two ellipses), and thus j ≥ i. Therefore it suffices to show
k+i
k + i + 2h + 2
j−i
+ 1 (k + 1)!
+ 1 ...
2
2
2
k − i − 2h − 2
k−i
k+j
k−j
≥
+ 1 ...
+1
+1 !
!.
2
2
2
2
We first observe that the following holds
k+i
k + i + 2h + 2
k−j
+ 1 ...
≥
!
2
2
2
49
This is true since: first j = k−(2h+2) and hence the RHS factorial contains h+1 terms which
is the same number of terms contained on the LHS; and second by the fact that (i, j) ∈ N
we have i + j + 2h + 2 ≥ 0 and therefore each of the terms in the product on the LHS is
greater than or equal to a corresponding term in the product of the RHS. Next, note that
j−i
k−i−2h−2
+ 1 since j = k − (2h + 2). Hence it is enough to show
2 +1=
2
k − i − 2h − 2
k−i
k+j
(k + 1)! ≥
+ 2 ...
+1
+1 !
2
2
2
The latter is equivalent to
k − i − 2h − 2
k−i
k+j
+ 2 . . . (k + 1) ≥
+ 2 ...
+1 ,
2
2
2
which holds true since the number of terms on both sides is the same, and i ≥ −k.
iv. Assume (b
j + 2, bi − 2) 6∈ S and (bi, b
j) ∈ S, but (j + 2, i − 2) 6∈ S. This case follows by iii.
v. Assume (b
j + 2, bi − 2) 6∈ S and (j + 2, i − 2) ∈ S but (bi, b
j) 6∈ S. This case follows by case i.
since its calculation implies
k
k
k
k
− k+i−2
≥ 0.
k+j
k+j+2
k+i
2
2
2
2
We have now shown that (D.2) holds. The proof will be completed if we showed that by
iterating over points in the set (E1 \ E2 ) ∩ S ∩ N exhausts all points from the remaining three sets
— (E1 \ E2 ) ∩ S ∩ S, (E2 \ E1 ) ∩ S ∩ W and (E2 \ E1 ) ∩ S ∩ E. We show this below.
First assume there is a point (i, j) ∈ (E2 \ E1 ) ∩ S ∩ E so that (j + 2, i − 2) 6∈ (E1 \ E2 ) ∩ S ∩ N .
Since (i, j) ∈ (E2 \ E1 ) ∩ E we necessarily have (j + 2, i − 2) ∈ (E1 \ E2 ) ∩ N . Therefore by our
assumption (j + 2, i − 2) 6∈ S. By the fact that (i, j) ∈ E we know that i ≥ j + 2 and therefore
k ≥ j + 2 ≥ j ≥ −k. Moreover i = −k implies that j ≤ −k − 2 contradicting (i, j) ∈ S, and
therefore i ≥ −k + 2. This implies that (j + 2, i − 2) ∈ S which is a contradiction.
Next, let (i, j) ∈ (E2 \ E1 ) ∩ S ∩ W so that (−j − 2h − 2, −i − 2h − 2) 6∈ (E1 \ E2 ) ∩ S ∩ N .
The fact that (i, j) ∈ (E2 \ E1 ) ∩ W ensures (−j − 2h − 2, −i − 2h − 2) ∈ (E1 \ E2 ) ∩ N and thus
(−j − 2h − 2, −i − 2h − 2) 6∈ S. Since (i, j) ∈ W we have that i ≤ −j − 2h − 2 and j ≤ −i − 2h − 2
hence since (i, j) ∈ S we also have −i − 2h − 2 ≥ −k and −j − 2h − 2 ≥ −k. Also −i − 2h − 2 ≤ k,
as otherwise we would have i ≤ −k − 2h − 2 < −k which would be a contradiction. Similarly
−j − 2h − 2 ≤ k. Thus (−j − 2h − 2, −i − 2h − 2) ∈ S which is a contradiction.
Finally assume that (i, j) ∈ (E1 \ E2 ) ∩ S ∩ S but (−i − 2h, −j − 2h − 4) 6∈ (E1 \ E2 ) ∩ S ∩ N .
Since (i, j) ∈ (E1 \ E2 ) ∩ S it can be checked that (−i − 2h, −j − 2h − 4) ∈ (E1 \ E2 ) ∩ N and thus
by assumption (−i − 2h, −j − 2h − 4) 6∈ S. Note that −i − 2h ≤ k and −j − 2h − 4 ≤ k or otherwise
i < −k − 2h ≤ k and j ≤ −k − 2h − 4 < −k which contradicts our assumption that (i, j) ∈ S.
We also have that −k + 2 ≤ j + 2 ≤ −i − 2h by the assumption that (i, j) ∈ S ∩ S. Moreover
−i − 2h − (−j − 2h − 4) = 4 + j − i ≤ 2 and therefore (−j − 2h − 4) ≥ −k + 2 − 2 = −k. Hence
(−i − 2h, −j − 2h − 4) ∈ S, which is a contradiction.
With this the proof is complete.
50
θ
−θ
Figure 9: The graph G0 is depicted with solid edges. A particular Gj,k can be constructed if we
add the two dashed edges to G0 .
Proof of Proposition 4.3. To compare the null and alternative hypothesis we construct the following
graphs. Let G0 = (Vd , E0 ), where
b ds c−2
E0 :=
[
j=0
d
c−2
[ b s[
{(u, v)}sj+1≤u<v≤s(j+1)
{(sj, sj + 1)}
j=1
[
{(j, j + 1)}.
s(b ds c−1)+1≤j≤d−1
which is a union of b ds c − 2, s-cliques connected via a path, and a disconnected path graph. See
Figure 9 for a visualization.
Based on G0 , we construct the ferromagnetic Ising measure PG0 which attaches an interaction
of magnitude θ > 0 to each edge in G0 so that:
X
PG0 (X) ∝ exp θ
Xu Xv .
(D.3)
(u,v)∈E0
Next construct the connected graphs Gj,k = (Vd , Ej,k ) for 1 ≤ j ≤ b ds c − 2 and 3 ≤ k ≤ s − 1, where
[ d
d
Ej,k := E0
.
s(b c − 1) + j, sj + 2 , s(b c − 1) + j, sj + k
s
s
We define the Ising measure associated with PGj,k by
X
PGj,k (X) ∝ exp θ
Xu Xv + θXs(b d c−1)+j (Xsj+2 − Xsj+k ) ,
s
(D.4)
(u,v)∈E0
which clearly belongs to model class (4.1) but not to (3.1). For two alternative graphs Gj1 ,k1 and
Gj2 ,k2 we evaluate the ratio:
EG0
PGj1 ,k1 PGj2 ,k2
PG0
PG0
=
EGj1 ,k1 exp(θXj̄2 (Xsj2 +2 − Xsj2 +k2 ))
EG0 exp(θXj̄2 (Xsj2 +2 − Xsj2 +k2 ))
where for brevity we set j̄2 := s(b ds c − 1) + j2 . Using identity (A.1) we have:
EG0
PGj1 ,k1 PGj2 ,k2
PG0
=1+
PG0
=
EGj1 ,k1 (1 + tanh(θ)Xj̄2 Xsj2 +2 )(1 − tanh(θ)Xj̄2 Xsj2 +k2 )
EG0 (1 + tanh(θ)Xj̄2 Xsj2 +2 )(1 − tanh(θ)Xj̄2 Xsj2 +k2 )
tanh(θ)EGj1 ,k1 [Xj̄2 Xsj2 +2 − Xj̄2 Xsj2 +k2 ] + tanh2 (θ)[EG0 Xsj2 +2 Xsj2 +k2 − EGj1 ,k1 Xsj2 +2 Xsj2 +k2 ]
1 − tanh2 (θ)EG0 Xsj2 +2 Xsj2 +k2
51
.
where we used EG0 Xj̄2 Xsj2 +k2 = EG0 Xj̄2 Xsj2 +2 = 0. We distinguish two cases.
Case I. First assume that j1 6= j2 . Since there are no simple paths connecting sj2 + 2 with sj2 + k2
which pass through the vertex j̄1 = s(b ds c − 1) + j1 , by Proposition A.3 we have that:
EG0 Xsj2 +2 Xsj2 +k2 = EGj1 ,k1 Xsj2 +2 Xsj2 +k2 .
What is more, due to the symmetry of the construction we also have that:
EGj1 ,k1 Xj̄2 Xsj2 +2 = EGj1 ,k1 Xj̄2 Xsj2 +k2 .
Hence when j1 6= j2 we have:
EG0
PGj1 ,k1 PGj2 ,k2
PG0
PG0
= 1.
Case II. Now consider the case j1 = j2 =: j. Below we control the terms EGj,k1 [Xj̄ Xsj+2 −
Xj̄ Xsj+k2 ] (here j̄ = s(b ds c − 1) + j) and EG0 Xsj+2 Xsj+k2 − EGj,k1 Xsj+2 Xsj+k2 , starting from the
former. The simple identity |a − b| = 1 − ab for any a, b ∈ {±1} yields the estimate
EGj,k1 [Xj̄ Xsj+2 − Xj̄ Xsj+k2 ] ≤ 1 − EGj,k1 Xsj+2 Xsj+k2 .
We have
PGj,k1 (Xsj+2 Xsj+k2 = 1)
PGj,k1 (Xsj+2 Xsj+k2 = −1)
≥ min
ξ∈{±1}
≥
PGj,k1 (Xsj+2 Xsj+k2 = 1|Xj̄ = ξ)
PGj,k1 (Xsj+2 Xsj+k2 = −1|Xj̄ = ξ)
PG0 (Xsj+2 Xsj+k2 = 1)
exp(2θbs/4c)
,
≥
exp(4θ)PG0 (Xsj+2 Xsj+k2 = −1)
exp(4θ)(2bs/4c − 1)
where the last inequality follows by (B.7), since the vertices sj + 2 and sj + k2 belong to the same sclique in the graph G0 , and hence they are on the same side of a 2bs/4c × 2bs/4c-biclique. Putting
the last two inequalities together with the identity EGj,k1 Xsj+2 Xsj+k2 = PGj,k1 (Xsj+2 Xsj+k2 =
1) − PGj,k1 (Xsj+2 Xsj+k2 = −1) gives:
EGj,k1 [Xj̄ Xsj+2 − Xj̄ Xsj+k2 ] ≤ 1 − EGj,k1 Xsj+2 Xsj+k2 ≤
2(2bs/4c − 1)
.
exp((2bs/4c − 4)θ) + (2bs/4c − 1)
Now we focus on the term EG0 Xsj+2 Xsj+k2 − EGj,k1 Xsj+2 Xsj+k2 . Note that
EG0 Xsj+2 Xsj+k2 − EGj,k1 Xsj+2 Xsj+k2 ≤ 1 − EGj,k1 Xsj+2 Xsj+k2 ,
and therefore the previous bound applies. We have established
EG0
PGj,k1 PGj,k2
PG0
PG0
tanh(θ) + tanh2 (θ)
2(2bs/4c − 1)
2
exp((2bs/4c − 4)θ) + (2bs/4c − 1)
1 − tanh (θ)
4(2bs/4c − 1)
≤1+
.
(1 − tanh2 (θ))(exp((2bs/4c − 4)θ) + (2bs/4c − 1))
≤1+
52
Let P
⊗n
be the measure such that P
⊗n
=
1
(b ds c−2)(s−3)
P
j,k
P⊗n
Gj,k . We have:
⊗n
P⊗n
X
Gj1 ,k1 PGj2 ,k2
1
Dχ2 (P
=
E ⊗n
P⊗n
[(b ds c − 2)(s − 3)]2 j ,j ,k ,k PG0 P⊗n
G0
G0
1 2 1 2
n
4(2bs/4c−1)
1 + (1−tanh2 (θ))(exp((2bs/4c−4)θ)+(2bs/4c−1))
(b ds c − 2)2 − (b ds c − 2)
≤
+
(b ds c − 2)2
(b ds c − 2)(s − 3)2
4(2bs/4c−1)
exp n (1−tanh2 (θ))(exp((2bs/4c−4)θ)+(2bs/4c−1)) − (s − 3)2
≤1+
.
(b ds c − 2)(s − 3)2
⊗n
, P⊗n
G0 )
Hence if the last quantity is 1 + o(1) the risk is 1. The latter is ensured under condition (4.4) from
the statement, and hence the proof is complete.
E
Correlation Testing for General Models
Proof of Proposition 5.1. First consider the case when the true graph θ ∈ S0 (θ, Θ). Then by (5.3)
we have that
b u Xv − E eXu Xv | ≤ ε(δ),
max |EX
θ
u,v
with probability at least 1 − δ. Next, suppose that θ ∈ S1 (θ, Θ). Let
(u∗ , v ∗ ) = argmax |EθeXu Xv − Eθ Xu Xv |
u,v
We have
b u Xv − E eXu Xv | ≥ |EX
b u∗ Xv∗ − E eXu∗ Xv∗ |
max |EX
θ
θ
u,v
b u∗ Xv∗ − Eθ Xu∗ Xv∗ |
≥ |EθeXu∗ Xv∗ − Eθ Xu∗ Xv∗ | − |EX
≥ |EθeXu∗ Xv∗ − Eθ Xu∗ Xv∗ | − τ
≥
sinh2 (θ/4)
− τ,
2sΘ(3 exp(2sΘ) + 1)
where the next to last inequality holds with probability at least 1 − δ by Lemma C.1, and the last
inequality holds by Lemma 6 of Santhanam and Wainwright (2012) since θe ∈ S0 (θ, Θ) and hence
e This completes the proof.
G(θ) 6= G(θ).
Proof of Lemma 5.2. The proof is a direct consequence of Lemma C.1.
Proof of Proposition 5.3. Clearly the output of Algorithm 6 satisfies θe ∈ S0 (θ, Θ) and thus for
e we have P(Te) = 0 by design. Hence it remains to argue that (5.3) holds with the
Te = G(θ)
specified ε(δ). Using Lemma C.1 we have that
b u Xv − Eθ Xu Xv | ≤ τ.
max |EX
u,v
53
(E.1)
with probability at least 1 − δ. Below we assume this event holds, so that all statements we make
should be understood to hold with probability at least 1 − δ.
We begin by arguing that Te = G(θ). First we will show that the set E(Te) \ E(θ) = ∅. Suppose
the contrary, and consider the first step, say m, at which an edge e = (u, v) from the set E(Te)\E(θ)
was added to Te. Let P (possibly P = ∅) denote the path connecting u with v in the graph G(θ).
b u Xv | ≤ τ and since tanh θ − τ ≥ τ this edge is
First suppose P = ∅. Then Eθ Xu Xv = 0, so |EX
pruned at the end of the procedure which is a contradiction. So P 6= ∅ and thus |P | ≥ 2 (if P was
a single edge it follows that e ∈ E(θ) which is a contradiction). By Proposition A.3 and Lemma
A.4 we know
Y
Eθ Xk X` ,
Eθ Xu Xv =
(k,`)∈P
th
b u Xv | ≤
and therefore |EX
(k,`)∈P |Eθ Xk X` | + τ . Since we have added e at the m step of building
the MST, and e 6∈ P , then there exists an edge from e0 ∈ P , say e0 = (u0 , v 0 ), which does not belong
to Te, since otherwise adding e would make a cycle in Te. Adding edge e over e0 means that the
b u0 Xv0 | ≤ |EX
b u Xv | holds. Hence
inequality |EX
Y
b u0 Xv0 | ≤ |EX
b u Xv | ≤
|Eθ Xu0 Xv0 | − τ ≤ |EX
|Eθ Xk X` | + τ ≤ |Eθ Xu0 Xv0 | tanh(Θ) + τ.
Q
(k,`)∈P
However, |Eθ Xu0 Xv0 | ≥ tanh(θ) and since tanh(θ)(1−tanh(Θ)) > 2τ this contradicts the inequality
in the preceding display. We conclude that E(Te) ⊆ E(θ). If there exists an edge e ∈ E(θ) \ E(Te),
then when building the MST this edge was either pruned at the end of the procedure, or a no-edge
was preferred to it. However these two scenarios are also impossible using the same logic as before.
Hence Te = G(θ).
b u Xv , or
By our construction we have that for (u, v) ∈ E(θ) we either have EθeXu Xv = EX
b u Xv | > tanh(Θ) or |EX
b u Xv | < tanh(θ) in which case |E eXu Xv | = tanh(Θ) or |E eXu Xv | =
|EX
θ
θ
tanh(θ) respectively. Putting these two cases together we conclude that
max
(u,v)∈E(θ)
b u Xv − E eXu Xv | ≤ τ.
|EX
θ
b u Xv − E eXu Xv | ≤ τ .
Clearly if two vertices u and v are disconnected in G(θ), we also have |EX
θ
Next take two connected vertices u, v and let P denote the (unique) path between them. Consider
Y
b u Xv −
b u Xv − E eXu Xv | = EX
|EX
EθeXk X`
θ
(k,`)∈P
b u Xv − Eθ Xu Xv | + |
≤ EX
≤τ +|
Y
(k,`)∈P
Eθ Xk X` −
Y
(k,`)∈P
Y
Eθ Xk X` −
Y
EθeXk X`
(k,`)∈P
EθeXk X` .
(k,`)∈P
Suppose that the length of P is l. For brevity put τuv := Eθ Xu Xv − EθeXu Xv . By (E.1), and the
b u Xv to E eXu Xv will only result in better estimation of Eθ Xu Xv we
fact that our corrections of EX
θ
know that |τuv | ≤ τ for all (u, v) ∈ E(θ) = E(Te).
54
Note that condition tanh(θ)(1 − tanh(Θ)) > 2τ implies that tanh(Θ) < 1 and 1 > 2τ + tanh(Θ).
A simple expansion gives
Y
Y
Y
Y
EθeXk X` =
EθeXk X`
Eθ Xk X` −
(EθeXk X` + τk` ) −
|
(k,`)∈P
(k,`)∈P
(k,`)∈P
(k,`)∈P
≤
X
Y
S∪R=P
(k,`)∈S
S∩R=∅,|S|<l
|EθeXk X` |
Y
(k,`)∈R
|τk` |
≤ (τ + tanh(Θ))l − tanh(Θ)l
≤
≤
l
X
τ (τ + tanh(Θ))l−j tanh(Θ)j−1
j=1
τ
,
1 − tanh(Θ)
where the next to last inequality follows easily by induction in l, and the last inequality is due to
the fact that τ + tanh(Θ) ≤ 1 and tanh(Θ) < 1. Hence we have successfully argued that
b u Xv − E eXu Xv | ≤ τ 2 − tanh(Θ) ,
max |EX
θ
u,v∈Vd
1 − tanh(Θ)
which completes the proof.
55
| 10math.ST
|
arXiv:1501.01307v2 [math.NT] 17 Aug 2016
Integrality in the Steinberg module and
the top-dimensional cohomology of GLn OK
Thomas Church, Benson Farb, and Andrew Putman∗
August 18, 2016
Abstract
We prove a new structural result for the spherical Tits building attached to GLn K
for many number fields K, and more generally for the fraction fields of many Dedekind
domains O: the Steinberg module Stn (K) is generated by integral apartments if and only
if the ideal class group cl(O) is trivial. We deduce this integrality by proving that the
complex of partial bases of On is Cohen–Macaulay. We apply this to prove new vanishing
and nonvanishing results for Hνn (GLn OK ; Q), where OK is the ring of integers in a number field and νn is the virtual cohomological dimension of GLn OK . The (non)vanishing
depends on the (non)triviality of the class group of OK . We also obtain a vanishing theorem for the cohomology Hνn (GLn OK ; V ) with twisted coefficients V . The same results
hold for SLn OK as well.
1
Introduction
1.1
The Tits building and the Steinberg module
One of the most fundamental geometric objects attached to the general linear group GLn K
over a field K is its associated Tits building, denoted Tn (K). The space Tn (K) is the (n − 2)dimensional simplicial complex whose p-simplices are flags of subspaces
0 ( V0 ( · · · ( Vp ( K n .
The group GLn K acts on Tn (K) by simplicial automorphisms. By the Solomon–Tits theorem
[So], the space Tn (K) is homotopy equivalent to a wedge of (n − 2)-dimensional spheres, so
it has only one interesting homology group.
Definition 1.1 (The Steinberg module). The Steinberg module Stn (K) is the GLn Ke n−2 (Tn (K); Z).
module given by the top homology group H
The Solomon–Tits theorem also gives a generating set for Stn (K). A frame for K n is
a set L = {L1 , . . . , Ln } of lines in K n such that K n = L1 ⊕ · · · ⊕ Ln . The apartment AL
corresponding to L is the full subcomplex of Tn (K) on the 2n − 2 subspaces span(Li | i ∈ I)
for ∅ ( I ( {1, . . . , n}. The apartment AL is homeomorphic to an (n − 2)-sphere, so its
fundamental class determines an apartment class
e n−2 (Tn (K); Z) = Stn (K)
[AL ] ∈ H
which is defined up to ±1. The Solomon–Tits theorem states that Stn (K) is generated by
the set of apartment classes.
∗
TC was supported by NSF grants DMS-1103807 and DMS-1350138 and the Alfred P. Sloan Foundation;
BF was supported by NSF grant DMS-1105643; AP was supported by NSF grant DMS-1255350 and the Alfred
P. Sloan Foundation.
1
1.2
Integrality and non-integrality
Consider a Dedekind domain O with field of fractions K. While the action of GLn K on Tn (K)
is transitive, the action of GLn O is usually not, and indeed this action encodes arithmetic
information about O. For example, the number of orbits of the GL2 O-action on T2 (K) is
the class number |cl(O)| of O. In this context, we have the following natural notion.
Definition 1.2 (Integral apartment). Let O be a Dedekind domain with field of fractions
K. A frame L = {L1 , . . . , Ln } of K n is integral, and AL is an integral apartment, if
On = (L1 ∩ On ) ⊕ · · · ⊕ (Ln ∩ On ).
In the case O = Z, and more generally when O is Euclidean, Ash–Rudolph [AR, Theorem 4.1] proved that Stn (K) is generated by the fundamental classes of integral apartments.
Our first theorem extends this result of Ash–Rudolph to a wide class of Dedekind domains
of arithmetic type, whose definition we recall.
Definition 1.3 (Dedekind domain of arithmetic type). Let K be a global field, i.e.
either a number field ([K : Q] < ∞) or a function field in one variable over a finite field
([K : Fq (T )] < ∞). Let S be a finite nonempty set of places of K; if K is a number field,
assume that S contains all infinite places. The ring of S-integers
OS := {x ∈ K | ordp (x) ≥ 0 for all p 6∈ S}
is a Dedekind domain, and we say that OS is a Dedekind domain of arithmetic type.
Theorem A (Integrality Theorem). Let O be a Dedekind domain with |cl(O)| = 1 and
field of fractions K. Assume either that O is Euclidean, or that O = OS is a Dedekind
domain of arithmetic type such that |S| > 1 and S contains a non-imaginary place. Then
Stn (K) is spanned by integral apartment classes.
Under the conditions of Theorem A, the set of integral apartments in Tn (K) is precisely
the GLn O-orbit of a single standard apartment. Thus Theorem A implies that Stn (K), which
is obviously cyclic as a GLn K-module, is in fact cyclic as a GLn O-module.
Remark 1.4. Ash–Rudolph’s proof [AR] of Theorem A when O is Euclidean is based on a
beautiful generalization of the method of continued fractions to higher dimensions. Using the
Euclidean function on O as a measure of “complexity”, they give an algorithm to write a nonintegral apartment class as a sum of integral apartment classes. Our proof is quite different,
even in the special case that O is Euclidean: non-integral apartments never actually show up
in our proof, and we do not even make use of the fact that Stn (K) is generated by apartment
classes.
By Dirichlet’s theorem, the group of units OS× has rank |S| − 1, so the assumption |S| > 1
is equivalent to OS× = ∞. The assumptions on S rule out only two families of Dedekind
domains of arithmetic type, for which Theorem A need not hold (see Remark 4.1 below):
• the ring of integers OK in a totally imaginary number field K;
• those OS in positive characteristic with finitely many units, such as Fq [T ].
The assumption in Theorem A that |cl(OS )| = 1 obviously excludes many more examples.
However, this assumption is necessary in a strong sense, as the following theorem shows.
Theorem B (Non-Integrality Theorem). Let O be a Dedekind domain with field of fractions K. If 1 < |cl(O)| < ∞ and n ≥ 2 then Stn (K) is not generated by integral apartment
classes.
Theorem B applies in particular to all number rings with nontrivial class group.
2
1.3
The top-degree cohomology of GLn OK and SLn OK
When OK is the ring of integers in a number field K, the Steinberg module Stn (K) is directly
connected to the cohomology of GLn OK and SLn OK , as we now explain.
Let r1 (resp. 2r2 ) be the number of real (resp. complex) embeddings of OK . Then
OK ⊗ R ∼
= Rr1 ⊕ Cr2 and GLn (OK ⊗ R) ∼
= (GLn R)r1 × (GLn C)r2 . The quotient MK of
the associated symmetric space by the discrete group GLn OK is a Riemannian orbifold with
H∗ (MK ; Q) ∼
= H ∗ (GLn OK ; Q).
The computation of these cohomology groups is a fundamental problem in group theory,
topology and number theory. The existence of MK implies that Hi (GLn OK ; Q) = 0 for
i > dim MK . But since MK is not compact, it is not even clear what the largest i for which
Hi (GLn OK ; Q) 6= 0 is. The first progress on this basic problem came in 1973 by Borel and
Serre [BS]. Recall that the virtual cohomological dimension of a virtually torsion-free group
Γ is
vcd(Γ) := max{k | Hk (Γ; V ⊗ Q) 6= 0 for some Γ-module V }.
By constructing and analyzing a compactification
of the orbifold MK , Borel–Serre proved that
2 − n. In particular Hk (GL O ; Q) = 0 for
vcd(GLn OK ) is equal to νn := r1 n+1
+
r
·
n
2
n K
2
k > νn . The coefficient module V used by Borel–Serre to certify that Hνn (GLn OK ; V ) 6= 0 is
infinite-dimensional. This left open the question as to whether or not GLn OK has untwisted
rational cohomology in the top dimension νn .
In 1976 Lee–Szczarba [LS, Theorem 1.3] answered this question in the special case that
OK is Euclidean, proving that Hνn (GLn OK ; Q) = 0 for these OK . Except for some explicit
computations in low-dimensional cases (sometimes with the help of computers), there seems
to have been no progress on this question since Lee–Szczarba.
Our next theorem extends the theorem of Lee–Szczarba from Euclidean number rings to
many number rings with class number 1 (we remark that even in the Euclidean case our proof
is different from that of Lee–Szczarba). In addition to untwisted coefficients Q, it also gives
a result for twisted coefficient systems arising from rational representations of the algebraic
group GLn . Let νn′ := vcd(SLn OK ); an explicit formula for νn′ is given in §4 below. For
λ = (λ1 , . . . , λn ) ∈ Zn with λ1 ≥ · · ·P
≥ λn , let Vλ be the rational GLn K-representation with
highest weight λ, and define kλk = ni=1 (λi − λn ).
Theorem C (Vanishing Theorem). Let OK be the ring of integers in an algebraic number
field K with |cl(OK )| = 1. Suppose that K has a real embedding or that OK is Euclidean.
Then
′
Hνn (SLn OK ; Vλ ) = Hνn (GLn OK ; Vλ ) = 0
′
for n ≥ 2 + kλk. In particular, Hνn (SLn OK ; Q) = Hνn (GLn OK ; Q) = 0 for n ≥ 2.
For twisted coefficients, Theorem C seems to be new even in the case when OK is Euclidean. As we discuss in Remark 4.1 below, the assumption that K has a real embedding is
necessary if OK is not Euclidean.
In contrast to Theorem C, a different phenomenon occurs when |cl(OK )| =
6 1.
Theorem D (Non-Vanishing Theorem). Let OK be the ring of integers in an algebraic
number field K. Then for n ≥ 2,
′
dim Hνn (SLn OK ; Q) ≥ dim Hνn (GLn OK ; Q) ≥ (|cl(OK )| − 1)n−1 .
Theorem D is classical for n = 2; see Example 5.6. For n ≥ 3 the result and method of
proof are new. Unlike Theorem C, there is no version of Theorem D for twisted coefficients.
For example, if K is any field of characteristic 0 then Hi (GLn K; K n ) = 0 for all i ≥ 0.
3
Remark 1.5. The cohomology of arithmetic groups such as SLn OK has been fruitfully
studied via the theory of automorphic forms, see e.g. [F1, FS]. However, it can be difficult
to extract concrete statements from them. For example, it is not clear how to see from [F1]
even Borel–Serre’s result that the rational cohomology of SLn OK vanishes above its vcd, let
alone the sharper vanishing of Theorem C.
Franke [F2] has also given a more practicable description of the Hecke-trivial
subspace
√
∗
of H (SLn OK ; Q). For example, for a quadratic imaginary field K = Q( −d), it can be
deduced from [F2, Theorem 9] that H i (SLn Ok ; Q) contains no Hecke-trivial cohomology for
i > νn′ − n + 1. But as Theorem D shows, this cannot capture all the cohomology; indeed,
the Hecke action on the classes we construct in Theorem D will be twisted by nontrivial ideal
class characters.
1.4
Relationship between Integrality and Vanishing/Non-Vanishing
Theorems C (Vanishing) and D (Non-Vanishing) relate to the Integrality Theorem (Theorem A) as follows. While GLn OK does not satisfy Poincaré duality, Borel–Serre [BS] proved
that it does satisfy Bieri–Eckmann duality with rational dualizing module Stn (K). This
means that for any Q GLn OK -module V and any i ≥ 0 we have
Hνn −i GLn OK ; V ∼
= Hi GLn OK ; Stn (K) ⊗ V .
This implies in particular that
Hνn (GLn OK ; Q) ∼
= H0 (GLn OK ; Stn (K) ⊗ Q) ∼
= (Stn (K) ⊗ Q)GLn OK .
(1.1)
Equation (1.1) converts the problem of computing Hνn (GLn OK ; Q) to the problem of understanding the GLn OK -action on Stn (K). This understanding is precisely what is given by
the Integrality Theorem, which is applied directly to prove Theorem C.
As for Theorem D, the projection Tn (K) → Tn (K)/ GLn OK induces a map
e n−2 (Tn (K)/ GLn OK ; Z) ∼
Stn (K) → H
= ZN
where N = (|cl(OK )| − 1)n−1 .
By Equation (1.1), this map rationally factors through Hνn (GLn OK ; Q). It turns out that
Tn (K)/ GLn OK is also a spherical building (see Proposition 5.4 below). The hard part
in proving Theorem D is showing that this map is surjective. This is difficult because the
apartments of Tn (K) that generate Stn (K) are made up of n! simplices, while the apartments
of Tn (K)/ GLn OK have only 2n−1 simplices. Since n! ≫ 2n−1 , most apartments of Tn (K) will
be draped over a huge number of apartments in the quotient, and it is difficult to describe the
resulting homology class. However, we show that by making careful choices, each apartment
of Tn (K)/ GLn OK can be lifted to a certain special apartment of Tn (K) for which all of the
remaining n! − 2n−1 simplices exactly cancel each other out, leaving the desired homology
class. This requires a delicate combinatorial argument.
1.5
The complex Bn (O) of partial bases
The first step in our proof of Theorem A is to consider a certain “integral model” Bn (O) for
the Tits building Tn (K). The complex of partial bases Bn (O) is the simplicial complex whose
maximal simplices are bases of On . Letting Cn−1 (Bn (O)) be the (n − 1)-chains of Bn (O), we
define in §3 an integral apartment class map
φ : Cn−1 (Bn (O)) → Stn (K)
4
sending a basis for On to the integral apartment class determined by the corresponding frame
of K n . Theorem A is the assertion that φ is surjective. We prove that φ is surjective as long
as Bn (O) is as highly connected as possible. Our last main theorem is that this connectivity
does indeed hold.
A d-dimensional complex is d-spherical if it is (d − 1)-connected, in which case it is
homotopy equivalent to a wedge of d-spheres. A simplicial complex X is Cohen–Macaulay
(abbreviated CM) of dimension d if:
1. X is d-spherical, and
2. for every k-simplex σ k of X, the link LinkX (σ k ) is (d − k − 1)-spherical.
The following theorem is the main technical result of this paper, and is of independent
interest.
Theorem E (Bn (OS ) is Cohen–Macaulay). Let OS be a Dedekind domain of arithmetic
type. Assume that |S| > 1 and that S contains a non-imaginary place. Then Bn (OS ) is CM
of dimension n − 1 for all n ≥ 1.
Note that, unlike our other results, in Theorem E we make no assumption about the
class group cl(OS ). Theorem E applies to any number ring OK possessing a real embedding
OK ֒→ R (with the exception of Z), as well as to every OS with OS× = ∞ in which some
prime p is invertible. Maazen [Ma] proved earlier that for any Euclidean domain O, the
complex of partial bases Bn (O) is CM of dimension n − 1.
The proof of Theorem E is given in §2. It is a complicated inductive argument on the rank
n. To make the induction work, we need to prove that a more general class of “complexes of
I-bases” relative to an ideal I ⊂ O are CM. It is in proving this stronger statement in the
base case n = 2 that the specific arithmetic hypotheses on OS are used.
Remark 1.6. Resolving a conjecture of Quillen, Van der Kallen [vdK, Theorem 2.6] proved
that if a ring R satisfies Bass’s stable range condition SRd , then Bn (R) is (n − d)-connected.
Any Dedekind domain O satisfies Bass’s stable range condition SR3 , so by Van der Kallen’s
result it was known that Bn (O) is (n − 3)-connected. However, Theorem E includes the
stronger assertion that Bn (O) is (n − 2)-connected; therefore to prove Theorem E requires us
to go beyond these known results and show that Bn (O) is as highly connected as if O were
a field or local ring!
Our proof of Theorem E uses a stronger condition that can be thought of as “SR2.5 ”
which Reiner proved holds for all Dedekind domains. One might hope that Theorem E could
be proved by an argument mimicking [vdK] by simply substituting this “SR2.5 ” condition in
place of SR3 . But this cannot be so, because the story is more complicated: in Theorem 2.1
below we show that Theorem E is not true for all Dedekind domains, nor even those of
arithmetic type. New ideas are needed. See §2.1 for a high-level discussion of the difficulties
that arise in extending Van der Kallen’s results to this situation, and an explanation of where
the hypotheses of Theorem E come from.
Outline of paper. In §2 we prove that Bn (O) is CM for appropriate O (Theorem E). We
apply this result in §3 to the integral apartment class map to prove the Integrality Theorem
(Theorem A). This result is applied in §4 to prove the Vanishing Theorem (Theorem C) for the
top degree cohomology of SLn OK and GLn OK . The Non-Vanishing Theorem (Theorem D)
and the Non-Integrality Theorem (Theorem B) are proved in §5.
Acknowledgments. We are grateful to an anonymous referee for their careful reading,
which improved this paper greatly, and in particular was invaluable for the overview in §2.1.
5
Bn(OS ) is CM of dimension n − 1
2
In this section we prove Theorem E. We also prove the following contrasting result.
Theorem 2.1. Let OS belong to one of the families of Dedekind domains of arithmetic type:
√
1. K = Q( d) and OS = OK for d < 0 squarefree with d 6∈ {−1, −2, −3, −7, −11}.
2. [K : Fq (T )] < ∞ with q odd, S = {degf }, and either |cl(OS )| =
6 1 or deg f ≥ 2.
Then the (n − 1)-dimensional complex Bn (OS ) is not Cohen–Macaulay for any n ≥ 2.
In particular, Theorem 2.1 implies that B2 (OS ) is not connected,
so Van der Kallen’s
√
√
1+ −43
result is sharp. For example, neither B2 (Z[√ −5]) nor B2 (Z[ 2 ]) is connected, even
though both are Dedekind domains and Z[ 1+ 2−43 ] is even a PID. Similarly, for K = F3 (T )
√
√
and f = degT 2 +1 we have Of ∼
= F3 [X, X − X 2 ], so B2 (F3 [X, X − X 2 ]) is not connected.
Theorem 2.1 is proved in §2.4.
Outline. We begin in §2.1 with a high-level overview of the proof of Theorem E; this overview
is separate from the actual proof, and can be skipped if desired, but for both experts and nonexperts it should be useful in understanding the structure of the proof. In §2.2 we summarize
basic properties of Dedekind domains and Cohen–Macaulay complexes.
We prove Theorem E by induction, but to do this we must strengthen the inductive
hypothesis to a stronger result (Theorem E′ ) applying to a more general family of simplicial
complexes; we introduce these in §2.3. The body of the proof then occupies §2.4, §2.5, §2.6,
§2.7, and §2.8.
2.1
A bird’s-eye view of the proof of Theorem E
In this subsection we give an overview of our approach to Theorem E, including a comparison
with related classical results and an explanation of the differences that arise in our situation.
This overview is written at a higher level than the proof itself, which is given in §2.3, §2.4,
§2.5, §2.6, §2.7, and §2.8. It may be especially useful for experts who are already familiar
with past results on complexes of partial bases, such as the work of Quillen–Wagoner [Wa]
or Van der Kallen [vdK]. Nevertheless, the overview should be useful for any reader in
understanding the big-picture structure of the argument. We have kept the notation in this
section consistent with the remainder of the paper. For simplicity, some of the definitions in
this subsection may be missing or imprecise; precise definitions of all terms used are given in
the proofs themselves.
Notation. All the connectivity bounds we obtain or desire in this overview are of the same
form: if a complex is (N + 1)-dimensional, our goal is always that it is N -connected, or in
other words that it is spherical. As a result, in this outline there is no need to keep track of
the precise connectivity or dimensions that arise; instead, the reader is encouraged to think
of “spherical” as meaning “as highly connected as appropriate”, or simply “nice”. To foster
this, by an abuse of notation we will similarly say (in this section only) that the link of a
simplex, or a map between complexes, is spherical if it has the appropriate connectivity.1
This will allow us to describe the structure of the argument without getting bogged down in
numerics.
1
The reader will have no need of the precise numerics, but for completeness: we say a link of a (k−1)-simplex
in an (N +1)-dimensional complex is spherical if it is (N −k)-connected, and a map between (N +1)-dimensional
complexes is spherical if the map is N -connected.
6
Approach to Theorem E. Our main goal is to prove by induction on n that the complex
of partial bases Bn (O) is spherical. We will defer the discussion of the base case until later,
since certain details would be misleading at this point.
As a stepping-stone, we will make use of the full subcomplex Bn (0) on vectors whose last
coordinate is either 0 or 1. The inductive hypothesis implies straightforwardly that Bn (0) is
spherical (by comparing with the full subcomplex on vectors whose last coordinate is 0, which
is isomorphic to Bn−1 (O)). Therefore it suffices to show that the inclusion of the subcomplex
Bn (0) → Bn (O) is spherical.
Classical approach. It is sometimes possible to verify the connectivity of such an inclusion
“simplex-by-simplex”, in the following way. Say that the 0-link of a partial basis {v1 , . . . , vk }
in Bn (O) is its link inside the subcomplex Bn (0), i.e. the intersection of its link with this
subcomplex. If one could verify that the 0-link of each partial basis {v1 , . . . , vk } in Bn (O) is
spherical, it would follow that the injection itself is spherical.
We point out that the full link in Bn (O) of a partial basis {v1 , . . . , vk } is not so hard to
understand. The quotient map from On onto On /hv1 , . . . , vk i ∼
= On−k induces a map from
the link of {v1 , . . . , vk } to Bn−k (O). This map is not an isomorphism, but it “fibers” the
former over the latter in a convenient way (cf. Def. 2.17), from which one can show that
the former is spherical if and only if the latter is. In particular, we know by induction that
all links in Bn (O) are spherical; however the question remains of whether the 0-links are
spherical.
In the case of a local ring or semi-local ring, where the result was proved by Quillen [Wa,
Proposition 1] and Van der Kallen [vdK, Theorem 2.6] respectively, this approach does indeed
work: by relating with Bn−k (O), one can show that the 0-link of every partial basis is
spherical, so the desired connectivity follows. However, for Dedekind domains this is simply
false. There are two distinct problems that arise.
Problem 1: some partial bases are bad. The first problem is that even in the simplest
cases, there are some partial bases for which there is no hope that their 0-links will be
spherical. Consider for example the vertex v = (2, 5) of B2 (Z): to be spherical its 0-link
would need to be (−1)-connected, i.e. nonempty, but it is easy to check that its 0-link is
empty (the key property is that 2 ∈
/ Z× + 5Z).
To get around this, we single out a special class of partial bases whose 0-link has a better
chance of having the appropriate connectivity: we say that a partial basis is good if it is
contained in some basis where some vector has last coordinate equal to 1 (Def. 2.10). We
warn the reader that the “good subcomplex” formed by these good partial bases is definitely
not a full subcomplex (Prop. 2.14).
Problem 2: strengthening the inductive hypothesis. The other problem, even after
passing to this good subcomplex, is that our inductive hypothesis is no longer strong enough.
The issue comes from the restriction on the last coordinate. We saw above that the link of
{v1 , . . . , vk } fibers over Bn−k (O). We are now interested in the 0-link, which is the subcomplex
including only those vectors whose last coordinate is 0 or 1. This subcomplex no longer fibers
over Bn−k (O); most simplices in Bn−k (O) are inaccessible by such vectors. Indeed, if J ⊂ O
is the ideal generated by the last coordinates of v1 , . . . , vk , the only vectors in Bn−k (O) in
the image of the 0-link are those whose last coordinate is congruent to 0 or 1 modulo J.
To deal with this, we introduce a new complex Bm (I) defined relative to an ideal I ⊂ O,
consisting only of vectors whose last coordinate is ≡ 0 or 1 mod I (with some additional
conditions, see Def. 2.4). To continue the induction, we are forced to go back and strengthen
7
our inductive hypothesis to the stronger claim (Thm E′ ) that Bm (I) is spherical for every
ideal I ⊂ O.
In most cases this resolves the issue: the 0-link in Bn (O) fibers over Bn−k (J), so the
strengthened inductive hypothesis gives us the necessary connectivity.
However there is a further problem in one case. If J = O (a case that one might have
expected to be easier ), although the 0-link does map to Bn−k (O), it is not fibered in the same
convenient way. To handle this, we are forced to introduce a weaker notion of “fibered relative
to a core” (Def. 2.17); although this condition is quite technical, it seems to be necessary
to push the induction through in this case. We point out that this step is one of the main
reasons that in Thm E/E′ we have to prove that Bn (O) is Cohen–Macaulay, and not just
highly connected (cf. Remark 2.22).
Returning to the base case. Let us return to the discussion of the base case, which we set
aside earlier. The key base case for us is n = 2, in which case Thm E′ states that the graph
B2 (I) should be connected for any ideal I. The connectivity of this graph is directly related
to the question of elementary generation for congruence subgroups of SL2 (O) relative to the
ideal I (Prop. 2.7). Therefore we can use the work of Vaserstein–Liehl and Bass–Milnor-Serre
on relative elementary generation to show that B2 (I) is indeed connected; our conditions on
O come directly from the hypotheses of Vaserstein–Liehl and Bass–Milnor-Serre.
We can now see why we deferred the base case until the end: if we had discussed it before
introducing the complexes B2 (I) and Thm E′ , it would have seemed that the base case of
Thm E needs only the connectivity of B2 (O). But the connectivity of B2 (O) only needs
elementary generation for SL2 (O) itself, rather than for its congruence subgroups; the former
is a much easier condition, and holds for a broader class of Dedekind domains. In particular,
one could not see from this perspective where the restrictions on O in Thm E come from.
The n = 2 base case also plays another completely separate role in our proof: it turns
out to be key in the first part of the inductive step, letting us show that we really can reduce
to the “good subcomplex” discussed above. This is also where we use the “SR2.5 ” condition
mentioned in Remark 1.6.
Structure of the proof. §2.2 summarizes elementary facts about Dedekind domains and
Cohen–Macaulay complexes; experts can skip this section. In §2.3 we introduce the complexes
Bn (I) and state Theorem E′ , which includes Theorem E as a special case.
The remainder of §2 is occupied by the proof of Theorem E′ , by induction on n. In §2.4
we establish the base cases n ≤ 2. In §2.5 we define a subcomplex Bngd (I) ⊂ Bn (I) consisting
of “good simplices” and prove that (Bn (I), Bngd (I)) is (n − 2)-connected. In §2.6 we prove
that (Bngd (I), Bn (0)) is (n − 2)-connected. In §2.7 we prove that Bn (0) is (n − 2)-connected.
In §2.8 we prove the connectivity of links, and assemble these pieces to prove Theorem E′ .
2.2
2.2.1
Dedekind domains and Cohen–Macaulay complexes
Dedekind domains
Let O be a Dedekind domain with field of fractions K. We recall some basic facts about O;
for proofs see e.g. [Mi].
Projective modules. Every submodule of a projective O-module is projective. If M is a
finitely generated projective O-module, the rank of M is defined by rk(M ) := dimK (M ⊗O K).
The following standard lemma summarizes some properties of projective O-modules.
8
Lemma 2.2. Let O be a Dedekind domain with field of fractions K and let M be a finitely
generated projective O-module of rank n ≥ 1.
(a) A submodule U ⊂ M is a direct summand of M if and only if M/U is torsion-free.
(b) If U and U ′ are summands of M and U ⊂ U ′ , then U is a summand of U ′ .
(c) The assignment U 7→ U ⊗O K defines a bijection
{Direct summands of M } ←→ {K-subspaces of M ⊗O K}
with inverse V 7→ V ∩ M .
The class group. The class group cl(O) is the set of isomorphism classes of rank 1 projective
O-modules. The tensor product endows cl(O) with the structure of an abelian group, which
we write additively; the identity element is given by the free module O. For every rank n
projective O-module M there is a unique rank 1 projective O-module I ∈ cl(O) such that
M ∼
= On−1 ⊕ I. We write [M ] = I ∈ cl(O) for this element. For two finite rank projective
O-modules M and M ′ , we have the identity [M ⊕ M ′ ] = [M ] + [M ′ ]. Two such modules M
and M ′ are isomorphic if and only if rk(M ) = rk(M ′ ) and [M ] = [M ′ ].
The complex of partial bases. A partial basis for On is a set {v1 , . . . , vk } of elements of
On that can be completed to a basis {v1 , . . . , vk , vk+1 , . . . , vn } for On . As we mentioned in
the introduction, the complex of partial bases Bn (O) is the simplicial complex whose (k − 1)simplices are partial bases {v1 , . . . , vk } for On . Steinitz proved in 1911 that when O is a
Dedekind domain, a vector v = (a1 , . . . , an ) ∈ On lies in a basis for On if and only if v is
unimodular, that is, the coordinates {a1 , . . . , an } generate the unit ideal of O; see [Re]. The
vertices of Bn (O) are therefore the unimodular vectors in On .
Lemma 2.3. Let O be a Dedekind domain.
(a) A subset {v1 , . . . , vk } ⊂ On is a partial basis of On if and only if span(v1 , . . . , vk ) is a
rank k direct summand of On .
(b) Let M be a free summand of On . A subset {v1 , . . . , vk } ⊂ M is a partial basis of M if
and only if {v1 , . . . , vk } is a partial basis of On .
Proof. Set U = span(v1 , . . . , vk ). If {v1 , . . . , vk } is a partial basis for On then U is a rank k
direct summand of On . We must prove the converse, so assume that U is a rank k direct
summand of On . Since U is a rank k projective module generated by k elements, it must be
free. Write On = U ⊕ W . Since [W ] = [On ] − [U ] = 0 − 0 = 0, we have W ∼
= On−k . Adjoining
a basis for W then yields a basis for On , so {v1 , . . . , vk } is a partial basis and (a) follows.
Part (b) follows from the characterization in part (a) together with Lemma 2.2(b).
2.2.2
Cohen–Macaulay complexes
A space is (−1)-connected if and only if it is nonempty; every space is ℓ-connected for ℓ ≤ −2.
A simplicial complex X is d-dimensional if it contains a d-simplex but no (d + 1)-simplices.
It is d-spherical if it is d-dimensional and (d − 1)-connected. Finally, X is Cohen–Macaulay
(CM) of dimension d if X is d-spherical and for every k-simplex σ k of X, the link LinkX (σ k )
is (d − k − 1)-spherical.
Low dimensions. A simplicial complex X is CM of dimension 0 if X is 0-dimensional. X
is CM of dimension 1 if X is 1-dimensional and connected.
9
Links are CM. If X is CM of dimension d, then LinkX (σ k ) is CM of dimension d − k − 1 for
all σ k ∈ X (where the definition requires only (d − k − 1)-spherical); see [Q, Proposition 8.6].
A sufficient condition to be CM. If a d-dimensional simplicial complex X has the property
that every simplex is contained in a d-simplex, then LinkX (σ k ) is (d − k − 1)-dimensional for
all σ k ∈ X. Therefore, to prove that X is CM of dimension d, it suffices to show that X is
(d − 1)-connected, and that for every k-simplex σ k ∈ X the link LinkX (σ k ) is (d − k − 2)connected.
2.3
The complex Bn (I) of partial I-bases
One of the key insights in our proof of Theorem E is that to induct on n, one must strengthen
the inductive hypothesis by working with a more general family of complexes.
Definition 2.4 (Partial I-basis). Fix a surjection L : On ։ O. Given an ideal I ⊂ O, a
basis {v1 , . . . , vn } for On is called an I-basis if
L(vi ) ≡ 0 mod I or L(vi ) ≡ 1 mod I
for all 1 ≤ i ≤ n.
A partial basis is called a partial I-basis if it can be completed to an I-basis. The complex
of partial I-bases, denoted Bn (I), is the complex whose (k − 1)-simplices are partial I-bases
{v1 , . . . , vk }.
Remark 2.5. The definition of Bn (I) depends on the choice of L. However, if L′ : On ։ O
is a different surjection, then both ker L and ker L′ are projective summands with rk(ker L) =
rk(ker L′ ) = n − 1 and [ker L] = [ker L′ ] = [On ] − [O] = 0 ∈ cl(O). Therefore there exists an
∼
=
→ Bn (O)
isomorphism Ψ : On → On such that L = L′ ◦Ψ. The resulting isomorphism Bn (O) −
induces an isomorphism between the complexes Bn (I) defined with respect to L and L′ .
Because of this, we will not concern ourselves with the choice of L except when it is necessary
to clarify our proofs.
We record some observations regarding I-bases and partial I-bases.
(a) When I = O, the condition of Definition 2.4 is vacuous, so Bn (I) = Bn (O).
(b) If I ⊂ J then any I-basis is a J-basis by definition. Therefore any partial I-basis is a
partial J-basis, so Bn (I) is a subcomplex of Bn (J). In particular, Bn (I) is a subcomplex
of Bn (O).
(c) We allow the case I = (0), in which case we simplify notation by writing Bn (0) for
Bn (I). The subcomplex Bn (0) is contained in Bn (I) for any ideal I since (0) ⊂ I. We
will see in the proof of Proposition 2.15 below that Bn (0) is the full subcomplex of
Bn (O) on the unimodular vectors v ∈ On which satisfy L(v) = 0 or L(v) = 1.
(d) A partial basis {v1 , . . . , vk } may not be a partial I-basis even if L(vi ) is congruent to
0 or 1 modulo I for all 1 ≤ i ≤ k. For example, let O = Z, let I = 5Z, and take
L : Z2 ։ Z to be L(b, c) = c. The vector v := (2, 5) ∈ Z2 determines a partial basis
{v} since {(2, 5), (1, 3)} is a basis for Z2 . However, although L(v) = 5 ≡ 0 mod I, the
partial basis {v} is not a partial I-basis. Indeed, any basis {(2, 5), (b, c)} must have
2c − 5b = ±1, which implies that c ≡ 2 or 3 mod 5Z. Therefore v is not contained in
any I-basis for Z2 , so {v} is not a vertex of B2 (I).
(e) A consequence of the previous paragraph is that for n ≥ 3 the complex of partial Ibases Bn (I) is not a full subcomplex of Bn (O) in general. With O = Z and I = 5Z
10
as before, take L : Z3 ։ Z to be L(a, b, c) = c. Using the additional room in Z3 , the
vector v := (0, 2, 5) can be extended to an I-basis such as {(0, 2, 5), (3, 1, 0), (−1, 0, 1)}.
The same is true of the vector w := (1, 2, 5) so both v and w belong to B3 (I). Since
{v, w, (0, 1, 3)} is a basis, the vertices v and w determine an edge of B3 (O). But they
do not form a partial I-basis, because as before any basis {(0, 2, 5), (1, 2, 5), (a, b, c)}
must have 2c − 5b = ±1. Thus this edge of B3 (O) is absent from B3 (I), even though
its endpoints do belong to B3 (I).
(f) In the previous two remarks, one might be concerned that Z does not satisfy the hy1
],
potheses of Theorem E. But in both remarks we can obtain an example in O = Z[ 19
which does satisfy the hypotheses of Theorem E, by taking the same vectors; the nec1
essary property is simply that 2 ∈
/ O× + 5O, which still holds when O = Z[ 19
].
Our main result concerning the complexes Bn (I) is as follows.
Theorem E′ . Let OS be a Dedekind domain of arithmetic type. Assume that |S| > 1 and S
contains a non-imaginary place. Then Bn (I) is CM of dimension n − 1 for any ideal I ⊂ OS .
Theorem E is the special case I = OS of Theorem E′ .
2.4
Step 1: Base case, n ≤ 2
Our inductive proof of Theorem E′ has two base cases. The first is easy.
Proposition 2.6. For any Dedekind domain O and any ideal I ⊂ O, the complex B1 (I) is
CM of dimension 0.
Proof. A surjection L : O1 ։ O must be an isomorphism, so there exists a unique v ∈ O1
with L(v) = 1. Since {v} is an I-basis of O1 for any I, the complex B1 (I) contains the
0-simplex {v}, and thus is 0-dimensional as desired.
The second base case is more involved. It is the key place where the arithmetic hypotheses
on OS enter into the proof of Theorem E′ .
Proposition 2.7. Let O = OS be a Dedekind domain of arithmetic type. Assume that
|S| > 1 and that S does not consist solely of imaginary places. Then for any ideal I ⊂ O,
the complex B2 (I) is CM of dimension 1.
Remark 2.8. Our proof of Proposition 2.7 could be shortened slightly if we were only
interested in Proposition 2.7; we go a bit further here to simplify the proof of Theorem 2.1
afterwards.
Proof of Proposition 2.7. The complex B2 (I) is 1-dimensional, so it is enough to show that
B2 (I) is connected. Since L is surjective we can fix a basis {e1 , e2 } for O2 such that L(e1 ) = 0
and L(e2 ) = 1. This is an I-basis, so it defines an edge e = {e1 , e2 } of B2 (I).
Let Γ1 (O, I) be the group
a b
a b ≡ ( ∗ ∗ ) mod I .
Γ1 (O, I) :=
01
c d
c d ∈ SL2 O
Letting SL2 O act on O2 via the basis {e1 , e2 }, the subgroup Γ1 (O, I) consists of the autoL
morphisms preserving the composition O2 → O ։ O/I. In particular, the Γ1 (O, I)-orbit of e
consists of I-bases and thus is contained in B2 (I). Let Z ⊂ B2 (I) be the subgraph determined
by this orbit.
11
If I 6= O then the action of Γ1 (O, I) on Z is without inversions. Indeed, in this case
vertices with L(v) ≡ 0 mod I cannot be exchanged with those with L(v) ≡ 1 mod I. The
edge e is a fundamental domain for the action of Γ1 (O, I) on Z. The Γ1 (O, I)-stabilizers of e1
and e2 are exactly the elementary matrices that lie in Γ1 (O, I), so by a standard argument in
geometric group theory (see, e.g., [Se, Lemma I.4.1.2]) the complex Z is connected if and only
if the group Γ1 (O, I) is generated by elementary matrices. This argument does not apply
verbatim when I = O, but it is easy to verify nevertheless that Z is connected if and only if
SL2 O is generated by elementary matrices.
Vaserstein (see the Main Theorem of [Va], and Liehl [L] for a correction) proved that if
O = OS is a Dedekind domain of arithmetic type with |S| > 1, then for any ideal I ⊂ O
the subgroup E of Γ1 (O, I) generated by elementary matrices is normal, and the quotient
Γ1 (O, I)/E is isomorphic to the relative K-group SK1 (O, I). We remark that Vaserstein’s
result does not apply to I = (0), but in this case Γ1 (O, I) consists solely of elementary
matrices. In their resolution of the Congruence Subgroup Problem, Bass–Milnor–Serre [BMS,
Theorem 3.6] proved that if O is a Dedekind domain of arithmetic type that is not totally
imaginary, then SK1 (O, I) = 0 for all ideals I ⊂ O. Our assumptions guarantee that both of
these results apply, so Γ1 (O, I) is generated by elementary matrices and Z is connected.
It is thus enough to show that the pair (B2 (I), Z) is 0-connected, i.e. that every vertex
v ∈ B2 (I) can be connected by a path to a vertex in Z. By definition, any vertex v ∈ B2 (I)
is contained in an I-basis {v, w} for O2 . We first show that if L(v) ≡ 1 mod I then v lies in
Z.
By possibly replacing w by w − v, we can assume that L(w) ≡ 0 mod I. There exists a
unique g ∈ GL2 O such that g · e1 = w and g · e2 = v. Let u = (det g)−1 w. The element γ
satisfying γ · e1 = u and γ · e2 = v lies in SL2 O. Since L(u) ≡ 0 mod I and L(v) ≡ 1 mod I,
we see that γ ∈ Γ1 (O, I). Therefore v lies in the Γ1 (O, I)-orbit of e2 and v ∈ Z.
If instead L(v) 6≡ 1 mod I, then since {v, w} is an I-basis we must have L(w) ≡ 1 mod I,
so w lies in Z. We conclude that every vertex of B2 (I) is connected to Z by a path of length
at most 1, as desired.
Before moving on to Step 2, we take a moment to prove Theorem 2.1.
Proof of Theorem 2.1. We first prove the theorem in the case n = 2. We can break the
hypotheses of Theorem 2.1 into the following three cases, and Cohn proved in each that
GL2 OS is not generated by elementary matrices together with diagonal matrices: for the
quadratic imaginary case see [C, Theorem 6.1], for the function field case with |cl(OS )| =
6 1
see [C, Corollary 5.6], and for the function field case with deg f ≥ 2 see [C, Theorem 6.2].
Thus certainly SL2 OS cannot be generated by elementary matrices. We saw in the proof of
Proposition 2.7 that the 1-dimensional complex B2 (OS ) is connected if and only if SL2 OS
is generated by elementary matrices. Therefore B2 (OS ) is disconnected and not Cohen–
Macaulay.
Now assume that n ≥ 3. Choose a basis {e1 , . . . , en } for On ; let σ n−3 = {e1 , . . . , en−2 } and
W = hen−1 , en i. Let BW (OS ) denote the complex of partial bases of W ; since W ∼
= OS2 there
is an isomorphism BW (OS ) ∼
= B2 (OS ). By the previous paragraph, BW (OS ) is disconnected.
Consider the subcomplex L := LinkBn (OS ) (σ n−3 ). If Bn (OS ) were CM of dimension n − 1
then L would be (n − 1) − (n − 3) − 1 = 1-spherical and therefore connected. Since {w1 , w2 }
is a basis of W if and only if {e1 , . . . , en−2 , w1 , w2 } is a basis of On , there is a natural
inclusion i : BW (OS ) ֒→ L. Let π : On ։ W be the projection with kernel he1 , . . . , en−2 i.
Since {e1 , . . . , en−2 , v1 , v2 } is a basis of On if and only if {π(v1 ), π(v2 )} is a basis of W , the
projection π induces a simplicial map π : L → BW (OS ). By definition i : BW (OS ) ֒→ L is a
12
section of this projection, so π : L → BW (OS ) is surjective. Since BW (OS ) is disconnected,
this shows that L = LinkBn (OS ) (σ n−3 ) is disconnected, so the (n − 1)-dimensional complex
Bn (OS ) is not Cohen–Macaulay.
Remark 2.9. Other examples can be deduced from Geller [G]; for example Bn (Of ) is not
Cohen–Macaulay for Of ∼
= F2 [X, Y ]/(X 2 + XY + Y 2 + X), which occurs as Of ⊂ F2 (T ) for
2
f = T + T + 1. In contrast, Vaserstein [Va] proved that when |S| ≥ 2 the group SL2 OS is
generated by elementary matrices, so in that case B2 (OS ) is indeed connected.
2.5
Step 2: (Bn (I), Bngd (I)) is (n − 2)-connected
We begin by defining certain bases to be “good bases”.
Definition 2.10. We say that a basis {v1 , . . . , vn } for On is good if L(vi ) = 1 for at least
one 1 ≤ i ≤ n. We say that {v1 , . . . , vk } is a partial good I-basis if it is contained in a good
I-basis. We denote by Bngd (I) the complex of partial good I-bases.
The following proposition is the main result of this section.
Proposition 2.11. Let O be a Dedekind domain and let I ⊂ O an ideal. Assume that B2 (I)
is CM of dimension 1. Then for any n ≥ 3 the pair (Bn (I), Bngd (I)) is (n − 2)-connected.
2.5.1
The link of an (n − 2)-simplex
We need a few preliminary results before starting the proof of Proposition 2.11. The first
ingredient in the proof of Proposition 2.11 is the following lemma.
Definition 2.12. Given a simplex σ = {v1 , . . . , vk } of Bn (O), define Vσ to be the direct
summand of On with basis {v1 , . . . , vk }, and define Iσ to be the ideal L(Vσ ) ⊂ O.
Lemma 2.13. Let n ≥ 3. Consider an (n − 2)-simplex σ n−2 ∈ Bn (I), and let v be a vertex
of σ with L(v) 6= 0. If Iσ 6⊂ I, assume additionally that L(v) ≡ 1 mod I. Then there exists a
vertex w of LinkBn (I) (σ) such that I{v,w} = O.
Proof. Since σ n−2 ∈ Bn (I) is a partial I-basis, there exists some z ∈ On such that σ ∪ {z} is
an I-basis. Let b1 = L(v) and b2 = L(z); by hypothesis b1 6= 0. Write σ n−2 = {v, v3 , . . . , vn }
and let bi = L(vi ) for 3 ≤ i ≤ n. Since σ ∪ {z} is a basis for On , the set {b1 , . . . , bn } ⊂ O
generates the unit ideal. Since n ≥ 3 and b1 6= 0 and {b1 , . . . , bn } generates the unit ideal, we
can apply a theorem of Reiner [Re] to conclude that there exists c3 , . . . , cn ∈ O such that the
two elements b1 and ξ := b2 + c3 b3 + c4 b4 + · · · + cn bn generate the unit ideal (this theorem of
Reiner is a strengthening for Dedekind domains of Bass’s stable range condition SR3 ). Set
y = z + c3 v3 + · · · cn vn . The ideal I{v,y} is generated by L(v) = b1 and L(y) = ξ, so by
construction I{v,y} = O. Since σ ∪ {z} is a basis for On , so is σ ∪ {y}.
If Iσ ⊂ I then L(v3 ) ≡ · · · ≡ L(vn ) ≡ 0 mod I, so L(y) ≡ L(z) mod I. Since z ∈ Bn (I),
we know that L(z) is congruent to 0 or 1 modulo I. Therefore when Iσ ⊂ I, the basis σ ∪ {y}
is an I-basis. We conclude that w := y ∈ LinkBn (I) (σ) satisfies the conditions of the lemma.
When Iσ 6⊂ I, the basis σ∪{y} may not be an I-basis since L(y) = ξ may not be congruent
to 0 or 1 modulo I. In this case, by hypothesis L(v) ≡ 1 mod I, so the vector w := y − ξv
satisfies
L(w) = L(y) − ξ · L(v) ≡ ξ − ξ = 0 mod I.
Therefore the basis σ ∪ {w} is an I-basis, so w ∈ LinkBn (I) (σ). Since
I{v,w} = (L(v), L(w)) = (L(v), L(y) − ξ · L(v)) = (L(v), L(y)) = (b1 , ξ) = O,
the vector w satisfies the conditions of the lemma.
13
2.5.2
Characterizing good partial bases
The second ingredient in the proof of Proposition 2.11 is the following characterization of
good partial bases.
Proposition 2.14. Let σ k = {v1 , . . . , vk+1 } be a k-simplex of Bn (O) with k < n − 1.
(1) Assume that L(vi ) ≡ 0 or 1 mod I for i = 1, . . . , k + 1. Then the following are equivalent.
(i) {v1 , . . . , vk+1 } is contained in a good basis.
(ii) {v1 , . . . , vk+1 } is contained in a good I-basis {v1 , . . . , vn } s.t. L(vi ) = 1 for i ≥ k+2.
(iii) There exists a complement W such that On = Vσ ⊕ W and L(W ) = O.
(2) If Iσ = (0) or Iσ = O or k < n − 2, then σ is contained in a good basis.
Before proving Proposition 2.14, we introduce some terminology. We say that σ is a good
simplex if the three conditions of Proposition 2.14(1) are satisfied, and call the subspace W
of condition (iii) a good complement for σ. In this case we denote by BW (I) the complex of
partial I-bases for W , whose simplices are subsets of W that can be extended to an I-basis
for W (with respect to L). Since [W ] = [On ] − [Vσ ] = 0, any good complement is a free
O-module, and L|W : W ։ O is surjective by definition. Thus by Remark 2.5, there exists
an isomorphism BW (I) ∼
= Bℓ (I), where ℓ = rk(W ).
Proof of Proposition 2.14. To prove the first part of the proposition we will show that
(i) =⇒ (ii) =⇒ (iii) =⇒ (i).
Assume (i), so σ = {v1 , . . . , vk+1 } is contained in a good basis
{v1 , . . . , vk+1 , uk+2 , . . . , un }.
By definition, some vector v lying in this good basis satisfies L(v) = 1. Define vi =
ui + (1 − L(ui ))v for k + 2 ≤ i ≤ n. The set {v1 , . . . , vn } is then a basis satisfying L(vi ) = 1
for k + 2 ≤ i ≤ n. Since k + 1 < n, this is a good basis, and since L(vi ) ≡ 0 or 1 mod I for
1 ≤ i ≤ k + 1 by assumption, it is also an I-basis. Therefore (i) =⇒ (ii).
Assume (ii), and define W = hvk+2 , . . . , vn i. We certainly have On = Vσ ⊕ W , and since
L(vn ) = 1 we have L(W ) = O. Therefore (ii) =⇒ (iii).
Finally, assume (iii), and choose any w ∈ W with L(w) = 1. We then have a splitting
W = hwi ⊕ ker L|W . The submodule ker L|W is free since [ker L|W ] = [W ] − [O] = 0, so we
can choose a basis {uk+3 , . . . , un } for ker L|W . Then
{v1 , . . . , vk+1 , w, uk+3 , . . . , un }
is a good basis containing σ. Therefore (iii) =⇒ (i).
We now prove the second part of the proposition. By Lemma 2.3(a), Vσ has a complement
U ⊂ On such that On = Vσ ⊕ U . In particular O = L(On ) = L(Vσ ) + L(U ) = Iσ + L(U ).
If Iσ = (0), this implies that L(U ) = O, so U is a good complement for σ. Otherwise,
choose a basis {uk+2 , . . . , un } for U . If Iσ = O, there exists v ∈ Vσ with L(v) = 1. Then
W := huk+2 + (1 − L(uk+2 ))v, uk+3 , . . . , un i is a good complement for σ.
It remains to handle the case when (0) 6= Iσ 6= O and k < n − 2, so rk(U ) ≥ 2. Since
Iσ 6= O we cannot have L(ui ) = 0 for all k + 2 ≤ i ≤ n, so assume without loss of generality
that L(un ) 6= 0. Consider the (n − 2)-simplex
σ n−2 := {v1 , . . . , vk+1 , uk+3 , . . . , un } ∈ Bn (O).
Applying Lemma 2.13 with I = O gives a vertex w in LinkBn (O) (σ) such that I{w,un } = O
and {v1 , . . . , vk+1 , w, uk+3 , . . . , un } is a basis for On . It follows that W := hw, uk+3 , . . . , un i
is a good complement for σ.
14
2.5.3
Completing the proof of Step 2
The pieces are now in place for the proof of Proposition 2.11.
Proof of Proposition 2.11. Given a k-simplex σ k ∈ Bn (I) with k ≤ n − 3, Proposition 2.14(2)
implies that σ is contained in Bngd (I). In other words, Bngd (I) contains the (n − 3)-skeleton
of Bn (I). Thus (Bn (I), Bngd (I)) is (n − 3)-connected, and every (n − 2)-simplex σ n−2 ∈ Bn (I)
has ∂σ n−2 ⊂ Bngd (I). To prove that (Bn (I), Bngd (I)) is (n − 2)-connected, it suffices to show
that every (n − 2)-simplex of Bn (I) that does not lie in Bngd (I) is homotopic relative to its
boundary into Bngd (I).
Let σ n−2 be an (n−2)-simplex of Bn (I) that does not lie in Bngd (I). By Proposition 2.14(2),
we must have (0) ( Iσ ( O. Choose a vertex v of σ with L(v) 6= 0; if Iσ 6⊂ I then choose
v such that L(v) 6≡ 0 mod I. Using Lemma 2.13, choose a vertex w of Bn (I) such that
σ n−2 ∪ {w} ∈ Bn (I) and I{v,w} = O. Write σ n−2 = σ0 ∪ {v}, let V = Vσ0 and let W = hv, wi.
The 1-dimensional complex BW (I), whose edges consist of I-bases for W , is contained in
LinkBn (I) (σ0 ). As we discussed following Proposition 2.14, identifying W with O2 induces
an isomorphism BW (I) ∼
= B2 (I). We have assumed that B2 (I) is connected, so BW (I) is
connected.
Fix a vertex y of BW (I) such that L(y) = 1, and let P be an embedded path in BW (I)
from v to y. Let v = p1 , . . . , pm = y be an enumeration of the vertices of P . For 1 ≤ i < m,
the set {pi , pi+1 } is a basis for W , so (L(pi ), L(pi+1 )) = L(W ) = O.
Regarding P as a triangulated closed interval that is mapped into LinkBn (I) (σ0 ), we obtain
an embedding of the join σ0 ∗ P into Bn (I). Its boundary is
∂(σ0 ∗ P ) = (∂σ0 ∗ P ) ∪ (σ0 ∗ ∂P ) .
Since ∂P = {v} ⊔ {y}, the second term σ0 ∗ ∂P consists of σ0 ∗ {v} = σ and σ0 ∗ {y}. Since
L(y) = 1, σ0 ∗ {y} is a partial good I-basis by Proposition 2.14(2), so σ0 ∗ {y} ∈ Bngd (I).
The first term ∂σ0 ∗ P can be written as
G
∂σ0 ∗ P =
∂σ0 ∗ {pi , pi+1 }.
i
Since I{pi ,pi+1} = O, every (n − 2)-simplex τ occurring in the join ∂σ0 ∗ {pi , pi+1 } has Iτ = O,
and thus lies in Bngd (I) by Proposition 2.14(2). We conclude that σ is the unique face of
σ0 ∗ P that is not contained in Bngd (I). The subdivided (n − 1)-simplex σ0 ∗ P thus provides
the desired homotopy of σ into Bngd (I).
2.6
Step 3: (Bngd (I), Bn (0)) is (n − 2)-connected
Since every vertex v ∈ Bn (0) has L(v) = 0 or L(v) = 1, every simplex σ ∈ Bn (0) has Iσ = (0)
or Iσ = O. By Proposition 2.14(2), every such simplex is good, so Bn (0) is contained in
Bngd (I). The main result of this section is the following proposition.
Proposition 2.15. Let O be a Dedekind domain, and fix n ≥ 3. Assume that for all
1 ≤ m < n and all ideals I ⊂ O, the complex Bm (I) is CM of dimension m − 1. Then
for any ideal I ⊂ O, the pair (Bngd (I), Bn (0)) is (n − 2)-connected.
Before proving Proposition 2.15 we establish three results that will be necessary in the proof.
15
2.6.1
Pushing into a subcomplex
The first ingredient is the following lemma on the connectivity of a pair.
Lemma 2.16. Let A be a full subcomplex of a simplicial complex X, and fix m ≥ 0. Assume
that for every k-simplex σ k ∈ X which is disjoint from A, the intersection LinkX (σ k ) ∩ A is
(m − k − 1)-connected. Then the pair (X, A) is m-connected.
The proof of Lemma 2.16 will use the language of posets. Recall that a poset P is said
to be d-connected if its geometric realization |P | is d-connected. Similarly, the homology
H∗ (P ; Z) is defined to be H∗ (|P | ; Z). Of course, a simplicial complex X can be regarded
as a poset P(X) whose elements are the simplices of X under inclusion. The canonical
homeomorphism X ∼
= |P(X)| shows that P(X) is d-connected if and only if X is, and there
is a canonical isomorphism H∗ (X; Z) ∼
= H∗ (P(X); Z).
Proof of Lemma 2.16. We can replace X by its (m + 1)-skeleton without affecting the truth
of the conclusion, so we can assume that X is finite-dimensional. Let A be the full complex
on the set of vertices A(0) ⊂ X (0) , so P(A) = {τ ∈ P(X) | τ ⊂ A(0) }. Define B := X (0) \ A(0) .
A simplex σ ∈ P(X) is disjoint from A if σ ∩ A(0) = ∅, i.e. if σ ⊂ B; in this case, LinkX (σ)∩ A
is {η ∈ A | σ ∪ η ∈ P(X)}.
We will use PL Morse theory as described in [Be]. Consider the PL Morse function
f : P(X) → N × −N defined by
f (σ) = |σ ∩ B|, −|σ ∩ A(0) | ,
with the lexicographic order on N × −N. If σ ( σ ′ , then we must have |σ ∩ B| < |σ ′ ∩ B| or
σ ∩ A(0) < σ ′ ∩ A(0) , so f (σ) 6= f (σ ′ ). This verifies that f has no “horizontal edges”. Our
assumption that X is finite-dimensional implies that the image of f is finite.
Note that f −1 ({0} × −N) = {σ ∈ X | |σ ∩ B| = 0} = P(A). For any σ ∈ P(X), define
the downward link to be
Link↓ (σ) := {τ ∈ P(X) | f (τ ) < f (σ) and either τ ( σ or τ ) σ}.
The fundamental theorem of PL Morse theory states that if Link↓ (σ) is (m−1)-connected
for all σ ∈ f −1 (N>0 × −N), then (X, A) is m-connected. We wish to verify this condition, so
consider σ ∈ f −1 (N>0 × −N).
We can uniquely write σ as the disjoint union of σB := σ ∩ B and σA := σ ∩ A(0) , where
|σB | > 0 by assumption. Define subposets of P(X) by
T := {τ ∈ P(X) | τ ( σ, |τ ∩ B| < |σ ∩ B|},
R := {ρ ∈ P(X) | ρ ) σ, |ρ ∩ B| = |σ ∩ B|, |ρ ∩ A(0) | > |σ ∩ A(0) |}.
The downward link Link↓ (σ) is the sub-poset T ∪ R of P(X); since every τ ∈ T and every
ρ ∈ R satisfy τ ( ρ, in fact the downward link is the join T ∗ R.
Now, either σA 6= ∅ or σA = ∅. In the first case, the poset T is contractible via the
homotopy that takes τ to τ ∪ σA and then to σA . We conclude that in this case Link↓ (σ) =
T ∗ R is contractible, and in particular is (m − 1)-connected.
So suppose that σA = ∅. In other words, σ is disjoint from A. In this case, T is the
poset of proper nonempty subsets of the (k + 1)-element set σ k , so T ∼
= P(∂∆k ) ≃ S k−1 ; in
particular, T is (k − 1)-connected. The poset R consists of ρ ∈ P(X) that can be written as
ρ = σ ∪ η for some nonempty η ∈ P(A). The assignment σ ∪ η 7→ η gives an isomorphism of
R with P(LinkX (σ) ∩ A). Our assumption guarantees that R is (m − k − 1)-connected, so
Link↓ (σ) = T ∗ R is (m − 1)-connected, as desired.
16
2.6.2
Topology of nicely fibered complexes
The second ingredient we will need concerns the topology of simplicial complexes that are
“fibered” over a CM complex in the following sense.
Definition 2.17. Let F : X → Y be a simplicial map between simplicial complexes. Let
C ⊂ X (0) be a subset of the vertices of X with F (C) = Y (0) .
We say that X is fibered over Y by F with core C if the following condition holds: a
subset U = {x0 , . . . , xk } ⊂ X (0) forms a k-simplex of X if and only if
(1) F (U ) = {F (x0 ), . . . , F (xk )} is a k-simplex of Y , and
(2) if F (U ) is a maximal simplex of Y , then U ∩ C 6= ∅.
If C = X (0) , we say X is fully fibered over Y by F ; in this case condition (2) is vacuous.
The main result concerning fibered complexes is as follows.
Lemma 2.18. Assume that X is fibered over Y by F : X → Y with core C ⊂ X (0) . If Y is
CM of dimension d then X is CM of dimension d.
To prove Lemma 2.18 we will make use of a theorem of Quillen that describes the topology
of a poset map in terms of the topology of the fibers.
Given a poset A, recall that the height ht(a) of an element a is the largest k for which
there exists a chain a0
a1
···
ak = a. When a k-simplex σ k ∈ X is regarded as an
element of the poset P(X), it has height k. Given a map of posets F : A → B and element
b ∈ B, the fiber F≤b is the subposet of A defined by F≤b := {a ∈ A | F (a) ≤ b}.
A poset map F : A → B is strictly increasing if a < a′ implies F (a) < F (a′ ). Finally, a
poset P is said to be CM of dimension d if its geometric realization |P | is CM of dimension d
as a simplicial complex. A simplicial complex X is CM of dimension d if and only if P(X) is
CM of dimension d (this is not automatic since being CM is not homeomorphism-invariant;
nevertheless it follows from [Q, (8.5)]).
Theorem 2.19 (Quillen [Q, Theorem 9.1, Corollary 9.7]). Let F : A → B be a strictly
increasing map of posets. Assume that B is CM of dimension d and that for all b ∈ B, the fiber
e d (A; Z) → H
e d (B; Z)
F≤b is CM of dimension ht(b). Then A is CM of dimension d and F∗ : H
is surjective.
We also record the following observation, which follows directly from Definition 2.17.
Lemma 2.20. Let X be fibered over Y by F with core C, and consider a simplex σ ∈ X.
• If σ ∩ C 6= ∅, then the link LinkX (σ) is fully fibered over LinkY (F (σ)) by F .
• If σ ∩ C = ∅, then the link LinkX (σ) is fibered over LinkY (F (σ)) by F with core
C ∩ LinkX (σ)(0) .
Proof of Lemma 2.18. We prove the lemma by induction on d. The base case d = 0 is trivial,
so assume that d > 0 and that the lemma is true for all smaller values of d. We must prove
two things about X: that for every σ k ∈ X the link LinkX (σ k ) is CM of dimension d − k − 1,
and that X itself is d-spherical. By condition (1), the k-simplex σ k ∈ X projects to a ksimplex F (σ k ) ∈ Y . By Lemma 2.20, the complex LinkX (σ k ) is fibered over LinkY (F (σ k ))
by F (with some core). Since Y is CM of dimension d, the complex LinkY (F (σ k )) is CM of
dimension d − k − 1. Applying the inductive hypothesis, we conclude that LinkX (σ k ) is CM
of dimension d − k − 1, as desired.
It remains to prove that X is d-spherical. We begin by proving the lemma in the special
case when X is fibered over a single simplex Y = ∆d with vertices {y0 , . . . , yd }. Define Z to be
17
the complex with Z (0) = X (0) and where U ⊂ Z (0) forms a simplex if and only if U contains
at most one element from each F −1 (yi ). The complex Z is the join of the 0-dimensional
complexes F −1 (yi ). Since F (C) = Y (0) , each F −1 (yi ) is nonempty and thus 0-spherical. The
join of an r-spherical and an s-spherical complex is (r + s + 1)-spherical [Q, Example 8.1], so
the (d + 1)-fold join Z = F −1 (y0 ) ∗ · · · ∗ F −1 (yd ) is d-spherical.
Observe that X ⊂ Z. Since Z is d-spherical, it is enough to prove that the pair (Z, X) is
(d − 1)-connected. Since Y is itself a d-simplex, the only maximal simplex of Y is Y itself,
and hence condition (2) applies only to d-simplices of X. Therefore X and Z coincide in
dimensions up to d − 1 and Z is obtained from X by adding certain d-simplices σ d ∈ Z
with σ d 6∈ X. It thus suffices to show for each such σ d that the boundary ∂σ d ⊂ X is
null-homotopic in X.
By condition (1), such a d-simplex can be written as σ d = {x0 , . . . , xd } with F (xi ) = yi
for 0 ≤ i ≤ d, and by condition (2) we must have xi 6∈ C for all 0 ≤ i ≤ d. Since F (C) = Y (0) ,
for each 0 ≤ i ≤ d we can choose x′i ∈ C with F (x′i ) = yi . Consider the full subcomplex of Z
on {x0 , x′0 , . . . , xd , x′d }. By construction, this is the (d + 1)-fold join of the 0-spheres {xi , x′i },
and thus is homeomorphic to a d-sphere. Every simplex of this d-sphere except σ contains
some x′i , and thus lies in X. Therefore the full subcomplex of X on {x0 , x′0 , . . . , xd , x′d } is a
subdivided d-disk D with ∂D = ∂σ d , as desired.
This proves that X is d-spherical if it is fibered over a d-simplex with any core; combined
with the first paragraph, X is CM of dimension d in this case.
We now prove the lemma for a general CM complex Y of dimension d. By condition
(1), the map F induces a height-preserving poset map F : P(X) → P(Y ); we will verify the
conditions of Theorem 2.19 for F .
Fix a simplex τ k ∈ P(Y ). Let T ≃ ∆k be the subcomplex of Y determined by τ k and
let Te be the subcomplex of X consisting of σ ℓ ∈ X with F (σ ℓ ) ⊂ τ k . The fiber F≤τ k is the
subposet of P(X) consisting of simplices of Te, so F≤τ k = P(Te). Since F is height-preserving,
if k < d then no d-simplex is contained in Te, so condition (2) is vacuous; in this case Te is
fully fibered over T ≃ ∆k by F . If τ d is a d-simplex, condition (2) is not vacuous, and in this
case Te is fibered over T ≃ ∆d by F with core C ∩ Te(0) . In either case the base T is a single
simplex, so the special case of the lemma implies that Te is CM of dimension ht(τ k ) = k,
and thus so is F≤τ k = P(Te). Since P(Y ) is CM of dimension d by assumption, this verifies
the hypotheses of Theorem 2.19 for the map F : P(X) → P(Y ). Theorem 2.19 implies that
P(X) is CM of dimension d, so X is CM of dimension d as well.
2.6.3
A criterion for being fibered
The final ingredient we need is the following condition to verify that a complex is fibered
over another.
Lemma 2.21. Let X and Y be d-dimensional simplicial complexes with the property that
every simplex is contained in a d-simplex. Let F : X → Y be a simplicial map and C ⊂ X (0)
be a subset such that the following two conditions hold.
(i) A subset U = {x0 , . . . , xd } forms a d-simplex of X
⇐⇒ F (U ) is a d-simplex of Y and U ∩ C 6= ∅.
(ii) For each d-simplex τ d ∈ Y there exists U ⊂ C with F (U ) = τ d .
Then X is fibered over Y by F with core C.
Proof. The maximal simplices of Y are precisely the d-simplices, so when |U | = d+1 condition
(i) is equivalent to conditions (1) and (2) of Definition 2.17. Since every vertex of Y is
18
contained in a d-simplex, condition (ii) implies that F (C) = Y (0) . It remains to verify
conditions (1) and (2) for subsets U ⊂ X (0) with |U | < d + 1.
Every simplex τ k ∈ X is contained in a d-simplex σ d ∈ X. By condition (i), its image
F (σ d ) is a d-simplex of Y , so F (τ k ) is a k-simplex of Y . Conversely, consider a subset
U = {x0 , . . . , xk } with k < d such that F (U ) is a k-simplex of Y . Then F (U ) is contained
in a d-simplex η d = F (U ) ∪ {yk+1 , . . . , yd } ∈ Y . Using the assumption that F (C) = Y (0) ,
choose lifts xi ∈ C with F (xi ) = yi for k + 1 ≤ i ≤ d. By construction, U ′ := {x0 , . . . , xd }
has F (U ′ ) = η d and U ′ ∩ C 6= ∅, so by condition (i) U ′ is a d-simplex of X. It follows that U
is a k-simplex of X, as desired.
2.6.4
Completing the proof of Step 3
We are now ready to prove Proposition 2.15.
Proof of Proposition 2.15. We want to show that (Bngd (I), Bn (0)) is (n − 2)-connected. We
will prove this by applying Lemma 2.16 with X = Bngd (I) and A = Bn (0) and m = n − 2.
To apply Lemma 2.16 we need to verify the two hypotheses: first, that Bn (0) is a full
subcomplex of Bngd (I), and second, that for any k-simplex σ k of Bngd (I) that is disjoint from
Bn (0), the intersection
Link0 (σ k ) := Bn (0) ∩ LinkBngd (I) (σ k )
is (n − k − 3)-connected.
We begin by proving that Bn (0) is the full subcomplex of Bn (O) on the vertex set
Bn (0)(0) := {v ∈ Bn (O) | L(v) = 0 or L(v) = 1} .
Let σ = {v1 , . . . , vk } ∈ Bn (O) be a partial basis with vi ∈ Bn (0)(0) for 1 ≤ i ≤ k. We
have either Iσ = O (if some vi has L(vi ) = 1) or Iσ = (0) (if all vi have L(vi ) = 0). By
Proposition 2.14(2), σ is contained in a good basis. By Proposition 2.14(1)(ii) this implies
that σ is contained in a good (0)-basis, so σ ∈ Bn (0). This proves that Bn (0) is a full
subcomplex of Bn (O), and thus of Bngd (I), verifying the first hypothesis.
For the second hypothesis, consider a k-simplex σ k ∈ Bngd (I) that is disjoint from Bn (0)(0) ;
we will prove that Link0 (σ k ) is CM of dimension n − k − 2. Using Proposition 2.14(1)(iii),
choose a good complement W . By definition, we have On = Vσ ⊕ W and L(W ) = O. Let
π : On ։ W be the projection with kernel Vσ .
We will prove the following claims, using the terminology of Definition 2.17.
1. If Iσ 6= O, then Link0 (σ) is fully fibered over BW (Iσ ) by π.
2. If Iσ = O, then Link0 (σ) is fibered over BW (O) by π with core
C := {v ∈ Link0 (σ) | L(v) = 1} ⊂ Link0 (σ)(0) .
The bulk of the argument does not depend on whether Iσ = O or not, so we begin without
making any assumption on Iσ .
By definition, every simplex of BW (Iσ ) is contained in an (n−k−2)-simplex, corresponding
to an Iσ -basis of W . Similarly, if η ∈ Link0 (σ k ) then σ ∪ η ∈ Bngd (I) is a partial good Ibasis. By Proposition 2.14(1)(ii) we can enlarge this to a good I-basis σ ∪ τ with τ ∈ Bn (0).
Therefore every simplex η ∈ Link0 (σ k ) is contained in an (n − k − 2)-simplex τ n−k−2 ∈
Link0 (σ k ). This means that in verifying the claim, we can work only with (n − k − 2)simplices, and appeal to Lemma 2.21 for the general case.
A subset U = {uk+2 , . . . , un } ⊂ Bn (0)(0) forms an (n − k − 2)-simplex of Link0 (σ k ) if
σ ∪ U = {v1 , . . . , vk+1 , uk+2 , . . . , un } is a good I-basis, or equivalently if σ ∪ U is a good basis.
19
Since σ is a basis of Vσ and W is a complement to Vσ , we know that σ ∪ U is a basis for On if
and only if π(U ) is a basis for W . Moreover, any u ∈ Bn (0) has L(u) = 0 or L(u) = 1, which
implies L(π(u)) ≡ 0 mod Iσ or L(π(u)) ≡ 1 mod Iσ . Therefore σ ∪ U is a basis for On if and
only if π(U ) is an Iσ -basis for W .
We now prove the two cases of the claim separately. First, assume that Iσ = O. Since σ
is disjoint from Bn (0) by assumption, a basis σ ∪ U with U ⊂ Bn (0)(0) is good if and only if
U ∩ C 6= 0. This means that U forms an (n − k − 2)-simplex of Link0 (σ) if and only if π(U )
is an (n − k − 2)-simplex of BW (O) and U ∩ C 6= 0, verifying condition (i) of Lemma 2.21.
It remains to verify condition (ii) of Lemma 2.21. Given a basis τ = {wk+2 , . . . , wn }
of W , choose for each i = k + 2, . . . , n some vi′ ∈ Vσ with L(vi′ ) = 1 − L(wi ), and define
ui = wi + vi′ . By construction we have L(ui ) = 1 and π(ui ) = π(wi ) + π(vi′ ) = wi . Setting
U = {uk+2 , . . . , un } we see that π(U ) is the basis τ of W , and therefore U ⊂ C. We can
therefore apply Lemma 2.21 to conclude that Link0 (σ) is fibered over BW (Iσ ) by π with core
C, verifying Claim 1.
Second, assume that Iσ 6= O. In this case any basis σ ∪ U with U ⊂ Bn (0)(0) must be
good, because if L(ui ) = 0 for all k + 2 ≤ i ≤ n we would have L(On ) = Iσ 6= O. Therefore
σ ∪ U is a good basis if and only if π(U ) is an Iσ -basis of W . This verifies condition (i) of
Lemma 2.21, so it remains to check condition (ii).
Let τ = {wk+2 , . . . , wn } be an Iσ -basis of W . For i = k + 2, . . . , n, if L(wi ) ≡ 0 mod Iσ ,
choose vi′ ∈ Vσ with L(vi′ ) = −L(wi ); if L(wi ) ≡ 1 mod Iσ , choose vi′ ∈ Vσ with L(vi′ ) =
1 − L(wi ). In either case, define ui = wi + vi′ ; by construction, we have L(ui ) = 0 or L(ui ) = 1
and π(ui ) = π(wi ) + π(vi′ ) = wi . Setting U = {uk+2 , . . . , un } we see that π(U ) is the Iσ -basis
τ of W ; therefore U ⊂ Link0 (σ). We again apply Lemma 2.21 to conclude that Link0 (σ) is
fully fibered over BW (Iσ ) by π, verifying Claim 2.
Now that Claims 1 and 2 established, we finish the proof of the proposition. Since
W ∼
= On−k−1 , Remark 2.5 yields an isomorphism BW (Iσ ) ∼
= Bn−k−1 (Iσ ). Our hypothesis
guarantees that Bn−k−1 (Iσ ) is CM of dimension n − k − 2, so the same is true of BW (Iσ ). We
established in Claims 1 and 2 above that Link0 (σ) is fibered over BW (Iσ ) by π (either with
core C or with core Link0 (σ)(0) ). Lemma 2.18 thus implies that Link0 (σ) is CM of dimension
n − k − 2; in particular Link0 (σ) is (n − k − 3)-connected. This verifies the hypothesis of
Lemma 2.16 for the pair (Bngd (I), Bn (0)) with m = n − 2; we conclude that (Bngd (I), Bn (0)) is
(n − 2)-connected, as desired.
Remark 2.22. Many of the ingredients of the proof that Link0 (σ k ) is (n − k − 3)-connected
were introduced only for this step; this step thereby accounts for much of the complexity of
our proof of Theorem E.
• Even if we had only been interested in Bn (O), we were forced in the proof of Proposition 2.15 to consider the complex Bn−k−1 (Iσ ). This is the entire reason that we
introduced the complexes Bn (I) relative to nontrivial ideals of O.
• In the proof of Proposition 2.15 we needed the assumption that Bm (Iσ ) was CM of
dimension m−1 for m < n. This is the reason that we are forced to prove in Theorem E′
that Bn (I) is Cohen–Macaulay of dimension n − 1, rather than simply that Bn (I) is
(n − 1)-spherical. This weaker statement would not suffice as an inductive hypothesis;
Lemma 2.18 is not true if one only assumes that Y is d-spherical.
• In the case Iσ = O above, we applied Lemma 2.18 in a case where the core C only
contained some of the vertices of the domain. This is the reason that we introduced the
notion of “fibered with core C”; we will apply Lemma 2.18 in Step 4 and in the proof
of Theorem E as well, but only in the fully fibered case.
20
2.7
Step 4: Bn (0) is (n − 2)-connected
The main result of this section is the following proposition.
Proposition 2.23. Let O be a Dedekind domain, and assume that Bn−1 (O) is CM of dimension n − 2. Then Bn (0) is (n − 2)-connected.
Proof. Define Zn to be the full subcomplex of Bn (0) spanned by those vertices v of Bn (0) with
L(v) = 0. The complex Zn consists of partial bases contained in the summand ker L ∼
= On−1 .
In fact, by Lemma 2.3(b) these are precisely the partial bases of the summand ker L, so we
have an isomorphism Zn ∼
= Bn−1 (O). Therefore our assumption guarantees that Zn is CM
of dimension n − 2.
Consider a k-simplex σ k = {v0 , v1 , . . . , vk } ∈ Bn (0) disjoint from Zn , i.e. with L(v0 ) =
· · · = L(vk ) = 1. Let η k−1 = {v1 − v0 , . . . , vk − v0 } and τ k = {v0 } ∪ η ∈ Bn (0). Since Vσ = Vτ ,
the links LinkBn (0) (σ k ) and LinkBn (0) (τ k ) coincide. Furthermore, given ν n−k−2 ∈ Zn the set
{v0 } ∪ η ∪ ν is a basis for On if and only if η ∪ ν is a basis for ker L. In other words,
LinkBn (0) (τ k ) ∩ Zn = LinkZn (η k−1 ).
Since Zn is CM of dimension n − 2, we know that LinkZn (η k−1 ) is CM of dimension
(n − 2) − (k − 1) − 1 = n − k − 2, and is thus (n − k − 3)-connected. We now apply
Lemma 2.16 with X = Bn (0) and A = Zn and m = n − 2; the conclusion is that (Bn (0), Zn )
is (n − 2)-connected.
Fix a vector v ∈ On with L(v) = 1. Since {u2 , . . . , un } ⊂ ker L is a basis for ker L if and
only if {v, u2 , . . . , un } is a basis for On , every simplex of Zn is contained in LinkBn (0) (v). We
conclude that the image of Zn ֒→ Bn (0) can be contracted to v. Together with the fact that
(Bn (0), Zn ) is (n − 2)-connected, this implies that Bn (0) is (n − 2)-connected, as desired.
2.8
Assembling the proof: Bn (I) is CM of dimension n − 1
We are finally ready to combine the steps of the previous sections and prove Theorem E′ .
Proof of Theorem E′ . We prove by induction on n that Bn (I) is CM of dimension n − 1 for
all ideals I ⊂ O. The base cases are n = 1 and n = 2. Proposition 2.6, which holds for any
Dedekind domain O, states that B1 (I) is CM of dimension 0. Proposition 2.7 states that
under precisely our hypotheses, B2 (I) is CM of dimension 1.
Now fix n ≥ 3 and an ideal I ⊂ O, and assume that Bm (J) is CM of dimension m−1 for all
m < n and all J ⊂ O. Under these assumptions, Proposition 2.11 states that (Bn (I), Bngd (I))
is (n − 2)-connected; Proposition 2.15 states that (Bngd (I), Bn (0)) is (n − 2)-connected; and
Proposition 2.23 states that Bn (0) itself is (n − 2)-connected. Together these imply that the
(n − 1)-dimensional complex Bn (I) is (n − 2)-connected, and thus (n − 1)-spherical.
It remains to prove that for every k-simplex σ k ∈ Bn (I), the space LinkBn (I) (σ k ) is CM
of dimension n − k − 2. If k = n − 1 this is vacuous. If k = n − 2, we must prove that
LinkBn (I) (σ k ) is nonempty. But the partial I-basis σ n−2 is by definition contained in an
I-basis σ n−2 ∪ {v}, so LinkBn (I) (σ k ) contains the 0-simplex {v}.
Now assume that k < n − 2. By Lemma 2.14(1), σ has a good complement W . Let
π : On ։ W be the projection with kernel Vσ . We will prove that LinkBn (I) (σ) is fully fibered
over BW (I + Iσ ) by π using Lemma 2.21.
By definition, every simplex of LinkBn (I) (σ) or BW (I + Iσ ) is contained in an (n − k − 2)simplex. Given a subset U = {uk+2 , . . . , un } with L(ui ) ≡ 0 or 1 mod I for each i = k +
2, . . . , n, the set σ ∪ U is a basis of On if and only if π(U ) is a basis of W . Moreover
21
L(π(u)) ≡ L(u) mod Iσ , so in this case π(U ) is an (I + Iσ )-basis of W . Therefore such a
subset U is an (n − k − 2)-simplex of LinkBn (I) (σ) if and only if π(U ) is an (n − k − 2)-simplex
of BW (I + Iσ ), verifying condition (i) of Lemma 2.21.
It remains to check condition (ii) of Lemma 2.21, so consider an (I + Iσ )-basis τ =
{wk+2 , . . . , wn } of W . For each i we have L(wi ) ≡ 0 or 1 mod I + Iσ , so there exists vi ∈ Vσ
such that L(wi + vi ) ≡ 0 or 1 mod I. The set U := {wk+2 + vk+2 , . . . , wn + vn } projects under
π to τ , so it is an (n − k − 2)-simplex of LinkBn (I) (σ).
Lemma 2.21 now implies that LinkBn (I) (σ) is fully fibered over BW (I + Iσ ) by π. Since
W ∼
= On−k−1 , the complex BW (I + Iσ ) is isomorphic to Bn−k−1 (I + Iσ ), which we have
assumed by induction is CM of dimension n − k − 2. Applying Lemma 2.18 shows that
LinkBn (I) (σ k ) is CM of dimension n − k − 2, as desired.
We have proved that Bn (I) is (n − 1)-spherical and that LinkBn (I) (σ k ) is CM of dimension
n − k − 2 for every σ k ∈ Bn (I), so Bn (I) is CM of dimension n − 1. Since the ideal I was
arbitrary, this concludes the inductive step, and thus concludes the proof of Theorem E′ .
3
Integral apartments
In this section we use Theorem E to prove Theorem A.
Proof of Theorem A. Let C∗ be the reduced chain complex of the simplicial complex Bn (O),
so C−1 ∼
= Z. We define the integral apartment class map
φ : Cn−1 → Stn (K)
as follows. An ordered basis v = (v1 , . . . , vn ) of On determines an (n − 1)-simplex of Bn (O),
and thus a generator [v] = [v1 , . . . , vn ] of Cn−1 . Reordering the basis simply changes the
orientation of this simplex, so we have [vτ (1) , . . . , vτ (n) ] = (−1)τ [v1 , . . . , vn ] for τ ∈ Sn ; there
are no other relations between the elements [v] ∈ Cn−1 . Given v, define lines L1 , . . . , Ln
in K n via the formula Li := spanK (vi ). Since {v1 , . . . , vn } is a basis of On , Lemma 2.2(c)
implies that Li ∩ On = spanO (vi ) and On = spanO (vi ) ⊕ · · · ⊕ spanO (vi ), so the frame
Lv := {L1 , . . . , Ln } is integral.
Let Av := ALv be the integral apartment determined by this integral frame, and define
φ([v]) = [Av ] ∈ Stn (K). Reordering the basis vectors gives the same frame, and thus the
same apartment, but possibly with reversed orientation: [Aτ (v) ] = (−1)τ [Av ]. This shows
that the map φ is well-defined. Our goal is to prove that φ is surjective, i.e. that Stn (K) is
generated by integral apartment classes. We will do this by factoring the map φ as follows.
∼
=
For a simplicial complex X, the isomorphism b : H∗ (X; Z) → H∗ (P(X); Z) determined by
the homeomorphism between X and its barycentric subdivision is induced by the chain map
X
[{x1 , . . . , xk }] 7→
(−1)τ {xτ (1) } ⊂ {xτ (1) , xτ (2) } ⊂ · · · ⊂ {xτ (1) , . . . , xτ (k) } .
τ ∈Sk
Let Bn (O)(n−2) be the (n − 2)-skeleton of Bn (O) and let P (n−2) := P(Bn (O)(n−2) ) be the
poset of simplices of Bn (O)(n−2) . In the introduction we considered Tn (K) to be a simplicial
complex, but let us define Tn (K) to be the poset of proper nonzero K-subspaces of K n ; the
simplicial complex of flags from the introduction is then its geometric realization. Define the
poset map F : P (n−2) → Tn (K) by
F ({v1 , . . . , vr }) = spanK (v1 , . . . , vr ).
22
Since spanK (v1 , . . . , vr ) is an r-dimensional subspace of K n , the restriction to simplices of
Bn (O)(n−2) (i.e. to simplices satisfying 0 < r < n) guarantees that this is a proper nonzero
subspace, so F is well-defined and also height-preserving.
Our claim is that the integral apartment class map φ : Cn−1 → Stn (K) factors as the
composition
∂
e n−2 (Bn (O)(n−2) ; Z)
φ : Cn−1 −
→ ker(∂ : Cn−2 → Cn−3 ) = H
F
b
∗ e
e n−2 (P (n−2) ; Z) −→
H
Hn−2 (Tn (K); Z) = Stn (K).
−
→
∼
(3.1)
=
Indeed, consider an ordered basis P
v = (v1 , . . . , vn ) for On . Under the boundary map
∂ : Cn−1 → Cn−2 , we have ∂[v] = ni=1 (−1)i [v(i) ], where v(i) = {v1 , . . . , vbi , . . . , vn }. Under b, this is taken to
X
(−1)σ {xσ(1) } ⊂ · · · ⊂ {xσ(1) , . . . , xσ(n−1) } .
b(∂[v]) =
σ∈Sn
Under F∗ , the term corresponding to σ ∈ Sn is taken to Lσ(1) ⊂ Lσ(1) + Lσ(2) ⊂ · · · ⊂
Lσ(1) + · · · + Lσ(n−1) . These are precisely the (n − 2)-simplices making up the apartment
Av (compare with §5.1 and Example 5.1 below), and so F∗ (b(∂[v])) = [Av ], as claimed.
To prove that φ is surjective, we prove that the maps ∂ and F∗ occurring in (3.1) are
surjective. We first establish that under our assumptions, Bn (O) is CM of dimension n − 1.
In the case when |S| > 1 and S contains a non-imaginary place, this is the statement of
Theorem E. In the remaining case when O is Euclidean, this was proved by Maazen [Ma]
(see [vdKL, Appendix] and [DP, Proof of Theorem B] for published proofs). By definition,
e n−2 (Bn (O); Z). We have just estabthe cokernel of ∂ : Cn−1 → ker(∂ : Cn−2 → Cn−3 ) is H
lished that Bn (O) is CM of dimension n − 1, and thus certainly (n − 2)-acyclic. Therefore
e n−2 (Bn (O); Z) = 0 and ∂ : Cn−1 → ker(∂ : Cn−2 → Cn−3 ) is surjective.
H
To show that F∗ is surjective, we will apply Theorem 2.19 to the height-preserving poset
map F , so we must verify that its hypotheses are satisfied. The Solomon–Tits theorem says
that Tn (K) is CM of dimension (n − 2) (see e.g. [Q, Example 8.2]). Given V ∈ Tn (K), let
ℓ = dim(V ), so ht(V ) = ℓ − 1. Lemma 2.2(c) implies that V ∩ On is a rank-ℓ direct summand
of On . Since |cl(O)| = 1, every f.g. projective O-module is free, so V ∩ On is isomorphic to
Oℓ . Lemma 2.3(b) states that {v1 , . . . , vk } ∈ F≤V exactly if {v1 , . . . , vk } is a partial basis
of V ∩ On , so we have an isomorphism F≤V ∼
= P(Bℓ (O)). We established in the previous
paragraph that P(Bℓ (O)) is CM of dimension ℓ − 1 = ht(V ), so this verifies the remaining
hypothesis of Theorem 2.19. Applying Theorem 2.19, we deduce that the map
e n−2 (P (n−2) ; Z) → H
e n−2 (Tn (K); Z) = Stn (K)
F∗ : H
is surjective. Since all other maps in (3.1) are isomorphisms, this shows that the integral
apartment class map φ : Cn−1 → Stn (K) is surjective.
4
Vanishing Theorem
In this section we deduce Theorem C from Theorem A. We first remark that by Dirichlet’s
× ∼ r1 +r2 −1
× µK . This is a virtual Poincaré duality group with virtual
unit theorem, OK
=Z
cohomological dimension r1 + r2 − 1, so Bieri–Eckmann [BE, Theorem 3.5] implies that
×
), i.e. that
vcd(SLn OK ) = vcd(GLn OK ) − vcd(OK
vcd(SLn OK ) = νn′ := r1 n+1
− 1 + r2 n2 − 1 − (n − 1) .
2
23
Proof of Theorem C. In the special case λ = 0, Theorem C states that Hνn (GLn OK ; K) = 0
for all n ≥ 2 (and similarly for SLn OK ). Since H∗ (GLn OK ; K) ∼
= H∗ (GLn OK ; Q) ⊗Q K, the
last claim of the theorem follows from the first ones, so those are all we need to prove.
Fix λ = (λ1 , . . . , λn ) ∈ Zn with λ1 ≥ · · · ≥ λn . As we explained in §1.3, Borel–Serre
duality implies that Stn (K) is the rational dualizing module for both SLn OK and GLn OK ,
which implies that
Hνn (GLn OK ; Vλ ) = (Stn (K) ⊗ Vλ )GLn OK ,
′
Hνn (SLn OK ; Vλ ) = (Stn (K) ⊗ Vλ )SLn OK .
Since SLn OK ⊂ GLn OK , the coinvariants (Stn (K) ⊗ Vλ )GLn OK are a quotient of (Stn (K) ⊗
Vλ )SLn OK . The upshot is that to prove Theorem C, it is enough to prove that (Stn (K) ⊗
Vλ )SLn OK = 0 when n ≥ 2 + kλk.
Define λ′ = (λ′1 , . . . , λ′n−1 , 0) by λ′i = λi − λn ; observe that kλ′ k = kλk. As representations
of SLn OK , the representations Vλ and Vλ′ are isomorphic; they only differ as representations
of GLn OK . Using Schur–Weyl duality, we can embed the SLn K-representation Vλ′ as a
direct summand of (K n )⊗ℓ with ℓ := kλk. It follows that it is enough to prove that (Stn (K) ⊗
(K n )⊗ℓ )SLn OK = 0 when n ≥ 2 + ℓ.
Fix an integral frame L = {L1 , . . . , Ln } with integral apartment class [AL ] ∈ Stn (K). We
next show that [AL ] ⊗ (K n )⊗ℓ vanishes in the quotient (Stn (K) ⊗ (K n )⊗ℓ )SLn OK if n ≥ 2 + ℓ.
n is a rank 1 projective O -module. Since
For each 1 ≤ i ≤ n, the intersection Li ∩ OK
K
n . The
|cl(OK )| = 1 this module is free, so we can choose a generator vi for each Li ∩ OK
assumption that L is an integral frame means precisely that {v1 , . . . , vn } is an OK -basis for
n . In particular {v , . . . , v } is a K-basis for K n , so S = {v ⊗ · · · ⊗ v | 1 ≤ i ≤ n} is a
OK
j
1
n
i1
iℓ
basis for (K n )⊗ℓ .
Consider an arbitrary basis vector s = vi1 ⊗ · · ·⊗ viℓ ∈ S. Under the assumption n ≥ 2+ ℓ,
we can find 1 ≤ j < j ′ ≤ n which are distinct from all of the i1 , . . . , iℓ . There exists a unique
g ∈ SLn OK such that
vj ′
if i = j,
g(vi ) = −vj if i = j ′ ,
vi
otherwise.
We have g(s) = s and g(L) = L by construction. However g exchanges the two lines Lj and
Lj ′ , so it reverses the orientation of the apartment AL , i.e. g([AL ]) = −[AL ]. This means
that [AL ] ⊗ s and −[AL ] ⊗ s become equal, and thus vanish, in the coinvariant quotient
(Stn (K) ⊗ (K n )⊗ℓ )SLn OK . Since s was arbitrary, [AL ] ⊗ (K n )⊗ℓ vanishes in this quotient
for every integral apartment AL . Theorem A states that Stn (K) is generated by integral
apartment classes, and so we conclude that (Stn (K) ⊗ (K n )⊗ℓ )SLn OK = 0, as desired.
Remark 4.1. Some condition on K beyond |cl(OK )| = 1 is necessary in Theorem C. Indeed,
for d < 0√squarefree let Od denote the ring of integers in the quadratic imaginary field
Kd = Q( d). Those d < 0 for which Od is non-Euclidean but satisfies |cl(Od )| = 1 are
exactly d ∈ {−19, −43, −67, −163}. Although H2 (SL2 O−19 ; Q) = 0, we have
H2 (SL2 (O−43 ); Q) = Q,
H2 (SL2 (O−67 ); Q) = Q2 ,
H2 (SL2 (O−163 ); Q) = Q6 .
For these calculations see Vogtmann [Vo] and Rahm [Ra]. Presumably, similar things occur
for SLn Od for n ≥ 3, but we could not find such calculations in the literature. It is likely
that these four fields provide the only such counterexamples to the conclusion of Theorem C;
indeed, Weinberger [We] proved that the generalized Riemann hypothesis implies that if OK
has class number 1 and infinitely many units, then OK is Euclidean.
24
5
Non-Vanishing and Non-Integrality Theorems
In this section we prove Theorem B and Theorem D; in fact we prove stronger versions of
these results in Theorems B′ and D′ , which apply to projective modules M that need not be
free. After summarizing in §5.1 some basic properties of spherical buildings, in §5.2 we prove
Proposition 5.5, the key technical theorem of this section. We apply this proposition in §5.3
to prove Theorems B′ and D′ .
5.1
Spherical buildings
In this section we briefly review the classical theory of spherical buildings. See [Br] for a
general reference.
A d-dimensional spherical building Y is a simplicial complex satisfying certain axioms.
Associated to Y is a finite Coxeter group W of rank d+1 called the Weyl group and a canonical
collection of subcomplexes A ⊂ Y called apartments. Each apartment is isomorphic to the
Coxeter complex of W and is thus acted upon by W . In particular, each apartment A is
e d (Y ). The class [A] is
homeomorphic to a sphere S d and defines a fundamental class [A] ∈ H
only defined up to ±1, but for our purposes this ambiguity will never pose a problem, so we
do not mention it again. A d-simplex C in an apartment A is called P
a chamber of A. The
apartment A is the union of the translates wC for w ∈ W , and [A] = w∈W (−1)w [wC].
e d (Y )
Any d-dimensional spherical building Y is CM of dimension d. The top homology H
is spanned by the fundamental classes [A] of apartments; indeed, for any fixed chamber C, a
e d (Y ) is given by the apartment classes [A] of all apartments A containing C.
basis for H
The following three examples will be of the most interest to us. A poset P is said to be
a d-dimensional spherical building if its realization |P | is.
Example 5.1 (Poset of subspaces). Given an n-dimensional vector space W over a field
K, let T (W ) be the poset of proper, nonzero K-subspaces of W . Its realization is the
spherical Tits building associated to GL(W ), an (n − 2)-dimensional spherical building with
Weyl group W = Sn . The chambers of T (W ) are the complete flags
0 ( W1 ( · · · ( Wn−1 ( W.
The Coxeter complex of Sn can be identified with the barycentric subdivision of ∂∆n−1 . As
discussed in the introduction, apartments of T (W ) are in bijection with frames of W , i.e. sets
of lines L = {L1 , . .P
. , Ln } such that W = L1 ⊕ · · · ⊕ Ln . The apartment AL corresponding to
L satisfies [AL ] = σ∈Sn (−1)σ Vσ , where Vσ is the chamber
Lσ(1) ( Lσ(1) + Lσ(2) ( · · · ( Lσ(1) + · · · + Lσ(n−1) .
(5.1)
Example 5.2 (Poset of summands). Given a finite rank projective O-module M over a
Dedekind domain O, let T (M ) denote the poset of proper nonzero O-submodules U ⊂ M
which are direct summands of M . Recall from Lemma 2.2(c) that the assignment V 7→ V ∩M
gives a bijection between the K-subspaces V ( M ⊗O K and the direct summands of the
O-module M . This assignment preserves containment, so it gives a canonical isomorphism
T (M ⊗O K) ∼
= T (M ). In particular, if M has rank n, then T (M ) is an (n − 2)-dimensional
spherical building. The apartments of T (M ) correspond to frames L = {L1 , . . . , Ln } of
M ⊗O K, or equivalently to sets I = {I1 , . . . , In } of rank 1 summands of M such that
I1 + · · · + In has rank n. The formula (5.1) might suggest that the chambers of AI are given
by e.g. I1 ( I1 + I2 ( · · · , but this is not the case: indeed I1 + I2 need not be a summand of
25
M at all! The corresponding summand is M ∩ ((I1 + I2 ) ⊗O K), the “saturation” of I1 + I2 .
For n ∈ N, let [n] = {1, . . . , n}, and given X ⊂ [n] set
X
UX := M ∩ (
Ix ⊗O K).
(5.2)
x∈X
The chambers of AI are then
Vσ :
Uσ([1]) ( Uσ([2]) ( · · · ( Uσ([n−1])
for σ ∈ Sn ,
σ
e n−2 (T (M ); Z) is given by [AI ] = P
and the apartment class [AI ] ∈ H
σ∈Sn (−1) Vσ . A frame
I = {I1 , . . .L
, In } is M -integral if I1 + · · · + In = M . In this case we have M = I1 ⊕ · · · ⊕ In
and UX = x∈X IX , and the chambers of the M -integral apartment AI are indeed given by
Iσ(1) ( Iσ(1) ⊕ Iσ(2) ( · · · ( Iσ(1) ⊕ · · · ⊕ Iσ(n−1) .
Example 5.3 (Join of discrete posets). Let T be a finite set. Define a poset Xm (T ) =
{1, . . . , m} × T by declaring (p, t) < (p′ , t′ ) exactly when p < p′ . The poset Xm (T ) is an
(m − 1)-dimensional spherical building with Coxeter group W = (Z/2)m . The chambers of
Xm (T ) are chains
(1, t1 ) < (2, t2 ) < . . . < (m, tm )
with ti ∈ T for 1 ≤ i ≤ m. We will denote this chamber by (t1 , . . . , tm ). The apartments
of Xm (T ) are in bijection with ordered sequences S = (S1 , . . . , Sm ), where Si = {ai , bi } is a
2-element subset of T with ai 6= bi . For e = (ε1 , . . . , εm ) ∈ (Z/2)m , let cek = bk if εk = 0
and cek = ak if εk = 1 (this convention will simplify some of our later formulas). Denote
the corresponding chamber by Ce = (ce1 , . . . , cem ). The apartment AS is the union of the 2m
chambers Ce , and its fundamental class is
P
X
[AS ] =
(−1)e [Ce ]
where (−1)e = (−1) i εi .
e
Xm (T ) is the m-fold join of the discrete set T , so it is homotopy equivalent to a wedge of
e m−1 (Xm (T ); Z) ∼
(|T | − 1)m copies of S m−1 . In particular, H
= ZN with N = (|T | − 1)m .
5.2
Cycles in T (M) and the class group cl(O)
Let M be a rank n projective O-module. In this section we relate the Tits building T (M )
to the much simpler complex Xn−1 (cl(O)) that we introduced in
Example 5.3. Define a map
ψ : T (M ) → Xn−1 (cl(O)) via the formula ψ(U ) = rk(U ), [U ] . Given U, U ′ ∈ T (M ) with
U ( U ′ , Lemma 2.2(b) says that U is a summand of U ′ , so rk(U ) < rk(U ′ ). Thus ψ is a
poset map. Elements g ∈ GL(M ) and U ∈ T (M ) determine an isomorphism g : U ∼
= g(U ),
so ψ(g(U )) = ψ(U ). Therefore ψ induces a map ψ : |T (M )| / GL(M ) → |Xn−1 (cl(O))|.
Proposition 5.4. Let M be a rank n projective O-module. The map ψ : |T (M )| / GL(M ) →
|Xn−1 (cl(O))| defined above is then an isomorphism of cell complexes.
Proof. The proposition is equivalent to the following two claims; the first implies that ψ is
surjective and the second implies that ψ is injective.
1. Let (j1 , c1 )
···
(jp , cp ) be a chain in Xn−1 (cl(O)). Then there exists a chain
U1 ( · · · ( Up in T (M ) such that ψ(Ui ) = (ji , ci ) for all 1 ≤ i ≤ p.
2. For ℓ = 1, 2 let V1ℓ ( · · · ( Vqℓ be a chain in T (M ). Assume that ψ(Vi1 ) = ψ(Vi2 ) for all
1 ≤ i ≤ q. Then there exists some g ∈ GL(M ) such that g(Vi1 ) = Vi2 for all 1 ≤ i ≤ q.
26
We begin with the first claim. Set (j0 , c0 ) = (0, 0) and (jp+1 , cp+1 ) = (n − jp , [M ]). For
bi be a projective O-module of rank ji with [U
bi ] = ci − ci−1 . Then
1 ≤ i ≤ p + 1, let U
b1 ⊕ · · · ⊕ U
bp+1 is a projective O-module of rank n satisfying
U
b1 ⊕ · · · ⊕ U
bp+1 ] = c1 + (c2 − c1 ) + · · · + (cp − cp−1 ) + ([M ] − cp ) = [M ],
[U
b1 ⊕ · · · ⊕ U
bp+1 → M . The desired summands Ui are then
so there exists an isomorphism η : U
b1 ) ⊕ · · · ⊕ η(U
bi ).
Ui := η(U
ℓ
= M.
We now turn to the second claim. For both ℓ = 0 and ℓ = 1, set V0ℓ = 0 and Vq+1
ℓ
ℓ
For 1 ≤ i ≤ q + 1, Lemma 2.2(b) implies that Vi−1 is a summand of Vi , so there exists some
ℓ ⊕V
b ℓ . We thus have V ℓ = Vb ℓ ⊕ · · · ⊕ Vb ℓ . Observe that
Vbiℓ ⊂ Viℓ such that Viℓ = Vi−1
1
i
i
i
1
2
[Vbi1 ] = [Vi1 ] − [Vi−1
] = ci − ci−1 = [Vi2 ] − [Vi−1
] = [Vbi2 ].
and similarly rk(Vbi1 ) = rk(Vbi2 ). Therefore there exists an isomorphism ζi : Vbi1 → Vbi2 . The
desired g ∈ GL(M ) is then the composition
1
M = Vb11 ⊕ · · · ⊕ Vbq+1
ζ1 ⊕···⊕ζq+1
−→
2
Vb12 ⊕ · · · ⊕ Vbq+1
= M.
We now turn to the main technical theorem of §5, which asserts that the map
e n−2 (T (M ); Z) → H
e n−2 (Xn−1 (cl(O)); Z)
ψ∗ : H
is surjective. We will use the notation for apartments and chambers in these buildings that
was introduced in Examples 5.2 and 5.3. The reason that this result is so hard to prove
is that for a general apartment AI in T (M ⊗O K), it seems quite difficult to describe the
image ψ∗ ([AI ]) of its fundamental class in the homology of Xn−1 (cl(O)). For one thing, we
saw in Example 5.2 that the summands UX that make up the chambers of AI are not easily
understood; for example, knowing the images ψ(Ik ) of the rank 1 summands does not allow
us to determine ψ(UX ). For another, the apartment AI consists of n! chambers, while the
apartments in Xn−1 (cl(O)) are made up of only 2n−1 chambers. Since n! ≫ 2n−1 , in general
we expect that ψ(AI ) will be draped over many apartments in Xn−1 (cl(O)). However, the
following proposition guarantees that we can find certain special apartments AI whose image
is just a single apartment AS , with the excess n! − 2n−1 chambers all “folded away” and
canceling each other out precisely, leaving only the desired homology class.
Proposition 5.5 (Detecting spaces of cycles). Let O be a Dedekind domain, let M
be a rank n projective O-module, and let AS be an apartment of Xn−1 (cl(O)). Then there
e n−2 (Xn−1 (cl(O)); Z). In particular,
exists a frame I of M such that ψ∗ ([AI ]) = [AS ] ∈ H
e n−2 (T (M ); Z) → H
e n−2 (Xn−1 (cl(O)); Z) is surjective.
ψ∗ : H
Proof. Write S = (S1 , . . . , Sn−1 ) with Si = {ai , bi } for some ai , bi ∈ cl(O). For e =
(ε1 , . . . , εn−1 ) ∈ (Z/2)n−1 , let cek = bk if εk = 0, and cek = ak if εk = 1 (as we noted in
Example 5.3, this convention will simplify some of our formulas). The apartment AS is the
union of the 2n−1 chambers Ce = (ce1 , . . . , cen−1 ), and its fundamental class is
P
X
[AS ] =
(−1)e [Ce ]
with (−1)e = (−1) i εi .
e
(i)
Using Proposition 5.4 we may choose summands Bk for 1 ≤ k ≤ n − 1 and Ak for
1 ≤ i ≤ k ≤ n − 1 with the following properties:
0 = B0 ( B1 ( · · · ( Bn−1 ( Bn = M
(i)
(i)
rk(Bk ) = k, [Bk ] = bk .
(i)
Bi−1 ( Ai ( Ai+1 ( · · · ( An−1 ( M
27
(i)
(i)
(i)
rk(Ak ) = k, [Ak ] = ak , Ak ⊂ Bk+1
(i)
(i)
(First, choose all the Bk . Then for each i, choose the Ak inductively using Bi−1 ( Ai (
(i)
(i)
Bi+1 and Ak−1 ( Ak ( Bk+1 .) We now define
I1 := B1
and Ik :=
k−1
\
(i)
Ak−1
i=1
for 1 < k ≤ n.
Set I = {I1 , . . . , In }. We will prove that I is a frame of M and that ψ∗ ([AI ]) = [AS ].
Claim 1. For each 1 ≤ k ≤ n the intersection Ik is a rank 1 summand of M . Also,
I1 + · · · + Ik is a rank k O-submodule of M (not necessarily a summand). In particular, the
rank of I1 + · · · In is n, so I is a frame of M .
Proof of Claim 1. The bijection V 7→ M ∩ V of Lemma 2.2(c) preserves intersections, so
the intersection of summands of M is also a summand of M . In particular, each Ik is a
Tk−1 (i)
Ak−1 ⊗O K. The intersection of the k − 1 hyperplanes
summand of M and Ik ⊗O K = i=1
(1)
(k−1)
Ak−1 ⊗O K, . . . , Ak−1 ⊗O K inside the k-dimensional vector space Bk ⊗O K cannot vanish,
so rk(Ik ) ≥ 1. We will prove equality holds, but not until the end of the proof.
(i)
Since each Ak−1 is contained in Bk , we have Ik ⊂ Bk . Since Bj ⊂ Bk for j < k, we see
that I1 + · · · + Ik ⊂ Bk for all k.
We next claim that Ik ∩ Bk−1 = 0 for all k. The proof will be by induction on k. The
base case k = 1 is trivial since I1 = B1 , so assume that k > 1. For 1 ≤ i ≤ k − 2, we have
(i)
(i)
(i)
[Ak−1 ] = ak−1 6= bk−1 = [Bk−1 ], so Ak−1 6= Bk−1 . It follows that Ak−1 ∩ Bk−1 is a summand
(i)
of rank at most k −2. But by construction, the rank k −2 summand Ak−2 is contained in both
(i)
(i)
(i)
Ak−1 and in Bk−1 , so it must be their intersection: Ak−1 ∩ Bk−1 = Ak−2 for 1 ≤ i ≤ k − 2.
(k−1)
In a similar way, we have Ak−1 ∩ Bk−1 = Bk−2 . Combining these, we find that
Ik ∩ Bk−1 =
=
k−1
\
k−1
\
(i)
Ak−1
i=1
(i)
Ak−2 ∩ Bk−2 = Ik−1 ∩ Bk−2 .
i=1
k−2
\
∩ Bk−1 =
i=1
(i)
(Ak−1 ∩ Bk−1 )
By induction, this is zero, as claimed.
Finally, we prove Ik has rank 1 and that I1 + · · · + Ik has rank k. The proof of this is by
induction on k. The base case k = 1 is trivial, so assume that k > 1, that Ik−1 has rank 1,
and that I1 + · · · + Ik−1 has rank k − 1. From the containments
I1 + · · · + Ik−1 ⊂ I1 + · · · + Ik ⊂ Bk
we see that k−1 ≤ rk(I1 +· · ·+Ik ) ≤ k. Assume for the sake of contradiction that rk(I1 +· · ·+
Ik ) < k or that rk(Ik ) > 1. In either case, the subspaces Ik ⊗O K and (I1 + · · · + Ik−1 ) ⊗O K
must intersect nontrivially inside the k-dimensional vector space Bk ⊗O K. Multiplying by
an appropriate denominator, it follows that (I1 + · · · + Ik−1 ) ∩ Ik 6= 0. However we proved
above that Bk−1 ∩ Ik = 0; since I1 + · · · + Ik−1 ⊂ Bk−1 , this is a contradiction. Therefore
rk(I1 + · · · + Ik ) = k and rk(Ik ) = 1, as desired. This concludes the proof of Claim 1.
It remains to prove that ψ∗ ([AI ]) = [AS ]. As explained in Example 5.2, the chambers of
the apartment AI are given by
Vσ :=
Uσ([1])
Uσ([2])
···
28
Uσ([n−1])
for σ ∈ Sn .
The image ψ(Vσ ) ⊂ ψ(AL ) ⊂ Xn−1 (cl(O)) is the chamber determined by the sequence
[Uσ([1]) ], [Uσ([2]) ], . . . , [Uσ([n−1]) ] .
For most subsets X ⊂ [n], we cannot say anything about the summand UX or its class [UX ],
but there are two exceptions.
Claim 2 (Trichotomy).
1. U[k] = Bk , so [U[k] ] = bk .
(i)
2. If X = [k + 1] \ {i} for some 1 ≤ i ≤ k, then UX = Ak , so [UX ] = ak .
3. For other subsets X ⊂ [n], we know nothing.
Proof of Claim 2. Case 1: We saw above that I1 + · · · + Ik ⊂ Bk , and Claim 1 states that
rk(I1 +· · ·+Ik ) = k, so (I1 +· · ·+Ik )⊗O K = Bk ⊗O K. Since U[k] = M ∩((I1 +· · ·+Ik )⊗O K)
by definition, this means that U[k] = M ∩ (Bk ⊗O K). But Bk is a summand of M , so
M ∩ (Bk ⊗O K) = Bk .
(i)
Case 2: We begin by observing that Ix ⊂ Ak whenever x ≤ k + 1 and i 6= x. For i < x
(i)
(i)
(i)
this follows from Ix ⊂ Ax−1 (by the definition of Ix ) and Ax−1 ⊂ Ak . For i > x, we use
(i)
(i)
instead that Ix ⊂ Bx ⊂ Bi−1 ⊂ Ai ⊂ Ak . Let X = [k + 1] \ {i} for some 1 ≤ i ≤ k. It
follows from Claim 1 that I1 + · · · + Ibi + · · · + Ik+1 has rank k, and the beginning of this
(i)
paragraph shows that this submodule is contained in the rank k summand Ak . Just as
(i)
above, this implies that UX = M ∩ ((I1 + · · · + Ibi + · · · + Ik+1 ) ⊗O K) coincides with Ak .
This concludes the proof of Claim 2.
We can analyze this trichotomy in terms of σ ∈ Sn , as follows. Given σ ∈ Sn , think of σ
as a function σ : [n] → [n], and define sσ : [n] → [n] by
sσ (k) = sup{σ(i) | 1 ≤ i ≤ k}.
The function sσ is monotone increasing and satisfies k ≤ sσ (k) ≤ n. If sσ (k) = k then
it must be that σ([k]) = [k], so [Uσ([k]) ] = [Bk ] = bk . If sσ (k) = k + 1, it must be that
(i)
σ([k]) = [k + 1] − {i} for some 1 ≤ i ≤ k, so [Uσ([k]) ] = [Ak ] = ak . To sum up, we have
if sσ (k) = k
bk
[Uσ([k]) ] = ak
if sσ (k) = k + 1
unknown otherwise
(5.3)
Accordingly, we divide the permutations σ ∈ Sn into two subsets: we say that σ ∈ Sn is good
if sσ (k) ∈ {k, k + 1} for all k, and say that σ is bad otherwise.
Claim 3. The good permutations are the 2n−1 permutations of the form
σe := (1 2)ε1 (2 3)ε2 · · · (n − 1 n)εn−1 ∈ Sn
for some e = (ε1 , . . . , εn−1 ) ∈ (Z/2)n−1 . For such a good permutation σ = σe ,
sσ (k) = k ⇐⇒ εk = 0
sσ (k) = k + 1 ⇐⇒ εk = 1
29
for k < n.
Proof of Claim 3. The proof is by induction on n. The base case is n = 1, where the claim
is trivial. Assume now that n ≥ 2 and that the claim is true for smaller n. Let σ ∈ Sn be
a good permutation. Since sσ (k) ≥ σ(k), the condition sσ (k) ≤ k + 1 for all k implies that
σ(k) 6= n for any k < n − 1. Therefore we have either σ(n) = n or σ(n − 1) = n. In the first
case, the restriction σ|[n−1] is a good permutation in Sn−1 , so the claim holds by induction
with εn−1 = 0; since σ(n) = n in this case, we must have sσ (n − 1) = n − 1. In the second
case, we have sσ (n − 1) = n. Moreover, the permutation σ ′ := σ(n − 1 n) satisfies σ ′ (n) = n
and is good (since σ ′ |[n−2] = σ|[n−2] ), so applying the first case to σ ′ proves Claim 3.
We can now divide the chambers of AI between the two chains defined by
X
X
[AI ]good :=
(−1)σ Vσ and [AI ]bad :=
(−1)σ Vσ ,
σ∈Sn good
σ∈Sn bad
so [AI ] = [AI ]good + [AI ]bad as chains in Cn−2 (T (M )).
Claim 4. We have ψ∗ ([AI ]good ) = [AS ].
Proof of Claim 4. Consider e = (ε1 , . . . , εn−1 ) ∈ (Z/2)n−1 . By Claim 3, sσe (k) = k + εk for
all 1 ≤ k < n. By (5.3) this implies [Uσe ([k])] = cek for all k (recall that cek is either bk or
ak depending on whether εk is 0 or 1). Therefore the image ψ(Vσe ) of the chamber Vσe is
precisely the chamber Ce . Since
P
(−1)σe = (−1)
εi
= (−1)e ,
the orientations of Vσe and Ce agree, proving Claim 4.
To deal with [AI ]bad , we define an involution σ 7→ σ on the set of bad permutations σ ∈ Sn
as follows. Given a bad permutation σ, define
xσ = sup{x ∈ {2, . . . , n − 1} | sσ (x − 1) > x},
yσ = inf{y ∈ {xσ + 1, . . . , n} | sσ (y) = y}.
There exists some x with sσ (x − 1) > x if and only if σ is bad, so xσ is well-defined. The
second condition is always satisfied by y = n, so yσ is well-defined as well. We define
σ := σ(xσ yσ ) ∈ Sn .
Claim 5. The map σ 7→ σ is an involution on the set of bad permutations.
Proof of Claim 5. Let σ ∈ Sn be a bad permutation. We will show that sσ = sσ ; since xσ
and yσ only depend on sσ , this implies that xσ = xσ and yσ = yσ and thus that σ = σ.
For k < xσ or k ≥ yσ we have σ([k]) = σ([k]), so sσ (k) = sσ (k) automatically. Considering
the remaining terms we find that
for xσ ≤ k < yσ ,
we must have sσ (k) = k + 1.
(5.4)
Indeed, to have sσ (k) > k + 1 would contradict the definition of xσ , and to have sσ (k) = k
would contradict the definition of yσ . It follows that σ(k) = k + 1 for xσ < k < yσ . By
the pigeonhole principle, σ(yσ ) ≤ xσ + 1. But if σ(yσ ) were xσ + 1 we could not have
sσ (xσ ) = xσ + 1 = sσ (xσ − 1), so in fact σ(yσ ) ≤ xσ . For the same reason, we must have
σ(xσ ) ≤ xσ .
Since σ(xσ ) ≤ xσ and σ(yσ ) ≤ xσ while sup σ([k]) = sσ (k) = k + 1 > xσ for xσ ≤ k < yσ ,
we see that neither σ(xσ ) nor σ(yσ ) ever realizes this supremum. Therefore exchanging σ(xσ )
and σ(yσ ) will not change sup σ([k]) for any k. This shows that sσ = sσ , proving Claim 5.
30
Claim 6. We have ψ∗ ([AI ]bad ) = 0.
Proof of Claim 6. Let σ ∈ Sn be a bad permutation. By definition, we have (−1)σ = (−1) ·
(−1)σ , so it is enough to prove that that the chambers ψ(Vσ ) and ψ(Vσ ) coincide. In other
words, we must show that [Uσ([k]) ] = [Uσ([k]) ] for all 1 ≤ k ≤ n − 1. For all k < xσ or k ≥ yσ ,
we have (xσ yσ ) [k]) = [k], so σ([k]) = σ([k]) and the claim is automatic. For the remaining
xσ ≤ k < yσ , we showed in (5.4) that sσ (k) = sσ (k) = k + 1. By (5.3), for such k we have
[Uσ([k]) ] = ak = [Uσ([k]) ]. This concludes the proof of Claim 6.
By definition [AI ] = [AI ]good + [AI ]bad , and Claims 4 and 6 state that
ψ∗ ([AI ]good ) = [AS ] and ψ∗ ([AI ]bad ) = 0,
so ψ∗ ([AI ]) = [AS ], as desired. This concludes the proof of Proposition 5.5.
5.3
Proof of the Non-Vanishing and Non-Integrality Theorems
In this section we prove Theorems D and B. In fact, we prove the stronger Theorems D′
and B′ , which apply to the automorphism groups of arbitrary (not necessarily free) finite
rank projective modules. If M is a rank n projective OK -module, its automorphism group
GL(M ) is also a lattice in GLn (OK ⊗ R), commensurable with but not in general isomorphic
to GLn OK . As before, νn = vcd(GLn OK ) and νn′ = vcd(SLn OK ).
Theorem D′ . Let K be a number field with ring of integers OK and let M be a rank n
projective OK -module. Then for all n ≥ 2,
′
dim Hνn (SL(M ); Q) ≥ dim Hνn (GL(M ); Q) ≥ (|cl(OK )| − 1)n−1 .
n.
Theorem D is the special case of Theorem D′ with M ∼
= OK
e n−2 (T (M ); Z). The automorphism group GL(M ) acts naturally
Proof. Define St(M ) := H
on T (M ), and thus on St(M ). Since SL(M ) is commensurable with SLn OK , it follows from
[BE, Theorems 3.2, 3.3] that GL(M ) and SL(M ) are virtual duality groups with rational
dualizing module St(M ). Moreover, vcd(GL(M )) = vcd(GLn OK ) = νn and vcd(SL(M )) =
vcd(SLn OK ) = νn′ . Therefore just as in Equation (1.1), we have
Hνn (GL(M ); Q) ∼
= (St(M ) ⊗ Q)GL(M ) ,
′
Hνn (SL(M ); Q) ∼
= (St(M ) ⊗ Q)SL(M ) .
In particular, the former is a quotient of the latter, so this already proves the first inequality
of the theorem.
Consider the map ψ : T (M ) → Xn−1 (cl(O)) from the previous section, which is invariant
under the action of GL(M ) on T (M ). Recall from Example 5.3 that the realization of
Xn−1 (cl(O)) is homotopy equivalent to a (|cl(O)| − 1)n−2 -fold wedge of (n − 2)-spheres, so
e n−2 (Xn−1 (cl(O)); Z) ∼
H
= ZN with N = (|cl(OK )| − 1)n−2 . Proposition 5.5 states that the
map
e n−2 (T (M ); Z) → H
e n−2 (Xn−1 (cl(O)); Z) ∼
ψ∗ : St(M ) = H
= ZN
is surjective. Since ψ is GL(M )-invariant, this map factors through the coinvariants, yielding
a surjection St(M )GL(M ) ։ ZN . Tensoring with Q yields a surjection
St(M )GL(M ) ⊗ Q ∼
= (St(M ) ⊗ Q)GL(M ) ։ QN ,
so the dimension of (St(M ) ⊗ Q)GL(M ) is at least (|cl(OK )| − 1)n−2 , as claimed.
31
Example 5.6. The cohomology classes we construct in Theorem D√were known classically
when n = 2. To illustrate this, consider a quadratic imaginary field Q( −d) as in Remark 4.1.
The Bianchi group SL2 Od is a lattice in SL2 C, and thus acts by isometries on H3 . The associated locally symmetric space Xd = SL2 Od \H3 is a noncompact arithmetic 3-dimensional
hyperbolic orbifold of cohomological dimension 2. The cusps of Xd are in bijection with the
SL2 Od -conjugacy
classes of parabolic subgroups in SL2 Od , and thus with the SL2 Od -orbits
√
1
of P (Q( −d)), and thus with the ideal class group cl(Od ). An embedded path in Xd connecting one cusp to another defines an element of the locally finite homology group Hlf1 (Xd ; Q),
which is dual to H2 (X; Q) ∼
= H2 (SL2 Od ; Q). Since there are |cl(Od )| cusps, intersecting with
such paths gives a (|cl(Od )| − 1)-dimensional projection of H2 (X; Q) ∼
= H2 (SL2 Od ; Q). A
similar procedure works for SL2 OK for any number field K. However the case n ≥ 3 is more
complicated: the cusps overlap in complicated ways, so this simple argument does not work.
We now give a strengthening of Theorem B; Theorem B is the special case when M ∼
= On .
Theorem B′ . Let O be a Dedekind domain with 1 < |cl(O)| < ∞, and let M be a rank n
projective O-module for some n ≥ 2. If |cl(O)| = 2, assume that M is not the unique non-free
O-module of rank 2. Then St(M ) is not generated by M -integral apartment classes.
Proof. Let StInt (M ) be the subspace of St(M ) spanned by M -integral apartment classes.
e n−2 (Xn (cl(O)); Z) is surjective. Thus to prove
By Proposition 5.5, the map ψ∗ : St(M ) → H
Int
that St (M ) 6= St(M ) it is enough to prove that ψ∗ (StInt (M )) is a proper subspace of
e n−2 (Xn (cl(O)); Z).
H
Consider an M -integral frame I = {I1 , . . . , In } with [Ik ] = ck and M = I1 ⊕ · · · ⊕ In ; the
latter property implies that c1 + · · · + cn = [M ] ∈ cl(O). We claim first that ψ∗ ([AI ]) only
depends on {c1 , . . . , cn }, and not on the particular integral frame I. To see this, recall from
Example 5.2 that the chambers of an integral apartment AI are of the form
Vσ = Iσ(1) ( Iσ(1) ⊕ Iσ(2) ( · · · ( Iσ(1) ⊕ · · · ⊕ Iσ(n−1)
L
P
for σ ∈ Sn . Given X ⊂ [n], we have [ k∈X Ik ] = k∈X ck , so the chamber Vσ is taken by π
to the chamber
Cσ = cσ(1) , cσ(1) + cσ(2) , . . . , cσ(1) + · · · + cσ(n−1) .
(5.5)
The claim follows.
Next, assume for the moment that ci = cj for some i 6= j. Define the involution σ 7→ σ
b
on Sn by σ
b = (i j)σ. From (5.5) we see that Cσb will coincide with Cσ for all σ ∈ Sn . Since
(−1)σb = (−1) · (−1)σ , the chamber Vσb occurs with opposite orientation from Vσ in [AI ]. We
conclude that ψ∗ ([AI ]) = 0 unless the classes c1 , . . . , cn ∈ cl(O) are distinct, in which case
(up to sign) ψ∗ ([AI ]) only depends on the unordered set {c1 , . . . , cn } ⊂ cl(O). This already
shows that when n > |cl(O)|, the image ψ∗ (StInt (M )) vanishes.
We complete the proof by an elementary counting argument. By Example 5.3, the
e n−2 (Xn (cl(O)); Z) is (|cl(O)| − 1)n−1 . The precise number of n-element subsets
rank of H
{c1 , . . . , cn } ⊂ cl(O) satisfying c1 + · · · + cn = [M ] is hard to calculate and depends on
the group structure of cl(O). One case is easy, however: when cl(O) = Z/2Z, the relation c1 + · · · + cn = [M ] can never be satisfied by distinct elements ci ∈ cl(O) for any
n ≥ 2, excepting the single case n = 2 and [M ] 6= 0. Therefore, whenever |cl(O)| = 2
(excepting the excluded case when n = 2 and [M ] 6= 0) we have ψ∗ (StInt (M )) = 0. Since
e n−2 (Xn (cl(O)); Z) = Z when |cl(O)| = 2, this concludes the proof in that case.
H
For all other nontrivial class groups, the relation c1 + · · · + cn = [M ] implies that cn
is determined by {c1 , . . . , cn−1 }. This lets us bound the rank of ψ∗ (StInt (M )) by |cl(O)|
n−1 .
32
Elementary estimates show that
n−1
(|cl(O)| − 1)
>
|cl(O)|
n−1
for all n ≥ 2 whenever |cl(O)| > 2. Therefore ψ∗ (StInt (M )) must be a proper subspace of
e n−2 (Xn (cl(O)); Z), as desired.
H
In the excluded case when cl(O) ∼
= Z/2Z, let I ⊂ O be a non-principal ideal, so that
M ∼
= O ⊕ I. It can be verified as in Proposition 2.7 that St(M ) is spanned by M -integral
apartment classes if and only if the group
SL(M
)
is generated
by diagonal matrices together
1
I
1
0
and
. It would be interesting to know
with elementary matrices of the form
0 1
I −1 1
when this condition holds for SL(M ); we are not aware of any results on this subject.
References
[AR]
A. Ash and L. Rudolph, The modular symbol and continued fractions in higher dimensions, Invent.
Math. 55 (1979), no. 3, 241–250. doi:10.1007/bf01406842.
[BMS]
H. Bass, J. Milnor, and J.-P. Serre, Solution of the congruence subgroup problem for SLn (n ≥ 3) and
Sp2n (n ≥ 2), Inst. Hautes Études Sci. Publ. Math. No. 33 (1967), 59–137. doi:10.1007/bf02684586.
[Be]
M. Bestvina, PL Morse theory, Math. Commun. 13 (2008), no. 2, 149–162. Available at
hrcak.srce.hr/file/48930
[BE]
R. Bieri and B. Eckmann, Groups with homological duality generalizing Poincaré duality, Invent.
Math. 20 (1973), 103–124. doi:10.1007/bf01404060.
[Bo]
A. Borel, Stable real cohomology of arithmetic groups, Ann. Sci. École Norm. Sup. (4) 7 (1974),
235–272. Available at www.numdam.org/item?id=ASENS_1974_4_7_2_235_0
[BS]
A. Borel and J.-P. Serre, Corners and arithmetic groups, Comment. Math. Helv. 48 (1973), 436–491.
doi:10.1007/bf02566134.
[Br]
K. S. Brown, Buildings, Springer, New York, 1989. Freely available at
http://www.math.cornell.edu/~ kbrown/buildings/
[C]
P. M. Cohn, On the structure of the GL2 of a ring, Inst. Hautes Études Sci. Publ. Math. No. 30
(1966), 5–53. doi:10.1007/bf02684355.
[DP]
M. Day and A. Putman, The complex of partial bases for Fn and finite generation of the
Torelli subgroup of Aut(Fn ), Geom. Dedicata 164 (2013), 139–153. doi:10.1007/s10711-012-9765-6.
arXiv:1012.1914.
[F1]
J. Franke, Harmonic analysis in weighted L2 -spaces, Ann. Sci. École Norm. Sup. (4) 31 (1998), no.
2, 181–279. doi:10.1016/s0012-9593(98)80015-3.
[F2]
J. Franke, A topological model for some summand of the Eisenstein cohomology of congruence subgroups, in Eisenstein series and applications, 27–85, Progr. Math., 258, Birkhäuser Boston, Boston,
MA, 2008. doi:10.1007/978-0-8176-4639-4 2.
[FS]
J. Franke and J. Schwermer, A decomposition of spaces of automorphic forms, and the Eisenstein
cohomology of arithmetic groups, Math. Ann. 311 (1998), no. 4, 765–790. doi:10.1007/s002080050208.
[G]
S. Geller, On the GEn of a ring, Illinois J. Math. 21 (1977), no. 1, 109–112.
projecteuclid.org/euclid.ijm/1256049506
[LS]
R. Lee and R. H. Szczarba, On the homology and cohomology of congruence subgroups, Invent.
Math. 33 (1976), no. 1, 15–53. doi:10.1007/bf01425503.
[L]
B. Liehl, On the group SL2 over orders of arithmetic type, J. Reine Angew. Math. 323 (1981),
153–171. doi:10.1515/crll.1981.323.153.
33
[vdK]
W. van der Kallen, Homology stability for linear groups, Invent. Math. 60 (1980), no. 3, 269–295.
doi:10.1007/bf01390018.
[vdKL]
W. van der Kallen and E. Looijenga, Spherical complexes attached to symplectic lattices, Geom.
Dedicata 152 (2011), 197–211. doi:10.1007/s10711-010-9553-0. arXiv:1001.0883
[Ma]
H. Maazen, Homology Stability for the General Linear Group, thesis, University of Utrecht (1979).
Available at www.persistent-identifier.nl/?identifier=URN:NBN:NL:UI:10-1874-237657
[Mi]
J. Milnor, Introduction to algebraic K-theory, Annals of Mathematics Studies 72, Princeton University Press, Princeton, N.J.; University of Tokyo Press, Tokyo, 1971. doi:10.1515/9781400881796.
[Q]
D. Quillen, Homotopy properties of the poset of nontrivial p-subgroups of a group, Adv. in Math. 28
(1978), no. 2, 101–128. doi:10.1016/0001-8708(78)90058-0.
[Ra]
A. D. Rahm, The homological torsion of PSL2 of the imaginary quadratic integers, Trans. Amer.
Math. Soc. 365 (2013), no. 3, 1603–1635. doi:10.1090/s0002-9947-2012-05690-x. arXiv:1108.4608.
[Re]
I. Reiner, Unimodular complements, Amer. Math. Monthly 63 (1956), 246–247. doi:10.2307/2310351.
[Se]
J.-P. Serre, Trees, translated from the French by John Stillwell, Springer, Berlin, 1980.
doi:10.1007/978-3-642-61856-7.
[So]
L. Solomon, The Steinberg character of a finite group with BN -pair, in Theory of Finite Groups
(Symposium, Harvard Univ., Cambridge, Mass., 1968), 213–221, Benjamin, New York.
[Va]
L. N. Vaseršteı̆n, On the group SL2 over Dedekind rings of arithmetic type, Math. USSR Sbornik 18
(1972) 2, 321–332. doi:10.1070/SM1972v018n02ABEH001775.
[Vo]
K. Vogtmann, Rational homology of Bianchi groups, Math. Ann. 272 (1985), no. 3, 399–419.
doi:10.1007/bf01455567.
[Wa]
J. B. Wagoner, Stability for homology of the general linear group of a local ring, Topology 15 (1976),
no. 4, 417–423. doi:10.1016/0040-9383(76)90035-5.
[We]
P. J. Weinberger, On Euclidean rings of algebraic integers, in Analytic number theory (Proc. Sympos. Pure Math., Vol. XXIV, St. Louis Univ., St. Louis, Mo., 1972), 321–332, Amer. Math. Soc.,
Providence, RI. doi:10.1090/pspum/024/0337902.
Thomas Church
Department of Mathematics
Stanford University
450 Serra Mall
Stanford, CA 94305
[email protected]
Benson Farb
Department of Mathematics
University of Chicago
5734 University Ave.
Chicago, IL 60637
[email protected]
34
Andrew Putman
Department of Mathematics
University of Notre Dame
279 Hurley Bldg
Notre Dame, IN 46556
[email protected]
| 4math.GR
|
Revisiting Exception Handling Practices with
Exception Flow Analysis
Guilherme B. de Pádua
Weiyi Shang
Department of Computer Science and Software Engineering Department of Computer Science and Software Engineering
Concordia University - Montreal, QC, Canada
Concordia University - Montreal, QC, Canada
Email: g [email protected]
Email: [email protected]
Prior studies on exception handling aim to understand the
practices of exception handling in its different components: exception sources and handling code. Yet, the observed findings
were scattered and diverse. Recent empirical studies on exception handling practices have advocated the suboptimal use of
exception handling features in open source software [9], [10],
[11], [12]. Moreover, in our previous research, we observe the
prevalence of exception handling anti-patterns. These research
findings imply the lack of a thorough understanding of the
practice of exception handling [13].
Abstract—Modern programming languages, such as Java and
C#, typically provide features that handle exceptions. These
features separate error-handling code from regular source code
and aim to assist in the practice of software comprehension and
maintenance. Having acknowledged the advantages of exception
handling features, their misuse can still cause reliability degradation or even catastrophic software failures. Prior studies on
exception handling aim to understand the practices of exception
handling in its different components, such as the origin of the
exceptions and the handling code of the exceptions. Yet, the
observed findings were scattered and diverse. In this paper, to
complement prior research findings on exception handling, we
study its features by enriching the knowledge of handling code
with a flow analysis of exceptions. Our case study is conducted
with over 10K exception handling blocks, and over 77K related
exception flows from 16 open-source Java and C# (.NET) libraries
and applications. Our case study results show that each try block
has up to 12 possible potentially recoverable yet propagated
exceptions. More importantly, 22% of the distinct possible
exceptions can be traced back to multiple methods (average
of 1.39 and max of 34). Such results highlight the additional
challenge of composing quality exception handling code. To make
it worse, we confirm that there is a lack of documentation of
the possible exceptions and their sources. However, such critical
information can be identified by exception flow analysis on welldocumented API calls (e.g., JRE and .NET documentation).
Finally, we observe different strategies in exception handling code
between Java and C#. Our findings highlight the opportunities
of leveraging automated software analysis to assist in exception
handling practices and signify the need of more further in-depth
studies on exception handling practice.
Index Terms—source code analysis; exception flow analysis;
exception handling; software engineering
Therefore, in this paper, we re-visit exception handling
practices by conducting an in-depth study on 16 open-source
Java and C# libraries and applications. To understand and
analyze the state-of-the-practice of exception handling in these
projects, we perform source code analysis to track the flow
of exceptions from the source of exceptions, through method
invocations, to the attempting blocks of exceptions (try block)
and the exception handling block (catch block). With such
flow analysis, we extract information about exception handling
practice in over 10K exception handling blocks, and over 77K
related exception flows from the studied subject systems.
Our case study focuses on four aspects of the exception
handling practices: 1) the quantity of exceptions, 2) the
diversity of exceptions, 3) the sources of exceptions and 4) the
exception handling strategies and actions. Table I summarizes
our findings and their corresponding implications. Such results
confirm the challenge of composing quality exception handling
code. For example, we find a considerable amount of potentially recoverable yet propagated exceptions. However, more
importantly, we highlight the opportunities of leveraging our
automated source code analysis to complement the information
that is valuable for developers when handling exceptions. More
in-depth analyses are needed to ensure and improve the quality
and usefulness of exception handling in practice.
I. I NTRODUCTION
Modern programming languages, such as Java and C#,
typically provide exception handling features, such as throw
statements and try-catch-finally blocks. These features separate
error-handling code from regular source code and are leveraged widely in practice to support software comprehension
and maintenance [1], [2].
Having acknowledged the advantages of exception handling
features, their misuse can still cause catastrophic software
failures, such as application crashes [3], or reliability degradation, such as information leakage [4], [5]. A large portion
of systems has suffered from system crashes that were due
to exceptions [6]. Additionally, the importance of exception
handling source code has been illustrated in prior research
and surveys [7], [8].
The rest of the paper is organized as follows: Section II
discusses the related prior research of this work. Section III
presents the methodology of the exception flow analysis
through an illustrative example and our case study setup.
Section IV to VII presents the results of our case study. Section VIII discusses the threats to the validity of our findings.
Finally, Section IX concludes the paper and discusses potential
future research directions based on our research results.
1
TABLE I
O UR FINDINGS AND IMPLICATIONS ON EXCEPTION HANDLING PRACTICES .
Quantity of Exceptions (Section IV)
Implications
(1) There often exist multiple possible exceptions in each try block,
and, out of those, many are propagated.
Current state-of-the-practice may not provide information to developers about all possible exceptions. Automated techniques may
help developers be aware of all possible exceptions to make
exception handling decisions.
(2) There exists a considerable amount of potentially recoverable
exceptions that are propagated, even though they are recommended
to be handled by Java and C#.
Exception flow analysis can provide automated tooling support
to alert developers about not handling potentially recoverable
exceptions.
Diversity of Exceptions (Section V)
Implications
(3) With a significant amount of exceptions existing in each project,
many possible exception types appear in only one try block.
Developers may not need to be aware of all exception types in a
project by receiving automated suggestions of the exceptions that
he/she needs to understand.
Sources of Exceptions (Section VI)
Implications
(4) Over 22% of the exceptions are traced from different methods.
Automated tools are needed to help developers understand the
source of the exception if it is traced back to different methods.
(5) The libraries used by the systems can provide documentation
to most of the possible exceptions.
Developers should leverage automated analyses to understand possible exceptions.
Exception Handling Strategies and Actions (Section VII)
Implications
(6) Only a small portion of the exceptions are handled with the
Specific strategy.
Developers should be guided to prioritize on handling exceptions
with the Specific strategy, since developers cannot optimize the handling of the exception without knowing its exact type information.
(7) Java and C# have differences in leveraging various actions when
handling exceptions.
More in-depth analysis and user studies are needed to further
understand the rationale of differences of Java and C# exception
handling practices.
(8) Actions that are taken when handling exceptions with specific
or subsumption manners are not statistically significantly different.
Research and tooling support are needed to guide how to handle
exceptions, especially with the specific strategy.
(9) With statistical significance, all top 10 Java and 2 out of top
10 C# exceptions have at least one action that is taken differently
from the rest of the exceptions.
Developers may consider leveraging automated suggestion of exception handling actions.
II. R ELATED W ORK
A. Empirical studies on exception handling practices
Prior research studied exception handling based on source
code and issue trackers. Cabral and Marques [14] studied
exception handling practices from 32 projects in both Java and
.Net without considering the flow of exceptions. Prior work by
Jo et al. [15] focuses on uncaught exceptions of Java Checked
exceptions. They proposed an inter-procedural analysis based
on set-based framework without using declared exceptions.
Coelho et al. [16] assessed exception handling strategy
with exception flows from Aspect-oriented systems and objectoriented systems. They evaluated the number of uncaught
exceptions, exceptions caught by subsumption, and exceptions
caught with specialized handlers.
Sena et al. [17] investigated sampled exception flows from
656 Java libraries for flow characteristics, handler actions, and
handler strategies. We extend their work by looking into a
higher number of flows per system (e.g. in Apache ANT we
identified 930 catch blocks, compared to 2), by considering
applications besides libraries and including C# .NET systems.
Some studies reveal that developers consider exception
handling hard to learn and to use and tend to avoid it or
misuse it [9], [10], [11]. Bonifacio et al. [12] also surveyed
C++ developers encountering revelations of educational issues.
It also has been noted that there is a lack of documentation
of exceptions. Kechagia and Spinellis [18] found that 69%
of the methods had undocumented exceptions and 19% of
crashes could have been caused by insufficient documentation.
Sena et al. [17]’s findings confirm that API runtime exceptions
are poorly documented. Cabral and Marques [19] identify that
infrastructure (20%) and libraries (15%) have better exception
handling documentation when compared to applications (2%).
Significant research aimed to indicate exception handling
problems and their impacts. Sinha et al. [20] leveraged exception flow analyses to study the existence of 11 anti-patterns
in four Java systems. Other research [2], [8], [7] classified
exception-handling related bugs by mining software issue
tracking. In our previous research, we observe the prevalence
of exception handling anti-patterns [13].
Cacho et al. [21], [22] studied the evolution of the behavior
of exception handling in Java and C# source code changes.
Their results highlight the impact of the programming language design differences in the maintenance and robustness
of exception handling mechanisms.
Our study revisits and combines different aspects of the
studies mentioned above. Moreover, we present new findings
(see Table I) that are not yet highlighted in prior research.
2
They do analysis in call stack levels per caught exception, but
not at the try block level. It cannot be told what is the level
in a common scenario.
Although Coelho et al. [9] considered the exception flows
in Aspect oriented systems, they introduced measurements that
we extended in our work. Similar measurements were applied
at the try handling
block perspective,
and in different programming
B. Improving exception
practices
languages.
Robillard and Murphy
[23]Murphy
created
a toola tool
to analyze
exRobillard and
[10] created
to analyse exception flows
in Java programs
at at point
of the system.
Ourinwork
ception flows in Java
programs,
including
a graphical
user
is similar in terms of approach since we also use AST and call
terface. Similarly,graph
Garcia
and Cacho
proposed
differentthe
navigation.
We also [24]
share the
ability ofaidentifying
approach for .NET
related
languages.
Garcia
Cacho’s
tool
flows
at any point
of the program.
Weand
complement
their
work
by including
other different
analysis
and including
different
supports visualization
of metrics
over the
application
history.
projects and programming languages in our datasets. They
To support thestudied
software
development
lifecycle,
Sinha et
subsumption
but did not relate
to the actions.
al. [20] provided automated
for [11]
development,
mainteA tool calledsupport
eFlowMining
uses a different
approach
.NET related languages,
and exception
focus on thehandling.
evolution of
nance and testingforrequirements
related to
exception handling code. They applied their tool in smaller
To improve how
developers would deal with exception
systems, with a maximum of 45 try blocks, using different
handling complexity,
Kechagia et al. [25] discuss and propose
measurements.
Java control
inappropriate
coding
patterns
improvements in the
designflows,
of including
exception
handling
mechawere evaluated in [12]. They also provide schematic views of
nisms. Zhang andpossible
Krintzexception
[26] propose
an as-if-serial exception
types at any point in the program, however
handling mechanism
for parallel
programming.
they don’t
present any
relation with documentation or data
at the
try block level.
The burden ofanalysis
writing
exception
handling code has been
pointed out by Cabral and Marques [27]. They showed that a
system with an automated set of recovery actions is capable
of achieving better error resilience than a traditional system.
Barbosa et al. developed strategies with heuristics for
recommending exception handling code as a semi-automated
approach. Zhu et al. [28] proposed an approach that suggests
logging decisions for exception handling.
By conducting an in-depth study on 16 open source projects,
our findings illustrate the opportunities of leveraging various
analysis to combine information from different sources to understand and assist in exception handling flows and practices.
Our results are valuable to complement and assist in improving
existing exception handling techniques.
A requires
execute
a method
other
means,to has
the knowledge
otherissues:
means,
has thea knowledge
two
1) having
invalid pa
two issues:instead
1) having
a invalid pa
Therefore,
of executing
a
Therefore,
instead
executing
possibly
throw
two of
types
of excea
possibly
throw twowhich
types of
exce
and
IOException,
corresp
and with
IOException,
deal
the two which
possiblecorres
exce
deal needs
with the
two possible
exce
oper
to either
handle the
oper needsactions
to either
handle
alternative
when
suchthe
exc
alternative
actions
exc
the
exception
such when
that asuch
differe
source of exception 2: method declaration
the exception
suchexample,
that a differe
exception.
In our
devel
public
void by
A()
IOException
is propagated
A, throws
that is IOException,
represented by the InvalidPathException
exception. In our example,
deve
exception
a
{
is propagated
by
A,arrow
that is
IOException,
represented by the InvalidPathException
bottom
double
line
in
Figure
??.
exception a
2) Raising
method
call and propagating ex
try
bottom
double line
arrowcomplements:
in Figure ??. raising, documenting velopers
2) Handling
practice
2) Raising
andknowledge
propagating
have the
thatex
{
exception
Handling
raising,
documenting
and2)propagating
practice:complements:
We mentioned
that the
developer sponding
velopers
have
thepossible
knowledge
that
B();practice
to two
exceptio
flow(s)
and propagating
Wetwo
mentioned
that
the However,
developer be
}
already
know that practice:
B can throw
exception
types.
sponding
to tworaised
possible
exceptio
either newly
or propaga
catch
(InvalidPathException
already
know that
B can
two exception
types.
source
ofnewly
such
exception
types
are throw
not always
so clearex)
at the However,
develop- Bbecalls
either
raisedCorand
propaga
(e.g.,
method
the m
{
such moment.
exception
types are Bnot
always
so clear at
thefunctions
develop- can
ment
Moreover,
could
be
executing
other
B exception
calls
methodblocks
C and to
then
use (e.g.,
the throws
e.printStackTrace();
mentare
moment.
Moreover,
B could be executing other functions our
that
the issues.
can example,
use the throws
blocks
to IO
n
developers
raise
}propagating
documentation
that
issues. consult the source code of B block
In}are
ourpropagating
example, if the
a developer
our example,
developers
raise
I
in the method
C. Such
new
In our??)
example,
if aidentify
developer
the source
code of B propagated
}
(listing
he would
the consult
propagation
of IOException
block in thetomethod
Such new
all the C.
methods
tha
(listing
??) heblocks.
wouldHe
identify
IOException
in
the throws
wouldthe
alsopropagation
discover theofcalls
to C and propagated
to all the
methods thI
3) Documenting
exceptions:
Listing 1. An exception handling example.
in the throws
blocks.
He would
alsoconsulted.
discover
the
to C and can3)consult
IOException
getPath.
These
methods
are then
C iscalls
represented
Documenting
exceptions:
the source
code of B
getPath.
These
methodsthat
areBthen
is represented
InvalidPathException
in
listing
??.exceptions
can consult the
source
code
of B
in the
throw
stateme
The
two
can consulted.
throw
are Cexception
flows IOException
in listing
IOException
in the throw
always available
for a statem
metho
that
arrive??.
at A. Figure ?? represents such flows. One exception not
public void B() throws IOException
not always
available
for aBmetho
source of exception
4:
getPath
called
in method
is de
flow
A, that
is InvalidPathException,
represented
{is handled
public
void at
B()
throws
IOException
getPath called
in method
B is d
external
Developers
would
need to consul
by the
top single line arrow in Figure ??. One exception
flowdocumentation
{ C();
Developersofwould
need totoconsu
fileSystem.getPath(invalidPath);
JavaDoc)
the method
disc
C();
getPath can throw
} fileSystem.getPath(invalidPath);
JavaDoc)InvalidPathException.
of the method to disc
possible
InvalidPathException
}
possible
InvalidPathException.
There could
be cases when an
Listing 2. An exception handling example. B propagates two exceptions from
two
different
couldin be
cases when anT
Listing
2. Anmethods.
exception handling example. B IOException
propagates two exceptions from notThere
available
documentation.
two different methods.
not available
in (like
documentation.
ming
languages
Java), a poss
/**
ming
languages
(like Java),
a pos
the
method
declaration.
For exam
/***
source of exception 3:
the possible
method declaration.
Forisexam
the
exception that
prop
* @throws IOException documentation in comment
the possible
exception
that is prop
method
declaration.
However,
som
Method
called when disk failed.
IOException
* @throws
method declaration.
However, so
Method called when disk failed.
undeclared
(like InvalidPathExce
*
*/
undeclared (like InvalidPathExce
B. Exception flow analysis
public
static void C() throws IOException
*/
{
B.InException
flow analysis
public
static void C() throws IOException
the previous
subsection, w
a function called A. This function requires to execute the
function B. However, the developer, by other means, already
know that B can potentially face two issues and, in that case,
instead of executing as expected, B would possibly throw two
exception types. To deal with the possible exception types the
developer needs to either determine the alternative actions that
A will execute or propagate the exception so that a different
method IOException
would take care
it. The
listing
Fig. 1.ofFlows
related
to A.?? depicts the above
scenario.
Fig. 1. Flows related to A.
{
}
}
throw new IOException();
In the previous subsection, w
throw new IOException(); source of exception 1: a exception handling scenario w
a exception
handling
w
section,
we present
our scenario
methodolo
throw statement
Listing 3. An exception handling example. C raises and propagates one
exception.
Listing 3. An exception handling example. C raises and propagates one
Fig. 1. An illustrative example.
exception.
III. M ETHODOLOGY
III. M ETHODOLOGY
section,
we exceptions
present ourand
methodol
all
possible
their fl
all possible
exceptions
Our
automated
tool is and
builttheir
by lfl
Our.NET
automated
toolPlatform
is built by
and
Compiler
(“R
2) Raising and propagating exceptions: As mentioned,
the respectively.
andsource
.NET Compiler
Platform (“R
C#
code,
As
In this section, we present the methodology of our study. In C# source code, respectively. As
tool
consistent
of
three
steps.
S
developer
knows
that
B
can
have
two
issues,
corresponding
In this
section,
present theofmethodology
of our study.
In
order
to ease
the we
explanation
our methodology,
we first
tool consistent
of of
three
steps.ex
all sources
possible
touse
two
possible
exceptions.
These
two
be
order
ease
the explanation
of our
methodology,
weexceptions
first identifycan
antoillustrative
example
with
which
we then
introduce
identify
all sources
of blocks
possible
the
exception
handling
(cae
use exception
an newly
illustrative
example
which
we from
thentheintroduce
our
flow
analysis.
Finally,
we present
subject the
either
raised
or with
propagated
another
method
that
exception
handling
are
handled.
Finally,
weblocks
recover(ca
th
our
exception
flow
analysis.
Finally,
we
present
the
subject
of our study.
are
handled.
Finally,
we is
recover
t
structing
call
graph
that
relevan
Bprojects
calls
(e.g.,
method
C
and
method
getPath).
Developers
projects of our study.
1structing call graph that is releva
A.
An
illustrative
example
of
exception
handling
practices
can
use
a throw
statement
tohandling
raise practices
an exception.
In our
1
A. InAnthis
illustrative
example
of exception
subsection,
we explain
an illustrative
example that
1
III. M ETHODOLOGY
In this section, we present the methodology of our study.
Aiding the explanation of our methodology, we first consider
an illustrative example. Second, we introduce our exception
flow analysis. Finally, we discuss the subject projects used.1
code, binaries and Tablea
example,
developers
newly
raise The
IOException
with1Our
thesource
throw
In thisraises
subsection,
we explain
an illustrative
example also
that available
handles,
and propagates
exceptions.
example
online code,
at https://guipadua.github
Our source
binaries and Table
statement
in and
thepropagates
methodexceptions.
C. Moreover,
if the
happens,
handles, raises
The example
alsoissue
available
online at https://guipadua.githu
such exception will propagate to all the methods that call C.
3) Documenting exceptions: From our example, a developer could consult the source code of B and C to identify
the rise of IOException in the throw statement. However,
the source code is not always available for a method. For
example, the method getPath called in method B is declared
by an external API. Developers would need to consult the
documentation (such as JavaDoc) of the method to discover
the propagation of the possible InvalidPathException.
There could be cases when an exception is thrown but
still not available in the documentation. Therefore, in some
programming languages (like Java), a possible exception can
be part of the method declaration. For example, method A,
B, and C declares that a possible exception can propagate in
the throws block of the method declaration. However, some
exceptions can remain undeclared (like InvalidPathException).
A. An illustrative example of exception handling practices
In this subsection, we explain an illustrative example that
handles, raises and propagates exceptions (see Figure 1). The
example also illustrates the means of documenting exceptions.
1) Handling possible exceptions: In this example, a developer would like to implement a method named A. The method
A requires to execute method B. The developer, by other
means, has the knowledge that B can face two issues: 1) having
an invalid path as input and 2) I/O faults. Therefore, instead
of executing as expected, method B would possibly throw two
types of exceptions: InvalidPathException and IOException,
which correspond to the two issues, respectively. To deal
with the two possible exceptions in method B, the developer
needs to either handle the exception, i.e., determine the alternative actions when such exception happens, or propagate
the exception such that a different method would manage
the issue. In our example, the developer decides to handle
InvalidPathException only and to propagate IOException.
B. Exception flow analysis
In the previous subsection, we presented an example of
an exception handling scenario with its related flows. In
this section, we present our methodology that automatically
1 Source
code, binaries, statistical tests and Tableau visualizations with raw
data are available online at https://guipadua.github.io/scam2017.
3
extracts possible exceptions and their flows in Java and C#
projects. We build an automated tool using Eclipse JDT Core
and .NET Compiler Platform (“Roslyn”) to parse the Java and
C# source code, respectively. As an overview, our analysis
consists of three main steps. First, we identify the exception
handling blocks (catch blocks). Second, we recover the flow
of exceptions by constructing the call graph that is relevant to
the identified catch blocks. Finally, by traversing the flow of
exceptions, we identify the sources of possible exceptions.
As a building block, we obtain the abstract syntax tree
(AST) from the source code. In this step, we include not
only the source code but also the binary files of dependencies
from the Java Virtual Machine (JVM) for Java or the .NET
Global Assembly Cache (GAC) for C#. The dependencies of
third party libraries used by the projects are also included.
These dependencies enrich the analysis by providing binding
information, which draws connections between the different
parts of a program (i.e. any method call and its origin, either
if part of an internal declaration or external dependency). Also,
we enrich the AST by parsing the documentation of the dependencies mentioned above as another source of information.
1) Identifying the handling of exceptions: We collect all
the exception handling scenarios through all the catch blocks
available in the AST. At the catch block, we use the AST
elements to identify the methods that are executed to handle
each exception as handling actions. We also obtain the related
try blocks, which provide a list of called methods. These
methods are necessary since they might raise or propagate
the exceptions that the catch block potentially handles. In
our example, we identify the catch block in method A. The
method printStackTrace is the handling action of the exception
InvalidPathException. From this catch block, we obtain the try
block in which we find the call to method B. Method B can
potentially propagate InvalidPathException and IOException.
2) Constructing call graph: Exceptions are propagated in
method calls. Therefore, we leverage call graphs to recover the
flow of exceptions. To handle polymorphism without risking
over-estimation, we only consider the possible exceptions of
the method that is declared in the parent class, since they are
more generic and often called within the derived methods.
Based on the previous step, for each identified method we
traverse its call graph in a depth-first manner to find its
possible exceptions. In our example, we traverse the call graph
of method B and find two possible exceptions: IOException
from method C and InvalidPathException from method getPath. Hence, based on the examples’ catch block, we know
that the InvalidPathException is handled in method A while
IOException is propagated without handling.
3) Identifying sources of exceptions: During the call graph
traverse and based on the AST, we identify four sources of
exceptions. They are: 1) The newly raised exception by the
throw statement, 2) the declared exception in the throws of
the method declaration (only for Java), 3) the documentation
as comments in the source code (like JavaDoc comments), and
4) the external documentation. In our illustrative example, we
can identify the newly raised IOException in a throw statement
TABLE II
A N OVERVIEW OF THE SELECTED SUBJECT PROJECTS .
Project
C#
Java
Release Version Type # Try # Catch # Method (K) KLOC
Glimpse
1.8.6
App.
56
57
1
31
Google API
v1.15.0
Lib.
22
30
16
628
OpenRA
release-20160508 App.
138
143
7
125
ShareX
v11.1.0
App.
334
341
7
177
SharpDevelop
5.0.0
App.
940
1060
41
923
SignalR
2.2.1
Lib.
94
105
2
38
Umbraco-CMS
release-7.5.0
App.
595
615
15
362
Apache ANT
rel/1.9.7
App.
Eclipse JDT Core
I20160803-2000 Lib.
Elasticsearch
v2.4.0
Guava
934
1139
11
158
1,424
1655
25
383
App.
385
408
12
108
v19.0
Lib.
263
317
10
79
Hadoop Common
rel/release-2.6.4
Lib.
975
1144
14
147
Hadoop HDFS
44
rel/release-2.6.4
App.
525
586
4
Hadoop MapReduce rel/release-2.6.4
App.
293
367
6
57
Hadoop YARN
rel/release-2.6.4
Lib.
1,192
1529
29
257
Spring Framework
v4.3.2.RELEASE Lib.
1,940
2301
30
349
Total 10,110 11,797
230 3,866
in method C (source 1), the declaration of the IOException in
methods A, B, and C (source 2), and the JavaDoc documentation of method C for IOException (source 3). In addition,
since we include the information from external libraries, our
tool can also identify that the method getPath called in method
B is a source of a possible InvalidPathException (source 4).
Some exceptions can be identified from multiple sources.
For example, IOException is identified by three separate
sources. We do not consider the multiple sources as different
exceptions if the exceptions are associated with the same
method call (e.g. method B). We label the separate sources
of an exception as detailed information for each method call.
C. Subject projects
Table II depicts the studied subject projects. Our study
considers Java and C# due to their popularity and prior
research (see Section II). Moreover, we include C# because of
its different approach compared to Java exception handling.
To facilitate replication of our work, we chose open source
projects that are available on GitHub.
We leverage GitHub filters on the number of contributors
(i.e. projects with multiple contributors) and the number of
stargazers (i.e. projects with more than ten stargazers), as they
can achieve a good precision for selecting engineered software
projects [29]. To narrow down the number of projects we
also sorted the projects in descending order of the number of
stargazers. Moreover, to potentially investigate the differences
in exception handling practices and increase generalizability,
we picked projects based on the filtering mentioned above. After reading the official description of the projects, we selected
multiple applications and multiple libraries (i.e. project type),
as well as multiple projects for different business domains
(i.e. project purpose). From each project, we selected the most
recent stable version of the source code at the moment of data
collection for analysis.
4
IV. Q UANTITY OF E XCEPTIONS
In this section, we study the quantity of all possible exceptions that are in each try block.
B. Potentially recoverable yet propagated exceptions
In the previous subsection, we find that a significant amount
of the possible exceptions are propagated. However, not all
exceptions are easy to recover or, more importantly, should
even be recovered. For example, exceptions such as ThreadDeath in Java, and OutOfMemoryException in C# cannot
feasibly be recovered. In fact, both Java and C# define the
recoverability level of exceptions in their documentation [31],
[32]. In particular, they suggest that developers should handle
potentially recoverable exceptions while developers may not
handle potentially unrecoverable ones. Hence, we first group
all the propagated exceptions into either potentially recoverable or potentially unrecoverable, according to the specific
guidance on Java or C# documentation. Then we count the
number of propagated exceptions with potential recoverability.
We find that almost 8% (117) of C# and more than 19%
(1,359) of Java try blocks have at least one potentially recoverable yet propagated exception. For example, a method
named rename in Hadoop HDFS for file renaming features has
a possible and potentially recoverable exception called FileAlreadyExistsException. This exception indicates the situation
where a file is renamed to another existing file. However, this
potentially recoverable exception is not handled by any catch
block in that method.
⇤
A. All and propagated possible exceptions
Ideally, developers should be aware of all possible exceptions to decide between handling or propagating them. To do
that, developers need to navigate the call graph of a system
that could extend to multiple ramifications. Hence, the more
exceptions there are, the more challenging (i.e. exponential
growth) it is for developers to comprehend and decide about
exception handling. Besides that, missing possible exceptions
can be a reason for the lack of a handler that should exist,
which is considered one of the top causes of exception
handlings bugs [8]. For those reasons, we study the quantity
of total and propagated possible exceptions in each exception
handling block.
As described in our methodology (see Section III), we
collect all the methods called in each try block. For those
methods, we can recover the possible exceptions. Afterward,
we can measure the quantity of possible exception by counting
the unique types of exceptions in each try block.
We find that there typically exist multiple possible exceptions in each try block (see Figure 2). The median number of
distinct possible exception per try block is four and two, for
C# and Java respectively. More than 48% (C#) and 38% (Java)
of try blocks can throw in between two and five exceptions.
Moreover, more than 36% (C#) and 24% (Java) of the try
blocks have six or more possible exceptions. For example, an
important method named processCompiledUnits(int,boolean)
in Eclipse JDT Core in the Compiler class has a try block with
33 distinct possible exceptions. Developers should properly
handle exceptions in such an important method, to ensure
reliability.
Among all possible exceptions, there often exist possible
exceptions that are not handled by any of the catch blocks
that are associated with a try block [30]. These unhandled
exceptions may increase the challenge of exception handling
practice since they will be propagated and will need to
be handled elsewhere. If the propagated exceptions remain
uncaught across the whole system, there could be a risk
of system failures [20], [7], [8]. Therefore, we identify the
possible exceptions that are not handled by the catch block.
Figure 2 presents the possible propagated exceptions for
each try statement. We find that there may exist a large
number (up to 34) of possible exceptions that are unhandled in
each try block. For example, a method execute(List,int) from
class org.apache.tools.ant.taskdefs.optional.junit.JUnitTask in
Apache ANT has 25 possible exceptions while the corresponding catch block only handles IOException.
⇤
⇥
⇥
Finding 2: There exists a considerable amount of potentially
recoverable exceptions that are propagated, even though they are
recommended to be handled by Java and C#.
Implications: Exception flow analysis can provide automated
tooling support to alert developers about not handling potentially
recoverable exceptions.
V. D IVERSITY OF E XCEPTIONS
There can be a diverse set of exceptions being used across
try blocks. Prior research discusses that the use of a high
number of distinct exception types might represent a greater
concern with exception handling [12]. Therefore, in this
section, we study the diversity of exceptions in our subject
projects.
We count the total number of distinct exception types in
each project, and the amount of try blocks in which each
type of exception appears. Table III shows the percentage of
the exception types of each project that appear in different
quantities of try blocks. Despite the large number (up to 97 in
C# and 249 in Java) of distinct exception types, there exist a
considerable amount of exception types that only appear in few
try blocks. In fact, over half of the exception types in C#, and
almost 1/3 of the exception types in Java only appear in one
try block. Such results imply that although the high number
of distinct exception types may be a burden to developers, the
burden may not be as high since a considerable amount of the
exception types would only affect a small portion of the code.
Finding 1: There often exist multiple possible exceptions in each
try block, and, out of those, many are propagated.
Implications: Current state-of-the-practice may not provide information to developers about all possible exceptions. Automated
techniques may help developers be aware of all possible exceptions to make exception handling decisions.
5
Language / Project
# Propagated
# Prop. and Potenti..
C#
Java
20
15
10
5
0
20
Median: 0
Median: 0
15
10
5
0
20
Median: 1
Median: 0
# Total
15
10
5
Median: 4
Median: 2
0
Glimpse
Google
API
OpenRA
ShareX SharpDe.. SignalR Umbraco Apache
ANT
E. JDT Elasticse.. Guava
Core
H.
H. HDFS H. MapR H. YARN
Common
educe
Spring
Fig. 2. Number of possible exceptions per try block for each project broken down by Propagated and Potentially Recoverable, Propagated, and Total.
A. Multiple sources of the same exception
TABLE III
T OTAL AMOUNT OF DISTINCT EXCEPTION TYPES AND THE PERCENTAGES
The multiple sources of exceptions may increase the complexity of exception handling. Consequently, a developer
would need to comprehend and investigate more methods in
the source code to effectively handle exceptions. For example,
developers may encounter a FileNotFoundException due to
missing an input file or configuration file. However, developers
may need different actions to handle such an exception since
missing an input file may be caused by users’ mistake while
missing a configuration file is a critical issue of the software.
Multiple sources of the same exception may also impact testers
since they would need to properly test the exception behavior
as well as the multiple possible paths of control flow.
We group each possible exception by the distinct methods
that act as a source of exceptions. We only consider distinct
methods since the same method may not need different ways
to handle the exception while the exception propagated from
various methods may need to be handled differently. In Table IV, we present the percentage of possible exceptions that
are traced back from zero, one, two and more than two distinct
methods. The first group is from zero methods, which means
that these possible exceptions were traced back to explicit
throw invocations, not method invocations.
Although most of the possible exceptions (above 76%) are
traced back to a single method, we observe that more than 22%
of the exceptions are traced back to multiple invoked methods.
The try blocks with the highest number of methods can have
from two to 17 among C# projects; while, for Java, it is between five and 34. For example, the Umbraco C# class called
TypeFinder performs lazy accesses to all assemblies inside a
single try block and therefore System.ArgumentNullException
can be traced back from 14 different invoked methods. We
also noticed that exceptions that are super classes of other
exceptions (e.g., IOException) have a higher than average
chances of being from multiple sources.
OF DISTINCT EXCEPTION TYPES THAT APPEAR IN DIFFERENT QUANTITY
OF TRY BLOCKS .
C#
Java
⇤
⇥
Try blocks
Project
1
2
3
Glimpse
27.78%
38.89%
11.11%
Google API
28.00%
40.00%
12.00%
OpenRA
35.71%
4.76%
7.14%
ShareX
13.04%
8.70%
SharpDevelop
19.59%
SignalR
4
5
>5
Total
22.22%
18
4.00%
16.00%
25
16.67%
2.38%
33.33%
42
6.52%
2.17%
6.52%
63.04%
46
8.25%
6.19%
4.12%
2.06%
59.79%
97
56.67%
16.67%
3.33%
6.67%
16.67%
30
Umbraco
27.69%
9.23%
4.62%
1.54%
56.92%
65
Total
55.88%
25.00%
13.97%
9.56%
5.15%
47.79%
214
Apache ANT
15.73%
6.74%
6.74%
3.37%
3.37%
64.04%
89
E. JDT Core
5.56%
2.78%
2.78%
4.17%
1.39%
83.33%
72
Elasticsearch
27.78%
12.50%
16.67%
9.72%
4.17%
29.17%
72
Guava
24.00%
12.00%
16.00%
2.00%
10.00%
36.00%
50
H. Common
14.53%
15.12%
9.88%
11.63%
8.14%
40.70%
172
H. HDFS
27.50%
13.75%
16.25%
7.50%
1.25%
33.75%
80
H. MapReduce
21.74%
8.70%
4.35%
8.70%
2.17%
54.35%
46
H. YARN
17.53%
4.12%
11.34%
6.19%
3.09%
57.73%
97
Spring
22.09%
12.05%
7.63%
10.04%
4.82%
43.37%
249
Total
32.93%
17.76%
15.97%
14.77%
8.38%
42.32%
662
37.83%
19.31%
15.54%
13.66%
7.69%
43.49%
876
Grand Total
Finding 3: With a large amount of exceptions exist in each
project, many possible exception types appear in only one try
block.
Implications: Developers may not need to be aware of all
exception types in a project by receiving automated suggestions
of the exceptions that he/she needs to understand.
VI. S OURCES OF E XCEPTIONS
The same exception may be traced back from different
sources. In this section, we study the sources of exceptions
per try block.
6
TABLE IV
P ERCENTAGE OF DISTINCT EXCEPTIONS THAT ARE TRACED BACK TO ONE ,
TWO OR OVER TWO METHODS .
0
⇤
Distinct methods
1
2
>2
Total
C#
1.05%
76.11%
14.05%
8.80%
7,638
Java
0.61%
77.18%
13.00%
9.20%
28,854
Finding 4: Over 22% of the exceptions are traced from different
methods.
Implications: Automated tools are needed to help developers
understand the source of the exception if it is traced back to
⇥ different methods.
(a) Java Unchecked Exceptions.
B. Sources of exception documentation
Prior studies revealed that lacking immediate documentation
is one of the challenges of exception handling [14], [17], [18].
Prior studies observed a small number of documented exceptions. As shown in our illustrative example (see Section III-A),
possible exceptions can be recovered from up to four different
sources. They are: 1) the newly raised exception by the throw
statement in the source code, 2) the declared exception in
the throws of the method declaration (only for Java), 3) the
documentation as comments in the source code (like JavaDoc
comments) and 4) the external documentation. By recovering
the sources of each possible exception using exception flow
analysis, we may be able to provide the documentation of
possible exceptions. For Java, we only recover documentation
for unchecked exceptions since checked exceptions must be
specified in method declarations.
We find that, from all the possible exceptions that we
identify, 93% for Java and 71% for C# can be retrieved from
the external documentation of dependencies. Figure 3 depicts
the sets of possible exceptions per try block that were retrieved
by our exception flow analysis. Our findings show that the
challenge of having a low amount of documented exceptions
can be well addressed by applying exception flow analysis
with the information from external documents of libraries. We
find that such rich documents are typically from the exceptions
that are provided by the system libraries. Therefore, the high
availability of such documentation can be expected to assist
developers not only for our subject systems but also for the
majority of Java and C# projects.
⇤
⇥
(b) C# Exceptions
Fig. 3. Quantity of identified possible exceptions per source of information.
A. Exception handling strategies
Exception handling strategy describes the manner in which
an exception is handled. In particular, the relationship between
the possible exception in a try block and the handler exception
in the corresponding catch blocks. There exist in total two
handling strategies:
• Specific, is the strategy when the type of a possible
exception is exactly the same as the handler exception.
• Subsumption, is the strategy when the handler exception
is a superclass of a possible exception.
Since there can be multiple possible exceptions, it can be overwhelming for a developer to handle each possible exception
with a Specific strategy. On the other hand, the Subsumption
strategy may introduce uncertainty to the caught exception.
To study the handling strategies, we compare each possible
exception with the handler exception in the corresponding
catch block. Figure 4 depicts the quantity of distinct possible
exception per try block that is handled according to each
strategy.
The majority of the exceptions are handled with a subsumption strategy, while only a small portion of the exceptions
are handled specifically. The results show that developers
Finding 5: The libraries used by the systems can provide
documentation to most of the possible exceptions.
Implications: Developers should leverage automated flow analysis to understand possible exceptions.
VII. E XCEPTION H ANDLING S TRATEGIES AND ACTIONS
In this section, we study the strategy and actions in exception handling practices considering the exception flows of each
handler.
7
using each type of action in each C# and Java project. Hence,
we examined if there exists statistically significant difference
(i.e. p-value < 0.05) between Java and C#. A p-value < 0.05
means that the difference is likely not by chance. We choose
Wilcoxon Rank Sum test since it does not have an assumption
on the distribution of the data.
We find statistically significant difference between Java
and C# for exception handing with actions “Throw Wrap”,
“Throw New”, “Nested Try”, “Continue” and “Todo”. Among
these actions, “Throw Wrap”, “Throw New”and “Todo” may
indicate that the exceptions are not effectively handled but
rather propagated or ignored. All these three actions show up
more in Java than C#. We consider the reason may be that
Java compiler forces developers to explicitly manage checked
exceptions while developers may not have the knowledge of
how to handle them properly. To simply make the program
compile, developers potentially take these actions. Further
studies should investigate why such actions are chosen more
in Java than C#.
⇤
Lang.. Project
C#
Glimpse
Google API
OpenRA
ShareX
SharpDevel..
SignalR
Umbraco
Total
Java
Median: 2
Median: 0
Apache ANT
E. JDT Core
Elasticsearch
Guava
H. Common
H. HDFS
H. MapRedu..
H. YARN
Spring
Total
Median: 0
0
5
10
Subsumption
Median: 1
15 0
5
10
Specific
15
Fig. 4. Quantity of possible exceptions per try block per project by handling
strategy.
tend to over-catch exceptions. The extreme case of subsumption strategy is the “Catch Generic” exception handling
anti-pattern [13], where developers simply use an exception
type which can catch any exceptions in the software, e.g.,
Exception in Java. Such practice is heavily discussed in prior
research [20], [17] and is considered to be harmful since
developers cannot optimize the handling of the exception
based on the exact type of the exception, but rather only know
that there may exist some exceptions during run-time.
⇤
⇥
⇥
Finding 7: Java and C# have differences in leveraging various
actions when handling exceptions.
Implications: More in-depth analysis and user studies are needed
to further understand the rationale of differences of Java and C#
exception handling practices.
We also compare the actions that are taken when the exceptions are handled with either specific or subsumption strategy.
We perform Wilcoxon Rank Sum test similar as when comparing Java and C#. This time, for each particular programming
language and type of action, the test compared the percentage
of specific handling in each project with subsumption handling
in each project. Nonetheless, we observe only one action, i.e.,
Log, that is handled differently (statistically significant) with
specific or subsumption strategy.
⇤
Finding 6: Only a small portion of the exceptions are handled
with the Specific strategy.
Implications: Developers should be guided to prioritize on
handling exception with the Specific strategy, since developers
cannot optimize the handling of an exception without knowing
its exact type information.
⇥
B. Exception handling actions
Finding 8: Only one action, Log in Java, is taken differently when
exceptions are handled with specific or subsumption strategy.
Implications: Research and tooling support are needed to guide
how to handle exceptions, especially with the specific strategy.
We would like to know if any particular actions are taken
when handling some special possible exception. With such
knowledge, we may be able to suggest actions automatically
to developers handling exceptions. We gather a list of the ten
most handled types of possible exceptions in Java and C#,
respectively (see Table VI). We also obtain the percentage of
possible exceptions that are handled using each action for each
exception type. Similarly, we use Wilcoxon Rank Sum test to
compare. For each particular programming language, action,
and exception type, the test compared the percentage of the
given type in each project with the combined value of all other
types of exceptions in each project. We find with statistical
significance that, for Java, all top exceptions have at least one
action that is taken differently from the rest of the exceptions,
and, for C#, two top exceptions has such difference.
During the exception flow analysis (see Section III), we
collect a set of method calls in each catch block to know
how each exception is handled. Prior studies [3], [17], [14],
[16], [28] propose a list of actions based on the combination of
method calls in the catch block as Exception handling actions.
Table V presents the list of actions that are defined in prior
research and are used in this paper. To further understand
how exceptions are handled, we study the exception handling
actions in our subject projects.
Figure 5 presents the percentages of possible exceptions of
each project that are handled using a particular type of action.
We observe that Java and C# have differences in executing
various actions when handling exceptions. To determine the
differences, we perform Wilcoxon Rank Sum test [33] to
compare the percentage of possible exceptions that are handled
8
Language
Method
TABLE VI
L IST OF TOP 10 COMMON EXCEPTION TYPES IN THE STUDIED PROJECTS .
Java
C#
Throw
Wrap
50.78%
Java
C#
C#
Log
Java
C#
Throw
New
26.64%
Java
C#
Throw
Current
19.56%
Java
C#
Return
11.85%
Java
C#
Nested
Try
10.24%
Java
C#
Empty
4.67%
Java
C#
Continue
4.23%
Java
java.lang.IllegalArgumentException†
System.NotSupportedException†
java.lang.NullPointerException†
System.ArgumentOutOfRangeException
java.lang.IndexOutOfBoundsException†
System.InvalidOperationException
java.lang.SecurityException†
System.FormatException
java.lang.IllegalStateException†
System.IO.IOException
java.lang.ExceptionInInitializerError†
System.IO.PathTooLongException†
java.lang.ArrayStoreException†
System.Security.SecurityException
java.lang.IllegalAccessException†
System.ObjectDisposedException
java.lang.ClassNotFoundException†
0.26%
C#
generalize to other projects, languages or commercial systems.
Replicating our study on other subjects may address this threat
and further understand the state-of-the-practice of exception
handling.
Internal validity. We aim to include all possible sources
of information in our automated exception flow analysis.
However, our analysis may still miss possible exceptions, if
there is a lack of documentation or the source code is not
compilable. Also, the documentation of the exception may be
incorrect or outdated. In our analysis, we trust the content of
documentation. Therefore, we cannot claim that our analysis
fully recovers all possible exceptions nor that the recovered
information is impeccable. Further studies may perform deeper
analysis on the quality of exception handling documentation
to address this threat.
Construct validity. Our study may not cover all possible
handling actions. We selected actions based on the previous
research in the subject [3], [17], [14], [16], [28]. Some actions
are not included in our study if they are either 1) require
heuristic to detect or 2) are not well explained in details in
related work.
Our possible exception identification approach is based on
a call graph approximation from static code analysis. We
may still miss possible exceptions due to under-estimation
for polymorphism or unresolved method overload. Although
such approximation may impact our findings, our choice of
under-estimation would not significantly alter the existence of
observed challenges of exception handling, i.e., the challenge
may appear even worse without the under-estimation. Nevertheless, to complement our study, dynamic analysis on the
exception flow may be carried out to understand the system
exceptions during run-time.
0.13%
Java
C#
Default
0.09%
Java
C#
0.00%
0.00%
5.00%
10.00% 15.00% 20.00% 25.00% 30.00% 35.00% 40.00% 45.00% 50.00% 55.00% 60.00% 65.00% 70.00%
Value
Fig. 5. Percentage of possible exceptions that are handled using each type
of action.
⇥
java.io.IOException†
System.ArgumentException
Java
Todo
⇤
System.ArgumentNullException
†The exception has at least one actions that taken statistically significantly differently from the rest exceptions.
C#
Abort
Java
48.08%
Finding 9: All top 10 Java and 2 out of top 10 C# exceptions
have at least one action that is taken statistically significantly
differently from the rest exceptions.
Implications: Developers may consider leveraging automated
suggestions of exception handling actions.
TABLE V
L IST OF THE DETECTED ACTIONS .
Action
Short Description
Abort
The handler contains an abort statement [3].
Continue
The handler contains a continue statement [14].
Default
The handler contains the IDE suggested method (Java only).
Empty
The handler is empty [3], [17], [14], [16].
Log
The handler displays or logs some information [17], [14], [16].
Method
The handler contains a method invocation different
than the other actions listed in this table. [14], [17].
Nested Try
The handler contains a new try statement [28].
Return
The handler contains a return statement [17], [14].
Throw Current
The handler contains a throw statement
without a new exception instantiation [17], [14], [16].
Throw New
The handler contains a throw statement
with a new exception instantiation [17], [14], [16].
Throw Wrap
The handler contains a throw statement using the original
exception or its associated information [16].
Todo
The handler contains TODO or FIXME comments [3].
IX. C ONCLUSION
Exception handling is an important feature in modern
programming languages. However, prior studies unveil the
suboptimal usage of exception handling features in practice.
In this paper, we revisit practice of exception handling in 16
open source software in Java and C#. Although we confirm
that there exist suboptimal manners of exception handling,
VIII. T HREATS TO VALIDITY
In this section, we discuss the threats to validity of our
findings.
External validity. Our study is based on a set of open source
Java and C# projects from GitHub. Our findings may not
9
more importantly, we highlight the opportunities of performing
source code analysis to recover exception flows to help practitioners tackle various complex issues of handling exceptions.
In particular, the contributions of our paper are:
1) We design an automated tool that recovers exception
flows from both Java and C#.
2) We present empirical evidence to illustrate the challenges and complexity of exception handling in open
source systems.
3) Our exception flow analysis, as an automated tool, can
already provide valuable information to assist developers
better understand and make exception handling decisions.
This paper highlights the opportunities and urgency of providing automated tooling to help developers make exception
handling decisions during the development of quality and
reliable software systems.
[16] R. Coelho, A. Rashid, A. Garcia, F. Ferrari, N. Cacho, U. Kulesza,
A. von Staa, and C. Lucena, “Assessing the Impact of Aspects on
Exception Flows: An Exploratory Study,” in ECOOP 2008 – ObjectOriented Programming: 22nd European Conference Paphos, Cyprus,
July 7-11, 2008 Proceedings, 2008, vol. 5142 LNCS, pp. 207–234.
[17] D. Sena, R. Coelho, U. Kulesza, and R. Bonifácio, “Understanding the
exception handling strategies of Java libraries,” in Proceedings of the
13th International Workshop on Mining Software Repositories - MSR
’16, 2016, pp. 212–222.
[18] M. Kechagia and D. Spinellis, “Undocumented and unchecked: exceptions that spell trouble,” in Proceedings of the 11th Working Conference
on Mining Software Repositories - MSR 2014. New York, New York,
USA: ACM Press, 2014, pp. 312–315.
[19] B. Cabral, P. Sacramento, and P. Marques, “Hidden truth behind .NET’s
exception handling today,” Software, IET, vol. 2, no. 1, pp. 233–250,
2007.
[20] S. Sinha, A. Orso, and M. Harrold, “Automated support for development,
maintenance, and testing in the presence of implicit flow control,” in
Proceedings. 26th International Conference on Software Engineering,
no. September, 2004, pp. 336–345.
[21] N. Cacho, T. César, T. Filipe, E. Soares, A. Cassio, R. Souza, I. Garcia,
E. A. Barbosa, and A. Garcia, “Trading robustness for maintainability:
an empirical study of evolving c# programs,” Proceedings of the 36th
International Conference on Software Engineering - ICSE 2014, no. iii,
pp. 584–595, 2014.
[22] N. Cacho, E. A. Barbosa, J. Araujo, F. Pranto, A. Garcia, T. Cesar,
E. Soares, A. Cassio, T. Filipe, and I. Garcia, “How Does Exception
Handling Behavior Evolve? An Exploratory Study in Java and C#
Applications,” in 2014 IEEE International Conference on Software
Maintenance and Evolution, no. 1. IEEE, sep 2014, pp. 31–40.
[23] M. P. Robillard and G. C. Murphy, “Analyzing Exception Flow in Java
Programs,” in Proceedings of the 7th European Software Engineering
Conference Held Jointly with the 7th ACM SIGSOFT International
Symp. on Foundations of Software Engineering, 1999, pp. 322–337.
[24] I. Garcia and N. Cacho, “eFlowMining: An Exception-Flow Analysis
Tool for .NET Applications,” in 2011 Fifth Latin-American Symposium
on Dependable Computing Workshops, no. i. IEEE, apr 2011, pp. 1–8.
[25] M. Kechagia, T. Sharma, and D. Spinellis, “Towards a context dependent
java exceptions hierarchy,” in Proceedings of the 39th International
Conference on Software Engineering Companion, ser. ICSE-C ’17.
Piscataway, NJ, USA: IEEE Press, 2017, pp. 347–349.
[26] L. Zhang and C. Krintz, “As-if-serial exception handling semantics for
Java futures,” Sci. of Computer Programming, vol. 74, pp. 314–332,
2009.
[27] B. Cabral and P. Marques, “A transactional model for automatic exception handling,” Computer Languages, Systems and Structures, vol. 37,
no. 1, pp. 43–61, 2011.
[28] J. Zhu, P. He, Q. Fu, H. Zhang, M. R. Lyu, and D. Zhang, “Learning
to Log: Helping Developers Make Informed Logging Decisions,” in
2015 IEEE/ACM 37th IEEE International Conference on Software
Engineering, vol. 1. IEEE, may 2015, pp. 415–425.
[29] N. Munaiah, S. Kroh, C. Cabrey, and M. Nagappan, “Curating GitHub
for engineered software projects,” Empirical Software Engineering, pp.
1–35, 2017.
[30] M. P. Robillard and G. C. Murphy, “Static analysis to support the
evolution of exception structure in object-oriented systems,” ACM Transactions on Software Engineering and Methodology, vol. 12, no. 2, pp.
191–221, 2003.
[31] “Chapter 11. Exceptions - Java SE Specification.” [Online].
Available: http://docs.oracle.com/javase/specs/jls/se8/html/jls-11.html,
accessed 2017-03-29
[32] “Handling and Throwing Exceptions - .NET Framework
Documentation.” [Online]. Available: https://msdn.microsoft.com/enus/library/5b2yeyab(v=vs.110).aspx
[33] F. Wilcoxon and R. A. Wilcox, Some rapid approximate statistical
procedures. Lederle Laboratories, 1964.
R EFERENCES
[1] P. M. Melliar-Smith and B. Randell, “Software Reliability: The Role
of Programmed Exception Handling,” Reliable Computer Systems, pp.
143–153, 1985.
[2] C.-T. Chen, Y. C. Cheng, C.-Y. Hsieh, and I.-L. Wu, “Exception handling
refactorings: Directed by goals and driven by bug fixing,” Journal of
Systems and Software, vol. 82, no. 2, pp. 333–345, feb 2009.
[3] D. Yuan, Y. Luo, and X. Zhuang, “Simple Testing Can Prevent Most
Critical Failures,” 11th USENIX Symposium on Operating Systems
Design and Implementation (OSDI 14), 2014.
[4] T. Cargill, C++ Gems, S. B. Lippman, Ed. New York, NY, USA: SIGS
Publications, Inc., 1996.
[5] B. Zhang and J. Clause, “Lightweight automated detection of unsafe
information leakage via exceptions,” Proceedings of the 2014 International Symposium on Software Testing and Analysis - ISSTA 2014, pp.
327–338, 2014.
[6] F. Cristian, “Exception Handling and Software Fault Tolerance,” Computers, IEEE Transactions on, vol. C-31, no. 6, pp. 531–540, 1982.
[7] E. A. Barbosa, A. Garcia, and S. D. J. Barbosa, “Categorizing Faults
in Exception Handling: A Study of Open Source Projects,” in 2014
Brazilian Symposium on Software Engineering, no. September. IEEE,
sep 2014, pp. 11–20.
[8] F. Ebert, F. Castor, and A. Serebrenik, “An exploratory study on
exception handling bugs in Java programs,” Journal of Systems and
Software, vol. 106, pp. 82–101, aug 2015.
[9] S. Nakshatri, M. Hegde, and S. Thandra, “Analysis of exception handling
patterns in Java projects,” in Proceedings of the 13th International
Workshop on Mining Software Repositories - MSR ’16. New York,
New York, USA: ACM Press, 2016, pp. 500–503.
[10] M. B. Kery, C. Le Goues, and B. A. Myers, “Examining programmer
practices for locally handling exceptions,” in Proceedings of the 13th
International Workshop on Mining Software Repositories - MSR ’16.
New York, New York, USA: ACM Press, 2016, pp. 484–487.
[11] M. Asaduzzaman, M. Ahasanuzzaman, C. K. Roy, and K. A. Schneider,
“How developers use exception handling in Java?” in Proceedings of the
13th International Workshop on Mining Software Repositories - MSR
’16. New York, New York, USA: ACM Press, 2016, pp. 516–519.
[12] R. Bonifacio, F. Carvalho, G. N. Ramos, U. Kulesza, and R. Coelho,
“The use of C++ exception handling constructs: A comprehensive
study,” in 2015 IEEE 15th International Working Conference on Source
Code Analysis and Manipulation (SCAM). IEEE, sep 2015, pp. 21–30.
[13] G. B. de Pádua and W. Shang, “Studying the Prevalence of Exception
Handling,” 25th IEEE International Conference on Program Comprehension (ICPC ’17), pp. 1–4, 2017.
[14] B. Cabral and P. Marques, “Exception Handling: A Field Study in Java
and .NET,” in ECOOP 2007 Object-Oriented Programming, vol. 4609,
Berlin, Heidelberg, 2007, pp. 151–175.
[15] J. W. Jo, B. M. Chang, K. Yi, and K. M. Choe, “An uncaught exception
analysis for Java,” Journal of Systems and Software, vol. 72, no. 1, pp.
59–69, 2004.
10
| 6cs.PL
|
A Combinatorial Toolbox for Protein Sequence Design and
Landscape Analysis in the Grand Canonical Model
arXiv:cs/0101015v1 [cs.CE] 17 Jan 2001
James Aspnes∗
Julia Hartling†
Ming-Yang Kao‡
Junhyong Kim§
Gauri Shah¶
December 26, 2017
Abstract
In modern biology, one of the most important research problems is to understand how protein
sequences fold into their native 3D structures. To investigate this problem at a high level, one
wishes to analyze the protein landscapes, i.e., the structures of the space of all protein sequences
and their native 3D structures. Perhaps the most basic computational problem at this level is
to take a target 3D structure as input and design a fittest protein sequence with respect to one
or more fitness functions of the target 3D structure. We develop a toolbox of combinatorial
techniques for protein landscape analysis in the Grand Canonical model of Sun, Brem, Chan, and
Dill. The toolbox is based on linear programming, network flow, and a linear-size representation
of all minimum cuts of a network. It not only substantially expands the network flow technique
for protein sequence design in Kleinberg’s seminal work but also is applicable to a considerably
broader collection of computational problems than those considered by Kleinberg. We have used
this toolbox to obtain a number of efficient algorithms and hardness results. We have further used
the algorithms to analyze 3D structures drawn from the Protein Data Bank and have discovered
some novel relationships between such native 3D structures and the Grand Canonical model.
∗
Department of Computer Science, Yale University, New Haven, CT 06520-8285, USA. Email: [email protected].
Supported in part by NSF Grant CCR-9820888.
†
Department of Ecology and Evolutionary Biology, Yale University, New Haven, CT 06520-8285, USA. Email:
[email protected].
‡
Department of Electrical Engineering and Computer Science, Tufts University, Medford, MA 02155, USA. Email:
[email protected]. Supported in part by NSF Grant CCR-9531028.
§
Department of Ecology and Evolutionary Biology, Department of Molecular, Cellular, and Developmental Biology,
and Department of Statistics, Yale University, New Haven, CT 06520-8285, USA. Email: [email protected].
Supported in part by Merck Genome Research Institute Grant and NSF Grant DEB-9806570.
¶
Department of Computer Science, Yale University, New Haven, CT 06520-8285, USA. Email: [email protected].
1
1
Introduction
In modern biology, one of the most important research problems is to understand how protein sequences fold into their native 3D structures [23]. This problem can be investigated at two complementary levels. At a low level, one wishes to determine how an individual protein sequence folds.
A fundamental computational problem at this level is to take a protein sequence as input and find
its native 3D structure. This problem is sometimes referred to as the protein structure prediction
problem and has been shown to be NP-hard (see, e.g., [1, 5, 6]). At a high level, one wishes to analyze the protein landscapes, i.e., the structures of the space of all protein sequences and their native
3D structures. Perhaps the most basic computational problem at this level is to take a target 3D
structure as input and ask for a fittest protein sequence with respect to one or more fitness functions
of the target 3D structure. This problem has been called the protein sequence design problem and
has been investigated in a number of studies [3, 7, 9, 16, 27, 30, 33, 35].
The focus of this paper is on protein landscape analysis, for which several quantitative models have
been proposed in the literature [7, 30, 33]. As some recent studies on this topic have done [3, 18, 24],
this paper employs the Grand Canonical (GC) model of Sun, Brem, Chan, and Dill [33], whose
definition is given in Section 2. Generally speaking, the model is specified by (1) a 3D geometric
representation of a target protein 3D structure with n amino acid residues, (2) a binary folding code in
which the amino acids are classified as hydrophobic (H) or polar (P) [8, 19], and (3) a fitness function
Φ defined in terms of the target 3D structure that favors protein sequences with a dense hydrophobic
core and with few solvent-exposed hydrophobic residues.
In this paper, we develop a toolbox of combinatorial techniques for protein landscape analysis
based on linear programming, network flow, and a linear-size representation of all minimum cuts of
a network [26]. This toolbox not only substantially expands the network flow technique for protein
sequence design in Kleinberg’s seminal paper [18] but also is applicable to a considerably broader
collection of computational problems than those considered by Kleinberg. We have used this toolbox
to obtain a number of efficient algorithms and hardness results. We have further used the algorithms
to analyze 3D structures drawn from Protein Data Bank at http://www.rcsb.org/pdb and have
discovered some novel relationships between such native 3D structures and the Grand Canonical
model (Figure 1). Specifically, we report new results on the following problems, where ∆ is the
number of terms in the fitness function or functions as further defined in Section 3.1. Many of the
results depend on computing a maximum network flow in a graph of size O(∆); in most cases, this
network flow only needs to be computed once for each fitness function Φ.
P1 Given a 3D structure, find all its fittest protein sequences. Note that there can be exponentially
many fittest protein sequences. We show that these protein sequences together have a representation of size O(∆) that can be computed in O(∆) time after a certain maximum network flow
is computed (Theorem 5), and that individual fittest protein sequences can be generated from
this representation in O(n) time per sequence (Theorem 9).
P2 Given f 3D structures, find the set of all protein sequences that are the fittest simultaneously
for all these 3D structures. This problem takes O(∆) time after f maximum network flow
computations (Theorem 8).
P3 Given a protein sequence x̂ and its native 3D structure, find the set of all fittest protein sequences
that are also the most (or least) similar to x̂ in terms of unweighted (or weighted) Hamming
distances. This problem takes O(∆) time after a certain maximum network flow is computed
(Theorem 7).
2
P4 Count the number of protein sequences in the solution to each of Problems P1, P2, and P3.
These counting problems are computationally hard (Theorem 17).
P5 Given a 3D structure and a bound e, enumerate the protein sequences whose fitness function
values are within an additive factor e of that of the fittest protein sequences. This problem takes
polynomial time to generate each desired protein sequence (Theorem 12).
P6 Given a 3D structure, determine the largest possible unweighted (or weighted) Hamming distance between any two fittest protein sequences. This problem takes O(∆) time after a certain
maximum network flow is computed (Theorem 10).
P7 Given a protein sequence x̂ and its native 3D structure, find the average unweighted (or weighted)
Hamming distance between x̂ and the fittest protein sequences for the 3D structure. This
problem is computationally hard (Theorem 17).
P8 Given a protein sequence x̂, its native 3D structure, and two unweighted Hamming distances d1
and d2 , find a fittest protein sequence whose distance from x̂ is also between d1 and d2 . This
problem is computationally hard (Theorem 18(1)).
P9 Given a protein sequence x̂, its native 3D structure, and an unweighted Hamming distance d,
find the fittest among the protein sequences which are at distance d from x̂. This problem is
computationally hard (Theorem 18(2)). We have a polynomial-time approximation algorithm
for this problem (Theorem 13).
P10 Given a protein sequence x̂ and its native 3D structure, find all the ratios between the scaling
factors α and β in Equation 1 in Section 2 for the GC model such that the smallest possible
unweighted (or weighted) Hamming distance between x̂ and any fittest protein sequence is
minimized over all possible α and β. (This is a problem of tuning the GC model.) We have a
polynomial-time algorithm for this problem (Theorem 16).
P11 Given a 3D structure, determine whether the fittest protein sequences are connected, i.e., whether
they can mutate into each other through allowable mutations, such as point mutations, while
the intermediate protein sequences all remain the fittest [2, 8, 17, 20, 22, 29, 31]. This problem
takes O(∆) time after a certain maximum network flow is computed (Theorem 11).
P12 Given a 3D structure, in the case that the set of all fittest protein sequences is not connected,
determine whether two given fittest protein sequences are connected. This problem takes O(∆)
time after a certain maximum network flow is computed (Theorem 11).
P13 Given a 3D structure, find the smallest set of allowable mutations with respect to which the
fittest protein sequences (or two given fittest protein sequences) are connected. This problem
takes O(∆) time after a certain maximum network flow is computed (Theorem 11).
Previously, Sun et. al. [33] developed a heuristic algorithm to search the space of protein sequences for a fittest protein sequence without a guarantee of optimality or near-optimality. Hart
[16] subsequently raised the computational tractability of constructing a single fittest protein sequence as an open question. Kleinberg [18] gave the first polynomial-time algorithm for this problem,
which is based on network flow. In contrast, Problem P1 asks for all fittest protein sequences and
yet can be solved with the same time complexity. Kleinberg also formulated more general versions
of Problems P11 and P12 by extending the fitness function to a submodular function and gave
polynomial-time algorithms. Our formulations of these two problems and Problem P13 are directly
3
based on the fitness function of the GC model; furthermore, as is true with several other problems
above, once a solution to Problem P1 is obtained, we can solve these three problems in O(∆) time.
Among the above thirteen problems, those not yet mentioned in this comparison were not considered
by Kleinberg.
The remainder of this paper is organized as follows. Section 2 defines the GC model and states
the basic computational assumptions. Section 3 describes our three basic tools based on linear
programming, network flow, and an O(∆)-size representation of minimum cuts. Section 4 extends
these tools to optimize multiple objectives, analyze the structures of the space of all fittest protein
sequences, and generate near-fittest protein sequences. Section 5 gives some hardness results related
to counting fittest protein sequences and finding fittest protein sequences under additional restrictions.
Finally, Section 6 discusses our analysis of empirical 3D structures from the Protein Data Bank.
2
The Grand Canonical Model and Computational Assumptions
The Original Model Throughout this paper, all protein sequences are of n residues, unless explicitly stated otherwise. The GC model is specified by a fitness function Φ over all possible protein
sequences x with respect to a given 3D structure of n residues [18, 33]. In the model, to design a
protein sequence x is to specify which residues are hydrophobic (H) and which ones are polar (P ).
Thus, we model x as a binary sequence x1 , . . . , xn or equivalently as a binary vector (x1 , . . . , xn ),
where the i-th residue in x is H (respectively, P ) if and only if xi = 1 (respectively, 0). Then, Φ(x)
is defined as follows, where the smaller Φ(x) is, the fitter x is, as the definition is motivated by the
requirements that H residues in x (1) should have low solvent-accessible surface area and (2) should
be close to one another in space to form a compact hydrophobic core.
Φ(x) = α
X
g(di,j ) + β
i,j∈H(x),i<j−2
= α
X
g(di,j )xi xj + β
X
X
si
(1)
i∈H(x)
si xi , where
(2)
i
i<j−2
• H(x) = {i | xi = 1},
• the scaling parameters α < 0 and β > 0 have default values −2 and
require tuning for specific applications (see Section 4.4),
1
3
respectively and may
• si ≥ 0 is the area of the solvent-accessible contact surface for the residue (in Å) [10, 11],
• di,j > 0 is the distance between the residues i and j (in Å), and
• g is a sigmoidal function, defined by
g=
(
1
1+exp(di,j −6.5)
0
when di,j ≤ 6.5
when di,j > 6.5.
Extending the Model with Computational Assumptions Let opt(Φ) be the set of all protein sequences x that minimize Φ. This paper is generally concerned with the structure of opt(Φ).
Our computational problems assume that Φ is given as input; in other words, the computations of
α, β, si , g(di,j ) are not included in the problems. Also, for the sake of computational generality and
notational simplicity, we assume that α may be any nonpositive number, β any nonnegative number,
4
si any arbitrary number, and g(di,j ) any arbitrary nonnegative number; and that the terms g(di,j )
may range over 1 ≤ i < j ≤ n, unless explicitly stated otherwise. Thus, in the full generality of these
assumptions, Φ need not correspond to an actual protein 3D structure. Note that the relaxation
that si is any number is technically useful for finding Φ-minimizing protein sequences x that satisfy
additional constraints.
We write ai,j = −α·g(di,j ) ≥ 0 and bi = β·si and further assume that the coefficients ai,j and bi
are rational with some common denominator, that these coefficients are expressed with a polynomial
number of bits, and that arithmetic operations on these coefficients take constant time.
With these assumptions, we define the following sets of specific assumptions about Φ to be used
at different places of this paper.
P
P
F1 Let Φ(x) = − 1≤i<j≤n ai,j xi xj + 1≤i≤n bi xi , where ai,j ≥ 0, bi is arbitrary, and m of the
coefficients ai,j are nonzero. Let ∆ = n + m.
P
F2 For each β ≥ 0, let Φβ (x) = − 1≤i<j≤n ai,j xi xj + β
of the coefficients ai,j are nonzero. Let ∆ = n + m.
P
1≤i≤n si xi ,
F3 For each ℓ from 1 to f , let the ℓ-th fitness function Φℓ (x) = −
where aℓi,j ≥ 0 and bℓi is arbitrary. Let ∆ = f n2 .
P
where ai,j ≥ 0, si ≥ 0, and m
ℓ
1≤i<j≤n ai,j xi xj
+
P
ℓ
1≤i≤n bi xi ,
Sometimes we measure the dissimilarity between a fittest protein sequence x and a target protein
sequence x̂ in terms of Hamming distance. This distance is essentially the count of the positions i
where xi 6= x̂i and can be measured in two ways. The unweighted Hamming distance is |x − x̂|, where
P
P
|y| denotes the norm of vector y, i.e., ni=1 |yi |. The weighted Hamming distance is ni=1 wi ·|xi − x̂i |.
Throughout this paper, the weights w1 , . . . , wn are all arbitrary unless explicitly stated otherwise.
3
Three Basic Tools
This section describes our basic tools for computing fittest and near-fittest protein sequences. For
instance, Lemma 1 gives a representation of the problem of minimizing Φ as a linear program.
Lemma 2 further gives a representation of this problem as a minimum-cut problem, which generalizes
a similar representation of Kleinberg [18]. Theorem 5 gives a compact representation of the space
opt(Φ) using a Picard-Queyranne graph [26].
3.1
Linear Programming
From Equation 2, minimizing Φ(x) is an optimization problem in quadratic programming. Fortunately, because all the coefficients ai,j are nonnegative, it can be converted to a linear program, as
shown in Lemma 1.
Lemma 1 (characterizing Φ via linear program) Let Φ be as defined in Assumption F1. Consider the following linear program whose variables consist of the variables xi , together with new
variables yi,j for all i, j with ai,j 6= 0:
P
minimize Φ′ (x, y) = − ai,j yi,j +
subject to
0 ≤ xi ≤ 1 ∀i
0 ≤ yi,j ≤ 1
yi,j ≤ xi
∀i, j : ai,j 6= 0
yi,j ≤ xj
5
P
bi x i
(3)
There is a one-to-one correspondence that preserves x between the protein sequences that minimize
Φ(x) and the basic optimal solutions to Linear Program (3).
Proof: First, we show that for each 0-1 assignment to x there is a unique value of y that
minimizes Φ′ (x, y). Choose some yi,j , and suppose that either xi or xj is 0. Then yi,j is also 0 by
the constraint yi,j ≤ xi or yi,j ≤ xj . Alternatively, suppose xi and xj are both 1; then if yi,j is 0, Φ′
can be decreased by ai,j by setting yi,j to 1 without violating any constraints. Thus, in any optimal
integral solution to Linear Program (3), yi,j = min(xi , xj ) = xi xj .
P
P
Note that substituting xi xj for yi,j in Φ′ gives precisely − i,j ai,j xi xj + i bi xi = Φ(x); thus
minimizing Φ′ (x, y) is equivalent to minimizing Φ(x).
We now must show that all solutions to Linear Program (3) are integral. Every element of the
constraint matrix is either zero or ±1. Each row has either a single nonzero element (e.g, for the 0-1
bounds) or consists of zeroes and exactly one +1 and one −1. Thus the matrix is totally unimodular,
e.g., using [25, Theorem 13.3]. Since the right-hand side is integral, any vertex of the polytope
defined by Linear Program (3) is integral [25, Theorem 13.2]. Thus, all basic feasible solutions to
Linear Program (3) are 0-1 vectors.
So if (x, y) is a basic optimal solution to Linear Program (3), then x ∈ opt(Φ). Conversely, if
x ∈ opt(Φ), then the vector (x, y) in which yi,j = xi xj whenever ai,j is nonzero is an optimal solution
to Linear Program (3), which is a basic optimal solution since an appropriate subset of the constraints
0 ≤ xi ≤ 1 and 0 ≤ yi,j ≤ 1 form a basis.
Note that any xi with a negative coefficient bi is set to 1 in any optimal solution, as in this case
all terms containing xi have negative coefficients and are minimized when xi = 1. So an alternative
to allowing negative coefficients is to prune out any xi with a negative coefficient. This process must
be repeated recursively, since setting xi to 1 reduces terms of the form −ai,j xi xj to −ai,j xj , and may
yield more degree-1 terms with negative coefficients. To simplify our discussion, we let the linear
program (or, in Section 3.2, the minimum-cut algorithm) handle this pruning.
3.2
Network Flow
Recall that an s-t cut is a partition of the nodes of a digraph into two sets Vs and Vt , with s ∈ Vs and
t ∈ Vt . Also, a minimum s-t cut is an s-t cut with the smallest possible total capacity of all edges
from nodes in Vs to nodes in Vt .
In Kleinberg’s original construction [18], Φ(x) was minimized by solving an s-t minimum cut
problem in an appropriate digraph G. Lemma 2 describes a more general construction that includes
additional edges (s, vi ) to handle negative values for bi .
Lemma 2 (characterizing Φ via network flow) Let Φ be as defined in Assumption F1. Let GΦ
be a graph with a source node s, a sink node t, a node vi for each i, and a node ui,j for each i, j with
ai,j 6= 0, for a total of n + m + 2 = ∆ + 2 nodes. Let the edge set of GΦ consist of
• (s, ui,j ) for each ui,j , with capacity ai,j ,
• (vi , t) for each vi with bi > 0, with capacity bi ,
• (s, vi ) for each vi with bi < 0, with capacity −bi , and
• (ui,j , vi ) and (ui,j , vj ), for each ui,j , with infinite capacity,
6
for a total of Θ(∆) edges.
There is a one-to-one correspondence between the minimum s-t cuts in GΦ and the protein sequences in opt(Φ), such that vi is in the s-component of a cut if and only if xi = 1 in the corresponding
protein sequence.
Proof: We will show that the minimum s-t cuts in GΦ correspond to Φ-minimizing protein
sequences via Linear Program (3) of Lemma 1. Given a minimum s-t cut in G, let xi be 1 if vi
is in the s component, and 0 otherwise. Similarly, let yi,j be 1 if ui,j is in the s component, and
0 otherwise. Since no infinite-capacity edge (ui,j , vi ) or (ui,j , vj ) can appear in the cut, if ui,j is in
the s-component then vi and vj are as well. In terms of the x and y variables, we have yi,j ≤ xi
and yi,j ≤ xj whenever ai,j is nonzero, precisely the same constraints as in Linear Program (3).
Conversely, any 0-1 assignment (x, y) for which these constraints hold defines an s-t cut that does
not include any infinite-capacity edge.
Turning to the objective function, the total capacity of all edges in the cut is
X
ai,j (1 − yi,j ) +
ai,j 6=0
=
X
i,j
ai,j −
X
X
bi >0
bi −
bi <0
bi xi +
X
X
−bi (1 − xi )
bi <0
ai,j yi,j +
X
bi xi = K + Φ′ (x, y),
i
i,j
where K is a constant and Φ′ (x, y) is the objective function of Linear Program (3). Thus, the capacity
of the cut is minimized when Φ′ (x, y) is. The rest follows from Lemma 1.
Lemma 3 Let Φ be as defined in Assumption F1. Given Φ as the input, we can find an x ∈ opt(Φ)
in O(∆2 log ∆) time.
Proof: Given a digraph G = (V, E) as input, the Goldberg-Tarjan maximum-flow algorithm
takes O(|V ||E| log(|V |2 /|E|)) time [14]. We first apply Lemma 2 to Φ to obtain GΦ . We next use
this maximum-flow algorithm to find a minimum s-t cut in GΦ and then an optimal x from this cut.
All these steps take O(∆2 log ∆) total time.
3.3
A Compact Representation of Minimum Cuts
A given Φ may have more than one fittest protein sequence. Theorem 5 shows that opt(Φ) can be
summarized compactly using the Picard-Queyranne representation of the set of all minimum s-t cuts
in a digraph G [26], which is computed by the following steps:
1. computing any maximum flow φ in G;
2. computing strongly connected components in the residual graph Gφ whose edge set consists of
all edges in G that are not saturated by φ, plus edges (v, u) for any edge (u, v) that has nonzero
flow in φ;
3. contracting Gφ by contracting into single supernodes the set of all nodes reachable from s, the
set of all nodes that can reach t, and each strongly connected component in the remaining
graph.
The resulting graph Gs,t is a dag in which s and t are mapped to distinct supernodes by the
contraction. Furthermore, there is a one-to-one correspondence between the minimum s-t cuts in G
and the ideals in Gs,t , where an ideal is any node set I with the property that any predecessor of a
node in I is also in I.
7
Lemma 4 (see [26]) Given a digraph G with designated nodes s and t, there is a graph Gs,t together
with a mapping κ from V (G) to V (Gs,t ) with the following properties:
1. |V (Gs,t )| ≤ |V (G)|.
2. The node κ(s) has out-degree 0 while κ(t) has in-degree 0.
3. Given G as the input, Gs,t and κ can be computed using one maximum-flow computation and
O(|E(G)|) additional work.
4. A partition (Vs , Vt ) of V (G) is an s-t minimum cut in G if and only if Vt = κ−1 (I) for some
ideal I of Gs,t that contains κ(t) but not κ(s).
Combining Lemmas 2 and 4 gives the desired compact representation of the space of all fittest
protein sequences, as stated in the next theorem.
Theorem 5 (characterizing Φ via a dag) Let Φ be as defined in Assumption F1. There exists a
′
′
Φ
dag GΦ
s,t with designated nodes s and t and a mapping ρ from {1, . . . , n} to V (Gs,t ) with the following
properties:
1. GΦ
s,t has at most n + 2 nodes.
2
2. Given Φ as the input, GΦ
s,t and ρ can be computed in O(∆ log ∆) time.
3. There is a one-to-one correspondence between the protein sequences x ∈ opt(Φ) and the ideals of
b Φ = GΦ − {s′ , t′ }, in which xi = 0 if and only if ρ(i) = t′ or ρ(i) is in the ideal corresponding
G
s,t
s,t
to x.
Proof: The graph GΦ
s,t is obtained by applying Lemmas 2 and 4. Let κ be the contraction map
from Lemma 4. Let s′ = κ(s) and t′ = κ(t). The mapping ρ(i) is defined as κ(vi ).
Φ
To show that GΦ
s,t has at most n + 2 nodes, consider any node ui,j in G . Let φ be the maximum
Φ
flow used to define Gs,t . If φ(s, ui,j ) = 0, then ui,j is reachable from s in the residual graph GΦ
φ , and
ui,j is contracted onto the κ(s) supernode; if φ(s, ui,j ) > 0, then at least one of φ(ui,j , vi ) or φ(ui,j , vj )
is nonzero and ui,j is in the same strongly connected component in GΦ
φ as at least one of vi and vj .
In either case ui,j is contracted onto a supernode that contains s or some vi ; since the same thing
happens to all ui,j , there are at most n + 2 supernodes in GΦ
s,t : one for each vi , plus one for each of
s and t.
b Φ is justified by the observation that requiring t′ to be in an ideal and s′ to be
Using ideals of G
s,t
out of it has no effect on the presence or absence of other nodes, as t′ has no predecessors and s′ has
′
′
no successors in GΦ
s,t ; thus there is a one-to-one correspondence preserving all nodes except s and t
Φ
′
′
Φ
b .
between the ideals of Gs,t containing t but not s and the ideals of G
s,t
Remark. At some additional cost in time, Assumption F1 can be replaced in Theorem 5 by the
weaker assumption that Φ is submodular (i.e., that Φ(X ∪ Y ) + Φ(X ∩ Y ) ≤ Φ(X) + Φ(Y ) for all
X, Y , where each protein sequence x in Φ’s domain is regarded as the set X = {i | xi = 1} for
the purposes of taking unions and intersections). The reason is that a representation similar to the
Picard-Queyranne graph exists for the set of minima of any such submodular function. These minima
form a family closed under union and intersection, and any such family corresponds to the ideals of
an appropriate digraph [15, Proposition 10.3.3]. Such a representation can be computed efficiently,
as shown by Gabow [12].
8
Intuitively, what Theorem 5 says is the following. For any Φ, the residues in fittest protein
sequences are grouped into clusters, where the cluster ρ−1 (s) is always H, the cluster ρ−1 (t) is always
P , and for each of the remaining clusters, all residues in the cluster are either all H or all P . In
b Φ , such that if a cluster corresponding to the
addition, there is a dependence given by the edges of G
s,t
source of an edge is all H then the cluster at the other end is also all H.
There is no additional restriction on the structure of the space of all fittest protein sequences beyond those that follow from correspondence with the ideals of some digraph. As shown in Theorem 6,
b Φ , with any number of residues mapped to each supernode.
any graph may appear as G
s,t
Theorem 6 (characterizing a dag via Φ) Let Ĝ be an arbitrary digraph with n nodes, labeled 1 to
n, and m edges. Let Ĝ0 be the component graph of Ĝ obtained by contracting each strongly connected
component of Ĝ to a single supernode through a contraction map κ. Then, there exists some Φ as
defined in Assumption F1 such that for the GΦ
s,t and ρ defined in Theorem 5, an isomorphism exists
Φ
b
between Gs,t and Ĝ0 mapping each ρ(i) to κ(i).
Proof: Represent each node i in Ĝ by the variable xi . For each i, j, let ei,j = 1 if there is a
directed edge (i, j) in Ĝ, and 0 otherwise. To define Φ, let ai,j = ei,j + ej,i for each i < j; and, for
P
each i, let bi equal i’s out-degree δ+ (i) = j ei,j . Apply Lemma 2 to the resulting function Φ to get
a graph GΦ . Define a flow φ in GΦ as follows:
φ(s, ui,j )
φ(ui,j , vi )
φ(ui,j , vj )
φ(vi , t)
=
=
=
=
ei,j + ej,i
ei,j
ej,i
δ+ (i)
∀ui,j
∀ui,j
∀ui,j
∀vi
Note that this flow is a maximum flow because it saturates all edges leaving s as well as all edges
entering t. (It happens that this is the unique maximum flow, but we do not need this fact, as the
Picard-Queyranne construction works for any maximum flow.)
Φ
Our next goal is to show that the residual graph GΦ
φ of this flow contracts to Ĝ0 . Gφ has the
following classes of edges:
(ui,j , s)
(ui,j , vi )
(ui,j , vj )
(vi , ui,j )
(vj , ui,j )
(vi , t)
∀ui,j
when (i, j) ∈ Ĝ
when (j, i) ∈ Ĝ
∀vi
Φ
Since s has no successors and t has no predecessors in GΦ
φ , the supernodes in Gs,t containing s and t
consist of only s and t, respectively. Each node ui,j is in the same strongly-connected component as at
least one of vi or vj , so no other supernodes exist in GΦ
s,t that do not contain at least one of the nodes vi .
Note that every node-simple path from vi to vj in GΦ
φ is of the form vi , ui,q1 , vq1 , uq1 ,q2 , . . . , uqk−1 ,j , vj
(with the subscripts of each uqi ,qi+1 possibly reversed), which corresponds to a node-simple path
i = q1 , q2 , . . . , qk = j in Ĝ. The converse also holds. Therefore, i and j are in the same strongly
connected component in Ĝ if only if vi and vj are in the same strongly connected component in GΦ
s,t .
Φ
Now recall that for each i, ρ(i) is defined in Lemma 2 as the supernode in Gs,t containing vi . So an
Φ
edge from ρ(i) to ρ(j) in GΦ
s,t corresponds to a node-simple path from vi to vj Gs,t . By the above
Φ
path-to-path correspondence, every directed edge from ρ(i) to ρ(j) in Gs,t corresponds to an edge
9
b Φ is isomorphic to Ĝ0 , with the correspondence
from κ(i) to κ(j), and vice versa. In summary, G
s,t
ρ(i) ↔ κ(i) for all i.
4
Further Tools for Protein Landscape Analysis
4.1
Optimizing Multiple Objectives
We can extend the results of Section 3 beyond optimizing a single fitness function.
With more than one fittest protein sequence to choose from, we may wish to find a fittest protein
sequence x that is the closet to some target protein sequence x̂ in unweighted or weighted Hamming
distance. Theorem 7 shows that this optimization problem is as easy as finding an arbitrary fittest
protein sequence.
We may also wish to consider what protein sequences are simultaneously the fittest for more than
one fitness function. Theorem 8 shows how to compute a representation of this set similar to that
provided by Theorem 5.
Theorem 7 (optimizing Hamming distances and H-residue counts over opt(Φ)) Let Φ be
as defined in Assumption F1.
1. Given a target protein sequence x̂, some weights wi , and Φ as the input, we can find in
P
O(∆2 log ∆) time an x ∈ opt(Φ) with the minimum weighted Hamming distance i wi |xi − x̂i |
over opt(Φ).
2. Given Φ as the input, we can find in O(∆2 log ∆) time an x ∈ opt(Φ) with the largest (or
smallest) possible number of H residues over opt(Φ).
Proof: The statements are proved as follows.
Statement 1. Let ǫ be a positive constant at most 4W1nc , where W ≥ max |wi | and c is the common
denominator of all coefficients ai,j and bi . Below we show how to find a desired fittest protein sequence
P
by minimizing Φǫ (x) = Φ(x) + i ǫwi |xi − x̂i |.
First of all, since x and x̂ are 0-1 sequences, |xi − x̂i | = (xi − x̂i )2 = xi − 2x̂i xi + x̂i . Then, since
x̂ is given, Φǫ (x) can be minimized using Lemma 3 in O(∆2 log ∆) time.
Now suppose that y and z are two protein sequences with y ∈ opt(Φ) and z 6∈ opt(Φ). Then,
P
1
1
. Therefore, Φǫ (z) − Φǫ (y) ≥ 1c − 2W n ≥ 2c
.
Φ(z) − Φ(y) ≥ 1c . Also, | i wi |xi − x̂i || ≤ W n ≤ 4c
Thus every x that minimizes Φǫ (x) must also minimize Φ(x). The Hamming distance term in Φǫ
guarantees that from all x that do minimize Φ(x), minimizing Φǫ (x) selects one that also minimizes
this distance.
Statement 2. To find an x ∈ opt(Φ) with the largest (respectively, smallest) possible number of
H residues, apply Statement 1 with all wi = 1 and all x̂i = 1 (respectively, x̂i = 0).
Suppose we are given fitness functions Φ1 , . . . , Φf corresponding to multiple 3D structures, and
we wish to find protein sequences that are simultaneously optimal for each 3D structure. A simple
approach is to observe that the function Φ∗ = Φ1 + · · · + Φf satisfies Assumption F1, and that any
protein sequence that simultaneously optimizes each Φℓ optimizes Φ∗ . However, we must check any
minimum solution for Φ∗ to see that it is in fact a minimum solution for each Φℓ , as it may be that
the sets of minimum solutions of the Φℓ have empty intersection. Performing both the optimization
of Φ∗ and of the f individual fitness functions requires f + 1 network flow computations.
It turns out that we can reduce this cost to f network flows at the cost of some additional work
∗
Φℓ
to compute a composite Picard-Queyranne graph GΦ
s,t directly from the individual graphs Gs,t . This
10
approach, described in Theorem 8, is especially useful if we have already computed the individual
graphs for some other purpose.
Theorem 8 (minimizing multiple fitness functions) Let Φ1 , . . . , Φf be as defined in Assumpℓ
ℓ
ℓ
tion F3. For each ℓ, let GΦ
s,t and ρ be the dag and map computed from Φ in Theorem 5. Given
ℓ
ℓ
all GΦ
s,t and ρ as the input, there is an O(∆)-time algorithm that either (a) determines that there
∗
is no protein sequence x that simultaneously minimizes Φ1 through Φf , or (b) constructs a dag GΦ
s,t
∗
with designated nodes s′ and t′ and a mapping ρ∗ from {1, . . . , n} to V (GΦ
s,t ), such that there is a
one-to-one correspondence between the protein sequences x that simultaneously minimize all Φℓ (x)
b Φ∗ = GΦ∗ − {s′ , t′ }, in which xi = 0 if and only if ρ∗ (i) = t′ or ρ∗ (i) is in the ideal
and the ideals of G
s,t
s,t
corresponding to x.
Proof: By Theorem 5, the conditions below are necessary and sufficient for x to minimize Φℓ :
• xi = 1 if ρℓ (i) = s′ .
• xi = 0 if ρℓ (i) = t′ .
• xi = xj if ρℓ (i) = ρℓ (j).
b Φℓ .
• xi ≤ xj if (ρℓ (i), ρℓ (j)) is an edge in G
s,t
We will build a graph G whose nodes are s, t, and 1, . . . , n, and put in an edge (u, v) between
any nodes for which the constraint u ≤ v is required to minimize some Φℓ . In particular, we have the
following classes of edges, for each ℓ, where each class represents one of the above conditions:
• (s, v) and (v, s) whenever ρℓ (v) = s′ .
• (t, v) and (v, t) whenever ρℓ (v) = t′ .
• (u, v) and (v, u) whenever ρℓ (u) = ρℓ (v).
ℓ
b Φ )).
• (u, v) whenever (ρℓ (u), ρℓ (v) ∈ E(G
s,t
An assignment of 1 to s, 0 to t, and xi to each node i satisfies u ≤ v whenever (u, v) ∈ E(G) if
and only if all of the constraints required for x to simultaneously minimize all Φℓ are satisfied. Note
that such an assignment might not exist.
∗
To convert G into the desired graph GΦ
s,t , and to check whether there exist any assignments
meeting the constraints, contract each strongly connected component of G. If s and t are in the same
∗
strongly connected component, no simultaneous fittest solutions exist. Otherwise, let s′ in GΦ
s,t be
the supernode that contains s from G; let t′ be the supernode that contains t. Also, for each u in
1, . . . , n, let ρ∗ (u) be the supernode into which u is contracted. Then x simultaneously minimizes all
Φℓ if and only if xi = 1 when ρ∗ (i) = s′ , xi = 0 when ρ∗ (i) = t′ , and xi ≤ xj when (ρ∗ (i), ρ∗ (j)) is an
∗
Φ∗
edge in GΦ
s,t —precisely the condition that the zeroes in x correspond to nodes in some ideal of Gs,t
that contains t′ but not s′ .
To show the running time, observe that constructing the graph G takes O(∆) time, which dominates the contraction step.
11
4.2
The Space of All Fittest Protein Sequences
This section discusses some applications of the representation of the space opt(Φ) given by Theorem 5.
Theorem 9 gives an algorithm to enumerate this space. Theorem 10 gives an algorithm to compute the
diameter of the space in nonnegatively weighted Hamming distance. Theorem 11 gives an algorithm
to determine connectivity properties of the space with respect to various classes of mutations.
Theorem 9 (enumerating all protein sequences) Let Φ be as defined in Assumption F1. Given
the GΦ
s,t and ρ defined in Theorem 5 as the input, the protein sequences in opt(Φ) can be enumerated
in O(n) time per protein sequence.
b Φ |) = O(n)
b Φ in time O(|V (G
Proof: An algorithm of Steiner [32] enumerates the ideals of G
s,t
s,t
per ideal. For each ideal, invert the mapping ρ (in O(n) time) to recover the corresponding protein
sequence x.
Theorem 10 (computing the diameter) Let Φ be as defined in Assumption F1. Given the GΦ
s,t
and ρ defined in Theorem 5 as the input, it takes O(n) time to compute the diameter of opt(Φ) in
weighted Hamming distance where the weights wi are all nonnegative.
Proof: Any two fittest protein sequences x and y can differ only at indices i where ρ(i) ∈
/ {s, t}.
b Φ )). Then, d is an upper bound on the diameter.
Let d be the total weight of indices i ∈ ρ−1 (V (G
s,t
b Φ ) are both ideals of G
b Φ , and these ideals correspond to two
It is also a lower bound, as ∅ and V (G
s,t
s,t
protein sequences at distance d from each other.
b Φ to determine whether opt(Φ) is connected for various models of mutations. For
We can use G
s,t
instance, we can determine whether the space is connected for one-point mutations, in which at most
one residue changes with each mutation and all intermediate protein sequences must remain the
fittest. More generally, we can determine the minimum k so that the space is connected where each
mutation modifies at most k residues.
We adopt a general model proposed by Kleinberg [18]. In the model, there is a system Λ of subsets
of {1, . . . , n} that is closed downward, i.e., if A ⊆ B ∈ Λ, then A ∈ Λ. Two protein sequences x and y
are Λ-adjacent if they are in opt(Φ) and differ exactly at the positions indexed by elements of some
member of Λ. A Λ-chain is a sequence of protein sequences in opt(Φ) where each adjacent pair is
Λ-adjacent. Two protein sequences x and y are Λ-connected if there exists a Λ-chain between x and
y. A set of protein sequences is Λ-connected if every pair of elements of the set are Λ-connected. We
would like to tell for any given Λ and Φ whether particular protein sequences are Λ-connected and
whether the entire opt(Φ) is Λ-connected.
Kleinberg [18] gives polynomial-time algorithms for these problems that take Λ as input (via oracle
calls) and depend only on the fact that Φ is submodular. We describe a much simpler algorithm
b Φ from Theorem 5. This algorithm not only determines whether two protein sequences
that uses G
s,t
(alternatively, all protein sequences in opt(Φ)) are connected for any given Λ, but also determines
the unique minimum Λ for which the desired connectivity holds. Almost all of the work is done in
b Φ ; once we have this representation, we can read off the connectivity of opt(Φ)
the computation of G
s,t
directly.
Theorem 11 (connectivity via mutations) Let Φ be as defined in Assumption F1. The following
problems can both be solved in O(n) time.
12
′
1. Given the GΦ
s,t and ρ defined in Theorem 5 and two protein sequences x and x in opt(Φ) as the
input, compute the maximal elements of the smallest downward-closed set system Λ such that
x and x′ are Λ-connected.
2. Given the GΦ
s,t and ρ defined in Theorem 5 as the input, compute the maximal elements of the
smallest downward-closed set system Λ such that opt(Φ) is Λ-connected.
Proof: The statements are proved as follows.
b Φ such that the sets of zeros in x, x′ are ρ−1 (I), ρ−1 (I ′ ),
Statement 1. Let I, I ′ be the ideals in G
s,t
respectively. Let the maximal elements of Λ be the sets ρ−1 (v) over all v ∈ I ⊖ I ′ , where ⊖ is the
symmetric difference operator. Thus, Λ consists of these sets and all of their subsets. We will show
that Λ is the smallest downward-closed set system such that there is a Λ-chain between x and x′ in
opt(Φ).
First, consider some set system Λ′ where for some v ∈ I ⊖ I ′ , A = ρ−1 (v) is not in Λ′ . Recall
that x must be constant at the positions indexed by elements of A, where the constant depends on
b Φ corresponding to x. Partition opt(Φ) into sets Ω0 and Ω1 ,
whether or not v is in the ideal in G
s,t
j
where Ω consists of all z with zi = j for all i ∈ A. Since x and x′ differ on ρ−1 (v), one of them is in
Ω0 and the other is in Ω1 . However, since A ∈
/ Λ′ , no protein sequence in Ω0 is Λ′ -adjacent to one in
1
′
′
Ω . So there is no Λ -chain between x and x .
Conversely, to exhibit a Λ-chain from x to x′ , it suffices to show by iterations a Λ-chain from x to
the protein sequence y whose zeroes are given by ρ−1 (I ∩ I ′ ); the case of x′ is symmetric. Let I0 = I.
If at any iteration Ii = I ∩ I ′ , we are done. Otherwise, let v be a maximal element in Ii − (I ∩ I ′ ).
Then, Ii+1 = Ii − {v} is also an ideal. Since v ∈ I ⊖ I ′ , ρ−1 (v) ∈ Λ and the protein sequences
corresponding to Ii and Ii+1 are Λ-adjacent. After at most n such iterations, we reach y.
Statement 2. Let x and x′ be the protein sequences in opt(Φ) with the largest and the smallest
b Φ and its
possible numbers of H residues, respectively. In other words, x and x′ correspond to G
s,t
−1
Φ
b
empty ideal, respectively. If Λ includes ρ (v) for all v ∈ V (Gs,t ), then by Statement 1, there are
Λ-chains between any y ∈ opt(Φ) and x′ and thus between any two protein sequences in opt(Φ). If
it does not, then there is no Λ-chain between x and x′ . In summary, the maximal elements of Λ are
b Φ ).
the sets ρ−1 (v) over all v ∈ V (G
s,t
4.3
Generating Near-Fittest Protein Sequences
Finding good protein sequences other than the fittest is trickier, as Lemma 1 breaks down if we are
not looking at the fittest protein sequences. This section gives two algorithms that avoid this problem.
Theorem 12 describes an algorithm to generate all protein sequences x in order of increasing Φ(x).
Theorem 13 describes an algorithm to generate the fittest protein sequences at different unweighted
Hamming distances, which is useful for examining the trade-off between fitness and distance.
The algorithm for generating all protein sequences x in increasing order by Φ(x) is based on
Lemma 3 and a general technique for enumerating suboptimal solutions to combinatorial optimization
problems due to Lawler [21]. It is similar to an algorithm of Vazirani and Yannakakis [34] for
enumerating suboptimal cuts. We cannot use the Vazirani-Yannakakis algorithm directly because
suboptimal cuts in GΦ might include cuts corresponding to assignments in which yi,j is not equal to
xi xj for some i, j.
Theorem 12 (enumerating all protein sequences) Let Φ be as defined in Assumption F1. With
Φ as the input, we can enumerate all protein sequences x in order of increasing Φ(x) in time
O(n∆2 log ∆) per protein sequence.
13
Proof: For any length-k 0-1 sequence y = y1 , y2 , . . . , yk , let Ay be the set of all length-n 0-1
sequences x with xi = yi for each yi with 1 ≤ i ≤ k. Let ε be the empty sequence. Then, Aε is the set
of all length-n sequences. Observe that we can find an element z ∈ Ay that minimizes Φ(z) over Ay
in time O(∆2 log ∆) by setting zi = yi in Φ(z) for each yi and applying Lemma 3. Furthermore, the
set Ay − {z} is the disjoint union of the sets Ari for k + 1 ≤ i ≤ n, where ri = z1 , z2 , . . . , zi−1 , (1 − zi ).
To enumerate x in order of increasing Φ(x), we maintain a data structure that represents all
protein sequences less those already returned as a disjoint union of sets of the form Ay , together with
an Φ-minimizing element z for each, organized as a priority queue with key Φ(z). Initially, the queue
contains only (Aε , z), where z ∈ opt(Φ) is computed using Lemma 3 in time O(∆2 log ∆). At each
step, the smallest pair (Ay , z) is removed from the priority queue and is replaced by up to n pairs
(Ari , pi ), where pi is an Φ-minimizing element of Ari ; z is then returned. Each such step requires no
more than n applications of Lemma 3, and the cost of the at most n + 1 priority queue operations is
at most O(n log(2n )) = O(n2 ), giving a total cost of O(n∆2 log ∆) per value returned.
Let x̂ be a target protein sequence. For d ∈ {0, . . . , n}, let F (d) be the smallest Φ(x) over all
protein sequences x at unweighted Hamming distance d from x̂. A basic task of landscape analysis is
to plot the graph of F . As Theorem 18(2) in Section 5 shows, this task is computationally difficult
in general. Therefore, one way to plot the graph of F would be to use Theorem 12 to enumerate
all protein sequences x in order of increasing Φ(x) until for each d, at least one protein sequence at
distance d from x̂ has been enumerated. This solution may require processing exponentially many
protein sequences before F is fully plotted. As an alternative, Theorem 13 gives a tool for plotting
F approximately in polynomial time.
Theorem 13 (approximately plotting the energy-distance landscape) Let Φ be as defined
in Assumption F1. For each ǫ, let Φǫ (x) = Φ(x) + ǫ·|x − x̂|. Let Φ(ǫ) be the minimum Φǫ (x)
over all x.
1. Φ is a continuous piecewise linear concave function defined on R with at most n + 1 segments
and thus at most n + 1 corners.
2. Let (ǫ1 , Φ(ǫ1 )), . . . , (ǫk , Φ(ǫk )) be the corners of Φ, where ǫ1 < · · · < ǫk . Let di be the slope of
the segment immediately to the right of ǫi . Let d0 be the slope of the segment immediately to
the left of ǫ1 . Then, n = d0 > d1 > · · · > dk = 0.
3. Let d ∈ {0, 1, . . . , n}.
(a) F (d0 ) = Φ(ǫ1 ) − ǫ1 ·d0 . F (dk ) = Φ(ǫk ) − ǫk ·dk . For 0 < i < k, F (di ) = Φ(ǫi ) − ǫi ·di =
Φ(ǫi+1 ) − ǫi+1 ·di .
(b) For di > d > di+1 with 0 ≤ i < k, F (d) ≥ λF (di ) + (1 − λ)F (di+1 ), where λ =
d−di+1
di −di+1 .
4. Given Φ and x̂ as the input, we can compute (ǫ1 , Φ(ǫ1 )), . . . , (ǫk , Φ(ǫk )) and d0 , . . . , dk in
O(n∆2 log ∆) time.
Proof: The statements are proved as follows.
Statement 1. The concavity follows from the minimality of Φ(ǫ) and the fact that for any fixed
x, Φǫ (x) is linear in ǫ with slope |x − x̂|. Then, the continuous piecewise linearity and the counts of
segments and corners follow from the fact that |x − x̂| ∈ {0, 1, . . . , n}.
P
P
Statement 2. By the concavity of Φ, d0 > d1 > · · · > dk . Let W = 1 + 1≤i<j≤n ai,j + ni=1 |si |.
For all ǫ ≤ −W , Φǫ (x) is minimized if and only if x is at distance n from x̂. Similarly, for all ǫ ≥ W ,
Φǫ (x) is minimized if and only if x is at distance 0 from x̂. Therefore, d0 = n and dk = 0.
14
Statement 3. Case 3a is straightforward. To prove Case 3b, let x be a protein sequence that has the
smallest Φ(x) over all protein sequences at distance d from x̂. Then, F (d) = Φ(x), and Φǫi+1 (x) =
F (d) + ǫi+1 ·d. On the other hand, by the minimality of Φ, Φǫi+1 (x) ≥ Φ(ǫi+1 ). Furthermore, by
Case 3a, Φ(ǫi+1 ) = F (di+1 ) + ǫi+1 ·di+1 = F (di ) + ǫi+1 ·di . Thus,
F (di+1 ) ≤ F (d) + ǫi+1 ·d − ǫi+1 ·di+1 ;
F (di ) ≤ F (d) + ǫi+1 ·d − ǫi+1 ·di .
Case 3b follows from algebra and these two inequalities.
Statement 4. For given ǫ and x, let line(ǫ, x) be the line through the point (ǫ, Φǫ (x)) and with
slope |x − x̂|. Let Lǫ (respectively, Rǫ ) be the protein sequence x such that |x − x̂| is the largest
(respectively, smallest) possible over opt(Φǫ ). Note that Lǫ , Rǫ , line(ǫ, Lǫ ), and line(ǫ, Rǫ ) can be
computed in O(∆2 log ∆) total time using Theorem 7. Furthermore, line(ǫ, Lǫ ) and line(ǫ, Rǫ ) contain
the segments of Φ immediately to the left and the right of ǫ, respectively. Consequently, ǫ is a corner
of Φ if and only if line(ǫ, Lǫ ) 6= line(ǫ, Rǫ ).
To compute the corners and slopes of Φ, we first describe a recursive corner-slope finding subroutine as follows. The subroutine takes as input an interval [ǫ′ , ǫ′′ ] where ǫ′ < ǫ′′ together with
line(ǫ′ , Rǫ′ ) and line(ǫ′′ , Lǫ′′ ). It outputs all the corners (ǫ, Φ(ǫ)) of Φ together with slopes |Lǫ − x̂|
and |Rǫ − x̂| where ǫ′ < ǫ < ǫ′′ . There are two cases.
Case 1: line(ǫ′ , Rǫ′ ) = line(ǫ′′ , Lǫ′′ ). Then, there is no corner over the interval (ǫ′ , ǫ′′ ), and thus
the subroutine call ends without reporting any new corner or slope.
Case 2: line(ǫ′ , Rǫ′ ) 6= line(ǫ′′ , Lǫ′′ ). Then, compute ǫ′′′ at which line(ǫ′ , Rǫ′ ) and line(ǫ′′ , Lǫ′′ )
intersect; by the concavity of Φ stated in Statement 1, ǫ′ < ǫ′′′ < ǫ′′ . Also, compute line(ǫ′′′ , Lǫ′′′ ) and
line(ǫ′′′ , Rǫ′′′ ). There are two subcases:
Case 2a: line(ǫ′′′ , Lǫ′′′ ) 6= line(ǫ′′′ , Rǫ′′′ ). Then the subroutine returns (ǫ′′′ , Φ(ǫ′′′ )) as a new corner
together with slopes |Lǫ′′′ − x̂| and |Rǫ′′′ − x̂| and recurses on the intervals [ǫ′ , ǫ′′′ ] and [ǫ′′′ , ǫ′′ ].
Case 2b: line(ǫ′′′ , Lǫ′′′ ) = line(ǫ′′′ , Rǫ′′′ ). The subroutine returns no new corner or slope but
recurses on the intervals [ǫ′ , ǫ′′′ ] and [ǫ′′′ , ǫ′′ ]. In this case, the subroutine has found the line containing
a new segment of Φ, i.e., the segment through the point (ǫ′′′ , Φ(ǫ′′′ )).
This completes the description of the subroutine. The running time of this subroutine is dominated
by that for computing Lǫ′′′ and Rǫ′′′ and thus is O(∆2 log ∆).
With this subroutine, we can find the corners and slopes of Φ as follows. Recall W from the
proof of Statement 2. Note that if ǫ ≤ −W or ǫ ≥ W , then Φ has no corner at ǫ. So we compute
line(−2W, R−2W ) and line(2W, L2W ) and apply the subroutine to the interval [−2W, 2W ] to find all
the corners and slopes of Φ. This algorithm makes O(n) recursive calls to the subroutine since by
Statement 1, there are only O(n) corners and segments, and each recursive call finds at least one
new corner or segment. The running time of the algorithm is dominated by the total running time
of these calls and thus is O(n∆2 log ∆) as stated in the statement.
4.4
Tuning the Parameters of the GC Model
This section shows how to systematically tune the parameters α and β so that a fittest protein
sequence for a given 3D structure matches the 3D structure’s native protein sequence as closely as
possible in terms of unweighted or weighted Hamming distance. For this purpose, we assume si ≥ 0.
Furthermore, since the fitness function does not have an absolute scale, we may fix α at −1 and vary
β. In summary, this section adopts Assumption F2.
Let ΠΦ be the set of indices i with si 6= 0. The next lemma shows that for any fittest protein
sequence x of Φβ , the set H(x) ∩ ΠΦ of H residues with nonzero surface area si is monotone in β.
15
Then, as shown in Theorem 16, to tune β, we only need to consider at most n + 1 possible values of
β.
Lemma 14 Let Φ be as defined in Assumption F2. Let y and z be any fittest protein sequences for
Φβ1 and Φβ0 , respectively. If β1 < β0 , then H(y) ∩ ΠΦ ⊇ H(z) ∩ ΠΦ .
Proof: Let H1 = H(y) and H0 = H(z). Let H10 = H1 − H0 , i.e., the set of indices i for which
xi changes from 1 to 0 when β changes from β1 to β0 . Similarly, let H01 = H0 − H1 . Further, let
P
P
B10 = i∈H10 si and B01 = i∈H01 si .
P
P
Let A = i,j∈H0 ai,j − i,j∈H1 ai,j . Then, Φβ (z) − Φβ (y) = −A + β(B01 − B10 ). Let A01 =
P
i,j∈H0 ∧ {i,j}∩H01 6=∅ ai,j . Let A10 = −(A − A01 ), which is the sum of the terms ai,j that Φβ loses
when xi changes from 1 to 0. Similarly, A01 is the sum of the terms ai,j that Φβ gains when xi
changes from 1 to 0.
To show H(y) ∩ ΠΦ ⊇ H(z) ∩ ΠΦ , we need to prove H01 ∩ ΠΦ = ∅ or equivalently B01 = 0. To do
so by contradiction, suppose B01 > 0. There are two cases:
Case 1: −A01 + β0 B01 > 0. Notice that the protein sequence z ′ with H(z ′ ) = H0 − H01 has a
smaller fitness value for β0 than z does, contradicting the minimality of z.
Case 2: −A01 + β0 B01 ≤ 0. Then, −A01 + β1 B01 < 0. Therefor, the protein sequence y ′ with
H(y ′ ) = H1 ∪ H01 has a smaller fitness value for β1 than y does, contradicting the minimality of y.
Let Φ(β) be the minimum Φβ (x) over all x. The next lemma characterizes the structure of Φ(β).
This structure is then used to tune β in Theorem 16.
Lemma 15 Let Φ be as defined in Assumption F2.
1. Φ is a continuous piecewise linear concave function defined on [0, ∞) with at most n+1 segments
and thus at most n + 1 corners.
2. For all β1 < β3 < β4 < β2 where (β1 , Φ(β1 )) and (β2 , Φ(β2 )) are adjacent corners of Φ, we
have opt(Φβ3 ) = opt(Φβ4 ), opt(Φβ3 ) ⊆ opt(Φβ1 ), and opt(Φβ3 ) ⊆ opt(Φβ2 ). Similarly, for all
β1 < β3 < β4 where (β1 , Φ(β1 )) is the rightmost corner, we have opt(Φβ3 ) = opt(Φβ4 ) and
opt(Φβ3 ) ⊆ opt(Φβ1 ).
3. Given Φ as the input, it takes O(n∆2 log ∆) time to find the set of all β such that (β, Φ(β)) is
a corner of Φ.
Proof: The statements are proved as follows.
Statement 1. The concavity follows from the minimality of Φ(β) and the fact that for any fixed
P
x, Φβ (x) is linear in β with slope i∈H(x)∩ΠΦ si . Then, the continuous piecewise linearity and the
counts of segments and corners follow from Lemma 14.
Statement 2. The proofs for the case that (β1 , Φ(β1 )) is the rightmost corner and the complementary case are similar. So we only detail the proof for the former. Note that (β3 , Φ(β3 )) is not a
corner. So for every fixed x ∈ opt(Φβ3 ), the line Φβ (x) goes through the corner (β1 , Φ(β1 )) and the
point (β4 , Φ(β4 )). Thus, x ∈ opt(Φβ1 ) and x ∈ opt(Φβ4 ). By symmetry, for every y ∈ opt(Φβ4 ), we
have y ∈ opt(Φβ3 ). In summary, opt(Φβ3 ) = opt(Φβ4 ) and opt(Φβ3 ) ⊆ opt(Φβ1 ).
Statement 3. For any given β and x, let line(β, x) be the line through the point (β, Φβ (x))
P
and with slope i∈H(x)∩ΠΦ si . For each β, let Lβ (respectively, Rβ ) be the protein sequence x
such that H(x) has the largest (respectively, smallest) possible cardinality over opt(Φβ ). Note that
Lβ , Rβ , line(β, Lβ ), and line(β, Rβ ) can be computed in O(∆2 log ∆) total time using Theorem 7.
16
Furthermore, line(β, Lβ ) and line(β, Rβ ) contain the segments of Φ immediately to the left and the
right of β, respectively. Consequently, for β > 0, β is a corner of Φ if and only if line(β, Lβ ) 6=
line(β, Rβ ). Also, (0, Φ(0)) is the leftmost corner, and the segment of Φ to the right of 0 is contained
by line(0, R0 ).
To compute the corners of Φ, we first describe a recursive corner-finding subroutine as follows.
The subroutine takes as input an interval [β1 , β2 ] where β1 < β2 together with line(β1 , Rβ1 ) and
line(β2 , Lβ2 ). It outputs all the corners (β, Φ(β)) of Φ with β1 < β < β2 . There are two cases.
Case 1: line(β1 , Rβ1 ) = line(β2 , Lβ2 ). Then, there is no corner over the interval (β1 , β2 ), and thus
the subroutine call ends without reporting any new corner.
Case 2: line(β1 , Rβ1 ) 6= line(β2 , Lβ2 ). Then, compute β3 at which line(β1 , Rβ1 ) and line(β2 , Lβ2 )
intersect; by the concavity of Φ stated in Lemma 15(1), β1 < β3 < β2 . Also, compute line(β3 , Lβ3 )
and line(β3 , Rβ3 ). There are two subcases:
Case 2a: line(β3 , Lβ3 ) 6= line(β3 , Rβ3 ). Then the subroutine returns (β3 , Φ(β3 )) as a new corner
and recurses on the intervals [β1 , β3 ] and [β3 , β2 ].
Case 2b: line(β3 , Lβ3 ) = line(β3 , Rβ3 ). The subroutine returns no new corner but recurses on
the intervals [β1 , β3 ] and [β3 , β2 ]. In this case, the subroutine has found the line containing a new
segment of Φ, i.e., the segment through the point (β3 , Φ(β3 )).
This completes the description of the subroutine. The running time of this subroutine is dominated
by that for computing Lβ3 and Rβ3 and thus is O(∆2 log ∆).
With this subroutine, we can find the corners of Φ as follows. If every si = 0, then (0, Φ(0)) is
P
the only corner. Otherwise, let β∞ be 1 + 1≤i<j≤n ai,j divided by the smallest nonzero si . Note
that for every β ≥ β∞ , Φ has no corner at β, Then, we compute line(0, L0 ) and line(0, Rβ∞ ). We
report the leftmost corner (0, Φ(0)) and apply the subroutine to the interval [0, β∞ ] to find all the
other corners of Φ.
This algorithm makes O(n) recursive calls to the subroutine since by Lemma 15, there are only
O(n) corners and segments, and each recursive call finds at least one new corner or segment. The
running time of the algorithm is dominated by the total running time of these calls and thus is
O(n∆2 log ∆) as stated in the lemma.
Theorem 16 (tuning α and β) Let Φ be as defined in Assumption F2. Given a target protein
sequence x̂ and Φ as the input, we can find in O(n∆2 log ∆) time the set of all β where the closest
unweighted (or weighted) Hamming distance between x̂ and any protein sequence in opt(Φβ ) is the
minimum over all possible β.
Proof: The proofs for the cases of unweighted and weighted Hamming distances are similar. So
we only detail the proof for the unweighted case. Our algorithm for finding all distance-minimizing
choices of β has three stages.
Stage 1. Use Lemma 15(3) to find all the corners of Φ.
Stage 2. For each corner (β, Φ(β)), use Theorem 7 to compute the closest Hamming distance dβ
between x̂ and any protein sequence in opt(Φβ ). Let dmin be the smallest dβ over all corners. Then,
by Lemma 15(2), report all β with dβ = dmin as desired choices of distance-minimizing β.
Stage 3. Consider each segment of Φ. Let β1 and β2 be the vertical coordinates of the left and
right endpoints of the segment. Find a suitable β3 in the open interval (β1 , β2 ) as follows. If β2
is finite, then set β3 = (β1 + β2 )/2; otherwise, set β3 = β1 + 1. Then use Theorem 7 to compute
the closest unweighted or weighted Hamming distance dβ3 between x̂ and any protein sequence in
opt(Φβ3 ). If dβ3 = dmin , then by Lemma 15(2), report that every β in the interval (β1 , β2 ) is a desired
distance-minimizing β.
17
This completes the description of the algorithm. By Lemma 15(3), Stage 1 takes O(n∆2 log ∆)
time. By Theorem 7 and Lemma 15 (1), Stages 2 and 3 also take O(n∆2 log ∆) time. Thus, the total
running time is as stated in the theorem.
5
Computational Hardness Results
Theorem 17 (hardness of counting and averaging) Let Φ be as defined in Assumption F1.
The following problems are all #P-complete:
1. Given Φ as the input, compute the cardinality of opt(Φ).
2. Given Φ1 , . . . , Φf as the input, where f is any fixed positive integer and Φ1 , . . . , Φf are as defined
in Assumption F3, compute the number of protein sequences x that simultaneously minimize
Φℓ (x) for all ℓ = 1, . . . , f .
3. Given Φ as the input, compute the average norm |x|, i.e., the average number of H residues in
x, over all x ∈ opt(Φ).
4. Given Φ and a target protein sequence x̂ as the input, compute the average unweighted Hamming
distance |x − x̂| over all x ∈ opt(Φ).
5. Given Φ, a target protein sequence x̂, and an integer d as the input, compute the number of
protein sequences in opt(Φ) at unweighted Hamming distance d from x̂.
Proof: Note that each of the problems is in #P, because we can recognize an element of opt(Φ)
in polynomial time using Lemma 3. So to prove #P-completeness we must only show that each
problem is #P-hard.
Statement 1. Reduce from the problem of counting the number of ideals in a dag, which is #Pb Φ is isomorphic to Ĝ.
hard [28]. Given a dag Ĝ, apply Theorem 6 to get a function Φ for which G
s,t
bΦ ∼
By Theorem 5, counting opt(Φ) is then equivalent to counting the number of ideals of G
s,t = Ĝ.
Statement 2. The problem in Statement 1 is a special case of the problem in this statement.
Statement 3. Using the same construction as in Statement 1, we can reduce from the problem
of computing the average cardinality of an ideal in Ĝ. To see that this latter problem is #P-hard,
suppose that we can compute the average cardinalities of ideals in an n-node Ĝ and in an augmented
graph Ĝ′ obtained from Ĝ by adding a single new node s and edges from every v ∈ V (Ĝ) to s. Let c
be the average for Ĝ and c′ the average for Ĝ′ . Let N be the number of ideals in Ĝ. Then c = K/N
for some K, while c′ = (K + n + 1)/(N + 1), since the only new ideal in Ĝ′ consists of s and all other
nodes, and thus has size n + 1. Solving for N gives N = (n + 1 − c′ )/(c′ − c), which can be computed
from c, c′ , and n.
Statement 4. This problem has the problem in Statement 3 as a special case with all x̂i = 0.
Statement 5. To reduce the problem of counting protein sequences in opt(Φ) to counting protein
b Φ given by Theorem 5, and add
sequences at a given unweighted Hamming distance, take the dag G
s,t
to each node vi a node qi with edges (vi , qi ) and (qi , vi ). Apply Theorem 6 to this new graph Ĝ′ to
obtain a function Φ′ for which opt(Φ′ ) is in one-to-one correspondence with the set of ideals of the
strongly-connected component graph Ĝ′0 of Ĝ′ . Each strongly connected component of Ĝ′ consists
b Φ . Now we choose x̂ with x̂v = 0 and x̂q = 1.
of vi and qi for some i, so Ĝ′0 is isomorphic to G
s,t
i
i
Then, the contribution to |x − x̂| of each pair xvi , xqi is 1 regardless of their common value. Thus, the
18
number of Φ-minimizing protein sequences equals the number of Φ′ -minimizing protein sequences at
bΦ .
distance d from x̂, where d is the number of nodes in G
s,t
Theorem 18 (hardness of plotting the energy-distance landscape) Let Φ be as defined in
Assumption F1.
1. Given Φ and two integers d1 , d2 as the input, it is NP-complete to determine whether there is
an Φ-minimizing x with d1 ≤ |x| ≤ d2 .
2. Let x̂ be a target protein sequence. For d ∈ {0, . . . , n}, let F (d) be the smallest Φ(x) over all
protein sequences x at unweighted Hamming distance d from x̂. Given Φ and d as the input, it
is NP-hard to compute F (d).
Proof: Statement 2 follows from the fact that the problem in Statement 1 can be reduced to
the problem in this statement in polynomial time. Statement 1 is proved as follows.
Since we can recognize Φ-minimizing protein sequences using Lemma 3, the problem is clearly in
NP. To show that it is NP-hard, we reduce from PARTIALLY ORDERED KNAPSACK, problem
MP12 from Garey and Johnson [13, pp. 247–248].
The input to PARTIALLY ORDERED KNAPSACK consists of a partially-ordered set U , each
element u of which is assigned a size s(u) ∈ Z+ and a value v(u) ∈ Z+ , together with a upper bound
B1 on total size and a lower bound B2 on total value. The problem is to determine whether there
P
P
exists an ideal I in U such that u∈I s(u) ≤ B1 and u∈I v(u) ≥ B2 . Garey and Johnson note
that the problem, even with s(u) = v(u) for all u ∈ U , is NP-complete in the strong sense (meaning
that there is some polynomial bound on the size of all numbers in the input with which it remains
NP-complete).
Given an instance of PARTIALLY ORDERED KNAPSACK with s(u) = v(u) for all u and all
numbers bounded by some polynomial p, build a graph Ĝ where each u ∈ U is represented by a
clique C(u) of s(u) nodes, and there is an edge from C(u) to C(u′ ) if and only if u ≺ u′ . Note
that because s(u) ≤ p, Ĝ has polynomial size. Apply Theorem 6 to generate a function Φ (in
b Φ is isomorphic to the component graph Ĝ0 obtained by contracting
polynomial time) such that G
s,t
all strongly connected components of Ĝ. Since the strongly connected components of Ĝ are precisely
bΦ ∼
bΦ
the cliques C(u), G
U , interpreted as a dag. In particular any ideal of G
s,t = Ĝ0 is isomorphic to
s,t
P
corresponds to an ideal I of U . Let N = u∈U s(u). The norm of the corresponding vector |x| is
P
P
P
N − u∈I |C(u)| = N − u∈I s(u) = N − u∈I v(u). Set d1 = N − B1 , d2 = N − B2 , and we have
the problem stated in the theorem.
6
Applications to Empirical Protein 3D Structures
To demonstrate our algorithms, we chose 34 proteins with known 3D structures from the Protein
Data Bank (PDB) at http://www.rcsb.org/pdb. These 3D structures included 8 from Kleinberg’s
study [18] but excluded the protein fragments and multimeric proteins used in that study. The chosen
3D structures were then represented by centroids for each side chain calculated from the coordinates
of each atom in the side chain; in the case of 3D structures solved by NMR, hydrogen atoms were
included into centroid calculations. For glycine, the centroid was taken to be the position of Cα . For
each side chain, the area of solvent accessible surface was computed via the Web interface of the ASC
program with default parameters [10]. In accordance to the GC model, each of the chosen native
19
Name
1a7m
1a8y
1ab3
1ab7
1agi
1air
1b71
1ble
1bpi
1bw3
1clh
1ehs
1gym
1nar
1prn
1thv
1xnb
2aak
2bnh
2cba
2erl
2stv
6yas
8cho
9rat
1aaj
1aba
1bba
1brq
1cis
1hel
1ifb
3cln
3rn3
Solvent/Length
51.23
81.37
399.05
451.48
384.30
173.96
374.23
498.45
1453.75
732.31
607.48
2178.29
392.97
447.48
498.38
741.41
871.63
1130.39
412.43
773.98
5064.83
1153.80
904.99
2054.72
2126.95
51.89
125.39
584.25
272.54
780.88
539.14
583.57
835.80
1018.66
Length
180
338
88
89
125
352
191
161
58
125
166
48
296
289
289
207
185
150
456
258
40
184
256
125
124
105
87
36
174
66
129
131
143
124
α/β
-295.8
-155.7
-326.7
-79.5
-93
-0.3
-23.7
-269.4
-31.5
-33.9
-9.3
-32.1
-4.5
-6.3
-2.7
-10.5
-135.6
-149.1
-253.8
-2.4
-47.7
-3.6
-12.6
-423.6
-89.4
-155.1
-245.4
-69.6
-27.3
-454.8
-18.6
-40.5
-309.9
-88.8
% Similarity
74.44
73.37
78.41
80.90
77.60
69.89
69.11
72.67
68.97
70.40
69.28
72.92
73.99
75.78
56.40
71.01
65.95
78.67
78.51
73.64
80.00
64.67
67.19
72.80
75.00
70.48 (72)
78.16 (70)
66.67 (58)
72.99 (71)
69.70 (64)
76.74 (78)
79.39 (70)
72.03 (70)
72.58 (69)
Description
cytokine
Ca binding protein
ribosomal protein
ribonuclease inhibitor
endonuclease
pectate lyase
electron transport
phosphotransferase
proteinase inhibitor
lectin
cyclophilin
enterotoxin
phospholipase
plant seed protein
porin
sweet tasting protein
glycosidase
ubiquitin conjugation
ribonuclease inhibitor
lyase
pheromone
viral coat protein
lyase
isomerase
ribonuclease A
electron transport
electron transport
pancreatic hormone
retinol transport
lysozyme
fatty acid bind protein
Ca binding protein
ribonuclease
electron transport
Table 1: This table gives some statistics of our computed fittest H/P protein sequences for the 34
empirical protein 3D structures chosen from PDB. Column 1 contains the PDB name of the native
protein sequence, where the proteins from Kleinberg [18] are in bold. Column 2 shows the length
P
normalized solvent accessibility, i.e., i∈H(x) si /n, of the computed fittest protein sequence. Column
4 lists the protein sequence length n, i.e., the count of amino acids. Column 4 provides the computed
optimal ratio of the α and β parameters (see text). Column 5 displays the percentage of similarity
of the computed fittest protein sequence to the native protein sequence. Column 6 describes the
function of the native protein.
20
%Similarity
80
2bnh
70
G
N
60
0
100
200
PFAM
Figure 1: This plot displays the relationship between the percent similarity of computed fittest protein
sequences to native protein sequences versus the PFAM family size of native protein sequences.
Globular native proteins are shown as open circles while nonglobular ones are shown as crosses.
protein sequences was converted into a binary H/P sequence following Sun et. al. [33], where A, C,
F, I, L, M, V, W, Y are H, and the other amino acids are P .
We used Equation 1 in the GC model to calculate fitness values of protein sequences to determine
minimal energy values and consequently to compute the “fittest” protein sequences. These fitness
values consist of two terms in Equation 1. The first term accounts for the idea that hydrophobic
residues tend to cluster together due to stacking forces from the solvent. The second term accounts
for the idea that hydrophobic residues tend to avoid solvent accessible surfaces of the molecule.
The arbitrary parameters α and β represent scaling factors for the relative importance of these two
tendencies. We expected the appropriate ratio of these two values to depend on the type of a protein
(globular, nonglobular, and membrane) and the length of the protein. Therefore, we optimized the
scale of the two parameters to find a ratio that maximizes the similarity of a fittest protein sequence
to the native protein sequence, similarly in spirit to Kleinberg’s scaling algorithm [18].
Results of this optimization are shown in Table 1. As anticipated, our algorithms computed
fittest protein sequences that are closer to native protein sequences than found by Kleinberg [18],
whose results are shown in parenthesis in Table 1. (Note that protein 1aaj is an exception to this
improvement on proximity—perhaps due to the fact that the input data are not exactly the same.)
However, proximity to native protein sequences is not a good proxy for biological relevance of the
algorithms. Determination of protein 3D structure involves an energy landscape given by a statistical
thermodynamic energy function E(X , D) where X is the set of amino acid sequences and D is the set
of possible folded 3D structures. On the one hand, for a fixed protein sequence x, the 3D structure D
21
is determined by a temperature-dependent folding process that minimizes E(x, D) over D. Ab initio
solutions to this problem for protein sequences of practical length currently do not exist. On the other
hand, for a fixed 3D structure D, no known thermodynamic reasons connect the folding process with
a protein sequence x that minimizes E(x, D) over X . However, we can invoke evolutionary processes
as possibly selecting for those protein sequences that produce the most stable 3D structures, i.e.,
those with the lowest E(x, D), at a given temperature. In other words, given a suite of protein
sequences that fold into a particular 3D structure to perform a biological function, there might be
selection for the most thermodynamically stable protein sequences.
From the view point of evolutionary selection, the difference between computed fittest protein
sequences and native protein sequences may be attributed to three factors. First, the GC toy thermodynamic model is inappropriate. Second, the biological function of a protein actually requires
structural lability. Last, a native protein is part of a diverse family and other members of the family
lie closer to the computed fittest protein sequence. This last factor can be augmented by the argument
that if a computed fittest protein sequence exists in nature but is very different from the native protein
sequence, it is likely that many other protein sequences (thus a diverse family of protein sequences)
exist in nature and fold into the same or similar 3D structures. All of these factors are likely to play
in the data shown in Table 1. However, we conjectured a significant relationship between a computed
fittest protein sequence’s similarity to a native protein sequence and the diversity of the native protein
in nature. Such a relationship would be highly intriguing biologically. We examined this conjecture
by assessing the diversity of native proteins using the database PFAM at http://pfam.wustl.edu,
which is a database of protein families determined through Hidden Markov Models [4]. The database
contained information on the putative family size of 25 of our 34 chosen native proteins. Figure 1
shows the plot of the percent similarity of computed fittest protein sequences to native protein sequences versus the PFAM family size of native proteins. In the figure, globular proteins are shown
as circles and nonglobular ones as crosses. There is a negative linear trend as suggested by our conjecture. Linear regression is nearly significant at 0.05 level with p = 0.088. The figure shows three
outliers, 2bnh, 2stv, and 8cho. Of these, 2bnh is an exceedingly strange 3D structure with a protein
sequence of alpha helixes forming a horseshoe shaped sheet, resulting in a 3D structure that is very
deviant from globular proteins which are the genesis of the original thermodynamic model. Leaving
out this outlier results in a significant linear regression with p = 0.015.
There is still considerable uncertainty about the appropriateness of the GC toy model. The
average percentage of the hydrophobic residues is 42% in the native protein sequences compared to
35% in the computed fittest protein sequences. More importantly, the standard deviation of the
percentage of hydrophobic residues is 0.054 in the native protein sequences compared to 0.143 for
the computed fittest protein sequences. Thus, the percentage of hydrophobic residues is relatively
constant in the native protein sequences, reflecting perhaps a functional need or unknown structural
factors. In contrast, the percentage of hydrophobic residues in a computed fittest protein sequence
tends to vary depending on the 3D structure. This suggests that it might be important to introduce
a hydrophobic residue percentage constraint into optimization algorithms in the future as suggested
in the sliding algorithm of Kleinberg [18]. Nevertheless, our preliminary results show that even such
a simplified toy model might be useful for exploratory investigations of protein evolution especially
when coupled to computationally efficient algorithms to allow systematic investigation of the roughly
13,000 empirical protein 3D structures. We are currently planning a large-scale analysis of further
empirical protein 3D structures; the results will be reported in a subsequent paper.
22
Acknowledgments
We wish to thank Jon Kleinberg for generously providing LATEX entries for many of the references;
thank Lisa Fleischer and Hal Gabow for their help in tracking down references used to generalize
Theorem 5 for submodular functions; and thank Mark Gerstein for helpful discussions.
References
[1] J. Atkins and W. E. Hart. On the intractability of protein folding with a finite alphabet of amino
acids. Algorithmica, 25(2-3):279–294, 1999.
[2] A. Babajide, I. Hofacker, M. Sippl, and P. Stadler. Neutral networks in protein space: A
computational study based on knowledge-based potentials of mean force. Folding and Design,
2:261–269, 1997.
[3] J. Banavar, M. Cieplak, A. Maritan, G. Nadig, F. Seno, and S. Vishveshwara. Structure-based
design of model proteins. Proteins: Structure, Function, and Genetics, 31:10–20, 1998.
[4] A. Bateman, E. Birney, R. Durbin, S. R. Eddy, K. L. Howe, and E. L. L. Sonnhammer. PFAM– A
database of protein domain family alignments and HMMs. Nucleic Acids Research, 28:263–266,
2000.
[5] B. Berger and T. Leighton. Protein folding in the hydrophobic-hydrophilic (HP) model is NPcomplete. Journal of Computational Biology, 5(1):27–40, 1998.
[6] P. Crescenzi, D. Goldman, C. Papadimitriou, A. Piccolboni, and M. Yannakakis. On the complexity of protein folding. Journal of Computational Biology, pages 423–466, 1998.
[7] J. M. Deutsch and T. Kurosky. New algorithm for protein design. Physical Review Letters,
76:323–326, 1996.
[8] K. A. Dill, S. Bromberg, K. Yue, K. Fiebig, D. Yee, P. Thomas, and H. S. Chan. Principles of
protein folding — A perspective from simple exact models. Protein Science, 4:561–602, 1995.
[9] K. E. Drexler. Molecular engineering: An approach to the development of general capabilities
for molecular manipulation. Proceedings of the National Academy of Sciences of the U.S.A.,
78:5275–5278, 1981.
[10] F. Eisenhaber and P. Argos. Improved strategy in analytic surface calculation for molecular systems: Handling of singularities and computational efficiency. Journal of Computational
Chemistry, 14(N11):1272–1280, 1993.
[11] F. Eisenhaber, P. Lijnzaad, P. Argos, C. Sander, and M. Scharf. The double cube lattice method:
Efficient approaches to numerical integration of surface area and volume and to dot surface
contouring of molecular assemblies. Journal of Computational Chemistry, 16(N3):273–284, 1995.
[12] H. N. Gabow. Applications of a poset representation to edge connectivity and graph rigidity. In
Proceedings of the 32nd Annual IEEE Symposium on Foundations of Computer Science, pages
812–821, 1991.
[13] M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of NPCompleteness. Freeman, New York, NY, 1979.
23
[14] A. V. Goldberg and R. E. Tarjan. A new approach to the maximum-flow problem. Journal of
the ACM, 35(4):921–940, Oct. 1988.
[15] M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial Optimization, volume 2 of Algorithms and Combinatorics. Springer-Verlag, New York, NY, 1988.
[16] W. E. Hart. On the computational complexity of sequence design problems. In Proceedings of
the 1st Annual International Conference on Computational Molecular Biology, pages 128–136,
1997.
[17] M. Kimura. The Neutral Theory of Molecular Evolution. Cambridge University Press, Cambridge, United Kingdom, 1983.
[18] J. M. Kleinberg. Efficient algorithms for protein sequence design and the analysis of certain
evolutionary fitness landscapes. In Proceedings of the 3rd Annual International Conference on
Computational Molecular Biology, pages 226–237, 1999.
[19] K. F. Lau and K. A. Dill. A lattice statistical mechanics model of the conformational and
sequence spaces of proteins. Macromolecules, 22:3986–3997, 1989.
[20] K. F. Lau and K. A. Dill. Theory for protein mutability and biogenesis. Proceedings of the
National Academy of Sciences of the U.S.A., 87:638–642, 1990.
[21] E. L. Lawler. A procedure for computing the k best solutions to discrete optimization problems
and its application to the shortest path problem. Management Science, 18:401–405, 1972.
[22] D. Lipman and W. Wilbur. Modeling neutral and selective evolution of protein folding. Proceedings of Royal Society of London Series B, 245:7–11, 1991.
[23] K. M. Merz and S. M. L. Grand, editors. The Protein Folding Problem and Tertiary Structure
Prediction. Birkhauser, Boston, MA, 1994.
[24] C. Micheletti, F. Seno, A. Maritan, and J. Banavar. Design of proteins with hydrophobic and
polar amino acids. Proteins: Structure, Function, and Genetics, 32:80–87, 1998.
[25] C. H. Papadimitriou and K. Steiglitz. Combinatorial Optimization: Algorithms and Complexity.
Prentice-Hall, Upper Saddle River, NJ, 1982.
[26] J.-C. Picard and M. Queyranne. On the structure of all minimum cuts in a network and applications. Mathematical Programming Study, (13):8–16, 1980.
[27] J. Ponder and F. M. Richards. Tertiary templates for proteins. Journal of Molecular Biology,
193:63–89, 1987.
[28] J. S. Provan and M. O. Ball. The complexity of counting cuts and of computing the probability
that a graph is connected. SIAM Journal on Computing, 12(4):777–788, Nov. 1983.
[29] C. Reidys, P. Stadler, and P. Schuster. Generic properties of combinatory maps: Neutral networks of RNA secondary structures. Bulletin of Mathematical Biology, 59:339–397, 1997.
[30] E. I. Shakhnovich and A. M. Gutin. A new approach to the design of stable proteins. Protein
Engineering, 6:793–800, 1993.
24
[31] J. M. Smith. Natural selection and the concept of a protein space. Nature, 225:563–564, 1970.
[32] G. Steiner. An algorithm to generate the ideals of a partial order. Operations Research Letters,
5:317–320, 1986.
[33] S. J. Sun, R. Brem, H. S. Chan, and K. A. Dill. Designing amino acid sequences to fold with
good hydrophobic cores. Protein Engineering, 8(12):1205–1213, Dec. 1995.
[34] V. V. Vazirani and M. Yannakakis. Suboptimal cuts: Their enumeration, weight and number
(extended abstract). In W. Kuich, editor, Lecture Notes in Computer Science 623: Proceedings of
the 19th International Colloquium on Automata, Languages, and Programming, pages 366–377.
Springer-Verlag, New York, NY, 1992.
[35] K. Yue and K. A. Dill. Inverse protein folding problem: Designing polymer sequences. Proceedings
of the National Academy of Sciences of the U.S.A., 89:4163–4167, 1992.
25
| 5cs.CE
|
1
Control of active power for synchronization and
transient stability of power grids
arXiv:1802.06647v1 [math.OC] 19 Feb 2018
Randall Martyr, Benjamin Schäfer, Christian Beck, Vito Latora
Abstract—As the energy transition transforms power grids
across the globe it poses several challenges regarding grid design
and control. In particular, high levels of intermittent renewable
generation complicate the job of continuously balancing power
supply and demand, which is necessary for the grid’s stability.
Although there exist several proposals to control the grid, most
of them have not demonstrated to be cost efficient in terms of
optimal control theory. Here, we mathematically formulate the
control problem for stable operation of power grids, determining
the minimal amount of control in the active power needed to
achieve the constraints, and minimizing a suitable cost function
at the same time. We investigate the performance of the optimal
control method with respect to the uncontrolled scenario and
we compare it to a simple linear control case, for two types of
external disturbances. Considering case studies with two and five
nodes respectively, we find that the linear control can improve
the synchronization and transient stability of the power grid.
However, if the synchronized angular velocity after a disturbance
is allowed to differ from its initial steady state value, the linear
control performs inefficiently in comparison to the optimal one.
Index terms— Optimal control, Power system control,
Power system dynamics, Power system stability
I. I NTRODUCTION
The electrical power grid is undergoing drastic changes
due to the energy transition [1]–[3] and sophisticated control
approaches are necessary to ensure a reliable and stable
operation [4]. The generation side of the grid is changing
as additional renewable generators are installed to mitigate
climate change, introducing fluctuations in a time scale of
days [5] to sub-seconds [6]. In addition, the demand side is
changing due to the ongoing electrification of heating and
transport [7] and the introduction of demand control [8].
Regardless of these changing conditions, the grid needs to
stay within strict operational boundaries to guarantee a stable
electricity supply and to prevent damage to sensitive electronic
devices [4].
A fundamental aspect of power system stability is the ability
of interconnected synchronous machines of a power system to
remain synchronized [4]. Transient stability is the ability to
The first, second and fourth authors received financial support from the
UK Engineering and Physical Sciences Research Council (EPSRC) via grant
EP/N013492/1. The second author received support from the German Federal
Ministry of Education and Research (BMBF Grant No 03SF0472F).
Randall Martyr, Christian Beck and Vito Latora are with the School
of Mathematical Sciences, Queen Mary University of London, Mile End
Road, London E1 4NS, United Kingdom (e-mail: [email protected];
[email protected]; [email protected]).
Benjamin Schäfer is with the Chair for Network Dynamics, Center
for Advancing Electronics Dresden (cfaed), the Institute for Theoretical
Physics, Technical University of Dresden, Dresden and Network Dynamics,
Max Planck Institute for Dynamics and Self-Organization (MPIDS), 37077
Göttingen, Germany (e-mail: [email protected]).
maintain synchronism in the face of severe transient disturbances [4], and is of great importance in preventing cascading
failures [9]. Control mechanisms that balance active power
and regulate frequency in the grid are key to maintaining
these stability conditions. Primary controls respond within
a few seconds of an event to stabilize the frequency within
its permissible operating limits, after which secondary and
tertiary controls restore the frequency to its nominal value
[10]. Schäfer et al [11], [12] recently proposed a decentralized
linear local frequency controller that regulates demand and
supply on a power grid through economic incentives in order
to improve the grid’s transient stability. However, it remains
unclear how well and at what cost this linear control ensures
that the system’s operational constraints are satisfied. In fact,
some decentralized controllers for frequency response are
known to be inefficient [13].
In this paper, we formulate an optimal control problem [14],
[15] to keep the power grid within operational boundaries
when facing known changes in demand and generation. Note
that there are several predictable sources of perturbations
for which a designated control reaction may be designed.
For example, a recent study analysed power grid frequency
fluctuations in grids of different synchronous regions and
found a significant impact of trading on the frequency quality
with large fluctuations every fifteen minutes [16]. Furthermore,
there are predictable changes even in volatile renewable generation. For instance, in a system with large amounts of installed
solar capacity, the rising or setting sun will cause similar
power ramps as a recent solar eclipse caused in today’s system
[17]. The control problem we formulate is well-suited to keep
the grid within operational boundaries for such deterministic
perturbations.
In the following section we present a third order “fluxdecay” model which we use to describe the power grid dynamics in terms of a state vector x of phase angles, frequencies and
voltage amplitudes. Let U be a suitable set of control variables
u. For a given u ∈ U, we quantify its cost through a cost
function J(u), and evaluate its performance with respect to
various operational constraints Cη (u) and their tolerances εη .
The optimal control problem for power grid synchronization
is expressed mathematically as follows.
Problem:
minimize J(u) subject to:
i) ẋ(t) = f (t, x(t), u(t)),
ii) u ∈ U;
x(0) = x0 ;
iii) Cη (u) ≤ εη for η = 1, . . . , N + 2,
(1)
2
where function f governs the intrinsic dynamics of
the state of the grid, and N ≥ 2 is the number of
nodes in its representation as a network.
Problem (1) is solved numerically using a control parametrization method [18], [19]. We illustrate the efficiency of the
optimal control compared to the decentralized local frequency
control for the elementary two-node system and a five-node
network motif. In particular, we compute any violations of
the operational constraints and quantify the costs of control
actions. Finally, we close with a conclusion and outlook.
II. A N
OPTIMAL CONTROL PROBLEM FOR POWER SYSTEM
SYNCHRONIZATION
TABLE I
L IST OF NOTATION USED IN THE PAPER . V ECTORS AND MATRICES ARE
DENOTED IN BOLDFACE .
Notation
N
Ω
T
U
B
x
σ(ω)
hωi
θi
θij
ωi
ui
Vi
Pi
Viref
γi
αi , βi
J, Cη , εη
Description
the set of nodes {1, . . . , N } where N ≥ 2
is the number of nodes in the network
grid reference
control time horizon
set of control variables
N × N -dimensional coupling matrix
3N -dimensional state vector
variance of network angular velocities
mean value of network angular velocities
rotor angle at node i relative to the grid
reference
angular difference between nodes i and j,
θij = θi − θj
angular velocity at node i
controlled power added to node i
normalized machine voltage at node i
normalized mechanical power at node i
normalized voltage reference at node i
damping parameter at node i
machine parameters at node i
cost functional, constraint functional, constraint tolerance
Units
–
Hz
s
–
s−2
–
(rad · s−1 )2
rad · s−1
rad
set of differential equations for the rotor angle θi , angular
velocity ωi , and voltage Vi at each node,
(i) θ̇i (t) = ωi (t)
(ii) ω̇i (t) = −γi ωi (t) + Pi (t) + ui (t)
−
N
X
j=1
Bij Vi (t)Vj (t) sin θij (t)
1
(iii) V̇i (t) =
Viref − Vi (t)
αi
N
βi X
Bij Vj (t) cos θij (t) ,
+
αi j=1
(2)
for i = 1, 2, . . . , N.
Table I describes the notation used in this model. In
particular, θij (t) = θi (t) − θj (t), and the entries Bij of
matrix B describe the coupling topology of the network.
Positive values of Pi (t) indicate net generation at node i,
whilst negative values of Pi (t) indicate net consumption. If
Pi (t) is always positive (respectively, negative) then node i is
called a generator (respectively, consumer or motor) node. We
refer to positive changes in the value of the control variable
ui (t) as incremental actions [21] since they correspond to an
increase in generation or an equivalent decrease in demand.
Similarly, we refer to negative changes in the value of ui (t) as
decremental actions [21] since they correspond to a decrease
in generation or an equivalent increase in demand.
B. Decentralized control for improved power grid stability
rad
rad · s−1
s−2
1
s−2
1
s−1
s, s2
1
A. Dynamics for transient stability analysis
Before focusing on optimal control, we need to discuss
how to describe the intrinsic dynamics of the power grid.
According to [4, p. 27], deviations in rotor angles may also
lead to a progressive drop in the nodal voltages. This in turn
affects the angular velocities and rotor angle values. A realistic
model of the power grid should therefore take into account the
effect of the rotor angles’ deviations on the voltage amplitudes.
This allows us to analyse slower phenomena such as large
deviations in voltage or frequency, as typically done in midterm stability studies [4, p. 34]. Therefore, in this paper we use
the model in [20] which describes the power grid as a network
of N ≥ 2 synchronous generators and motors governed by a
Decentralized controllers for large-scale and complex networks such as the power grid are popular choices since
they rely only on local measurements and typically have
low computation and communication complexity [13]. Three
decentralized controllers that appear in the literature on power
grid synchronization and stability are mentioned below. They
all assume that at initial time t = 0 the power grid, as
described by the dynamics (2), is synchronized and at a steady
state with ω(0) = 0 in particular.
Linear local frequency controller: The basic modelling
assumption for the linear local frequency (LLF) controller in
[11], [12] is that a generator or consumer node acts in response
to its nodal price of electricity, which is a linear function of
the angular velocity at that node. Upon linearising the response
in power output at node i with respect to the nodal price,
the authors of [11], [12] derive a control ui that is directly
proportional to the angular velocity ωi ,
(3)
ui (t) := −νi (ωi (t) − ωi (0)) = −νi ωi (t) ,
where νi > 0 measures the willingness at node i to change the
active power level. This type of control effectively results in
the increase of the damping parameter from γi to γi +νi in (2).
Under the assumption of constant voltages, β ≡ 0, the authors
show that by achieving an equilibrium between demand and
supply through market price signals, their model can ensure
synchronization and stability of the power grid over time.
MARTYR et al.: CONTROL OF ACTIVE POWER FOR SYNCHRONIZATION AND TRANSIENT STABILITY OF POWER GRIDS
Integral local frequency controller: In [22], [23] the
following integral frequency control is proposed,
Z t
ui (t) := −Ki
ωi (τ )dτ,
0
where Ki > 0. Since ωi = θ̇i by (2) above, the integral
frequency control can be described equivalently by,
ui (t) = −Ki (θi (t) − θi (0)),
showing that it is directly proportional to the rotor angle’s
deviation from the steady state value θi (0). In the case
β ≡ 0, this control can improve transient stability and be
economically efficient in a particularly defined sense (see [22],
[23]). A similar statement for the general dynamics (2) with
time-varying voltages is not known [24].
Bandgap local frequency controller: In [25], the following bandgap control is proposed,
max
ωi −ωlow
min(1, ωhigh
ωi > ωlow ,
−ωlow ),
u i
−ωi −ωlow
max
ui (t) = −ui
min(1, ωhigh −ωlow ), ωi < −ωlow ,
0,
otherwise,
where 0 ≤ ωlow < ωhigh and umax
> 0 are parameters.
i
In this case, control is applied at node i ∈ N only if the
angular velocity deviation is sufficiently high, |ωi | > ωlow ,
and ramps up until a maximum value |umax
| which is reached
i
when |ωi | ≥ ωhigh . Under the assumption of constant voltages,
β ≡ 0, the authors show that the bandgap control can improve
grid stability.
C. Operational constraints of the power grid
Let x = (x1 , . . . , x3N ) denote the 3N -dimensional controlled state variable obtained from (2) with components given
by
xi = θi , xN +i = ωi , x2N +i = Vi for i ∈ N .
(4)
The dynamics of x in (2) can be written compactly as
ẋ(t) = f (t, x(t), u(t)),
(5)
where expressions for the components of f = (f1 , . . . , f3N )
are obtained from (2) using the assignment given in (4).
Each component of the control variable u = (u1 , . . . , un )
corresponds to the amount of additional active power injected
or withdrawn at an individual node in the network. We assume
that controls are bounded: for each i ∈ N we have ui (t) ∈ Ui
where:
Ui =
[umin
, umax
],
i
i
−∞ <
umin
i
<
umax
i
< ∞.
(6)
Let U denote the set of all such control functions.
Synchronization: Synchronization of the power grid
occurs when all of the rotor angles have the same velocity
[4, p. 19]. Consequently,
for every i ∈ N , we should have
PN
ω
.
ωi = hωi = N1
j=1 j Letting ω = (ω1 , . . . , ωN ) denote
the vector of angular velocities, the variance of ω, defined by
!
N
1 X
2
(7)
(ωi − hωi) ,
σ(ω) =
N i=1
3
measures the total dispersion of ω from its mean hωi, and
synchronization is achieved when
σ(ω) = 0.
(8)
Let 0 < T < ∞ denote the length of the control horizon [0, T ]
in seconds. Define the synchronization constraint loss function
by
ψ1 (x) = −σ(ω),
(9)
and the total synchronization loss on [0, T ] by
Z T
2
C1 (u) =
(min(0, ψ1 (x(t)))) dt
0
2
+ λ1 min 0, ψ1 (x(T )
Z T
=
σ(ω(t))2 dt + λ1 σ(ω(T ))2 ,
0
where λ1 ≥ 0 is a weight parameter which emphasizes the
relative importance of the constraint at the final time T . Note
that we weight the variance as given by eq. (7) quadratically in
this integral, which means small variances do not contribute
much whereas large variances are strongly penalized. Other
weighting schemes are also possible.
Mean angular velocity operational limits: The angular
velocity ωi quantifies the deviation of the frequency at node
i from the grid reference Ω. In the United Kingdom and
many other countries, the nominal frequency is 50 Hz, which
gives a reference value of Ω = 2π · 50 Hz. For reasons
related to the quality of electricity supply, the frequency must
respect certain operational limits. Therefore, values of the
mean angular velocity hωi should be constrained,
ωmin ≤ hωi ≤ ωmax .
(10)
In the United Kingdom, the statutory limits are ±0.5Hz of the
nominal value, but the operational limits are set to the stricter
range of ±0.2Hz from the nominal value [26]. In our setting,
the operational limits on the mean angular velocity become
ωmin = 2π · −0.2 Hz and ωmax = 2π · 0.2 Hz. Define the
mean angular velocity constraint loss function by,
ψ2 (x) = (ωmax − hωi)(hωi − ωmin ),
(11)
and the total loss on [0, T ] for violating this constraint by,
Z T
(min(0, ψ2 (x(t))))2 dt
C2 (u) =
0
2
+ λ2 min 0, ψ2 (x(T )) ,
where λ2 ≥ 0 is a weight parameter. Note that only when
ψ2 (x) is negative in eq. (11) we get a contribution.
Voltage operational limits: Since the voltages in our
model are also time dependent, it is important to also take into
account appropriate operational constraints on these variables.
Regulations in the United Kingdom require that the steady
state voltages should be kept within ±6% of the nominal
voltage for systems between 1 and 132 kV, or ±10% of the
nominal voltage for systems above 132 kV [27]. In our model
we can take this into account with the following constraint,
Vimin ≤ Vi ≤ Vimax , i = 1, . . . , N,
(12)
4
where Vimin < Vimax . We define a loss function for the
voltage constraint at each node i ∈ N by
ψ2+i (x) = (Vimax − Vi )(Vi − Vimin ),
(13)
and the total loss on [0, T ] for violating this constraint by
Z T
2
C2+i (u) =
(min(0, ψ2+i (x(t)))) dt
0
2
+ λ2+i min 0, ψ2+i (x(T )) ,
where λ2+i ≥ 0 are weight parameters.
D. Formulation of the optimal control problem
For η = 1, . . . , N + 2 the total loss Cη is non-negative, and
is equal to zero if, equivalently, the η-th constraint is satisfied
on [0, T ]. We relax this by introducing tolerance parameters
εη ≥ 0, η ∈ {1, . . . , N + 2}, and say that a control u ∈ U is
feasible if it satisfies
Cη (u) ≤ εη for η = 1, . . . , N + 2.
(14)
At an initial time t = 0, the power grid is synchronized
and at a steady state, ẋ = 0, in which various operational
constraints are satisfied. If an external disturbance ξ to the
mechanical power causes the grid to become unsynchronized,
we would like the control u to return it close to a synchronized
state within T seconds with a “minimal cost” that ensures the
constraint conditions (14) are satisfied. Letting I N denote the
N × N identity matrix, we define the following total cost for
a control u ∈ U,
Z T
J(u) =
u(t)tr I N u(t)dt,
(15)
0
where tr denotes the transpose operator. The rate function in
(15), utr I N u, is the square of the Euclidean norm of the
vector u,
N
X
(ui )2 .
utr I N u =
i=1
Therefore, higher cost is assigned to controls u ∈ U which
exert large amounts of effort over time. Moreover, adjustments
in demand and generation of the same magnitude are penalized
equally due to the symmetry utr I N u = (−u)tr I N (−u).
Note also that quadratic costs are standard in the frequency
control literature [23], [24]. Using the total cost (15) and the
previously defined constraint losses we formulate the optimal
control problem (1).
Existence of a minimising control: We study the case
in which the mechanical power P (t) is of the form,
Each constraint function Cη (u) in the optimal control
problem (1) represents one of the pure state constraints (8),
(10) and (12) in canonical form [18, p. 70]:
Z T
Cη (u) =
(min(0, ψη (x(t))))2 dt
(16)
0
2
+ λη min 0, ψη (x(T )) .
This formulation is convenient as a penalty function approach
for solving the optimal control problem numerically [18], [28].
There is a corresponding maximum principle which provides
necessary conditions for a solution to the optimal control
problem [14], but does not prove that there exists such a
solution. For completeness we provide the following result.
Theorem 1. Suppose the canonical form constraints in the
optimal control problem (1) are such that,
Cη (u) ≤ εη ∀η =⇒ (t, x(t)) ∈ Q ∀t ∈ [0, T ],
where Q is a compact set. Then there exists a state-control
pair (x∗ , u∗ ) solving Problem 1.
In physical terms, Theorem 1 tells us that there is a feasible
control strategy that minimizes the cost. Condition (17) is
a technical assumption needed to apply a general existence
theorem from optimal control theory (see [15, p. 483] for a
discussion on this). It is also a reasonable one since we aim to
control the power grid within a bounded range of operational
values. When ε = 0 all operational constraints are satisfied
at all times exactly, and (17) is easily verified from the pure
state constraints (8), (10) and (12).
Proof. First note that the dynamics for x given explicitly by
(2) are finitely generated in the following sense [15, p. 478]:
f (t, x(t), u(t)) = g 0 (t, x(t)) +
N
X
g k (t, x(t))uk (t), (18)
k=1
where g k , k = 0, 1, . . . , N , is a 3N -dimensional vectorvalued function. In particular, for k = 1, . . . , N we have
g k ≡ eN +k where eN +k is the 3N -dimensional vector with 1
at the (N + k)-th coordinate and zeros elsewhere. The optimal
control problem also has the following additional properties:
• Each function g k (k = 0, 1, . . . , N ) in (18) is measurable
in time t, continuous in the state x and satisfies,
(t, x) ∈ Q =⇒ kgk (t, x)k ≤ K(1 + kxk),
P (t) = P̄ + ξ(t),
where P̄ = (P̄1 , . . . , P̄N ) is an N -dimensional vector of
steady state values at the initial time 0, and ξ is a bounded
piecewise continuous N -dimensional deterministic function
that models the external disturbance at the nodes. Let uξ =
uξ1 , . . . , uξN denote the exact control, uξi (t) = −ξi (t) for all
i ∈ N . Since the system is assumed to be at a steady state at
time t = 0, the control uξ trivially ensures the inequalities (14)
are satisfied.
(17)
•
•
•
(19)
where K > 0 is a constant and k·k is the Euclidean norm.
For k = 1, . . . , N condition (19) is clearly satisfied since
g k ≡ eN +k . For k = 0 the growth condition (19) follows
from the dynamics (2), boundedness of the mechanical
power P , and compactness of Q
Q.
The set of control values U = N
i=1 Ui described by (6)
is compact and convex.
The mapping u 7→ utr I N u is continuous, non-negative
and convex.
The exact control, uξ , and its corresponding solution to
(5), xξ , satisfy conditions i) to iii) of (1). Furthermore,
the corresponding cost J uξ is finite since the disturbance ξ is bounded by assumption.
MARTYR et al.: CONTROL OF ACTIVE POWER FOR SYNCHRONIZATION AND TRANSIENT STABILITY OF POWER GRIDS
If there exists a compact set E such that the following stronger
version of (17) holds:
(
(t, x(t)) ∈ Q ∀t ∈ [0, T ],
Cη (u) ≤ εη ∀η ⇐⇒
(x(0), x(T )) ∈ E,
then the proof is completed by applying Theorem 23.11 in
[15]. Otherwise, we proceed by noticing that Cη (u) may be
written as,
Z T
Cη (u) =
Lη (t, x(t), u(t))dt,
(20)
0
where, in analogy to (18), Lη (t, x, u) is finitely generated,
measurable in t and continuous in x. If λη = 0 then (20)
clearly holds. If λη > 0 we use the fact that Ψη (x) =
2
λη min 0, ψη (x) is continuously differentiable then obtain
(20) by elementary calculus and using Ψη (x(0)) = 0.
Let x̃ be the (4N + 2)-dimensional vector with components
x̃j = xj for j = 1, . . . , 3N , and for η = 1, . . . , N + 2, x̃3N +η
has dynamics:
x̃˙ 3N +η (t) = Lη (t, x(t), u(t)),
Description of approximation method: The control
parametrization method [18], [19] is a numerical procedure
based on the maximum principle [14], [15], [28] which gives
an approximate solution to the optimal control problem. Let
np
denote a sequence of time points 0 = t0 ≤ . . . ≤
{tk }k=0
tnp = T used to partition the control horizon [0, T ] into
np > 0 subintervals, where np is an integer. Let Unp ⊂ U
denote the subset of controls u ∈ U of the form
ui (t) =
np
X
uki 1[tk−1 ,tk ) (t),
i = 1, . . . , N.
(21)
k=1
Each component ui , i ∈ N , of a control u ∈ Unp is
n
parametrized
by an np -dimensional vector (u1i , . . . , ui p ) ∈
Qnp
k=1 Ui . The optimal control problem is then approximated
by a constrained nonlinear optimization problem over the
bounded (N × np )-dimensional parameter space defining the
control vector u ∈ Unp . Standard optimization packages can
be used to solve this problem, and we used the Sequential
Least Squares Programming (SLSQP) routine in Python. For
the numerical simulations we use equidistant partitioning
points tk = k · nTp , 0 ≤ k ≤ np , with np = 2500.
x̃3N +η (0) = 0.
III. S IMULATIONS OF
Notice that x̃3N +η (T ) = Cη (u) and the dynamics for x̃ are
also finitely generated,
f̃ (t, x̃(t), u(t)) = g̃ 0 (t, x(t)) +
5
N
X
g̃ k (t, x(t))uk (t),
k=1
where g̃ k , k = 0, 1, . . . , N , is a (4N + 2)-dimensional vectorvalued function. We can now verify that there exist compact
sets Q̃ and Ẽ such that,
(
(t, x̃(t)) ∈ Q̃ ∀t ∈ [0, T ],
Cη (u) ≤ εη ∀η ⇐⇒
(x̃(0), x̃(T )) ∈ Ẽ,
and the proof is completed by applying Theorem 23.11 in
[15].
CONTROLLED NETWORKS
For the numerical simulations, we concentrate on the following two case studies. We model a two-node system and
the five-node network motif, shown in Figure 1a, consisting
of two generators at nodes 2 and 5 connected to three motors
at the remaining nodes. In Appendix A we list parameter
values for the control problem in Table III, and parameter
values for the two-node and five-node systems in Tables IV
and V respectively. We consider two types of disturbances
which only affect node 1: a persistent one and a temporary
one on the interval [0, 10]. For the two-node system, node
1 is perturbed by a disturbance of ξ1 = −0.7 s−2 . For the
five-node system a disturbance of ξ1 = −2.0 s−2 is used
as shown in Figure 1b. The simulated disturbances reflect a
sudden increase in demand, or equivalent loss in generation,
that can disappear after a few seconds (temporary case) or
persist for much longer (persistent case).
0.00
1
−0.25
3
ξ1 [s −2]
−0.50
5
−0.75
temporary
persistent
−1.00
−1.25
−1.50
−1.75
2
4
(a) Five-node network motif
−2.00
−5
0
5
10
15
Time [s]
20
25
30
(b) External power disturbances ξ1 applied at node 1
Fig. 1. The first figure (a) illustrates the five-node motif network with 2 generators (green squares) and 3 consumers (red circles). Parameters are given in
the Appendix. The second figure (b) illustrates the types of external power disturbances ξ1 applied at node 1. We consider a short temporary change (dashed)
of power and a persistent change (solid line). No disturbances are applied to the other nodes.
6
Additional active power applied to node 1
u1 [s −2]
u1 [s −2]
Additional active power applied to node 1
0.5
0.5
0.0
0.0
Additional active power applied to node 2
u2 [s −2]
u2 [s −2]
Additional active power applied to node 2
0.00
0.0
uncontrolled
linear
optimized
−0.2
−0.25
0
10
20
30
Time [s]
40
50
−0.4
60
0
10
20
(a) Temporary disturbance
30
Time [s]
40
50
60
(b) Persistent disturbance
Fig. 2. The controlled power applied at nodes 1 and 2 in the two-node system. Power profiles for the optimized and linear frequency controls are qualitatively
similar at node 1 but can be very different at node 2. For the optimized control, behaviour of the power profile at node 2 is almost opposite to that at node 1.
angular velocities is ω ∈ [−0.4π, 0.4π] s−1 , we see that the
range of implied values for the linear frequency control for
ν = 15 s−1 is already u ∈ [−19, 19] s−2 , which is very
large in comparison to the other active power parameters so
that consumers could be required to generate power instead of
consuming it.
Besides the exact and optimized controls, we consider a
linear frequency control which is defined by
∀i ∈ N : ui (t) = −ν ωi (t) , ν ≥ 0.
(22)
Note that in a synchronized system the same level of control
is applied to all nodes. Starting from ν = 0, which is the
uncontrolled case, we simulate the controlled system and
check whether the inequalities (14) are satisfied, incrementing
the value of ν by 1 until this becomes true. Table II below
shows the values of ν we obtained.
A. Two-node system
We show in Figures 2, 3 and 4 respectively, trajectories for
the controlled active power, average voltage, and the angular
velocity mean and standard deviation in the two-node system.
Uncontrolled system: The individual voltages and angular velocities both have periodic orbits that lie outside of the
operational constraints. For the voltages this is still reflected
when averaged across the nodes, as Figures 3a and 3b show.
Figures 4a and 4b show that the system loses its synchronism
and, even when the disturbance disappears, the unsynchronized
behaviour persists for the entire control horizon. For the
angular velocities, however, the oscillations largely cancel
out when averaged across the nodes so this behaviour is not
reflected in the top panels of Figures 4a and 4b.
Linear local frequency control: Trajectories for the
linear control applied to the temporary disturbance correspond
TABLE II
VALUES OF THE CONTROL PARAMETER ν IN s−1 WHICH SATISFIED THE
INEQUALITIES (14) WHEN USING THE LINEAR FREQUENCY CONTROL .
Temporary Disturbance
Persistent Disturbance
2 Nodes
15
280
5 Nodes
30
27
As discussed in Section II-A, the linear frequency control
effectively increases the damping parameter γ in the intrinsic
grid dynamics (2). Comparing the values of ν in Table II
with those for γ in Tables IV and V, we see that those for
ν are relatively large. Moreover, since the range of feasible
0.98
0.96
0.96
0.94
0.94
0.92
0.92
0.90
un⟩ontrolled
line⟨r
optimized
⟨ V⟩
⟨ V⟩
0.98
0
10
20
30
Time [s]
40
(a) Temporary disturbance
50
60
0.90
0
10
20
30
Time [s]
40
50
60
(b) Persistent disturbance
Fig. 3. Average voltage in the two-node system. Without a controlled response, the voltages exceed the lower operational limit demarcated by the dotted red
line. Contrarily, the control stabilizes the voltage.
MARTYR et al.: CONTROL OF ACTIVE POWER FOR SYNCHRONIZATION AND TRANSIENT STABILITY OF POWER GRIDS
Mean⟨angular⟨ elocity⟨in⟨rad/s
0
−1
√ σ⟩ω)
4
2
0.01
0
10
20
30
Time⟨[s]
40
50
Dispersion⟨of⟨ang lar⟨velocity⟨in⟨rad/s
0.02
4
√ σ⟩ω)
Dispersion⟨of⟨angular⟨ elocity⟨in⟨rad/s
0.02
0.00
Mean⟨ang lar⟨velocity⟨in⟨rad/s
1
0
−1
⟨ω ⟩
⟨ω ⟩
1
7
0.01
0
60
0.00
ncontrolled
linear
optimized
2
0
10
(a) Temporary disturbance
20
30
Time⟨[s]
40
0
60
50
(b) Persistent disturbance
p
Fig. 4. Angular velocity mean and deviation in the two-node system. Values of σ(ω) for the uncontrolled case are on the secondary y-axis. Red dotted
lines show operational limits. Without a controlled response, the system loses synchronism and the mean angular velocity hωi violates the operational limits.
This is particularly clear for the persistent disturbance.
to the coefficient value ν = 15 s−1 , whereas those for the
persistent one correspond to ν = 280 s−1 . Figure 2 shows
that the control initially adds power to node 1, which is nearly
equal in magnitude to the disturbance, and eventually the level
of additional power at this node decreases whilst the amount
added to the node 2 increases. For both disturbances, Figure 4
shows that the angular velocity hωi remains close to its initial
value. For the temporary disturbance, the system is already
close to being synchronized shortly after the disturbance ends,
whereas for the persistent one synchronization improves only
slowly over time.
Optimized control: The optimized control initially adds
about half the amount of power to node 1 as the linear one
whilst noticeably reducing the amount of power at node 2. This
results in a slower decrease in hωi as it approaches the lower
operational limit ωmin (cf. Figure 4). In the temporary case,
the system is returned close to its initial steady state values
after the disturbance has ended. In the persistent case, once hωi
reaches the lower operational limit ωmin it is kept there while
the disturbance persists. Figure 3 shows that the optimized
control takes a similar approach to ensuring the voltage stays
within its limits. Notice that in Figure 2 the optimized and
LLF controls are qualitatively similar and of the same sign at
node 1, but there is a sustained negative level of control at
node 2 during the time that there is a positive level for the
LLF control.
B. Five-node system
We show in Figures 5, 6 and 7 respectively, trajectories for
the controlled active power, average voltage, and the angular
velocity mean and standard deviation in the five-node system.
Uncontrolled system: Synchronization generally improves over time and the system eventually becomes resynchronized with acceptable voltages (Figures 6a and 6b) and
angular velocities (Figure 7a) except when the disturbance
persists (Figure 7b).
Linear local frequency control: The synchronization
results are similar to the two-node case, with angular velocities kept close to their initial values. The results show that
larger networks with higher levels of inertia can provide a
more distributed response, thereby requiring lower values of
the coefficient ν. Indeed, for persistent disturbance we used
ν = 27 s−1 , whereas for the two-node system we used
Additional active power applied to node 1
Additional active power applied to node 1
2
−2
1 [s ]
[s −2]
2
1
u
u
[s −2]
1
≠1
uj
0
−1
1
0
−1
j
j
0
Total additional active power applied to the other nodes
∑
≠1
uj
[s −2]
Total additional active power applied to the other nodes
∑
uncontrolled
linear
optimized
1
0
0
10
20
30
40
Time [s]
50
(a) Temporary disturbance
60
0
10
20
30
40
Time [s]
50
60
(b) Persistent disturbance
P
Fig. 5. The controlled power applied to the perturbed node 1, u1 , and the net total control applied to the remaining nodes, 5i=2 ui , in the five-node system.
Qualitatively, the linear frequency control and optimized one can be quite similar.
8
0.987
0.986
0.986
0.985
0.985
0.984
0.984
0.983
0.983
un⟩ontrolled
line⟨r
optimized
⟨V ⟩
⟨V ⟩
0.987
0.982
0.982
0.981
0.981
0.980
0.980
0.979
0.979
0
10
20
30
Time [s]
40
50
60
0
10
(a) Temporary disturbance
20
30
Time [s]
40
50
60
(b) Persistent disturbance
Fig. 6. Average voltage in the five-node system. Even without a controlled response the voltages stay well within the operational range [0.94, 1.06]
ν = 280 s−1 . Consequently, the combined response at the
unperturbed nodes leads to lower control costs.
Optimized control: The trajectories are quite similar to
those for the two-node system. In particular, as the optimized
control changes the net active power, the mean angular velocity
hωi follows its natural direction of descent or ascent within the
operational limits until a particular level. The angular velocity
is then kept at this level whilst the disturbance persists.
Furthermore, the combined action at the unperturbed nodes
is generally of the opposite type to that taken at the perturbed
node. That is, when there is an increase (respectively, decrease)
in P
u1 there is typically a decrease (respectively, increase)
5
in
i=2 ui at the same time. Somewhat surprising is the
extent to which the LLF and optimized controls agree, at least
qualitatively. Differences between the two mainly concern the
initial response to the disturbance:
• the LLF control causes a combined increase in active
power at the unperturbed nodes;
• the optimized control causes a combined decrease in
active power at the unperturbed nodes.
Another interesting observation is that the controlled power
and limiting angular velocities can all be different for the LLF,
uncontrolled and optimized cases, yet the limiting voltages
across these cases appear to be equal in Figure 6.
C. Comparison of control costs
In Figure 8 we show the cost J(u) for the linear, optimized
and exact controls associated with the trajectories displayed
above. While the optimized control satisfies the constraints
with smallest cost, we note that its costs are much lower than
the linear frequency control’s, partly because it is not forced
to keep the system at its initial steady state value. If solutions
that are sufficiently close to the initial steady state values are
more desirable, this can be taken into account in a variety of
ways, including:
•
•
J(u) =
10
20
30
Time⟨[s]
40
(a) Temporary disturbance
50
(23)
where ∆x(t) = x(t) − x0 and,
⟨ω ⟩
⟨ω ⟩
0
−2
0.0
60
0.050
0.2
0.025
0.000
ncontrolled
linear
optimized
Dispersion⟨of⟨ang lar⟨velocity⟨in⟨rad/s
√ σ⟩ω)
√ σ⟩ω)
0
qQ (u(t)) + qR (∆x(t))dt
Mean⟨ang lar⟨velocity⟨in⟨rad/s
0.2
0.025
T
+ qS (∆x(T )),
Dispersion⟨of⟨angular⟨ elocity⟨in⟨rad/s
0.050
Z
0
Mean⟨angular⟨ elocity⟨in⟨rad/s
1
0
−1
0.000
Adjusting the constraint costs Cη , η = 1, . . . , N + 2,
so that deviations from the initial state x0 are penalized
explicitly, perhaps increasingly as the remaining time for
control, T − t, decreases.
Penalising deviations from the initial state x0 by generalising (15) to
0
10
20
30
Time⟨[s]
40
50
0.0
60
(b) Persistent disturbance
p
Fig. 7. Angular velocity mean and deviation in the five-node system. Values of σ(ω) for the uncontrolled case are on the secondary y-axis. Red dotted
lines show operational limits.
MARTYR et al.: CONTROL OF ACTIVE POWER FOR SYNCHRONIZATION AND TRANSIENT STABILITY OF POWER GRIDS
30
25
250
exact
linear
optimized
200
( )
150
15
100
10
50
5
0
exact
linear
optimized
J u
J u
( )
20
9
temporary
persistent
(a) Two-node system
0
temporary
persistent
(b) Five-node system
Fig. 8. Comparison of control costs for the two-node (a) and the five-node (b) system using both temporary and persistent perturbations. The optimized
control has the lowest costs to keep the system within operational boundaries, but the difference becomes smaller in the larger network.
– for an M -dimensional square matrix A and an M dimensional vector y, qA (y) is the quadratic form
qA (y) = y tr Ay;
– Q, R and S are appropriately chosen diagonal
3N × 3N matrices with Q positive definite and R,
S positive semidefinite.
IV. C ONCLUSION
AND OUTLOOK
The linear local frequency control acts as a primary response
service to keep the grid frequency close to its nominal value.
In order to guarantee stable operation, its control coefficient
ν generally needs to be large. Using the analogy in [11],
[12], this means generators and consumers must have a high
willingness to adjust their output in response to changes in
the reference angular velocity. In terms of power output, this
leads to response profiles that almost exactly counteract the
disturbance, at least in the initial response phase.
Our results suggest that more efficient controllers distribute
the controlled response amongst all nodes in the power grid.
Moreover, this response need not be homogeneous throughout
the network, but could simultaneously involve incremental
actions (net increase in power) at some nodes and decremental
ones (net decrease in power) at others.
Trajectories associated with the optimized control show
that as it changes the net active power, the mean angular
velocity follows its natural direction of descent, or ascent
as appropriate, within the operational limits until a point is
reached, possibly at the boundary, at which the power grid is
synchronized and active power is balanced within the network.
For planned events, or events that will occur with very high
probability at an anticipated future time, an optimal centralized
control such as this one can be used as an initial response,
and can be coupled with a simple distributed or decentralized
control, such as the linear local frequency control, to smooth
out additional unknown perturbations.
Designing decentralized and distributed controllers that are
economically efficient is subject of ongoing work (see [13],
[23], [29], for instance). Decentralized stochastic control (see
[30], [31]), which generalizes our methodology by incorporating uncertainties and different information structures amongst
multiple controllers, is likely to become an important theoretical tool for understanding how these controllers work.
Its relevance to the stability and control of power systems
has been illustrated recently in [32]. While the numerical
results presented here to illustrate how the method works were
obtained for two simple case studies, namely a 2-node and a
5-node network, our method can be applied to more realistic
and larger networks, for which the problem discussed in our
paper is of utmost relevance: controlling the system not by
arbitrary means but in the most cost efficient way, such that
the operational constraints of the power grid keep on being
satisfied.
A PPENDIX
TABLE III
PARAMETER VALUES FOR THE SIMULATIONS . V ECTORS AND MATRICES
ARE DENOTED IN BOLDFACE .
Parameter
Value
Units
T
λ
ε
60
106
10−6
−4
4
0.94
1.06
1
−0.4π
0.4π
s
1
1
U min
U max
V min
V max
V ref
ωmin
ωmax
s−2
s−2
1
1
1
rad · s−1
rad · s−1
10
TABLE IV
PARAMETER VALUES FOR THE TWO - NODE SYSTEM BASED ON [33,
PP. 2–3]. V ECTORS AND MATRICES ARE DENOTED IN BOLDFACE .
Parameter
γ
α
β
B
P̄
θ̄
ω̄·
V̄·
Value
0.2
2
0.1
!
−0.8
1
1
−0.8
(−0.7, 0.7)
(0, 0.7953)
0
0.9901
[13]
Units
s−1
s
s2
[14]
[15]
s−2
[16]
s−2
rad
rad · s−1
1
[17]
[18]
TABLE V
PARAMETER VALUES FOR THE FIVE - NODE SYSTEM . V ECTORS AND
MATRICES ARE DENOTED IN BOLDFACE .
Parameter
Value
Units
γ
α
β
0.2
2
0.02
−5
1.63 1.63
0
1.63
1.63
−5
1.63 1.63
0
1.63 1.63
−5
1.63
0
1.63 1.63
−5
1.63
0
1.63
0
0
1.63
−5
(−1.0, 1.5, −1.0, −1.0, 1.5)
(−0.1689, 0.1087, −0.2872, −0.1689, 0.3353)
0
(0.9916, 0.9928, 0.9946, 0.9916, 0.9605)
s−1
s
s2
B
P̄
θ̄
ω̄
V̄
[19]
[20]
[21]
s−2
[22]
s−2
rad
rad · s−1
1
[23]
[24]
R EFERENCES
[25]
[1] G. Boyle, Renewable Energy: Power for a Sustainable Future, 2nd ed.
Oxford, UK: Oxford University Press, 2004.
[2] F. Ueckerdt, R. Brecha, and G. Luderer, “Analyzing major challenges
of wind and solar variability in power systems,” Renew. Energ., vol. 81,
pp. 1–10, Sept. 2015. DOI: 10.1016/j.renene.2015.03.002.
[3] J. A. Turner, “A realizable renewable energy future,” Science, vol. 285,
no. 5428, pp. 687–689, Jul. 1999. DOI: 10.1126/science.285.5428.687.
[4] P. Kundur, Power System Stability and Control, N. J. Balu and M. G.
Lauby, eds. New York, NY, USA: McGraw-Hill, 1994.
[5] D. Heide, L. Von Bremen, M. Greiner, C. Hoffmann, M. Speckmann,
and S. Bofinger, “Seasonal optimal mix of wind and solar power in a
future, highly renewable Europe,” Renew. Energ., vol. 35, no. 11, pp.
2483–2489, Nov. 2010. DOI: 10.1016/j.renene.2010.03.012.
[6] P. Milan, M. Wächter, and J. Peinke, “Turbulent character of wind
energy,” Phys. Rev. Lett., vol. 110, no. 13, Mar. 2013, art. no. 138701.
DOI: 10.1103/PhysRevLett.110.138701.
[7] K. Dennis, K. Colburn, and J. Lazar, “Environmentally beneficial electrification: The dawn of emissions efficiency,” The Electricity Journal,
vol. 29, no. 6, pp. 52–58, Jul. 2016. DOI: 10.1016/j.tej.2016.07.007.
[8] P. Palensky and D. Dietrich, “Demand Side Management: demand
response, intelligent energy systems, and smart Loads,” IEEE T. Ind. Inform., vol. 7, no. 3, pp. 381–388, Aug. 2011. 10.1109/TII.2011.2158841.
[9] B. Schäfer, D. Witthaut, M. Timme and V. Latora, “Dynamically
induced cascading failures in supply networks,” arXiv preprint, Jul.
2017. arXiv:1707.08018.
[10] D. Kirschen and G. Strbac, Fundamentals of Power System Economics.
Chicester, UK: John Wiley & Sons, 2004.
[11] B. Schäfer, M. Matthiae, M. Timme and D. Witthaut, “Decentral smart
grid control,” New J. Phys., vol. 17, art. no. 015002, Jan. 2015.
[12] B. Schäfer, C. Grabow, S. Auer, J. Kurths, D. Witthaut and M. Timme,
“Taming instabilities in power grid networks by decentralized control,”
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
Eur. Phys. J-Spec. Top., vol. 225, no. 3, pp. 569–582, May 2016. DOI:
10.1140/epjst/e2015-50136-y.
D. K. Molzahn, F. Dörfler, H. Sandberg, S. H. Low, S. Chakrabarti, R.
Baldick and J. Lavaei, “A survey of distributed optimization and control
algorithms for electric power systems,” IEEE T. Smart Grid, vol. 8, no.
6, pp. 2941–2962, Nov. 2017. DOI:10.1109/TSG.2017.2720471.
M. R. Hestenes, Calculus of Variations and Optimal Control Theory.
New York, NY, USA: John Wiley & Sons, 1966.
F. Clarke, Functional Analysis, Calculus of Variations and Optimal Control (Graduate Texts in Mathematics Vol. 264). London, UK: Springer,
2013.
B. Schäfer, C. Beck, K. Aihara, D. Witthaut and M. Timme, “NonGaussian power grid frequency fluctuations characterized by Lévy-stable
laws and superstatistics,” Nature Energy, to be published, 2018. DOI:
10.1038/s41560-017-0058-z.
R.G. Harrison and E. Hanna, “The solar eclipse: a natural meteorological
experiment,” Phil. Trans. R. Soc. A., vol. 374, no. 2077, Aug. 2016. DOI:
10.1098/rsta.2015.0225.
K. L. Teo and C.J. Goh, A Unified Computational Approach to Optimal
Control Problems. Essex, UK: Longman Scientific & Technical, 1991.
V. Rehbockt, K. L. Teo, L. S. Jennings and H. W. J. Lee, “A survey of the
control parametrization and control parametrization enhancing methods
for constrained optimal control problems,” in Progress in Optimization
(Applied Optimization Vol. 30). Boston, MA, USA: Springer US, 1999,
pp. 247–275. DOI:10.1007/978-1-4613-3285-5 13.
K. Schmietendorf, J. Peinke, R. Friedrich and O. Kamps, “Self-organized
synchronization and voltage stability in networks of synchronous machines,” Eur. Phys. J-Spec. Top., vol. 223, no. 12, pp. 2577–2592, Oct.
2014. DOI: 10.1140/epjst/e2014-02209-8.
D. Z. Szabó and R. Martyr, “Real option valuation of a decremental
regulation service provided by electricity storage,” Phil. Trans. R. Soc.
A., vol. 375, no. 2100, Aug. 2017. DOI: 10.1098/rsta.2016.0300.
C. Zhao, E. Mallada, and F. Dörfler, “Distributed frequency control
for stability and economic dispatch in power networks,” in Proc.
Am. Control Conf., Chicago, IL, USA, 2015, pp. 2359–2364. DOI:
10.1109/ACC.2015.7171085.
E. Weitenberg, Y. Jiang, C. Zhao, E. Mallada, C. De Persis and
F. Dörfler, “Robust decentralized secondary frequency control in
power systems: merits and trade-offs,” arXiv preprint, Nov. 2017.
arXiv:1711.07332.
S. Trip, M. Bürger and C. De Persis, “An internal model approach to (optimal) frequency regulation in power grids with timevarying voltages,” Automatica, vol. 64, pp. 240–253, Feb. 2016. DOI:
10.1016/j.automatica.2015.11.021.
S. Auer, C. Roos, J. Heitzig, F. Hellmann and J. Kurths, “The contribution of different electric vehicle control strategies to dynamical grid
stability,” arXiv preprint, Aug. 2017. arXiv:1708.03531.
Electricity
Transmission
System
Operations.
National
Grid
UK,
2017.
[Online].
Available:
http://www2.nationalgrid.com/uk/industry-information/electricity-transmission-system-o
Accessed on: July, 2017.
The Electricity Supply Regulations 1988. The National Archives, 2017.
http://www.legislation.gov.uk/uksi/1988/1057/made/data.xht?wrap=true.
Accessed on: July, 2017.
R. F. Hartl, S. P. Sethi and R. G. Vickson, “A survey of the maximum
principles for optimal control problems with state constraints,” SIAM
Rev., vol. 37, no. 2, pp. 181–218, Jun. 1995. DOi: 10.1137/1037043.
T. Stegink, C. De Persis and A. van der Schaft, “A unifying energybased approach to stability of power grids with market dynamics,” IEEE
T. Automat. Contr., vol. 62, no. 6, pp. 2612–2622, Jun. 2017. DOI:
10.1109/TAC.2016.2613901.
A. Mahajan and M. Mannan, “Decentralized stochastic control,”
Ann. Oper. Res., vol. 241, no. 1–2, pp. 109–126, Jun. 2016. DOI:
10.1007/s10479-014-1652-0.
C. D. Charalambous and N. U. Ahmed, “Centralized versus decentralized optimization of distributed stochastic differential decision systems with different information structures-part I: a general theory,”
T. Automat. Contr., vol. 62, no. 3, pp. 1194–1209, Mar. 2017. DOI:
10.1109/TAC.2016.2575818.
R. Singh, P. R. Kumar and L. Xie, “Decentralized control via dynamic
stochastic prices: the independent system operator problem,” arXiv
preprint, Jun. 2016. arXiv:1605.08926v2.
K. Schmietendorf, J. Peinke and O. Kamps, “On the stability and quality
of power grids subjected to intermittent feed-in,” arXiv preprint, Nov.
2016. arXiv:1611.08235.
| 3cs.SY
|
1
Stochastic Stability Analysis of Discrete Time
System Using Lyapunov Measure
arXiv:1503.04438v2 [math.DS] 8 Mar 2016
Umesh Vaidya, Senior Member, IEEE,
Abstract—In this paper, we study the stability problem of a
stochastic, nonlinear, discrete-time system. We introduce a linear
transfer operator-based Lyapunov measure as a new tool for
stability verification of stochastic systems. Weaker set-theoretic
notion of almost everywhere stochastic stability is introduced
and verified, using Lyapunov measure-based stochastic stability
theorems. Furthermore, connection between Lyapunov functions,
a popular tool for stochastic stability verification, and Lyapunov
measures is established. Using the duality property between the
linear transfer Perron-Frobenius and Koopman operators, we
show the Lyapunov measure and Lyapunov function used for
the verification of stochastic stability are dual to each other.
Set-oriented numerical methods are proposed for the finite
dimensional approximation of the Perron-Frobenius operator;
hence, Lyapunov measure is proposed. Stability results in finite
dimensional approximation space are also presented. Finite
dimensional approximation is shown to introduce further weaker
notion of stability referred to as coarse stochastic stability. The
results in this paper extend our earlier work on the use of
Lyapunov measures for almost everywhere stability verification of
deterministic dynamical systems (“Lyapunov Measure for Almost
Everywhere Stability”, IEEE Trans. on Automatic Control, Vol. 53,
No. 1, Feb. 2008).
Index Terms—Stochastic stability, almost everywhere, computational methods.
I. I NTRODUCTION
Stability analysis and control of stochastic systems are a
problem of theoretical and applied interests. For stochastic
systems, there are various notions of stabilities. Among the
most popular notions of stabilities are almost sure and moment
stability [1], [2]. Almost sure notion of stability implies
stability of sample-path trajectories of stochastic systems and
is a weaker notion of stability compared to the moment
stability definition. Moment stability definition deals with the
steady state probability density function of a stochastic system,
in particular, the integrability of it with respect to various
powers of state. There is extensive literature on stochastic
stability and stabilization in control and dynamical system
literature. Some of the classic results on stability analysis and
control of stochastic dynamical system are found in [1]–[3].
Mao [4] presents a systematic summary of results on various
stochastic stability definitions and Lyapunov function-based
verification techniques for a stochastic system. The topic of
stochastic stability of switched system and the Markov jump
system has also attracted lots of attention with applications in
network controlled dynamical systems [5]–[11]. The results in
Financial support from the National Science Foundation grant ECCS1150405 and CNS-1329915 is gratefully acknowledged. U. Vaidya is with the
Department of Electrical and Computer Engineering, Iowa State University,
Ames, IA, 50011 USA e-mail: [email protected].
these papers address not only the stochastic stability problem,
but also robust control synthesis and fundamental limitations
issues that arise in stabilization and estimation of dynamical
systems in the presence of stochasticity in a feedback loop.
More generally, stochastic stability and stabilization problems
for a nonlinear system with a multiplicative noise process are
addressed in [12]–[14].
As in a deterministic system, most of the existing methods
for stochastic stability verification are based directly or indirectly on the Lyapunov function. The Lyapunov function is
used for both almost sure and moment stability verification.
The Lyapunov function and Lyapunov function-based methods
are also applied for stochastic stabilization [15]–[20]. While
there is extensive literature on the numerical procedure for the
construction of a Lyapunov function for stability verification
and stabilization of a deterministic dynamical system [21]–
[30], the literature on numerical methods for stochastic stability verification is very scant [31], [32]. The contributions of
this paper are two-fold. We provide a novel operator theoretical
framework for stability verification of a stochastic dynamical
system. We also show the proposed operator theoretic framework is amicable to computations thereby providing systematic
numerical procedures for stochastic stability verification.
The operator theoretic framework that we introduce in this
paper is used to verify a weaker set-theoretic notion of almost
everywhere stability of a stochastic system. The notion of
almost everywhere stability was introduced for the first time
in the work of Rantzer for continuous time deterministic
dynamical systems [33]. This notion of stability was later
extended to continuous time stochastic systems in [34]. The
notion of a.e. stability essentially implies the set of points in
the state space starting from which system trajectories are not
attracted to the attractor set is a measure zero set. The almost
everywhere notion of stability was also used for the design
of stabilizing feedback controller and for solving verification
problems for a nonlinear system [22]. The input-output version
of almost everywhere stability was developed in [35]. In [36],
linear transfer operator-based framework was introduced to
verify this weaker set-theoretic notion of a.e. stability for
discrete-time dynamical system. Transfer Perron-Frobenius
operator-based Lyapunov measure was introduced as a new
tool to verify a.e stability of a nonlinear system. Duality
between the Lyapunov function and Lyapunov measure was
established using the duality in the transfer Perron-Frobenius
and Koopman operators. Application of the Lyapunov measure
for the design of stabilizing and optimal feedback controller
was proposed in [37], [38]. Operator theoretical framework
involving spectral analysis of Koopman operator was proposed
2
in [39] for stability analysis of deterministic nonlinear systems.
The results in this paper can be viewed as a natural extension
of the results from [36] towards a.e. stability verification of
discrete time stochastic dynamical system.
Associated with the stochastic dynamical system are two
linear transfer operators, the Perron-Frobenius (P-F) and
Koopman operators. These operators are used to study the
evolution of ensembles of points in the form of measures or
density supported on state space. These operators are dual to
each other and are used in the dynamical system literature
for the analysis of deterministic and stochastic dynamical
systems [40]–[47]. One of the main contributions of this
paper is the introduction of a linear transfer P-F operator
based Lyapunov measure for a.e. stochastic stability verification of stochastic dynamical system. We introduce a.e.
almost sure notion of stability and provide Lyapunov measurebased stability theorems to verify this stability. The Lyapunov
function for stochastic stability verification is shown intimately
connected with Koopman operator formalism. In particular,
analytical formulas for the computation of Lyapunov measure
and the Lyapunov function are obtained in terms of resolvent
of the Perron-Frobenius and Koopman operator respectively.
By exploiting the duality relationship between Koopman and
Perron-Frobenius operators, we show the Lyapunov function
and Lyapunov measure are dual to each other in a stochastic
setting as well. The results presented in this paper are extended
version of results appeared in [48], [49].
While results exist for the application of operator theoretical
methods for the stability verification of a stochastic system in
more general Markov chain settings [50], the main motivation
of this work is to provide computational methods for the
construction of stability certificate in the form of a Lyapunov
measure. Towards this goal, set-oriented methods are used for
the finite dimensional approximation of the linear transfer P-F
operator and for computation of the Lyapunov measure [51],
[52]. The finite dimension approximation introduces a further
weaker notion of a.e. stability by allowing stable dynamics
in the complement of the attractor set, but their domain of
attraction is smaller than the size of discretization cells used in
the finite dimensional approximation. This notion of stability is
referred to as coarse stochastic stability. The finite dimensional
approximation of the P-F operator arises as a Markov matrix
and provides for various alternate formulas for the computation
of the Lyapunov measure.
Organization of this paper is as follows. In section II,
we discuss the preliminaries of the transfer operators and
introduce various set-theoretic stochastic stability definitions.
The main results of this paper on Lyapunov measure-based,
stochastic stability theorems are proven in section III. The
formula for obtaining the Lyapunov function for a stochastic
system in terms of the resolvent of the Koopman operator is
presented in section IV. The connection between the Lyapunov
measure and the Lyapunov function is discussed in section V.
The set-oriented numerical method for the finite dimensional
approximation of P-F operator and stability results using the
finite dimensional approximation are discussed in VI and VII
respectively. Simulation examples and discussed in Section
VIII followed by conclusions in section IX.
II. P RELIMINARIES
The set-up and preliminaries for this section are adopted
from [40]. Consider the discrete-time stochastic dynamical
system,
xn+1 = T (xn , ξn ),
(1)
where xn ∈ X ⊂ Rd is a compact set. We denote B(X) as
the σ-algebra of Borel subsets of X. The random vectors,
ξ0 , ξ1 , . . ., are assumed independent identically distributed
(i.i.d) and takes values in W with the following probability
distribution,
Prob(ξn ∈ B) = v(B), ∀n, B ⊂ W,
and is the same for all n and v is the probability measure.
The system mapping T (x, ξ) is assumed continuous in x and
for every fixed x ∈ X, it is measurable in ξ. The initial
condition, x0 , and the sequence of random vectors, ξ0 , ξ1 , . . .,
are assumed independent. Let W × W = W 2 . Then, the two
times composition of the stochastic dynamical system, denoted
by T 2 : X × W 2 → X, is given by
xn+2 = T (T (xn , ξn ), ξn+1 ) =: T 2 (xn , ξnn+1 ),
where ξnn+1 ∈ W 2 . Since the sequence of random vectors {ξn }
is assumed i.i.d, the probability measure on W 2 will simply
be the product measure, v × v := v 2 . Similarly, the n-times
composition of a stochastic system (1), T n : X × W n → X,
is denoted by xn+1 = T n (x0 , ξ0n ), where ξ0n ∈ W n with
probability measure v n .
The basic object of study in our proposed approach to
stochastic stability is a linear transfer, the Perron-Frobenius
operator, defined as follows:
Definition 1 (Perron-Frobenius (P-F) operator): Let
M(X) be the space of finite measures on X. The PerronFrobenius operator, P : M(X) → M(X), corresponding to
the stochastic dynamical system (1) is given by
R R
[PT µ](A) = X W χA (T (x, y))dv(y)dµ(x)
h
i
= Eξ µ(Tξ−1 (A)) ,
(2)
for µ ∈ M(X), and A ∈ B(X), where χA (x) is an indicator
function of set A.
The expectations on ξ are taken with respect to the probability
measure, v, and Tξ−1 (A) is the inverse image of set A under
the mapping T (x, ξ) for a fixed value of ξ, i.e.,
Tξ−1 (A) = {x : T (x, ξ) ∈ A}.
Furthermore, if we denote the P-F operator corresponding to
the dynamical system, Tξ : X → X, for a fixed value of ξ as
Z
[PTξ µ](A) =
χA (T (x, ξ))dµ(x) = µ(Tξ−1 (A)),
X
then, the P-F operator (2) can be written as
[PT µ](A) = Eξ [PTξ µ](A).
The Koopman operator is dual to the P-F operator and defined
as follows:
Definition 2 (Koopman operator): Let h ∈ C 0 (X) be the
space of continuous function. The Koopman operator, UT :
3
C 0 (X) → C 0 (X), corresponding to the stochastic system (1)
is defined as follows:
Z
[UT h](x) =
h(T (x, y))dv(y) = Eξ [h(T (x, ξ))],
(3)
W
where the expectations are taken with respect to probability
measure, v.
Let h be a measurable function
and µ ∈ M(X), define the
R
inner product as hh, µi = X hdµ(x). With respect to this
inner product, the Koopman and P-F operators are dual to
each other. Using the inner product definition, we can write
(2) as follows:
hχA , PT µi = hUT χA , µi .
ϑ denoted as µ ≺ ϑ, if µ(B) = 0 for all B ∈ B(X) with
ϑ(B) = 0.
Definition 8 (Equivalent measure): The two measures, µ
and ϑ, are equivalent (µ ≈ ϑ) provided µ(B) = 0, if and
only if ϑ(B) = 0 for B ∈ B(X).
A. Decomposition of the P-F operator
Let E = {0}. Hence, E c = X \ E. We write T : E ∪ E c ×
W → X. For any set B ∈ B(E c ), we can write
R R
[PT µ](B) = X W χB (T (x, y))dv(y)dµ(x)
R R
= E c W χB (T (x, y))dv(y)dµ(x).
(4)
Due to the linearity of the scalarPproduct, this implies
n
hgn , PT µi = hUT gn , µi, where gn = i=1 αi χAi is the sum
of a simple function. Since, every measurable function, h, can
be approximated by a sequence {gn } of simple functions, we
obtain in the limit, hh, PT µi = hUT h, µi .
Assumption 3: We assume x = 0 is an equilibrium point of
system (1), i.e., T (0, ξn ) = 0, ∀n, for any given sequence
of random vectors {ξn } taking values in set W .
Assumption 4 (Local Stability): We assume the trivial solution, x = 0, is locally stochastic, asymptotically stable. In
particular, we assume there exists a neighborhood O of x = 0,
such that for all x0 ∈ O,
This is because T (x, ξ) ∈ B implies x ∈
/ E. Since set E is
invariant, we define the restriction of the P-F operator on the
complement set E c . Thus, we can define the restriction of the
P-F operator on the measure space M(E c ) as follows:
Z Z
χB (T (x, y))dv(y)dµ(x),
[P1 µ](B) =
Prob{T n (x0 , ξ0n ) ∈ O} = 1, ∀n ≥ 0,
where µ ∈ M(E) and B ⊂ B(E).
The above considerations suggest a representation of the PF operator, P, in terms of P0 and P1 . Indeed, this is the case,
if one considers a splitting of the measured space,
and
Prob{ lim T n (x0 , ξ0n ) = 0} = 1.
n→∞
Ec
c
W
for any set B ∈ B(E ) and µ ∈ M(E c ).
Next, the restriction T : E × W → E can also be used to
define a P-F operator denoted by
Z
[P0 µ](B) =
χB (T (x, y))dv(y)dµ(x),
B
Assumption 3 is used in the decomposition of the P-F operator
M(X) = M0 ⊕ M1 ,
(5)
in section (II-A) and Assumption 4 is used in the proof of
Theorem (10). In the following, we will use the notation U () where M0 := M(E), M1 := M(E c ), and ⊕ stands for
to denote the neighborhood of the origin for any positive direct sum.
Then it follows the splitting defined by Eq. (5), the P-F
value of > 0. We have 0 ∈ U () ⊂ O.
We introduce the following definitions for stability of operator has a lower-triangular matrix representation given by
stochastic dynamical systems (1).
P0 0
P=
.
(6)
Definition 5 (Almost everywhere (a.e.) almost sure stability):
× P1
The equilibrium point, x = 0, is said to be almost everywhere,
almost sure stable with respect to finite measure, m ∈ M(X),
III. LYAPUNOV MEASURE AND STOCHASTIC STABILITY
if for every δ() > 0, we have
We begin the main results section with the following
Lemma.
m{x ∈ X : P rob{ lim T n (x, ξ0n ) 6= 0} ≥ δ} = 0.
n→∞
Lemma 9: Let
Definition 6 (a.e. stochastic stable with geometric decay):
ξ0n = {ξ0 , . . . , ξn } ∈ W × . . . × W =: W n
For any given > 0, let U () be the neighborhood of the
{z
}
|
n
equilibrium point, x = 0. The equilibrium point, x = 0,
is said to be almost everywhere, almost sure stable with and F (x, ξ n ) := T n (x, ξ n ) : X × W n → X be the notation
0
0
geometric decay with respect to finite measure, m ∈ M(X), for the n times composition of the map T : X × W → X.
if there exists 0 < α() < 1, 0 < β < 1, and K() < ∞, Then, the Perron-Frobenius operator, PF : M(X) → M(X),
such that
corresponding to system mapping F is given by
m{x ∈ X : P rob{T n (x, ξ0n ) ∈ B} ≥ αn } ≤ Kβ n ,
for all sets B ∈ B(X \ U ()), such that m(B) > 0.
We introduce the following definition of absolutely continuous and equivalent measures.
Definition 7 (Absolutely continuous measure): A measure
µ is absolutely continuous with respect to another measure,
PF = PT PT . . . PT =: PnT .
|
{z
}
n
Proof: Let ϑ(A) = [PT µ](A), it then follows from
Definition 1 for the P-F operator that
h
i
[PT µ](A) = ϑ(A) = Eξ0 µ(Tξ−1
(A))
,
(7)
0
4
where expectation is taken with respect to probability measure
v and Tξ−1
(A) is the inverse image of set A under the mapping
0
T (x, ξ0 ) for fixed values of ξ0 , i.e.,
Tξ−1
(A) = {x ∈ X : T (x, ξ0 ) ∈ A}.
0
Hence,
[P2T µ](B) = [PT PT µ](B) = [PT ϑ](A) = Eξ1 [ϑ(Tξ−1
(B))].
1
Tξ−1
(B)
1
Following (7) and defining
= A, we obtain
h
h
ii
−1
[P2T µ](B) = Eξ1 Eξ0 µ(Tξ−1
(T
(B)))
.
ξ
0
1
Since ξ0 and ξ1 are independent with the same probability
distribution v, we obtain
h
i
−1
[P2T µ](B) = Eξ01 µ(Tξ−1
(T
(B)))
ξ
0
1
R R
= W 2 X χB (T (T (x, y0 ), y1 ))dv(y0 )dv(y1 )dµ(x)
= [PT 2 (x,ξ01 ) µ](B).
(8)
local neighborhood of x = 0. However, limn→∞ [Pn1 m](B) =
0 for any set B ∈ B(X \ U ()) and, in particular, for B = S̃.
Hence, we have m(S̃) = m(S) = 0.
Theorem 11: The x = 0 solution for system (1) is a.e.
stochastic stable with geometric decay with respect to finite
measure, m ∈ M(X), if and only if there exists a positive
constant, K(), and 0 < β < 1, such that
[Pn1 m](B) ≤ Kβ n , ∀n ≥ 0
for every set B ∈ B(X \ U ()), such that m(B) > 0.
Proof: We first prove the sufficient part.
Z Z
[Pn1 m](B) =
χB (T n (x, ξ0n ))dv0n (ξ0n )dm(x) ≤ Kβ n .
X
R R
[Pn1 m](B) = X W n χB (T n (x, ξ0n ))dv0n (ξ0n )dm(x)
R
= X P rob{T n (x, ξ0n ) ∈ B}dm(x).
(11)
Choose 0 < α < 1, such that
We now state the first main results of the paper on the
stochastic stability expressed in terms of asymptotic behavior
of Pn1 .
Theorem 10: The equilibrium point, x = 0, for system (1)
is almost everywhere, almost sure stable with respect to finite
measure, m ∈ M(X), if
lim [Pn1 m](B) = 0,
n→∞
for every set B ∈ B(X \ U ()), such that m(B) > 0. U () is
the neighborhood of the equilibrium point, x = 0, for any
given > 0.
Proof: For any given δ() > 0, consider the following
sets,
β
α
< 1. Let
Sn = {x ∈ X : P rob{T n (x, ξ0n ) ∈ B} ≤ αn }
The main results of this lemma then follow from induction.
Using the lower triangular structure of the P-F operator in
Eq. (6), one can write the PnT as follows:
n
P0 0
PnT =
.
(9)
× Pn1
Wn
and
S̄n = {x ∈ X : P rob{T n (x, ξ0n ) ∈ B} ≥ αn }.
Then,
[Pn1 m](B)
Z
=
ZSn
+
P rob{T n (x, ξ0n ) ∈ B}dm(x)
P rob{T n (x, ξ0n ) ∈ B}dm(x).
S̄n
Since integrals over both sets are positive, we have
Z
[Pn1 m](B) ≥
P rob{T n (x, ξ0n ) ∈ B}dm(x) ≥ αn m(S̄n ).
S̄n
Hence, we have,
m(S̄n ) ≤ K
βn
,
αn
which is equivalent to
m{x ∈ X : P rob{T n (x, ξ0n ) ∈ B} ≥ αn } ≤ K β̄ n .
For the necessary part, we assume the system is a.e. stochastic
Sn = {x ∈ X : P rob(T k (x, ξ0k ) ∈ X\U ()) ≥ δ, for some k > n}
stable with geometric decay (Definition 6). Construct the
and
sets, Sn and S̄n , from the sufficiency part of the proof.
S = ∩∞
n=1 Sn .
The parameter, α, for this construction now comes from the
So, set S consists of points with probability larger than δ to stability definition (Definition 6). We have
Z
end up in set X \ U (). Now, since is arbitrary small and
n
[P
m](B)
=
P rob{T n (x, ξ0n ) ∈ B}dm(x)
1
from the local stability property of equilibrium point, we have
ZSn
0 ∈ U () ⊂ O. The proof of this theorem follows, if we show
+
P rob{T n (x, ξ0n ) ∈ B}dm(x).
m(S) = 0. Let S̃ = S∩(X\O). Then, from the property of the
S̄n
local neighborhood, m(S) = m(S̃). From the construction of
set S, it follows x ∈ S, if and only if P rob(T (x, ξ) ∈ S) = 1. Hence,
Hence,
[Pn1 m](B) ≤ αn m(Sn ) + m(S̄n ) ≤ αn m(Sn ) + Kβ n .
Z
[PT m](S) =
P rob(χS (T (x, ξ)))dm(x) = m(S). (10) Now, since X is assumed compact, we have m(Sn ) < M for
X
some M < ∞. Hence, we have
Now, S̃ ⊂ S with m(S̃) = m(S). Since T is continuous in x
[Pn1 m](B) ≤ K̄ β̄ n ,
and measurable in ξ, we have [P1 m](S̃) = [P1 m](S). Using
(10), we obtain [P1 m](S̃) = m(S̃). S̃ ⊂ X lies outside some where K̄ = max(M, K) and β̄ = max(β, α) < 1.
5
Following is the definition of Lyapunov measure introduced
for stability verification of a stochastic system.
Definition 12 (Lyapunov measure): A Lyapunov measure,
µ̄ ∈ M(X \ U ()), is defined as any positive measure finite
outside the neighborhood of equilibrium point and satisfies
[P1 µ̄](B) < γ µ̄(B)
(12)
for 0 < γ ≤ 1 and for all sets B ∈ B(X \ U ()).
Theorem 13: Consider the stochastic dynamical system (1)
with x = 0 as a locally stable, equilibrium point. Assume
there exists a Lyapunov measure, µ̄, satisfying Eq. (12) with
γ < 1. Then,
1) x = 0 is almost everywhere almost sure stable with
respect to finite measure, m, which is absolutely continuous with respect to the Lyapunov measure µ̄.
2) x = 0 is almost everywhere stochastic stable with
geometric decay with respect to measure any finite
measure, m ≺ κµ̄, for some constant, κ > 0.
Proof: 1) Using the definition of Lyapunov measure with
γ < 1, we obtain
[PnT µ̄](B) ≤ γ n µ̄(B) =⇒ lim [PnT µ̄](B) = 0.
n→∞
Now, since m ≺ µ̄, we have limn→∞ [PnT m](B) = 0. The
proof then follows by applying the results from Theorem 10.
2) We have
[PnT m](B) ≤ κ[PnT µ̄](B) ≤ κγ n µ̄(B) ≤ K()γ n ,
where K() = κµ̄(X \ U ()), which, by definition of the Lyapunov measure, is finite. The proof then follows by applying
the results from Theorem 11.
The following theorem provides for the construction of
the Lyapunov measure as an infinite series involving the PF operator.
Theorem 14: Let the equilibrium point, x = 0, be almost
everywhere, stochastic stable with geometric decay with respect to measure, m. Then, there exists a Lyapunov measure
(Definition 12) µ̄ with γ < 1. Furthermore, the Lyapunov
measure is equivalent to measure m (i.e., µ̄ ≈ m) and the Lyapunov measure dominates measure m (i.e., µ̄(B) ≥ m(B)).
Proof: Following the results of Theorem 11, we know
there exists a positive constant, K, and β < 1, such that
[Pn1 m](B) ≤ Kβ n .
Let β = β1 β2 , such that β1 < 1 and β2 < 1. Hence, we have
αn [Pn1 m](B) ≤ Kβ1n
with α =
follows:
1
β2
> 1. Now, construct the Lyapunov measure as
µ̄(B) = m + α[P1 m] + α2 + . . . (B)
P∞
= k=0 αk [Pk1 m](B).
IV. KOOPMAN OPERATOR AND LYAPUNOV FUNCTION
In this section, we describe the connection between the
Koopman operator and Lyapunov function for a stochastic
dynamical system.
Definition 15 (pth Moment Stability): The equilibrium solution x = 0 is said pth moment exponentially stable with
p ∈ Z+ , if there exist a positive constants, K < ∞, and
β < 1, such that
Eξ0n [k xn+1 kp ] ≤ Kβ n k x0 kp , ∀n ≥ 0
for all initial conditions x0 ∈ X.
The results of the following theorem are not new [1], [4].
However, the proposed construction of the Lyapunov function
in terms of the Koopman operator is new. Furthermore, the
connection between the Koopman operator and Lyapunov
function also brings out clearly the dual nature of the Lyapunov function and the Lyapunov measure for stochastic
stability verification.
Theorem 16: The equilibrium solution, x = 0, is pth
moment exponentially stable, if and only if there exists a
nonnegative function V : X → R+ satisfying
a k x kp ≤ V (x) ≤ b k x kp ,
(13)
[UT V ](x) ≤ cV (x),
where a, b, c are positive constants with c < 1. Furthermore,
V can be expressed in terms of resolvent of the Koopman
operator as follows:
V (x) = (I − UT )−1 f (x),
where f (x) =k x kp .
Proof: We first prove the sufficient part. We have
Eξ0n [k xn+1 kp ] ≤ a1 Eξ0n [V (xn+1 )] = a1 Eξn−1 [[UT V ](xn )]
0
≤ ac Eξn−1 [V (xn )] ≤
0
cn
a V
(x0 ) ≤ ab cn k x0 kp .
(14)
Let, K = ab and c = β. Thus, we obtain the desired condition
for the pth moment exponential stability. For the necessary
part, let
PN
PN
VN (x0 ) = k=0 Eξ0k [f (xk+1 )] = k=0 Eξk−1 [[UT f ](xk )]
0
PN
k
= k=0 [UT f ](x0 ).
(15)
The uniform bound on VN (x) follows from pth moment
exponential stability. Hence, V (x) = limN →∞ VN (x) is well
defined. Furthermore,
V (x) = lim
The above infinite series is well-defined and converges because
αn [Pn1 m](B) ≤ Kβ1n . Multiplying both sides of (13) with
I − αP1 , we obtain
α[P1 µ̄](B) ≤ µ̄(B) =⇒ [P1 µ̄](B) ≤ β2 µ̄(B).
Allowing γ = β2 , we satisfy the requirements for the Lyapunov measure. The equivalence of measure, m, the Lyapunov
measure, µ̄, and the dominance of the Lyapunov measure
follows from the construction of the Lyapunov measure as
a infinite series formula Eq. (13).
N →∞
N
X
[UkT f ](x0 ) = (I − UT )−1 f.
(16)
k=0
The bounds on the function, V , and the inequality,
[UT V ](x) ≤ cV (x), follow from the construction of the
function, V.
The P-F operator and Koopman operator are shown dual to
each other. In Eqs. (13) and (16), the Lyapunov measure and
6
the Lyapunov function are expressed in terms of an infinite
series involving the P-F and Koopman operators, respectively.
Using the duality relationship between the P-F and Koopman
operators, it then follows the Lyapunov function and the
Lyapunov measure used for verifying stochastic stability are
dual to each other. In the following section, we establish
a precise connection between the two stochastic, stability
verification tools.
Hence, (following Theorem 14), there exists a Lyapunov
measure satisfying
dµ̄(x) − αd[P1 µ](x) = g(x)dm(x).
Furthermore, the density function corresponding to ρ(x) satisfies
a1 k x k−p ≤ ρ(x) ≤ a2 k x k−p .
V. R ELATION BETWEEN LYAPUNOV MEASURE AND
FUNCTION
In this section, we relate Lyapunov measure and Lyapunov
function. We impose additional assumptions, the system mapping (1), i.e., T : X × W → X. We assume the system
mapping T is C 1 invertible diffeomorphism with respect to
state variable, x, for any fixed value of noise parameter, ξ.
For the diffeomorphism, we define
Jξ−1 (x) =
dTξ−1
dx
(x) ,
where | · | stands for the determinant.
Lemma 17: Let PT be the P-F operator for the system
mapping T : X × W → X. Then,
Then, the x = 0 solution is pth exponentially stable with
Lyapunov function V obtained as V (x) = ρ(x)−1 .
2) Let x = 0 be the pth moment exponentially stable with the Lyapunov function V (x) satisfying V (x) <
βEξ [Jξ (x)V (Tξ−1 (x))] for some β < 1. Then, the measure,
Z
1
dm(x),
(18)
µ̄(B) =
γ
B V (x)
is a Lyapunov measure satisfying Eξ [µ̄(Tξ−1 (B))] < κµ̄(B)
for some κ < 1 and for all B ∈ M(X \U ()) with m(B) > 0.
γ ≥ 1 is a suitable constant chosen, such that V1γ is integrable.
Proof: 1)
dµ̄(x) − αd[P1 µ](x) = g(x)dm(x)
d[PT m](x) = Eξ [Jξ−1 (x)]dm(x).
If ρ(x) is the density of an absolutely continuous measure,
µ, with respect to the Lebesgue measure, m, i.e., dµ(x) =
ρ(x)dm(x), then
d[PT µ](x) = Eξ [Jξ−1 (x)ρ(Tξ−1 (x))]dm(x).
Proof:
Z
Z
[PT m](B) =
χB (Ty (x))dm(x)dv(y)
W
Z
Z
χB (x)dm(Ty−1 (x))dv(y).
=
W
Z
Z
X
Jy−1 (x)dm(x)dv(y) =
χB (x)
X
X
W
Z
X
dµ̄(x) = ρ(x)dm(x).
Z Z
[PT µ̄](B) =
χB (Ty (x))dv(y)ρ(x)dm(x)
X
Z Z
=
X
= V −1 (x)dm(x) − αEξ [Jξ−1 (x)V −1 (Tξ−1 (x))]dm(x).(19)
Hence, we have V −1 (x) − αEξ [Jξ−1 (x)V −1 (Tξ−1 (x))] =
g(x) ≥ 0. Now, since Jξ (x) < 1 or Jξ−1 (x) > 1 for all
values of ξ ∈ W , we have
#
"
1
1
.
(20)
≥ αEξ
V (x)
V (Tξ−1 (x))
Now, using Holder’s inequality, we have
"
#
"
#
V (Tξ−1 (x))
1
1 = Eξ
≤ Eξ
Eξ [V (Tξ−1 (x))].
V (Tξ−1 (x))
V (Tξ−1 (x))
χB (x)E[Jξ−1 (x)]dm(x).
d[PT m] = Eξ [Jξ−1 (x)]dm(x).
W
χB (x)dv(y)ρ(Ty−1 (x))Jy−1 (x)dm(x).
W
d[PT µ̄] = E[ρ(Tξ−1 (x))Jξ−1 (x)]dm(x).
We have the following theorem connecting the Lyapunov
measure and the Lyapunov function for a stochastic system.
Theorem 18: Let Jξ (x) < ∆ < 1 for all ξ ∈ W and x ∈ X.
1) Let the equilibrium point, x = 0, be almost everywhere,
stochastic stable with geometric decay with respect to the
Lebesgue measure, m. Assume the Lyapunov measure is
absolutely continuous with respect to the Lebesgue measure,
m, with density function ρ(x), i.e., dµ̄(x) = ρ(x)dm(x).
(17)
"
Eξ
1
V
(Tξ−1 (x))
#
≥
1
.
Eξ [V (Tξ−1 (x))]
(21)
Combining (20) and (21), we obtain
V (x) ≤ ᾱEξ [V (Tξ−1 (x))], ᾱ =
1
< 1.
α
(22)
Since V (x) = ρ−1 (x) is assumed radially unbounded (Eq.
(17)), then using the results from Theorem (16) and inequality
(22), we achieve the desired results, the x = 0 solution is the
pth moment exponentially stable.
2) We have V (x)γ < β γ Eξ [Jξ (x)V (Tξ−1 (x))]γ . Using the
Holder inequality, we obtain
V (x)γ < β̄Eξ [Jξγ (x)V γ (Tξ−1 (x))],
where β̄ = β γ < 1.
1
β̄
>
.
γ
γ
V (x)
Eξ [Jξ (x)V γ (Tξ−1 (x))]
7
Now, since Jξ < ∆ < 1 is uniformly bounded, there exists P , is now obtained as
a γ > 1, sufficiently large, such that
νj = [Pµ](Dj )
#
"
Q
L Z
1
β̄
1
X
X
dm(x)
>
.
≥
E
ξ
γ
=
p`
δTw` (x) (Dj )µi
V (x)γ
Eξ [Jξ (x)V γ (Tξ−1 (x))]
Jξ (x)V γ (Tξ−1 (x))
m(D
i)
D
Integrating over set B ∈ B(X \ U ()), we obtain
=
Z
β̄
V
−γ
Z
(x)dm(x) ≥
B
B
`=1
Q
X
`=1
Jξ−1 (x)V −γ (Tξ−1 (x))dm(x).
Using the results from Lemma (17) and (18), we obtain the
desired results by letting κ = β̄, i.e.,
[P1 µ̄](B) ≤ κµ̄(B).
p`
i=1
L
X
i=1
where
Pijw` =
i
µi Pijw` =
L
X
i=1
µi
Q
X
p` Pijw` ,
`=1
m(Tw−1
(Dj ) ∩ Di )
`
,
m(Di )
(23)
m the Lebesgue measure. The resulting matrix is non-negative
and because Tw` : Di → X for ` = 1, . . . , Q,
L
X
Pijw` = 1, ` = 1, . . . , Q,
j=1
VI. D ISCRETIZATION OF THE P-F OPERATOR
In this section, we will discuss the set-oriented numerical
methods proposed for the finite dimensional approximation of
the P-F operator. For the finite dimensional approximation of
the P-F operator, we will make the following assumption on
the stochastic process.
Assumption 19: We assume the random vector ξn takes
finitely many vector values w1 , . . . , wQ with the following
probabilities
Prob{ξn = w` } = p` ∀n,
` = 1, . . . , Q.
For the finite dimension approximation of the P-F operator,
we consider the finite partition of the state space, X, as
follows:
X = {D1 , . . . , DL },
i.e., P is a Markov or a row-stochastic matrix.
Computationally, several short-term trajectories are used
to compute the individual entries Pijw` . The mapping Tw`
is used to transport M “initial conditions” chosen to be
uniformly distributed within a set Di . The entry Pijw` is then
approximated by the fraction of initial conditions in box, Dj ,
after one iterate of the mapping. In the remainder of the paper,
the notation of this section is used, whereby P represents
the finite-dimensional Markov matrix corresponding to the
infinite dimensional P-F operator P. The finite dimensional
approximation of the P-F operator can be used to study the
approximated dynamics of the stochastic dynamical system. In
particular, the eigenvector with eigenvalue one of the matrix
P , i.e.,
L
X
µi = 1,
µP = µ, µi ≥ 0,
i=1
where ∪j Dj = X. The infinite dimensional measure, µ, is
approximated by ascribing a real number µj to each cell of
the partition, X . In this way, the infinite dimensional measure,
µ, is approximated by a finite dimensional vector, µ ∈ RL .
The finite dimensional approximation of the P-F operator
will arise as a Markov matrix of size L × L on the finite
dimensional vector space. Using Assumption 19, the stochastic
dynamical system, T (x, ξ) : X × W → X, is parameterized
by finitely many values that the random variable takes. In
particular, for each fixed value of a random variable, i.e.,
ξn = wk , we have T (x, wk ) for k = 1, . . . , Q. For notational
convenience, we will also write the mapping for each fixed
value of random variable as T (x, wk ) =: Twk (x). Hence,
we have Twk : X → X for k = 1, . . . , Q. With the finite
parameterizations in the noise space, the finite dimensional
approximation of the P-F operator follows exactly along the
lines of a deterministic setting. In particular, corresponding to
a vector, µ = (µ1 , · · · , µL ) ∈ RL , define a measure on X as
dµ(x) =
L
X
i=1
µi κi (x)
dm(x)
,
m(Di )
where m is the Lebesgue measure and κj denotes the indicator
function with support on Dj . The approximation, denoted by
captures the steady state dynamics of the system.
VII. S TABILITY IN FINITE DIMENSION
The finite dimensional approximation of the P-F operator
presented in the previous section can be decomposed into
the lower triangular form similar to the decomposition of the
infinite dimensional P-F operator in Eq. (6). With no loss of
generality, we assume the equilibrium point at the origin is
contained inside the cell, D1 , and the partition is sufficiently
fine so D1 ⊂ O, where O is the local neighborhood of
the stable equilibrium point. We decompose X into two
complimentary partition as follows,
X0 = {D1 }, X1 = {D2 , ..., DL },
∪L
k=2 Dk .
(24)
with domains X0 = D1 and X1 =
Since the local
stable equilibrium point at the origin is contained in the cell,
D1 , there exists a left eigenvector with eigenvalue one to
the Markov matrix, P , of the form µ0 = (1, 0 . . . , 0), i.e.,
µ0 = µ0 P . The existence of the eigenvector can be used
to decompose the Markov matrix, P , in the following upper
triangular structure. Let M ∼
= RL , M0 ∼
= R, and M1 ∼
= RL−1
denote the finite dimensional measure space associated with
the partition X , X0 , and X1 , respectively and as defined in
8
(24). Then, for the splitting M = M0 ⊕ M1 , the P matrix has
the following lower triangular representation.
P0 0
P =
,
× P1
where P0 = 1 and maps P0 : M0 → M0 and P1 : M1 → M1
is the sub-Markov matrix with row sum less than or equal to
one. We refer the readers to [36] for details on the decomposition in finite dimension. Our goal is to study stability with
respect to initial condition starting from partition X1 using the
sub-Markov matrix, P1 . In a discrete finite dimensional setting,
stability is expressed in terms of the transient property of the
stochastic matrix P1 .
Definition 20 (Transient states): A sub-Markov matrix P1
has only transient states, if P1n → 0, element-wise, as n → ∞.
Transience of P1 is shown to imply a weaker notion of stability
referred to as coarse stochastic stability and defined as follows.
where m is the Lebesgue measure. We need to show P1 < 1.
Denote,
Sk = {x ∈ DL : Twk (x) ∈ DL }, k = 1, . . . , Q.
Clearly, Sk ⊂ DL and the existence of Lyapunov measure µ̄
satisfying Eq. (25) implies
k
µ̄(Sk ) = [Pw
1 µ̄](DL ) < µ̄(DL ), k = 1, . . . , Q.
Now,
Q
X
pk µ̄(Sk ) =
k=1
Q
X
[P1 µ̄](B) < µ̄(B)
(25)
for all B ⊂ B(X1 ), and additionally, µ̄ ≡ m, the
Lebesgue measure. Then, the finite-dimensional approximation, P1 , is transient.
2) Suppose P1 is transient, then the equilibrium point, x =
0, is coarse stable with respect to the initial conditions
in X1 .
Proof: Before stating the proof, we claim for any two
sets, S1 and S, such that S1 ⊂ S, if µ ≈ m, then
µ(S1 ) = µ(S)
⇐⇒
m(S1 ) = m(S).
(26)
S1c
:= S \S1 as the complement set. We have, µ(S1 ) =
Denote
µ(S) implies µ(S1c ) = 0 which, in turn, implies m(S1c ) = 0
and, thus, m(S1 ) = m(S).
1. We first present a proof for the simplest case where the
partition, X1 , consists of precisely one cell, i.e., X1 = {DL }.
In this case, P1 ∈ [0, 1] is a scalar given by
P1 =
Q
X
k=1
pk
m(Tw−1
(DL ) ∩ DL )
k
,
m(DL )
(27)
k
pk [Pw
1 µ̄](DL )] < µ̄(DL ).
k=1
Using (26), m(Sk ) 6= m(DL ) and since Sk ⊂ DL , we have
m(Sk ) < m(DL ) for k = 1, . . . , Q. Hence,
P1 =
Definition 21 (Stochastic Coarse Stability): Consider the
finite partition, X1 , of the complement set, X1 = X \ X0 .
The equilibrium point, x = 0, is said to be stochastic coarse
stable with respect to the initial condition in X1 , if for an
attractor set, B ⊂ U ⊂ X1 there exists no subpartition
S = {Ds1 , . . . , Dsl } in X1 with domain S = ∪lk=1 Dsk such
that B ⊂ S ⊂ U and for all x ∈ S P rob{T (x, ξ) ∈ S} = 1 .
For typical partitions, coarse stability means stability modulo attractor sets, B, with domain of attraction, U , smaller
than the size of cells within the partition. In the infinitedimensional limit, where the cell size (measure) goes to zero,
one obtains stability modulo attractor sets with measure 0
domain of attraction.
Theorem 22: Let the equilibrium point, x = 0 ∈ X0 ⊂ X,
with local domain of attraction contained inside X0 , P1 is the
sub-Markov operator on M(E c ). P1 be its finite dimensional
approximation obtained with respect to the partition X1 of the
complement set X1 = X \X0 . For this, we have the following
1) Suppose a Lyapunov measure, µ̄, exists such that
(28)
Q
X
pk
k=1
m(Sk )
<1
m(DL )
is transient.
We prove the result for the general case, where X1 is a finite
partition, by contradiction. Suppose P1 is not transient. Then,
using the general result from the theory of finite Markov chains
[53], there exists at least one non-negative invariant probability
vector, ν, such that
ν · P1 = ν.
(29)
Let,
S = {x ∈ Di : νi > 0},
S` = {x ∈ S : Tw` (x) ∈ S},
for ` = 1, . . . , Q. It is claimed
m(S` ) = m(S), ` = 1, . . . , Q.
(30)
We first assume the claim to be true and show the desired
contradiction. Clearly, S` ⊂ S and if the claim were true, (26)
shows
µ̄(S` ) = µ̄(S), ` = 1, . . . , Q.
(31)
Next, because S ⊂ X1 ,
PQ
P1 µ̄(S)
= `=1 p` µ̄(Tw−1
(S) ∩ X1 )
`
PQ
PQ
−1
≥ `=1 p` µ̄(Tw` (S) ∩ S) = `=1 p` µ̄(S` ).
This, together with Eq. (31), gives
P1 µ̄(S) ≥ µ̄(S)
for a set S with positive Lebesgue measure. This contradicts
Eq. (25) and proves the theorem.
It remains to show the claim. Let {ik }lk=1 be the indices
with νik > 0. Equation . (29) gives
l
X
k=1
νik
Q
X
p` [P1w` ]ik jr = νjr for r = 1, . . . , l.
`=1
Taking a summation,
l
X
k=1
νik
Pl
r=1 ,
on either side gives
Q
l X
X
r=1 `=1
p` [P1w` ]ik jr = 1.
9
Since individual entries are non-negative and ν is a probability
vector, this implies
Q
l X
X
p` [P1w` ]ik jr = 1 k = 1, . . . , l.
Using formula (23) for the individual matrix entries, this gives
p` m(Tw−1
(Djr ) ∩ Dik ) = m(Dik ),
`
r=1 `=1
Q
X
p`
`=1
l
X
m(Tw−1
(Djr ) ∩ Dik ) = m(Dik ).
`
r=1
Since Djr are disjoint for any fixed w` ,
∪r {x : Tw` (x) ∈ Djr } = {x : Tw` (x) ∈ ∪r Djr },
i.e., ∪Tw−1
(Djr ) = Tw−1
(∪Djr ). Furthermore, since Tw` is
`
`
(Djr1 ) ∩ Dik ) ∩ (Tw−1
one-to-one, we have (Tw−1
(Djr2 ) ∩
`
`
Dik ) = ∅ as
{x ∈ Dik : Tw` (x) ∈ Djr1 }∩{x ∈ Dik : Tw` (x) ∈ Djr2 } = ∅
PQ
−1
`
for r1 6= r2 . Hence,
`=1 p` m(Tw` (∪r=1 Djr ) ∩ Dik ) =
m(Dik ). Therefore,
Q
X
Q
X
p`
`=1
r=1 `=1
Q
l X
X
whenever Dj ∈
/ S. Since, Tw` : S → S,
µ̄ = m · (I − P1 )−1 = m + m · P12 + m · P12 + . . . ,
for row vector m > 0 (element wise). For example m
could be taken as row vector of all ones corresponding
to the finite dimensional approximation of Lebesgue
measure.
2) Compute the Lyapunov measure as a solution of the
linear program
µ̄(αI − P1 ) > 0, α < 1.
`=1
VIII. E XAMPLES AND S IMULATION
However, by construction, S = ∪lr=1 Djr and, thus,
p` m(Tw−1
(S)
`
Example 23: We consider an inverted pendulum example
with a stochastic damping parameter, ξ.
∩ Dik ) = m(Dik ).
`=1
Taking a summation on both sides
Q
X
p` m(S` ) = m(S) =
`=1
Pl
Q
X
k=1 ,
we obtain
p` m(S).
(32)
`=1
Since S` ⊂ S, we have m(S` ) ≤ m(S). From (32), we obtain
Q
X
p` (m(S` ) − m(S)) = 0.
`=1
Since p` > 0 and m(S` ) ≤ m(S), we conclude m(S` ) =
m(S).
2. Suppose P1 is transient. To show the equilibrium point
is coarse stable, we proceed by contradiction. Using definition 21, if the equilibrium point is not coarse stable, then
there exists an attractor set B ⊂ U ⊂ X1 with a sub-partition
S = {Ds1 , ..., Dsl }, S = ∪lk=1 Dsk such that B ⊂ S ⊂ U and
P rob{T (x, ξ) ∈ S} = 1 for all x ∈ S.
P rob{T (x, ξ) ∈ S} = 1 =
Q
X
pk χS (Tw` (x))
`=1
for x ∈ S. This implies S is left invariant by each Tw` for
` = 1, . . . , Q. Hence,
Pswk`j =
m(Tw−1
(Dj ) ∩ Dsk )
`
= 0,
m(Dsk )
i = 1, ..., l,
j=1
i.e., P1 is a Markov matrix with respect to the finite partition,
S. From the general theory of the Markov matrix [53], there
exists an invariant probability vector, ν, such that ν · P1n = ν
for all n > 0 and P1 is not transient.
There are various different equivalent ways of computing
the finite dimensional approximation of Lyapunov measure
using the finite dimensional approximation of the P-F operator.
In particular, if P1 is transient, then the Lyapunov measure µ̄
can be computed using the following formulas.
1) Compute the Lyapunov measure using an infinite series
formula,
p` m(Tw−1
(∪`r=1 Djr ) ∩ Dik ) = m(Dik ).
`
Q
X
l
X
[P1w` ]si sj = 1
x˙1
= x2 .
x˙2
= − sin x1 − (ξ + 0.7)x2 .
(33)
For the purpose of simulation, the continuous system is
discretized with the time step of discretization, ∆t = 0.1.
The phase space, X, for simulation is taken to be equal to
X = [−π, π] × [−π, π]. For the purpose of compacting the
state space, we identify −π and π along the x and y axis.
The damping parameter,, ξ is assumed stochastic with zero
mean and uniform distribution; hence, of the form [−α, α]. In
the following, we present simulation results by changing the
variance, i.e., α, of the random variable, ξ, while keeping the
mean zero and verifying the stochastic stability of the origin.
In Figs. 1 and 2, we show the plot for the Lyapunov measure
with random variable, ξ, supported on interval [−0.5, 0.5]
and [−0.75, 0.75], respectively. The existence of Lyapunov
measure implies the origin is almost everywhere almost sure
stable. When support of the random parameter, ξ, is increased
to [−1, 1], the system fails to have Lyapunov measure and
is confirmed from the invariant measure plot shown in Fig.
3. From Fig. 3, we notice the invariant measure has support
around the origin. This is in contrast to the invariant measure
plots when ξ is supported on [−0.5, 0.5] and [−0.75, 0.75] and
shown in Figs. 1 and 2. The invariant measure in these cases
is supported at the origin and is denoted by a blue dot at the
origin in Figs. 1 and 2.
10
Fig. 1.
Lyapunov measure plot with support of ξ ∈ [−0.5, 0.5].
Fig. 2.
Lyapunov measure plot with support of ξ ∈ [−0.75, 0.75].
distributed between [−α, α]. The deterministic
√ system has four
equilibrium points (0, 0), (2, 0) and (3, ± 3) of which the
origin is stable equilibrium and the remaining are unstable. For
the stochastic system, two of the equilibrium points - one at
the origin and equilibrium point at (2, 0) are preserved. For the
purpose of simulation, the system is discretized in time with
the time step for discretization chosen to be equal to ∆t = 0.1.
The state space X is taken to be X = [−4, 4] × [−4, 4]. For
the purpose of compactness, −4 and 4 are identified along the
x−axis.
For better visualization, we plot the logarithm of the Lyapunov measure plot for varying variance of the random variable, ξ, by changing α, while retaining its zero mean. In Figs.
4 and 5, we show the log plot for the Lyapunov measure with ξ
supported between [−0.25, 0.25] and [−0.5, 0.5], respectively.
The existence of a Lyapunov measure in these plots implies
the origin is almost everywhere almost sure stable for these
statistics of random variable. When the support of the random
variable is changed to [−0.1, 0.1], the origin is no longer
almost everywhere almost sure stable. In Fig. 6, we show the
plot for the invariant measure supported in the neighborhood
of the origin. This is in contrast to the invariant measure plot
for the cases when ξ ∈ [0.25, 0.25] and ξ ∈ [0.5, 0.5]. For both
cases, the invariant measure is supported only at the single cell
containing the origin.
Fig. 4. Log plot of Lyapunov measure with noise support ξ ∈ [−0.25, 0.25].
Fig. 3.
Invariant measure plot with support of ξ ∈ [−1, 1].
Example 24: The second example is the stochastic counterpart of the deterministic almost everywhere stable system
example from [33].
ẋ
= −2x + x2 − y 2 ,
ẏ
= −6y(1 + ξ) + 2xy,
(34)
where ξ is a random variable with zero mean and uniformly
IX. C ONCLUSIONS
Weaker set-theoretic notion of almost everywhere stability
for stochastic dynamical system is introduced. Linear P-F
operator-based Lyapunov measure as a new tool for stochastic stability verification is introduced. Duality between the
Lyapunov function and the Lyapunov measure is established
in a stochastic setting. Set-oriented numerical methods are
proposed for the finite dimensional approximation of the
Lyapunov measure. The finite dimensional approximation introduce further weaker notion of stability refereed to as coarse
stochastic stability. The framework developed in this paper
can be easily extended to the case where the noise, ξn , forms
a Markov processes. This can be done by employing the PF operator defined for systems with Markov noise processes
[47].
11
Fig. 5.
Log plot of Lyapunov measure with noise support ξ ∈ [−0.5, 0.5].
Fig. 6.
Invariant measure plot for ξ ∈ [−1, 1].
X. ACKNOWLEDGMENTS
The author would like to acknowledge Venketash Chinde
from Iowa State University for help with simulations.
R EFERENCES
[1] R. Z. Has’minskiĭ, Stochastic Stability of differential equations. Germantown ,MD: Sijthoff & Noordhoff, 1980.
[2] L. Arnold, Random Dynamical Systems. Berlin, Heidenberg: Springer
Verlag, 1998.
[3] H. J. Kushner, Stocahstic Stability and Control. New York: Academic
Press, 1967.
[4] X. Mao, Exponential Stability of Stochastic Differential Equations.
Monographs and Textbooks in Pure and Applied Mathematics Series,
Marcel Dekker, Inc., 1994.
[5] O. Imer, S. Yuksel, and T. Basar, “Optimal control of LTI systems over
communication networks,” Automatica, vol. 42, no. 9, pp. 1429–1440,
2006.
[6] L. Schenato and B. Sinopoli and M. Franceschitti and K. Poolla and S.
Sastry, “Foundations of control and estimation over Lossy networks,”
Proceedings of IEEE, vol. 95, no. 1, pp. 163–187, 2007.
[7] N.Elia, “Remote stabilization over fading channels,” Systems and Control Letters, vol. 54, pp. 237–249, 2005.
[8] A. Diwadkar and U. Vaidya, “Limitation on nonlinear observation over
erasure channel,” IEEE Transactions on Automatic Control, vol. 58,
no. 2, pp. 454–459, 2013.
[9] A. Diwadkar and U. Vaidya, “Stabilization of linear time varying
systems over uncertain channels,” International Journal of Robust and
Nonlinear Control, vol. 24, no. 7, pp. 1205–1220, 2014.
[10] A. Diwadkar, S. Dasgupta, and U. Vaidya, “Control of systems in Lure
form over erasure channels,” Accepted for publication in International
Journal of Robust and Nonlinear Control, 2014. [Online]. Available:
http://dx.doi.org/10.1002/rnc.3231
[11] S. Pushpak, A. Diwadkar, and U. Vaidya, “Stochastic stability analysis
and controller systhesis for continuous time linear systems,” in Proceedings of IEEE Control and Decision Conference, Osaka, Japan, 2015.
[12] H. Deng, M. Kristić, and R. J. Williams, “Stabilization of stochastic
nonlinear systems driven by noise of unknown covariance,” IEEE
Transactions of Automatic Control, vol. 46, pp. 1237–1253, 2001.
[13] H. Deng and M. Kristić, “Output-feedback stabilization of stochastic
nonlinear systems driven by noise of unknown covariance,” Systems and
Control Letters, vol. 39, pp. 173–182, 2000.
[14] U. Vaidya and N. Elia, “Limitation on nonlinear stabilization over
packet-drop channels: Scalar case,” Systems and Control Letters, vol. 61,
no. 9, pp. 959–966, 2012.
[15] P. Florchinger, “Lyapunov like techniques for stochastic stability,” SIAM
J. Control Optim., vol. 33, pp. 1151–1169, 1995.
[16] ——, “A universal formula for the stabilization of control stochastic
differential equation,” Stochastic Annal. Appl., vol. 11, pp. 155–162,
1993.
[17] ——, “Feedback stabilization of affine in the control stochastic differential systems by the control lyapunov function method,” SIAM J. Control
Optim., vol. 35, p. 500511, 1997.
[18] H. Deng and M. Kristić, “Stocahstic nonlinear stabilization- part i:
backstepping design,” Systems and control letters, vol. 32, pp. 143–150,
1997.
[19] ——, “Stocahstic nonlinear stabilization- part II: Inverse optimality,”
Systems and control letters, vol. 32, pp. 151–159, 1997.
[20] ——, “Output-feedback stochastic nonlinear stabilization,” IEEE Trans.
Automat. Contr, vol. 44, p. 328333, 1999.
[21] D. Henrion and A. Garulli, Eds., Positive polynomials in control, ser.
Lecture Notes in Control and Information Sciences. Berlin: SpringerVerlag, 2005, vol. 312.
[22] S. Prajna, P. A. Parrilo, and A. Rantzer, “Nonlinear control synthesis by
convex optimization,” IEEE Transactions on Automatic Control, vol. 49,
no. 2, pp. 1–5, 2004.
[23] P. A. Parrilo, “Structured semidefinite programs and semialgebraic
geometry methods in robustness and optimization,” Ph.D. dissertation,
California Institute of Technology, Pasadena, CA, 2000.
[24] O. Junge and H. Osinga, “A set oriented approach to global optimal
control,” ESAIM: Control, Optimisation and Calculus of Variations,
vol. 10, no. 2, pp. 259–270, 2004.
[25] L. Grune and O. Junge, “A set oriented approach to optimal feedback
stabilization,” Systems Control Lett., vol. 54, no. 2, pp. 169–180, 2005.
[26] L. G. Crespo and J. Q. Sun, “Solution of fixed final state optimal control
problem via simple cell mapping,” Nonlinear dynamics, vol. 23, pp.
391–403, 2000.
[27] D. Hernández-Hernández, O. Hernández-Lerma, and M. Taksar, “The
linear programming approach to deterministic optimal control problem,”
Applicationes Mathematicae, vol. 24, no. 1, pp. 17–33, 1996.
[28] S. P. Meyn, “Algorithms for optimization and stabilization of controlled
Markov chains,” Sādhanā, vol. 24, no. 4-5, pp. 339–367, 1999.
[29] J. Lasserre, C. Prieur, and D. Henrion, “Nonlinear optimal control: Numerical approximation via moment and LMI-relaxations,” in Proceeding
of IEEE Conference on Decision and Control, Seville, Spain, 2005.
[30] W. McEneaney, Max-plus methods for nonlinear control and estimation.
Boston: Birkhauser, 2006.
[31] L. Grune, “Error estimation and adaptive discretization for the discrete
stochastic Hamilton-Jacobi-Bellman equation,” Numerische Mathematik,
vol. 99, pp. 85–112, 2004.
[32] L. G. Crespo and J. Q. Sun, “Stochastic optimal control via bellman
principle,” Automatica, vol. 39, pp. 2109–2114, 2003.
[33] A. Rantzer, “A dual to Lyapunov’s stability theorem,” Systems & Control
Letters, vol. 42, pp. 161–168, 2001.
[34] R. Van Handel, “Almost global stochastic stability,” SIAM Journal on
Control and Optimization, vol. 45, pp. 1297–1313, 2006.
[35] D. Angeli, “An almost global notion of input-to-state stability,” IEEE
Transactions on Automatic Control, vol. 49, pp. 866–874, 2004.
[36] U. Vaidya and P. G. Mehta, “Lyapunov measure for almost everywhere
stability,” IEEE Transactions on Automatic Control, vol. 53, no. 1, pp.
307–323, 2008.
[37] U. Vaidya, P. Mehta, and U. Shanbhag, “Nonlinear stabilization via
control lyapunov meausre,” IEEE Transactions on Automatic Control,
vol. 55, no. 6, pp. 1314–1328, 2010.
12
[38] A. Raghunathan and U. Vaidya, “Optimal stabilization using Lyapunov
measures,” IEEE Transactions on Automatic Control, vol. 59, no. 5, pp.
1316–1321, 2014.
[39] A. Mauroy and I. Mezic, “A spectral operator-theoretic framework for
global stability,” in Proc. of IEEE Conference of Decision and Control,
Florence, Italy, 2013.
[40] A. Lasota and M. C. Mackey, Chaos, Fractals, and Noise: Stochastic
Aspects of Dynamics. New York: Springer-Verlag, 1994.
[41] M. Dellnitz, O. Junge, W. S. Koon, F. Lekien, M. Lo, J. E. Marsden,
K. Padberg, R. Preis, S. D. Ross, and B. Thiere, “Transport in dynamical
astronomy and multibody problems,” International Journal of Bifurcation and Chaos, vol. 15, pp. 699–727, 2005.
[42] G. Froyland and M. Dellnitz, “Detecting and locating near-optimal
almost-invariant sets and cycles,” SIAM Journal on Scientific Computing,
vol. 24, no. 6, pp. 1839–1863, 2003.
[43] M. Budisic, R. Mohr, and I. Mezic, “Applied koopmanism,” Chaos,
vol. 22, pp. 047 510–32, 2012.
[44] I. Mezić and A. Banaszuk, “Comparison of systems with complex
behavior,” Physica D, vol. 197, pp. 101–133, 2004.
[45] I. Mezić and S. Wiggins, “A method for visualization of invariant sets
of dynamical systems based on ergodic partition,” Chaos, vol. 9, no. 1,
pp. 213–218, 1999.
[46] G. Froyland, K. Padberg, M. England, and A. M. Treguier, “Detection
of coherent oceanic structures using transfer operators,” Physical Review
Letters, vol. 98, pp. 2 245 031–4, 2007.
[47] G. Froyland, “Extracting dynamical behaviour via Markov models,”
in Nonlinear Dynamics and Statistics: Proceedings, Newton Institute,
Cambridge, 1998, A. Mees, Ed. Birkhauser, 2001, pp. 283–324.
[48] U. Vaidya, “Stochastic stability analysis of discrete-time system using
lyapunov measure,” in Proceedings of American Control Conference,
Chicago, IL, 2015, pp. 4646–4651.
[49] U. Vaidya and V. Chinde, “Computation of lyapunov measure for
stochastic stability verification,” in Accepted for publication in IEEE
Control and Decision Conference, Osaka, Japan, 2015.
[50] S. P. Meyn and R. L. Tweedie, Markov Chains and Stochastic Stability.
London: Springer-Verlag, 1993.
[51] M. Dellnitz and O. Junge, “On the approximation of complicated
dynamical behavior,” SIAM Journal on Numerical Analysis, vol. 36, pp.
491–515, 1999.
[52] ——, Set oriented numerical methods for dynamical systems. World
Scientific, 2000, pp. 221–264.
[53] J. Norris, Markov Chains, ser. Cambridge Series in Statistical and
Probabilistic Mathematics. Cambridge: Cambridge University Press,
1997.
| 3cs.SY
|
Journal of Machine Learning Research 17 (2016) 1-32
Submitted 10/15; Revised 4/16; Published 4/16
Stereo Matching by Training a Convolutional Neural
Network to Compare Image Patches
Jure Žbontar∗
[email protected]
arXiv:1510.05970v2 [cs.CV] 18 May 2016
Faculty of Computer and Information Science
University of Ljubljana
Večna pot 113, SI-1001 Ljubljana, Slovenia
Yann LeCun†
[email protected]
Courant Institute of Mathematical Sciences
New York University
715 Broadway, New York, NY 10003, USA
Editor: Zhuowen Tu
Abstract
We present a method for extracting depth information from a rectified image pair. Our
approach focuses on the first stage of many stereo algorithms: the matching cost computation. We approach the problem by learning a similarity measure on small image patches
using a convolutional neural network. Training is carried out in a supervised manner by
constructing a binary classification data set with examples of similar and dissimilar pairs
of patches. We examine two network architectures for this task: one tuned for speed, the
other for accuracy. The output of the convolutional neural network is used to initialize the
stereo matching cost. A series of post-processing steps follow: cross-based cost aggregation, semiglobal matching, a left-right consistency check, subpixel enhancement, a median
filter, and a bilateral filter. We evaluate our method on the KITTI 2012, KITTI 2015, and
Middlebury stereo data sets and show that it outperforms other approaches on all three
data sets.
Keywords: stereo, matching cost, similarity learning, supervised learning, convolutional
neural networks
1. Introduction
Consider the following problem: given two images taken by cameras at different horizontal
positions, we wish to compute the disparity d for each pixel in the left image. Disparity
refers to the difference in horizontal location of an object in the left and right image—an
object at position (x, y) in the left image appears at position (x − d, y) in the right image. If
we know the disparity of an object we can compute its depth z using the following relation:
z=
fB
,
d
∗. Jure Žbontar is also with the Courant Institute of Mathematical Sciences, New York University,
715 Broadway, New York, NY 10003, USA.
†. Yann LeCun is also with Facebook AI Research, 770 Broadway, New York, NY 10003, USA.
c 2016 Jure Žbontar and Yann LeCun.
Žbontar and LeCun
Left input image
Output disparity map
90 m
Right input image
20 m
1.7 m
Figure 1: The input is a pair of images from the left and right camera. The two input images
differ mostly in horizontal locations of objects (other differences are caused by
reflections, occlusions, and perspective distortions). Note that objects closer to
the camera have larger disparities than objects farther away. The output is a
dense disparity map shown on the right, with warmer colors representing larger
values of disparity (and smaller values of depth).
where f is the focal length of the camera and B is the distance between the camera centers.
Figure 1 depicts the input to and the output from our method.
The described problem of stereo matching is important in many fields such as autonomous driving, robotics, intermediate view generation, and 3D scene reconstruction.
According to the taxonomy of Scharstein and Szeliski (2002), a typical stereo algorithm
consists of four steps: matching cost computation, cost aggregation, optimization, and disparity refinement. Following Hirschmüller and Scharstein (2009) we refer to the first two
steps as computing the matching cost and the last two steps as the stereo method. The
focus of this work is on computing a good matching cost.
We propose training a convolutional neural network (LeCun et al., 1998) on pairs of
small image patches where the true disparity is known (for example, obtained by LIDAR
or structured light). The output of the network is used to initialize the matching cost. We
proceed with a number of post-processing steps that are not novel, but are necessary to
achieve good results. Matching costs are combined between neighboring pixels with similar
image intensities using cross-based cost aggregation. Smoothness constraints are enforced
by semiglobal matching and a left-right consistency check is used to detect and eliminate
errors in occluded regions. We perform subpixel enhancement and apply a median filter
and a bilateral filter to obtain the final disparity map.
The contributions of this paper are
• a description of two architectures based on convolutional neural networks for computing the stereo matching cost;
• a method, accompanied by its source code, with the lowest error rate on the KITTI
2012, KITTI 2015, and Middlebury stereo data sets; and
• experiments analyzing the importance of data set size, the error rate compared with
other methods, and the trade-off between accuracy and runtime for different settings
of the hyperparameters.
2
Stereo by Training a Network to Compare Patches
This paper extends our previous work (Žbontar and LeCun, 2015) by including a description of a new architecture, results on two new data sets, lower error rates, and more
thorough experiments.
2. Related Work
Before the introduction of large stereo data sets like KITTI and Middlebury, relatively few
stereo algorithms used ground truth information to learn parameters of their models; in
this section, we review the ones that did. For a general overview of stereo algorithms see
Scharstein and Szeliski (2002).
Kong and Tao (2004) used the sum of squared distances to compute an initial matching
cost. They then trained a model to predict the probability distribution over three classes:
the initial disparity is correct, the initial disparity is incorrect due to fattening of a foreground object, and the initial disparity is incorrect due to other reasons. The predicted
probabilities were used to adjust the initial matching cost. Kong and Tao (2006) later extend
their work by combining predictions obtained by computing normalized cross-correlation
over different window sizes and centers. Peris et al. (2012) initialized the matching cost
with AD-Census (Mei et al., 2011), and used multiclass linear discriminant analysis to learn
a mapping from the computed matching cost to the final disparity.
Ground-truth data was also used to learn parameters of probabilistic graphical models.
Zhang and Seitz (2007) used an alternative optimization algorithm to estimate optimal
values of Markov random field hyperparameters. Scharstein and Pal (2007) constructed a
new data set of 30 stereo pairs and used it to learn parameters of a conditional random
field. Li and Huttenlocher (2008) presented a conditional random field model with a nonparametric cost function and used a structured support vector machine to learn the model
parameters.
Recent work (Haeusler et al., 2013; Spyropoulos et al., 2014) focused on estimating the
confidence of the computed matching cost. Haeusler et al. (2013) used a random forest classifier to combine several confidence measures. Similarly, Spyropoulos et al. (2014) trained
a random forest classifier to predict the confidence of the matching cost and used the predictions as soft constraints in a Markov random field to decrease the error of the stereo
method.
A related problem to computing the matching cost is learning local image descriptors
(Brown et al., 2011; Trzcinski et al., 2012; Simonyan et al., 2014; Revaud et al., 2015;
Paulin et al., 2015; Han et al., 2015; Zagoruyko and Komodakis, 2015). The two problems
share a common subtask: to measure the similarity between image patches. Brown et al.
(2011) introduced a general framework for learning image descriptors and used Powell’s
method to select good hyperparameters. Several methods have been suggested for solving
the problem of learning local image descriptors, such as boosting (Trzcinski et al., 2012),
convex optimization (Simonyan et al., 2014), hierarchical moving-quadrant similarity (Revaud et al., 2015), convolutional kernel networks (Paulin et al., 2015), and convolutional
neural networks (Zagoruyko and Komodakis, 2015; Han et al., 2015). Works of Zagoruyko
and Komodakis (2015) and Han et al. (2015), in particular, are very similar to our own,
differing mostly in the architecture of the network; concretely, the inclusion of pooling and
subsampling to account for larger patch sizes and larger variation in viewpoint.
3
Žbontar and LeCun
3. Matching Cost
A typical stereo algorithm begins by computing a matching cost at each position p for all
disparities d under consideration. A simple method for computing the matching cost is the
sum of absolute differences:
X
CSAD (p, d) =
|I L (q) − I R (q − d)|,
(1)
q∈Np
where I L (p) and I R (p) are image intensities at position p in the left and right image and
Np is the set of locations within a fixed rectangular window centered at p.
We use bold lowercase letters p and q to denote image locations. A bold lowercase d
denotes the disparity d cast to a vector, that is, d = (d, 0). We use typewriter font for
the names of hyperparameters. For example, we would use patch size to denote the size
of the neighbourhood area Np .
Equation (1) can be interpreted as measuring the cost associated with matching a patch
from the left image, centered at position p, with a patch from the right image, centered at
position p − d. We want the cost to be low when the two patches are centered around the
image of the same 3D point, and high when they are not.
Since examples of good and bad matches can be constructed from publicly available
data sets (for example, the KITTI and Middlebury stereo data sets), we can attempt to
solve the matching problem by a supervised learning approach. Inspired by the successful
application of convolutional neural networks to vision problems, we used them to assess
how well two small image patches match.
3.1 Constructing the Data Set
We use ground truth disparity maps from either the KITTI or Middlebury stereo data sets
to construct a binary classification data set. At each image position where the true disparity
is known we extract one negative and one positive training example. This ensures that the
data set contains an equal number of positive and negative examples. A positive example
is a pair of patches, one from the left and one from the right image, whose center pixels are
the images of the same 3D point, while a negative example is a pair of patches where this
is not the case. The following section describes the data set construction step in detail.
L (p), P R (q) > denote a pair of patches, where P L (p) is an n × n patch
Let < Pn×n
n×n
n×n
R (q) is an n × n patch from the
from the left image centered at position p = (x, y), Pn×n
right image centered at position q, and d denotes the correct disparity at position p. A
negative example is obtained by setting the center of the right patch to
q = (x − d + oneg , y),
where oneg is chosen from either the interval [dataset neg low, dataset neg high] or, its
origin reflected counterpart, [−dataset neg high, −dataset neg low]. The random offset
oneg ensures that the resulting image patches are not centered around the same 3D point.
A positive example is derived by setting
q = (x − d + opos , y),
4
Stereo by Training a Network to Compare Patches
Similarity score
Dot product
Normalize
Normalize
Convolution
Convolution
Convolution, ReLU
Convolution, ReLU
Convolution, ReLU
Convolution, ReLU
Left input patch
Right input patch
Figure 2: The fast architecture is a siamese network. The two sub-networks consist of
a number of convolutional layers followed by rectified linear units (abbreviated
“ReLU”). The similarity score is obtained by extracting a vector from each of
the two input patches and computing the cosine similarity between them. In
this diagram, as well as in our implementation, the cosine similarity computation
is split in two steps: normalization and dot product. This reduces the running
time because the normalization needs to be performed only once per position (see
Section 3.3).
where opos is chosen randomly from the interval [−dataset pos, dataset pos]. The reason
for including opos , instead of setting it to zero, has to do with the stereo method used
later on. In particular, we found that cross-based cost aggregation performs better when
the network assigns low matching costs to good matches as well as near matches. In our
experiments, the hyperparameter dataset pos was never larger than one pixel.
3.2 Network Architectures
We describe two network architectures for learning a similarity measure on image patches.
The first architecture is faster than the second, but produces disparity maps that are slightly
less accurate. In both cases, the input to the network is a pair of small image patches and
the output is a measure of similarity between them. Both architectures contain a trainable
feature extractor that represents each image patch with a feature vector. The similarity
between patches is measured on the feature vectors instead of the raw image intensity
values. The fast architecture uses a fixed similarity measure to compare the two feature
vectors, while the accurate architecture attempts to learn a good similarity measure on
feature vectors.
3.2.1 Fast Architecture
The first architecture is a siamese network, that is, two shared-weight sub-networks joined
at the head (Bromley et al., 1993). The sub-networks are composed of a number of convolutional layers with rectified linear units following all but the last layer. Both sub-networks
output a vector capturing the properties of the input patch. The resulting two vectors are
5
Žbontar and LeCun
Similarity score
Fully-connected, Sigmoid
Fully-connected, ReLU
Fully-connected, ReLU
Fully-connected, ReLU
Concatenate
Convolution, ReLU
Convolution, ReLU
Convolution, ReLU
Convolution, ReLU
Convolution, ReLU
Convolution, ReLU
Left input patch
Righ input patch
Figure 3: The accurate architecture begins with two convolutional feature extractors. The
extracted feature vectors are concatenated and compared by a number of fullyconnected layers. The inputs are two image patches and the output is a single real
number between 0 and 1, which we interpret as a measure of similarity between
the input images.
compared using the cosine similarity measure to produce the final output of the network.
Figure 2 provides an overview of the architecture.
The network is trained by minimizing a hinge loss. The loss is computed by considering
pairs of examples centered around the same image position where one example belongs to
the positive and one to the negative class. Let s+ be the output of the network for the
positive example, s− be the output of the network for the negative example, and let m, the
margin, be a positive real number. The hinge loss for that pair of examples is defined as
max(0, m + s− − s+ ). The loss is zero when the similarity of the positive example is greater
than the similarity of the negative example by at least the margin m. We set the margin
to 0.2 in our experiments.
The hyperparameters of this architecture are the number of convolutional layers in each
sub-network (num conv layers), the size of the convolution kernels (conv kernel size),
the number of feature maps in each layer (num conv feature maps), and the size of the
input patch (input patch size).
3.2.2 Accurate Architecture
The second architecture is derived from the first by replacing the cosine similarity measure
with a number of fully-connected layers (see Figure 3). This architectural change increased
the running time, but decreased the error rate. The two sub-networks comprise a number
of convolutional layers, with a rectified linear unit following each layer. The resulting
two vectors are concatenated and forward-propagated through a number of fully-connected
6
Stereo by Training a Network to Compare Patches
layers followed by rectified linear units. The last fully-connected layer produces a single
number which, after being transformed with the sigmoid nonlinearity, is interpreted as the
similarity score between the input patches.
We use the binary cross-entropy loss for training. Let s denote the output of the network
for one training example and t denote the class of that training example; t = 1 if the example
belongs to the positive class and t = 0 if the example belongs to the negative class. The
binary cross-entropy loss for that example is defined as t log(s) + (1 − t) log(1 − s).
The decision to use two different loss functions, one for each architecture, was based on
empirical evidence. While we would have preferred to use the same loss function for both
architectures, experiments showed that the binary cross-entropy loss performed better than
the hinge loss on the accurate architecture. On the other hand, since the last step of the
fast architecture is the cosine similarity computation, a cross-entropy loss was not directly
applicable.
The hyperparameters of the accurate architecture are the number of convolutional layers in each sub-network (num conv layers), the number of feature maps in each layer
(num conv feature maps), the size of the convolution kernels (conv kernel size), the size
of the input patch (input patch size), the number of units in each fully-connected layer
(num fc units), and the number of fully-connected layers (num fc layers).
3.3 Computing the Matching Cost
The output of the network is used to initialize the matching cost:
CCNN (p, d) = −s(< P L (p), P R (p − d) >),
where s(< P L (p), P R (p − d) >) is the output of the network when run on input patches
P L (p) and P R (p − d). The minus sign converts the similarity score to a matching cost.
To compute the entire matching cost tensor CCNN (p, d) we would, naively, have to
perform the forward pass for each image location and each disparity under consideration.
The following three implementation details kept the running time manageable:
• The outputs of the two sub-networks need to be computed only once per location,
and do not need to be recomputed for every disparity under consideration.
• The output of the two sub-networks can be computed for all pixels in a single forward
pass by propagating full-resolution images, instead of small image patches. Performing
a single forward pass on the entire w × h image is faster than performing w · h forward
passes on small patches because many intermediate results can be reused.
• The output of the fully-connected layers in the accurate architecture can also be
computed in a single forward pass. This is done by replacing each fully-connected
layer with a convolutional layer with 1 × 1 kernels. We still need to perform the
forward pass for each disparity under consideration; the maximum disparity d is 228
for the KITTI data set and 400 for the Middlebury data set. As a result, the fullyconnected part of the network needs to be run d times, and is a bottleneck of the
accurate architecture.
7
Žbontar and LeCun
To compute the matching cost of a pair of images, we run the sub-networks once on each
image and run the fully-connected layers d times, where d is the maximum disparity under
consideration. This insight was important in designing the architecture of the network.
We could have chosen an architecture where the two images are concatenated before being
presented to the network, but that would imply a large cost at runtime because the whole
network would need to be run d times. This insight also led to the development of the
fast architecture, where the only layer that is run d times is the dot product of the feature
vectors.
4. Stereo Method
The raw outputs of the convolutional neural network are not enough to produce accurate
disparity maps, with errors particularly apparent in low-texture regions and occluded areas.
The quality of the disparity maps can be improved by applying a series of post-processing
steps referred to as the stereo method. The stereo method we used was influenced by Mei
et al. (2011) and comprises cross-based cost aggregation, semiglobal matching, a left-right
consistency check, subpixel enhancement, a median, and a bilateral filter.
4.1 Cross-based Cost Aggregation
Information from neighboring pixels can be combined by averaging the matching cost over
a fixed window. This approach fails near depth discontinuities, where the assumption of
constant depth within a window is violated. We might prefer a method that adaptively
selects the neighborhood for each pixel, so that support is collected only from pixels of the
same physical object. In cross-based cost aggregation (Zhang et al., 2009) we build a local
neighborhood around each location comprising pixels with similar image intensity values
with the hope that these pixels belong to the same object.
The method begins by constructing an upright cross at each position; this cross is used
to define the local support region. The left arm pl at position p extends left as long as the
following two conditions hold:
• |I(p) − I(pl )| < cbca intensity; the image intensities at positions p and pl should
be similar, their difference should be less than cbca intensity.
• kp − pl k < cbca distance; the horizontal distance (or vertical distance in case of top
and bottom arms) between positions p and pl is less than cbca distance pixels.
The right, bottom, and top arms are constructed analogously. Once the four arms are
known, we can compute the support region U (p) as the union of horizontal arms of all
positions q laying on p’s vertical arm (see Figure 4).
Zhang et al. (2009) suggest that aggregation should consider the support regions of both
images in a stereo pair. Let U L and U R denote the support regions in the left and right
image. We define the combined support region Ud as
Ud (p) = {q|q ∈ U L (p), q − d ∈ U R (p − d)}.
8
Stereo by Training a Network to Compare Patches
top arm
q
left arm pl
horizontal arms of q
p
right arm
bottom arm
Figure 4: The support region for position p is the union of horizontal arms of all positions
q on p’s vertical arm.
The matching cost is averaged over the combined support region:
0
CCBCA
(p, d) = CCNN (p, d),
X
1
i−1
i
(q, d),
CCBCA
(p, d) =
CCBCA
|Ud (p)|
q∈Ud (p)
where i is the iteration number. We repeat the averaging a number of times. Since the
support regions are overlapping, the results can change at each iteration. We skip crossbased cost aggregation in the fast architecture because it is not crucial for achieving a low
error rate and because it is relatively expensive to compute.
4.2 Semiglobal Matching
We refine the matching cost by enforcing smoothness constraints on the disparity image.
Following Hirschmüller (2008), we define an energy function E(D) that depends on the
disparity image D:
E(D) =
X
X
4
CCBCA
(p, D(p)) +
P1 · 1{|D(p) − D(q)| = 1}
p
q∈Np
+
X
P2 · 1{|D(p) − D(q)| > 1} ,
q∈Np
where 1{·} denotes the indicator function. The first term penalizes disparities with high
matching costs. The second term adds a penalty P1 when the disparity of neighboring pixels
differ by one. The third term adds a larger penalty P2 when the neighboring disparities
differ by more than one.
Rather than minimizing E(D) in all directions simultaneously, we could perform the
minimization in a single direction with dynamic programming. This solution would introduce unwanted streaking effects, since there would be no incentive to make the disparity
image smooth in the directions we are not optimizing over. In semiglobal matching we
9
Žbontar and LeCun
minimize the energy in a single direction, repeat for several directions, and average to obtain the final result. Although Hirschmüller (2008) suggested choosing sixteen direction, we
only optimized along the two horizontal and the two vertical directions; adding the diagonal
directions did not improve the accuracy of our system. To minimize E(D) in direction r,
we define a matching cost Cr (p, d) with the following recurrence relation:
Cr (p, d) =
4
CCBCA
(p, d)
− min Cr (p − r, k) + min Cr (p − r, d), Cr (p − r, d − 1) + P1 ,
k
Cr (p − r, d + 1) + P1 , min Cr (p − r, k) + P2 .
k
The second term is subtracted to prevent values of Cr (p, d) from growing too large and
does not affect the optimal disparity map.
The penalty parameters P1 and P2 are set according to the image gradient so that
jumps in disparity coincide with edges in the image. Let D1 = |I L (p) − I L (p − r)| and
D2 = |I R (p−d)−I R (p−d−r)| be the difference in image intensity between two neighboring
positions in the direction we are optimizing over. We set P1 and P2 according to the following
rules:
P1 = sgm P1,
P2 = sgm P2
if D1 < sgm D, D2 < sgm D;
P1 = sgm P1/sgm Q2, P2 = sgm P2/sgm Q2 if D1 ≥ sgm D, D2 ≥ sgm D;
P1 = sgm P1/sgm Q1, P2 = sgm P2/sgm Q1 otherwise.
The hyperparameters sgm P1 and sgm P2 set a base penalty for discontinuities in the disparity map. The base penalty is reduced by a factor of sgm Q1 if one of D1 or D2 indicate
a strong image gradient or by a larger factor of sgm Q2 if both D1 and D2 indicate a strong
image gradient. The value of P1 is further reduced by a factor of sgm V when considering
the two vertical directions; in the ground truth, small changes in disparity are much more
frequent in the vertical directions than in the horizontal directions and should be penalised
less.
The final cost CSGM (p, d) is computed by taking the average across all four directions:
CSGM (p, d) =
1X
Cr (p, d).
4 r
After semiglobal matching we repeat cross-based cost aggregation, as described in the
previous section. Hyperparameters cbca num iterations 1 and cbca num iterations 2
determine the number of cross-based cost aggregation iterations before and after semiglobal
matching.
4.3 Computing the Disparity Image
The disparity image D(p) is computed by the winner-takes-all strategy, that is, by finding
the disparity d that minimizes C(p, d),
D(p) = argmin C(p, d).
d
10
Stereo by Training a Network to Compare Patches
4.3.1 Interpolation
The interpolation steps attempt to resolve conflicts between the disparity map predicted
for the left image and the disparity map predicted for the right image. Let DL denote the
disparity map obtained by treating the left image as the reference image—this was the case
so far, that is, DL (p) = D(p)—and let DR denote the disparity map obtained by treating
the right image as the reference image. DL and DR sometimes disagree on what the correct
disparity at a particular position should be. We detect these conflicts by performing a
left-right consistency check. We label each position p by applying the following rules in
turn:
correct
if |d − DR (p − d)| ≤ 1 for d = DL (p),
mismatch if |d − DR (p − d)| ≤ 1 for any other d,
occlusion otherwise.
For positions marked as occlusion, we want the new disparity value to come from the
background. We interpolate by moving left until we find a position labeled correct and
use its value. For positions marked as mismatch, we find the nearest correct pixels in 16
different directions and use the median of their disparities for interpolation. We refer to the
interpolated disparity map as DINT .
4.3.2 Subpixel Enhancement
Subpixel enhancement provides an easy way to increase the resolution of a stereo algorithm.
We fit a quadratic curve through the neighboring costs to obtain a new disparity image:
DSE (p) = d −
C+ − C−
,
2(C+ − 2C + C− )
where d = DINT (p), C− = CSGM (p, d − 1), C = CSGM (p, d), and C+ = CSGM (p, d + 1).
4.3.3 Refinement
The final steps of the stereo method consist of a 5 × 5 median filter and the following
bilateral filter:
X
1
DBF (p) =
DSE (q) · g(kp − qk) · 1{|I L (p) − I L (q)| < blur threshold},
W (p)
q∈Np
where g(x) is the probability density function of a zero mean normal distribution with
standard deviation blur sigma and W (p) is the normalizing constant,
X
W (p) =
g(kp − qk) · 1{|I L (p) − I L (q)| < blur threshold}.
q∈Np
The role of the bilateral filter is to smooth the disparity map without blurring the edges.
DBF is the final output of our stereo method.
5. Experiments
We used three stereo data sets in our experiments: KITTI 2012, KITTI 2015, and Middlebury. The test set error rates reported in Tables 1, 2, and 4 were obtained by submitting
11
Žbontar and LeCun
Rank
1
2
3
4
5
6
7
8
9
10
Method
MC-CNN-acrt
Displets
MC-CNN
PRSM
MC-CNN-fst
SPS-StFl
VC-SF
Deep Embed
JSOSM
OSF
CoR
Setting
Accurate architecture
Güney and Geiger (2015)
Žbontar and LeCun (2015)
Vogel et al. (2015)
Fast architecture
Yamaguchi et al. (2014)
Vogel et al. (2014)
Chen et al. (2015)
Unpublished work
Menze and Geiger (2015)
Chakrabarti et al. (2015)
F, MV
F, MS
F, MV
F
Error
2.43
2.47
2.61
2.78
2.82
2.83
3.05
3.10
3.15
3.28
3.30
Runtime
67
265
100
300
0.8
35
300
3
105
3000
6
Table 1: The highest ranking methods on the KITTI 2012 data set as of October 2015.
The “Setting” column provides insight into how the disparity map is computed:
“F” indicates the use of optical flow, “MV” indicates more than two temporally
adjacent images, and “MS” indicates the use of epipolar geometry for computing
the optical flow. The “Error” column reports the percentage of misclassified pixels
and the “Runtime” column measures the time, in seconds, required to process one
pair of images.
the generated disparity maps to the online evaluation servers. All other error rates were
computed by splitting the data set in two, using one part for training and the other for
validation.
5.1 KITTI Stereo Data Set
The KITTI stereo data set (Geiger et al., 2013; Menze and Geiger, 2015) is a collection of
rectified image pairs taken from two video cameras mounted on the roof of a car, roughly
54 centimeters apart. The images were recorded while driving in and around the city of
Karlsruhe, in sunny and cloudy weather, at daytime. The images were taken at a resolution
of 1240 × 376. A rotating laser scanner mounted behind the left camera recorded ground
truth depth, labeling around 30 % of the image pixels.
The ground truth disparities for the test set are withheld and an online leaderboard
is provided where researchers can evaluate their method on the test set. Submissions are
allowed once every three days. Error is measured as the percentage of pixels where the
true disparity and the predicted disparity differ by more than three pixels. Translated into
distance, this means that, for example, the error tolerance is 3 centimeters for objects 2
meters from the camera and 80 centimeters for objects 10 meters from the camera.
12
Stereo by Training a Network to Compare Patches
Rank Method
1
2
3
4
5
6
7
8
9
10
MC-CNN-acrt
MC-CNN-fst
SPS-St
OSF
PR-Sceneflow
SGM+C+NL
SGM+LDOF
SGM+SF
ELAS
OCV-SGBM
SDM
Setting Error Runtime
Accurate architecture
Fast architecture
Yamaguchi et al. (2014)
Menze and Geiger (2015)
Vogel et al. (2013)
Hirschmüller (2008); Sun et al. (2014)
Hirschmüller (2008); Brox and Malik (2011)
Hirschmüller (2008); Hornacek et al. (2014)
Geiger et al. (2011)
Hirschmüller (2008)
Kostková and Sára (2003)
F
F
F
F
F
3.89
4.62
5.31
5.79
6.24
6.84
6.84
6.84
9.72
10.86
11.96
67
0.8
2
3000
150
270
86
2700
0.3
1.1
60
Table 2: The leading submission on the KITTI 2015 leaderboard as of October 2015. The
“Setting”, “Error”, and “Runtime” columns have the same meaning as in Table 1.
Two KITTI stereo data sets exist: KITTI 20121 and, the newer, KITTI 20152 . For the
task of computing stereo they are nearly identical, with the newer data set improving some
aspects of the optical flow task. The 2012 data set contains 194 training and 195 testing
images, while the 2015 data set contains 200 training and 200 testing images. There is a
subtle but important difference introduced in the newer data set: vehicles in motion are
densely labeled and car glass is included in the evaluation. This emphasizes the method’s
performance on reflective surfaces.
The best performing methods on the KITTI 2012 data set are listed in Table 1. Our
accurate architecture ranks first with an error rate of 2.43 %. Third place on the leaderboard
is held by our previous work (Žbontar and LeCun, 2015) with an error rate of 2.61 %. The
two changes that reduced the error from 2.61 % to 2.43 % were augmenting the data set
(see Section 5.4) and doubling the number of convolution layers while reducing the kernel
size from 5 × 5 to 3 × 3. The method in second place (Güney and Geiger, 2015) uses
the matching cost computed by our previous work (Žbontar and LeCun, 2015). The test
error rate of the fast architecture is 2.82 %, which would be enough for fifth place had the
method been allowed to appear in the public leaderboard. The running time for processing
a single image pair is 67 seconds for the accurate architecture and 0.8 seconds for the fast
architecture. Figure 5 contains a pair of examples from the KITTI 2012 data set, together
with the predictions of our method.
Table 2 presents the frontrunners on the KITTI 2015 data sets. The error rates of
our methods are 3.89 % for the accurate architecture and 4.46 % for the fast architecture,
occupying first and second place on the leaderboard. Since one submission per paper is
1. The KITTI 2012 scoreboard:
benchmark=stereo
2. The KITTI 2015 scoreboard:
benchmark=stereo
http://www.cvlibs.net/datasets/kitti/eval_stereo_flow.php?
http://www.cvlibs.net/datasets/kitti/eval_scene_flow.php?
13
Žbontar and LeCun
Left input image
Right input image
Ground truth
Census
Error: 4.63 %
Fast architecture
Error: 1.01 %
Accurate architecture
Error: 0.91 %
Left input image
Right input image
Ground truth
Census
Error: 14.12 %
Fast architecture
Error: 1.54 %
Accurate architecture
Error: 1.45 %
Figure 5: Examples of predicted disparity maps on the KITTI 2012 data set. Note how
some regions of the image (the white wall in the top example, and the asphalt
in the bottom example) cause problems for the census transform. The fast and
the accurate architecture perform better, with the accurate architecture making
fewer mistakes on average.
14
Stereo by Training a Network to Compare Patches
Year
Number of Image Pairs
Resolution
Maximum Disparity
2001
2003
2005
2006
2014
8
2
6
21
23
380 × 430
1800 × 1500
1400 × 1100
1400 × 1100
3000 × 2000
30
220
230
230
800
Table 3: A summary of the five Middlebury stereo data sets. The column “Number of
Image Pairs” counts only the image pairs for which ground truth is available. The
2005 and 2014 data sets additionally contain a number of image pairs with ground
truth disparities withheld; these image pairs constitute the test set.
allowed, only the result of the accurate architecture appears on the public leaderboard. See
Figure 6 for the disparity maps produced by our method on the KITTI 2015 data set.
5.2 Middlebury Stereo Data Set
The image pairs of the Middlebury stereo data set are indoor scenes taken under controlled
lighting conditions. Structured light was used to measure the true disparities with higher
density and precision than in the KITTI data set. The data sets were published in five
separate works in the years 2001, 2003, 2005, 2006, and 2014 (Scharstein and Szeliski, 2002,
2003; Scharstein and Pal, 2007; Hirschmüller and Scharstein, 2007; Scharstein et al., 2014).
In this paper, we refer to the Middlebury data set as the concatenation of all five data sets;
a summary of each is presented in Table 3.
Each scene in the 2005, 2006, and 2014 data sets was taken under a number of lighting conditions and shutter exposures, with a typical image pair taken under four lighting
conditions and seven exposure settings for a total of 28 images of the same scene.
An online leaderboard3 , similar to the one provided by KITTI, displays a ranked list
of all submitted methods. Participants have only one opportunity to submit their results
on the test set to the public leaderboard. This rule is stricter than the one on the KITTI
data set, where submissions are allowed every three days. The test set contains 15 images
borrowed from the 2005 and 2014 data sets.
The data set is provided in full, half, and quarter resolution. The error is computed at
full resolution; if the method outputs half or quarter resolution disparity maps, they are
upsampled before the error is computed. We chose to run our method on half resolution
images because of the limited size of the graphic card’s memory available.
Rectifying a pair of images using standard calibration procedures, like the ones present in
the OpenCV library, results in vertical disparity errors of up to nine pixels on the Middlebury
data set (Scharstein et al., 2014). Each stereo pair in the 2014 data set is rectified twice:
once using a standard, imperfect approach, and once using precise 2D correspondences for
perfect rectification (Scharstein et al., 2014). We train the network on imperfectly rectified
3. The Middlebury scoreboard: http://vision.middlebury.edu/stereo/eval3/
15
Žbontar and LeCun
Left input image
Right input image
Ground truth
Census
Error: 4.58 %
Fast architecture
Error: 2.79 %
Accurate architecture
Error: 2.36 %
Left input image
Right input image
Ground truth
Census
Error: 5.25 %
Fast architecture
Error: 3.91 %
Accurate architecture
Error: 3.73 %
Figure 6: Examples of predictions on the KITTI 2015 data set. Observe that vehicles in
motion are labeled densely in the KITTI 2015 data set.
16
Stereo by Training a Network to Compare Patches
Rank
1
2
3
4
5
6
7
8
9
10
Method
MC-CNN-acrt
MeshStereo
LCU
TMAP
IDR
SGM
LPS
LPS
SGM
SNCC
Accurate architecture
Zhang et al. (2015)
Unpublished work
Psota et al. (2015)
Kowalczuk et al. (2013)
Hirschmüller (2008)
Sinha et al. (2014)
Sinha et al. (2014)
Hirschmüller (2008)
Einecke and Eggert (2010)
Resolution
Error
Runtime
Half
Half
Quarter
Half
Half
Half
Half
Full
Quarter
Half
8.29
13.4
17.0
17.1
18.4
18.7
19.4
20.3
21.2
22.2
150
65.3
6567
2435
0.49
9.90
9.52
25.8
1.48
1.38
Table 4: The top ten methods on the Middlebury stereo data set as of October 2015. The
“Error” column is the weighted average error after upsampling to full resolution
and “Runtime” is the time, in seconds, required to process one pair of images.
image pairs, since only two of the fifteen test images (Australia and Crusade) are rectified
perfectly.
The error is measured as the percentage of pixels where the true disparity and the
predicted disparity differ by more than two pixels; this corresponds to an error tolerance
of one pixel at half resolution. The error on the evaluation server is, by default, computed
only on non-occluded pixels. The final error reported online is the weighted average over
the fifteen test images, with weights set by the authors of the data set.
Table 4 contains a snapshot of the third, and newest, version of the Middlebury leaderboard. Our method ranks first with an error rate of 8.29 % and a substantial lead over the
second placed MeshStereo method, whose error rate is 13.4 %. See Figure 7 for disparity
maps produced by our method on one image pair from the Middlebury data set.
5.3 Details of Learning
We construct a binary classification data set from all available image pairs in the training
set. The data set contains 25 million examples on the KITTI 2012, 17 million examples on
the KITTI 2015, and 38 million examples on the Middlebury data set.
At training time, the input to the network was a batch of 128 pairs of image patches.
At test time, the input was the entire left and right image. We could have used entire
images during training as well, as it would allow us to implement the speed optimizations
described in Section 3.3. There were several reasons why we preferred to train on image
patches: it was easier to control the batch size, the examples could be shuffled so that one
batch contained patches from several different images, and it was easier to maintain the
same number of positive and negative examples within a batch.
We minimized the loss using mini-batch gradient descent with the momentum term
set to 0.9. We trained for 14 epochs with the learning rate initially set to 0.003 for the
accurate architecture and 0.002 for the fast architecture. The learning rate was decreased
17
Žbontar and LeCun
Left input image
Right input image
Census
Error: 34.65 %
Fast architecture
Error: 16.19 %
Accurate architecture
Error: 7.25 %
Ground truth
Figure 7: An example of a particularly difficult image pair from the Middlebury data set; the
white wall in the background is practically textureless. The accurate architecture
is able to classify most of it correctly. The fast architecture doesn’t do as well
but still performs better than census.
18
Stereo by Training a Network to Compare Patches
KITTI 2012
Hyperparameter
input patch size
num conv layers
num conv feature maps
conv kernel size
num fc layers
num fc units
dataset neg low
dataset neg high
dataset pos
cbca intensity
cbca distance
cbca num iterations 1
cbca num iterations 2
sgm P1
sgm P2
sgm Q1
sgm Q2
sgm V
sgm D
blur sigma
blur threshold
KITTI 2015
Middlebury
fst
acrt
fst
acrt
fst
acrt
9×9
4
64
3
9×9
4
112
3
4
384
4
10
1
0.13
5
2
0
1.32
32
3
6
2
0.08
6
6
9×9
4
64
3
9×9
4
112
3
4
384
4
10
1
0.03
5
2
4
2.3
55.8
3
6
1.75
0.08
6
5
11 × 11
5
64
3
11 × 11
5
112
3
3
384
1.5
18
0.5
0.02
14
2
16
1.3
18.1
4.5
9
2.75
0.13
1.7
2
4
10
1
4
223
3
7.5
1.5
0.02
7.74
5
4
10
1
2.3
42.3
3
6
1.25
0.08
4.64
5
1.5
6
0.5
2.3
55.9
4
8
1.5
0.08
6
2
Table 5: The hyperparameter values we used for the fast and accurate architectures (abbreviated “fst” and “acrt”). Note that hyperparameters concerning image intensity
values (cbca intensity and sgm D) apply to the preprocessed images and not to
raw images with intensity values in the range from 0 to 255.
by a factor of 10 on the 11th epoch. The number of epochs, the initial learning rate, and the
learning rate decrease schedule where treated as hyperparameters and were optimized with
cross-validation. Each image was preprocessed by subtracting the mean and dividing by
the standard deviation of its pixel intensity values. The left and right image of a stereo pair
were preprocessed separately. Our initial experiments suggested that using color information
does not improve the quality of the disparity maps; therefore, we converted all color images
to grayscale.
The post-processing steps of the stereo method were implemented in CUDA (Nickolls
et al., 2008), the network training was done with the Torch environment (Collobert et al.,
2011) using the convolution routines from the cuDNN library (Chetlur et al., 2014). The
OpenCV library (Bradski, 2000) was used for the affine transformation in the data augmentation step.
19
Žbontar and LeCun
The hyperparameters where optimized with manual search and simple scripts that
helped automate the process. The hyperparameters we selected are shown in Table 5.
5.4 Data Set Augmentation
Augmenting the data set by repeatedly transforming the training examples is a commonly
employed technique to reduce the network’s generalization error. The transformations are
applied at training time and do not affect the runtime performance. We randomly rotate,
scale and shear the training patches; we also change their brightness and contrast. Since
the transformations are applied to patches after they have been extracted from the images,
the data augmentation step does not alter the ground truth disparity map or ruin the
rectification.
The parameters of the transformation are chosen randomly for each pair of patches,
and after one epoch of training, when the same example is being presented to the network
for the second time, new random parameters are selected. We choose slightly different
transformation parameters for the left and right image; for example, we would rotate the
left patch by 10 degrees and the right by 14. Different data sets benefited from different
types of transformations and, in some cases, using the wrong transformations increased the
error.
On the Middlebury data set we took advantage of the fact that the images were taken
under different lighting conditions and different shutter exposures by training on all available
images. The same data set augmentation parameters were used for the KITTI 2012 and
KITTI 2015 data sets.
The Middlebury test data sets contains two images worth mentioning: Classroom, where
the right image is underexposed and, therefore, darker than the left; and Djembe, where the
left and right images were taken under different light conditions. To handle these two cases
we train, 20 % of the time, on images where either the shutter exposure or the arrangements
of lights are different for the left and right image.
We combat imperfect rectification on the Middlebury data set by including a small
vertical disparity between the left and right image patches.
Before describing the steps of data augmentation, let us introduce some notation: in the
following, a word in typewriter is used to denote the name of a hyperparameter defining a
set, while the same word in italic is used to denote a number drawn randomly from that set.
For example, rotate is a hyperparameter defining the set of possible rotations and rotate
is a number drawn randomly from that set. The steps of data augmentation are presented
in the following list:
• Rotate the left patch by rotate degrees and the right patch by rotate + rotate diff
degrees.
• Scale the left patch by scale and the right patch by scale · scale diff.
• Scale the left patch in the horizontal direction by horizontal scale and the right patch
by horizontal scale · horizontal scale diff.
• Shear the left patch in the horizontal direction by horizontal shear and the right patch
by horizontal shear + horizontal shear diff.
20
Stereo by Training a Network to Compare Patches
KITTI 2012
Middlebury
Hyperparameter
Range
Error
Range
Error
rotate
scale
horizontal scale
horizontal shear
brightness
contrast
vertical disparity
rotate diff
horizontal scale diff
horizontal shear diff
brightness diff
contrast diff
[−7, 7]
2.65
[0.9, 1]
[0, 0.1]
[0, 0.7]
[1, 1.3]
2.62
2.61
2.61
2.63
[0, 0.3]
2.63
[−28, 28]
[0.8, 1]
[0.8, 1]
[0, 0.1]
[0, 1.3]
[1, 1.1]
[0, 1]
[−3, 3]
[0.9, 1]
[0, 0.3]
[0, 0.7]
[1, 1.1]
7.99
8.17
8.08
7.91
8.16
7.95
8.05
8.00
7.97
8.05
7.92
8.01
No data set augmentation
Full data set augmentation
2.73
2.61
8.75
7.91
Table 6: The hyperparameters governing data augmentation and how they affect the validation error. The “Error” column reports the validation error when a particular
data augmentation step is not used. The last two rows report validation errors
with and without data augmentation. For example, the validation error on the
KITTI 2012 is 2.73 % if no data augmentation is used, 2.65 % if all steps except
rotation are used, and 2.61 % if all data augmentation steps are used.
• Translate the right patch in the vertical direction by vertical disparity.
• Adjust the brightness and contrast by setting the left and right image patches to:
P L ← P L · contrast + brightness and
P R ← P R · (contrast · contrast diff) + (brightness + brightness diff),
with addition and multiplication carried out element-wise where appropriate.
Table 6 contains the hyperparameters used and measures how each data augmentation step
affected the validation error.
Data augmentation reduced the validation error from 2.73 % to 2.61 % on the KITTI
2012 data set and from 8.75 % to 7.91 % on the Middlebury data set.
5.5 Runtime
We measure the runtime of our implementation on a computer with a NVIDIA Titan X
graphics processor unit. Table 7 contains the runtime measurements across a range of
hyperparameter settings for three data sets: KITTI, Middlebury half resolution, and a
21
Žbontar and LeCun
new, fictitious data set, called Tiny, which we use to demonstrate the performance of our
method on the kind of images typically used for autonomous driving or robotics. The sizes
of images we measured the runtime on were: 1242 × 350 with 228 disparity levels for the
KITTI data set, 1500 × 1000 with 200 disparity levels for the Middlebury data set, and 320
× 240 with 32 disparity levels for the Tiny data set.
Table 7 reveals that the fast architecture is up to 90 times faster than the accurate
architecture. Furthermore, the running times of the fast architecture are 0.78 seconds on
KITTI, 2.03 seconds on Middlebury, and 0.06 seconds on the Tiny data set. We can also
see that the fully-connected layers are responsible for most of the runtime in the accurate
architecture, as the hyperparameters controlling the number of convolutional layer and the
number of feature maps have only a small effect on the runtime.
Training times depended on the size of the data set and the architecture, but never
exceeded two days.
5.6 Matching Cost
We argue that the low error rate of our method is due to the convolutional neural network
and not a superior stereo method. We verify this claim by replacing the convolutional neural
network with three standard approaches for computing the matching cost:
• The sum of absolute differences computes the matching cost according to Equation (1),
that is, the matching cost between two image patches is computed by summing the
absolute differences in image intensities between corresponding locations. We used
9 × 9 patches.
• The census transform (Zabih and Woodfill, 1994) represents each image position as a
bit vector. The size of this vector is a hyperparameter whose value, after examining
several, we set to 81. The vector is computed by cropping a 9×9 image patch centered
around the position of interest and comparing the intensity values of each pixel in the
patch to the intensity value of the pixel in the center. When the center pixel is brighter
the corresponding bit is set. The matching cost is computed as the hamming distance
between two census transformed vectors.
• Normalized cross-correlation is a window-based method defined with the following
equation:
P
L
R
q∈Np I (q)I (q − d)
CN CC (p, d) = qP
.
P
L
2
R
2
q∈Np I (q)
q∈Np I (q − d)
The normalized cross-correlation matching cost computes the cosine similarity between the left and right image patch, when the left and right image patches are
viewed as vectors instead of matrices. This is the same function that is computed
in the last two layers of the fast architecture (normalization and dot product). The
neighbourhood Np was set to a square 11 × 11 window around p.
The “sad”, “cens”, and “ncc” columns of Table 8 contain the results of the sum of absolute
differences, the census transform, and normalized cross-correlation on the KITTI 2012,
KITTI 2015, and Middlebury data sets. The validation errors in the last rows of Table 8
22
Stereo by Training a Network to Compare Patches
KITTI
Hyperparameter
num conv layers
num conv feature maps
Middlebury
Tiny
fst
acrt
fst
acrt
fst
acrt
1
2
3
4
5
6
0.23
0.26
0.30
0.34
0.38
0.42
66.1
66.2
66.3
66.4
66.5
66.7
0.70
0.82
0.97
1.11
1.24
1.37
74.9
75.2
75.4
75.6
75.7
76.0
0.01
0.01
0.02
0.03
0.03
0.04
1.7
1.8
1.8
1.8
1.9
1.9
16
32
48
64
80
96
112
128
0.09
0.15
0.25
0.34
0.44
0.53
0.61
0.71
59.4
60.4
61.5
62.7
64.0
65.3
66.4
67.7
0.27
0.51
0.94
1.24
1.63
1.93
2.28
2.61
64.8
66.2
68.2
70.0
72.0
73.9
75.7
77.8
0.01
0.01
0.02
0.03
0.04
0.04
0.05
0.06
1.6
1.6
1.7
1.7
1.8
1.8
1.8
1.9
num fc layers
1
2
3
4
5
16.3
32.9
49.6
66.4
82.9
25.3
50.7
75.7
101.2
126.4
0.5
0.9
1.4
1.8
2.3
num fc units
128
256
384
512
17.4
38.5
66.4
101.0
21.4
44.9
75.7
113.3
0.6
1.1
1.8
2.7
No stereo method
Full stereo method
0.34
0.78
66.4
67.1
1.24
2.03
75.7
84.8
0.03
0.06
1.8
1.9
Table 7: The time, in seconds, required to compute the matching cost, that is, the time
spent in the convolutional neural network without any post-processing steps. The
time does include computing the matching cost twice: once when the left image is
taken to be the reference image and once when the right image is taken to be the
reference image. We measure the runtime as a function of four hyperparameters
controlling the network architecture; for example, the first six rows contain the
runtime as the number of convolutional layers in the network increases from one
to six. The last row of the table contains the running time for the entire method,
including the post-processing steps. As before, we abbreviate the fast and accurate
architectures as “fst” and “acrt”.
23
Žbontar and LeCun
should be used to compare the five methods. On all three data sets the accurate architecture
performs best, followed by the fast architecture, which in turn is followed by the census
transform. These are the three best performing methods on all three data sets. Their error
rates are 2.61 %, 3.02 %, and 4.90 % on KITTI 2012; 3.25 %, 3.99 %, and 5.03 % on KITTI
2015; and 7.91 %, 9.87 %, and 16.72 % on Middlebury. The sum of absolute differences and
the normalized cross-correlation matching costs produce disparity maps with larger errors.
For a visual comparison of our method and the census transform see Figures 5, 6, and 7.
5.7 Stereo Method
The stereo method includes a number of post-processing steps: cross-based cost aggregation,
semiglobal matching, interpolation, subpixel enhancement, a median, and a bilateral filter.
We ran a set of experiments in which we excluded each of the aforementioned steps and
recorded the validation error (see Table 8).
The last two rows of Table 8 allude to the importance of the post-processing steps of the
stereo method. We see that, if all post-processing steps are removed, the validation error
of the accurate architecture increases from 2.61 % to 13.49 % on KITTI 2012, from 3.25 %
to 13.38 % on KITTI 2015, and from 7.91 % to 28.33 % on Middlebury.
Out of all post-processing steps of the stereo method, semiglobal matching affects the
validation error the strongest. If we remove it, the validation error increases from 2.61 % to
4.26 % on KITTI 2012, from 3.25 % to 4.51 % on KITTI 2015, and from 7.91 % to 11.99 %
on Middlebury.
We did not use the left-right consistency check to eliminate errors in occluded regions on
the Middlebury data set. The error rate increased from 7.91 % to 8.22 % using the left-right
consistency check on the accurate architecture, which is why we decided to remove it.
5.8 Data Set Size
We used a supervised learning approach to measure the similarity between image patches.
It is, therefore, natural to ask how does the size of the data set affect the quality of the
disparity maps. To answer this question, we retrain our networks on smaller training sets
obtained by selecting a random set of examples (see Table 9).
We observe that the validation error decreases as we increase the number of training
examples. These experiments suggest a simple strategy for improving the results of our
stereo method: collect a larger data set.
5.9 Transfer Learning
Up to this point the training and validation sets were created from the same stereo data
set, either KITTI 2012, KITTI 2015, or Middlebury. To evaluate the performance of our
method in the transfer learning setting, we run experiments where the validation error is
computed on a different data set than the one used for training. For example, we would
use the Middlebury data set to train the matching cost neural network and evaluate its
performance on the KITTI 2012 data set. These experiments give us some idea of the expected performance in a real-world application, where it isn’t possible to train a specialized
24
Stereo by Training a Network to Compare Patches
KITTI 2012
fst
acrt
sad
cens
ncc
Cross-based cost aggregation
Semiglobal matching
Interpolation
Subpixel Enhancement
Median filter
Bilateral filter
3.02
8.78
3.48
3.03
3.03
3.26
2.73
4.26
2.96
2.65
2.63
2.79
8.22
19.58
9.21
8.16
8.16
8.75
5.21
8.84
5.96
4.95
4.92
5.70
8.93
10.72
11.16
8.93
9.00
9.76
No stereo method
Full stereo method
15.70
3.02
13.49
2.61
32.30
8.16
53.55
4.90
22.21
8.93
KITTI 2015
fst
acrt
sad
cens
ncc
Cross-based cost aggregation
Semiglobal matching
Interpolation
Subpixel Enhancement
Median filter
Bilateral filter
3.99
8.40
4.47
4.02
4.05
4.20
3.39
4.51
3.33
3.28
3.25
3.43
9.94
19.80
10.39
9.44
9.44
9.95
5.20
7.25
5.83
5.03
5.05
5.84
8.89
9.36
10.98
8.91
8.96
9.77
No stereo method
Full stereo method
15.66
3.99
13.38
3.25
30.67
9.44
50.35
5.03
18.95
8.89
Middlebury
fst
acrt
sad
cens
ncc
Cross-based cost aggregation
Semiglobal matching
Interpolation
Subpixel Enhancement
Median filter
Bilateral filter
9.87
25.50
9.87
10.29
10.16
10.39
10.63
11.99
7.91
8.44
7.91
7.96
43.09
51.25
41.86
42.71
41.90
41.97
29.28
19.51
16.72
17.18
16.73
16.96
33.89
35.36
33.89
34.12
34.17
34.43
No stereo method
Full stereo method
30.84
9.87
28.33
7.91
59.57
41.86
64.53
16.72
39.23
33.89
Table 8: The numbers measure validation error when a particular post-processing step is
excluded from the stereo method. The last two rows of the tables should be
interpreted differently: they contain the validation error of the raw convolutional
neural network and the validation error after the complete stereo method. For
example, if we exclude semiglobal matching, the fast architecture achieves an error
rate of 8.78 % on the KITTI 2012 data set and an error rate of 3.02 % after applying
the full stereo method. We abbreviate the method names as “fst” for the fast
architecture, “acrt” for the accurate architecture, “sad” for the sum of absolute
differences, “cens” for the census transform, and “ncc” for the normalized crosscorrelation matching cost.
25
Žbontar and LeCun
KITTI 2012
KITTI 2015
Middlebury
Data Set Size (%)
fst
acrt
fst
acrt
fst
acrt
20
40
60
80
100
3.17
3.11
3.09
3.05
3.02
2.84
2.75
2.67
2.65
2.61
4.13
4.10
4.05
4.02
3.99
3.53
3.40
3.34
3.29
3.25
11.14
10.35
10.14
10.09
9.87
9.73
8.71
8.36
8.21
7.91
Table 9: The validation error as a function of training set size.
Test Set
KITTI 2012
Training Set
KITTI 2012
KITTI 2015
Middlebury
KITTI 2015
Middlebury
fst
acrt
fst
acrt
fst
acrt
3.02
3.60
3.16
2.61
4.28
3.07
4.12
3.99
4.48
3.99
3.25
4.49
12.78
13.70
9.87
11.09
14.19
7.91
Table 10: The validation error when the training and test sets differ. For example, the
validation error is 3.16 % when the Middlebury data set is used for training the
fast architecture and the trained network is tested on the KITTI 2012 data set.
network because no ground truth is available. The results of these experiments are shown
in Table 10.
Some results in Table 10 were unexpected. For example, the validation error on KITTI
2012 is lower when using the Middlebury training set compared to the KITTI 2015 training
set, even though the KITTI 2012 data set is obviously more similar to KITTI 2015 than
Middlebury. Furthermore, the validation error on KITTI 2012 is lower when using the fast
architecture instead of the accurate architecture when training on KITTI 2015.
The matching cost neural network trained on the Middlebury data set transfers well
to the KITTI data sets. Its validation error is similar to the validation errors obtained by
networks trained on the KITTI data sets.
5.10 Hyperparameters
Searching for a good set of hyperparameters is a daunting task—with the search space
growing exponentially with the number of hyperparameters and no gradient to guide us.
To better understand the effect of each hyperparameter on the validation error, we conduct
a series of experiments where we vary the value of one hyperparameter while keeping the
others fixed to their default values. The results are shown in Table 11 and can be summarized by observing that increasing the size of the network improves the generalization
26
Stereo by Training a Network to Compare Patches
KITTI 2012
Hyperparameter
num conv layers
num conv feature maps
KITTI 2015
Middlebury
fst
acrt
fst
acrt
fst
acrt
1
2
3
4
5
6
5.96
3.52
3.10
3.02
3.03
3.05
3.97
2.98
2.72
2.61
2.64
2.70
5.61
4.19
4.04
3.99
3.99
4.01
4.06
3.45
3.27
3.25
3.30
3.38
20.74
12.11
10.81
10.26
9.87
9.71
12.37
9.20
8.56
8.21
7.91
8.11
16
32
48
64
80
96
112
128
3.33
3.15
3.07
3.02
3.02
2.99
2.98
2.97
2.84
2.68
2.66
2.64
2.64
2.68
2.61
2.63
4.32
4.12
4.06
3.99
3.99
3.97
3.96
3.95
3.51
3.35
3.32
3.30
3.29
3.27
3.25
3.23
11.79
10.48
10.20
9.87
9.81
9.62
9.59
9.45
10.06
8.67
8.47
8.12
7.95
8.03
7.91
7.92
num fc layers
1
2
3
4
5
2.83
2.70
2.62
2.61
2.62
3.50
3.31
3.30
3.25
3.29
8.52
8.33
8.06
8.00
7.91
num fc units
128
256
384
512
2.72
2.65
2.61
2.60
3.36
3.28
3.25
3.23
8.44
8.03
7.91
7.90
dataset neg low
1.0
1.5
2.0
4.0
6.0
3.00
3.00
2.99
3.02
3.06
2.76
2.71
2.63
2.61
2.63
3.97
3.97
3.98
3.99
4.05
3.35
3.33
3.31
3.25
3.28
9.84
9.87
9.98
10.20
10.13
8.00
7.91
8.08
8.66
8.86
dataset neg high
6
10
14
18
22
3.00
3.02
3.04
3.07
3.07
2.72
2.61
2.61
2.60
2.61
3.98
3.99
4.02
4.06
4.05
3.30
3.25
3.25
3.23
3.24
9.87
9.97
10.00
9.98
10.16
8.59
8.23
8.05
8.11
7.91
dataset pos
0.0
0.5
1.0
1.5
2.0
3.04
3.02
3.02
3.04
3.04
2.67
2.65
2.61
2.62
2.66
4.00
3.99
3.99
4.04
4.04
3.26
3.28
3.25
3.27
3.29
9.92
9.87
9.86
10.00
10.16
7.97
7.91
8.04
8.34
8.51
Table 11: Validation errors computed across a range of hyperparameter settings.
27
Žbontar and LeCun
performance, but only up to a point, when presumably, because of the size of the data set,
the generalization performance starts do decrease.
Note that the num conv layers hyperparameter implicitly controls the size of the image
patches. For example, a network with one convolutional layer with 3 × 3 kernels compares
image patches of size 3 × 3, while a network with five convolutional layers compares patches
of size 11 × 11.
6. Conclusion
We presented two convolutional neural network architectures for learning a similarity measure on image patches and applied them to the problem of stereo matching.
The source code of our implementation is available at https://github.com/jzbontar/
mc-cnn. The online repository contains procedures for computing the disparity map, training the network, as well as the post-processing steps of the stereo method.
The accurate architecture produces disparity maps with lower error rates than any
previously published method on the KITTI 2012, KITTI 2015, and Middlebury data sets.
The fast architecture computes the disparity maps up to 90 times faster than the accurate
architecture with only a small increase in error. These results suggest that convolutional
neural networks are well suited for computing the stereo matching cost even for applications
that require real-time performance.
The fact that a relatively simple convolutional neural network outperformed all previous
methods on the well-studied problem of stereo is a rather important demonstration of the
power of modern machine learning approaches.
References
G. Bradski. The OpenCV library. Dr. Dobb’s Journal of Software Tools, 2000.
Jane Bromley, James W Bentz, Léon Bottou, Isabelle Guyon, Yann LeCun, Cliff Moore,
Eduard Säckinger, and Roopak Shah. Signature verification using a siamese time delay
neural network. International Journal of Pattern Recognition and Artificial Intelligence,
7(04):669–688, 1993.
Matthew Brown, Gang Hua, and Simon Winder. Discriminative learning of local image
descriptors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):
43–57, 2011.
Thomas Brox and Jitendra Malik. Large displacement optical flow: descriptor matching
in variational motion estimation. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 33(3):500–513, 2011.
Ayan Chakrabarti, Ying Xiong, Steven J. Gortler, and Todd Zickler. Low-level vision by
consensus in a spatial hierarchy of regions. IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), June 2015.
Zhuoyuan Chen, Xun Sun, Yinan Yu, Liang Wang, and Chang Huang. A deep visual correspondence embedding model for stereo matching costs. IEEE International Conference
on Computer Vision (ICCV), 2015.
28
Stereo by Training a Network to Compare Patches
Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan
Catanzaro, and Evan Shelhamer. cuDNN: Efficient primitives for deep learning. CoRR,
abs/1410.0759, 2014. URL http://arxiv.org/abs/1410.0759.
Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A matlab-like environment for machine learning. In BigLearn, NIPS Workshop, 2011.
Nils Einecke and Julian Eggert. A two-stage correlation method for stereoscopic depth
estimation. In Digital Image Computing: International Conference on Techniques and
Applications (DICTA), pages 227–234, 2010.
Andreas Geiger, Martin Roser, and Raquel Urtasun. Efficient large-scale stereo matching. In
Proceedings of the 10th Asian Conference on Computer Vision - Volume Part I, ACCV’10,
pages 25–38. Springer-Verlag, Berlin, Heidelberg, 2011.
Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics:
the KITTI dataset. International Journal of Robotics Research (IJRR), 2013.
Fatma Güney and Andreas Geiger. Displets: Resolving stereo ambiguities using object
knowledge. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
June 2015.
Ralf Haeusler, Rahul Nair, and Daniel Kondermann. Ensemble learning for confidence
measures in stereo vision. IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), June 2013.
Xufeng Han, Thomas Leung, Yangqing Jia, Rahul Sukthankar, and Alexander C Berg.
MatchNet: Unifying feature and metric learning for patch-based matching. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2015.
Heiko Hirschmüller. Stereo processing by semiglobal matching and mutual information.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328–341, 2008.
Heiko Hirschmüller and Daniel Scharstein. Evaluation of cost functions for stereo matching.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007.
Heiko Hirschmüller and Daniel Scharstein. Evaluation of stereo matching costs on images
with radiometric differences. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(9):1582–1599, 2009.
Michael Hornacek, Andrew Fitzgibbon, and Carsten Rother. SphereFlow: 6 DoF scene
flow from RGB-D pairs. IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), June 2014.
Dan Kong and Hai Tao. A method for learning matching errors for stereo computation.
British Machine Vision Conference (BMVC), 2004.
Dan Kong and Hai Tao. Stereo matching via learning multiple experts behaviors. British
Machine Vision Conference (BMVC), 2006.
29
Žbontar and LeCun
Jana Kostková and Radim Sára. Stratified dense matching for stereopsis in complex scenes.
British Machine Vision Conference (BMVC), 2003.
Jedrzej Kowalczuk, Eric T Psota, and Lance C Perez. Real-time stereo matching on CUDA
using an iterative refinement method for adaptive support-weight correspondences. IEEE
Transactions on Circuits and Systems for Video Technology, 23(1):94–104, 2013.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Yunpeng Li and Daniel P Huttenlocher. Learning for stereo vision using the structured
support vector machine. IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), June 2008.
Xing Mei, Xun Sun, Mingcai Zhou, Haitao Wang, Xiaopeng Zhang, et al. On building an
accurate stereo matching system on graphics hardware. IEEE International Conference
on Computer Vision Workshops (ICCV Workshops), pages 467–474, 2011.
Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), June 2015.
John Nickolls, Ian Buck, Michael Garland, and Kevin Skadron. Scalable parallel programming with CUDA. Queue, 6(2):40–53, 2008.
Mattis Paulin, Matthijs Douze, Zaid Harchaoui, Julien Mairal, Florent Perronin, and
Cordelia Schmid. Local convolutional features with unsupervised training for image retrieval. In IEEE International Conference on Computer Vision (ICCV), pages 91–99,
2015.
Martin Peris, Atsuto Maki, Sara Martull, Yasuhiro Ohkawa, and Kazuhiro Fukui. Towards
a simulation driven stereo vision system. In 21st International Conference on Pattern
Recognition (ICPR), pages 1038–1042, 2012.
Eric T Psota, Jedrzej Kowalczuk, Mateusz Mittek, and Lance C Perez. Map disparity
estimation using hidden markov trees. IEEE International Conference on Computer
Vision (ICCV), 2015.
Jerome Revaud, Philippe Weinzaepfel, Zaid Harchaoui, and Cordelia Schmid. Deepmatching: Hierarchical deformable dense matching. ArXiv e-prints, 1(7):8, 2015.
Daniel Scharstein and Chris Pal. Learning conditional random fields for stereo. IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), June 2007.
Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame
stereo correspondence algorithms. International Journal of Computer Vision, 47(1-3):
7–42, 2002.
Daniel Scharstein and Richard Szeliski. High-accuracy stereo depth maps using structured
light. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June
2003.
30
Stereo by Training a Network to Compare Patches
Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nešić,
Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate
ground truth. German Conference on Pattern Recognition (GCPR), September 2014.
Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Learning local feature descriptors using convex optimisation. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 36(8):1573–1585, 2014.
Sudipta N Sinha, Daniel Scharstein, and Richard Szeliski. Efficient high-resolution stereo
matching using local plane sweeps. IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), June 2014.
Aristotle Spyropoulos, Nikos Komodakis, and Philippos Mordohai. Learning to detect
ground control points for improving the accuracy of stereo matching. IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), June 2014.
Deqing Sun, Stefan Roth, and Michael J Black. A quantitative analysis of current practices in optical flow estimation and the principles behind them. International Journal of
Computer Vision, 106(2):115–137, 2014.
Tomasz Trzcinski, Mario Christoudias, Vincent Lepetit, and Pascal Fua. Learning image
descriptors with the boosting-trick. In Advances in neural information processing systems,
pages 269–277, 2012.
Christoph Vogel, Konrad Schindler, and Stefan Roth. Piecewise rigid scene flow. IEEE
International Conference on Computer Vision (ICCV), 2013.
Christoph Vogel, Stefan Roth, and Konrad Schindler. View-consistent 3D scene flow estimation over multiple frames. European Conference on Computer Vision (ECCV), September
2014.
Christoph Vogel, Konrad Schindler, and Stefan Roth. 3D scene flow estimation with a
piecewise rigid scene model. International Journal of Computer Vision, pages 1–28,
2015.
Koichiro Yamaguchi, David McAllester, and Raquel Urtasun. Efficient joint segmentation,
occlusion labeling, stereo and flow estimation. European Conference on Computer Vision
(ECCV), September 2014.
Ramin Zabih and John Woodfill. Non-parametric local transforms for computing visual
correspondence. European Conference on Computer Vision (ECCV), 1994.
Sergey Zagoruyko and Nikos Komodakis. Learning to compare image patches via convolutional neural networks. IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), June 2015.
Jure Žbontar and Yann LeCun. Computing the stereo matching cost with a convolutional
neural network. IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
June 2015.
31
Žbontar and LeCun
Chi Zhang, Zhiwei Li, Yanhua Cheng, Rui Cai, Hongyang Chao, and Yong Rui. Meshstereo:
A global stereo model with mesh alignment regularization for view interpolation. IEEE
International Conference on Computer Vision (ICCV), 2015.
Ke Zhang, Jiangbo Lu, and Gauthier Lafruit. Cross-based local stereo matching using
orthogonal integral images. IEEE Transactions on Circuits and Systems for Video Technology, 19(7):1073–1079, 2009.
Li Zhang and Steven M Seitz. Estimating optimal parameters for MRF stereo from a single
image pair. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(2):
331–342, 2007.
32
| 9cs.NE
|
arXiv:1706.09776v1 [math.NA] 29 Jun 2017
Numerical assessment of two-level domain
decomposition preconditioners for incompressible
Stokes and elasticity equations
Gabriel R. Barrenechea1 , Michal Bosy
1
∗1
, and Victorita Dolean1
Department of Mathematics and Statistics, University of Strathclyde, 26 Richmond Street, G1 1XH
Glasgow, United Kingdom
June 30, 2017
Abstract
Solving the linear elasticity and Stokes equations by an optimal domain decomposition
method derived algebraically involves the use of non standard interface conditions. The onelevel domain decomposition preconditioners are based on the solution of local problems. This
has the undesired consequence that the results are not scalable, it means that the number of
iterations needed to reach convergence increases with the number of subdomains. This is the
reason why in this work we introduce, and test numerically, two-level preconditioners. Such
preconditioners use a coarse space in their construction. We consider the nearly incompressible
elasticity problems and Stokes equations, and discretise them by using two finite element
methods, namely, the hybrid discontinuous Galerkin and Taylor-Hood discretisations.
Key words. Stokes problem, nearly incompressible elasticity, Taylor-Hood, hybrid discontinuous Galerkin methods, domain decomposition, coarse space, optimized restricted additive Schwarz
methods
1
Introduction
In [BBD+ 16] the one-level domain decomposition methods for Stokes equations were introduced
in conjunction with the non standard interface conditions. Although it can be observed there
the lack of scalability with respect to the number of subdomains. It means that by splitting the
problem in a larger number of subdomains leads to the increase of size of the plateau region in the
convergence of an iterative method (see Figure 1) when using the one-level domain decomposition
methods. This is caused by the lack of global information, as subdomains can only communicate
with their neighbours. Hence, when the number of subdomains increases in one direction, the
length of the plateau also increases. Even in cases when the local problems are of the same size,
the iteration count grows with the increase of the number of subdomain. This can be also observed
in all experiments in this manuscript in case of one-level methods.
The remedy for this is the use of a second level in the preconditioner or a coarse space correction
that adds the necessary global information. Two-level algorithms have been analysed for several
classes of problems in [TW05]. The key point of these kind of methods is to choose the appropriate
coarse space. The classical coarse space introduced by Nicolaides in [Nic87] is defined by vectors
that support is in each subdomain. Hence the coarse space has the size equal to a number of
∗ Corresponding
author: E-mail: [email protected]
1
0
-1
-1
-2
-2
log(error)
log(error)
0
-3
-4
-3
-4
-5
-5
-6
-6
-7
-7
0
200
400
600
800
1000
1200
0
50
100
150
200
250
Iterations
Iterations
(a) Taylor-Hood
(b) hdG
300
350
400
450
Figure 1: Increase of the size of the plateau region for increasing number of subdomains.
subdomains. A coarse space construction, named spectral coarse space, was motivated by the
complexity of the problems that classical coarse space performance was not satisfying. This
construction allows to enrich a bigger size of the coarse space, but can be also reduced to the
classical one. This idea was introduced for the first time in [BHMV99] in the case of multigrid
methods. It relies on solving local generalised eigenvalue problems allowing to choose suitable
vectors for the coarse space.
For overlapping domain decomposition preconditioners, a similar idea was introduced in the
case of Darcy equations in [GE10a, GE10b]. The authors of [NXD10] consider also the heterogeneous Darcy equation and presented a different generalised eigenvalue problem based on local
Dirichlet-to-Neumann maps. The method has been analysed in [DNSS12] and proved to be very
robust in the case of small overlaps. The same idea was extended numerically to the heterogeneous
Helmholtz problem in [CDKN14]. The authors of [LNS15] apply the coarse space associated with
low-frequency eigenfunctions of the subdomain Dirichlet-to-Neumann maps for the generalisation
of the optimised Schwarz methods, named 2-Lagrange multiplier methods.
The first attempt to extend this spectral approach to general symmetric positive definite problems was made in [EGLW12] as an extension of [GE10a, GE10b]. Since some of the assumptions of
the previous framework are hard to fulfil, authors of [SDH+ 14] proposed slightly different approach
for symmetric positive definite problems. Their idea of constructing partition of unity operator
associated with degrees of freedom allows to work with various finite element spaces. An overview
of different kinds of two-level methods can be found in [DJN15, Chapters 5 and 7].
Despite the fact that all these approaches provide satisfying results, there is no universal treatment to build efficient coarse spaces in the case of non definite problems such as Stokes equations.
The spectral coarse spaces that we use in this work are inspired by those proposed in [HJN15].
The authors introduced and tested numerically symmetrised two-level preconditioners for overlapping algorithms which use Robin interface conditions between the subdomains (see (5.27) for
details). They have applied these preconditioners to solve saddle point problems such as nearly
incompressible elasticity and Stokes discretised by Taylor-Hood finite elements. In our case, we
use non standard interface conditions. Therefore the use of spectral coarse spaces could lead to
an important gain.
In this work, we test this improvement in case of nearly incompressible elasticity and Stokes
equations that are discussed in Section 2. As the discretisations we use the Taylor-Hood [GR86,
Chapter II, Section 4.2] and hybrid discontinuous Galerkin method [CGL09, CG09] presented in
Section 3. In Section 4 we introduce the two-level domain decomposition preconditioners. Sections 5 and 6 present the two and three dimensional numerical experiments, respectively. Finally,
a summary is outlined in Section 7.
2
2
The differential equations
Let Ω be an open polygon in R2 or an open Lipschitz polyhedron in R3 , with Lipschitz boundary
Γ := ∂Ω. We use d = 2, 3 to denote the dimension of the space. We use bold for tensor or vector
variables. In addition we denote normal and tangential components as follows un := u · n, and
ut := u − un n, where n is the outward unit normal vector to the boundary Γ .
For D ⊂ Ω, we use the standard L2 (D) space and C 0 (D̄) denotes the set of all continuous
functions on the closure of a set D. Let us define following Sobolev spaces
H m (D) := v ∈ L2 (D) : ∀ |α| ≤ m ∂ α v ∈ L2 (D) for m ∈ N,
1
H 2 (∂D) := ṽ ∈ L2 (∂D) : ∃v ∈ H 1 (D) ṽ = tr(v) ,
H (div, D) := v ∈ [L2 (D)]d : ∇ · v ∈ L2 (D) ,
where, for α = (α1 , ..., αd ) ∈ Nd and |α| =
1
Pd
i=1
αi we denote ∂ α =
∂ |α|
α ,
α
∂x1 1 ...∂xd d
and tr : H 1 (Ω) →
H 2 (∂Ω) is the trace operator. In addition, we use the following notation of the space including
boundary and average conditions
Z
L20 (D) := v ∈ L2 (D) :
v dx = 0 ,
D
n
o
1
1
HΓ̃ (D) := v ∈ H (D) : v = 0 on Γ̃ ,
where Γ̃ ⊂ ∂D. If Γ̃ = ∂D, then HΓ̃1 (D) is denoted H01 (D).
Now we present the two differential problems considered in this work.
2.1
Stokes equation
Let us start with d-dimensional, d = 2, 3, Stokes problem
−ν∆u + ∇p
= f
(2.1)
∇·u = 0
in Ω
,
in Ω
where u : Ω̄ → Rd is the velocity field, p : Ω̄ → R the pressure, ν > 0 the viscosity which
is considered to be constant and f ∈ [L2 (Ω)]d is a given function. We define the stress tensor
1
σ := ν∇u − pI and the flux as σn := σ n. For uD ∈ [H 2 (Γ)]d and g ∈ L2 (Γ) we consider three
types of boundary conditions:
• Dirichlet (non-slip)
(2.2)
u = uD on Γ;
• tangential-velocity and normal-flux (TVNF)
ut =
(2.3)
σnn =
0
g
on Γ
;
on Γ
• normal-velocity and tangential-flux (NVTF)
un = 0
(2.4)
σnt = g
on Γ
.
on Γ
The third type of boundary condition has already been considered for the Stokes problem in [AdDBM+ 14].
3
2.2
Nearly incompressible elasticity equation
From a mathematical point of view, the nearly incompressible elasticity problem is very similar to
the Stokes equations. The difference
is that instead of considering the gradient ∇v, the symmetric
gradient ε(v) := 21 ∇v + ∇T v is used. We want to solve the following d-dimensional, d = 2, 3,
problem
−2µ∇ · ε(u) + ∇p
= f
in Ω
(2.5)
− ∇ · u = λ1 p in Ω
where u : Ω̄ → Rd is the displacement field, p : Ω → R the pressure, f ∈ [L2 (Ω)]d is a given
function, λ and µ are the Lamé coefficients defined by
λ=
Eν
,
(1 + ν)(1 − 2ν)
µ=
E
,
2(1 + ν)
where E is the Young modulus and ν the Poisson ratio. We define the stress tensor as σ sym :=
2µε(u) − pI and its normal component as σ sym
:= σ sym n. For g ∈ L2 (Γ) we consider three types
n
of boundary conditions:
• mixed such that Γ = ΓD ∪ ΓN and ΓD ∩ ΓN = ∅
u = 0
(2.6)
sym
σn
= 0
on ΓD
;
on ΓN
• tangential-displacement and normal-normal-stress (TDNNS)
ut = 0 on Γ
(2.7)
;
sym
σnn
= g on Γ
• normal-displacement and tangential-normal-stress (NDTNS)
un = 0 on Γ
(2.8)
.
sym
σnt
= g on Γ
The second type of boundary condition has already been considered for linear elasticity equation
in [PS11].
3
The numerical methods
Let {Th }h>0 be a regular family of triangulations of Ω̄ made of simplices. For each triangulation
Th , Eh denotes the set of its facets (edges for d = 2, faces for d = 3). In addition, for each element
K ∈ Th , hK := diam(K), and we denote h := maxK∈Th hK . We define the following broken
Sobolev spaces on the set of all edges in Eh (for d = 2)
L2 (Eh ) := v : v|E ∈ L2 (E) ∀ E ∈ Eh .
Moreover, for D ⊂ Ω, Pk (D) denotes the space of polynomials of total degree smaller than, or
equal to, k on the set D.
We now present the two discretisations to be used in the numerical experiments.
3.1
Taylor-Hood discretisation
We first consider the Taylor-Hood discretisation using the following approximation spaces
T Hhk := vh ∈ [H 1 (Ω)]d : vh |K ∈ [Pk (K)]d ∀ K ∈ Th ,
Rhk−1 := qh ∈ C 0 (Ω̄) : qh |K ∈ Pk−1 (K) ∀ K ∈ Th
4
where k ≥ 2 (see [GR86, Chapter II, Section 4.2]).
If (2.1) is supplied with the homogeneous boundary conditions (2.2), then the discrete problem
reads:
Find (uh , ph ) ∈ T Hhk ∩ [H01 (Ω)]d × Rhk−1 ∩ L20 (Ω)
s.t. for all (vh , qh ) ∈ T Hhk ∩ [H01 (Ω)]d × Rhk−1 ∩ L20 (Ω)
Z
ν∇uh : ∇vh dx
(3.9)
−
Ω
−
Z
ph ∇ · vh dx
ZΩ
Z
=
f vh dx
Ω
∇ · uh qh dx
=
0.
Ω
In case of TVNF boundary conditions (2.3), we define Vt := v ∈ [H 1 (Ω)]d : vt = 0 on Γ , and
the discrete problem reads:
Find (uh , ph ) ∈ T Hhk ∩ Vt × Rhk−1
s.t. for all (vh , qh ) ∈ T Hhk ∩ Vt × Rhk−1
(3.10)
Z
ν∇uh : ∇vh dx
−
Ω
−
Z
ZΩ
ph ∇ · vh dx
=
Z
f vh dx +
Ω
∇ · uh qh dx
=
Z
Γ
g (vh )t ds
0.
Ω
If NVTF
boundary conditions (2.4) are used, then we define the following space
Vn := v ∈ [H 1 (Ω)]d : vn = 0 on Γ , and the discrete problem reads:
Find (uh , ph ) ∈ T Hhk ∩ Vn × Rhk−1 ∩ L20 (Ω)
s.t. for all (vh , qh ) ∈ T Hhk ∩ Vn × Rhk−1 ∩ L20 (Ω)
(3.11)
Z
ν∇uh : ∇vh dx
−
Ω
−
Z
ZΩ
ph ∇ · vh dx
=
Z
f vh dx +
Γ
Ω
∇ · uh qh dx
=
Z
g (vh )n ds
0.
Ω
In similar way, if the problem (2.5) is supplied with the boundary conditions (2.6), then the
discrete problem reads
Find (uh , ph ) ∈ T Hhk ∩ [HΓ1D (Ω)]d × Rhk−1
s.t. for all (vh , qh ) ∈ T Hhk ∩ [HΓ1D (Ω)]d × Rhk−1
(3.12)
Z
2µε (uh ) : ε (vh ) dx
Z
Ω
∇ · uh qh dx
−
Ω
−
−
Z
ph ∇ · vh dx
ΩZ
1
ph qh dx
λ
=
Z
f vh dx
Ω
=
0.
Ω
The rest of the discrete problems associated with (2.5) that is supplied with TDNNS boundary
conditions (2.7) or NDTNS boundary conditions (2.8) are similar to (3.10) or (3.11), respectively.
5
3.2
Hybrid discontinuous Galerkin discretisation
We restrict the discussion of this to two dimensional case d = 2. This method has been presented
and analysed in [BBD+ 16]. The velocity is approximated using the Brezzi-Douglas-Marini spaces
(see [BBF13, Section 2.3.1]) of degree k given by
o
n
BDMhk := vh ∈ H (div, Ω) : vh |K ∈ [Pk (K)]2 ∀ K ∈ Th ,
n
o
2
k
:=
v
∈
H
(div,
Ω)
:
v
|
∈
[P
(K)]
∀
K
∈
T
∧
(v
)
=
0
on
Γ̃
,
BDMh,
h
h
K
k
h
h
n
Γ̃
k
k
where Γ̃ ⊂ ∂Ω. If Γ̃ = ∂Ω, then BDMh,
is denoted BDMh,0
.
Γ̃
The pressure is approximated in the space
Qhk−1 := qh ∈ L2 (Ω) : qh |K ∈ Pk−1 (K) ∀ K ∈ Th .
Finally, a Lagrange multiplier, aimed at approximating the tangential component of the velocity
is introduced. The space swhere this multipliers is sought are given by
Mhk−1 := ṽh ∈ L2 (Eh ) : ṽh |E ∈ Pk−1 (E) ∀ E ∈ Eh ,
n
o
k−1
k−1
Mh,
:=
ṽ
∈
M
:
ṽ
=
0
on
Γ̃
,
h
h
h
Γ̃
k−1
k−1
where Γ̃ ⊂ ∂Ω. If Γ̃ = ∂Ω, then Mh,
is denoted Mh,0
. Furthermore, we introduce for all E ∈ Eh
Γ̃
k−1
the L2 (E)-projection ΦE
: L2 (E) → Pk−1 (E) defined by
Z
Z
k−1
(w̃)ṽh ds =
w̃ṽh ds ∀ ṽh ∈ Pk−1 (E) ,
(3.13)
ΦE
E
E
k−1
for all E ∈ Eh .
and we denote Φk−1 : L2 (Eh ) → Mhk−1 defined as Φk−1 |E := ΦE
If (2.1) is supplied with the homogeneous boundary conditions (2.2), then the discrete problem
reads:
k−1
k
Find (uh , ũh , ph ) ∈ BDMh,0
× Mh,0
× Qhk−1 ∩ L20 (Ω)
k−1
k
s.t. for all (vh , ṽh , qh ) ∈ BDMh,0
× Mh,0
× Qhk−1 ∩ L20 (Ω) ,
(3.14)
where
a ((uh , ũh ) , (vh , ṽh )) +
=
f vh dx
Ω
b ((uh , ũh ) , qh ) =
a ((wh , w̃h ) , (vh , ṽh )) :=
X Z
K∈Th
−
(3.15)
b ((vh , ṽh ) , ph )
Z
Z
Z∂K
ν∇wh : ∇vh dx
K
ν (∂n wh )t (vh )t − ṽh ds
ν (wh )t − w̃h (∂n vh )t ds
∂K
Z
τ
+ν
Φk−1 (wh )t − w̃h Φk−1 (vh )t − ṽh ds ,
hK ∂K
−
and τ > 0 is a stabilisation parameter, and
(3.16)
,
0
b ((vh , ṽh ) , qh ) := −
X Z
K∈Th
qh ∇ · vh dx.
K
If TVNF boundary conditions (2.3) are used, then the discrete problem reads:
6
k−1
Find (uh , ũh , ph ) ∈ BDMhk × Mh,0
× Qhk−1
k−1
s.t. for all (vh , ṽh , qh ) ∈ BDMhk × Mh,0
× Qhk−1 ,
(3.17)
a ((uh , ũh ) , (vh , ṽh )) +
b ((vh , ṽh ) , ph )
Z
=
f vh dx +
Ω
b ((uh , ũh ) , qh )
Z
Γ
= 0
g (vh )n ds
.
In case of NVTF boundary conditions (2.4), the discrete problem reads:
k
Find (uh , ũh , ph ) ∈ BDMh,0
× Mhk−1 × Qhk−1 ∩ L20 (Ω)
k
s.t. for all (vh , ṽh , qh ) ∈ BDMh,0
× Mhk−1 × Qhk−1 ∩ L20 (Ω) ,
(3.18)
a ((uh , ũh ) , (vh , ṽh )) +
b ((vh , ṽh ) , ph )
=
Z
f vh dx +
Ω
b ((uh , ũh ) , qh ) =
Z
Γ
0
gṽh ds
.
In similar way, if the problem (2.5) is supplied with the mixed boundary conditions (2.6), then
the discrete problem reads:
k−1
k
× Qhk−1
× Mh,Γ
Find (uh , ũh , ph ) ∈ BDMh,Γ
D
D
k−1
k
s.t. for all (vh , ṽh , qh ) ∈ BDMh,ΓD × Mh,ΓD × Qhk−1
(3.19)
where
as ((uh , ũh ) , (vh , ṽh )) +
Z
f vh dx
Ω
b ((uh , ũh ) , qh ) +
as ((wh , w̃h ) , (vh , ṽh )) :=
X Z
K∈Th
−
(3.20)
b ((vh , ṽh ) , ph ) =
Z
Z∂K
c(ph , qh )
=
0,
2µε(wh ) : ε(vh ) dx
K
2µ (εn (wh ))t (vh )t − ṽh ds
2µ (wh )t − w̃h (εn (vh ))t ds
∂K
Z
τ
+2µ
Φk−1 (wh )t − w̃h Φk−1 (vh )t − ṽh ds ,
hK ∂K
−
b is defined by (3.16), and
c(rh , qh ) := −
1
λ
Z
rh qh ds.
Ω
The rest of the discrete problems associated with (2.5) that is supplied with TDNNS boundary
conditions (2.7) or NDTNS boundary conditions (2.8) are similar to (3.17) or (3.18), respectively.
4
The domain decomposition preconditioners
Let us assume that we have to solve the following linear system AU = F where A is the matrix
arising from discretisation of the Stokes or linear elasticity equation on the domain Ω, U is the
vector of unknowns, and F is the right hand side. To accelerate the performance of an iterative
Krylov method [DJN15, Chapter 3] applied to this system we will consider domain decomposition
preconditioners which are naturally parallel. They are based on an overlapping decomposition of
the computational domain.
7
(a) L-shaped domain
(b) T-shaped domain
Figure 2: Partition of the domain for 8 subdomains by METIS.
Let {Th,i }N
i=1 be a partition of the triangulation Th (see examples in Figure 2). For an integer
l N
l
value l ≥ 0, we define an overlapping decomposition {Th,i
}i=1 such that Th,i
is a set of all triangles
l−1
l−1
l−1
from Th,i and all triangles from Th \ Th,i that have non-empty intersection with Th,i
, and
0
Th,i = Th,i . With this definition the width of the overlap will be of 2l. Furthermore, if Wh stands
l
l
for the finite element space associated to Th , Wh,i
is the local finite element spaces on Th,i
, which
is a triangulation of Ωi .
Let N be the set of indices of degrees of freedom of Wh and Nil the set of indices of degrees
l
l
of freedom of Wh,i
for l ≥ 0. Moreover, we define the restriction operator Ri : Wh → Wh,i
as a
l
rectangular matrix |Ni | × |N | such that if V is the vector of degrees of freedom of vh ∈ Wh , then
l
l
Ri V is the vector of degrees of freedom of Wh,i
in Ωi . The extension operator from Wh,i
to Wh
T
and its associated matrix are both given by Ri . In addition we introduce a partition of unity Di
as a diagonal matrix |Nil | × |Nil | such that
(4.21)
Id =
N
X
RTi Di Ri ,
i=1
where Id ∈ R|N |×|N | is the identity matrix.
We first recall the Modified Restricted Additive Schwarz (MRAS) preconditioner introduced
in [BBD+ 16] for the Stokes equation. This preconditioner is given by
(4.22)
M−1
MRAS =
N
X
RTi Di B−1
i Ri ,
i=1
where Bi is the matrix associated to a discretisation of Stokes equation (2.1) in Ωi where we
impose either TVNF (2.3) or NVTF (2.4) boundary conditions in Ωi . In case of a discretisation
of elasticity equation (2.5) in Ωi , we impose either TDNNS (2.7) or NDTNS (2.8) boundary
conditions in Ωi .
We new introduce a symmetrised variant of (4.22) called Symmetrised Modified Restricted
Additive Schwarz (SMRAS), given by
(4.23)
M−1
SMRAS =
N
X
RTi Di B−1
i Di Ri .
i=1
8
4.1
Two-level methods
A two-level version of the SMRAS and MRAS preconditioners will be based on a spectral coarse
space obtained by solving the following local generalised eigenvalue problems
Find (V jk , λjk ) ∈ R|Nj | \ {0} × R s.t.
(4.24)
Ãj V jk = λjk Bj V jk ,
where Ãj are local matrices associated to a discretisation of local Neumann boundary value prob|N |
lem in Ωj . Let θ > 0 be a user-defined threshold.
as the vector space
We define ZGenEO ⊂ R
T
spanned by the family of vectors Rj Dj V jk λ <θ , 1 ≤ j ≤ N , corresponding to eigenvalues
jk
smaller than θ. The value of θ is chosen such that for a given problem and preconditioner, the
behaviour of the method should be robust in the sense that, its convergence should not depend,
or depends very weakly, on the number of subdomains.
We are now ready to introduce the two-level method with coarse space ZGenEO . Let P0 be
the A-orthogonal projection onto the coarse space ZGenEO . The two-level SMRAS preconditioner
is defined as
(4.25)
−1
T
M−1
+ (Id − P0 )M−1
SMRAS,2 = P0 A
SMRAS (Id − P0 ).
Furthermore, if R0 is a matrix whose rows are a basis of the coarse space ZGenEO , then
−1
P0 A−1 = RT0 R0 ART0
R0 .
In similar way, we can introduce the two-level MRAS preconditioner
(4.26)
5
−1
T
M−1
+ (Id − P0 )M−1
MRAS,2 = P0 A
MRAS (Id − P0 ).
Numerical results for two dimensional problems
In this section we assess the performance of the preconditioners defined in Section 4.1. We will
compare the newly introduced ones with that of ORAS and SORAS introduced in [HJN15]. These
kind of preconditioners are associated with the Robin interface conditions and require an optimised parameter as it can be seen in (5.27) below. The big advantage of SMRAS and MRAS
preconditioners from the previous section is that they are parameter-free. We consider the partial
differential equation model for nearly incompressible elasticity and Stokes flow as problems of similar mixed formulation. Each of these problems is discretised by using the Taylor-Hood methods
from Section 3.1 and the hdG discretisation from Section 3.2.
Our experiments will be based on the classical weak scaling test. This test is built as follows.
A domain Ω̄ is split into a triangulation Th . For each of element K ∈ Th , hK = diam(K), and
we denote the mesh size by h := maxK∈Th hK . Then, this triangulation is split into overlapping
subdomains of size H, in such a way H
h remains constant. In the absence of a second level in the
preconditioner, if the number of subdomains grows, then the convergence gets slower. A coarse
space provides a global information and leads to a more robust behaviour.
The simplest way to build a coarse space is to consider the zero energy modes. More precisely
they are the eigenvectors associated with the zero eigenvalues of (4.24) on a floating subdomains.
Hence, by a floating subdomain we mean a subdomain without Dirichlet boundary condition on
any part of the boundary. Then the matrix on the left hand side of (4.24) is singular and there are
zero eigenvalues. These zero energy modes are the rigid body motions (three in two dimensions,
six in three dimensions) for the elasticity problem, and the constants (two in two dimensions, three
in three dimensions) for the Stokes equations. Unfortunately for some cases, this choice is not
sufficient, so we have collected the smallest M eigenvalues for each subdomain and build a coarse
9
space by including the eigenvectors associated to them. The different values of M are presented
in the table in brackets.
All experiments have been made by using FreeFem++ [Hec12], which is a free software specialised in variational discretisations of partial differential equations. We use GMRES [SS86] as
an iterative solver. Generalized eigenvalue problems to generate the coarse space are solved using
ARPACK [LSY98]. The overlapping decomposition into subdomains can be uniform (Unif) or
generated by METIS (MTS) [KK98]. In each of the examples in this section we consider decomposition with two layers of mesh size h in the overlap. Tables show the number of iterations
kU −U k
needed to achieve a relative l2 norm of the error smaller than 10−6 , kU −U n0 k l22 < 10−6 , where
l
U is the solution of global problem given by direct solver and U m denotes m-th iteration of the
iterative solver. In addition, DOF stands for number of degrees of freedom and N for the number
of subdomains in all tables.
5.1
Taylor-Hood discretisation
In this section we consider the Taylor-Hood discretisation from Section 3.1 with different values
of k ≥ 2 for nearly incompressible elasticity and Stokes equations.
5.1.1
Nearly incompressible elasticity
Since we consider the preconditioners with various interface conditions we need to comment the
way of imposing them. ORAS and SORAS preconditioners follow [HJN15] and use Robin interface
conditions. This means, the weak formulation of the linear elasticity problem contains the following
term
Z
Z
µ(2µ + λ)
sym
uh vh ds
(5.27)
σ n (vh )n ds +
2α
λ + 3µ
∂Ωi \Γ
∂Ωi \Γ
where again, following [HJN15] we choose α = 10. Fortunately, the MRAS and SMRAS preconditioners are parameter-free. For all numerical experiments associated in this section we use zero
as an initial guess for the GMRES iterative solver. Moreover, the overlapping decomposition into
subdomains is generated by METIS.
ΓD
ΓN
(a) Domain and boundary
(b) Discrete solution
Figure 3: L-shaped domain problem.
Test case 1 (The L-shaped domain problem). We consider the L-shaped domain Ω = (−1, 1)2 \
{(0, 1) × (−1, 0)} clamped on the left side and partly from a top and bottom as it is depicted in
10
Figure 3a. This example is similar to the one in [CS15]. The associated boundary value problem
is
−2µ∇ · ε(u) + ∇p
= (0, −1)T
in Ω
− ∇·u
= λ1 p
in Ω
.
(5.28)
u(x, y)
= (0, 0)T
on ΓD
sym
T
σ n (x, y) = (0, 0)
on ΓN
The physical parameters are E = 105 and ν = 0.4999 (nearly incompressible). In Figure 3b we
plot the mesh of the bent domain.
We choose k = 3 for the Taylor-Hood discretisation. In Figure 4 we plot the eigenvalues of
one floating subdomain. The clustering of small eigenvalues of the generalised eigenvalue problem
defined in (4.24) suggests the number of eigenvectors to be added to the coarse space. The three
zero eigenvalues correspond to the zero energy modes.
Table 1: Comparison of preconditioners for Taylor-Hood discretisation (T Hh3 , Rh2 ) - the L-shaped
domain problem.
DOF
124 109
478 027
933 087
1 899 125
3 750 823
N
4
16
32
64
128
ORAS
26
57
84
130
209
SORAS
60
131
180
293
412
DOF
124 109
478 027
933 087
1 899 125
3 750 823
N
4
16
32
64
128
ORAS
18
37
49
65
83
SORAS
40
52
57
64
64
DOF
124 109
478 027
933 087
1 899 125
3 750 823
N
4
16
32
64
128
ORAS
15
31
40
49
69
SORAS
32
41
48
51
54
DOF
124 109
478 027
933 087
1 899 125
3 750 823
N
4
16
32
64
128
ORAS
14
26
31
39
58
SORAS
33
41
43
47
49
One-level
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
26
60
30
69
143
65
109
221
104
181
362
161
302
568
251
Two-level (3 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
19
36
24
40
57
46
56
67
53
70
75
61
74
77
75
Two-level (5 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
17
35
24
31
47
42
38
52
53
45
53
64
49
54
70
Two-level (7 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
16
30
24
25
38
42
25
42
49
30
39
59
30
43
61
TDNNS-SMRAS
59
140
211
312
510
TDNNS-SMRAS
41
56
66
74
72
TDNNS-SMRAS
37
47
51
56
53
TDNNS-SMRAS
35
44
46
50
50
The results of Table 1 show a clear improvement in the scalability of the two-level preconditioners over the one-level ones. In fact, using five eigenvectors per subdomain, the number of
iterations is virtually unaffected by the number of subdomains. All two-level preconditioners show
a comparable performance. For this case, increasing the dimension of the coarse space beyond
5 × N eigenvectors does not seem to improve the results dramatically.
Test case 2 (The heterogeneous beam problem). We consider a heterogeneous beam with ten
layers of steel and rubber. Five layers are made from steel with the physical parameters E =
210 · 109 and ν = 0.3, and other five are made from rubber with the physical parameters E = 108
and ν = 0.4999 as it is depicted in Figure 5a. A similar example was considered in [HJN15]. The
computational domain is the rectangle Ω = (0, 5) × (0, 1). The beam is clamped on its left side,
11
0.3
0.25
0.2
0.2
0.15
0.1
Eigenvalue
0.2
Eigenvalue
Eigenvalue
0.25
0.15
0.1
0.15
0.1
0.05
0.05
0.05
0
0
1
2
3
4
5
6
7
8
9
10
0
1
2
Number of eigenvalue
3
4
5
6
7
8
9
10
1
2
Number of eigenvalue
(a) Robin interface conditions
3
4
5
6
7
8
9
10
Number of eigenvalue
(b) NDTNS interface conditions
(c) TDNNS interface conditions
Figure 4: Eigenvalues on one of the floating subdomains in case of uniform decomposition and
Taylor-Hood discretisation (T Hh3 , Rh2 ) - the L-shaped domain problem.
(a) Steel and rubber layers
(b) Discrete solution
Figure 5: Heterogeneous beam.
hence we consider the following problem
−2µ∇ · ε(u) + ∇p
− ∇·u
(5.29)
u(x, y)
σ sym
n (x, y)
= (0, −1)T
= λ1 p
= (0, 0)T
= (0, 0)T
in Ω
in Ω
.
on ∂Ω ∩ {x = 0}
on ∂Ω \ {x = 0}
In Figure 5b we plot the mesh of the bent beam. Because of the heterogeneity of the problem,
we do not notice a clear clustering of the eigenvalues (see Figure 6). In such case it is well known
that coarse space including only three zero energy modes is not sufficient [DNSS12]. That is why
we consider a coarse space built using 5 or 7 eigenvectors per subdomain.
10 -3
0.014
7
0.03
0.012
6
0.025
0.01
4
3
Eigenvalue
Eigenvalue
Eigenvalue
5
0.02
0.015
0.008
0.006
2
0.01
0.004
1
0.005
0.002
0
0
1
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(a) Robin interface conditions
0
1
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(b) NDTNS interface conditions
1
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(c) TDNNS interface conditions
Figure 6: Eigenvalues on one of the floating subdomains in case of METIS decomposition and
Taylor-Hood discretisation (T Hh3 , Rh2 ) - the heterogeneous beam.
As in the previous example, the introduction of a coarse space provides a significant improvement in the number of iterations needed for convergence. Due to the high heterogeneity of this
problem, more eigenvectors per subdomain are needed to obtain scalable results. We notice an
important improvement of the convergence when using two-level methods (see Table 2). Although
we get a stable number of iterations only when considering a coarse space which is sufficiently big.
12
Table 2: Comparison of preconditioners for Taylor-Hood discretisation (T Hh3 , Rh2 ) - the heterogeneous beam.
DOF
44 963
87 587
177 923
347 651
707 843
1 385 219
N
8
16
32
64
128
256
ORAS
168
226
373
615
973
>1000
SORAS
301
490
711
>1000
>1000
>1000
DOF
44 963
87 587
177 923
347 651
707 843
1 385 219
N
8
16
32
64
128
256
ORAS
109
136
193
260
412
379
SORAS
160
204
291
304
356
414
DOF
44 963
87 587
177 923
347 651
707 843
1 385 219
N
8
16
32
64
128
256
ORAS
76
106
157
178
140
119
SORAS
118
146
202
191
114
86
5.1.2
One-level
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
160
267
177
245
462
229
447
684
440
728
>1000
746
>1000
>1000
>1000
>1000
>1000
>1000
Two-level (5 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
136
147
148
192
200
181
296
275
326
363
282
491
420
369
601
448
400
711
Two-level (7 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
124
115
133
166
138
159
203
185
302
225
170
326
153
112
266
118
77
259
TDNNS-SMRAS
264
424
672
>1000
>1000
>1000
TDNNS-SMRAS
136
184
276
299
346
317
TDNNS-SMRAS
103
123
214
182
122
94
Stokes equation
We now turn to the Stokes discrete problem given in Sections 3.1. Once again in case of ORAS
and SORAS we choose α = 10 as in [HJN15] for the Robin interface conditions (5.27). In the first
case we consider a random initial guess for the GMRES iterative solver. Later with the second
example we use zero as an initial guess.
(a) Velocity field
(b) Pressure
Figure 7: Numerical solution of the driven cavity problem - the driven cavity problem.
Test case 3 (The driven cavity problem). The test case is the driven cavity. We consider the
13
following problem on the unit square Ω = (0, 1)2
= f
−∆u + ∇p
− ∇·u
= 0
(5.30)
u(x, y) = (1, 0)T
u(x, y) = (0, 0)T
in Ω
in Ω
.
on ∂Ω ∩ {y = 1}
on ∂Ω \ {y = 1}
In Figure 7 we plot the vector field and pressure, after solving numerically the problem.
We start with the two energy modes only (see Figure 8). This already provides some improvement. Then, we add more eigenvectors to see if they bring improvement.
0.4
0.4
0.5
0.35
0.35
0.3
0.3
0.2
Eigenvalue
0.3
Eigenvalue
Eigenvalue
0.4
0.25
0.2
0.15
0.25
0.2
0.15
0.1
0.1
0.05
0.05
0.1
0
0
1
2
3
4
5
6
7
8
9
10
0
1
Number of eigenvalue
2
3
4
5
6
7
8
9
10
1
Number of eigenvalue
(a) Robin interface conditions
(b) NVTF interface conditions
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(c) TVNF interface conditions
Figure 8: Eigenvalues on one of the floating subdomains in case of uniform decomposition and
Taylor-Hood discretisation (T Hh2 , Rh1 ) - the driven cavity problem.
Table 3: Comparison of preconditioners for Taylor-Hood discretisation (T Hh2 , Rh1 ) - the driven
cavity problem.
DOF
N
91 003
362 003
813 003
1 444 003
2 728 003
5 768 003
4
16
36
64
121
256
DOF
N
91 003
362 003
813 003
1 444 003
2 728 003
5 768 003
4
16
36
64
121
256
DOF
N
91 003
362 003
813 003
1 444 003
2 728 003
5 768 003
4
16
36
64
121
256
ORAS
Unif
MTS
12
17
28
35
39
75
53
91
80
278
>1000 >1000
SORAS
Unif MTS
24
34
56
67
92
103
120
144
180
212
271
317
ORAS
Unif
MTS
10
14
20
25
27
33
31
42
39
103
300
849
SORAS
Unif MTS
18
22
32
37
36
44
38
53
39
51
46
54
ORAS
Unif
MTS
9
12
16
20
23
27
26
36
35
41
66
60
SORAS
Unif MTS
13
16
21
24
25
26
27
29
29
32
32
33
One-level
NVTF-MRAS NVTF-SMRAS
Unif
MTS
Unif
MTS
22
22
34
40
52
53
90
106
85
91
165
185
120
135
254
283
182
280
412
580
303
452
917
955
Two-level (2 eigenvectors)
NVTF-MRAS NVTF-SMRAS
Unif
MTS
Unif
MTS
19
17
26
30
33
34
50
62
47
49
62
86
104
66
85
114
74
81
85
133
109
108
146
132
Two-level (5 eigenvectors)
NVTF-MRAS NVTF-SMRAS
Unif
MTS
Unif
MTS
16
15
18
20
27
22
28
37
33
30
39
40
40
34
35
45
43
38
34
48
56
41
60
49
TVNF-MRAS
Unif
MTS
22
25
54
53
91
88
132
132
199
213
322
319
TVNF-SMRAS
Unif
MTS
30
40
70
84
118
136
169
206
251
439
397
695
TVNF-MRAS
Unif
MTS
27
20
60
40
79
53
85
52
92
86
91
78
TVNF-SMRAS
Unif
MTS
21
26
42
51
59
63
62
79
62
93
63
90
TVNF-MRAS
Unif
MTS
25
20
56
37
65
41
77
45
84
72
88
61
TVNF-SMRAS
Unif
MTS
16
18
26
35
28
37
28
42
29
47
29
44
The conclusions remain the same as for the L-shaped domain problem for the nearly incompressible elasticity equation discretised by Taylor-Hood method (T Hh3 , Rh2 ) since Tables 3 and 1
show similar results.
14
Test case 4 (The T-shaped domain problem). Finally, we consider a T-shaped domain Ω =
(0, 1.5) × (0, 1) ∪ (0.5, 1) × (−1, 1), and we impose mixed boundary conditions given by
(4y(1 − y), 0)T if x = 0 or x = 1.5
(5.31)
u(x, y) =
(0, 0)T
otherwise.
The numerical solution of this problem is depicted in Figure 9. The overlapping decomposition
(a) Velocity field u
(b) Pressure p
Figure 9: Numerical solution - the T-shaped problem.
into subdomains is generated by METIS.
Once again a clustering of small eigenvalues of generalised eigenvalue problem defined in (4.24)
is a motivation of the size of the coarse space (see Figure 10).
0.45
0.25
0.3
0.25
0.2
0.15
Eigenvalue
0.25
Eigenvalue
Eigenvalue
0.3
0.3
0.4
0.35
0.2
0.15
0.2
0.15
0.1
0.1
0.05
0.05
0.1
0.05
0
0
1
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(a) Robin interface conditions
0
1
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(b) NVTF interface conditions
1
2
3
4
5
6
7
8
9
10
Number of eigenvalue
(c) TVNF interface conditions
Figure 10: Eigenvalues on one of the floating subdomains in case of METIS decomposition and
Taylor-Hood discretisation (T Hh3 , Rh2 ) - the T-shaped problem.
The same as in all examples for Taylor-Hood discretisation we notice an important improvement
of the convergence when using two-level methods. Although from Table 4 we can see that the
coarse spaces containing five eigenvectors seem to be sufficient.
15
Table 4: Comparison of preconditioners for Taylor-Hood discretisation (T Hh3 , Rh2 ) - the T-shaped
problem.
DOF
33 269
138 316
269 567
553 103
1 134 314
2 201 908
N
4
16
32
64
128
256
ORAS
13
36
59
92
146
232
SORAS
20
51
85
132
208
328
DOF
33 269
138 316
269 567
553 103
1 134 314
2 201 908
N
4
16
32
64
128
256
ORAS
10
21
29
35
42
47
SORAS
14
27
35
45
52
56
DOF
33 269
138 316
269 567
553 103
1 134 314
2 201 908
N
4
16
32
64
128
256
ORAS
8
15
14
16
17
16
SORAS
13
16
19
20
22
21
5.2
One-level
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
12
19
13
33
52
31
52
85
49
83
136
78
132
223
117
209
357
189
Two-level (2 eigenvectors)
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
9
15
12
19
24
22
30
38
25
34
43
33
47
58
34
69
76
38
Two-level (5 eigenvectors)
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
8
13
12
14
16
20
20
22
24
18
19
29
23
24
30
34
37
35
TVNF-SMRAS
19
45
75
115
188
293
TVNF-SMRAS
15
24
30
35
41
45
TVNF-SMRAS
14
18
19
20
22
24
hdG discretisation
In this section we discretise nearly incompressible elasticity equation and the Stokes flow by using
the lowest order hdG discretisation introduced in Section 3.2.
5.2.1
Nearly incompressible elasticity
In case of ORAS and SORAS we consider the Robin interface conditions as in [HJN15] with
α = 10. For all numerical experiments in this section we use zero as an initial guess for the
GMRES iterative solver. Moreover, the overlapping decomposition into subdomains is generated
by METIS.
Test case 5 (The L-shaped domain problem). We consider the L-shaped domain discrete problem (5.28).
Table 5 shows an important improvement of the convergence that is brought by the twolevel methods. We cannot conclude that SMRAS preconditioners are much better than SORAS.
Although we can note that coarse space improvement is visible for MRAS preconditioners and not
for ORAS. For symmetric preconditioners (SMRAS and SORAS) five eigenvectors seem to lead
already to satisfying results. While for the non-symmetric ones a bigger coarse space is required.
On the other hand, we state the fact that the new preconditioners are parameter-free, which makes
them easier to use as no parameter is required.
16
Table 5: Comparison of preconditioners for hdG discretisation - the L-shaped domain problem.
DOF
238 692
466 094
948 921
1 874 514
3 856 425
N
8
16
32
64
128
ORAS
61
123
267
622
>1000
SORAS
158
232
331
477
752
DOF
238 692
466 094
948 921
1 874 514
3 856 425
N
8
16
32
64
128
ORAS
48
89
250
535
>1000
SORAS
98
99
130
135
152
DOF
238 692
466 094
948 921
1 874 514
3 856 425
N
8
16
32
64
128
ORAS
43
77
197
429
>1000
SORAS
81
82
100
103
118
DOF
238 692
466 094
948 921
1 874 514
3 856 425
N
8
16
32
64
128
ORAS
35
61
153
423
934
SORAS
67
79
90
95
110
One-level
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
64
174
77
101
259
109
160
415
179
243
685
254
479
>1000
523
Two-level (3 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
52
98
61
71
123
75
110
158
118
135
155
129
172
176
181
Two-level (5 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
44
74
61
51
92
63
79
119
96
102
122
110
122
129
141
Two-level (7 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
38
71
44
47
80
51
58
95
74
71
93
72
79
104
108
TDNNS-SMRAS
177
306
473
657
>1000
TDNNS-SMRAS
116
148
173
159
192
TDNNS-SMRAS
94
103
121
138
167
TDNNS-SMRAS
82
95
115
111
133
Test case 6 (The heterogeneous beam problem). We consider the heterogeneous beam with ten
layers of steel and rubber that is defined as a problem (5.29).
Table 6: Comparison of preconditioners for hdG discretisation - the heterogeneous beam.
DOF
46 777
88 720
179 721
353 440
704 329
1 410 880
N
8
16
32
64
128
256
ORAS
196
317
537
899
>1000
>1000
SORAS
440
602
>1000
>1000
>1000
>1000
DOF
46 777
88 720
179 721
353 440
704 329
1 410 880
N
8
16
32
64
128
256
ORAS
168
244
385
514
835
>1000
SORAS
255
313
525
444
557
567
DOF
46 777
88 720
179 721
353 440
704 329
1 410 880
N
8
16
32
64
128
256
ORAS
148
205
318
403
490
>1000
SORAS
197
201
337
262
168
116
One-level
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
189
402
186
330
582
326
574
>1000
587
847
>1000
846
>1000
>1000
>1000
>1000
>1000
>1000
Two-level (5 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
162
230
161
273
299
262
442
458
469
551
526
590
782
684
765
>1000
694
844
Two-level (7 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
149
192
158
286
187
283
385
301
433
397
247
460
447
182
558
387
138
473
TDNNS-SMRAS
463
666
>1000
>1000
>1000
>1000
TDNNS-SMRAS
275
346
587
558
832
821
TDNNS-SMRAS
231
273
419
389
443
298
We notice an improvement only when using a coarse space which is sufficiently big (see Table 6).
17
Furthermore, we get a stable number of iterations only for the symmetric preconditioners (SMRAS
and SORAS), and the coarse space improvement in case of ORAS preconditioner is much less
visible than in case of MRAS preconditioners. This may be due to the fact we have not chosen an
optimized parameter in the Robin interface conditions (5.27).
5.2.2
Stokes equation
We now turn to the Stokes discrete problem given in 3.2. Once again in case of ORAS and SORAS
we choose α = 10 as in [HJN15] for the Robin interface conditions (5.27). In the first case we
consider a random initial guess for the GMRES iterative solver. Later with the second example
we use zero as an initial guess.
Test case 7 (The driven cavity problem). We consider the driven cavity defined as a problem (5.30).
Table 7: Comparison of preconditioners for hdG discretisation - the driven cavity problem.
DOF
N
93 656
373 520
839 592
1 491 872
2 819 432
5 963 072
4
16
36
64
121
256
DOF
N
93 656
373 520
839 592
1 491 872
2 819 432
5 963 072
4
16
36
64
121
256
DOF
N
93 656
373 520
839 592
1 491 872
2 819 432
5 963 072
4
16
36
64
121
256
ORAS
Unif
MTS
17
18
76
122
152
327
261
587
364
>1000
592
>1000
SORAS
Unif MTS
37
38
75
84
120
133
162
176
229
256
367
398
ORAS
Unif
MTS
12
14
81
80
236
228
395
463
840
>1000
>1000 >1000
SORAS
Unif MTS
30
28
47
57
61
60
67
71
73
86
80
87
ORAS
Unif
MTS
10
12
27
35
135
84
278
212
607
584
>1000 >1000
SORAS
Unif MTS
25
24
38
37
45
41
49
45
56
49
62
55
One-level
NVTF-MRAS NVTF-SMRAS
Unif
MTS
Unif
MTS
24
22
44
44
52
54
107
111
91
96
194
200
130
143
294
303
199
213
504
649
326
477
>1000
>1000
Two-level (2 eigenvectors)
NVTF-MRAS NVTF-SMRAS
Unif
MTS
Unif
MTS
18
18
33
32
36
40
61
73
57
65
97
104
79
85
139
129
113
127
188
178
171
179
283
287
Two-level (5 eigenvectors)
NVTF-MRAS NVTF-SMRAS
Unif
MTS
Unif
MTS
14
16
23
22
27
29
38
41
35
37
51
50
44
42
58
55
46
56
58
62
52
64
57
69
TVNF-MRAS
Unif
MTS
32
25
68
67
112
115
159
158
238
251
392
404
TVNF-SMRAS
Unif
MTS
48
50
122
126
206
210
292
286
628
643
995
740
TVNF-MRAS
Unif
MTS
40
23
100
49
132
66
142
70
157
86
167
108
TVNF-SMRAS
Unif
MTS
38
37
85
82
112
107
128
122
127
139
132
148
TVNF-MRAS
Unif
MTS
52
22
117
39
145
49
157
59
162
81
166
75
TVNF-SMRAS
Unif
MTS
29
26
53
53
64
61
64
64
65
75
65
75
The conclusions remain the same as in the case of nearly incompressible elasticity equation
for the L-shaped domain. Although Table 7 shows that coarse spaces containing five eigenvectors
seem to decrease the number of iteration even in the case of MRAS preconditioners that are not
fully scalable.
Test case 8 (The T-shaped domain problem). Finally, we consider a T-shaped domain Ω =
(0, 1.5)×(0, 1)∪(0.5, 1)×(−1, 1), and we impose mixed boundary conditions (5.31). The numerical
solution of this problem is depicted in Figure 9.
In this case scalable results can be only observed for the preconditioners associated with the
non standard interface conditions (MRAS and SMRAS), and when using a coarse space which
is sufficiently big (see Table 8). In the case of ORAS or SORAS, one possibility is to choose a
different parameter α, but the proof of this, and even the question of whether this would have a
positive impact, are open problems.
18
Table 8: Comparison of preconditioners for hdG discretisation - the T-shaped problem.
DOF
38 803
154 606
311 369
616 772
1 246 136
2 451 365
N
4
16
32
64
128
256
ORAS
22
111
265
568
>1000
>1000
SORAS
45
98
144
238
494
712
DOF
38 803
154 606
311 369
616 772
1 246 136
2 451 365
N
4
16
32
64
128
256
ORAS
16
113
254
510
>1000
>1000
SORAS
35
69
99
153
221
286
DOF
38 803
154 606
311 369
616 772
1 246 136
2 451 365
N
4
16
32
64
128
256
ORAS
14
155
159
426
955
>1000
SORAS
30
54
55
88
113
182
6
One-level
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
36
49
22
83
172
83
133
262
130
212
410
195
333
665
313
464
>1000
477
Two-level (2 eigenvectors)
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
31
37
21
73
75
38
103
176
93
171
273
121
242
252
155
343
515
189
Two-level (5 eigenvectors)
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
27
27
28
54
45
25
72
59
29
106
83
37
115
99
43
138
101
54
TVNF-SMRAS
51
182
266
412
602
889
TVNF-SMRAS
38
75
162
140
138
231
TVNF-SMRAS
30
44
52
76
72
73
Numerical results for three dimensional problems
In this section we again assess the performance of the preconditioners as in Section 5, but this
time in case of three dimensional problems. We consider the partial differential equation model
for nearly incompressible elasticity and Stokes flow as three dimensional problems of similar mixed
formulation. Each of these problems is discretised by using the Taylor-Hood methods from Section 3.1. In addition, we use the same tools as in Section 5. For both test cases we use zero as an
initial guess.
6.1
Taylor-Hood discretisation
In this section we consider the Taylor-Hood discretisation from Section 3.1 with k = 2 for nearly
incompressible elasticity and Stokes equations.
6.1.1
Nearly incompressible elasticity
In three dimensional space, ORAS and SORAS preconditioners also require an optimized parameter. We follow [HJN15] and use Robin interface conditions (5.27) with α = 10.
Test case 9 (The homogeneous beam problem). We consider a homogeneous beam with the
physical parameters E = 108 and ν = 0.4999. The computational domain is the rectangle Ω =
(0, 5) × (0, 1) × (0, 1). The beam is clamped on one side, hence we consider the following problem
−2µ∇ · ε(u) + ∇p
= (0, 0, −1)T
in Ω
in Ω
− ∇·u
= λ1 p
.
(6.32)
u(x, y)
= (0, 0, 0)T
on ∂Ω ∩ {x = 0}
T
σ sym
on ∂Ω \ {x = 0}
n (x, y) = (0, 0, 0)
The results of Table 9 show a clear improvement in the scalability of the two-level preconditioners over the one-level ones. In fact, using only zero energy modes, the number of iterations
19
Table 9: Comparison of preconditioners for Taylor-Hood discretisation (T Hh2 , Rh1 ) - the homogeneous beam.
DOF
32 446
73 548
139 794
299 433
549 396
N
8
16
32
64
128
ORAS
21
31
43
55
78
SORAS
45
70
99
143
192
DOF
32 446
73 548
139 794
299 433
549 396
N
8
16
32
64
128
ORAS
10
11
13
15
17
SORAS
17
22
26
27
28
DOF
32 446
73 548
139 794
299 433
549 396
N
8
16
32
64
128
ORAS
9
10
11
14
16
SORAS
16
19
21
24
27
One-level
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
29
36
27
38
64
26
74
94
66
161
140
149
314
192
229
Two-level (6 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
13
18
12
16
25
16
25
28
17
19
27
24
20
25
21
Two-level (8 eigenvectors)
NDTNS-MRAS NDTNS-SMRAS TDNNS-MRAS
12
17
12
15
24
14
17
23
17
17
24
21
18
23
20
TDNNS-SMRAS
37
67
91
139
199
TDNNS-SMRAS
17
22
26
28
26
TDNNS-SMRAS
16
20
21
23
22
is virtually unaffected by the number of subdomains. All two-level preconditioners show a comparable performance. For this case, increasing the dimension of the coarse space beyond 6 × N
eigenvectors does not seem to improve the results dramatically.
6.1.2
Stokes equation
We now turn to the Stokes discrete problem given in 3.1. Once again in case of ORAS and SORAS
we choose α = 10 as in [HJN15] for the Robin interface conditions (5.27).
Test case 10 (The driven cavity problem). The test case is the three-dimensional version of the
driven cavity problem. We consider the following problem on the unit cube Ω = (0, 1)3
= f
in Ω
−∆u + ∇p
− ∇·u
= 0
in Ω
(6.33)
.
T
u(x,
y)
=
(1,
0,
0)
on
∂Ω
∩
{y
=
1}
u(x, y) = (0, 0, 0)T on ∂Ω \ {y = 1}
The conclusions remain the same as for the homogeneous beam example for the nearly incompressible elasticity equation discretised by Taylor-Hood method (T Hh3 , Rh2 ) since Tables 10 and 9
show similar results.
20
Table 10: Comparison of preconditioners for Taylor-Hood discretisation (T Hh2 , Rh1 ) - the driven
cavity problem.
DOF
38 229
76 542
158 818
325 293
643 137
N
8
16
32
64
128
ORAS
12
18
23
28
37
SORAS
24
34
45
60
79
DOF
38 229
76 542
158 818
325 293
643 137
N
8
16
32
64
128
ORAS
10
11
13
15
18
SORAS
17
20
24
27
31
DOF
38 229
76 542
158 818
325 293
643 137
N
8
16
32
64
128
ORAS
9
10
11
13
15
SORAS
16
17
19
19
21
7
One-level
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
12
22
11
18
31
15
20
45
19
36
64
25
64
91
33
Two-level (3 eigenvectors)
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
10
18
11
11
19
14
13
24
16
15
27
19
17
32
22
Two-level (7 eigenvectors)
NVTF-MRAS NVTF-SMRAS TVNF-MRAS
9
16
12
10
18
15
11
20
17
13
21
20
16
22
22
TVNF-SMRAS
23
31
45
60
88
TVNF-SMRAS
18
19
23
26
31
TVNF-SMRAS
17
17
20
20
22
Conclusion
We tested numerically two-level preconditioners with spectral coarse spaces for nearly incompressible elasticity and Stokes equations. We considered two finite element methods, namely,
Taylor-Hood (Section 3.1) and the hdG (Section 3.2) discretisations.
In the case of the homogeneous nearly incompressible elasticity the two-level methods coupled
with SORAS preconditioner defined in [HJN15] and SMRAS preconditioner defined by (4.23)
allowed us to achieve good scalability results for both discretisations. Furthermore, for these
symmetric preconditioners coarse spaces containing only zero energy modes seem to be enough for
two and three dimensional problems. For the heterogeneous problem we also achieved scalability
for two-level SORAS and SMRAS preconditioners, but, as expected, only in the case when the
size of the coarse space is sufficiently big.
The improvement of the convergence in the case of the Stokes flow is visible only when the
coarse space contains more eigenvectors than only constants. For the Taylor-Hood discretisation,
taking sufficient big coarse space we were able to achieve good scalability for all preconditioners.
It is remarkable that these good results occur even when using the hdG discretisation, despite the
fact the optimized parameter to be used in SORAS and ORAS is not available.
We can conclude that the two-level preconditioners associated with non standard interface
conditions are at least as good as the two-level ones in conjunction with Robin interface conditions
using optimised parameters. It shows an important advantage of newly introduced preconditioners
as they are parameter-free.
Numerical tests have shown that the coarse spaces bring an important improvement in the
convergence, but the size of the coarse space depends on the problem. Building as small as
possible coarse spaces is important from computational point of view. Thus, it is necessary to
investigate what could be an optimal criterion for choosing the eigenvectors for a coarse space.
As we mentioned before, the theoretical foundation of the two-level preconditioners has not
been extended to saddle point problems. Hence, future research will be devoted to this topic.
21
Acknowledgements
This research was supported supported by the Centre for Numerical Analysis and Intelligent
Software (NAIS). We thank Frédéric Nataf and Pierre-Henri Tournier for many helpful discussions
and insightful comments, and Ryadh Haferssas and Frédéric Hecht for their assistance with the
FreeFem++ codes.
References
[AdDBM+ 14] B. Ayuso de Dios, F. Brezzi, L. D. Marini, J. Xu, and L. Zikatanov. A simple
preconditioner for a discontinuous Galerkin method for the Stokes problem. Journal
of Scientific Computing, 58(3):517–547, 2014.
[BBD+ 16]
G. R Barrenechea, M. Bosy, V. Dolean, F. Nataf, and P.-H. Tournier. Hybrid discontinuous Galerkin discretisation and preconditioning of the Stokes problem with non
standard boundary conditions. preprint, https://arxiv.org/abs/1610.09207,
October 2016.
[BBF13]
D. Boffi, F. Brezzi, and M. Fortin. Mixed finite element methods and applications,
volume 44 of Springer Series in Computational Mathematics. Springer, Heidelberg,
2013.
[BHMV99]
M. Brezina, C. I Heberton, J. Mandel, and P. Vanek. An iterative method with
convergence rate chosen a priori, ucd/ccm report 140. Technical report, Center for
Computational Mathematics, University of Colorado at Denver, 1999.
[CDKN14]
L. Conen, V. Dolean, R. Krause, and F. Nataf. A coarse space for heterogeneous
Helmholtz problems based on the Dirichlet-to-Neumann operator. J. Comput. Appl.
Math., 271:83–99, 2014.
[CG09]
B. Cockburn and J. Gopalakrishnan. The derivation of hybridizable discontinuous
Galerkin methods for Stokes flow. SIAM J. Numer. Anal., 47(2):1092–1125, 2009.
[CGL09]
B. Cockburn, J. Gopalakrishnan, and R. Lazarov. Unified hybridization of discontinuous Galerkin, mixed, and continuous Galerkin methods for second order elliptic
problems. SIAM J. Numer. Anal., 47(2):1319–1365, 2009.
[CS15]
C. Carstensen and M. Schedensack. Medius analysis and comparison results for firstorder finite element methods in linear elasticity. IMA J. Numer. Anal., 35(4):1591–
1621, 2015.
[DJN15]
V. Dolean, P. Jolivet, and F. Nataf. An introduction to domain decomposition
methods. Society for Industrial and Applied Mathematics (SIAM), Philadelphia,
PA, 2015. Algorithms, theory, and parallel implementation.
[DNSS12]
V. Dolean, F. Nataf, R. Scheichl, and N. Spillane. Analysis of a two-level Schwarz
method with coarse spaces based on local Dirichlet-to-Neumann maps. Comput.
Methods Appl. Math., 12(4):391–414, 2012.
[EGLW12]
Y. Efendiev, J. Galvis, R. Lazarov, and J. Willems. Robust domain decomposition
preconditioners for abstract symmetric positive definite bilinear forms. ESAIM
Math. Model. Numer. Anal., 46(5):1175–1199, 2012.
[GE10a]
J. Galvis and Y. Efendiev. Domain decomposition preconditioners for multiscale
flows in high-contrast media. Multiscale Model. Simul., 8(4):1461–1483, 2010.
[GE10b]
J. Galvis and Y. Efendiev. Domain decomposition preconditioners for multiscale
flows in high contrast media: reduced dimension coarse spaces. Multiscale Model.
Simul., 8(5):1621–1644, 2010.
[GR86]
V. Girault and P. A. Raviart. Finite element methods for Navier-Stokes equations,
volume 5 of Springer Series in Computational Mathematics. Springer-Verlag, Berlin,
1986. Theory and algorithms.
[Hec12]
F. Hecht. New development in FreeFem++. J. Numer. Math., 20(3-4):251–265,
2012.
[HJN15]
R. Haferssas, P. Jolivet, and F. Nataf. An additive Schwarz method type
theory for Lions’ algorithm and Optimized Schwarz Methods.
preprint,
https://hal.archives-ouvertes.fr/hal-01278347, December 2015.
[KK98]
G. Karypis and V. Kumar. A software package for partitioning unstructured graphs,
partitioning meshes, and computing fill-reducing orderings of sparse matrices. Technical report, University of Minnesota, Department of Computer Science and Engineering, Army HPC Research Center, Minneapolis, MN, 1998.
[LNS15]
Sébastien Loisel, Hieu Nguyen, and Robert Scheichl. Optimized Schwarz and 2Lagrange multiplier methods for multiscale elliptic PDEs. SIAM J. Sci. Comput.,
37(6):A2896–A2923, 2015.
[LSY98]
R. B. Lehoucq, D. C. Sorensen, and C. Yang. ARPACK users’ guide, volume 6 of
Software, Environments, and Tools. Society for Industrial and Applied Mathematics
(SIAM), Philadelphia, PA, 1998. Solution of large-scale eigenvalue problems with
implicitly restarted Arnoldi methods.
[Nic87]
R. A. Nicolaides. Deflation of conjugate gradients with applications to boundary
value problems. SIAM J. Numer. Anal., 24(2):355–365, 1987.
[NXD10]
F. Nataf, H. Xiang, and V. Dolean. A two level domain decomposition preconditioner based on local Dirichlet-to-Neumann maps. C. R. Math. Acad. Sci. Paris,
348(21-22):1163–1167, 2010.
[PS11]
A. Pechstein and J. Schöberl. Tangential-displacement and normal-normal-stress
continuous mixed finite elements for elasticity. Math. Models Methods Appl. Sci.,
21(8):1761–1782, 2011.
[SDH+ 14]
N. Spillane, V. Dolean, P. Hauret, F. Nataf, C. Pechstein, and R. Scheichl. Abstract
robust coarse spaces for systems of PDEs via generalized eigenproblems in the
overlaps. Numer. Math., 126(4):741–770, 2014.
[SS86]
Y. Saad and M. H. Schultz. GMRES: a Generalized Minimal Residual algorithm for
solving nonsymmetric linear systems. SIAM J. Sci. Statist. Comput., 7(3):856–869,
1986.
[TW05]
A. Toselli and O. Widlund. Domain Decomposition Methods - Algorithms and
Theory, volume 34 of Springer Series in Computational Mathematics. Springer,
2005.
23
| 5cs.CE
|
Pseudorandom number generator based on the Bernoulli map on
cubic algebraic integers
Asaki Saito1, ∗ and Akihiro Yamaguchi2
arXiv:1706.08472v1 [math.NT] 20 Jun 2017
1
Future University Hakodate, 116-2 Kamedanakano-cho,
Hakodate, Hokkaido 041-8655, Japan
2
Fukuoka Institute of Technology, 3-30-1 Wajiro-higashi,
Higashi-ku, Fukuoka 811-0295, Japan
(Dated: June 27, 2017)
Abstract
We develop a method for generating pseudorandom binary sequences using the Bernoulli map
on cubic algebraic integers. The distinguishing characteristic of our generator is that it generates
chaotic true orbits of the Bernoulli map by exact computation. In particular, we clarify a way
to properly prepare a set of initial points (i.e., seeds), which is needed when generating multiple
pseudorandom sequences. With this seed selection method, we can distribute the initial points
almost uniformly in the unit interval and can also guarantee that the orbits starting from them
do not merge. We also report results of a large variety of tests indicating that the generated
pseudorandom sequences have good statistical properties as well as an advantage over what is
probably the most popular generator, the Mersenne Twister MT19937.
PACS numbers: 05.45.-a
∗
Electronic address: [email protected]
1
I.
INTRODUCTION
A random sequence is a sequence of numbers that are a typical sample of independently
identically distributed random variables, and it cannot be generated by a deterministic algorithm (cf., e.g., Refs. [1, 2]). A pseudorandom sequence, i.e., a computer-generated sequence
that appears similar to a random one, is therefore not random at all, but has a wide range of
applications, such as Monte Carlo methods, probabilistic algorithms, and cryptography [3].
In order to generate such pseudorandom sequences, various pseudorandom number generators have been proposed, including linear congruential generators [4], linear feedback shift
registers [5], and the Mersenne Twister [6]. Of all the generators, MT19937 [6], a version of
the Mersenne Twister, is probably the most popular one at this time. MT19937 can produce, at very high speed, a pseudorandom sequence having an “astronomically” long period
of length 219937 − 1 and having a high-dimensional (623-dimensional) equidistribution property, which makes the generator very useful, especially for Monte Carlo simulations. Even
if a generator has these remarkable properties, however, there is no guarantee that independence, which is the greatest characteristic of random sequences, is preferable (cf. Sec. IV B).
In this paper, we deal with the issue of how we generate pseudorandom sequences having
the best possible statistical properties even if such generation increases the computational
cost to some extent.
Among random sequences, the most fundamental ones are (uniform) random binary sequences. One of the mathematically simplest and soundest ways to generate (pseudo-)
random binary sequences is to use the Bernoulli map. Also known as the doubling map,
the dyadic map, or the 2x modulo 1 map, the Bernoulli map is a map on the half-open unit
interval [0, 1) given by
MB (x) =
2x
if x ∈ [0, 1/2)
2x − 1
if x ∈ [1/2, 1) .
Note that the repeated tossing of a fair coin is modeled by the one-sided Bernoulli shift on
{0, 1}N with 0 and 1 having equal weight 1/2, and this Bernoulli shift is measure-theoretically
isomorphic to MB (cf., e.g., Ref. [7]). Thus, MB can produce binary sequences equivalent to
those obtained by tossing a fair coin. However, it is well known that one cannot simulate MB
with conventional simulation methods such as those using double-precision binary floatingpoint numbers or arbitrary-precision rational numbers (see, e.g., Refs. [8, 9]). This is because,
2
for MB , finite binary decimals on [0, 1) are eventually fixed points (i.e., points that reach
the fixed point at x = 0 after finitely many iterations) and because rational numbers on
[0, 1) are eventually periodic points (i.e., points that reach a periodic point after finitely
many iterations). For this reason, a computational method that realizes pseudorandom
number generation using MB has not been proposed [10] (except for our previous study
[11]), although pseudorandom number generators based on chaotic dynamics have been very
widely studied for many decades [12–14].
On the other hand, orbit computations using algebraic numbers other than rational ones
have been performed in the fields of number theory and arithmetic dynamics (e.g., Refs. [15–
19]). Also, by using our methods to achieve exact simulations of piecewise linear and linear
fractional maps [20, 21], one can generate errorless true orbits displaying the same statistical
properties as typical orbits of MB (as well as those of the tent map and the baker’s transformation; cf. [10]). In particular, by using true orbits on quadratic algebraic integers, we
succeeded in realizing a pseudorandom number generator using MB [11]. To our knowledge,
the generator of Ref. [11] is the only one that has a direct connection to the repeated tossing
of a fair coin, but we can expect that we can establish such generators having good statistical properties also by using algebraic integers of degree three or more. In order to realize
such generators, however, we particularly need to resolve the issue described below. When
proposing a pseudorandom number generator, it is desirable to simultaneously disclose how
one can properly perform seed selection, especially in the case where one needs multiple
seeds to generate more than one pseudorandom sequence. In particular, such a method
for selecting initial points (i.e., seeds) is indispensable for a generator based on true orbits:
In true orbit computations, the longer a true orbit, the higher the computational cost of
generating it. Therefore, the computational cost can be markedly lowered by generating a
number of relatively short true orbits. We could establish such a seed selection method in
the case of quadratic algebraic integers, but algebraic numbers of different degrees are quite
distinct from each other, and it is unclear even whether such a seed selection method exists
in the case of algebraic integers of degree three or more.
In this paper, we realize a pseudorandom number generator using chaotic true orbits of
the Bernoulli map on cubic algebraic integers. We also devise, for the cubic case, a seed
selection method for generating multiple pseudorandom binary sequences. Moreover, we
demonstrate the ability of our generator by performing two kinds of computer experiments:
3
extensive statistical testing and a comparison with MT19937.
II.
PROPOSED PSEUDORANDOM NUMBER GENERATOR
In this study, we use cubic algebraic integers to simulate the Bernoulli map MB . A cubic
algebraic integer is a complex number that is a root of a monic irreducible cubic polynomial
x3 + bx2 + cx + d with b, c, d ∈ Z (see, e.g., Ref. [22] for a detailed explanation of algebraic
integers). MB maps any cubic algebraic integer in the open unit interval (0, 1) to a cubic
algebraic integer in (0, 1).
Let us introduce two sets, S̄ and S, and a map π from S̄ to S. Let S̄ be the set of all
(b, c, d) ∈ Z3 satisfying the following three conditions:
(i) b2 − 3c ≤ 0
(ii) d < 0
(iii) 1 + b + c + d > 0
Figure 1 shows part of S̄. If we consider a function f : R → R, given by f (x) = x3 + bx2 +
cx + d with (b, c, d) ∈ S̄, we see from (i) that f is strictly monotonically increasing. Thus,
f has a unique real root, denoted by α. We also see from (ii) and (iii) that f (0) < 0 and
f (1) > 0, which implies α ∈ (0, 1). Since α ∈
/ Z, we see that α is a cubic algebraic integer.
Also, let S be the set of all cubic algebraic integers in (0, 1) that are roots of x3 + bx2 + cx + d
with (b, c, d) ∈ S̄. We can define a map π from S̄ to S by assigning each (b, c, d) ∈ S̄ the
unique real root α ∈ S of x3 + bx2 + cx + d. It is easy to see that π : S̄ → S is a bijection.
In the following, we represent α ∈ S with (b, c, d) = π −1 (α) ∈ S̄.
MB maps α ∈ S to α′ = 2α mod 1. As already mentioned, α′ is a cubic algebraic integer
in (0, 1). Moreover, we can see α′ ∈ S as follows. Let (b, c, d) be the representation of α,
and let x3 + b′ x2 + c′ x + d′ be the minimal polynomial of α′ . Then, the coefficients b′ , c′ , d′
are given as follows:
If α ∈ (0, 1/2),
′
b
2 0 0
b
′
c = 0 4 0 c .
d
0 0 8
d′
4
(1a)
10
5c
0
5
-15
-10
d -5
0
b
0
-5
FIG. 1: Part of S̄. The dots represent elements of S̄.
If α ∈ (1/2, 1),
3
b
2 0 0
b
′
c = 4 4 0 c + 3 .
1
d
2 4 8
d′
′
(1b)
Equation (1a) (resp. Eq. (1b)) is obtained by substituting α = α′ /2 (resp. α = (α′ + 1)/2)
into x3 + bx2 + cx + d = 0. It is easy to confirm that (b′ , c′ , d′ ) satisfies the conditions (i),
(ii), and (iii), which implies α′ ∈ S.
Equation (1) gives the explicit form of π −1 ◦ MB ◦ π, i.e., the transformation on S̄ corre-
sponding to MB . We denote this transformation by M̄B . Note that M̄B gives the representation (b′ , c′ , d′) of α′ from the representation (b, c, d) of α. This transformation is exactly
computable by using only integer arithmetic. Incidentally, we see easily that (b, c, d) with
b2 − 3c < 0 is mapped to (b′ , c′ , d′ ) with b′2 − 3c′ < 0, and that (b, c, d) with b2 − 3c = 0 is
mapped to (b′ , c′ , d′ ) with b′2 − 3c′ = 0.
One has to exactly determine whether a given α ∈ S is in (0, 1/2) or (1/2, 1) in order to
generate true orbits of MB on S, thereby obtaining pseudorandom binary sequences. Let α
be represented by (b, c, d) ∈ S̄, and let f (x) = x3 + bx2 + cx + d. This determination can
be made easily by evaluating the sign of f (1/2). In fact, if f (1/2) > 0, then α ∈ (0, 1/2);
otherwise, i.e., if f (1/2) < 0, then α ∈ (1/2, 1). To evaluate the sign of f (1/2), it is sufficient
to evaluate that of 1 + 2b + 4c + 8d. Therefore, one can exactly determine whether α ∈
(0, 1/2) or not by using only integer arithmetic. Consequently, one can generate a true orbit
5
{(bn , cn , dn )}n=0,1,2,··· of M̄B starting from an initial point (b0 , c0 , d0 ) ∈ S̄, where (bn , cn , dn ) =
M̄Bn (b0 , c0 , d0 ). In order to obtain a pseudorandom binary sequence {ǫn }n=0,1,2,···, all one has
to do is let ǫn = 0 if 1 + 2bn + 4cn + 8dn > 0 and ǫn = 1 if 1 + 2bn + 4cn + 8dn < 0, in the
course of generating a true orbit.
III.
SEED SELECTION METHOD
In this section, we consider how to select initial points (i.e., seeds). Because α represented
by (b, c, d) ∈ S̄ is irrational, its binary expansion is guaranteed to be nonperiodic. Thus,
one can choose any (b, c, d) ∈ S̄ as an initial point in the sense that at least one obtains
a nonperiodic binary sequence. It is worth noting that the binary sequence obtained from
(b, c, d) ∈ S̄ is not only guaranteed to be nonperiodic. In fact, it is widely believed that
every irrational algebraic number is a normal number (Borel’s conjecture [23]). Recall that
α ∈ R is said to be normal if, for any integer b ≥ 2, every word of length l ≥ 1 on the
alphabet {0, 1, . . . , b − 1} occurs in the base-b expansion of α with asymptotic frequency
b−l . Also, it is proved that the base-b expansion of any irrational algebraic number cannot
have a regularity so simple that it can be generated by a finite automaton [24]. Moreover,
our previous studies strongly suggest that the base-b expansion of any irrational algebraic
number has the same statistical properties as those of almost all real numbers [11, 20, 21].
For the generation of more than one pseudorandom sequence, it is necessary to prepare
an initial point set I¯ ⊂ S̄. One can consider a variety of conditions that I¯ should satisfy;
here, we impose the following two conditions (Conditions 1 and 2) on it.
Condition 1. The elements of I ⊂ S corresponding to I¯ are uniformly distributed in the
unit interval.
This condition is for unbiased sampling of initial points and is a natural one also from
the viewpoint of applications, such as the Monte Carlo method. However, it is a nontrivial
question as to how we can construct I¯ satisfying Condition 1, because α depends on (b, c, d)
in a very complicated way. In fact, α takes the following complex form:
6
If b2 − 3c < 0,
p
√ √
3
−2b3 + 9bc − 27d + 3 3 −b2 c2 + 4c3 + 4b3 d − 18bcd + 27d2
√
α =
332
√
3
2 (−b2 + 3c)
b
p
− 3
− .
√ √
3 −2b3 + 9bc − 27d + 3 3 −b2 c2 + 4c3 + 4b3 d − 18bcd + 27d2 3
If b2 − 3c = 0,
α =
√
3
−2b3 + 9bc − 27d b
− .
3
3
Condition 2. The orbits starting from the elements of I¯ do not merge.
Even if one selects two different points as the elements of S̄, the latter parts of the resulting
binary sequences may coincide with each other. In fact, this happens if the two points are on
the same orbit or, more generally, if the orbits starting from them merge. When generating
multiple pseudorandom sequences, it is desirable that the binary sequences derived from
I¯ are as different from each other as possible, and it is obviously desirable that I¯ satisfies
¯ we need to make it clear how we can
Condition 2. However, in order to realize such an I,
select the elements of I¯ while avoiding such orbital overlaps.
In what follows, we show that we can construct an I¯ satisfying Conditions 1 and 2.
Concerning Condition 1, the following fact holds.
Fact 1. Let c be a sufficiently large positive integer, and let
I¯b,c = (b, c, d) ∈ S̄ | d ∈ {−1, −2, · · · , −(b + c)} .
(3)
Then, the elements of Ib,c ⊂ S corresponding to I¯b,c are distributed almost uniformly (equidistantly) in the unit interval.
Proof. Since |b| ≤
√
3c, |b| ≪ c holds for sufficiently large c. Let (b, c, d) ∈ I¯b,c , αd =
π(b, c, d), and fd (x) = x3 + bx2 + cx + d. We see that f−1 (0) = −1, limc→∞ f−1 (2/c) = 1,
limc→∞ f−(b+c) (1 − 2/c) = −1, and f−(b+c) (1) = 1. Thus, we have limc→∞ α−1 = 0 and
limc→∞ α−(b+c) = 1. We also see easily that αd < αd−1 and fd (αd−1 ) = 1 hold for d ∈
{−1, −2, · · · , −(b + c) + 1}. Let ∆d = αd−1 − αd (d ∈ {−1, −2, · · · , −(b + c) + 1}). By the
mean value theorem, there exists β ∈ (αd , αd−1 ) such that fd′ (β) = ∆−1
d . It is easy to see
7
that −2|b| + c < fd′ (x) < 3 + 2|b| + c holds for x ∈ (0, 1). Thus, for sufficiently large c, we
have (3 + 2|b| + c)−1 < ∆d < (−2|b| + c)−1 , which implies
−1
−1
2|b| + 3
2|b|
∆d
1+
< −1 < 1 −
.
c
c
c
We note that −x + 2 ≤ x−1 holds for x ≥ 1 and that x−1 ≤ −2x + 3 holds for 1/2 ≤ x ≤ 1.
Thus, for sufficiently large c, we have
1−
2|b| + 3
∆d
4|b|
< −1 < 1 +
,
c
c
c
which implies
lim
max
c→∞ d∈{−1,··· ,−(b+c)+1}
∆d
− 1 = 0.
c−1
Therefore, if we take a sufficiently large c, the elements of Ib,c are distributed across the unit
interval almost equidistantly, with distances approximately equal to c−1 .
An important characteristic of M̄B on S̄ (or equivalently, MB on S) when considering
I¯ satisfying Condition 2 is its injectivity. The inverse image of (b′ , c′ , d′ ) ∈ S̄ under M̄B is
uniquely determined if it exists: If b′ (or c′ or d′ ) is even, (b′ , c′ , d′ ) is derived from Eq. (1a).
If odd, it is derived from Eq. (1b). Let us call an element of S̄ a source point if it does
not have an inverse image in S̄. It is clear that two different source points do not exist on
the same orbit. Also, the injectivity prevents the merging of orbits starting from different
source points. Concerning the source points, the following fact holds.
Fact 2. There is no inverse image for (b, c, d) ∈ S̄ if and only if one of the following
conditions holds:
(i) b, c, d are neither all even nor all odd.
(ii) b, c, d are all even, but c 6≡ 0 (mod 4) or d 6≡ 0 (mod 8).
(iii) b, c, d are all odd, but −2b + c 6≡ 1 (mod 4) or b − c + d 6≡ 1 (mod 8).
Proof. If (b, c, d) ∈ S̄ has an inverse image, then by Eq. (1) b, c, d are either all even or all
odd. We can easily verify that a necessary and sufficient condition for (b, c, d) ∈ S̄ with b,
c, d all even to have an inverse image is that both c ≡ 0 (mod 4) and d ≡ 0 (mod 8) hold.
Similarly, we can verify that a necessary and sufficient condition for (b, c, d) ∈ S̄ with b, c, d
all odd to have an inverse image is that both −2b + c ≡ 1 (mod 4) and b − c + d ≡ 1 (mod 8)
8
hold. Therefore, (b, c, d) ∈ S̄ has no inverse image if and only if one of the conditions (i)–(iii)
holds.
The orbits starting from the elements of I¯ do not merge if one lets I¯ consist of only source
points.
Consequently, on the basis of Facts 1 and 2, we can construct I¯ satisfying Conditions 1
and 2: The simplest way is to choose b to be an even integer and c to be a large positive
odd integer, or b to be an odd integer and c to be a large positive even integer, and to let
I¯ be the I¯b,c given by Eq. (3). Note, however, that consisting of only source points is not a
necessary condition for I¯ to be free from orbital mergers. For example, I¯b,c with b = 0 and
c = 8 contains a point that is not a source point, but mergers do not occur with I¯0,8 (cf.
next paragraph).
Condition 2 is equivalent to the condition that latter parts of the binary sequences derived
from I¯ do not coincide, which in turn is equivalent to the condition that, even if each of the
binary sequences is transformed by any multi-bit shift operation that is expressible as a map
x 7→ 2n x mod 1 (n ∈ Z≥0 ), none of the resulting sequences are identical. With computer
assistance, one can reveal that many, but not all, of I¯b,c have a much more desirable property
than Condition 2. Namely, for many of I¯b,c , Q(α) 6= Q(β) holds for all α, β ∈ Ib,c with α 6= β
(i.e., each element of Ib,c belongs to a different cubic field). In particular, we experimentally
confirmed that all of Ib,c with b = 0 and c in 1 ≤ c ≤ 5 × 104 have this desirable property,
which leads us to the following conjecture:
Conjecture 1. Let c ∈ Z>0 . Then, Q(α) 6= Q(β) holds for all α, β ∈ I0,c with α 6= β.
If Ib,c has such a property, the binary sequences derived from I¯b,c are significantly different
from each other in the following sense: Even if each of the binary sequences is transformed by
any operation expressible as a rational map with rational coefficients (except those mapping
elements of Ib,c to rational numbers), the resulting sequences include no identical sequences.
Such operations include not only multi-bit shifts, but a wide variety of operations, e.g.,
all-bit inversion, which is expressible as the map x 7→ 1 − x.
9
IV.
A.
EXPERIMENTAL RESULTS
Statistical testing
We evaluated our generator using three statistical test suites: DIEHARD [25], NIST
statistical test suite [26], and TestU01 [27]. We summarize their results in Table I.
We performed DIEHARD and NIST tests on the binary sequences of length 106 derived
from I¯0,1001 . For TestU01, we prepared test data as follows: We generated the binary sequences of length 1000032 using I¯0,12000001 . We then removed the first 32 bits of each sequence
and concatenated the resulting sequences in descending order of d value. We removed the
first 32-bit blocks in order to avoid introducing correlations among them, because each of
these blocks stores information regarding the position of the initial point.
Here we briefly explain the three statistical test suites and report their results.
DIEHARD [25] contains 234 statistical tests classified into 18 categories. The results
for 6 of the 18 categories are further tested by checking the uniformity of the resulting P values. (That is, DIEHARD consists of 234 first-level tests and 6 second-level ones.) Using
DIEHARD version “DOS, Jan 7, 1997”, we performed all 240 tests with a significance level
of 0.01. As a result, 238 of the 240 tests were passed.
NIST statistical test suite [26] contains 188 first-level tests. In NIST testing, each of 188
first-level tests is performed 103 times, and the results of each first-level test are further
tested in two ways: (i) The proportion of passing sequences is tested using a significance
level of 0.001540 (cf. Ref. [11]). (ii) The uniformity of P -values is tested using a significance
level of 0.0001. For this procedure, we used version 2.1.2 of the NIST statistical test suite.
As a result, 187 of the 188 second-level tests based on the proportion of passing sequences
were passed. As for the second-level tests based on the uniformity of P -values, all 188 tests
were passed.
TestU01 [27] offers several predefined sets of tests, including SmallCrush, Crush, and
BigCrush, which consist of 15, 144, and 160 tests, respectively. In TestU01, the result of
each test is interpreted as clear failure if the P -value for the test is less than 10−10 or greater
than 1 − 10−10 . The result is interpreted as suspicious if the P -value is in [10−10 , 10−4 ) or
(1 − 10−4 , 1 − 10−10 ]. In all other cases, the test is considered as passed. Using version 1.2.3
of TestU01, we applied SmallCrush, Crush, and BigCrush to the test data described above.
10
TABLE I: Results of statistical testing.
Statistical testing
Number of:
Tests Passed tests Suspicious tests Failed tests
DIEHARD First-level tests
234
232
—
2
6
6
—
0
Second-level tests (proportion)
188
187
—
1
Second-level tests (uniformity)
188
188
—
0
SmallCrush
15
15
0
0
Crush
144
144
0
0
BigCrush
160
160
0
0
Second-level tests
NIST STS
TestU01
As a result, all tests of SmallCrush, Crush, and BigCrush were passed.
Consequently, all tests were passed for NIST’s second-level testing based on the uniformity
of P -values and TestU01’s SmallCrush, Crush, and BigCrush, while a few tests were failed
for DIEHARD and NIST’s second-level testing based on the proportion of passing sequences.
Note that the numbers of failed tests (i.e., two for DIEHARD and one for NIST’s second-level
testing based on the proportion of passing sequences) are within relevant ranges because they
are close to the expected numbers of failed tests (i.e., 2.40 for DIEHARD and 0.29 for NIST’s
second-level testing based on the proportion of passing sequences). From these results, we
can confirm that our generator has good statistical properties.
B.
Comparison with the Mersenne Twister MT19937
Here we attempt a comparison between our generator and MT19937.
As described in Sec. I, MT19937 is a highly practical generator that produces, at very high
speed, a pseudorandom sequence having a period of length 219937 − 1 and a 623-dimensional
equidistribution property. In spite of these marked characteristics, this generator has been
reported to fail linear complexity tests and birthday spacings tests with specific lags [27–29].
This is due to the fact that the generator is based on a linear recurrence over the two-element
field F2 = {0, 1}.
MT19937 generates a sequence of 32-bit unsigned integers. In the following, we will
11
identify a 32-bit unsigned integer with an element of F32
2 . Also, we will not distinguish
between row and column vectors except in that a vector postmultiplying a matrix will be
regarded as a column vector. MT19937 is one of the multiple-recursive matrix methods
[30, 31], and any sequence {yn }n=0,1,2,··· in F32
2 generated by MT19937 obeys the following
recurrence relation (cf. Ref. [29]):
yn = yn−227 + Ayn−623 + Byn−624 , n ≥ 624,
(4)
where y0 , y1 , · · · y623 are initial vectors, and A and B are 32 × 32 matrices with elements in
F2 . The explicit forms of A and B are given in Appendix A.
From Eq. (4), we can grasp the regularity of the sequence generated by MT19937. For
example, the most significant 8 bits of yn and those of yn−227 coincide if an integer n with
n ≥ 624 satisfies the following two conditions:
(a) The inner product of the ith row vector of A and yn−623 equals zero for every i with
1 ≤ i ≤ 8.
(b) The inner product of the second row vector of B and yn−624 equals zero.
Note that condition (b) is equivalent to the condition that Byn−624 = 0 (cf. the form of
P8
8−i
B in Appendix A). Let yn = (yn,1, yn,2, · · · , yn,32 ) ∈ F32
for n ≥ 0.
2 and Yn =
i=1 yn,i 2
We generated a sequence {yn }n=0,1,2,··· ,312499 of 32-bit unsigned integers using MT19937 [32],
and plotted, in Fig. 2, the points (Yn−227 , Yn ) for n satisfying conditions (a) and (b). All
the points are on the diagonal line Yn = Yn−227 , but, obviously, this cannot happen with a
random sequence.
On the other hand, our generator produced a binary sequence of length 107 , using
(0, 1, −1) ∈ S̄ as an initial point. Then, by partitioning it into nonoverlapping binary subsequences of length 32, we transformed it into a sequence {yn }n=0,1,2,··· ,312499 of 32-bit unsigned
integers. Also for this {yn }n=0,1,2,··· ,312499 , we plotted, in Fig. 2, the points (Yn−227 , Yn ) for n
satisfying conditions (a) and (b), which was similar to what we did for {yn }n=0,1,2,··· ,312499 obtained by MT19937. We can see from Fig. 2 that the points obtained from our pseudorandom
sequence are almost uniformly distributed on the square. Although the computational cost
of our generator is significantly higher than that of MT19937, our pseudorandom sequence
displays the same behavior as true (uniform) random sequences.
12
250
200
Yn
150
100
50
0
0
50
100 150
Yn-227
200
250
FIG. 2: Plot of the points (Yn−227 , Yn ) for n satisfying conditions (a) and (b). Dots represent points
obtained from a pseudorandom sequence produced by our generator. Crosses represent those by
MT19937.
V.
CONCLUSION
In this paper, we have introduced a pseudorandom number generator using chaotic true
orbits of the Bernoulli map on cubic algebraic integers. Although this generator has a high
computational cost, it exactly simulates the Bernoulli map that can generate ideal random
binary sequences. We also have clarified a seed selection method that can select initial
points (i.e., seeds) without bias and can avoid overlaps in latter parts of the pseudorandom
sequences derived from them. Moreover, we have obtained experimental results supporting
the conjecture that the initial point sets I0,c with c ∈ Z>0 have a more desirable property
such that each element of I0,c belongs to a different cubic field. In order to demonstrate the
capabilities of our generator, we have performed two kinds of computer experiments: Firstly,
we have tested our generator using three statistical test suites—DIEHARD, NIST statistical
test suite, and TestU01—and have shown that it has good statistical properties. Secondly,
we have examined the independence property of pseudorandom numbers and have clarified
an advantage that our generator has over what is probably the most popular generator, the
Mersenne Twister MT19937.
13
Acknowledgments
We thank Shigeki Akiyama, Shunji Ito, Teturo Kamae, Jun-ichi Tamura, Shin-ichi Yasutomi, and Masamichi Yoshida for their suggestions. This research was supported by JSPS
KAKENHI Grant Number 15K00342.
14
Appendix A: Explicit forms of matrices A and B in Eq. (4)
In this Appendix, we provide the explicit forms of matrices A and B in the recurrence
relation (4) for the Mersenne Twister MT19937.
Matrix A:
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
1
1
0
1
1
1
1
1
0
0
1
0
1
0
1
1
1
1
0
1
0
1
0
1
0
1
0
1
1
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0
1
1
1
1
1
0
1
0
0
1
1
1
1
1
1
0
1
1
1
0
1
1
1
0
0
0
0
0
0
0
0
0
1
1
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
1
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
1
0
0
1
1
0
0
0
0
0
0
0
1
1
1
0
1
1
0
0
1
0
1
1
0
0
1
1
1
1
1
0
1
0
1
0
0
1
1
0
1
1
0
1
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
1
1
1
1
1
1
1
1
1
0
0
0
0
1
0
1
1
1
0
0
1
0
1
0
1
1
1
0
1
0
0
0
1
0
0
0
0
1
0
0
1
0
0
0
1
0
0
1
0
0
1
0
0
0
0
1
0
0
1
0
0
0
1
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
0
0
1
0
1
0
1
1
1
1
0
1
1
1
0
1
0
1
0
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
0
0
1
0
0
0
0
0
0
0
1
1
1
0
1
1
1
0
1
0
1
0
0
1
1
1
1
1
0
0
1
1
1
0
1
0
1
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
1
0
0
0
1
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
1
0
0
0
0
1
0
0
0
0
0
0
1
0
0
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
1
0
0
0
1
1
1
1
1
1
1
1
0
1
0
0
0
0
1
1
1
1
1
0
0
1
0
1
0
1
0
1
0
1
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Matrix B:
15
[1] M. Li and P. Vitányi, An Introduction to Kolmogorov Complexity and Its Applications, 2nd
ed. (Springer, New York, 1997).
[2] H. Sugita, Monte Carlo Method, Random Number, and Pseudorandom Number (Mathematical
Society of Japan, Tokyo, 2011).
[3] D.E. Knuth, The Art of Computer Programming, 3rd ed. (Addison-Wesley, Reading, MA,
1998), Vol. 2, Chap. 3.
[4] D. H. Lehmer, in Proc. 2nd Symp. on Large-Scale Digital Calculating Machinery (Harvard
University Press, 1951), p. 141.
[5] S.W. Golomb, Shift Register Sequences (Aegean Park Press, Laguna Hills, CA, 1982).
[6] M. Matsumoto and T. Nishimura, ACM Trans. on Modeling and Computer Simulation 8, 3
(1998).
[7] P. Billingsley, Ergodic Theory and Information (Wiley, New York, 1965).
[8] E. Atlee Jackson, Perspectives of Nonlinear Dynamics (Cambridge University Press, Cambridge, 1991), Vol. 1, Chap. 4.
[9] A. Saito, Prog. Theor. Phys. Supplement 161, 328 (2006).
[10] For a similar reason, there has also been no proposed pseudorandom number generator using
the tent map on [0, 1] given by MT (x) = 1 − |2x − 1| or the baker’s transformation on [0, 1)2
given by
y
2x,
if x ∈ [0, 1/2)
Mb (x, y) = 2 y + 1
,
2x − 1,
if x ∈ [1/2, 1)
2
although these maps, together with MB , provide (literally) textbook examples of chaotic maps.
[11] A. Saito and A. Yamaguchi, Chaos 26, 063122 (2016).
[12] S.M. Ulam and J. von Neumann, Bull. Amer. Math. Soc. 53, 1120 (1947).
[13] T.Y. Li and J.A. Yorke, Nonlinear Anal. 2, 473 (1978).
[14] S. Oishi and H. Inoue, Trans. IECE Japan E65, 534 (1982).
[15] S. Lang and H. Trotter, J. Reine Angew. Math. 255, 112 (1972); Addendum, J. Reine Angew.
Math. 267, 219 (1974).
[16] F. Vivaldi, Nonlinearity 5, 941 (1992).
16
[17] J.H. Lowenstein, G. Poggiaspalla, and F. Vivaldi, Dynamical Systems 20, 413 (2005).
[18] S. Akiyama, Actes des rencontres du CIRM 1, 3 (2009).
[19] M. Furukado, S. Ito, A. Saito, J. Tamura, and S. Yasutomi, Experimental Math. 23, 390
(2014).
[20] A. Saito and S. Ito, Physica D 268, 100 (2014).
[21] A. Saito, S. Yasutomi, J. Tamura, and S. Ito, Chaos 25, 063103 (2015).
[22] E. Hecke, Lectures on the Theory of Algebraic Numbers (Springer, New York, 1981), Chap. 5.
[23] É. Borel, C. R. Acad. Sci. Paris 230, 591 (1950).
[24] B. Adamczewski and Y. Bugeaud, Annals of Mathematics 165, 547 (2007).
[25] G. Marsaglia, DIEHARD: A battery of tests of randomness, 1996.
[26] A. Rukhin, J. Soto, J. Nechvatal, M. Smid, E. Barker, S. Leigh, M. Levenson, M. Vangel, D.
Banks, A. Heckert, J. Dray, and S. Vo, A Statistical Test Suite for Random and Pseudorandom Number Generators for Cryptographic Applications, NIST Special Publication 800-22
Revision 1a (2010).
[27] P. L’Ecuyer and R. Simard, ACM Trans. Math. Softw. 33, 22 (2007).
[28] F. Panneton, P. L’Ecuyer, and M. Matsumoto, ACM Trans. Math. Softw. 32, 1 (2006).
[29] S. Harase, Math. Comput. Simul. 100, 103 (2014).
[30] H. Niederreiter, Linear Algebra Appl. 192, 301 (1993).
[31] H. Niederreiter, Finite Fields Appl. 1, 3 (1995).
[32] For initialization, we carried out the same procedure as the one adopted in the program available at http://www.math.sci.hiroshima-u.ac.jp/∼m-mat/MT/MT2002/CODES/mt19937ar.c.
17
| 7cs.IT
|
The parameterised complexity of list problems on graphs of
bounded treewidth
Kitty Meeks∗ and Alexander Scott†
Mathematical Institute, University of Oxford, UK
arXiv:1110.4077v3 [cs.CC] 4 Aug 2016
Abstract
We consider the parameterised complexity of several list problems on graphs, with parameter
treewidth or pathwidth. In particular, we show that List Edge Chromatic Number and
List Total Chromatic Number are fixed parameter tractable, parameterised by treewidth,
whereas List Hamilton Path is W[1]-hard, even parameterised by pathwidth. These results
resolve two open questions of Fellows, Fomin, Lokshtanov, Rosamond, Saurabh, Szeider and
Thomassen (2011).
1
Introduction
Many graph problems that are know to be NP-hard in general are fixed parameter tractable when
parameterised by the treewidth k of the graph, that is they can be solved in time f (k) · nO(1) for
some function f . Often there even exists a linear-time algorithm to solve the problem on graphs of
fixed treewidth [2, 6, 10, 28].
This is the case for a number of graph colouring problems. Although it is NP-hard to determine
whether an arbitrary graph is 3-colourable, the chromatic number of a graph of treewidth at most
k (where k is a fixed constant) can be computed in linear time (Arnborg and Proskurowski [2]).
Similarly, while it is NP-hard to determine the edge-chromatic number of cubic graphs (Holyer [19])
and more generally d-regular graphs for any d ≥ 3 (Leven and Galil [24]), this problem can again
be solved in linear time on graphs of bounded treewidth (Zhou, Nakano and Nishizeki [32]). The
pattern is the same for the total colouring problem: this is NP-hard even for regular bipartite graphs
(McDiarmid and Sánchez-Arroyo [27]) but there exists a linear-time algorithm to solve the problem
on graphs of bounded treewidth (Isobe, Zhou and Nishizeki [20]).
However, list versions of such problems cannot always be solved so efficiently on graphs of bounded
treewidth. The well-known list variant of vertex colouring is clearly NP-hard in general (as it generalises the standard colouring problem), but can also be solved in polynomial time on graphs of fixed
treewidth k, and even in linear time on such graphs if the number of colours is bounded (Jansen
and Scheffler [21]). However, in contrast with the linearity result for Chromatic Number, it has
been shown that when the number of colours is unbounded List Colouring is W[1]-hard and
so not (up to certain complexity theoretic assumptions) fixed parameter tractable (Fellows, Fomin,
Lokshtanov, Rosamond, Saurabh, Szeider and Thomassen [14]). The list variants of edge-colouring
and total colouring are harder still on graphs of bounded treewidth: both problems are NP-hard
on series-parallel graphs (Zhou, Matsuo and Nishizeki [31]), which have treewidth at most two, and
∗
Present address:
School of Mathematics
[email protected]; Corresponding author.
†
Email: [email protected]
and
1
Statistics,
University
of
Glasgow;
Email:
list edge-colouring is also NP-hard on outerplanar graphs (Marx [26]), another class of graphs with
treewidth at most two.
Perhaps surprisingly, however, it can be easier to determine the minimum length of lists required
to guarantee the existence of a proper list colouring for a graph G than to determine whether G
admits a proper colouring with a particular set of lists. Alongside the negative result about the
complexity of List Colouring, in [14] the authors also use Courcelle’s theorem [10] to prove the
following result.
Theorem 1.1 ([14]). The List Chromatic Number problem, parameterised by the treewidth bound
t, is fixed-parameter tractable, and solvable in linear time for any fixed t.
In Section 2 we show that the same pattern extends to the list chromatic number problems for
edge-colouring and total colouring: although these problems are both NP-hard on graphs of treewidth
two, it is possible to determine the list edge chromatic number and list total chromatic number of
graphs of bounded treewidth in linear time, answering an open question from [14].
The situation is similar for the problem of determining whether a given graph contains a Hamilton
path. Hamilton Path is known to be computationally difficult in general, remaining NP-hard when
restricted to planar, cubic, 3-connected graphs (Garey, Johnson and Tarjan [17]) or bipartite graphs
(Krishnamoorthy [23]), but can be solved in linear time on graphs of bounded treewidth (Arnborg
and Proskurowski [2]). In Section 3 we consider a list variant of the problem, and show that it is
unlikely to be fixed parameter tractable, even parameterised by pathwidth, answering another open
question from [14].
In the remainder of this section we define formally the problems whose complexity we consider,
and give some background on the treewidth bound and the theory of parameterised complexity.
1.1
Problems considered
A proper edge-colouring of a graph G is an assignment of colours to the edges of G so that no two
incident edges receive the same colour. If a set L = {Le : e ∈ E(G)} of lists of permitted colours
is given, a proper list edge-colouring of (G, L) is a proper edge-colouring of G in which each edge e
receives a colour from its list Le . The list edge chromatic number of G, ch0 (G), is the smallest integer
c such that, for any assignment of colour lists to the edges of G in which each list has length at least
c, there exists a proper list edge-colouring of G. We define the following problem.
List Edge Chromatic Number
Input: A graph G = (V, E).
Question: What is ch0 (G)?
A proper total colouring of a graph G = (V, E) is an assignment of colours to the vertices and
edges of G such that no two adjacent vertices or incident edges have the same colour, and no edge
has the same colour as either of its endpoints. If a set L = {Lx : x ∈ V ∪ E} of lists of permitted
colours is given, a proper list total colouring of (G, L) is a proper total colouring of G in which every
vertex and edge receives a colour from its list. The list total chromatic number of G, chT (G), is the
smallest integer c such that, for any assignment of colour lists to the vertices and edges of G in which
each list has length at least c, there exists a proper list total colouring of G. We define the following
problem.
List Total Chromatic Number
Input: A graph G = (V, E).
Question: What is chT (G)?
2
We also consider a list version of Hamilton Path, introduced in [14], where each vertex has a list
of permissible positions on the path.
List Hamilton Path
Input: A graph G = (V, E), and a set of lists L = {Lv ⊆ {1, . . . , |V |} : v ∈ V } of permitted
positions.
Question: Does there exist a path P = P [1] . . . P [|G|] in G such that, for 1 ≤ i ≤ |G|, we have
i ∈ LP [i] ?
This introduction of lists to the Hamilton path problem is perhaps most naturally interpreted
as the addition of timing constraints; we illustrate this with the Itinerant Lecturer Problem, which
we define as follows. A professor has recently proved an exciting new result, and plans to spend
n days travelling the country, giving a lecture about his work at a different university each day.
We represent each institution he plans to visit with a vertex; each pair of institutions which can
potentially be visited on consecutive days (the distance not being too great) is connected by an edge.
Each institution also places different restrictions on which days he can visit (perhaps they require
that he comes on the correct day of the week to speak at their regular seminar, or it might be that his
host is away for part of the month) which give rise to the list of permitted positions for each vertex.
The Itinerant Lecturer is then seeking to find a solution to the List Hamilton Path problem for
this input.
1.2
Treewidth and Pathwidth
We consider the complexity of these problems restricted to graphs of bounded treewidth or pathwidth.
Given a graph G, we say that (T, D) is a tree decomposition of G if T is a tree and D = {D(t) : t ∈ T }
is a collection of non-empty subsets of V (G) (or bags), indexed by the nodes of T , satisfying:
S
1. V (G) = t∈T D(t),
2. for every e = uv ∈ E(G), there exists t ∈ T such that u, v ∈ D(t),
3. for every v ∈ V , if T (v) is defined to be the subgraph of T induced by nodes t with v ∈ D(t),
then T (v) is connected.
The width of the tree decomposition (T, D) is defined to be maxt∈T |D(t)|−1, and the treewidth of G is
the minimum width over all tree decompositions of G. A path decomposition is a tree decomposition
(P, D) in which the indexing tree is a path, and the pathwidth of G is the minimum width over all
path decompositions of G.
Given a tree decomposition (T, D) of G, we assume that an arbitrary node r ∈ V (T ) is chosen
to be the root of T , and define the height, h(t) of any t ∈ T to be the distance from r to t. For any
v ∈ V (G), we then define tv to be the unique node t of minimum height such that v ∈ D(t) (i.e. tv
is the node of minimal height in the subtree T (v)).
We will make use of the following well-known bound on the number of nodes required in the tree
indexing the tree decomposition of a graph G (see [15, Lemma 11.9] for a proof); we shall assume
throughout that tree decompositions of this form are given.
Lemma 1.2. Let G be a graph of order n and treewidth k. Then there exists a width k tree decomposition (T, D) for G with |T | ≤ n. Moreover, given any tree decomposition (T 0 , D0 ) for G, a tree
decomposition (T, D) such that |T | ≤ n can be computed in linear time.
Graphs having treewidth at most k can alternatively be characterised as partial k-trees, as in
[2, 32]. It follows immediately from this equivalent definition that if a graph G of order n is a partial
k-tree (i.e. it has treewidth at most k) then G has at most kn edges.
3
1.3
Parameterised complexity
When considering the class of problems which are solvable in polynomial time on graphs of treewidth
at most k, we aim, as mentioned above, to distinguish those which can be solved in time f (k) ·
nO(1) . Problems in this subclass are said to be fixed parameter tractable, parameterised by treewidth
(and belong to the parameterised complexity class FPT). A standard method of showing that a
parameterised problem does not belong to FPT (and so the best known algorithm has running time
nf (k) for some unbounded function f ) is to prove that it is hard for some level of the W-hierarchy;
in Section 3 we show that List Hamilton Path is hard for W[1], the first level of this hierarchy.
In order to show that a problem is W[1]-hard, we give a parameterised reduction from a problem
that is known to be W[1]-hard. Given a W[1]-hard decision problem Π with parameter k, this involves
constructing an instance (I 0 , k 0 ) of Π0 such that (I 0 , k 0 ) is a yes-instance for Π0 if and only if (I, k) is
a yes-instance for Π; for a parameterised reduction we require that k 0 is bounded by some function
of k, in addition to the fact that (I 0 , k 0 ) can be computed from (I, k) in time polynomial in |I| and
that |I 0 | is bounded by a polynomial function of |I|.
One useful W[1]-hard problem, which we use for a reduction in Section 3, is Multicolour Clique
(shown to be W[1]-hard by Fellows, Hermelin, Rosamond and Vialette in [13]): given a graph G,
properly coloured with k colours, the problem is to determine whether there exists a clique in G
containing one vertex of each colour.
In fact, by considering the relationship between the parameters k and k 0 in the parameterised
reduction, it is sometimes possible to prove conditional lower bounds on the running time of any
algorithm for a specific parameterised problem. Assuming the exponential time hypothesis, a lower
bound on the running time for any algorithm to solve the Clique problem is known.
Theorem 1.3 ([9, 25]). Assuming the exponential time hypothesis, there is no f (k)no(k) algorithm
for Clique.
Since the reduction from Clique to Multicolour Clique in [13] does not change the value
of the parameter, the same is true for Multicolour Clique; our parameterised reduction to List
Hamilton Path in Section 3 only increases the parameter value by a constant factor, so we are
able to deduce that the same conditional lower bound also holds for List Hamilton Path.
For further background on the theory of parameterised complexity, we refer the reader to [12].
2
List Chromatic Number Problems
The main result of this section is the following theorem.
Theorem 2.1. List Edge Chromatic Number and List Total Chromatic Number are
fixed parameter tractable, parameterised by the treewidth bound k, and are solvable in linear time for
any fixed k.
The key technical tool we use to prove this result is Theorem 2.8, which determines the list edge
chromatic number and list total chromatic number for graphs with fixed treewidth and large maximal
degree.
2.1
Background
We begin by recalling some existing results about the edge chromatic number and list edge chromatic
number of a graph. It is easy to see that, for any graph G, we have
∆(G) ≤ χ0 (G) ≤ ch0 (G) ≤ 2∆(G) − 1,
4
where χ0 (G) denotes the edge chromatic number of G and ∆(G) is the maximum degree of G. The
lower bound comes from the fact that every edge incident with a single vertex must receive a different
colour; for the upper bound, note that every edge is incident with at most 2(∆(G) − 1) others, so
lists of length 2∆(G) − 1 guarantee that we can colour greedily. For the edge chromatic number, we
have the much stronger result of Vizing:
Theorem 2.2 ([29]). χ0 (G) is equal to either ∆(G) or ∆(G) + 1.
In order to give a linear time algorithm to solve Edge Chromatic Number on graphs of
bounded treewidth, Zhou, Nakano and Nishizeki [32] prove that, for graphs of fixed treewidth k and
maximum degree ∆ ≥ 2k, the edge chromatic number must in fact be equal to ∆.
There is no direct analogue of Vizing’s theorem for the list edge chromatic number. The List
(Edge) Colouring Conjecture (discussed in [1, 5, 18]) asserts that χ0 (G) = ch0 (G) for any graph G,
and would immediately imply Vizing’s conjecture (1976) that ch0 (G) ≤ ∆(G) + 1. However, neither
of these conjectures has been proved except for certain special classes of graphs, and the best general
bound on the list edge chromatic number is due to Kahn.
Theorem 2.3 ([22]). For any > 0, if ∆(G) is sufficiently large,
ch0 (G) ≤ (1 + )∆(G).
In Theorem 2.8, we show that for a graph G of bounded treewidth and large maximum degree,
ch0 (G) = χ0 (G) = ∆(G),
proving a special case of the List (Edge) Colouring Conjecture. Using this result, the List Edge
Chromatic Number problem on graphs of bounded treewidth can be reduced to the case in which
the maximum degree of the graph is bounded.
We prove an analogous result for the list total chromatic number. Once again, there exist trivial
bounds for the (list) total chromatic number of an arbitrary graph:
∆(G) + 1 ≤ χT (G) ≤ chT (G) ≤ 2∆(G) + 1,
where χT (G) denotes the total chromatic number of the graph. It is a long-standing but unproved
conjecture (the Total Colouring Conjecture [3, 30]) that
χT (G) ≤ ∆(G) + 2.
Theorem 2.8 further demonstrates that, for a graph G of bounded treewidth and large maximum
degree,
chT (G) = χT (G) = ∆(G) + 1,
and so once again it suffices to solve the problem for graphs with bounded maximum degree.
Of course, there is a correspondence between these colouring problems and the vertex-colouring
problems discussed above. For any graph G = (V, E), the line graph L(G) of G has vertex set E, and
e, f ∈ E are adjacent in L(G) if and only if e and f are incident in G. Then solving (for example) List
Edge Chromatic Number for the graph G is equivalent to solving List Chromatic Number
for L(G). Similarly, we define the total graph T (G) of G to have vertex set V ∪ E, and edge set
E ∪ {ef : e, f ∈ E and e, f incident in G} ∪ {ve : v is an endpoint of e}, and solving List Total
Chromatic Number for G is then equivalent to solving List Chromatic Number for T (G).
However, as the treewidth of L(G) or T (G) can in general be arbitrarily large even when G itself
has small treewidth, results about the parameterised complexity of vertex colouring problems do not
immediately transfer to the edge and total colouring cases.
If the maximum degree of G is bounded, however, the following result (proved in [11]) tells us
that the treewidth of L(G) is bounded by a constant multiple of that of G.
5
Lemma 2.4. Let G be a graph of treewidth at most k, and maximum degree at most ∆. Then L(G)
has treewidth at most (k + 1)∆.
A similar result holds for the treewidth of T (G).
Lemma 2.5. Let G be a graph of treewidth at most k, and maximum degree at most ∆. Then T (G)
has treewidth at most (k + 1)(∆ + 1).
Proof. If (T, D) is a width k tree decomposition for G, it is easy to verify that (T, D0 ), where
D0 (t) = D(t) ∪ {uv ∈ E : {u, v} ∩ D(t) 6= ∅}, is a tree decomposition of T (G) of width at most
(k + 1)(∆ + 1).
We will need two further results for our proofs. First, a theorem of Galvin concerning the list
edge chromatic number of bipartite graphs:
Theorem 2.6 ([16]). Let G be a bipartite graph. Then
ch0 (G) = χ0 (G) = ∆(G).
Finally, we make use of an algorithm of Bodlaender:
Theorem 2.7 ([4]). For all k ∈ N, there exists a linear-time algorithm that tests whether a given
graph G = (V, E) has treewidth at most k and, if so, outputs a tree-decomposition of G with treewidth
at most k.
2.2
Results and Proofs
In this section we prove our technical results about the list edge chromatic number and list total
chromatic number of graphs with bounded treewidth and large maximum degree, and then give
linear-time algorithms to solve List Edge Chromatic Number and List Total Chromatic
Number on graphs of bounded treewidth.
Our algorithms rely on the following theorem, which is a special case of both the List (Edge)
Colouring Conjecture and the Total Colouring Conjecture.
Theorem 2.8. Let G be a graph with treewidth at most k and ∆(G) ≥ (k + 2)2k+2 . Then ch0 (G) =
∆(G) and chT (G) = ∆(G) + 1.
This result can be derived from earlier work by Borodin, Kostochka and Woodall [7, Theorem 7]
(with a quadratic rather than exponential dependence on k). For the sake of completeness, however,
we will give a direct, self-contained proof of this result. Moreover, since the dissemination of an initial
version of this paper, the techniques introduced in our proof have been refined by Bruhn, Lang and
Stein [8] to demonstrate that only a linear dependence on k is in fact required.
We prove the result by means of two lemmas, concerning the list edge chromatic number and list
total chromatic number respectively; Theorem 2.8 follows immediately from these auxiliary results.
We start by considering the list edge chromatic number.
Lemma 2.9. Let G be a graph with treewidth at most k. Then ch0 (G) ≤ max{∆(G), (k + 2)2k+2 }.
Proof. We proceed by contradiction. Suppose that the result does not hold, and let G be a counterexample with as few edges as possible, so there exists a set L = (Le )e∈E(G) of colour-lists, all of
length ∆0 = max{∆(G), (k + 2)2k+2 }, such that there is no proper list edge-colouring of (G, L). We
may assume without loss of generality that G contains no isolated vertices and so, by edge-minimality
of G, we must have ch0 (G0 ) ≤ max{∆(G0 ), (k + 2)2k+2 } ≤ ∆0 for any proper subgraph G0 of G.
6
We may assume that every edge e ∈ E(G) is incident with at least ∆0 others: if e is incident with
fewer than ∆0 other edges then we can extend any proper list edge-colouring of (G − e, L \ {Le }) to
a proper list edge-colouring of (G, L). We will show that, under this assumption, there must exist a
nonempty set of vertices U such that any proper list edge-colouring of (G \ U, (Le )e∈E(G\U ) ) can be
extended to a proper list edge-colouring of (G, L), contradicting the choice of G as an edge-minimal
counterexample (as the fact there are no isolated vertices in G means e(G \ U ) < e(G)).
Let us define L ⊆ V (G) to be the set of vertices of degree at least ∆0 /2 (the set of vertices of
“large” degree), and note that every edge is incident with at least one vertex from L (as otherwise
it can be incident with only ∆0 − 1 other edges). Thus S = V (G) \ L (the set of vertices of “small”
degree) is an independent set. Fix a width k tree decomposition (T, D) of G, and choose v ∈ L such
that h(tv ) = maxx∈L h(tx ). We then set T 0 to be the subtree of T rooted at tv , that is the subgraph
of T induced by nodes u S
such that the path from u to the root contains tv .
Set X ⊆ V (G) to be t∈T 0 D(t), and X 0 = X \ D(tv ). We then make the following observations
(where, for any U ⊆ V (G), Γ(U ) denotes the set of neighbours of vertices in U ).
1. L ∩ X ⊆ D(tv ): if any vertex z ∈ L appears in a bag indexed by T 0 but does not appear in
D(tv ), we must have h(tz ) > h(tv ), contradicting the choice of v. This implies immediately
that X 0 ⊆ S.
2. Γ(X 0 ) ⊆ D(tv ): no vertex from X 0 can appear in a bag of the decomposition not indexed by
T 0 , so clearly we have Γ(X 0 ) ⊆ X; but also, as X 0 ⊂ S, we have Γ(X 0 ) ⊆ L and so we see
Γ(X 0 ) ⊆ L ∩ X ⊆ D(tv ).
3. |X 0 | ≥ ∆0 /2 − k: since v appears only in bags indexed by T 0 , we have Γ(v) ⊆ X \ {v},
implying |X| − 1 ≥ d(v) ≥ ∆0 /2, and we know |X| − |X 0 | = |D(tv )| ≤ k + 1, so we have
|X 0 | ≥ |X| − k − 1 ≥ ∆0 /2 − k.
As the neighbourhood of any vertex x ∈ X 0 is contained in D(tv ), there are at most 2k+1 possibilities
for the neighbourhood of such a vertex. Therefore there must exist some subset U ⊆ X such that
every vertex in U has the same neighbourhood, and
|U | ≥
∆0 /2 − k
(k + 2)2k+1 − k
|X 0 |
≥
≥
≥ k + 1 ≥ |Γ(U )|.
2k+1
2k+1
2k+1
Now let φ be a proper edge-colouring of (G \ U, (Le )e∈E(G\U ) ). If we can extend φ to a proper
edge-colouring of G in which every edge incident with U also receives a colour from its list, then we
have a proper list edge-colouring of (G, L), giving the required contradiction.
Set W = Γ(U ), say W = {w1 , . . . , wr } (where r ≤ |U |), and let H be the complete bipartite
subgraph of G induced by U ∪ W . Suppose, for 1 ≤ i ≤ r, that Fi is the set of colours already used
by φ on edges incident with wi , and for each u ∈ U define the list L0uwi to be Luwi \ Fi . If we can
properly colour the edges of H in such a way that each edge e ∈ E(H) is given a colour from L0e ,
then we can extend φ as required.
Observe that, for each i, |Fi | ≤ ∆(G) − |U |, and so we have |L0uwi | ≥ ∆0 − ∆(G) + |U | ≥ |U |. But
as H is bipartite, with maximum degree |U |, we have (by Theorem 2.6)
ch0 (H) = ∆(H) = |U | .
Therefore, as each list L0e contains at least |U | colours, there exists a proper edge colouring of H in
which every edge e receives a colour from its list L0e , completing the proof.
We use a very similar argument to prove an analogous result for the list total chromatic number.
Lemma 2.10. Let G be a graph with treewidth at most k. Then chT (G) ≤ max{∆(G), (k+2)2k+2 }+1.
7
Proof. Again, we proceed by contradiction. Let ∆0 = max{∆(G), (k + 2)2k+2 }. Suppose that the
result does not hold, and let G be a counterexample with as few edges as possible, so there exists a
set L = (Lx )x∈V (G)∪E(G) of colour-lists, all of length ∆0 + 1, such that there is no proper list total
colouring of (G, L). We may assume without loss of generality that G contains no isolated vertices
and so, by edge-minimality of G, we must have chT (G0 ) ≤ max{∆(G0 ), (k + 2)2k+2 } + 1 ≤ ∆0 + 1 for
any proper subgraph G0 of G.
We may assume that every edge e ∈ E(G) is incident with at least ∆0 − 1 others: if e is incident
with fewer than ∆0 − 1 other edges then at most ∆0 colours can appear on vertices or edges incident
with e, and so any proper list total colouring of (G − e, L \ {Le }) can be extended to a proper
list total colouring of (G, L). Thus, if we define L0 ⊆ V (G) to be the set of vertices of degree at
least (∆0 − 1)/2, every edge is incident with at least one vertex from L0 , and S 0 = V (G) \ L0 is an
independent set.
Exactly as in the proof of Lemma 2.9, we can find a subset U ⊆ S 0 such that all vertices of U
have the same neighbourhood of size at most k + 1, and
(∆0 − 1)/2 − k
2k+1
((k + 2)2k+2 − 1)/2 − k
≥
2k+1
>k+1
≥ |Γ(U )|.
|U | ≥
As G \ U is a proper subgraph of G, there exists a proper list total colouring φ of (G \ U, L0 )
where L0 = (Lx )x∈E(G\U )∪V (G\U ) . If we can extend φ to a proper total colouring of G in which every
vertex from U and every edge incident with U also receives a colour from its list, then we have a
proper list total colouring of (G, L), giving the desired contradiction.
As in the proof of Lemma 2.9, set W = Γ(U ), say W = {w1 , . . . , wr } (where r ≤ |U |), and let
H be the complete bipartite subgraph of G induced by U ∪ W . Suppose, for 1 ≤ i ≤ r, that Fi is
the set of colours already used by φ on wi and the edges incident with wi , and for each u ∈ U define
L0uwi to be Luwi \ Fi .
Observe that, for each i, |Fi | ≤ 1 + ∆(G) − |U |, and so we have
|Luwi | ≥ ∆0 + 1 − 1 − ∆(G) + |U | ≥ |U |.
As H is bipartite, with maximum degree |U |, we have (by Theorem 2.6) ch0 (H) = ∆(H) = |U |.
Therefore, as each list L0e contains at least |U | colours that are not already used by φ on edges or
vertices incident with e, we can extend φ to a proper list colouring φ0 including the edges incident
with U . If we can then colour the vertices of U in such a way that no u ∈ U receives a colour used
by φ0 on an adjacent vertex or incident edge, we can indeed extend φ to a proper list total colouring
of G.
Note that every u ∈ U has degree at most k + 1, and so is adjacent to or incident with at most
2(k + 1) vertices and edges of G. Thus, as |Lu | ≥ ∆0 + 1 > 2(k + 1), there is at least one colour
in Lu that is not used by φ0 on a vertex adjacent to u or on an edge incident with u, and we can
extend φ0 to the vertices of U greedily. This gives a proper list total colouring of (G, L), completing
the proof.
Together, Lemmas 2.9 and 2.10 prove Theorem 2.8. We can now prove our main complexity
result, based on Theorem 2.8.
Proof of Theorem 2.1. Let G be a graph on n vertices, with treewidth at most k, and set f (k) =
(k + 2)2k+2 . We can check in time O(f (k)n) whether ∆(G) ≥ f (k), and if this is the case then, by
8
Theorem 2.8, we know the exact value of ch0 (G) and chT (G). Thus it suffices to solve both problems
in the case that ∆(G) < f (k).
This is exactly the same as solving List Chromatic Number on L(G) or T (G), when ∆(G) <
f (k). But in this case, by Lemmas 2.4 and 2.5, L(G) and T (G) have bounded treewidth. Note that
both graphs can be computed from G in time
O(e(L(G)) + e(T (G))) = O((k + 1)(f (k) + 1)(|T (G)| + |L(G)|)
= O(k 2 f (k)|G|),
and so we can then use Bodlaender’s algorithm (Theorem 2.7) to find a tree decomposition of L(G)
or T (G), of width at most (k + 1)(f (k) + 1), in time O(|G|) for any fixed k. Given this decomposition
we can, by Theorem 1.1, compute the list chromatic number of L(G) or T (G), and hence the list
edge chromatic number or list total chromatic number of G, in linear time.
Our proof also implies a polynomial-time algorithm to compute a proper list edge-colouring of
any graph G of fixed treewidth and large maximum degree, provided every L ∈ L has length at least
∆(G). The same method can also be used to compute a list total colouring of such a graph, provided
every list has length at least ∆(G) + 1. Here we describe an O(n2 ) algorithm to achieve this; with
the use of suitable data structures this can probably be improved to a linear-time algorithm, but for
simplicity of presentation we do not seek to optimise the running time here.
Theorem 2.11. Let G be a graph of order n and treewidth k, with maximum degree at least (k +
2)2k+3 , and let L = (Le )e∈E(G) be a set of colour-lists such that |Le | ≥ ∆(G) for all e ∈ E(G). Then,
for fixed k, we can compute a proper list edge-colouring of (G, L) in time O(n2 ).
Proof. The idea of the algorithm is to delete edges or sets of vertices repeatedly, as in the proof of
Lemma 2.9, until the degree of the remaining graph is less than (k + 2)2k+2 < ∆(G)/2, so the edges
of this graph can be list-coloured greedily. Edges and vertices are then reinserted and coloured to
extend this colouring to the edges of G. Unlike in the proof of Lemma 2.9, where we identified a set
of at least k + 1 vertices to delete, in this algorithm we always delete a set of exactly k + 1 vertices.
Without loss of generality, we may also assume that every list Le has length exactly ∆(G) < n,
discarding additional colours if necessary.
We begin with some preprocessing. For each vertex, we construct a list of its neighbours, and
we also construct an n-element array storing the degree of each vertex. For each vertex of degree at
most k + 1, its list of neighbours is sorted into order; the list of vertices of degree at most k + 1 is
then also sorted, in order of neighbourhoods (so that vertices with the same neighbourhood occur
consecutively). All this can be done in time O(n2 ).
Note that each time we delete an edge, we can update this information in time O(n): we update
the neighbour lists of the edge’s endpoints and decrement their degrees, and if one or both of the
endpoints now has degree at most k + 1 its neighbours are sorted and it is inserted into the correct
position in the list of small-degree vertices. From the point of view of updating information, deleting
a set of k + 1 vertices can be regarded as a series of edge-deletions, each performed in time O(n). As
there are O(n) edges in total, this means we can perform all updates after deletions in time O(n2 ).
Given the array of degrees of all vertices, it is straightforward to identify in time O(n) an edge
incident with fewer than ∆(G) others, if such an edge exists. If there is no such edge, we know from
the proof of Lemma 2.9 that (provided the maximum degree of the graph is still at least (k + 2)2k+2 )
there exists a set of k +1 vertices with a common neighbourhood of size at most k +1. As the vertices
of degree at most k + 1 are sorted by their neighbourhoods, it is also possible to identify such a set of
vertices in linear time. Thus at each step we are able to identify the edge or set of vertices to delete
in time O(n), and so all deletions (and subsequent updating) can be performed in time O(n2 ).
It therefore remains to show that we can also perform the reinsertions in time O(n2 ). When we
reinsert an edge, we can simply colour it with the first available colour from its colour-list, taking
9
time O(n). When reinserting a set of k +1 vertices, we need to colour up to (k +1)2 edges which form
a complete bipartite subgraph. Recall from the proof of Lemma 2.9 that every such edge still has
at least k + 1 available colours from its list (i.e. colours that have not already been used on incident
edges), and that there exists a proper list edge-colouring of the bipartite graph if every edge has a
list of exactly k + 1 permitted colours. For each of the edges, we can compute in time O(kn) a list
of k + 1 colours from its colour-list which have not yet been used on incident edges. We can then
check in constant time all possible colourings of the (k + 1)2 edges in which each receives one of the
k + 1 colours from its list, to find a proper colouring of this bipartite graph, which is guaranteed to
extend the list edge-colouring as required.
Thus we can perform all reinsertions in time O(n2 ), completing the proof of the theorem.
3
List Hamilton Path
Recall from Section 1.1 the List Hamilton Path (or Itinerant Lecturer) problem:
List Hamilton Path
Input: A graph G = (V, E), and a set of lists L = {Lv ⊆ {1, . . . , |V |} : v ∈ V } of permitted
positions.
Question: Does there exist a path P = P [1] . . . P [|G|] in G such that, for 1 ≤ i ≤ |G|, we have
i ∈ LP [i] ?
Given a graph G and a set of lists L = {Lv ⊆ {1, . . . , |G|} : v ∈ V (G)} of permitted positions, we
say a path P = P [1] . . . P [|G|] in G is a valid Hamilton path if i ∈ LP [i] for every i. In this section,
we prove the following result.
Theorem 3.1. List Hamilton Path, parameterised by pathwidth, is W[1]-hard. Moreover, under the exponential time hypothesis, there is no algorithm to solve List Hamilton Path in time
f (k)no(k) on graphs of pathwidth k.
We prove our theorem by means of a reduction from Multicolour Clique. The reduction
involves an edge representation strategy: we select an edge by visiting the vertices of the associated
gadget in a specific order, and the permitted positions for other vertices in the graph ensure that the
edges selected in this way must in fact be the edges of a multicolour clique.
Suppose G is the k-coloured, n-vertex graph in an instance of Multicolour Clique: we may
assume without loss of generality that all k vertex classes have the same size, and also that the number
of edges between each pair of vertex classes is the same (as adding isolated edges and vertices does
not change the existence or otherwise of a multicolour clique). Let the vertex classes be V1 , . . . , Vk ,
where Vi contains vertices Vi [1], . . . , Vi [p], and assume that there are q edges between each pair of
vertex classes. For technical reasons, it will be useful to assume (without loss of generality) that
k, n ≥ 4.
We now describe the construction of a graph H of pathwidth at most 5k, and a set of lists
L = (Lv )v∈V (H) such that there is a valid Hamilton path in (H, L) if and only if G contains a
multicolour clique.
Our construction consists of k+1 paths, with some additional edges linking them: paths P1 , . . . , Pk
correspond to the vertex classes V1 , . . . , Vk , andan additional path Q is used to connect P1 , . . . , Pk .
Each path Pi has 2n2 vertices (so |Pi | ≥ 4q k2 = 4e(G)), and we denote the j th vertex of Pi by
Pi [j − 1]. The path Q = Q1 . . . Qk is the concatenation of k subpaths Q1 , . . . , Qk , each containing
n2 (n − 2) vertices and, for 1 ≤ i ≤ k, every vertex of Qi , except for the first vertex of Q1 and the
last vertex of Qk , is adjacent to both Pi [0] and Pi [2n2 − 1].
In addition, we have a number of edge-gadgets, consisting of edges between pairs of the paths
P1 , . . . , Pk . Suppose E(G) = {e0 , . . . , em−1 }. Then, for each edge er between Vi and Vj with i < j,
10
we have an edge-gadget G(er ), involving the rth group of four vertices in Pi and the corresponding
group of vertices from Pj . G(er ) has edges Pi [4r]Pj [4r + 1], Pj [4r]Pi [4r + 1], Pi [4r + 1]Pj [4r + 3] and
Pi [4r + 2]Pj [4r + 2], as illustrated in Figure 1.
Pi [4r]
Pi [4r+1]
Pi [4r+2]
Pi [4r+3]
Pj [4r]
Pj [4r+1]
Pj [4r+2]
Pj [4r+3]
Figure 1: The edge-gadget G(er ), where er has endpoints in Vi and Vj with i < j
This completes the construction of the graph H. We now define the list of permitted positions
for each vertex. For 1 ≤ i ≤ k, we set
LPi [0] = {(i − 1)n3 + 3αn2 : 1 ≤ α ≤ p},
and
LPi [2n2 −1] = {(i − 1)n3 + 3αn2 + (2n2 − 1) + k + 1 − 2i : 1 ≤ α ≤ p}.
We further define the list
L(i, t) = {(i − 1)n3 + 3αn2 + t + β : 1 ≤ α ≤ p, −(k − 1) ≤ β ≤ k − 1}.
For every internal vertex Pi [t], the list LPi [t] will contain L(i, t). In most cases we in fact set LPi [t] =
L(i, t), the only exceptions being three vertices in each edge-gadget G(e): if e = Vi [r]Vj [s] and the
vertices in G(e) are Pi [`], . . . , Pi [` + 3], Pj [`], . . . , Pj [` + 3], then the list for Pi [` + 1] will additionally
contain positions
{(j − 1)n3 + 3sn2 + (` + 1) + β : −(k − 1) ≤ β ≤ k − 1},
while the lists for Pj [` + 1] and Pj [` + 2] also contain
{(i − 1)n3 + 3rn2 + (` + 1) + β : −(k − 1) ≤ β ≤ k − 1}
and
{(i − 1)n3 + 3rn2 + (` + 2) + β : −(k − 1) ≤ β ≤ k − 1}
respectively. The first vertex on Q has singleton list {1} and the last vertex on Q has singleton list
{kn3 }; we place no restriction on the positions that the remaining vertices from Q can take in a valid
Hamilton path.
Intuitively, the idea is that any valid Hamilton path must (with the exception of a few vertices
belonging to the edge-gadgets) traverse P1 , . . . , Pk in that order, using sections of Q before and after
each Pi to connect the paths. In this construction, our choice of position for Pi [0] corresponds to
a choice of vertex from Vi : if Pi [0] takes position (i − 1)n3 + 3rn2 , we say that the Hamilton path
selects Vi [r]. Notice that, for any edge-gadget Ge (where e = Vi [r]Vj [s]), there are precisely two
possibilities for which edges within Ge are used by a Hamilton Path, as illustrated in Figure 2. If
we use the edge Pi [4r]Pj [4r + 1] (and so also use the edges Pj [4r + 2]Pi [4r + 2], Pj [4r]Pi [4r + 1] and
Pi [4r + 1]Pj [4r + 3]) we say that the Hamilton path selects the edge e.
We begin by explaining how, given a multicolour clique in G, we can construct a valid Hamilton
path in H.
11
Pi [4r]
Pi [4r+1]
Pi [4r+2]
Pi [4r+3]
Pj [4r]
Pj [4r+1]
Pj [4r+2]
Pj [4r+3]
Pi [4r]
Pi [4r+1]
Pi [4r+2]
Pi [4r+3]
Pj [4r]
Pj [4r+1]
Pj [4r+2]
Pj [4r+3]
Figure 2: Two possibilities for the edges used by a Hamilton path in the gadget Ge : in the first case,
we say that the path selects e.
Lemma 3.2. If G is a yes-instance for Multicolour Clique, then (H, L) is a yes-instance for
List Hamilton Path.
Proof. Suppose we have a multicolour clique that contains Vi [ri ] for 1 ≤ i ≤ k. We claim we can
then find a Hamilton path in H in which every vertex has a position from its permitted list.
In such a path, we give Pi [0] position (i − 1)n3 + 3ri n2 . Each path Pi is traversed from end to
end in order, except for the selection of precisely k2 edges: for each 1 ≤ i < j ≤ k we will select
the edge Vi [ri ]Vj [rj ] (which is possible since V1 [r1 ], . . . , Vk [rk ] induces a clique. Note that the lists
for each vertex in the gadget corresponding to the edge Vi [ri ]Vj [rj ] will always allow us to select this
edge: if the vertices in the gadget are Pi [`], . . . , Pi [` + 3], Pj [`], . . . , Pj [` + 3], then Pi [` + 1] is allowed
to take any position in {(j − 1)n3 + 3rj n2 + (` + 1) + β : −(k − 1) ≤ β ≤ k − 1}, whereas Pj [` + 1]
can take any position in {(i − 1)n3 + 3ri n2 + (` + 1) + β : −(k − 1) ≤ β ≤ k − 1} and Pj [` + 2] any
position in {(i − 1)n3 + 3ri n2 + (` + 2) + β : −(k − 1) ≤ β ≤ k − 1}.
The number of vertices lying on the path between Pi [0] and Pi [2n2 − 1] will then be equal to
2
2n − 2 + (k − i) − (i − 1) = 2n2 − 1 + k − 2i (since, in comparison with simply traversing Pi , this
segment of the path gains one vertex for each selected edge Vi [ri ]Vj [rj ] where j > i, and loses one
vertex for each selected edge Vi [ri ]Vj [rj ] where j < i). This means that Pi [2n2 − 1] will have position
(i − 1)n3 + 3ri n2 + (2n2 − 1) + k + 1 − 2i, which belongs to its list, and it is straightforward to verify
that every other vertex on each path Pi will receive a position from its list.
Finally, we connect up these path segments using segments of Q, so that each vertex Pi [0] has the
desired position in the final path. If Pi [0] occurs on the path immediately after some vertex Qi [αi ]
then Pi [2n2 − 1] is followed by Qi [α + 1], so every vertex of Q is included on the path; the first vertex
of Q is the first vertex on the path, and the last vertex of Q the last, as required by the lists for these
two vertices.
This path then includes every vertex in H, and gives each vertex a position from its permitted
list, so (H, L) is indeed a yes-instance for List Hamilton Path.
We now proceed to demonstrate that the converse is also true: if H contains a valid Hamilton
Path then G must contain a multicolour clique. In order to demonstrate this fact in Lemma 3.9, we
first prove a number of propositions about the structure of valid Hamilton paths in H. Throughout,
12
we shall assume that P is a valid Hamilton path in H.
Proposition 3.3. For any vertex Pi [`] in H, there is a unique pair of values (fP (Pi [`]), gP (Pi [`])) with
fP (Pi [`]) ∈ {1, . . . , k} and gP (Pi [`]) ∈ {1, . . . , p} such that Pi [`] takes position (fP (Pi [`]) − 1) · n3 + 3 ·
gP (Pi [`])·n2 +t+β for some t ∈ {0, . . . , 2n2 −1} and β ∈ {−(k−1), . . . , (k−1)}. Moreover, if vertices
Pi [`] and Pj [`0 ] occur consecutively on P , then fP (Pi [`]) = fP (Pj [`0 ]) and gP (Pi [`]) = gP (Pj [`0 ]).
Proof. It is straightforward to show that, if a1 6= a2 or b1 6= b2 , then, for any choice of t1 , t2 ∈
{0, . . . , 2n2 − 1} and β1 , β2 ∈ {−(k − 1), . . . , (k − 1)}, we have
(a1 · n3 + 3 · b1 · n2 + t1 + β1 ) − (a2 · n3 + 3 · b2 · n2 + t2 + β2 ) > 1,
which immediately implies the result.
Proposition 3.4. For each i ∈ {1, . . . , k}, the set of vertices fP−1 (i) occurs consecutively on P , and
there is a unique φ(i) ∈ {1, . . . , p} such that gP (Pi [`]) = φ(i) for every Pi [`] ∈ fP−1 (i).
Proof. Let us call a vertex belonging to the path Pi for some 1 ≤ i ≤ k a p-vertex, and a vertex
belonging to the path Q in H a q-vertex. The Hamilton path P can then be decomposed into
segments, where each segment is a maximal set of consecutive vertices of the same kind (either pvertices or q-vertices). Note that the first and last segments of P must consist of q-vertices, since we
know that the first vertex on Q must be the first vertex of P , and the last vertex of Q must be the
last vertex of P .
Now consider an arbitrary segment of p-vertices, and note that each of its endpoints must be
adjacent in P to a q-vertex. Notice that the set of p-vertices that have at least one q-vertex as a
neighbour is precisely {Pi [0], Pi [2n2 − 1] : 1 ≤ i ≤ k}, so the endpoints of every segment of p-vertices
must lie in this set. Since two distinct segments of p-vertices must have disjoint pairs of endpoints
in this set, and |{Pi [0], Pi [2n2 − 1] : 1 ≤ i ≤ k}| = 2k, it follows that there can be at most k distinct
segments of p-vertices. Moreover, by Proposition 3.3, all p-vertices belonging to the same segment
must map to the same value under both f and g.
Notice that, by definition of the lists for Pi [0] and Pi [2n2 −1], we have f (Pi [0]) = f (Pi [2n2 −1]) = i.
Thus, for each 1 ≤ i ≤ k, there must be precisely one segment, Si , of p-vertices such that fP (v) = i
for every vertex v belonging to the segment. It follows that the vertices of fP−1 (i) are exactly those of
Si , so all vertices of fP−1 (i) occur consecutively on P . Moreover, it therefore follows from Proposition
3.3 that there exists some φ(i) ∈ {1, . . . , p} such that gP (v) = φ(i) for every vertex v ∈ fP−1 (i).
Proposition 3.5. For each i ∈ {1, . . . , k}, we have |fP−1 (i)| = 2n2 + k + 1 − 2i.
Proof. Recall that fP (Pi [0]) = fP (Pi [2n2 − 1]) = i, so by Proposition 3.4 we have gP (Pi [0]) =
gP (Pi [2n2 −1]) = φ(i). It then follows from the lists of permitted positions for Pi [0] and Pi [2n2 −1] that
their positions on the path differ by precisely 2n2 +k−2i. Thus we must have |fP−1 (i)| ≥ 2n2 +k−2i+1.
It then follows that
k
X
i=1
|fP−1 (i)|
≥
k
X
i=1
1
2n2 + k − 2i + 1 = 2kn2 + k 2 + k − 2 · k(k + 1) = 2kn2 ,
2
with equality if and only if |fP−1 (i)| = 2n2 + k − 2i + 1 for each i. Since the sets {fP−1 (1), . . . , fP−1 (k)}
P
partition the p-vertices, of which there are in total 2kn2 , it therefore follows that ki=1 |fP−1 (i)| = 2kn2
and hence that |fP−1 (i)| = 2n2 + k − 2i + 1 for each i, as required.
Proposition 3.6. If P selects the edge e, we must have e = Vi [φ(i)]Vj [φ(j)] for some 1 ≤ i < j ≤ k.
13
Proof. Suppose that P selects the edge e, and that the vertices in the gadget Ge are Pi [`], . . . , Pi [` +
3], Pj [`], . . . , Pj [` + 3]. Since Pi [`] and Pj [` + 1] occur consecutively on P , it follows from Proposition
3.3 that fP (Pi [`]) = fP (Pj [` + 1]) and gP (Pi [`]) = gP (Pi [` + 1]). The list for Pi [`] is precisely
{(i − 1)n3 + 3αn2 + t + β : 1 ≤ α ≤ p, −(k − 1) ≤ β ≤ (k − 1)}, so we can see immediately that
fP (Pi [`]) = i and hence (by Proposition 3.4) gP (Pi [`]) = φ(i).
The fact that fP (Pj [` + 1]) = i means that the only possible positions in the list for Pj [` + 1] are
{(i − 1)n3 + 3rn2 + (` + 1) + β : −(k − 1) ≤ β ≤ (k − 1)}, where e = Vi [r]Vj [s] for some s ∈ {1, . . . , p}.
We know, however, that gP (Pj [` + 1]) = φ(i), so it follows that Vi [φ(i)] is one endpoint of e.
Applying symmetric reasoning to the Pj [`] and Pi [`+1] (which also occur consecutively on P ) tells
us that Vj [φ(j)] is the other endpoint of e. We therefore have e = Vi [φ(i)]Vj [φ(j)], as required.
Proposition 3.7. The set of edges selected by P is {Vi [φ(i)Vj [φ(j)] : 1 ≤ i < j ≤ k}.
Proof. First observe that the only vertices which can possibly belong to fP−1 (i) are vertices which
either belong to Pi , or else to an edge gadget Ge such that P selects e and one endpoint of e is in
Vi . Suppose that P selects the gadget Ge where e = Vi [r]Vj [s]. If j > i, fP−1 (i) will contain precisely
two vertices of Pj in Ge , but one vertex of Pi in Ge will not belong to fP−1 (i). If j < i, on the other
hand, then fP−1 (i) will contain exactly one vertex of Pj in Ge , but two vertices of Pi in Ge will not
belong to fP−1 (i).
We now prove, by induction on i, that the set of edges P selects with one endpoint in V1 ∪ · · · ∪
Vi is precisely {V` [φ(`)]Vj [φ(j)] : 1 ≤ ` ≤ i}. We know from Proposition 3.5 that |fP−1 (1)| =
2n2 + k − 1 so, as |Pi | = 2n2 , we must have additional vertices belonging to fP−1 (1) from selected edges with an endpoint in V1 . For each such edge that P selects, the number of vertices in fP−1 (i) will be increased by exactly one, so we must select exactly k − 1 such edges. By
Proposition 3.6, we know that the only edges we can select which have one endpoint in V1 are
V1 [φ(1)]V2 [φ(2)], V1 [φ(1)]V3 [φ(3)], . . . , V1 [φ(1)]Vk [φ(k)], so in fact all of these edges must be selected.
This completes the proof of the base case.
For the inductive step, we will assume that all edges Vi [φ(i)]Vj [φ(j)] with 1 ≤ i < k 0 and i < j ≤ k
are selected. We need to show that all edges Vk0 [φ(k 0 )]Vj [φ(j)] with k 0 < j ≤ k are also selected.
We know from Proposition 3.5 that |fP−1 (k 0 )| = 2n2 + k + 1 − 2k 0 . Since |Pk0 | = 2n2 , and we know
from the inductive hypothesis that exactly k 0 −1 edges Vi [φ(i)]Vk0 [φ(k 0 )] with i < k 0 are selected (each
of which reduces the number of vertices in fP−1 (k 0 ) by one), it follows that there must be exactly
2n2 + k + 1 − 2k 0 − (2n2 − (k 0 − 1)) = k − k 0
edges selected with one endpoint in Vk0 and one endpoint in Vj for some j > k 0 . By Proposition 3.6,
the only such edges which could be selected are Vk0 [φ(k 0 )]Vk0 +1 [φ(k 0 + 1)], . . . , Vk0 [φ(k 0 )]Vk [φ(k)]; it
follows that all of these edges must be selected by P .
This completes the proof by induction, and the result follows immediately.
Proposition 3.8. The set of vertices {Vi [φ(i)] : 1 ≤ i ≤ k} induces a multicolour clique in G.
Proof. It is immediate that this set of vertices is multicoloured, since it contains exactly one vertex from each vertex class V1 , . . . , Vk . Moreover, we know from Proposition 3.7 that P selects
Vi [φ(i)]Vj [φ(j)] for each 1 ≤ i < j ≤ k, so Vi [φ(i)]Vj [φ(j)] must be an edge of G, for each
1 ≤ i < j ≤ k. Hence {Vi [φ(i)] : 1 ≤ i ≤ k} induces a clique.
The fact that the existence of a valid Hamilton path in H implies the existence of a multicolour
clique in G now follows immediately.
Lemma 3.9. If (H, L) is a yes-instance for List Hamilton Path, then G is a yes-instance for
Multicolour Clique.
14
It now remains only to bound the pathwidth of the graph H.
Lemma 3.10. H has pathwidth at most 5k.
Proof. We construct a path decomposition of H, indexed by a path T with |T | = |Q| + 2n2 − 5.
Every bag of the decomposition contains the vertices Vend = {Pj [0], Pj [2n2 − 1] : 1 ≤ j ≤ k}. In
addition, for 1 ≤ i ≤ |Q| − 1, the bag indexed by the ith node of T contains the ith and (i + 1)th
vertices of Q, while, for |Q| ≤ i ≤ |Q| + 2n2 − 5, the bag indexed by the ith node of T contains all
vertices {Pj [i − |Q| + 1], Pj [i − |Q| + 2], Pj [i − |Q| + 3] : 1 ≤ j ≤ k}. Note that every bag contains at
most 5k vertices.
It is immediate from this construction that, for any vertex v ∈ V (H), the nodes indexing bags
that contain v induce a subpath of T . So it remains to show that, for every edge uv ∈ E(H), there
exists some bag of the decomposition that contains both u and v. For all edges within Q this is
clearly true. Note that, for 1 ≤ l ≤ 2n2 − 2, any vertex Pi [l] is only adjacent to vertices Pj [l0 ]
where |l − l0 | ≤ 2, and so any edge between internal vertices of the paths P1 , . . . , Pk must have both
its endpoints in some bag of the decomposition. All remaining edges are then incident with some
v ∈ Vend , but Vend is contained in every bag, and so the condition is also satisfied for these edges.
Hence we have a path decomposition of H of width at most 5k.
We are now ready to prove the main theorem of this section.
Proof of Theorem 3.1. It follows immediately from Lemmas 3.2 and 3.9 that (H, L) is a yes-instance
for List Hamilton Path if and only if G is a yes-instance for Multicolour Clique. H has
order polynomial in |G|, and can clearly be computed from G in polynomial time. Moreover, by
Lemma 3.10 we know that the pathwidth of H depends only on the parameter k, the number of
colours used in G. This completes the reduction to show that List Hamilton Path, parameterised
by treewidth, is W[1]-hard. Moreover, the construction given above demonstrates that if there is
an algorithm to solve List Hamilton Path in time f (k)no(k) on graphs of pathwidth k (for some
function f ) then Multicolour Clique and hence Clique can be solved in time g(k)no(k) (for
some function g); thus, by Theorem 1.3, there can be no such algorithm for List Hamilton Path
unless the Exponential Time Hypothesis fails.
4
Conclusions and Open Problems
We have proved that List Edge Chromatic Number and List Total Chromatic Number
are fixed parameter tractable, parameterised by treewidth, although the List Edge Colouring
and List Total Colouring problems are NP-hard on graphs of treewidth at most two. Thus, as
for vertex colouring, it is computationally easier to calculate list edge or total chromatic number of
a graph than to determine whether a given set of lists admits a proper colouring of the graph.
We also demonstrated that List Hamilton Path is W[1]-hard, even when parameterised by
pathwidth, giving another example of a problem that solvable in linear time on graphs of bounded
treewidth but has a W[1]-hard list version. Our reduction also implies f (k)no(k) lower bound on the
running time for any algorithm to solve this problem on graphs of treewidth k, under the exponential
time hypothesis. A natural open question is whether List Hamilton Path may belong to FPT
with respect to other parameters, such as the feedback vertex set size or the vertex cover number of
the graph.
More generally, it would be interesting to investigate whether there are further problems that are
fixed parameter tractable parameterised by treewidth but have a list version which is W[1]-hard in
this setting.
15
References
[1] Noga Alon, Restricted colorings of graphs, Surveys in Combinatorics, 1993, Proceedings, 14th
British Combinatorial Conference (K. Walker, ed.), London Mathematical Society Lecture Note
Series, vol. 187, Cambridge University Press, 1993, pp. 1–33.
[2] Stefan Arnborg and Andrzej Proskurowski, Linear time algorithms for NP-hard problems restricted to partial k-trees, Discrete Applied Mathematics 23 (1989), 11–24.
[3] M. Behzad, Graphs and their chromatic numbers, Ph.D. thesis, Michigan State University, 1965.
[4] Hans L. Bodlaender, A linear-time algorithm for finding tree-decompositions of small treewidth,
SIAM Journal on Computing 25 (1996), no. 6, 1305–1317.
[5] B. Bollobás and A. Harris, List-colourings of graphs, Graphs and Combinatorics 1 (1985), 115–
127.
[6] Richard B. Borie, R. Gary Parker, and Craig A. Tovey, Automatic generation of linear-time
algorithms from predicate calculus descriptions of problems on recursively constructed graph
families, Algorithmica 7 (1992), 555–581.
[7] O. V. Borodin, A. V. Kostochka, and D. R. Woodall, List edge and list total colourings of
multigraphs, Journal of Combinatorial Theory, Series B 71 (1997), no. 2, 184–204.
[8] Henning Bruhn, Richard Lang, and Maya Stein, List edge-coloring and total coloring in graphs
of low treewidth, Journal of Graph Theory 81 (2016), no. 3, 272–282.
[9] Jianer Chen, Benny Chor, Mike Fellows, Xiuzhen Huang, David Juedes, Iyad A. Kanj, and
Ge Xia, Tight lower bounds for certain parameterized NP-hard problems, Information and Computation 201 (2005), no. 2, 216 – 231.
[10] B. Courcelle, The monadic second-order logic of graphs I: recognizable sets of finite graphs,
Information and Computation 85 (1990), 12–75.
[11] Gruia Cǎlinescu, Cristina G. Fernandes, and Bruce Reed, Multicuts in unweighted graphs and
digraphs with bounded degree and bounded tree-width, Journal of Algorithms 48 (2003), 333–359.
[12] Rodney G. Downey and Michael R. Fellows, Fundamentals of parameterized complexity, Texts
in Computer Science, Springer, 2013.
[13] M. Fellows, D. Hermelin, F. Rosamond, and S. Vialette, On the fixed-parameter intractability
and tractability of multiple-interval graph properties, Theoretical Computer Science 410 (2009),
53–61.
[14] Michael R. Fellows, Fedor V. Fomin, Daniel Lokshtanov, Frances Rosamond, Saket Saurabh,
Stefan Szeider, and Carsten Thomassen, On the complexity of some colorful problems parameterized by treewidth, Information and Computation 209 (2011), 143–153.
[15] J. Flum and M. Grohe, Parameterized complexity theory, Springer, 2006.
[16] F. Galvin, The list chromatic index of a bipartite multigraph, Journal of Combinatorial Theory,
Series B 63 (1995), no. 1, 153–158.
[17] M. R. Garey, D. S. Johnson, and R. Endre Tarjan, The planar Hamilton circuit problem is
NP-complete, SIAM Journal on Computing 5 (1976), 704–714.
16
[18] Roland Häggkvist and Amanda Chetwynd, Some upper bounds on the total and list chromatic
numbers of multigraphs, Journal of Graph Theory 16 (1992), no. 5, 503–516.
[19] Ian Holyer, The NP-completeness of edge-coloring, SIAM J. on Computing 10 (1981), no. 4,
718–720.
[20] Shuji Isobe, Xiao Zhou, and Takao Nishizeki, Total colourings of degenerate graphs, Combinatorica 27 (2007), no. 2, 167–182.
[21] Klaus Jansen and Petra Scheffler, Generalized coloring for tree-like graphs, Discrete Applied
Mathematics 75 (1997), 135–155.
[22] Jeff Kahn, Asymptotics of the chromatic index for multigraphs, Journal of Combinatorial Theory,
Series B 68 (1996), no. 2, 233–254.
[23] M. S. Krishnamoorthy, An NP-hard problem in bipartite graphs, SIGACT News 7 (1975), 26–26.
[24] Daniel Leven and Zvi Galil, NP completeness of finding the chromatic index of regular graphs,
Journal of Algorithms 4 (1983), no. 1, 35–44.
[25] Daniel Lokshtanov, Dániel Marx, Saket Saurabh, et al., Lower bounds based on the exponential
time hypothesis, Bulletin of EATCS 3 (2013), no. 105.
[26] Dániel Marx, NP-completeness of list coloring and precoloring extension on the edges of planar
graphs, Journal of Graph Theory 49 (2005), 313–324.
[27] Colin J.H. McDiarmid and Abdón Sánchez-Arroyo, Total colouring regular bipartite graphs is
NP-hard, Discrete Mathematics 124 (1994), no. 1-3, 155–162.
[28] D. Seese, Tree-partite graphs and the complexity of algorithms, Fundamentals of Computation
Theory (Lothar Budach, ed.), Lecture Notes in Computer Science, vol. 199, Springer Berlin /
Heidelberg, 1985, pp. 412–421.
[29] V. G. Vizing, On an estimate of the chromatic class of a p-graph. (Russian), Diskret. Analiz 3
(1964), 25–30.
[30] V. G. Vizing, Some unsovled problems in graph theory (Russian), Uspekhi Math. Nauk. 23
(1968), 117–134.
[31] Xiao Zhou, Yuki Matsuo, and Takao Nishizeki, List total colorings of series-parallel graphs,
Journal of Discrete Algorithms 3 (2005), 47–60.
[32] Xiao Zhou, Shin-ichi Nakano, and Takao Nishizeki, Edge-colouring partial k-trees, Journal of
Algorithms 21 (1996), 598–617.
17
| 8cs.DS
|
1
Nonlinear Bayesian Estimation: From Kalman
Filtering to a Broader Horizon
arXiv:1712.01406v2 [cs.SY] 14 Dec 2017
Huazhen Fang, Ning Tian, Yebin Wang, MengChu Zhou, and Mulugeta A. Haile
Abstract—This article presents an up-to-date tutorial review
of nonlinear Bayesian estimation. State estimation for nonlinear
systems has been a challenge encountered in a wide range of
engineering fields, attracting decades of research effort. To date,
one of the most promising and popular approaches is to view and
address the problem from a Bayesian probabilistic perspective,
which enables estimation of the unknown state variables by
tracking their probabilistic distribution or statistics (e.g., mean
and covariance) conditioned on the system’s measurement data.
This article offers a systematic introduction of the Bayesian
state estimation framework and reviews various Kalman filtering
(KF) techniques, progressively from the standard KF for linear
systems to extended KF, unscented KF and ensemble KF for
nonlinear systems. It also overviews other prominent or emerging
Bayesian estimation methods including the Gaussian filtering,
Gaussian-sum filtering, particle filtering and moving horizon
estimation and extends the discussion of state estimation forward
to more complicated problems such as simultaneous state and
parameter/input estimation.
Index Terms—State estimation, nonlinear Bayesian estimation,
Kalman filtering, stochastic estimation.
I. I NTRODUCTION
S a core subject of control systems theory, state estimation for nonlinear dynamic systems has been undergoing active research and development for a few decades.
Considerable attention is gained from a wide community of
researchers, thanks to its significant applications in signal
processing, navigation and guidance, and econometrics, just to
name a few. When stochastic systems, i.e., systems subjected
to the effects of noise, are considered, the Bayesian estimation
approaches have evolved as a leading estimation tool enjoying
wide popularity. Bayesian analysis traces back to the 1763
essay [1], published two years after the death of its author,
Rev. Thomas Bayes. This seminal work was meant to tackle
the following question: “Given the number of times in which
an unknown event has happened and failed: Required the
chance that the probability of its happening in a single trial
lies somewhere between any two degrees of probability that
can be named”. Rev. Bayes developed a solution to examine
the case of only continuous probability, single parameter and
A
H. Fang and N. Tian are with the Department of Mechanical Engineering, University of Kansas, Lawrence, KS 66045, USA (e-mail: {fang,
ning.tian}@ku.edu).
Y. Wang is with the Mitsubishi Electric Research Laboratories, Cambridge,
MA 02139, USA (e-mail: [email protected]).
M. Zhou is with the Department of Electrical and Computer Engineering,
New Jersey Institute of Technology, Newark, NJ 07102, USA (e-mail:
[email protected]).
M.A. Haile is with the Vehicle Technology Directorate, US
Army Research Laboratory, Aberdeen, MD 21005, USA (e-mail:
[email protected]).
a uniform prior, which is an early form of the Bayes’ rule
known to us nowadays. Despite its preciousness, this work
remained obscure for many scientists and even mathematicians
of that era. The change came when the French mathematician
Pierre-Simon de Laplace rediscovered the result and presented
the theorem in the complete and modern form. A historical
account and comparison of Bayes’ and Laplace’s work can be
found in [2]. From today’s perspective, the Bayes’ theorem
is a probability-based answer to a philosophical question:
How should one update an existing belief when given new
evidence [3]? Quantifying the degree of belief by probability,
the theorem modifies the original belief by producing the
probability conditioned on new evidence from the initial
probability. This idea was applied in the past century from
one field to another whenever the belief update question arose,
driving numerous intriguing explorations. Among them, a
topic of relentless interest is Bayesian state estimation, which
is concerned with determining the unknown state variables of
a dynamic system using the Bayesian theory.
The capacity of the Bayesian analysis to provide a powerful
framework for state estimation has been well recognized now.
A representative method within the framework is the wellknown Kalman filter (KF), which “revolutionized the field
of estimation ... (and) opened up many new theoretical and
practical possibilities” [4]. KF was initially developed by using
the least squares in the 1960 paper [5] but reinterpreted from a
Bayesian perspective in [6], only four years after its invention.
Further envisioned in [6] was that “the Bayesian approach
offers a unified and intuitive viewpoint particularly adaptable
to handling modern-day control problems”. This investigation
and vision ushered a new statistical treatment of nonlinear
estimation problems, laying a foundation for prosperity of
research on this subject.
In this article, we offer a systematic and bottom-to-up
introduction to major Bayesian state estimators, with a particular emphasis on the KF family. We begin with outlining
the essence of Bayesian thinking for state estimation problems, showing that its core is the model-based prediction
and measurement-based update of the probabilistic belief of
unknown state variables. A conceptual KF formulation can
be made readily in the Bayesian setting, which tracks the
mean and covariance of the states modeled as random vectors
throughout the evolution of the system. Turning a conceptual
KF into executable algorithms requires certain approximations
to nonlinear systems; and depending on the approximation
adopted, different KF methods are derived. We demonstrate
three primary members of the KF family in this context:
extended KF (EKF), unscented KF (UKF), and ensemble
2
KF (EnKF), all of which have achieved proven success both
theoretically and practically. A review of other important
Bayesian estimators and estimation problems is also presented
briefly in order to introduce the reader to the state of the art
of this vibrant research area.
II. A BAYESIAN V IEW OF S TATE E STIMATION
We consider the following nonlinear discrete-time system:
(
xk+1 = f (xk ) + wk ,
(1)
yk = h(xk ) + vk ,
where xk ∈ Rnx is the unknown system state, and yk ∈ Rny
the output, with both nx and ny being positive integers.
The process noise wk and the measurement noise vk are
mutually independent, zero-mean white Gaussian sequences
with covariances Qk and Rk , respectively. The nonlinear
mappings f : Rnx → Rnx and h : Rnx → Rny represent the
process dynamics and the measurement model, respectively.
The system in (1) is assumed input-free for simplicity of
presentation, but the following results can be easily extended
to an input-driven system.
The state vector xk comprises a set of variables that fully
describe the status or behavior of the system. It evolves
through time as a result of the system dynamics. The process
of states over time hence represents the system’s behavior.
Because it is unrealistic to measure the complete state in
most practical applications, state estimation is needed to infer
xk from the output yk . More specifically, the significance of
estimation comes from the crucial role it plays in the study of
dynamic systems. First, one can monitor how a system behaves
with state information and take corresponding actions when
any adjustment is necessary. This is particularly important
to ensure the detection and handling of internal faults and
anomalies at the earliest phase. Second, high-performing state
estimation is the basis for the design and implementation of
many control strategies. The past decades have witnessed a
rapid growth of control theories, and most of them, including
optimal control, model predictive control, sliding mode control
and adaptive control, premise the design on the availability of
state information.
While state estimation can be tackled in a variety of
ways, the stochastic estimation has drawn remarkable attention
and been profoundly developed in terms of both theory and
applications. Today, it is still receiving continued interest
and intense research effort. From a stochastic perspective,
the system in (1) can be viewed as a generator of random
vectors xk and yk . The reasoning is as follows. Owing to
the initial uncertainty or lack of knowledge of the initial
condition, x0 can be considered as a random vector subject
to variation due to chance. Then, f (x0 ) represents a nonlinear
transformation of x0 , and its combination with w0 modeled
as another random vector generates a new random vector x1 .
Following this line, xk for any k is a random vector, and
the same idea applies to yk . In practice, one can obtain the
sensor measurement of the output at each time k, which can
be considered as a sample drawn from the distribution of the
random vector yk . For simplicity of notation, we also denote
the output measurement as yk and the measurement set at
time k as k := {y1 , y2 , · · · , yk }. The state estimation then
is to build an estimate of xk using k at each time k. To
this end, one’s interest then lies in how to capture p(xk | k ),
i.e., the probability density function (pdf) of xk conditioned
on k . This is because p(xk | k ) captures the information of
xk conveyed in k and can be leveraged to estimate xk .
A “prediction-update” procedure1 can be recursively executed to obtain p(xk | k ). Standing at time k − 1, we can
predict what p(xk | k−1 ) is like using p(xk−1 | k−1 ). When
the new measurement yk conveying information about xk arrives, we can update p(xk | k−1 ) to p(xk | k ). Characterizing
a probabilistic belief about xk before and after the arrival of
yk , p(xk | k−1 ) and p(xk | k ) are referred to as the a priori
and a posteriori pdf’s, respectively. Specifically, the prediction
at time k − 1, demonstrating the pass from p(xk−1 | k−1 ) to
p(xk | k−1 ), is given by
Z
p(xk | k−1 ) = p(xk |xk−1 )p(xk−1 | k−1 )dxk−1 . (2)
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Y
Let us explain how to achieve (2). By the ChapmanKolmogorov equation, it can be seen that
Z
p(xk | k−1 ) = p (xk , xk−1 | k−1 ) dxk−1 ,
Y
Y
which, according to the Bayes’ rule, can be written as
Z
p(xk | k−1 ) = p(xk |xk−1 , k−1 )p(xk−1 | k−1 )dxk−1 .
Y
Y
Y
Y
It reduces to (2), because p(xk |xk−1 , k−1 ) = p(xk |xk−1 ) as
a result of the Markovian propagation of the state. Then on the
arrival of yk , p(xk | k−1 ) can be updated to yield p(xk | k ),
which is governed by
Y
|xk )p(xk |Yk−1 )
.
Yk ) = p(ykp(y
k |Yk−1 )
p(xk |
Y
(3)
The above equation is also owing to the use of the Bayes’
rule:
Yk ) = p(xp(kY, Yk )k ) = p(xp(yk ,ky,kY, Yk−1k−1) )
p(yk |xk , Yk−1 )p(xk , Yk−1 )
=
p(yk , Yk−1 )
p(yk |xk , Yk−1 )p(xk |Yk−1 )
=
.
p(yk |Yk−1 )
Note that we have p(yk |xk , Yk−1 ) = p(yk |xk ) from the fact
p(xk |
that yk only depends on xk . Then, (3) is obtained. Together,
(2)-(3) represent the fundamental principle of Bayesian state
estimation for the system in (1), describing the sequential
propagation of the a priori and a posteriori pdf’s. The former captures our belief over the unknown quantities in the
presence of only the prior evidence, and the latter updates this
belief using the Bayesian theory when new evidence becomes
available. The two steps, prediction and update, are executed
alternately through time, as illustrated in Fig. 1.
1 The two steps are equivalently referred to as ‘time-update’ and
‘measurement-update’, or ‘forecast’ and ‘analysis’, in different literature.
3
the problem further, certain Gaussianity approximations can
be made because of the mathematical tractability and statistical soundness of Gaussian distributions (for the reader’s
convenience, several properties of the Gaussian distribution to
be used next are summarized in the Appendix.). Proceeding
in this direction, we can reach a formulation of the Kalman
filtering (KF) methodology, as shown below.
In order to predict xk at time k − 1, we consider the
minimum-variance unbiased estimation, which gives that the
best estimate of xk given
k−1 , denoted as x̂k|k−1 , is
E(xk | k−1 ) [7, Theorem 3.1]. That is,
Z
x̂k|k−1 = E(xk | k−1 ) = xk p(xk | k−1 )dxk .
(4)
Y
Y
Fig. 1. The Bayesian filtering principle. The running of a dynamic system
propagates the state xk through time and produces output measurement yk
at each time k. For the purpose of estimation, the Bayesian filtering principle
tracks the pdf of xk given the measurement set k = {y1 , y2 , · · · , yk }.
It consists of two steps sequentially implemented: the prediction step from
p(xk−1 | k−1 ) to p(xk | k−1 ), and the update step from p(xk | k−1 ) to
p(xk | k ) upon the arrival of yk .
Y
Y
Y
Y
Y
Y
Y
Inserting (2) into the above equation, we have
Z Z
x̂k|k−1 =
xk p(xk |xk−1 )dxk
Yk−1)dxk−1.
· p(xk−1 |
Looking at the above Bayesian filtering principle, we can
summarize three elements that constitute the thinking of
Bayesian estimation. First, all the unknown quantities or
uncertainties in a system, e.g., state, are viewed from the
probabilistic perspective. In other words, any unknown variable is regarded as a random variable. Second, the output
measurements of the system are samples drawn from a certain
probability distribution dependent on the concerned variables.
They provide data evidence for state estimation. Finally, the
system model represents transformations that the unknown and
random state variables undergo over time. Originating from
the philosophical abstraction that anything unknown, in one’s
mind, is subject to variations due to chance, the randomnessbased representation enjoys universal applicability even when
the unknown or uncertain quantities are not necessarily random
in physical sense. In addition, it can easily translate into a
convenient ‘engineering’ way for estimation of the unknown
variables, as will be shown in the following discussions.
By assuming that wk is a white Gaussian noise independent
R of xk , we have xk |xk−1 ∼ N (f (xk−1 ), Q) and
then xk p(xk |xk−1 )dxk = f (xk−1 ) according to (A.1).
Hence, (5) becomes
Z
x̂k|k−1 = f (xk−1 )p(xk−1 | k−1 )dxk−1
Y
Yk−1)] .
= E [f (xk−1 |
(6)
For x̂k|k−1 in (6), the associated prediction error covariance
is
x
Pk|k−1
= E (xk − x̂k|k−1 )(xk − x̂k|k−1 )>
Z
= (xk − x̂k|k−1 )(xk − x̂k|k−1 )> p(xk | k−1 )dxk . (7)
Y
With the use of (2) and (A.1), we can obtain
Z
x
>
Pk|k−1 = xk x>
k p(xk | k−1 )dxk − x̂k|k−1 x̂k|k−1
Z Z
=
xk x>
p(x
|x
)dx
p(xk−1 | k−1 )dxk−1
k
k−1
k
k
Y
Y
III. F ROM BAYESIAN F ILTERING TO K ALMAN F ILTERING
In the above discussion, we have shown the probabilistic
nature of state estimation and presented the Bayesian filtering
principle (2)-(3) as a solution framework. However, this does
not mean that one can simply use (2)-(3) to track the conditional pdf of a random vector passing through nonlinear transformations, because the nonlinearity often makes it difficult
or impossible to derive an exact or closed-form solution. This
challenge turns against the development of executable state
estimation algorithms, since a dynamic system’s state propagation and observation are based on the nonlinear functions
of the random state vector xk , i.e., f (xk ) and h(xk ). Yet for
the sake of estimation, one only needs the statistics (mean and
covariance) of xk conditioned on the measurements in most
circumstances, rather than a full grasp of its conditional pdf.
A straightforward and justifiable way is to use the mean as
the estimate of xk and the covariance as the confidence (or
equivalently, uncertainty) measure. Reducing the pdf tracking
to the mean and covariance tracking can significantly mitigate
the difficulty in the design of state estimators. To simplify
(5)
Z
− x̂k|k−1 x̂>
k|k−1
f (xk−1 )f > (xk−1 ) + Q p(xk−1 |
Z
− x̂k|k−1 x̂>
k|k−1
>
f (xk−1 ) − x̂k|k−1 f (xk−1 ) − x̂k|k−1
=
=
Yk−1)dxk−1
Yk−1)dxk−1 + Q
= Cov [f (xk−1 )|Yk−1 ] + Q.
· p(xk−1 |
(8)
When yk becomes available, we assume that
p(xk , yk | k−1 ) can be approximated by a Gaussian
distribution
xy
x
Pk|k−1
Pk|k−1
x̂
,
(9)
N k|k−1 , xy >
ŷk|k−1
P
Py
Y
k|k−1
k|k−1
where ŷk|k−1 is the prediction of yk given
as
Z
ŷk|k−1 = E(yk | k−1 ) = yk p(yk |
Y
Yk−1 and expressed
Yk−1)dyk .
(10)
4
Y
a Gaussian approximation can be constructed for p (xk | k ).
Its mean and covariance can be expressed as
−1
y
xy
(yk − ŷk|k−1 ), (16)
Pk|k−1
x̂k|k = x̂k|k−1 + Pk|k−1
|
{z
}
Kalman gain
>
−1
xy
y
xy
x
x
.
Pk|k−1
Pk|k−1
− Pk|k−1
= Pk|k−1
Pk|k
Fig. 2. A schematic sketch of the KF technique. KF performs the predictionupdate procedure recursively to track the mean and covariance of xk for
estimation. The equations show that its implementation depends on determining the mean and covariance of the random state vector through nonlinear
functions f (·) and h(·).
The associated covariance is
Z
>
y
Pk|k−1 =
yk − ŷk|k−1 yk − ŷk|k−1
Yk−1) dyk .
· p (yk |
(11)
It is noted that
Yk−1) =
Yk−1)dxk
Z
p(yk |
p(yk |xk )p(xk |
=
Yk−1)dxk .
IV. S TANDARD L INEAR K ALMAN F ILTER
(12)
Combining (10)-(11) with (12) yields
Z Z
ŷk|k−1 =
yk p(yk |xk )dyk p(xk | k−1 )dxk
Z
= h(xk )p(xk | k−1 )dxk = E [h(xk )| k−1 ] ,
Y
Y
Y
(13)
Z
h(xk ) − ŷk|k−1
h(xk ) − ŷk|k−1
>
Yk−1)dxk + R
= Cov [h(xk )|Yk−1 ] + R.
· p(xk |
(14)
The cross-covariance between xk and yk is
ZZ
>
xy
Pk|k−1
=
xk − x̂k|k−1 yk − ŷk|k−1
Z
· p (xk , yk |
Z
xk − x̂k|k−1
yk − ŷk|k−1
>
Yk−1) dxk
p (yk |xk ) dyk
· p (xk |
Z
=
xk − x̂k|k−1
= Cov [xk , h(xk )|
h(xk ) − ŷk|k−1
Yk−1] .
>
In this section, KF for linear systems is reviewed briefly
to pave the way for discussion of nonlinear KFs. Consider a
linear time-invariant discrete-time system of the form
(
xk+1 = F xk + wk ,
(18)
yk = Hxk + vk ,
where: 1) {wk } and {vk } are zero-mean white Gaussian noise
sequences with wk ∼ N (0, Q) and vk ∼ N (0, R), 2) x0 is
Gaussian with x0 ∼ N (x̄0 , P0 ), and 3) x0 , {wk } and {vk } are
independent of each other. Note that, under these conditions,
the Gaussian assumptions in Section III will exactly hold for
the linear system in (18).
The standard KF for the linear dynamic system in (18)
can be readily derived from the conceptual KF summarized
in Fig. 2. Since the system is linear and under a Gaussian
setting, p (xk−1 | k−1 ) and p (xk | k−1 ) are strictly Gaussian
according to the properties
of Gaussian
vectors. Specifically,
x
xk−1 | k−1 ∼ N x̂k−1|k−1 , Pk−1|k−1
and xk | k−1 ∼
x
N x̂k|k−1 , Pk|k−1
. According to (6) and (8), the prediction
is
Y
Yk−1) dxk dyk
=
Putting together (6)-(8), (13)-(14) and (15)-(17), we can
establish a conceptual description of the KF technique, which
is outlined in Fig. 2. Built in the Bayesian-Gaussian setting,
it conducts state estimation through tracking the mean and
covariance of a random state vector. It is noteworthy that
one needs to develop explicit expressions to enable the use
of the above KF equations. The key that bridges the gap is
to find the mean and covariance of a random vector passing
through nonlinear functions. For linear dynamic systems, the
development is straightforward, because, in the considered
context the involved pdf’s are strictly Gaussian and the linear
transformation of the Gaussian state variables can be readily
handled. The result is the standard KF to be shown in the
next section. However, complications arise in the case of
nonlinear systems. This issue has drawn significant interest
from researchers. A wide range of ideas and methodologies
have been developed, leading to a family of nonlinear KFs.
The three most representative among them are EKF, UKF,
and EnKF to be introduced following the review of the linear
KF.
p(xk , yk |
Z
y
Pk|k−1
=
(17)
Yk−1) dxk
p (xk |
(15)
For two jointly Gaussian random vectors, the conditional distribution of one given another is also Gaussian, which is summarized in (A.4) in Section IV. It then follows from (9) that
Y
Y
Y
Yk−1) = F x̂k−1|k−1,
Yk−1) + Q
x̂k|k−1 = E (F xk−1 |
(19)
x
Pk|k−1
= Cov (F xk−1 |
x
= F Pk−1|k−1
F > + Q,
(20)
The update can be accomplished along the similar lines.
y
Based on (13)-(15), we have ŷk|k−1 = H x̂k|k−1 , Pk|k−1
=
5
EKF and UKF in the field of control systems and the EnKF
popular in the data assimilation community.
A. Extended Kalman Filter
Fig. 3. A schematic of the KF/EKF structure, modified from [8]. KF/EKF
comprises two steps sequentially executed through time, prediction and
update. For prediction, xk is predicted by using the data up to time k − 1.
The forecast is denoted as x̂k|k−1 and subject to uncertainty quantified by the
prediction error covariance Pk|k−1 . The update step occurs upon the arrival
of the new measurement yk . In this step, yk is leveraged to correct x̂k|k−1
and produce the updated estimate x̂k|k . Meanwhile, Pk|k−1 is updated to
generate Pk|k to quantify the uncertainty imposed on x̂k|k .
xy
HPk|k−1 H > + R, and Pk|k−1
= Pk|k−1 H > . Then, as indicated by (16)-(17), the updated state estimate is
−1
x
x
x̂k|k = x̂k|k−1 + Pk|k−1
H > HPk|k−1
H> + R
{z
}
|
Kalman gain
x
Pk|k
· yk − H x̂k|k−1 ,
(21)
−1
x
x
x
= Pk|k−1
− Pk|k−1
H > HPk|k−1
H> + R
x
· HPk|k−1
.
(22)
Together, (19)-(22) form the linear KF. Through the predictionupdate procedure, it generates the state estimates and associated estimation error covariances recursively over time
when the output measurement arrives. From the probabilistic
perspective, x̂k|k and Pk|k together determine the Gaussian
distribution of xk conditioned on k . The covariance, quantifying how widely the random vector xk can potentially spread
out, can be interpreted as a measure of the confidence on or
uncertainty of the estimate. A schematic diagram of the KF is
shown in Fig. 3 (it can also be used to demonstrate EKF to
be shown next).
1
Given that (F, H) is detectable and (F, Q 2 ) stabilizable,
x
Pk|k−1
converges to a fixed point, which is the solution to a
discrete-time algebraic Riccati equation (DARE)
Y
X = F XF > − F XH > (HXH > + R)−1 HXF > + Q.
This implies that KF can approach a steady state after a few
time instants. With this idea, one can design a steady-state
KF by solving DARE offline to obtain the Kalman gain and
then apply it to run KF online, as detailed in [7]. Obviating
the need for computing the gain and covariance at every
time instant, the steady-state KF, though suboptimal, presents
higher computational efficiency than the standard KF.
V. R EVIEW OF N ONLINEAR K ALMAN F ILTERS
In this section, an introductory overview of the major
nonlinear KF techniques is offered, including the celebrated
EKF is arguably the most celebrated nonlinear state estimation technique, with numerous applications across a variety
of engineering areas and beyond [9]. It is based on the
linearization of nonlinear functions around the most recent
state estimate. When the state estimate x̂k−1|k−1 is generated,
consider the first-order Taylor expansion of f (xk−1 ) at this
point:
f (xk−1 ) ≈ f (x̂k−1|k−1 ) + Fk−1 xk−1 − x̂k−1|k−1 , (23)
∂f
.
(24)
Fk−1 =
∂x x̂k−1|k−1
Y
For simplicity, let p(xk−1 | k−1 ) be approximated by a disx
tribution with mean x̂k−1|k−1 and covariance Pk−1|k−1
. Then
inserting (23) to (6)-(8), we can readily obtain the one-stepforward prediction
(25)
x̂k|k−1 = E [f (xk−1 | k−1 )] ≈ f x̂k−1|k−1 ,
Y
x
Pk|k−1
Yk−1] + Q
= Cov [xk |
=
x
>
Fk−1 Pk−1|k−1
Fk−1
+ Q.
(26)
Looking into (23), we find that the Taylor expansion approximates the nonlinear transformation of the random vector
xk by an affine one. Proceeding on this approximation, we
can easily estimate the mean and covariance of f (xk−1 )
once provided the mean and covariance information of xk−1
conditioned on k−1 . This, after being integrated with the
effect of the noise wk on the prediction error covariance,
establishes a prediction of xk , as specified in (25)-(26). After
x̂k|k−1 is produced, we can investigate the linearization of
h(xk ) around this new operating point in order to update the
prediction. That is,
(27)
h(xk ) ≈ h x̂k|k−1 + Hk xk − x̂k|k−1 ,
∂h
Hk =
.
(28)
∂x x̂k|k−1
Y
Y
Similarly, we assume that p(xk | k−1 ) can be replaced by
x
a distribution with mean x̂k|k−1 and covariance Pk|k−1
.
Using (27),
the
evaluation
of
(13)-(15)
yields
ŷ
≈
k|k−1
y
xy
x
Hk> + R, and Pk|k−1
h x̂k|k−1 , Pk|k−1
≈ Hk Pk|k−1
≈
x
>
Pk|k−1 Hk .
Here, the approximate mean and covariance information of
h(xk ) and yk are obtained through the linearization of h(xk )
around x̂k|k−1 . With the aid of the Gaussianity assumption
in (9), an updated estimate of xk is produced as follows:
−1
x
x
Hk> + R
x̂k|k = x̂k|k−1 + Pk|k−1
Hk> Hk Pk|k−1
· yk − h x̂k|k−1 ,
(29)
−1
x
x
x
x
Pk|k
= Pk|k−1
− Pk|k−1
Hk> Hk Pk|k−1
Hk> + R
x
· Hk Pk|k−1
.
(30)
6
Then, EKF consists of (25)-(26) for prediction and (29)-(30)
for update. When comparing it with the standard KF in (19)(22), we can find that they share significant resemblance
in structure, except that EKF introduces the linearization
procedure to accommodate the system nonlinearities.
Since the 1960s, EKF has gained wide use in the areas of aerospace, robotics, biomedical, mechanical, chemical,
electrical and civil engineering, with great success in the
real world witnessed. This is often ascribed to its relative
easiness of design and execution. Another important reason
is its good convergence from a theoretical viewpoint. In
spite of the linearization-induced errors, EKF has provable
asymptotic convergence under some conditions that can be
satisfied by many practical systems [10–14]. However, it also
suffers from some shortcomings. The foremost one is the
inadequacy of its first-order accuracy for highly nonlinear
systems. In addition, the need for explicit derivative matrices
not only renders EKF futile for discontinuous or other nondifferentiable systems, but also pulls it away from convenient
use in view of programming and debugging, especially when
nonlinear functions of a complex structure are faced. This
factor, together with the computational complexity at O(n3x ),
limits the application of EKF to only low-dimensional systems.
Some modified EKFs have been introduced for improved
accuracy or efficiency. In this regard, a natural extension is
through the second-order Taylor expansion, which leads to
the second-order EKF with more accurate estimation [15–
17]. Another important variant, named iterated EKF (IEKF),
iteratively refines the state estimate around the current point
at each time instant [18–20]. Though requiring an increased
computational cost, it can achieve higher estimation accuracy
even in the presence of severe nonlinearities in systems.
B. Unscented Kalman Filter
As the performance of EKF degrades for systems with
strong nonlinearities, researchers have been seeking better
ways to conduct nonlinear state estimation. In the 1990s,
UKF was invented [21, 22]. Since then, it has been gaining
significant popularity among researchers and practitioners.
This technique is based on the so-called “unscented transform
(UT)”, which exploits the utility of deterministic sampling to
track the mean and covariance information of a random variable passing through a nonlinear transformation [23–25]. The
basic idea is to approximately represent a random variable by a
set of sample points (sigma points) chosen deterministically to
completely capture the mean and covariance. Then, projecting
the sigma points through the nonlinear function concerned,
one obtains a new set of sigma points and then use them to
form the mean and covariance of the transformed variable for
estimation.
To explain how UT tracks the statistics of a nonlinearly
transformed random variable, we consider a random variable
x ∈ n and a nonlinear function z = g(x). It is assumed that
the mean and covariance of x are x̄ and Px , respectively. The
UT proceeds as follows [23, 24]. First, a set of sigma points
{xi , i = 0, 1, · · · , 2n} are chosen deterministically:
R
x0 = x̄,
(31)
Fig. 4. A schematic sketch of the UT procedure, adapted from [23]. A set of
sigma points (blue dots) are generated first according to the initial mean (red
five-pointed star) and covariance (red ellipse) (left) and projected through
the nonlinear function to generate a set of new sigma points (right). The
new sigma points are then used to calculate the new mean (green star) and
covariance (green ellipse).
xi = x̄ +
xi+n = x̄ −
√
√
n+λ
hp
n+λ
hp
Px
Px
i
ii
, i = 1, 2, · · · , n,
(32)
, i = 1, 2, · · · , n,
(33)
i
where [·]i represents the i-th column of the matrix and the
√ √ >
matrix square root is defined by P P = P achievable
through the Cholesky decomposition. The sigma points spread
across the space around x̄. The width of spread is dependent
on the covariance P and the scaling parameter λ, where λ =
α2 (n + κ) − n. Typically, α is a small positive value (e.g.,
10−3 ), and κ is usually set to 0 or 3 − n [21]. Then the sigma
points are propagated through the nonlinear function g(·) to
generate the sigma points for the transformed variable z, i.e.,
z i = g xi , i = 0, 1, · · · , 2n.
The mean and covariance of z are estimated as
2n
X
(m)
z̄ = E [g(x)] ≈
Wi z i ,
(34)
i=0
h
i
>
Pz = E (g(x) − z̄) (g(x) − z̄)
≈
2n
X
(c)
Wi
z i − z̄
z i − z̄
>
,
(35)
i=0
where the weights are
(m)
W0
(c)
W0
(m)
Wi
λ
,
n+λ
λ
=
+ (1 − α2 + β),
n+λ
1
(c)
= Wi =
, i = 1, 2, · · · , 2n.
2(n + λ)
=
(36)
(37)
(38)
The parameter β in (37) can be used to include prior information on the distribution of x. When x is Gaussian, β = 2 is
optimal. The UT procedure is schematically shown in Fig. 4.
To develop UKF, it is necessary to apply UT at both
prediction and update steps, which involve nonlinear state
transformations based on f and h, respectively. For prediction,
suppose that the mean and covariance of xk−1 , x̂k−1|k−1 and
x
Pk−1|k−1
, are given. To begin with, the sigma points for xk−1
are generated:
(0)
x̂k−1|k−1 = x̂k−1|k−1 ,
(39)
7
(i)
x̂k−1|k−1 = x̂k−1|k−1 +
(i+n )
x
x̂k−1|k−1
= x̂k−1|k−1 −
p
p
nx + λ
hq
x
Pk−1|k−1
i
,
i
i = 1, 2, · · · , nx ,
(40)
i
hq
x
,
nx + λ
Pk−1|k−1
i
i = 1, 2, · · · , nx .
(41)
Then, they are propagated forward through the nonlinear
function f (·), that is,
−(i)
(i)
x̂k|k−1 = f x̂k−1|k−1 , i = 0, 1, · · · , 2nx .
(42)
These new sigma points are considered capable of capturing
the mean and covariance of f (xk−1 ). Using them, the prediction of xk can be achieved as follows:
2nx
X
Yk−1] = Wi(m)x̂−(i)
k|k−1 ,
i=0
x
= Cov [f (xk−1 ) |Yk−1 ] + Q
Pk|k−1
x̂k|k−1 = E [f (xk−1 ) |
=
2nx
X
(c)
Wi
−(i)
x̂k|k−1 − x̂k|k−1
(43)
>
−(i)
x̂k|k−1 − x̂k|k−1
i=0
+ Q.
(44)
By analogy, the sigma points for xk need to be generated
first in order to perform the update, which are
+(0)
x̂k|k−1 = x̂k|k−1 ,
+(i)
x̂k|k−1 = x̂k|k−1 +
+(i+n )
x̂k|k−1 x
(45)
p
hq
i
x
nx + λ
Pk|k−1
,
i
i = 1, 2, · · · , nx ,
hq
i
p
x
Pk|k−1
,
= x̂k|k−1 − nx + λ
(46)
i = 1, 2, · · · , nx .
(47)
i
Propagating them through h(·), we can obtain the sigma points
for h(xk ), given by
(i)
+(i)
(48)
ŷk|k−1 = h x̂k|k−1 , i = 0, 1, · · · , 2nx .
The predicted mean and covariance of yk and the predicted
cross-covariance between xk and yk are as follows:
2nx
X
(i)
Yk−1] = Wi(m)ŷk|k−1
,
i=0
y
Pk|k−1
= Cov [h(xk )|Yk−1 ] + R
ŷk|k−1 = E [yk |
=
2nx
X
(c)
Wi
(49)
>
(i)
(i)
ŷk|k−1 − ŷk|k−1 ŷk|k−1 − ŷk|k−1
+ R,
i=0
xy
Pk|k−1
2nx
X
=
Yk−1]
= Cov [xk , h(xk )|
(c)
Wi
+(i)
x̂k|k−1 − x̂k|k−1
(50)
(51)
>
(i)
ŷk|k−1 − ŷk|k−1 , (52)
i=0
With the above quantities becoming available, the Gaussian
update in (16)-(17) can be leveraged to enable the projection
from the predicted estimate x̂k|k−1 to the updated estimate
x̂k|k .
Summarizing the above equations leads to UKF sketched
in Fig. 5. Compared with EKF, UKF incurs a computational
Fig. 5. A schematic of UKF. Following the prediction-update procedure,
UKF tracks the mean and covariance of state xk using sigma points chosen
deterministically. A state estimate is graphically denoted by a red five-pointed
star mean surrounded by a covariance ellipse, and the sigma points are colored
in blue dots.
cost of the same order O(n3x ) but offers second-order accuracy [23], implying an overall smaller estimation error. In
addition, its operations are derivative-free, exempt from the
burdensome calculation of the Jacobian matrices in EKF. This
will not only bring convenience to practical implementation
but also imply its applicability to discontinuous undifferentiable nonlinear transformations. Yet, it is noteworthy that,
with a complexity of O(n3x ) and operations of 2nx + 1 sigma
points, UKF faces substantial computational expenses when
the system is high-dimensional with a large nx , thus unsuitable
for this kind of estimation problems.
Owing to its merits, UKF has seen a growing momentum
of research since its advent. A large body of work is devoted
to the development of modified versions. In this respect,
square-root UKF (SR-UKF) directly propagates a square root
matrix, which enjoys better numerical stability than squaring
the propagated covariance matrices [26]; iterative refinement
of the state estimate can also be adopted to enhance UKF
as in IEKF , leading to iterated UKF (IUKF). The performance of UKF can be improved by selecting the sigma
points in different ways. While the standard UKF employs
symmetrically distributed 2nx + 1 sigma points, asymmetric
point sets or sets with a larger number of points may bring
8
better accuracy [27–30]. Another interesting question is the
determination of the optimal scaling parameter κ, which is
investigated in [31]. UKF can be generalized to the so-called
sigma-point Kalman filtering (SPKF), which refers to the class
of filters that uses deterministic sampling points to determine
the mean and covariance of a random vector through nonlinear
transformation [32, 33]. Other SPKF techniques include the
central-difference Kalman filter (CDKF) and Gauss-Hermite
filter (GHKF), which perform sigma-point-based filtering and
can also be interpreted from the perspective of Gaussianquadrature-based filtering [34] (GHKF will receive further
discussion in Section VII).
where vki is generated as per the Gaussian distribution N (0, R)
to delineate the measurement noise vk . The sample mean of
this ensemble is
ŷk|k−1 =
Ns
1 X
(i)
ŷ
,
Ns i=1 k|k−1
with the associated sample covariance
N
y
=
Pk|k−1
s
X
1
(i)
ŷk|k−1 − ŷk|k−1
Ns − 1 i=1
>
(i)
· ŷk|k−1 − ŷk|k−1 .
The cross-covariance between xk and yk given
C. Ensemble Kalman Filter
Y
Y
The sample mean and covariance of this ensemble can be
calculated as:
Ns
X
Yk−1) ≈ N1s x̂(i)
k|k−1 ,
i=1
x
Pk|k−1
= Cov (xk |Yk−1 )
(54)
N
≈
s
>
X
1
(i)
(i)
x̂k|k−1 − x̂k|k−1 x̂k|k−1 − x̂k|k−1 ,
Ns − 1 i=1
(55)
which form the prediction formulae.
The update step begins with the construction of the ensemble for p (yk | k−1 ) by means of
(i)
(i)
(i)
ŷk|k−1 = h x̂k|k−1 + vk , i = 1, 2, · · · , Ns
(56)
Y
(58)
Yk−1 is
N
Since its early development in [35–37], ensemble Kalman
filter (EnKF) has established a strong presence in the field of
state estimation for large-scale nonlinear systems. Its design
is built on an integration of KF with the Monte Carlo method,
which is a prominent statistical method concerning simulationbased approximation of probability distributions using samples
directly drawn from certain distributions. Basically, EnKF
maintains an ensemble representing the conditional distribution of a random state vector given the measurement set.
The state estimate is generated from the sample mean and
covariance of the ensemble. In view of the sample-based
approximation of probability distributions, EnKF shares similarity with UKF; however, the latter employs deterministic
sampling while EnKF adopts non-deterministic sampling.
(i)
Suppose that there is an ensemble of samples, x̂k−1|k−1 for
i = 1, 2, · · · , Ns , drawn from p(xk−1 | k−1 ) to approximately
(i)
represent this pdf. Next, let an ensemble of samples, {wk−1 }
for i = 1, 2, · · · , Ns , be drawn independently and identically
from the Gaussian distribution N (0, Q) in order to account
(i)
for the process noise wk−1 at time k − 1. Then, x̂k−1|k−1 can
o
n
(i)
hence be projected to generate a priori ensemble x̂k|k−1
that represents p(xk | k−1 ) as follows:
(i)
(i)
(i)
x̂k|k−1 = f x̂k−1|k−1 + wk−1 , i = 1, 2, · · · , Ns (53)
x̂k|k−1 = E (xk |
(57)
xy
=
Pk|k−1
s
X
1
(i)
x̂k|k−1 − x̂k|k−1
Ns − 1 i=1
>
(i)
· ŷk|k−1 − ŷk|k−1 .
(59)
Once it arrives, the latest measurement yk can be applied to
update each member of a priori ensemble in the way defined
by (16), i.e.,
−1
(i)
(i)
(i)
xy
y
x̂k|k = x̂k|k−1 + Pk|k−1
Pk|k−1
yk − ŷk|k−1 ,
i = 1, 2, · · · , Ns .
(60)
n
o
(i)
This a posteriori ensemble x̂k|k can be regarded as an
approximate representation of p(xk | k ). Then, the updated
estimation of the mean and covariance of xk can be achieved
by
Y
x̂k|k =
Ns
1 X
x̂i ,
Ns i=1 k|k
(61)
N
x
Pk|k
s
>
X
1
(i)
(i)
.
x̂k|k − x̂k|k x̂k|k − x̂k|k
=
Ns − 1 i=1
(62)
The above ensemble-based prediction and update will be
repeated recursively, forming EnKF. Note that the computation
of estimation error covariance in (55) and (62) can be skipped
if a user has no interest in learning about the estimation
accuracy. This can further cut down EnKF’s storage and
computational cost.
EnKF is illustrated schematically in Fig. 6. It features
direct operation on the ensembles as a Monte Carlo-based
extension of KF. Essentially, it represents the pdf of a state
vector by using an ensemble of samples, propagates the
ensemble members and makes estimation by computing the
mean and covariance of the ensemble at each time instant. Its
complexity is O(n3y + n2y Ns + ny Ns2 + nx Ns2 ) (nx ny and
nx Ns for high-dimensional systems) [38], which contrasts
with O(n3x ) of EKF and UKF. This, along with the derivativefree computation and freedom from covariance matrix propagation, makes EnKF computationally efficient and appealing
to be the method of choice for high-dimensional nonlinear
systems. An additional contributing factor in this respect is
that the ensemble-based computational structure places it in an
advantageous position for parallel implementation [39]. It has
9
ψ̇dr = αLm ids − αψdr − ψqr ω,
ψ̇qr = αLm iqs + ψdr ω − αψqr ,
µ
TL
ω̇ = (−ids ψqr + ψdr iqs ) −
,
J
J
i
y = ds ,
iqs
Fig. 6. A schematic of EnKF. EnKF maintains an ensemble of sample points
for the state vector xk . It propagates and updates the ensemble to track the
distribution of xk . The state estimation is conducted by calculating the sample
mean (red five-pointed-star) and covariance (red ellipse) of the ensemble.
been reported that convergence of the EnKF can be fast even
with a reasonably small ensemble size [40, 41]. In particular,
its convergence to KF in the limit for large ensemble size and
Gaussian state probability distributions is proven in [41].
VI. A PPLICATION TO S PEED S ENSORLESS I NDUCTION
M OTORS
This section presents a case study of applying EKF, UKF
and EnKF to state estimation for speed sensorless induction
motors. Induction motors are used as an enabling component for numerous industrial systems, e.g., manufacturing
machines, belt conveyors, cranes, lifts, compressors, trolleys,
electric vehicles, pumps, and fans [42]. In an induction motor,
electromagnetic induction from the magnetic field of the stator
winding is used to generate the electric current that drives
the rotor to produce torque. This dynamic process must be
delicately controlled to ensure accurate and responsive operations. Hence, control design for this application was researched
extensively during the past decades, e.g., [42–44]. Recent years
have seen a growing interest in speed sensorless induction
motors, which have no sensors to measure the rotor speed to
reduce costs and increase reliability. However, the absence of
the rotor speed makes control design more challenging. To
address this challenge, state estimation is exploited to recover
the speed and other unknown variables. It is also noted that
an induction motor as a multivariable and highly nonlinear
system makes a valuable benchmark for evaluating different
state estimation approaches [44, 45].
The induction motor model in a stationary two-phase reference frame can be written as
i̇ds = −γids + αβψdr + βψqr ω + uds /σ,
i̇qs = −γiqs − βψdr ω + αβψqr + uqs /σ,
where (ψdr , ψqr ) is the rotor fluxes, (ids , iqs ) is the stator
currents, and (uds , uqs ) is the stator voltages, all defined in a
stationary d-q frame. In addition, ω is the rotor speed to be
estimated, J is the rotor inertia, TL is the load torque, and y
is the output vector composed of the stator currents. The rest
symbols are parameters, where σ = Ls (1 − L3m /Ls Lr ), α =
Rr /Lr , β = Lm /σLr , γ = Rs /σ + αβLm , µ = 3Lm /2Lr ;
(Rs , Ls ) and (Rr , Lr ) are the resistance-inductance pairs of
the stator and rotor, respectively; Lm is the mutual inductance.
As shown above, the state vector x comprises ids , iqs , ψdr ,
ψqr , and ω. The parameter setting follows [46]. Note that,
because of the focus on state estimation, an open-loop control
scheme is considered with uds (t) = 380 sin(100πt) and
uqs (t) = −380 sin(200πt). The state estimation problem is
then to estimate the entire state vector through time using the
measurement data of ids , iqs , uds and uqs .
In the simulation, the model is initialized with x0 =
[0 0 0 0 0]> . The initial state guess for all the filters is set to be
x̂0|0 = [1 1 1 1 1]> and P0x = 10−2 I. For EnKF, its estimation
accuracy depends on the ensemble size. Thus, different sizes
are implemented to examine this effect, with Ns = 40, 100,
200 and 400. To make a fair evaluation, EnKF with each Ns
is run for 100 times as a means to reduce the influence of
randomly generated
noise. The estimation error for each run
P
is defined as k kxk − x̂k|k k2 ; the errors from the 100 runs
are averaged to give the final estimation error for comparison.
Fig. 7 shows the estimation errors for ω along with ±3σ
bounds in a simulation run of EKF, UKF and EnKF with
ensemble size of 100 (here, σ stands for the standard deviation
associated with the estimate of ω, and ±3σ bounds correspond
to the 99% confidence region). It is seen that, in all three cases,
the error is large at the initial stage but gradually decreases
to a much lower level, indicating that the filters successfully
adapt their running according to their own equations. In
addition, UKF demonstrates the best estimation of ω overall.
The average estimation errors over 100 runs are summarized
in Table I. It also shows that UKF offers the most accurate
estimation when all state variables are considered. In addition,
the estimation accuracy of EnKF improves when the ensemble
size increases.
We draw the following remarks about nonlinear state estimation from our extensive simulations with this specific
problem and experience with other problems in our past
research.
•
The initial estimate can significantly impact the estimation accuracy. For the problem considered here, it is found
that EKF and EnKF are more sensitive to an initial guess.
It is noteworthy that an initial guess, if differing much
from the truth, can lead to divergence of filters. Hence,
one is encouraged to obtain a guess as close as possible to
10
50
50
Estimation error
3σ bounds
30
20
10
0
−10
30
20
10
0
−10
0
0.2
0.4
0.6
0.8
1
Estimation error
3σ bounds
40
Estimation error for ω
40
Estimation error for ω
Estimation error for ω
40
50
Estimation error
3σ bounds
30
20
10
0
−10
0
0.2
Time (second)
0.4
0.6
0.8
1
0
0.2
Time (second)
(a)
0.4
0.6
0.8
1
Time (second)
(b)
(c)
Fig. 7. Estimation error for ω: (a) EKF; (b) UKF; (c) EnKF with Ns = 100.
TABLE I
AVERAGE ESTIMATION ERRORS FOR EKF, UKF, AND E N KF.
Filter
EKF
UKF
Average estimation error
682.63
321.05
Ns = 40
347.24
EnKF
Ns = 100
Ns = 200
329.70
323.96
Ns = 400
320.30
TABLE II
C OMPARISON OF EKF, UKF AND E N KF.
•
•
•
Computational
complexity
Jacobian matrix
System dimensions
EKF
High
Needed
Low
UKF
High
Not needed
Low to medium
EnKF
Low
Not needed
High
the truth through using prior knowledge or trying different
guesses.
A filter’s performance can be problem-dependent. A filter
can provide estimation at a decent accuracy when applied
to a problem but may fail when used to handle another.
Thus, the practitioners are suggested to try different filters
whenever allowed to find out the one that performs the
best for his/her specific problem.
Successful application of a filter usually requires to tune
the covariance matrices and in some cases, parameters
involved in a filter (e.g., λ, α and β in UKF), because of
their important influence on estimation [47]. The trialand-error method is common in practice. Meanwhile,
there also exist some studies of systematic tuning methods, e.g., [48, 49]. Readers may refer to them for further
information.
In choosing the best filter, engineers need to take into
account all the factors relevant to the problem they are addressing, including but not limited to estimation accuracy,
computational efficiency, system’s structural complexity,
and problem size. To facilitate such a search, Table II
summarizes the main differences and application areas
of EKF, UKF and EnKF.
Applications
Guidance and navigation, flight control, attitude control, target
tracking, robotics (e.g., simultaneous localization and mapping),
electromechanical systems (e.g., induction motors and electric
drives), vibration control, biomedical signal processing, sensor
fusion, structural system monitoring, sensor networks, process
control, computer vision, battery management, HVAC systems,
econometrics
Meteorology, hydrology, weather forecasting, oceanography,
reservoir engineering, transportation systems, power systems
VII. OTHER F ILTERING A PPROACHES AND E STIMATION
P ROBLEMS
Nonlinear stochastic estimation remains a major research
challenge for the control research community. Continual research effort has been in progress toward the development of
advanced methods and theories in addition to the KFs reviewed
above. This section gives an overview of other major filtering
approaches.
Gaussian filters (GFs). GFs are a class of Bayesian filters
enabled by a series of Gaussian distribution approximations.
They bear much resemblance with KFs in view of their
prediction-update structure and thus, in a broad sense, belong
to the KF family. As seen earlier, the KF-based estimation
relies on the evaluation of a set of integrals indeed—for
example, the prediction of xk is attained in (6) by computing
the conditional mean of f (xk−1 ) on k−1 . The equation is
repeated here for convenience of reading:
Y
Y
x̂k|k−1 = E [f (xk−1 )| k−1 ]
Z
= f (xk−1 )p(xk−1 |
Y
Yk−1)dxk−1.
GFs approximate p(xk−1 | k−1 ) with a Gaussian distribution
x
having mean x̂k−1|k−1 and covariance Pk−1|k−1
. Namely,
11
Y
x
p(xk−1 | k−1 ) is replaced by N (x̂k−1|k−1 , Pk−1|k−1
) [34].
Continuing with this assumption, one can use the Gaussian
quadrature integration rules to evaluate the integral. A quadrature is a means of approximating a definite integral of a
function by a weighted sum of values obtained by evaluating
the function at a set of deterministic points in the domain
of integration. An example of a one-dimensional Gaussian
quadrature is the Gauss-Hermite quadrature, which plainly
states that, for a given function g(x),
Z ∞
m
X
g(x) · N (x; 0, 1)dx ≈
wi g(xi ),
−∞
i=1
where m is the number of points used, xi for i = 1, 2, · · · , m
the roots of the Hermite polynomial Hm (x), and wi the
associated weights
√
2m−1 m! π
wi =
2.
m2 [Hm−1 (xi )]
Exact equality holds for polynomials of order up to 2m − 1.
Applying the multivariate version of this quadrature, one
can obtain a filter in a KF form, which is named GaussHermite KF (GHKF) [34, 50]. GHKF reduces to UKF in
certain cases [34]. Besides, the cubature rules for numerical
integration can also be used in favor of a KF realization,
which yields a cubature Kalman filter (CKF) [51, 52]. It is
noteworthy that CKF is a special case of UKF given α = 1,
β = 0 and κ = 0 [53].
Gaussian-sum filters (GSFs). Though used widely in the
development of GFs and KFs, Gaussianity approximations are
often inadequate and performance-limiting for many systems.
To deal with a non-Gaussian pdf, GSFs represent it by a
weighted sum of Gaussian basis functions [7]. For instance,
the a posteriori pdf of xk is approximated by
m
X
i
Wki N xk ; x̂ik|k , Pk|k
p(xk | K ) =
,
Y
i=1
i
where Wki , x̂ik and Pk|k
are the weight, mean and covariance
of the i-th Gaussian basis function (kernel), respectively. This
can be justified by the Universal Approximation Theorem,
which states that a continuous function can be approximated
by a group of Gaussian functions with arbitrary accuracy
under some conditions [54]. A GSF then recursively updates
i
i
x̂ik|k , Pk|k
and Wki . In the basic form, x̂ik|k and Pk|k
for
i = 1, 2, · · · , m are updated individually through EKF, and
Wki updated according to the output-prediction accuracy of x̂ik .
The assumption for the EKF-based update is that the system’s
nonlinear dynamics can be well represented by aggregating
linearizations around a sufficient number of different points
(means). In recent years, more sophisticated GSFs have been
developed by merging the Gaussian-sum pdf approximation
with other filtering techniques such as UKF, EnKF, GFs and
particle filtering [34, 55–58] or optimization techniques [59].
Particle filters (PFs). The PF approach was first proposed
in the 1990s [60] and came to prominence soon after that owing to its capacity for high-accuracy nonlinear non-Gaussian
estimation. Today they have grown into a broad class of
filters. As random-sampling-enabled numerical realizations of
the Bayesian filtering principle, they are also known as the
sequential Monte Carlo methods in the literature. Here, we
introduce the essential idea with minimum statistical theory
to offer the reader a flavor of this approach. Suppose that
(i)
Ns samples, x̂k−1|k−1 for i = 1, 2, · · · , Ns are drawn from
p(xk−1 | k−1 ) at time k−1. The i-th sample is associated with
PNs
(i)
(i)
Wk−1 = 1. Then, p(xk−1 | k−1 )
a weight Wk−1 , and i=1
can be empirically described as
Y
Y
Yk−1) ≈
p(xk−1 |
Ns
X
(i)
(i)
Wk−1 δ xk−1 − x̂k−1|k−1 .
(63)
i=1
This indicates that the estimate of xk−1 is
Z
x̂k−1|k−1 = xk−1 p(xk−1 | k−1 )dxk−1
Y
=
N
X
(i)
(i)
Wk−1 x̂k−1|k−1 .
i=1
The samples can be propagated one-step forward to generate
a sampling-based description of xk , i.e.,
(i)
(i)
(i)
x̂k = f x̂k−1|k−1 + wk−1 , i = 1, 2, · · · , Ns ,
(i)
where wk−1 for i = 1, 2, · · · , Ns are samples drawn from the
distribution of wk−1 . After the propagation, each new sample
should take a different weight in order to be commensurate
with its probabilistic importance with
respect
to the others. To
(i)
account for this, one can evaluate p yk |x̂k , which quantifies
(i)
the likelihood of yk given the i-th sample x̂k . Then, the
weight can be updated and normalized on [0, 1] by
(i)
W
(i)
(i)
(i)
(i)
Wk = Wk−1 p yk |x̂k , Wk = PN k (i) .
s
i=1 Wk
Then, an empirical sample-based distribution is built for
p(xk | k ) as in (63), and the estimate of xk can be computed
as
Y
x̂k =
Ns
X
(i) (i)
Wk x̂k .
i=1
In practical implementation of the above procedure, the issue of degeneracy may arise, which refers to the scenario
that many or even most samples take almost zero weights.
Any occurrence of degeneracy renders the affected samples
useless. Remedying this situation requires the deployment
of resampling, which replaces the samples by new ones
drawn from the discrete empirical distribution defined by the
weights. Summarizing the steps of sample propagation, weight
update and resampling gives rise to a basic PF, which is
schematically shown in Fig. 8. While the above outlines a
reasonably intuitive explanation of the PF approach, a rigorous
development can be made on a solid statistical foundation, as
detailed in [16, 61, 62].
With the sample-based pdf approximation, PFs can demonstrate estimation accuracy superior to other filters given a
sufficiently large Ns . It can be proven that their estimation
error bound does not depend on the dimension of the system [64], implying applicability for high-dimensional systems.
12
+
k−1
X
wl> Q−1 wl +
l=k−N
subject to :
k
X
vl> R−1 vl ,
l=k−N
xl+1 = f (xl ) + wl ,
yl = h(xl ) + vl ,
x ∈ X, w ∈ W, v ∈ V,
Fig. 8. A graphic diagram of the PF technique modified from [63]. Suppose
that a set of samples (particles, as shown in gray color in the figure) are used to
approximate the conditional pdf of the state on the available measurements as
a particle discrete distribution. A one-step-froward propagation is implemented
to generate the samples for the state at the next time instant. On its arrival, the
new measurement will be used to update the weight of each sample to reflect
its importance relative to others. Some samples may be given almost zero
weight, referred to as degeneracy, and thus have meaningless contribution to
the state estimation. Resampling will then be performed to generate a new
set of samples.
A possible limitation is their computational complexity, which
comes at O(Ns n2x ) with Ns nx . Yet, a strong anticipation
is that the rapid growth of computing power tends to overcome
this limitation, enabling wider application of PFs. A plethora
of research has also been undertaken toward computationally efficient PFs [65]. A representative means is the RaoBlackwellization that applies the standard KF to the linear
part of a system and a PF to the nonlinear part and reduces
the number of samples to operate on [16]. The performance of
PFs often relies on the quality of samples used. To this end,
KFs can be used in combination to provide high-probability
particles for PFs, leading to a series of combined KF-PF
techniques [66–68]. A recent advance is the implicit PF, which
uses the implicit sampling method to generate samples capable
of an improved approximation of the pdf [69, 70].
Moving-horizon estimators (MHEs). MHEs are an emerging estimation tool based on constrained optimization. In
general, they aim to find the state estimate through minimizing
a cost function subject to certain constraints. The cost function
is formulated on the basis of the system’s behavior in a moving
horizon. To demonstrate the idea, we consider the Maximum
a Posteriori estimation (MAP) for the system in (1) during the
horizon [k − N, k] as shown in (64). Assuming wk ∼ N (0, Q)
and vk ∼ N (0, R) and using the logarithmic transformation,
the above cost function becomes
min
xk−N ,··· ,xk
Φ(xk−N ) +
k−1
X
wl> Q−1 wl
+
l=k−N
k
X
vl> R−1 vl ,
l=k−N
where Φ(xk−N ) is the arrival cost summarizing the past information up to the beginning of the horizon. The minimization
here should be subject to the system model in (1). Meanwhile,
some physically motivated constraints for the system behavior
should be incorporated. This produces the formulation of MHE
given as
{x̂k−N , · · · , x̂k } = arg min Φ(xk−N )
xk−N ,··· ,xk
where X, W and V are, respectively, the sets of all feasible
values for x, w and v and imposed as the constraints. It is
seen that MHE tackles the state estimation through constrained
optimization executed over time in a receding-horizon manner,
as shown in Fig. 9. For an unconstrained linear system,
MHE reduces to the standard KF. It is worth noting that the
arrival cost Φ(xk−N ) is crucial for the performance or even
success of MHE approach. In practice, an exact expression is
often unavailable, thus requiring an approximation [71, 72].
With the deployment of constrained optimization, MHE is
computationally expensive and usually more suited for slow
dynamic processes; however, the advancement of real-time
optimization has brought some promises to its faster implementation [73, 74].
Simultaneous state and parameter estimation (SSPE). In
state estimation problems, a system model is considered fully
known a priori. This may not be true in various real-world
situations, where part or even all of the model parameters are
unknown or subject to time-varying changes. Lack of knowledge of the parameters can disable an effort for state estimation
in such a scenario. Hence, SSPE is motivated to enable state
estimation self-adapting to the unknown parameters. Despite
the complications, a straightforward and popular way for SSPE
is through state augmentation. To deal with the parameters, the
state vector is augmented to include them, and on account of
this, the state-space model will be transformed accordingly
to one with increased dimensions. Then, a state estimation
technique can be applied directly to the new model to estimate
the augmented state vector, which is a joint estimation of
the state variables and parameters. In the literature, EKF,
UKF, EnKF and PFs have been modified using this idea for
a broad range of applications [20, 75–78]. Another primary
solution is the so-called dual Kalman filtering. By “dual”,
it means that the state estimation and parameter estimation
are performed in parallel and alternately. As such, the state
estimate is used to estimate the parameters, and the parameter
estimate is used to update the state estimation. Proceeding
with this idea, EKF, UKF and EnKF can be dualized [79–
82]. It should be pointed out that caution should be taken
when an SSPE approach is developed. Almost any SSPE
problem is nonlinear by nature with coupling between state
variables and parameters. The joint state observability and
parameter identifiability may be unavailable, or the estimation
may get stuck in local minima. Consequently, the convergence
can be vulnerable or unguaranteed, diminishing the chance of
successful estimation. Thus application-specific SSPE analysis
and development are recommended.
Simultaneous state and input estimation (SSIE). Some
practical applications entail not only unknown states but also
unknown inputs. An example is the operation monitoring for
an industrial system subject to unknown disturbance, where the
13
{x̂k−N , · · · , x̂k } = arg
= arg
= arg
max
p(xk−N , · · · , xk |yk−N , · · · , yk )
max
p(xk−N )
xk−N ,··· ,xk
xk−N ,··· ,xk
max
xk−N ,··· ,xk
k
Y
p(yl |xl )
l=k−N
p(xk−N )
k
Y
l=k−N
k−1
Y
p(xl+1 |xl )
l=k−N
p(vl )
k−1
Y
p(wl ).
(64)
l=k−N
approaches may find their use in solving SSPE problems.
Fig. 9. MHE performed over a moving observation horizon that spans
N + 1 time instants. For estimation at time k, the arrival cost Φ(xk−N )
is determined first, which summarizes the information of the system behavior
up to the beginning of the horizon. Then, the output measurements within
the horizon, yk−N , · · · , yk , are used, along with the arrival cost, to conduct
estimation of x̂k−N , · · · , x̂k through constrained optimization.
operational status is the state and the disturbance the input.
In maneuvering target tracking, the tracker often wants to
estimate the state of the target, e.g., position and velocity,
and the input, e.g., the acceleration. Another example is
the wildfire data assimilation extensively investigated in the
literature. The spread of wildfire is often driven by local
meteorological conditions such as the wind. This gives rise
to the need for a joint estimation of both the fire perimeters
(state) and the wind speed (input) toward accurate monitoring
of the fire growth.
The significance of SSIE has motivated a large body of
work. A lead was taken in [83] with the development of a
KF-based approach to estimate the state and external white
process noise of a linear discrete-time system [83]. Most recent
research builds on the existing state estimation techniques.
Among them, we highlight KF [84, 85], MHE [86], H∞ filtering [87], sliding mode observers [88, 89], and minimumvariance unbiased estimation [90–94]. SSIE for nonlinear
systems involves more complexity, with fewer results reported.
In [95, 96], SSIE is investigated for a special class of nonlinear
systems that consist of a nominally linear part and a nonlinear
part. However, the Bayesian statistical thinking has been
generalized to address this topic, exemplifying its power in
the development of nonlinear SSIE approaches. In [97, 98], a
Bayesian approach along with numerical optimization is taken
to achieve SSIE for nonlinear systems of a general form. This
Bayesian approach is further extended in [99, 100] to build
an ensemble-based SSIE method, as a counterpart of EnKF,
for high-dimensional nonlinear systems. It is noteworthy that
SSIE and SSPE would overlap if we consider the parameters
as a special kind of inputs to the system. In this case, the SSIE
VIII. C ONCLUSION
This article offered a state-of-the-art review of nonlinear
state estimation approaches. As a fundamental problem encountered across a few research areas, nonlinear stochastic
estimation has stimulated a sustaining interest during the past
decades. In the pursuit of solutions, the Bayesian analysis
has proven to be a time-tested and powerful methodology for
addressing various types of problems. In this article, we first
introduced the Bayesian thinking for nonlinear state estimation, showing the nature of state estimation from the perceptive
of Bayesian update. Based on the notion of Bayesian state
estimation, a general form of the celebrated KF is derived.
Then, we illustrated the development of the standard KF
for linear systems and EKF, UKF and EnKF for nonlinear
systems. A case study of state estimation for speed sensorless
induction motors was provided to present a comparison of the
EKF, UKF and EnKF approaches. We further extended our
view to a broader horizon including GF, GSF, PF and MHE
approaches, which are also deeply rooted in the Bayesian state
estimation and thus can be studied from a unified Bayesian
perspective to a large extent.
Despite remarkable progress made thus far, it is anticipated
that nonlinear Bayesian estimation continues to see intensive
research in the coming decades. This trend will be partially
driven by the need to use state estimation as a mathematical
tool to enable various emerging systems in contemporary
industry and society, stretching from autonomous transportation to sustainable energy and smart X (grid, city, planet,
geosciences, etc.). Here, we envision several directions that
may shape the future research in this area. The first one
lies in accurately characterizing the result of a nonlinear
transformation applied to a probability distribution. Many of
the present methods such as EKF, UKF and EnKF were
more or less motivated to address this fundamental challenge.
However, there still exists no solution generally acknowledged
as being satisfactory, leaving room for further exploration.
Second, much research is needed to deal with uncertainty.
Uncertainty is intrinsic to many practical systems because of
unmodeled dynamics, external disturbances, inherent variability of a dynamic process, and sensor noise, representing a
major threat to successful estimation. Although the literature
contains many results on state estimation with robustness to
uncertainty, the research has not reached a level of maturity
because of the difficulty involved. A third research direction is optimal sensing structure design. Sensing structure
14
or sensor deployment is critical for data informativeness and
thus can significantly affect the effectiveness of estimation.
An important question thus is how to achieve co-design
of a sensing structure and Bayesian estimation approach to
maximize estimation accuracy. Fourth, Bayesian estimation
in a cyber-physical setting is an imperative. Standing at the
convergence of computing, communication and control, cyberphysical systems (CPSs) are foundationally important and
underpinning today’s smart X initiatives. They also present
new challenges for estimation, which include communication
constraints or failures, computing limitations, and cyber data
attacks. The current research is particularly rare on nonlinear
Bayesian estimation for CPSs. Finally, many emerging industrial and social applications are data-intensive, thus asking
for a seamless integration of Bayesian estimation with big
data processing algorithms. New principles, approaches and
computing tools must be developed to meet this pressing need,
which will make an unprecedented opportunity to advance the
Bayesian estimation theory.
A PPENDIX
This appendix offers a summary of the properties of the
Gaussian distribution. Suppose z ∈ n is a Gaussian random
vector with z ∼ N (z̄, Pz ). The pdf of z is expressed as
1
exp −(z − z̄)Pz−1 (z − z̄)> .
p(z) = p
n
( (2π) |Pz |
R
Some useful properties of the Gaussian vectors are as follows [101].
1)
Z
Z
zp(z)dz = z̄,
(z − z̄)(z − z̄)> p(z)dz = Pz ,
Z
zz > p(z)dz = Pz + z̄ z̄ > .
(A.1)
2) The affine transformation of z, Az + b, is Gaussian, i.e.,
Az + b ∼ N Az̄ + b, APz A> .
(A.2)
3) The sum of two independent Gaussian random vectors is
Gaussian; i.e., if z1 ∼ N (z̄1 , Pz1 ) and z2 ∼ N (z̄2 , Pz2 )
and if z1 and z2 are independent, then
Az1 + Bz2 ∼ N Az̄1 + B z̄2 , APz1 A> + BPz2 B > .
(A.3)
4) For two random vectors jointly Gaussian, the conditional
distribution of one given the other is Gaussian. Specifically, if z1 and z2 are jointly Gaussian with
P
Pz1 z2
z1
z̄1
∼N
, >z1
,
z2
z̄2
Pz1 z2 Pz2
then
z1 |z2 ∼ N z̄1 + Pz1 z2 Pz−1
(z2 − z̄2 ),
2
>
Pz1 − Pz1 z2 Pz−1
P
.
z
z
2
1 2
(A.4)
R EFERENCES
[1] Mr. Bayes and Mr. Price, “An essay towards solving a problem in the doctrine of chances,” Philosophical Transactions,
vol. 53, pp. 370–418, 1763.
[2] A. Dale, “Bayes or Laplace? An examination of the origin and
early applications of bayes’ theorem,” Archive for History of
Exact Sciences, vol. 27, no. 1, pp. 23–47, 1982.
[3] D. Malakoff, “Bayes offers a ‘New’ way to make sense of
numbers,” Science, vol. 286, no. 5444, pp. 1460–1464, 1999.
[4] A. Jackson, “Kalman receives national medal of science,”
Notices of the AMS, vol. 57, no. 1, pp. 56–57, 2010.
[5] R. E. Kalman, “A new approach to linear filtering and prediction problems,” Journal of Basic Engineering, vol. 82, no. 1,
pp. 35–45, 1960.
[6] Y.-C. Ho and R. Lee, “A Bayesian approach to problems in
stochastic estimation and control,” IEEE Transactions on on
Automatic Control, vol. 9, no. 4, pp. 333–339, 1964.
[7] B. D. Anderson and J. B. Moore, Optimal filtering. PrenticeHall, 1979.
[8] Wikipedia contributors, “Kalman filter — Wikipedia,
the free encyclopedia,” [Online; accessed 24-September2016]. [Online]. Available: \url{https://en.wikipedia.org/wiki/
Kalman filter}
[9] F. Auger, M. Hilairet, J. Guerrero, E. Monmasson,
T. Orlowska-Kowalska, and S. Katsura, “Industrial applications of the Kalman filter: A review,” IEEE Transactions on
Industrial Electronics, vol. 60, no. 12, pp. 5458–5471, 2013.
[10] M. Boutayeb, H. Rafaralahy, and M. Darouach, “Convergence
analysis of the extended Kalman filter used as an observer for
nonlinear deterministic discrete-time systems,” IEEE Transactions on Automatic Control, vol. 42, no. 4, pp. 581–586, 1997.
[11] A. Krener, “The convergence of the extended Kalman filter,” in
Directions in Mathematical Systems Theory and Optimization,
A. Rantzer and C. Byrnes, Eds. Springer, 2003, vol. 286, pp.
173–182.
[12] S. Kluge, K. Reif, and M. Brokate, “Stochastic stability of the
extended Kalman filter with intermittent observations,” IEEE
Transactions on Automatic Control, vol. 55, no. 2, pp. 514–
518, 2010.
[13] S. Bonnabel and J.-J. Slotine, “A contraction theory-based
analysis of the stability of the deterministic extended Kalman
filter,” IEEE Transactions on Automatic Control, vol. 60, no. 2,
pp. 565–569, 2015.
[14] A. Barrau and S. Bonnabel, “The invariant extended kalman
filter as a stable observer,” IEEE Transactions on Automatic
Control, vol. 62, no. 4, pp. 1797–1812, 2017.
[15] H. Tanizaki, Nonlinear Filters: Estimation and Applications.
Springer-Verlag Berlin Heidelberg, 1996.
[16] S. Särkkä, Bayesian Filtering and Smoothing. Cambridge
University Press, 2013.
[17] M. Roth and F. Gustafsson, “An efficient implementation of
the second order extended Kalman filter,” in Proceedings
of the 14th International Conference on Information Fusion
(FUSION), 2011, pp. 1–6.
[18] A. H. Jazwinski, Stochastic processes and filtering theory.
New York, NY, USA: Academic Press, 1970.
[19] B. Bell and F. Cathey, “The iterated Kalman filter update
as a gauss-newton method,” IEEE Transactions on Automatic
Control, vol. 38, no. 2, pp. 294–297, 1993.
[20] H. Fang, Y. Wang, Z. Sahinoglu, T. Wada, and S. Hara, “State
of charge estimation for lithium-ion batteries: An adaptive
approach,” Control Engineering Practice, vol. 25, pp. 45–54,
2014.
[21] S. Julier, J. Uhlmann, and H. Durrant-Whyte, “A new approach
for filtering nonlinear systems,” in Proceedings of American
Control Conference, vol. 3, 1995, pp. 1628–1632.
[22] S. J. Julier and J. K. Uhlmann, “A new extension of the Kalman
filter to nonlinear systems,” in Proceedings of AeroSense: The
11th International Symposium on Aerospace/Defence Sensing,
Simulation and Controls, 1997, pp. 182–193.
15
[23] S. Julier and J. Uhlmann, “Unscented filtering and nonlinear
estimation,” Proceedings of the IEEE, vol. 92, no. 3, pp. 401–
422, 2004.
[24] E. Wan and R. van der Merwe, “The unscencted Kalman filter,”
in Kalman Filtering and Neural Networks, S. Haykin, Ed.
John Wiley & Sons, 2001, pp. 221–269.
[25] H. M. T. Menegaz, J. Y. Ishihara, G. A. Borges, and A. N.
Vargas, “A systematization of the unscented kalman filter
theory,” IEEE Transactions on Automatic Control, vol. 60,
no. 10, pp. 2583–2598, 2015.
[26] R. Van Der Merwe and E. Wan, “The square-root unscented
Kalman filter for state and parameter-estimation,” in Proceedings of IEEE International Conference on Acoustics, Speech,
and Signal Processing, vol. 6, 2001, pp. 3461–3464.
[27] S. Julier, “The spherical simplex unscented transformation,”
in Proceedings of the American Control Conference, vol. 3,
2003, pp. 2430–2434.
[28] W.-C. Li, P. Wei, and X.-C. Xiao, “A novel simplex unscented
transform and filter,” in Proceedings on International Symposium on Communications and Information Technologies, 2007,
pp. 926–931.
[29] P. Date, K. Ponomareva, and Z. Wang, “A new unscented
Kalman filter with higher order moment-matching,” in Proceedings of Mathematical Theory of Networks and Systems,
2010, pp. 926–931.
[30] A. Charalampidis and G. Papavassilopoulos, “Development
and numerical investigation of new non-linear Kalman filter
variants,” IET Control Theory Applications, vol. 5, no. 10, pp.
1155–1166, 2011.
[31] J. Dunik, M. Simandl, and O. Straka, “Unscented Kalman
filter: Aspects and adaptive setting of scaling parameter,” IEEE
Transactions on Automatic Control, vol. 57, no. 9, pp. 2411–
2416, 2012.
[32] J. Crassidis, “Sigma-point Kalman filtering for integrated GPS
and inertial navigation,” IEEE Transactions on Aerospace and
Electronic Systems, vol. 42, no. 2, pp. 750–756, 2006.
[33] R. van der Merwe, “Sigma-point Kalman filters for probabilistic inference in dynamic state-space models,” Ph.D. dissertation, Oregon Health & Science University, 2004.
[34] K. Ito and K. Xiong, “Gaussian filters for nonlinear filtering
problems,” IEEE Transactions on Automatic Control, vol. 45,
no. 5, pp. 910–927, 2000.
[35] G. Evensen, “Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to
forecast error statistics,” Journal of Geophysical Research:
Oceans, vol. 99, no. C5, pp. 10 143–10 162, 1994.
[36] G. Evensen and P. J. van Leeuwen, “Assimilation of geosat altimeter data for the agulhas current using the ensemble Kalman
filter with a quasigeostrophic model,” Monthly Weather Review, vol. 124, pp. 85–96, 1996.
[37] P. L. Houtekamer and H. L. Mitchell, “Data assimilation
using an ensemble Kalman filter technique,” Monthly Weather
Review, vol. 126, pp. 796–811, 1998.
[38] J. Mandel, “Efficient implementation of the ensemble Kalman
filter,” University of Colorado at Denver, Tech. Rep.
UCDHSC/CCM Report No. 231, May 2006.
[39] S. Lakshmivarahan and D. Stensrud, “Ensemble Kalman filter,”
IEEE Control Systems, vol. 29, no. 3, pp. 34–46, 2009.
[40] G. Evensen, Data Assimilation: The Ensemble Kalman Filter,
2nd ed. Springer, 2009.
[41] J. Mandel, L. Cobb, and J. Beezley, “On the convergence of the
ensemble Kalman filter,” Applications of Mathematics, vol. 56,
no. 6, pp. 533–541, 2011.
[42] M. Montanari, S. M. Peresada, C. Rossi, and A. Tilli, “Speed
sensorless control of induction motors based on a reducedorder adaptive observer,” IEEE Transactions on Control Systems Technology, vol. 15, no. 6, pp. 1049–1064, 2007.
[43] S. R. Bowes, A. Sevinc, and D. Holliday, “New natural
observer applied to speed-sensorless dc servo and induction
motors,” IEEE Transactions on Industrial Electronics, vol. 51,
no. 5, pp. 1025–1032, 2004.
[44] R. Marino, P. Tomei, and C. Verrelli, “A global tracking control
for speed-sensorless induction motors,” Automatica, vol. 40,
no. 6, pp. 1071–1077, 2004.
[45] Y. Wang, L. Zhou, S. A. Bortoff, A. Satake, and S. Furutani,
“High gain observer for speed-sensorless motor drives: Algorithm and experiments,” in Proceedings of IEEE International
Conference on Advanced Intelligent Mechatronics, 2016, pp.
1127–1132.
[46] L. Zhou and Y. Wang, “Speed sensorless state estimation for
induction motors: A moving horizon approach,” in Proceedings of American Control Conference, 2016, pp. 2229–2234.
[47] Q. Ge, T. Shao, Z. Duan, and C. Wen, “Performance analysis
of the kalman filter with mismatched noise covariances,” IEEE
Transactions on Automatic Control, vol. 61, no. 12, pp. 4014–
4019, 2016.
[48] L. A. Scardua and J. J. da Cruz, “Complete offline tuning of
the unscented kalman filter,” Automatica, vol. 80, pp. 54 – 61,
2017.
[49] B. F. La Scala, R. R. Bitmead, and B. G. Quinn, “An extended
kalman filter frequency tracker for high-noise environments,”
IEEE Transactions on Signal Processing, vol. 44, no. 2, pp.
431–434, 1996.
[50] B. Jia, M. Xin, and Y. Cheng, “Sparse Gauss-Hermite quadrature filter with application to spacecraft attitude estimation,”
AIAA Journal of Guidance, Control, and Dynamics, vol. 34,
no. 2, pp. 367–379, 2011.
[51] I. Arasaratnam and S. Haykin, “Cubature Kalman filters,”
IEEE Transactions on Automatic Control, vol. 54, no. 6, pp.
1254–1269, 2009.
[52] B. Jia, M. Xin, and Y. Cheng, “High-degree cubature Kalman
filter,” Automatica, vol. 49, no. 2, pp. 510–518, 2013.
[53] S. Särkkä, J. Hartikainen, L. Svensson, and F. Sandblom. On
the relation between Gaussian process quadratures and sigmapoint methods.
[54] J. Park and I. W. Sandberg, “Universal approximation using
radial-basis-function networks,” Neural Computation, vol. 3,
no. 2, pp. 246–257, 1991.
[55] O. Straka, J. Dunik, and M. Simandl, “Gaussian sum unscented
Kalman filter with adaptive scaling parameters,” in Proceedings of the 14th International Conference on Information
Fusion, 2011, pp. 1–8.
[56] L. Dovera and E. Della Rossa, “Multimodal ensemble Kalman
filtering using Gaussian mixture models,” Computational Geosciences, vol. 15, no. 2, pp. 307–323, 2011.
[57] J. Horwood and A. Poore, “Adaptive Gaussian sum filters for
space surveillance,” IEEE Transactions on Automatic Control,
vol. 56, no. 8, pp. 1777–1790, 2011.
[58] J. H. Kotecha and P. Djuric, “Gaussian sum particle filtering,”
IEEE Transactions on Signal Processing, vol. 51, no. 10, pp.
2602–2612, 2003.
[59] G. Terejanu, P. Singla, T. Singh, and P. Scott, “Adaptive
Gaussian sum filter for nonlinear Bayesian estimation,” IEEE
Transactions on Automatic Control, vol. 56, no. 9, pp. 2151–
2156, 2011.
[60] N. Gordon, D. Salmond, and A. Smith, “Novel approach to
nonlinear/non-Gaussian Bayesian state estimation,” Radar and
Signal Processing, IEE Proceedings F, vol. 140, no. 2, pp.
107–113, 1993.
[61] A. Doucet and A. M. Johansen, “Tutorial on particle filtering
and smoothing: Fifteen years later,” in The Oxford Handbook
of Nonlinear Filtering, D. Crisan and B. Rozovskii, Eds.
Oxford University Press, 2011, pp. 656–704.
[62] M. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A
tutorial on particle filters for online nonlinear/non-Gaussian
Bayesian tracking,” IEEE Transactions on Signal Processing,
vol. 50, no. 2, pp. 174–188, 2002.
[63] A. Parker, “Dealing with data overload in the scientific realm,”
pp. 4–11, January/February 2013.
16
[64] D. Crisan and A. Doucet, “Convergence of sequential Monte
Carlo methods,” 2000, technical Report, University of Cambridge.
[65] W. Yi, M. Morelande, L. Kong, and J. Yang, “A computationally efficient particle filter for multitarget tracking using an
independence approximation,” IEEE Transactions on Signal
Processing, vol. 61, no. 4, pp. 843–856, 2013.
[66] A. H. G. J. F. G. de Freitas, M. Niranjan and A. Doucet,
“Sequential monte carlo methods to train neural network
models,” Neural Computation, vol. 12, no. 4, pp. 955–993,
2000.
[67] R. van der Merwe, N. de Freitas, A. Doucet, and E. Wan, “The
unscented particle filter,” in Advances in Neural Information
Processing Systems 13, 2001, pp. 584–590.
[68] L. Li, H. Ji, and J. Luo, “The iterated extended Kalman particle
filter,” in Proceedings of IEEE International Symposium on
Communications and Information Technology, vol. 2, 2005,
pp. 1213–1216.
[69] A. J. Chorin and X. Tu, “Implicit sampling for particle filters,”
Proceedings of the National Academy of Sciences, vol. 106,
no. 41, pp. 17 429–17 254, 2009.
[70] A. Chorin, M. Morzfeld, and X. Tu, “Implicit particle filters for
data assimilation,” Communications in Applied Mathematics
and Computational Science, vol. 5, no. 2, pp. 221–240, 2010.
[71] C. Rao, J. Rawlings, and D. Mayne, “Constrained state estimation for nonlinear discrete-time systems: stability and moving
horizon approximations,” IEEE Transactions on Automatic
Control, vol. 48, no. 2, pp. 246–258, 2003.
[72] C. V. Rao, “Moving horizon strategies for the constrained
monitoring and control of nonlinear discrete-time systems,”
Ph.D. dissertation, University of Wisconsin Madison, 2000.
[73] M. Tenny and J. Rawlings, “Efficient moving horizon estimation and nonlinear model predictive control,” in Proceedings
of American Control Conference, vol. 6, 2002, pp. 4475–4480.
[74] M. Diehl, H. Ferreau, and N. Haverbeke, “Efficient numerical
methods for nonlinear mpc and moving horizon estimation,” in
Nonlinear Model Predictive Control, L. Magni, D. Raimondo,
and F. Allg´’ower, Eds. Springer Berlin Heidelberg, 2009, vol.
384, pp. 391–417.
[75] J. Ching, J. L. Beck, and K. A. Porter, “Bayesian state
and parameter estimation of uncertain dynamical systems,”
Probabilistic Engineering Mechanics, vol. 21, no. 1, pp. 81–
96, 2006.
[76] H. Fang, Y. Wang, Z. Sahinoglu, T. Wada, and S. Hara, “Adaptive estimation of state of charge for lithium-ion batteries,” in
Proceedings of American Control Conference, 2013, pp. 3485–
3491.
[77] G. Chowdhary and R. Jategaonkar, “Aerodynamic parameter
estimation from flight data applying extended and unscented
kalman filter,” Aerospace Science and Technology, vol. 14,
no. 2, pp. 106–117, 2010.
[78] T. Chen, J. Morris, and E. Martin, “Particle filters for state and
parameter estimation in batch processes,” Journal of Process
Control, vol. 15, no. 6, pp. 665–673, 2005.
[79] E. A. Wan and A. T. Nelson, “Dual extended Kalman filter methods,” in Kalman Filtering and Neural Networks,
S. Haykin, Ed. John Wiley & Sons, Inc., 2001, pp. 123–
173.
[80] E. A. Wan, R. V. D. Merwe, and A. T. Nelson, “Dual
estimation and the unscented transformation,” in Advances in
Neural Information Processing Systems. MIT Press, 2000,
pp. 666–672.
[81] M. C. Vandyke, J. L. Schwartz, and C. D. Hall, “Unscented
Kalman filtering for spacecraft attitude state and parameter
estimation,” in Proceedings of the AAS/AIAA Space Flight
Mechanics Conference, 2004.
[82] H. Moradkhani, S. Sorooshian, H. V. Gupta, and P. R. Houser,
“Dual stateparameter estimation of hydrological models using ensemble Kalman filter,” Advances in Water Resources,
vol. 28, no. 2, pp. 135–147, 2005.
[83] J. Mendel, “White-noise estimators for seismic data processing
in oil exploration,” IEEE Transactions on Automatic Control,
vol. 22, no. 5, pp. 694–706, 1977.
[84] C.-S. Hsieh, “Robust two-stage Kalman filters for systems with
unknown inputs,” IEEE Transactions on Automatic Control,
vol. 45, no. 12, pp. 2374–2378, 2000.
[85] ——, “On the optimality of two-stage Kalman filtering for
systems with unknown inputs,” Asian Journal of Control,
vol. 12, no. 4, pp. 510–523, 2010.
[86] L. Pina and M. A. Botto, “Simultaneous state and input estimation of hybrid systems with unknown inputs,” Automatica,
vol. 42, no. 5, pp. 755–762, 2006.
[87] F.-Q. You, F.-L. Wang, and S.-P. Guan, “Hybrid estimation
of state and input for linear discrete time-varying systems: A
game theory approach,” Acta Automatica Sinica, vol. 34, no. 6,
pp. 665–669, 2008.
[88] T. Floquet, C. Edwards, and S. K. Spurgeon, “On sliding mode
observers for systems with unknown inputs,” International
Journal of Adaptive Control & Signal Processing, vol. 21, no.
8-9, pp. 638–656, 2007.
[89] F. Bejarano, L. Fridman, and A. Poznyak, “Exact state estimation for linear systems with unknown inputs based on
hierarchical super-twisting algorithm,” International Journal
of Robust and Nonlinear Control, vol. 17, no. 18, pp. 1734–
1753, 2007.
[90] S. Gillijns and B. De Moor, “Unbiased minimum-variance
input and state estimation for linear discrete-time systems with
direct feedthrough,” Automatica, vol. 43, no. 5, pp. 934–937,
2007.
[91] ——, “Unbiased minimum-variance input and state estimation
for linear discrete-time systems,” Automatica, vol. 43, no. 1,
pp. 111–116, 2007.
[92] H. Fang, Y. Shi, and J. Yi, “A new algorithm for simultaneous input and state estimation,” in Proceedings of American
Control Conference, 2008, pp. 2421–2426.
[93] ——, “On stable simultaneous input and state estimation for
discrete-time linear systems,” International Journal of Adaptive Control and Signal Processing, vol. 25, no. 8, pp. 671–686,
2011.
[94] H. Fang and R. A. de Callafon, “On the asymptotic stability
of minimum-variance unbiased input and state estimation,”
Automatica, vol. 48, no. 12, pp. 3183–3186, 2012.
[95] M. Corless and J. Tu, “State and input estimation for a class of
uncertain systems,” Automatica, vol. 34, no. 6, pp. 757–764,
1998.
[96] Q. P. Ha and H. Trinh, “State and input simultaneous estimation for a class of nonlinear systems,” Automatica, vol. 40,
no. 10, pp. 1779–1785, 2004.
[97] H. Fang, R. A. de Callafon, and J. Cortés, “Simultaneous input
and state estimation for nonlinear systems with applications to
flow field estimation,” Automatica, vol. 49, no. 9, pp. 2805 –
2812, 2013.
[98] H. Fang, R. A. de Callafon, and P. J. S. Franks, “Smoothed
estimation of unknown inputs and states in dynamic systems
with application to oceanic flow field reconstruction,” International Journal of Adaptive Control and Signal Processing,
vol. 29, no. 10, pp. 1224–1242, 2015.
[99] H. Fang and R. de Callafon, “Simultaneous input and state
filtering: An ensemble approach,” in Proceedings of IEEE 52nd
Annual Conference on Decision and Control (CDC), 2013, pp.
7690–7695.
[100] H. Fang, T. Srivas, R. A. de Callafon, and M. A. Haile,
“Ensemble-based simultaneous input and state estimation for
nonlinear dynamic systems with application to wildfire data
assimilation,” Control Engineering Practice, vol. 63, pp. 104–
115, 2017.
[101] J. V. Candy, Bayesian Signal Processing: Classical, Modern
and Particle Filtering Methods. New York, NY, USA: WileyInterscience, 2009.
| 3cs.SY
|
Margins of discrete Bayesian networks
Robin J. Evans
arXiv:1501.02103v2 [math.ST] 30 Jan 2017
January 31, 2017
Abstract
Bayesian network models with latent variables are widely used in
statistics and machine learning. In this paper we provide a complete
algebraic characterization of Bayesian network models with latent variables when the observed variables are discrete and no assumption is made
about the state-space of the latent variables. We show that it is algebraically equivalent to the so-called nested Markov model, meaning that
the two are the same up to inequality constraints on the joint probabilities. In particular these two models have the same dimension. The nested
Markov model is therefore the best possible description of the latent variable model that avoids consideration of inequalities, which are extremely
complicated in general. A consequence of this is that the constraint finding algorithm of Tian and Pearl (2002) is complete for finding equality
constraints.
Latent variable models suffer from difficulties of unidentifiable parameters and non-regular asymptotics; in contrast the nested Markov model
is fully identifiable, represents a curved exponential family of known dimension, and can easily be fitted using an explicit parameterization.
1
Introduction
Directed acyclic graph (DAG) models, also known as Bayesian network models,
are widely used multivariate models in probabilistic reasoning, machine learning
and causal inference (Bishop, 2007; Darwiche, 2009; Pearl, 2009). These models
are defined by simple factorizations of the joint distribution, and in the case of
discrete or jointly Gaussian random variables, are curved exponential families
of known dimension. The inclusion of latent variables within Bayesian network
models can greatly increase their flexibility, and also account for unobserved
confounding. However, this flexibility comes at the cost of creating models that
are not easy to explicitly describe when considered as marginal models over the
observed variables. Latent variable models generally do not have fully identifiable parameterizations, and contain ‘singularities’ that lead to non-regular
asymptotics (Drton, 2009a). In addition, using them may force a modeller to
specify a parametric structure over the latent variables, introducing additional
assumptions that are generally difficult to test and may be unreasonable.
If no parametric assumptions are made about the latent variables, and no
assumption is made about their state-space, this leads to an implicitly defined
marginal model. The marginal model has the advantage of avoiding some of the
assumptions made by a parametric latent variable model, however no explicit
1
0
1
2
3
4
Figure 1: A directed acyclic graph on five vertices.
characterization of the model is available, and nor is there any obvious method
for fitting it to data.
Example 1.1. Consider the DAG on five vertices shown in Figure 1. The graph
represents a multivariate model over five random variables X0 , X1 , X2 , X3 and
X4 , with the restriction that the joint density factorizes as
p(x0 , x1 , x2 , x3 , x4 ) = p(x0 ) · p(x1 ) · p(x2 | x0 , x1 ) · p(x3 | x2 ) · p(x4 | x0 , x3 );
here, for example, p(x3 | x2 ) represents the conditional density of X3 given X2 .
This model arises naturally in the context of dynamic treatment regimes and
longitudinal exposures (Robins, 1986): X1 and X3 represent treatments and X2
and X4 some outcome of interest. The treatments are randomized, though the
second treatment X3 may depend upon the first outcome X2 , for example a dose
may be dynamically adjusted. Since the outcomes are measured on the same
patient, they are assumed to be correlated due to a common cause X0 , which
might represent an underlying health status, as well as genetic and lifestyle
factors.
If we treat X0 as a latent variable, the marginal model over the remaining
observed variables (X1 , X2 , X3 , X4 ) is the collection of probability distributions
that can be written in the form
p(x1 , x2 , x3 , x4 )
Z
p(x0 ) · p(x1 ) · p(x2 | x0 , x1 ) · p(x3 | x2 ) · p(x4 | x0 , x3 ) dx0 .
=
(1)
X0
That is, any (X1 , X2 , X3 , X4 )-margin of a distribution which factorizes according to the DAG over all five variables, for any state-space or distribution of
X0 1 .
⊥ X1 | X2
From (1) we can deduce that the conditional independence X3 ⊥
holds in the marginal model; i.e.
p(x3 | x1 , x2 ) = p(x3 | x2 ).
(2)
In addition this model satisfies the so-called Verma constraint of Robins (1986)
(see also Verma and Pearl, 1990), because the expression
X
p(x2 | x1 ) · p(x4 | x1 , x2 , x3 )
(3)
q(x4 | x3 ) ≡
x2
1 In
general it is sufficient to assume hidden variables are uniform on (0, 1); for this particular
graph, we can choose X0 to be finite and discrete without loss of generality provided it has a
sufficiently large number of states.
2
M
N
L
Figure 2: Diagrammatic representation of the probability simplex (dashed outline) and a marginal model (M, in blue) sitting strictly within the associated
nested model (N , in red); note the two models have the same dimension. Any
parametric latent variable model will be contained strictly within M, but it
may have a smaller dimension and be non-regular (an example is shown as L).
The ‘ordinary Markov model’ is not shown, but contains N and would generally
have larger dimension.
does not depend upon x1 (see Example 3.2).
The set of distributions satisfying both (2) and (3) is a so-called nested
Markov model (Richardson et al., 2017). If the four observed variables are binary these equations represent four independent constraints, and the nested
model is therefore an 11-dimensional subset of the 15-dimensional probability
simplex.
It is not immediately clear whether or not this nested model is the same
as the marginal model defined by (1): in principle the marginal model might
impose additional restrictions beyond (2) and (3). This begs the question, is
the set of distributions that satisfy (1) characterized by (2) and (3)?
The answer turns out to be ‘almost’, in the sense that the set of distributions
that can be written in the form (1) is a full-dimensional subset of the set that
satisfy (2) and (3), though there are additional inequality constraints. This
situation is represented by Figure 2, which shows the marginal model (M, in
blue) lying strictly within the nested model (N , in red), but the two having the
same dimension.
This paper shows that this result holds generally for all models of this kind.
That is, the constraints on the model are precisely those derived from the algorithm of Tian and Pearl (2002), as represented by the statistical model of
Richardson et al. (2017).
Existing approaches to the problem of describing Bayesian network models with hidden variables either make use of parametric structure on the latent
3
variables (for example, Silva and Ghahramani, 2009; Anandkumar et al., 2013),
or are restricted to testing conditional independences and do not consider constraints such as (3). This latter category includes the ancestral graph models of
Richardson and Spirtes (2002) and the equivalent2 models on acyclic directed
mixed graphs (ADMGs) of Richardson (2003); these pure conditional independence models, which we refer to as the ordinary Markov models, generally have
a larger dimension than any latent variable model, so using them as a proxy
leads to a loss of power to distinguish between certain kinds of model.
On the other hand parametric hidden variable models suffer from various
problems caused by the choice of state-space. They may be ‘too large’, in
the sense that the dimension of the full model is greater than the dimension
of the model over the observed data, thereby introducing identifiability problems. They may also be ‘too small’, in that unwanted additional restrictions are
implied by the parametric structure, and therefore the models have a smaller
dimension than the marginal model: this is depicted by the curve labelled L in
Figure 2.
Paradoxically, it may even be the case that a hidden variable model is ‘too
large’ and ‘too small’ at the same time! For example, if we use a latent variable
in Example 1.1 with the simplest possible state-space in which everything is
binary, then the full model over all five variables has dimension 12; however,
we have already established that the dimension of the marginal model over the
observed variables is at most 11, so the model is clearly over-parameterized. In
fact, it can be shown that the dimension of this latent variable model over the
observed variables is only 10, so an additional artificial restriction is present due
to the choice of a binary latent variable model (see Appendix A).
If X0 is given enough states, the latent variable model and the marginal
model coincide for graphs such as the one in Figure 1, a fact we will exploit
in our proofs. However, this latent variable model is less useful for statistical
inference because it is generally massively over-parameterized.
1.1
A Short Algebra Tutorial
This paper makes use of some results from real algebraic geometry, which provides powerful tools for analysing these complicated sets of distributions. All our
statistical models are collections of distributions within the probability simplex
that satisfy certain constraints. The constraints on a Bayesian network model
are conditional independences, and can be represented as the requirement that
certain polynomials in the probabilities are equal to zero; for example the conditional independence X1 ⊥
⊥ X3 | X2 is equivalent to
p(x2 ) · p(x1 , x2 , x3 ) − p(x1 , x2 ) · p(x2 , x3 ) = 0
∀x1 , x2 , x3 .
A set defined by the zeros of polynomials is said to be an algebraic variety, or
sometimes an algebraic set. In addition to equality constraints, these models will
satisfy polynomial inequalities; i.e. p(xV ) ≥ 0. A set defined by a combination of
polynomial equalities and inequalities is said to be semi-algebraic; this category
includes many common finite-dimensional statistical models. Semi-algebraic
sets have the nice property that when we eliminate one of the variables or
2 The models are equivalent if selection variables are not present, which is the case throughout this paper.
4
project onto a linear subspace, they remain semi-algebraic, generally known as
the Tarski-Seidenberg theorem (see, for example, Basu et al., 1996, Theorem
2.72). A consequence of this is that the margin of any model defined by a
semi-algebraic set is also defined by a semi-algebraic set.
The Zariski closure of a set is the smallest algebraic variety that contains
it; the fact that this is well-defined is a significant result in algebraic geometry.
For a semi-algebraic set one can informally think of its Zariski closure as the set
obtained by keeping the equality constraints and ‘throwing away’ the inequality
constraints.
Semi-algebraic sets have many interesting properties, but they are not necessarily ‘nice’ from a statistical perspective, in the sense of leading to regular
asymptotics. For this we need our set to be a manifold, i.e. to be locally Euclidean.
1.2
Contribution
In this paper we show that marginal models with finite discrete observed variables are algebraically equivalent to the appropriate nested Markov model, in
the sense that the Zariski closures of the marginal model and the nested model
are the same. A consequence of this is that a margin of a DAG model and
its nested counterpart have the same dimension, and differ only by inequality
constraints. The marginal model defined by (1) in Example 1.1 is indeed 11dimensional, and is algebraically defined by (2) and (3); however, the marginal
model also satisfies polynomial inequality constraints that the nested model
does not. The result can be interpreted as showing that the constraint finding
algorithm of Tian and Pearl (2002) is ‘complete’, in the sense that there are no
other equality constraints to find without making further assumptions.
This means that we have, for the first time, a full algebraic characterization
of margins of Bayesian network models. It also shows that the nested model represents a sensible and pragmatic approximation to the marginal model: inequality constraints are typically extremely complicated, so the nested model—which
has a factorization criterion, separation criteria, and a discrete parameterization (Richardson et al., 2017)—is much easier to work with, and can easily be
fitted with existing algorithms (Evans and Richardson, 2010). In addition, the
nested model inside the probability simplex is a manifold and therefore regular whenever the joint distribution is positive, whereas the marginal model
may have a boundary that lies strictly inside the simplex. The nested model
therefore has better statistical properties than the marginal model, in the sense
that data generated from any strictly positive distribution will lead to regular
asymptotics.
Causal discovery methods such as the FCI algorithm that use conditional
independence constraints could, in principle, be extended to the constraints
implied by nested models (Spirtes et al., 2000); our main result shows that is
‘as good as it gets’, in the sense that there are no other equality constraints
to test without making further (e.g. parametric) assumptions. Thus, this paper
probes the limits of what it is possible to learn about causal models with hidden
variables from observational data.
We work with a class of hyper-graphs called mDAGs, with which we associate marginals of DAG models (Evans, 2016). The remainder of the paper
is organized as follows: Section 2 introduces DAG models, their margins and
5
0
1
2
3
4
Figure 3: A conditional directed acyclic graph with three random vertices
(0, 2, 4) and two fixed vertices (1, 3).
mDAGs, and carefully defines the problem of interest. Section 3 describes the
nested Markov property. Section 4 describes latent variable models that can be
used to represent the marginal model without loss of generality, and Section 5
contains the main results of the paper. Finally, in Section 6 we show that a
large class of marginal models represent smooth manifolds, and provide some
discussion.
2
Directed Graphical Models
We begin with some elementary graphical definitions.
Definition 2.1. A directed graph, G(V, E), consists of a finite set of vertices, V ,
and a collection of edges, E, which are ordered pairs of distinct elements of V .
If (v, w) ∈ E we denote this by v → w, and say that v is a parent of w; the set
of parents of w is denoted by paG (w). Similarly w is a child of v, and the child
set is denoted by chG (v).
A directed graph is acyclic if there is no sequence of edges v1 → v2 → · · · →
vk → v1 for k > 1. We call such a graph a directed acyclic graph, or DAG.
Graphs are best understood visually: an example of a DAG with five vertices
and five edges is given in Figure 1. We will require the following generalization
of a DAG that allows for two separate types of vertex.
˙ 3
Definition 2.2. A conditional DAG G(V, W, E) is a DAG with vertices V ∪W
and edge set E, with the restriction that no vertex in W may have any parents.
The vertices in V are the random vertices, and W the fixed vertices; these two
sets are disjoint.
If W = ∅, this reduces to the ordinary definition of a DAG. We denote fixed
vertices with square nodes, and random ones with round nodes: see the example
in Figure 3.
2.1
Graphical Models
A graphical model arises from the identification of a graph with a collection of
multivariate probability distributions; see Lauritzen (1996) for an introduction.
Each vertex v ∈ V represents a random variable Xv taking values in a finite
state-space Xv , and a model for their joint distribution is determined by the
structure of the graph. With a conditional DAG G we associate a collection
3 Here
and throughout ∪˙ denotes a disjoint union of sets.
6
of probability measures P (· | xW ) on XV ≡ ×v∈V Xv , indexed by xW ∈ XW .
Mathematically, fixed nodes play a similar role to the ‘parameter nodes’ used
by Dawid (2002).
Following Lauritzen (1996), we say a probability kernel
P over XA given XB is
a non-negative function q : XA × XB → R such that xA q(xA | xB ) = 1 for
all xB ∈ XB . A kernel behaves much like a conditional probability distribution,
but no assumption is made about any distribution over the indexing set XB .
We apply the usual definitions for marginalizing and conditioning in kernels:
q(xA | xB ) ≡
X
q(xA , xC | xB ),
q(xA | xB , xC ) ≡
xC
q(xA , xC | xB )
.
q(xC | xB )
If q(xA | xB , xC ) does not depend upon xB then we will denote it q(xA | xC ),
and say that XA ⊥
⊥ XB | XC [q]. Here, and elsewhere, we use the shorthand
V W for V ∪ W in subscripts.
Definition 2.3. Let p(xV | xW ) be a probability kernel over XV indexed by
XW . We say that p obeys the factorization criterion with respect to a DAG G
if it factorizes into univariate kernels as
Y
p(xV | xW ) =
p(xv | xpa(v) ),
xV W ∈ XV W .
(4)
v∈V
The definition reduces to the familiar factorization criterion for DAGs if
W = ∅. The extra generality will be useful for discussing Markov properties
which involve factorization of the distribution into conditional pieces. The fixed
vertices are analogous to variables that have been conditioned upon; if p satisfies
(4) then, after renormalization, it also satisfies the factorization criterion for the
same DAG with all vertices random.
The definition of a Bayesian network can be extended to the case where no
joint density exists by insisting that each random variable Xv can be written
as a measurable function of Xpa(v) and an independent noise variable; we call
this the structural equation property. If the density exists the two criteria are
equivalent, and since we work with discrete variables this condition is always
satisfied. Although the factorization property is often simpler to work with
for practical purposes such as modelling and fitting, the structural equation
property is useful in proofs. The well-known global Markov property based on
d-separation is also equivalent to the structural equation property (Pearl, 2009).
Example 2.4. A distribution P with density p obeys the factorization criterion
for the graph in Figure 1 if the density has the form
p(x0 , x1 , x2 , x3 , x4 ) = p(x0 ) · p(x1 ) · p(x2 | x0 , x1 ) · p(x3 | x2 ) · p(x4 | x0 , x3 ).
Such distributions are precisely those which satisfy the conditional independences
X1 ⊥
⊥ X0 ,
X3 ⊥
⊥ X0 , X1 | X2 ,
X4 ⊥
⊥ X1 , X2 | X0 , X3 .
Example 2.5. A kernel p obeys the factorization criterion for the conditional
DAG in Figure 3 if it can be written as
p(x0 , x2 , x4 | x1 , x3 ) = p(x0 ) · p(x2 | x0 , x1 ) · p(x4 | x0 , x3 ).
7
1
2
3
4
Figure 4: An mDAG representing the DAG in Figure 1, with the vertex 0 treated
as unobserved.
2.2
Latent Variables and mDAGs
We now introduce the possibility that some of the random variables are unobserved or latent, leaving the marginal distribution over the remaining observed
variables. We represent the collection of margins of DAG models using a larger
class of hyper-graphs called mDAGs (‘marginal DAGs’). These avoid dealing
with latent variables directly, instead introducing additional edges to represent
them. For example, the DAG in Figure 1, with the vertex 0 treated as a latent
variable, is represented by the mDAG in Figure 4.
Define an abstract simplicial complex B over V as a collection of non-empty
subsets of V such that (i) {v} ∈ B for every v ∈ V , and (ii) if A ∈ B and B ⊆ A
with B 6= ∅, then B ∈ B.
Definition 2.6. An mDAG, G(V, W, E, B), is hyper-graph consisting of a conditional DAG with random vertices V , fixed vertices W and directed edge set
E, together with an abstract simplicial complex B over V , called the bidirected
faces.
We say that G ′ (V ′ , W ′ , E ′ , B ′ ) is a subgraph of G if V ′ ⊆ V , E ′ ⊆ E, B ′ ⊆ B,
and W ′ ⊆ V ∪ W : that is, each component is contained within the previous
one, but random vertices may become fixed.
The mDAG was introduced by Evans (2016), without the additional generality of fixed vertices. This aspect changes very little to the theory of these
graphs, but is necessary for understanding the nested Markov model; note that
bidirected faces only involve the random vertices. As with conditional DAGs,
when representing mDAGs graphically the fixed vertices are drawn as square
nodes and random vertices as circles.
The bidirected simplicial complex is represented by its maximal non-trivial
elements (i.e. those of size at least 2), called the bidirected hyperedges, or just
edges. These are drawn in red, as in Figure 5(a); in this case W = {6} and the
maximal sets of B are {1, 2}, {2, 3, 4}, and {3, 4, 5}.
With each mDAG, G, we can associate a conditional DAG Ḡ by replacing
each maximal element B ∈ B (of size at least 2) with a new random vertex u,
such that the children of u are precisely the vertices in B. The new vertex u
becomes the ‘unobserved’ variable represented by the bidirected edge B. We
call Ḡ the canonical DAG associated with G. The mDAG in Figure 5(a) is thus
associated with the canonical DAG in Figure 5(b).
Our interest in mDAGs lies in their representation of the margin of the
associated canonical DAG, and so we define our model in this spirit; see Evans
(2016).
˙ , and let Ḡ be the
Definition 2.7. Let G be an mDAG with vertices V ∪W
˙ ∪W
˙ . A kernel p over XV indexed by XW is
canonical DAG with vertices V ∪U
8
1
2
3
6
4
5
(a)
1
u1
2
3
u2
u3
6
4
5
(b)
Figure 5: (a) An mDAG, G, and (b) a DAG with hidden variables, Ḡ, representing the same model (the canonical DAG).
9
said to be in the marginal model for G if there exists a kernel q that factorizes
according to Ḡ, and
Z
q(xV , xU | xW ) dxU .
p(xV | xW ) =
XU
That is, the margin of q over XV is p. Denote the collection of such kernels by
M(G).
In other words, the marginal model is the collection of kernels that could be
constructed as the margin of a Bayesian network with latent variables replacing
the bidirected edges. If G is a DAG then the marginal model is just the usual
model defined by the factorization.
A latent variable model corresponding to a canonical DAG Ḡ (i.e. possibly
with parametric or distributional assumptions on the latent variables) always lies
within the marginal model corresponding to the mDAG G. This should not be
taken as meaning that the marginal model supersedes all latent variable models,
since sometimes the additional parametric assumptions made in a latent variable
model are crucial to their utility. For example, representing hidden Markov
models and phylogenetic tree models using an mDAG leads to a bidirected
hyper-edge containing all vertices; the marginal models are therefore saturated,
and rather uninteresting from the perspective statistical inference.
From the definitions above it may seem as though the set of marginal DAG
models that can be represented by mDAGs is restricted to cases where the latent
variables have no parents; in fact this does not cause any loss of generality, since
all marginal DAG models can be represented in this way (see Evans, 2016).
2.3
Districts and Sterile Vertices
Definition 2.8. A collection of random vertices C ⊆ V in an mDAG G is
bidirected-connected if for any distinct v, w ∈ C, there is a sequence of vertices v = v0 , v1 , . . . , vk = w all in C such that, for each i = 1, . . . , k, the pair
{vi−1 , vi } ∈ B.
A district of an mDAG is an inclusion maximal bidirected-connected set of
random vertices.
More informally, a district is a maximal set of random vertices joined by the
red edges in an mDAG. It is easy to see from the definition that districts form
a partition of the random vertices in an mDAG. The mDAG in Figure 4, for
example, contains three districts, {1}, {3} and {2, 4}. Districts inspire a useful
reduction of mDAGs, via the following special subgraph.
Definition 2.9. Let G be an mDAG containing random vertices C ⊆ V . Then
G[C] is the subgraph of G with
(i) random vertices C and fixed vertices paG (C) \ C;
(ii) those directed edges w → v such that v ∈ C (and w ∈ paG (C));
(iii) the bidirected simplicial complex BC ≡ {B ∩ C : B ∈ B(G)}.
G[C] is therefore the subgraph induced over C, together with parents of
C and edges directed towards C. Any edges (whether directed or bidirected)
between the newly fixed vertices are removed.
10
1
2
3
(b)
(a)
1
2
3
4
(c)
Figure 6: Subgraphs corresponding to factorization of the graph in Figure 4 into
districts. Parent nodes of the district are drawn as squares.
For the graph in Figure 4 the subgraphs G[{1}], G[{3}] and G[{2, 4}] are
shown in Figures 6(a), (b) and (c) respectively. Note in particular that the edge
2 → 3 is not in the subgraph G[{2, 4}].
Definition 2.10. Let G be an mDAG with random vertices V . For an arbitrary
set C ⊆ V , define sterileG (C) ≡ C \ paG (C). In words sterileG (C) is the subset
of C whose elements have no children in C. We say a set C is sterile if C =
sterileG (C).
Proposition 2.11. Let G be an mDAG with districts D1 , . . . , Dk . A probability
kernel p is in the marginal model for G if and only if
p(xV | xW ) =
k
Y
gi (xDi | xpa(Di )\Di ),
i=1
where each gi is a probability kernel in the marginal model for G[Di ].
In addition, p is in the marginal model for G only if for every v such that
chG (v) = ∅, the marginal distribution
X
p(xV | xW )
p(xV \v | xW ) =
xv
is in the marginal model for G[V \ {v}].
Proof. Consider the factorization of the canonical DAG Ḡ. The first result
follows from grouping the factors according to districts and noting that there is
no overlap in the variables being integrated out. The second result follows from
noting that if v has no children, the variable xv only appears in a single factor,
and that factor is a conditional distribution that integrates to 1.
It follows from this result that to characterize the marginal model we need
only consider mDAGs containing a single district, since other models can always
be reduced to combinations of such graphs.
11
1
2
1
3
2
(a)
3
(b)
Figure 7: (a) An mDAG on three vertices representing a saturated model; (b)
the bidirected 3-cycle, the simplest non-geared mDAG.
2.4
Relationship between mDAGs and ADMGs
Previous papers considering marginal and nested models for DAGs have used
acyclic directed mixed graphs, which are the restriction of mDAGs with random
vertices so that each bidirected edge has size two (Richardson, 2003; Shpitser et al.,
2012; Evans and Richardson, 2014; Richardson et al., 2017).
From the perspective of the nested Markov property this distinction is unimportant: if we replace any bidirected simplicial complex with all its subsets of
size 2, we obtain a conditional ADMG that represents the same model under
the nested Markov property. However, if we consider the marginal model the
models are not equal, as the restriction to pairwise independent latent parents
will sometimes introduce additional inequality constraints. The marginal model
for the mDAG in Figure 7(b) is strictly smaller than the one for 7(a) (Fritz,
2012, Proposition 2.13), for example. See Evans (2016) for a more detailed
discussion.
It follows from the results of this paper that there is no difference in equality
constraints between graphs that differ only in this manner; algebraically the
model defined by having a single latent parent for several variables is the same
as having separate parents for each pair of vertices. The mDAGs in Figure 7
both represent marginal models of full dimension, for example. Hence, in terms
of model dimension, nothing is lost by using ADMGs instead of mDAGs.
3
Nested Markov Property
The nested Markov property imposes constraints on a joint distribution that
mimic those satisfied by the marginal model, including conditional independences and the Verma constraint in Example 1.1 (Richardson et al., 2017). It is
defined in the following recursive way, which is a modification of the algorithm
of Tian and Pearl (2002).
Definition 3.1 (Nested Markov Property). A kernel p over XV indexed by XW
obeys the nested Markov property for an mDAG G(V, W ) if V = ∅, or both:
12
1. p factorizes over the districts D1 , . . . , Dl of G:
p(xV | xW ) =
l
Y
gi (xDi | xpa(Di )\Di )
i=1
where each gi is a kernel which (if l ≥ 2 or W \ paG (V ) 6= ∅) obeys the nested
Markov property with respect to G[Di ]; and
2. for each v ∈ V such that chG (v) = ∅, the marginal kernel
X
p(xV | xW )
p(xV \v | xW ) =
xv
obeys the nested Markov property with respect to G[V \ {v}].
The set of kernels that obey the nested Markov property for G is the nested
Markov model, denoted by N (G).
The condition that l ≥ 2 or W \ paG (V ) 6= ∅ in the first criterion of this
definition is simply to prevent an infinite recursion of the definition: all the
graphs invoked recursively have either fewer random vertices or fewer vertices
overall than their predecessor in the recursion. When we reach a graph with a
single random vertex v such that all fixed vertices are parents of v, then any
kernel p(xv | xpa(v) ) satisfies the nested Markov property.
Example 3.2. Consider again the mDAG in Figure 4. Applying criterion 1 to
this graph implies that
p(x1 , x2 , x3 , x4 ) = g1 (x1 ) · g24 (x2 , x4 | x1 , x3 ) · g3 (x3 | x2 )
for some g1 , g3 and g24 obeying the nested Markov property with respect to the
mDAGs in Figures 6(a), (b) and (c) respectively. Applying the second criterion
to g24 and the now childless vertex 2 (see Figure 6(c)) gives
X
g24 (x2 , x4 | x1 , x3 ) = h(x4 | x3 ),
x2
for some function h independent of x1 (by a further application of the first
criterion); this is precisely the Verma constraint.
The marginal model implies additional conditions on joint distributions because, although it satisfies the properties used to define the nested model, these
properties are not sufficient to describe it. In particular, for p to be in the
marginal model, the kernel g24 must satisfy Bell’s inequalities (see, for example,
ver Steeg and Galstyan, 2011, Section 4.1).
The nested Markov property is ‘sound’ with respect to marginal models, in
the sense that all constraints represented by the former also hold in the latter.
Theorem 3.3. For any mDAG G we have M(G) ⊆ N (G).
Proof. This follows from the fact that the nested Markov model is defined in
terms of constraints which are proven in Proposition 2.11 to be satisfied by the
marginal model.
13
3.1
Parameterizing Sets
Definition 3.4. Let G be an mDAG. A subset of random vertices S ⊆ V is
called intrinsic if S is a district in any graph which can be obtained by iteratively
applying graphical operations of the form 1 and 2 in Definition 3.1 (i.e. taking
the graph G[D] for a district D, or G[V \ {v}] for a sterile vertex v).
Given an intrinsic set, S, define H = sterileG (S) to be the recursive head, and
T = paG (S) the tail, associated with S (note that H and T are disjoint). The collection of all recursive heads in G is denoted by H(G). There is a one-to-one correspondence between intrinsic sets and recursive heads (Evans and Richardson,
2015). Throughout we will use H and T to indicate recursive heads and tails
respectively, with the context making it clear which intrinsic set is being referred
to.
Define
A(G) ≡ {H ∪ A | H ∈ H(G), A ⊆ T }
to be the parameterizing sets of G. This collection of sets is so-called because it
(locally) describes the set of distributions (or kernels) contained in the nested
and marginal models, as we prove in Section 5.
Conversely, non-empty sets not in A(G) are called the constrained sets, and
locally describe the set of constraints imposed by the nested and marginal models.
Example 3.5. The mDAG in Figure 4 has districts {1}, {3} and {2, 4}, so
these are all intrinsic sets. Further, in the subgraph G[{2, 4}] the vertices 2 and
4 have no children, so we can marginalize either to see that respectively {4} and
{2} are intrinsic sets. The corresponding recursive heads and tails are then:
S
{1}
H
{1}
T
∅
{2}
{3}
{2}
{3}
{1}
{2}
{4}
{2,4}
{4}
{2,4}
{3}
{1,3}
A
{1}
{2}, {1,2}
{3}, {2,3}
.
{4}, {3,4}
{2,4}, {1,2,4}, {2,3,4}, {1,2,3,4}
Note that every non-empty subset of V is represented in A except for {1, 3},
{1, 2, 3}, {1, 4} and {1, 3, 4}. These are the ‘constrained sets’; the first two
correspond to the conditional independence, X1 ⊥
⊥ X3 | X2 in (2), and the
others to the Verma constraint (3).
Intrinsic sets and recursive heads consist only of random vertices, while tails
may include both random and fixed vertices.
Proposition 3.6. Let C be a bidirected-connected set in an mDAG G; then
there exists an intrinsic set S such that C ⊆ S and sterileG (S) ⊆ sterileG (C).
Proof. The district containing C is intrinsic by definition, so there exists an
intrinsic set containing C; let S be a minimal intrinsic set (by inclusion) containing C. By the definition of intrinsic sets S is a district in some graph reached
by iteratively applying the operations 1 and 2 to G: applying operation 1 again
gives the graph G[S].
14
Suppose for contradiction that there exists v ∈ sterileG (S)\ sterileG (C); then
v∈
/ C, since otherwise some child of v would be in C, and therefore in S. In
addition, v is childless in the subgraph G[S], so we can remove v under operation
2 of Definition 3.1. In the resulting strictly smaller graph, C is still contained
within one district, say S ′ , since C is bidirected-connected; in addition S ′ is also
intrinsic, so we have found a strictly smaller intrinsic set S ′ ⊇ C, and reached
a contradiction.
We use the △ operator to denote the symmetric difference of two sets:
A△B ≡ (A \ B) ∪ (B \ A). Given a collection of sets Ai , i = 1, . . . , k indexed
by a finite set I, let
k
i
Ai ≡ A1 △A2 △ · · · △Ak .
i=1
denote the symmetric difference of all the Ai . That is, it is the set containing
precisely those elements a which appear in an odd number of the sets Ai .
The following result gives a characterization of the parameterizing sets in
terms of symmetric differences which will be fundamental to our proof of the
main results in this paper.
Lemma 3.7. A set A ∈ A(G) if and only if there exists a bidirected-connected
set C = {v1 , . . . , vk } in G, and sets Ai , i = 1, . . . , k, satisfying
{vi } ⊆ Ai ⊆ {vi } ∪ paG (vi ),
such that
A=
k
i
Ai = A1 △ · · · △Ak .
(5)
i=1
Proof. Suppose that A ∈ A(G); then H ⊆ A ⊆ H ∪ T for some head-tail pair
(H, T ), with associated intrinsic set S. Then let C = S, since intrinsic sets are
by definition bidirected-connected. Now consider sets A′ of the form (5); start
with Ai = {vi }, so that A′ = S, and we will adjust the sets Ai to obtain A′ = A.
We always have that A′ contains H, because each vi ∈ H appears in Ai and,
by sterility of heads, in no other set Aj . Each vertex t ∈ T is (by definition)
the parent of some vertex vj(t) ∈ S, so we can either include or exclude it from
A′ (as required) just by replacing Aj(t) = {vj(t) } by {vj(t) , t}. Hence we just do
this to include vertices in A \ S ⊆ T and exclude vertices in S \ A ⊆ T .
Conversely, suppose that A is of the form (5) for some bidirected-connected
set C; let S be an intrinsic set satisfying the conditions of Proposition 3.6,
and (H, T ) be its associated head-tail pair. Then the head H = sterileG (S) ⊆
sterileG (C). Each vi ∈ H ⊆ C appears in A, since vi ∈ Aj if and only if i = j.
Also A ⊆ C ∪ paG (C) ⊆ S ∪ paG (S) = H ∪ T , so A ∈ A(G).
3.2
Parameterization of the nested model
The nested Markov model can be parameterized with parameters indexed by
head-tail sets (Evans and Richardson, 2015), and the parameterization defines
a smooth bijection between an open subset of a real vector space (i.e. the parameter space) and the model (the set of probability distributions). This has
15
some nice consequences that we now state (for proofs see Evans and Richardson,
2015).
In particular, for a fixed state-space XV W the set N (G) is a smooth manifold
within the strictly positive probability simplex, and has dimension
Y
X
|XT |
(|Xh | − 1).
d(G, XV W ) ≡
h∈H
H∈H(G)
In the all-binary case this reduces to
d(G, XV W ) ≡
X
2|T | .
H∈H(G)
Our main result will show that M(G) always has the same dimension as N (G).
Indeed, the parameterization of N (G) will in principle also serve as a parameterization of M(G), except that one would also have to restrict the parameter
space in order to enforce the inequality constraints; of course, this is currently
impractical since the inequality constraints are not generally known.
4
Geared mDAGs
In this section we introduce a special class of mDAGs which we term ‘geared’.
For marginal models relating to such graphs, the state-space of the hidden
vertices can be restricted without loss of generality, making proofs considerably
easier. In Section 5 we prove our main result first for geared graphs, and then
extend the result to the general case.
Definition 4.1. Let G be an mDAG with bidirected simplicial complex B. We
say that G is geared if the maximal elements of B satisfy the running intersection
property. That is, there is an ordering of the edges B1 , . . . , Bk such that for
each j > 1, there exists s(j) < j with
[
Bi = Bj ∩ Bs(j) .
Bj ∩
i<j
In other words, the vertices that are contained in both Bj and any previous
edge are all contained within one such edge Bs(j) .
A particular ordering of the elements of B which satisfies running intersection
is called a gearing of G.4
Example 4.2. The simplest non-geared mDAG is the bidirected 3-cycle, depicted in Figure 7(b); there is no way to order the bidirected edge sets {1, 2},
{2, 3}, {1, 3} in a way which satisfies the running intersection property, since
whichever edge is placed last in the ordering shares a different vertex with each
of the two other edges.
4 The term ‘geared’ is chosen because a collection of bidirected edges which satisfies running intersection may appear rather like ‘cogs’ in a set of gears: see Figure 5. The definition is equivalent to the requirement that the simplicial complex B is vertex decomposable
(Provan and Billera, 1980), and is also closely related to the notion of decomposability in an
undirected or directed graph. Indeed the term ‘decomposable’ is used by Fox et al. (2014) to
describe the same idea. We avoid using this terminology because of its existing meaning in
connection with undirected and directed graphical models: for example, ordinary DAGs are
trivially geared, but they may or may not be decomposable in the original sense (Lauritzen,
1996).
16
The following fact about geared subgraphs of mDAGs will allow us to generalize our later results to graphs which are not geared.
Lemma 4.3. Let G be an mDAG with parameterizing sets A(G). For any
A ∈ A(G) there exists a geared mDAG G ′ ⊆ G, such that A ∈ A(G ′ ).
Proof. By Lemma 3.7, A is of the form (5) for some bidirected-connected set C.
Let G ′ have the same vertices (random and fixed) and directed edges as G, but
be such that the set C is singly connected by bidirected edges (i.e. the edges are
all of size 2 and removing any of them will cause C to be disconnected) chosen
to be a subgraph of G. Then G ′ is geared by standard properties of trees and
running intersection, and using Lemma 3.7 again we have A ∈ A(G ′ ).
4.1
Functional Models
The key property of geared graphical models is that we can find a finite discrete latent variable model that is the same (over the observed variables) as
the marginal model; that is, if the latent variables have a sufficiently large
state-space then they do not impose additional restrictions on the observed distribution. This is achieved by letting each observed variable be a deterministic
function of its latent and observed parents. We illustrate this with an example.
Example 4.4. Consider the mDAG in Figure 8(a) representing the instrumental variables model, used to model non-compliance in clinical trials; here,
for example, X1 represents a randomized treatment, X2 the treatment actually
taken, and X3 a patient’s outcome or response, such as survival. Suppose that
each of these quantities is binary, taking values in {0, 1}. Conceptually, it can be
useful to posit the existence of two different potential outcomes X3 (0), X3 (1) for
the survival response, one for each level of the treatment; X3 (0) is the patient’s
outcome given that they choose not to take the treatment (so that X2 = 0) and
X3 (1) is their outcome given that they do (X2 = 1). For example, if X3 (0) = 0
and X3 (1) = 1 then the patient survives if they take the treatment but dies
if they do not. This pair of values is known as a patient’s response type. Of
course, we can only ever observe one of these outcomes in a given patient, the
one corresponding to the observed value of X2 .
Similarly, we can conceive of two versions of the treatment X2 (0), X2 (1)
depending upon the assigned value of X1 , this pair being called the patient’s
compliance type. For example, X2 (0) = X2 (1) = 0 means that the patient will
not take the treatment, regardless of whether or not they are assigned to the
treatment group. These concepts have proved fruitful in causal inference, as they
enable discussion of whether treatments have effects at the level of individual
patients, rather than just over the entire population on average (Neyman, 1923;
Rubin, 1974; Richardson et al., 2011).
Now, since the latent variable (say U ) with children {2, 3} can take any
value, we can—without loss of generality—assume that it includes the pair
(X3 (0), X3 (1)), or equivalently a function f3 : X2 → X3 that determines, given
the observed X2 , which value X3 will take. In this case X3 is still a measurable
function of its parents U and X2 . Similarly we can assume U includes a function
f2 : X1 → X2 that determines X2 given an observed X1 .
An observation for a particular patient can be obtained by drawing a random
treatment assignment X1 , a random compliance type for the patient f2 , and a
17
f2 , f3
1
2
3
1
(a)
2
3
(b)
Figure 8: (a) An mDAG representing the instrumental variables model; (b)
a DAG with functional latent variables equivalent to the potential outcomes
model of instrumental variables.
random response type f3 , and then evaluating (X1 , X2 , X3 ) = (X1 , f2 (X1 ), f3 (f2 (X1 ))).
The key point is that one can place a distribution over (X1 , f2 , f3 ) and obtain
a distribution over the observed variables (X1 , X2 , X3 ). The only requirement
for the distribution to be Markov with respect to this particular graph is that
X1 ⊥
⊥ {f2 , f3 }, as depicted in Figure 8(b).
The functional construction outlined above is mathematically equivalent to
potential outcomes, and provides a model that is somewhat simpler to study
than the general latent variable model. In fact, any geared mDAG can be
reduced to a latent variable model in the way described above, something we
now proceed to show.
4.2
Remainder Sets
Given a single-district, geared mDAG with at least one bidirected edge and a
gearing B1 , . . . , Bk , define
[
Bi
Rj ≡ Bj \
i<j
(taking R1 ≡ B1 ) to be the remainder set associated with Bj . Remainder sets
partition V , so for a random vertex v ∈ V , define r(v) to be the unique j such
that v ∈ Rj .
Now say that an ordering < on the vertices in V respects the gearing if for
v ∈ Ri and w ∈ Rj , we have v < w whenever i > j; in other words, all the
vertices in Rk precede all those in Rk−1 , etc; such an ordering always exists.
For each v ∈ V with r(v) = j, let
[
Ri ;
π(v) =
i>j
v∈Bi
that is, the remainders associated with all bidirected edges which contain v and
are later than j in the ordering. Then define a collection of functions
Fv ≡ {f : Xpa(v) × Fπ(v) → Xv },
where FA = ×a∈A Fa and F∅ = X∅ = {1}. This is valid recursive definition,
since all the vertices in π(v) precede v in an ordering which respects the gearing.
18
Example 4.5. The mDAG in Figure 8(a) has only one bidirected edge and
therefore is trivially geared with R1 = B1 = {2, 3}. This leads to the sets
F2 = {f2 : X1 → X2 },
F3 = {f3 : X2 → X3 },
which are precisely the sets of functions for compliance type and response type
respectively.
Example 4.6. Consider the mDAG in Figure 5, and order the bidirected edges
as B1 = {1, 2}, B2 = {2, 3, 4} and B3 = {3, 4, 5}, giving respective remainder
sets R1 = {1, 2}, R2 = {3, 4} and R3 = {5}. The ordering 5 < 4 < 3 < 2 < 1 of
the random vertices respects the gearing, and we have
π(1) = π(5) = ∅,
π(3) = π(4) = {5},
π(2) = {3, 4}.
In this case then
F5 = {f : X3 → X5 }
F4 = {f : X2,3,6 × F5 → X4 }
F3 = {f : X1 × F5 → X3 }
F1 = {f : {1} → X1 }
F2 = {f : F3,4 → X2 }
Alternatively, if we order the bidirected edges as {2, 3, 4}, {1, 2}, {3, 4, 5}, then
we could take 5 < 1 < 2 < 3 < 4, and
π(1) = π(5) = ∅,
π(3) = π(4) = {5},
π(2) = {1};
this yields F2 = {f : F1 → X2 }, with other collections Fv remaining unchanged.
4.3
Functional Models for Geared Graphs
If a vertex v is contained within exactly one bidirected edge, B, then without loss
of generality we can assume that the latent variable corresponding to B contains
all the residual information about how Xv should behave given the values of its
visible parents, Xpa(v) . In other words, the latent variable associated with B
includes a (random) function fv : Xpa(v) → Xv which, once instantiated, ‘tells’
Xv = fv (Xpa(v) ) which value it should take for each value of its other parents,
exactly as in Example 4.5.5 All the randomness of Xv is collapsed into fv and
Xpa(v) .
If v is contained within two or more bidirected edges, say Bi and Bj , we
might say that Bi tells Xv what value to take for every value of its visible
parents and the other latent variables. However, it is not clear how to define
such a function until the state-space associated with the other latent parents
(i.e. Bj ) has already been fixed. The decomposable structure of geared graphs
makes it possible to iteratively fix state-spaces for latent variables without loss
of generality.
To see this, suppose we have a single-district, geared mDAG G with remainder sets R1 , . . . , Rk , and form the canonical DAG Ḡ by replacing each bidirected
edge Bi in G with a new vertex ui , such that chḠ (ui ) = Bi . Compare, for example, the structure of the graphs in Figures 5(a) and (b).
5 Equivalently,
one could take a deterministic function fv and introduce an ‘error term’ Ev
so that Xv = fv (Xpa(v) , Ev ), as in the non-parametric structural equation models of Pearl
(2009).
19
1
f1 , f2
2
3
f3 , f4
f5
6
4
5
Figure 9: A DAG with functional latent variables, associated with a gearing of
the mDAG in Figure 5(a).
2
3
f3 , f4
f5
6
4
Figure 10: Subgraph of the DAG in Figure 9 containing the vertex 4 and its
parents.
Note that each vertex v ∈ Rk has a single latent parent uk in Ḡ. Then,
without loss of generality, incorporate the function fv : XpaG (v) → Xv into
the latent variable Uk . We ‘replace’ Uk with the collection of such functions
f Rk ∈ F Rk .
Each vertex v ∈ Rk−1 has latent parent uk−1 and possibly also uk ; but since
the state-space of Uk has been fixed as FRk , we can define fv : Xpa(v) ×FRk → Xv
for those v with latent parents uk−1 and uk , and just fv : Xpa(v) → Xv otherwise.
These functions fRk−1 can be integrated into Uk−1 , and the process repeated
for i = k − 2, . . . , 1.
We end up with latent variables Ui taking values in FRi for i = 1, . . . , k.
For example, with the first gearing given in Example 4.6 for the graph in Figure
5(a), we would have
U1 = (f1 , f2 ),
U2 = (f3 , f4 ),
U3 = (f5 ).
Associating each variable Ui with the vertex ui leads to the DAG in Figure
9. Notice that, for each v ∈ V , the function fv is contained within a parent
variable of v. In addition, all the arguments of the function fv are also parents
of v. For example, take v = 4, whose parents are drawn separately in Figure 10.
The function f4 ∈ F4 is generated as part of the latent variable U2 = (f3 , f4 ),
20
and the associated vertex u2 is indeed a parent of 4. In addition, F4 = {f :
X2,3,6 × F5 → X4 }, so the arguments of the function f4 , namely X2 , X3 , X6
and f5 , all correspond to vertices which are also parents of 4. Thus, in setting
X4 = f4 (X2 , X3 , X6 , f5 ) we ensure that X4 is a well defined function of its
parent variables.
In fact using this construction we can set
Xv = fv (fπ(v) , Xpa(v) )
for every v ∈ V , which is well defined because the directed part of the original
mDAG is acyclic. The following result shows that the resulting conditional
distribution over XV given XW is in the marginal model for the original mDAG.
Theorem 4.7. Let G be a geared mDAG, and Ri , i = 1, . . . , k be the remainder
sets corresponding to some gearing of G. Suppose we generate functions fv ∈ Fv
according to a distribution in which
(fv | v ∈ Ri ) ⊥
⊥ (fw | w ∈ V \ Ri ),
for each i = 1, . . . , k, and then define
Xv = fv (fπ(v) , Xpa(v) ),
v ∈ V.
Then the induced conditional distribution on XV given XW is in the marginal
model for G.
Proof. For each bidirected edge Bi , define the random variable Ui = (fv | v ∈
Ri ). The Ui s are represented by exogenous variables on the DAG Ḡ, and the
conditions given in the statement of the theorem ensures they are all independent. The structural equation property for Ḡ will therefore be satisfied if each
Xv is a well defined function of its parents in the graph.
In other words, the three components fv , fπ(v) and Xpa(v) must all be determined from random variables which are parents of v in Ḡ. This holds for Xpa(v)
by definition. Additionally v ∈ Ri implies that v ∈ Bi , and that therefore the
variable Ui ≡ (fv : v ∈ Ri ) is a parent variable of Xv .
Lastly suppose w ∈ π(v); this happens if and only if w, v ∈ Bj for some
j > i, in which case w ∈ Rj for the minimal such j by the running intersection
property of the gearing. Then fw is contained in Uj , which is also a parent
variable of Xv .
In fact it is not hard to see that any distribution in the marginal model of
a geared graph can be generated in the way described in Theorem 4.7. Since
each of these latent variables takes values in a finite collection of functions,
this means that the marginal model of a geared graph is equivalent to a latent
variable model in which all the random variables (latent and observed) are
finite and discrete. It follows from this that marginal models for geared mDAGs
are semi-algebraic sets by the Tarski-Seidenberg theorem (Basu et al., 1996,
Chapter 2).
Example 4.8. Consider the marginal model for the graph in Figure 4. In this
case the vertices 2 and 4 are each contained in only one bidirected edge, so
21
without loss of generality this edge could be replaced in the canonical DAG
(Figure 1) with a latent variable taking values in F2 × F4 where
F2 ≡ {f : X1 → X2 },
F4 ≡ {f : X3 → X4 }.
That is, the latent variable may be assumed to be U = (f2 , f4 ), where f2 and f4
respectively assign values to X2 and X4 given particular values of X1 and X3 .
For non-geared graphs such as that in Figure 12(a), there is no clear way
to write the marginal model as a latent variable model without possible loss
of generality. It is therefore not possible for us to prove that marginal models
corresponding to non-geared mDAGs are semi-algebraic; however, we conjecture
that they indeed are. Our results will show that for a sufficiently large latent
state-space the dimension of the latent variable model becomes the same as that
of the nested model, but it is at least conceivable that there are non-polynomial
inequality constraints on the marginal model for non-geared graphs.
4.4
Generating Distributions for Geared mDAGs
Let G be a single-district, geared mDAG, with gearing given by remainder sets
R1 , . . . , Rk ; assign a probability distribution ρi to each collection of functions
Ui ≡ (fv | v ∈ Ri ). Suppose we draw values for variables Ui = (fv )v∈Ri independently according to ρi , and use them to generate values for the observed variables
XV for each possible value of the fixed vertices XW . The resulting (conditional)
distribution over XV given XW is, by Theorem 4.7, in the marginal model for
G.
S
Let π(Ri ) ≡ v∈Ri π(v) and fA ≡ (fv | v ∈ A). Define
p[ρk , . . . , ρ1 ](xV | xW ) =
X
ρk (fRk ) · · ·
X
ρ1 (fR1 ),
(6)
fR1 ∈Φ1 (fπ(R1 ) ,xV W )
fRk ∈Φk (xV W )
where
Φi (fπ(Ri ) , xV W ) = {fRi | fv (xpa(v) , fπ(v) ) = xv for each v ∈ Ri };
(7)
that is, Φi (fπ(Ri ) , xV W ) is precisely the set of functions fRi that, given the
indicated values of parents variables, jointly evaluate to xRi . Hence (6) is a
sum over all the combinations of functions fV that, given the input XW = xW ,
recursively evaluate to xV .
The function p[·] takes distributions over the functions fV and returns a
kernel over XV indexed by XW . For brevity we will generally denote this by
X
X
p[ρk , . . . , ρ1 ] =
ρk · · ·
ρ1 ,
Φk
Φ1
with the dependence upon xV W left implicit. It may be helpful to think of this
as an over-parameterized family of kernels for XV given XW , with parameters
ρ1 , . . . , ρk .
The mapping p[·] is clearly smooth (infinitely differentiable), and its image
defines the marginal model. Hence we will be able to deduce various aspects of
the model’s geometry by studying p[·] and its derivatives. Choosing ρi (fRi ) = 1
for each i (up to a constant of proportionality which, for simplicity, we do not
22
write explicitly) induces the uniform distribution on XV for each xW ∈ XW ;
we denote this kernel by p0 ≡ p[1, . . . , 1]. Clearly p0 is contained within M(G)
for any mDAG G—as, in fact, is any distribution corresponding to all variables
being independent.
Example 4.9. For the instrumental variables model in Figure 8 (if we consider
X1 to be fixed), we have
X
ρ(f2 , f3 )
p[ρ](x2 , x3 | x1 ) =
Φ(x123 )
where
Φ(x123 ) = {(f2 , f3 ) : f2 (x1 ) = x2 , f3 (x2 ) = x3 }.
Example 4.10. In the case of the mDAG in Figure 5(a) we have three bidirected edges and remainder sets, and the gearing used in Figure 9 gives
X
X
X
p[ρ3 , ρ2 , ρ1 ] =
ρ3 (f5 )
ρ2 (f3 , f4 )
ρ1 (f1 , f2 ),
Φ3
Φ2
Φ1
where
Φ1 = {(f1 , f2 ) | f1 = x1 , f2 (f3 , f4 ) = x2 }
Φ2 = {(f3 , f4 ) | f3 (x1 ) = x3 , f4 (x2 , x3 , x6 , f5 ) = x4 }
Φ3 = {f5 | f5 (x3 ) = x5 }.
5
Main Results
In this section we provide our main results, showing that the marginal model
M(G) has the same dimension as the nested model. This is done first for geared
mDAGs, and the result is then extended to general graphs. For geared graphs,
the marginal model is just the image of the infinitely differentiable function p[·]
described in the previous section. Such functions can be locally approximated at
a particular point, say p0 = p[1, . . . , 1], by the linear map given by the derivative
of p[·].
This column space of this linear map (also called the pushforward map) gives
the linear space that approximates the model at p0 , also known as the tangent
space. We will show that the tangent space to the marginal model at p0 is equal
to the tangent space of the nested model N (G) at p0 . To do this we take a basis
of the tangent space of N (G), and for every vector λ in the basis we explicitly
construct a vector δ such that the directional derivative of p[·] with respect to δ
is equal to λ. This shows that each λ is also contained in the tangent space of
M(G). Since the marginal model is contained within the nested model, it will
then follow from results in algebraic geometry that the two models coincide in
a neighbourhood of p0 .
For non-geared graphs we have do slightly more work, showing that we can
combine maps from different geared sub-graphs to obtain the same result.
23
5.1
Vector Spaces and Tangent Cones
A probability kernel p(xV | xW ) can be thought of equally as a vector with
entries indexed by XV W , or a real function with domain XV W . The following
decomposition of the vector space R|XV | will prove useful.
Definition 5.1. For any A ⊆ V , let ΛA be the subspace of R|XV | consisting of
vectors p such that
P
(i)
ya ∈Xa p(ya , xV \a ) = 0 for each a ∈ A and xV \{a} ∈ XV \{a} ;
(ii) p(xV ) = p(yV ) whenever xA = yA .
In other words, considered as a function p : XV → R, the value of p ∈ ΛA
only depends upon xA , and its sum over xa for a ∈ A (keeping the other
arguments fixed) is 0. In particular
Λ∅ is the subspace spanned by the vector of
Q
1s. The dimension of ΛA is a∈A (|Xa | − 1); in the case where all the variables
are binary, each ΛA has dimension one and is the same as the space spanned by
the corresponding column of a log-linear design matrix.
It is simple to check that the spaces ΛA are all orthogonal, and the real
vector space R|XV | can be decomposed as the direct sum
M
R|XV | =
ΛA .
A⊆V
Definition 5.2. Let A be a subset of Rk containing a point x. The tangent
cone of A at x is the set of vectors of the form
v = lim ηn−1 (vn − x)
n→∞
where ηn → 0 and each vn ∈ A.
A tangent cone is a cone, but may or may not be a vector space, depending
upon whether the set A is regular at x. If A is defined by the image of a differentiable bijective map then the tangent cone is a vector space, and the same as the
image of the pushforward map. This is the case with the nested model N (G),
which has an explicit and smooth parameterization (Evans and Richardson,
2015). Its tangent cone at the uniform distribution p0 is the vector space
M
ΛA ,
(8)
TSn0 ≡
A∈A(G)
where A(G) are the parameterizing sets; this can be deduced by looking directly
at the parameterization.
As noted in Section 4, any marginal model M(G) also contains the uniform
distribution
p0 (xV | xW ) ≡ |XV |−1 ,
xV ∈ XV , xW ∈ XW ,
at which point all variables are totally independent. The tangent cone of the
marginal model M(G) at p0 is also the vector space (8), which forms the main
result of this section.
24
Theorem 5.3. The tangent cone of M(G) at p0 , denoted TC0 , is the vector
space
M
TC0 = TSn0 ≡
ΛA .
A∈A
That TC0 ⊆ TSn0 follows from the fact that M(G) ⊆ N (G). The proof of
the reverse inclusion is delayed until the end of the section.
5.2
Results for Geared Graphs
Definition 5.4. Let λ : XA → R; we say that λ is A-degenerate (or just
degenerate) if for each a ∈ A, and xA\a ∈ XA\a ,
X
λ(ya , xA\a ) = 0.
ya
It is not hard to see that the set of A-degenerate functions is isomorphic to
the vector space ΛA ; both formulations will be useful.
Definition 5.5. Given a degenerate function εi : FRi → R, define
Di (εi ) = lim η −1 {p[1, . . . , 1 + ηεi , . . . , 1] − p[1, . . . , 1, . . . , 1]} ,
η↓0
so that Di (εi ) is a vector in R|XV W | , the directional derivative of the ith component of p[·] with respect to εi . For sufficiently small η > 0, the vector 1 + ηεi is
non-negative and therefore a valid distribution over FRi (up to the normalizing
constant); it follows that Di (εi ) ∈ TC0 , the tangent cone of M(G) at p0 .
Let Ti = {Di (εi ) | εi degenerate}. Since the function p[·] is differentiable at
[1, . . . , 1] it follows that Ti is a vector space, and also that the vector space T1 +
· · · + Tk is contained within the tangent cone of M at the uniform distribution.
We will show that T1 + · · · + Tk is in fact the same as (8).
It will be useful to define the following collection of supersets of Φi , for
B ⊆V:
ΦB
i (fπ(Ri ) , xV W ) ≡ {fRi | fv (xpa(v) , fπ(v) ) = xv for each v ∈ Ri ∩ B}.
(9)
That is, the collection of functions fRi such that, given inputs fπ(Ri ) and
xpa(Ri )\Ri , the values of fB∩Ri jointly evaluate to xB∩Ri . Note that ΦB
i = Φi
for any B ⊇ Ri .
Lemma 5.6. Let C ⊆ Ri , with sterileG (C) ⊆ A ⊆ C ∪ paG (C) and E ⊆ π(C).
Then for every degenerate function
λ : XA × FE → R,
there exists a degenerate function δ : FC → R such that
X
δ(fC ) = λ(xA , fE ),
fRi ∈Φi
where Φi is given by (7). In addition,
X
|XRi \B |λ(xA , fE ) if C ⊆ B
δ(fC ) =
0
otherwise.
B
fRi ∈Φi
25
Proof. See appendix, Section B.2.
Note that if we set E = ∅, the above result shows that for any λ ∈ ΛA there
exists a δ such that
η −1 {p[1, . . . , 1 + ηδ, . . . , 1] − p[1, . . . , 1, . . . , 1]}
X X
X
X
···
= η −1
ηδ(fC )
···
1
Φk
Φi
Φi−1
Φ1
= λ.
Hence ΛA ≤ Ti (i.e. ΛA is a subspace of Ti ) for any A such that sterileG (C) ⊆
A ⊆ C ∪ paG (C) and C ⊆ Ri .
The next result forms the backbone for proving Theorem 5.3: it extends
Lemma 5.6 to sets C that are not contained within a single remainder set.
Lemma 5.7. Let C be a bidirected-connected set, and for each i define Ci ≡
C ∩ Ri ; let I ≡ {i | Ci 6= ∅}. For sterileG (Ci ) ⊆ Ai ⊆ Ci ∪ paG (Ci ), let
i
A=
Ai .
i∈I
Then ΛA ≤ Tl , where l is the minimal element of I.
Proof. By Lemma B.5 (see Appendix), there exists a rooted tree Π with vertices
I, such that i → j in Π only if there exist vi ∈ Ri ∩ C and vj ∈ Rj ∩ C with
vj ∈ π(vi ). In particular i → j only if Cj ⊆ π(vi ).
Let l be the root node of Π, and for each j ∈ chΠ (l) denote by Πj the rooted
tree with root j formed only from the descendants of j.
Let λi : XAi → R be arbitrary Ai -degenerate functions for each i ∈ I. Then
starting with vertices which have no children (i.e. the leaves of the tree), and
using Lemma 5.6, recursively define δi for i ∈ I as the degenerate function of
fCi such that
X
Y
δj (fCj ),
δi (fCi ) = λi (xAi )
Φi
j∈chΠ (i)
where the empty product is defined to be equal to 1. Then
X
X
XX
X Y
δj (fCj ).
···
δl (fCl ) = λl (xAl )
···
Φk
Φl+1 Φl
Φk
Φl+1 j∈chΠ (l)
P
For each i ∈ I an expression of the form Φi δi (fCi ) is only a function of fCα
(and xAi ) for α ∈ chΠ (i) and Π is a tree, so the sum factorizes into components
only involving the descendants of each j ∈ chΠ (l):
X X
Y
X
···
δl (fCl ) = λl (xAl )
δj (fCj ).
Φk
Φ1
j∈chΠ (l)
Φs
s∈deΠ (j)
But then for each j the factor represents a disjoint sub-tree Πj with root node
j, so we can just iterate this process within each factor, and get
Y
=
(10)
λi (xAi ).
i∈I
26
f1 , f2
1
2
1
2
f3
3
4
f4
3
(a)
4
(b)
Figure 11: (a) an mDAG on 4 variables, and (b) a DAG with hidden variables
corresponding to a gearing of the mDAG in (a).
It follows that any function of the form (10) lies in Tl . Since ΛA is spanned by
such functions by Lemma B.3 (see Appendix B) it follows that ΛA ≤ Tl .
Corollary 5.8. For geared graphs G, we have
M
ΛA ≤ T 1 + · · · + T k .
A∈A(G)
Proof. Reformulating Lemma S
3.7 slightly,
S for any A ∈ A(G) there exists a
bidirected-connected set C = i Ci = i {vi1 , . . . , viki }, where Ci = C ∩ Ri
(we have changed nothing other than to label the vertices vij by which remainder set they are contained in). Then A is of the form
i j i i j
Ai
Ai =
A=
i
i,j
j
for some sets Aji such that {vij } ⊆ Aji ⊆ {vij } ∪ paG (vij ).
Applying
Lemma 3.7 in reverse to the bidirected-connected set Ci shows that
a
Ai ≡ j Aji is in A(G), and therefore satisfies sterileG (Ci ) ⊆ Ai ⊆ Ci ∪ paG (Ci ).
Then by Lemma 5.7 the space ΛA is contained in some Ti , i = 1, . . . , k.
Example 5.9. In the IV model from Figure 8 (see Examples 4.5 and 4.9) has
a saturated nested model with the following parameterizing sets:
head H
{1}
{2}
{3}
tail T
∅
{1}
{1, 2}
parametrizing sets A
{1}
.
{2}, {1,2}
{3}, {1,3}, {2,3}, {1,2,3}
Indeed, taking the functional parameterization suggested in Example 4.9,
one can see that altering the distribution of the compliance functions f2 will
affect the distribution of X2 conditional on X1 , which is why Λ2 and Λ12 are
contained in TC0 . For example, to introduce a correlation between X1 and X2
whilst keeping the marginal distributions fixed, we can increase the proportion
27
of ‘compliers’ (that is the people for whom f2 (0) = 0, f2 (1) = 1) and decrease
the proportion of ‘defiers’ (f2 (0) = 1, f2 (1) = 0).
Similarly, modifying the distribution of f3 gives us Λ3 and Λ23 . Obtaining
the directional derivatives in Λ13 and Λ123 requires modifying the distribution
of f2 , f3 jointly.
Example 5.10. Consider the single-district, geared mDAG in Figure 11(a);
the nested Markov model for this graph is saturated (i.e. has no constraints),
and thus its tangent space at p0 is of full dimension. Correspondingly, one can
check that A(G) consists of all non-empty subsets of {1, 2, 3, 4}:
head H
{1}
{2}
{1,2}
{3}
{1,3}
{4}
{2,4}
{3,4}
tail T
∅
∅
∅
{2}
{2}
{1}
{1}
{1,2}
parametrizing sets A
{1}
{2}
{1,2}
.
{3}, {2,3}
{1,3}, {1,2,3}
{4}, {1,4}
{2,4}, {1,2,4}
{3,4}, {1,3,4}, {2,3,4}, {1,2,3,4}
Let us now see how our previous results apply to the marginal model in this
case. Consider the gearing
B1 = {1, 2}
R1 = {1, 2}
B2 = {1, 3}
R2 = {3}
B3 = {2, 4}
R3 = {4}
and ordering 4 < 3 < 1 < 2 which respects this gearing. This leads to the
hidden variable model in Figure 11(b); here
f 3 : X2 → X3
f 4 : X1 → X4
f 1 : F3 → X1
f 2 : F4 → X2 .
Applying Lemma 5.6 to each remainder set in turn tells us that
Λ1 + Λ2 + Λ12 ≤ T1 ,
Λ3 + Λ23 ≤ T2 ,
Λ4 + Λ14 ≤ T3 .
We can apply Lemma 5.7 with the connected set C = {1, 2, 3, 4} to find that
ΛA ≤ T1 , where A is a set of the form A = {1, 2}△A2△A3 and
{3} ⊆ A2 ⊆ {2, 3},
{4} ⊆ A3 ⊆ {1, 4};
this gives us any A ∈ {{3, 4}, {1, 3, 4}, {2, 3, 4}, {1, 2, 3, 4}}, and so
Λ34 + Λ134 + Λ234 + Λ1234 ≤ T1 .
Repeating with C = {1, 2, 3} and {1, 2, 4} respectively gives Λ13 + Λ123 ≤ T1
and Λ24 + Λ124 ≤ T1 .
Thus for every non-empty A ⊆ {1, 2, 3, 4} there is some i ∈ {1, 2, 3} such that
ΛA ≤ Ti , and therefore the tangent cone of M(G) at the uniform distribution is
the same as that of the saturated model on four variables. In other words the
nested model and marginal model are both of full dimension.
Evans (2012) shows that the marginal model associated with this graph
induces some inequality constraints on the joint distribution, and so the nested
and marginal models are not identical.
28
1
2
1
2
1
2
4
3
4
3
4
3
(a)
(b)
(c)
Figure 12: (a) the bidirected 4-cycle, and (b), (c) two geared subgraphs.
5.3
Extension to non-geared graphs
Corollary 5.8 put us in a position to prove Theorem 5.3 for geared graphs;
however it does not so far extend to the general case, because we cannot fix
the state-spaces of the latent variables without a gearing. In this section we
will show that the tangent cone of a general marginal model at the uniform
distribution is the vector space spanned by the tangent cones of its geared
subgraphs, and that therefore the problem can be reduced to geared graphs.
Proposition 5.11. Let G be an arbitrary mDAG containing geared subgraphs
G1 , . . . , Gk . Suppose that, for each subgraph and a suitable gearing ΛAi ≤
TC0 (Gi ) as a consequence of the earlier results in this section. Then ΛA1 +
· · · + ΛAk ≤ TC0 (G).
In other words, the tangent cone of G includes the vector space spanned by
all the tangent cones of the subgraphs.
Proof. First consider the case W = ∅ and k = 2, from which the general result
will follow similarly.
Let p1 ∈ M(G1 ) ⊆ M(G) be formed by random functions fV according to a
gearing of G1 , and p2 ∈ M(G2 ) ⊆ M(G) by random functions f˜V according to a
gearing of G2 . Let Uv be independent Bernoulli( 21 ) variables, and define a new
distribution by setting
Zv = Uv fv (fπ(v) , Zpa(v) ) + (1 − Uv )f˜v (f˜π̃(v) , Zpa(v) );
i.e. we randomly (and independently of all other vertices) choose one of the
mechanisms fv or f˜v to generate Zv . Denote Xv ≡ fv (fπ(v) , Zpa(v) ) and Yv ≡
f˜v (f˜π(v) , Zpa(v) ). Note that although fv and f˜v are independent the values of
Xv and Yv are not, since they share parent variables.
Denote the resulting joint distribution of ZV by p. We have p ∈ M(G) since
we are still generating each variable as a random function of its parents and
some independent noise, which clearly satisfies the structural equation property
for Ḡ.
29
Splitting into cases indexed by B ≡ {v : Uv = 1}, we have
X
P (ZV = zV ) =
P (UB = 1, UV \B = 0, XB = zB , YV \B = zV \B )
B⊆V
=
1
2|V |
X
P (XB = zB , YV \B = zV \B ).
(11)
B⊆V
It follows from the proof of Lemma 5.7 that if A1 ∈ A(G1 ) and λA1 ∈ ΛA1 then
there exists a degenerate δ(fCi ) such that
X
X
X
···
δ(fCi ) · · ·
1 = λA1 (xA1 )
Φk
Φi
Φ1
and from Lemma 5.6 that
X
X X
|XV \B |λA1 (xA1 ) if Ci ⊆ B
1=
···
δ(fCi ) · · ·
0
otherwise.
B
B
B
Φk
(12)
Φ1
Φi
Similarly, given A2 ∈ A(G2 ) and λA2 ∈ ΛA2 there exists a δ̃(f˜Cj ) satisfying
equivalent conditions over f˜s.
Now since the functions used to generate XV and YV are independent,
P (XB = zB , YV \B = zV \B )
X
X
X
X
X
X
ρ̃1
ρ1
ρ̃k · · ·
(ρ̃j + η δ̃) · · ·
(ρi + ηδ) · · ·
ρk · · ·
=
ΦB
k
ΦB
1
ΦB
i
= |XB |−1 + ηc1 λA1
V \B
Φ̃k
|XV \B |−1 + ηc2 λA2
= |XV |−1 + η(c′1 λA1 + c′2 λA2 ) + O(η 2 ).
V \B
Φ̃j
V \B
Φ̃1
The first equality above follows just from consideration of which functions we
need to evaluate to which values in order to obtain XB = zB and YV \B =
zV \B (although this expression apparently factorizes, note that both factors
can depend upon all of zV ). Note that possibly c′i = 0, depending on which
of the conditions from (12) are satisfied. However, certainly c′i > 0 for some
subsets B, so plugging this back into (11) we get
P (ZV = zV ) = |XV |−1 + η(c′′1 λA1 + c′′2 λA2 ) + O(η 2 ),
where c′′i > 0. Then by an appropriate choice of scaling for each λAi we see that
ΛA1 + ΛA2 ≤ TC0 (G).
For non-empty W , we can draw ZW = XW = YW as a uniform random
variable, and then look at XV | XW ; the proof is otherwise the same.
Example 5.12. The bidirected 4-cycle in Figure 12(a) is not geared, and therefore we cannot apply our earlier results to it directly. The nested model for
this graph is equivalent to the model defined by the constraints X1 ⊥
⊥ X3 and
X2 ⊥
⊥ X4 , and has parameterizing sets
A(G) = {{1}, {2}, {1, 2}, {3}, {2, 3}, {1, 2, 3},
{4}, {1, 4}, {1, 2, 4}, {3, 4}, {1, 3, 4}, {2, 3, 4}, {1, 2, 3, 4}},
30
which are also the bidirected-connected sets of vertices. The two subgraphs in
Figures 12(b) and (c), say G1 and G2 , are geared, however, and have parameterizing sets
A(G1 ) = {{1}, {2}, {1, 2}, {3}, {2, 3}, {1, 2, 3},
{4}, {3, 4}, {2, 3, 4}, {1, 2, 3, 4}}
A(G2 ) = {{1}, {2}, {1, 2}, {3}, {4}, {1, 4},
{1, 2, 4}, {3, 4}, {1, 3, 4}, {1, 2, 3, 4}};
L
we have
A∈A(Gi ) ΛA ≤ TC0 (Gi ) for i = 1, 2 by Corollary 5.8. Note that
A(G1 ) ∪ A(G2 ) = A(G), and therefore by applying Proposition 5.11 with these
graphs, we find that
M
M
M
ΛA ≤ TC0 (G).
ΛA +
ΛA =
A∈A(G)
A∈A(G2 )
A∈A(G1 )
It follows that the marginal model is also defined by the independences X1 ⊥
⊥ X3
and X2 ⊥
⊥ X4 , possibly with some additional inequality constraints.
We are now in a position to put together these ideas and prove the main
result for general mDAGs.
Proof of Theorem 5.3. Suppose first that G is geared.
p[1, . . . , 1 + ηεi , . . . , 1] obeys the nested Markov property for any degenerate
function εi and η sufficiently small that 1 + ηεi is positive; it follows that Ti ≤
TC0 for each i, and that therefore using Corollary 5.8,
M
ΛA ≤ T 1 + · · · + T k
A∈A(G)
is also contained in TC0 , by the differentiability of p[·] at (1, . . . , 1).
Now for general G, and each A ∈ A(G), there exists a geared subgraph G ′ of
G such that ΛA ≤ TC0 (G ′ ) by Lemma 4.3. Then applying Proposition 5.11, we
see that the space spanned by these subspaces is contained within the tangent
cone for G:
M
ΛA ≤ TC0 (G).
A∈A(G)
If a distribution is in the marginal model then it is also in the nested model,
and therefore TC0 is contained within the tangent space TSn0 of N (G) at p0 ,
which has dimension
Y
X
|XT |
(|Xh | − 1)
dim(TSn0 ) =
h∈H
H∈H(G)
=
X
dim(ΛA );
A∈A(G)
Q
the second equality here follows from dim(ΛA ) = h∈A (|Xh | − 1) and
X
X
Y
Y
dim(ΛA ) =
(|Xh | − 1) = |XT |
(|Xh | − 1).
H⊆A⊆H∪T
H⊆A⊆H∪T h∈A
31
h∈H
Then combining
M
ΛA ≤ TC0 ⊆ TSn0
A∈A(G)
with the dimension of TSn0 gives the result.
As a corollary of this result, within a neighbourhood of p0 the models M(G)
and N (G) are the same. This is because N (G) is parametrically defined via
polynomials, and therefore its Zariski closure is an irreducible variety (see, e.g.
Cox et al., 2007, Proposition 4.5.5). For algebraic varieties V1 , V2 , if V1 ⊆ V2
and V2 is irreducible, then either V1 has a strictly smaller dimension than V2 ,
or they are identical. The results about the tangent space show that the Zariski
closures of M and N have the same dimension, and therefore they are the same.
This means that, locally to p0 , the models themselves are also the same.
6
Smoothness of the marginal model
The results of Section 5, together with the smoothness of the nested model,
allow us to show that for geared graphs, the interior of the marginal model is a
smooth manifold.
Theorem 6.1. For a geared graph G and state-space XV W , the relative interior
of the marginal model M(G) is a manifold of dimension d(G, XV W ), and its
boundary is described by a finite number of semi-algebraic constraints.
Proof. The nested Markov model is parametrically defined (with a polynomial
parameterization), and therefore its Zariski closure is an irreducible variety (see,
e.g. Cox et al., 2007, Proposition 4.5.5). Furthermore Evans and Richardson
(2015) give a diffeomorphism between the set of strictly positive distributions
obeying the nested Markov property, and an open parameter set. It follows that
N (G) is a manifold on the interior of the simplex (see, for example, Kass and Vos,
1997, Appendix A).
As noted in the previous section, the marginal model for a geared graph is a
semi-algebraic set. Since the M(G) ⊆ N (G) and these two sets have the same
Zariski closure, it follows that M(G) is defined from N (G) by a finite number of
additional polynomial inequalities. It further follows that it is also a manifold
at any point these inequality constraints are not active.
It follows from Theorem 6.1 that the interior of the marginal model for a
geared mDAG is a curved exponential family of dimension d(G, XV W ), and that
therefore the nice statistical properties of these models can be applied. For
example, the maximum likelihood estimator of a distribution within the model
will be asymptotically normal and unbiased, and the likelihood ratio statistic
for testing this model has an asymptotic χ2 -distribution.
For non-geared mDAGs we cannot assume that the latent variables are discrete without loss of generality, so it is conceivable that these marginal models
may be defined by non-polynomial inequalities on the probabilities. We conjecture, however, that a result akin to Theorem 6.1 does hold for general graphs.
For a point on the boundary defined by an active inequality constraint,
the asymptotic distribution of the likelihood ratio statistic may be much more
32
p̂
×
M
p̂m
N
p̂n
Figure 13: Diagramatic representation of estimation with the marginal model.
The thicker line represents the marginal model, and its thinner extension the
nested model. The unconstrainted MLE is shown as p̂, and its projection to
MLEs under the marginal and nested models as p̂m and p̂n respectively. Note
that p̂m is on the boundary of the model M; if the true data generating distribution is on the boundary this generally leads to irregular asymptotics.
complicated (Drton, 2009b); in general it is a mixture of χ2 -distributions, and
this mixture will vary depending upon the unknown truth. A possible advantage
of the nested model is that we can guarantee that the true distribution does not
lie on the boundary of N if the MLE consists of strictly positive probabilities,
because the boundary only consists of distributions with at least some zero
probabilities; the same cannot be said for M. This is depicted in Figure 2, in
which the MLE under the nested model (p̂n ) is in the interior of N , but the
MLE for the marginal model p̂m lies on the boundary of M.
Inequality constraints are generally much more complicated than equality
constraints, and efforts to characterize them fully in DAGs with latent variable
models have been limited by computational challenges. Evans (2012), generalizing a result first given by Pearl (1995), provides a graphical criterion for
obtaining some inequalities, but deriving a complete set of bounds may be an
NP-hard problem (ver Steeg and Galstyan, 2011).
6.1
Model Fitting
In theory we can exactly fit the marginal model for a geared graph using a
latent variable model of the kind derived in Section 4. In practice this model is
massively over-parameterized and unidentifiable, with the state-space of sets Fv
being potentially very large even for modest graphs; this will cause problems for
most standard fitting algorithms. We can restrict the state-space of the latent
variables to something more managable to obtain some latent variable model
L(G) ⊆ M(G); in general the inclusion will be strict, as in the example in the
introduction and as depicted in Figure 2. However, for any graph G—whether
geared or not—and any latent variable model we have L(G) ⊆ M(G) ⊆ N (G).
Fitting the nested model by maximum likelihood (ML) is straightforward using
the algorithm in Evans and Richardson (2010), and a latent variable model can
be fitted using (for example) an EM algorithm. A measure of goodness-of-fit for
these two models can be used to bound the goodness-of-fit of the marginal model,
and thus potentially used to confirm or refute the marginal model. Fitting the
marginal model directly is likely to be extremely difficult for general graphs: see
the discussion in Evans (2016).
33
Acknowledgements
We thank Angelos Armen for a very close reading and substantial comments, as
well as the associate editor and two anonymous referees for their suggestions.
References
A. Anandkumar, D. Hsu, A. Javanmard, and S. Kakade. Learning linear
Bayesian networks with latent variables. In Proceedings of The 30th International Conference on Machine Learning, volume 28, pages 249–257, 2013.
S. Basu, R. Pollack, and M.-F. Roy. Algorithms in real algebraic geometry.
Springer, 1996.
C. M. Bishop. Pattern recognition and machine learning. Springer, 2007.
D. Cox, J. Little, and D. O’Shea. Ideals, varieties, and algorithms: an introduction to computational algebraic geometry and commutative algebra. Springer,
2007. Third Edition.
A. Darwiche. Modeling and reasoning with Bayesian networks. Cambridge
University Press, 2009.
A. P. Dawid. Influence diagrams for causal modelling and inference. International Statistical Review, 70(2):161–189, 2002.
W. Decker, G.-M. Greuel, G. Pfister, and H. Schönemann.
Singular 4-1-0 — A computer algebra system for polynomial computations.
http://www.singular.uni-kl.de, 2016.
M. Drton. Discrete chain graph models. Bernoulli, 15(3):736–753, 2009a.
M. Drton. Likelihood ratio tests and singularities. Annals of Statistics, 37(2):
979–1012, 2009b.
R. J. Evans. Graphical methods for inequality constraints in marginalized
DAGs. In Machine Learning for Signal Processing (MLSP), 2012.
R. J. Evans. Graphs for margins of Bayesian networks. Scandinavian Journal
of Statistics, 43(3):625–648, 2016.
R. J. Evans and T. S. Richardson. Maximum likelihood fitting of acyclic directed mixed graphs to binary data. In Proceedings of the 26th conference on
Uncertainty in Artificial Intelligence (UAI-08), 2010.
R. J. Evans and T. S. Richardson. Markovian acyclic directed mixed graphs for
discrete data. Annals of Statistics, 42:1452–1482, 2014.
R. J. Evans and T. S. Richardson. Parameterization of the discrete nested
Markov model. In preparation, 2015.
C. J. Fox, A. Käufl, and M. Drton. On the causal interpretation of acyclic mixed
graphs under multivariate normality. Linear Algebra and its Applications,
2014.
34
T. Fritz. Beyond Bell’s Theorem: Correlation scenarios. New J. Phys, 14:
103001, 2012.
R. E. Kass and P. W. Vos. Geometrical foundations of asymptotic inference.
John Wiley & Sons, 1997.
S. L. Lauritzen. Graphical Models. Clarendon Press, Oxford, UK, 1996.
J. Neyman. Sur les applications de la théorie des probabilités aux experiences
agricoles: Essai des principes. Roczniki Nauk Rolniczych, 10:1–51, 1923. In
Polish; English translation by D. Dabrowska and T. Speed in Statist. Science
5 463–472, 1990.
J. Pearl. On the testability of causal models with latent and instrumental
variables. In Proceedings of the 11th conference on Uncertainty in Artificial
Intelligence (UAI-95), pages 435–443, 1995.
J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University
Press, second edition, 2009.
J. Provan and L. Billera. Decompositions of simplicial complexes related to
diameters of convex polyhedra. Mathematics of Operations Research, 5(4):
576–594, 1980.
T. S. Richardson. Markov properties for acyclic directed mixed graphs. Scand.
J. Statist., 30(1):145–157, 2003.
T. S. Richardson and P. Spirtes. Ancestral graph Markov models. Annals of
Statistics, 30(4):962–1030, 2002.
T. S. Richardson, R. J. Evans, and J. M. Robins. Transparent parameterizations
of models for potential outcomes. Bayesian Statistics, 9:569–610, 2011.
T. S. Richardson, R. J. Evans, J. M. Robins, and I. Shpitser. Nested Markov
properties for acyclic directed mixed graphs. arXiv:1701.06686, 2017.
J. Robins. A new approach to causal inference in mortality studies with a sustained exposure period-application to control of the healthy worker survivor
effect. Mathematical Modelling, 7(9):1393–1512, 1986.
D. B. Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5):688–701, 1974.
I. Shpitser, T. S. Richardson, J. M. Robins, and R. J. Evans. Parameter and
structure learning in nested Markov models. In UAI-12 (Workshop on Causal
Structure Learning), 2012.
R. Silva and Z. Ghahramani. The hidden life of latent variables: Bayesian learning with mixed graph models. The Journal of Machine Learning Research,
10:1187–1238, 2009.
P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction and Search.
MIT press, 2000.
35
J. Tian and J. Pearl. On the testable implications of causal models with hidden
variables. In Proceedings of the 18th Conference on Uncertainty in Artificial
Intelligence (UAI-02), pages 519–527, 2002.
G. ver Steeg and A. Galstyan. A sequence of relaxations constraining hidden
variable models. In Proceedings of the 27th Conference on Uncertainty in
Artificial Intelligence (UAI-11), 2011.
T. S. Verma and J. Pearl. Equivalence and synthesis of causal models. In
Proceedings of the 6th Conference on Uncertainty in Artificial Intelligence
(UAI-90), pages 255–268, 1990.
A
Degenerate Binary State-Space
Consider the model from Example 1.1 where we take all five variables to be
binary. The model can be written as
X
p(x0 ) · p(x1 ) · p(x2 | x1 , x0 ) · p(x3 | x2 ) · p(x4 | x3 , x0 )
p(x1 , x2 , x3 , x4 ) =
x0
= p(x1 ) · p(x3 | x2 ) ·
X
p(x0 ) · p(x2 | x1 , x0 ) · p(x4 | x3 , x0 ) .
x0
|
{z
p∗ (x2 ,x4 |x1 ,x3 )
}
The dimension of the model is therefore the sum of the dimensions of these three
factors (of which the first two are 1 and 2 respectively). For the final factor,
assuming X0 is binary means it may be written as
X
p(x0 ) · p(x2 | x1 , x0 ) · p(x4 | x3 , x0 )
p∗ (x2 , x4 | x1 , x3 ) =
x0
= α · q(x2 | x1 ) · q(x4 | x3 ) + (1 − α) · r(x2 | x1 ) · r(x4 | x3 )
(13)
for some distributions q, r. This is a parametric definition of a variety over the
probabilities p(x2 , x4 | x1 , x3 ), and using the computational algebra package
Singular (Decker et al., 2016) we explicitly found the polynomial constraints
that define it6 . It turns out that the set of such probabilities that can be written
in the form (13) has dimension 7, and therefore the total dimension of the latent
variable model is 1 + 2 + 7 = 10.
B
B.1
Technical Proofs
Degenerate Functions
We present a series of Lemmas which build up to showing that we can construct
degenerate functions from finite sums and products of degenerate functions with
simpler argument sets.
6 For
the
code
used
and
the
resulting
polynomial
constraints,
see
http://www.stats.ox.ac.uk/~ evans/bell.html.
As well as the usual marginal independence constraints, the latent variable model implies an additional cubic polynomial
constraint on the observed conditional probabilities (for which we were unable to find a nice
interpretation).
36
Lemma B.1. Let λ be a discrete (A ∪ B)-degenerate function, for A ∩ B = ∅.
Then λ can be written as a finite sum
X
λiA λiB
λ=
i
of A-degenerate functions
λiA ,
and B-degenerate functions λiB .
Proof. Since a matrix can be written as a sum of rank one matrices, clearly we
can find (not necessarily degenerate) functions such that the result holds. But
now suppose that the λiA are not degenerate over a ∈ A, and consider
!
X
X
i
i
λA (xA\a , ya ) λiB (xB )
λA (xA ) −
ya
i
=
X
λiA (xA )λiB (xB )
−
XX
ya
i
= λ(xA , xB ) −
X
λiA (xA\a , ya )λiB (xB )
i
λ(ya , xA\a , xB )
ya
= λ(xA , xB ).
Thus we can replace each λiA with the degenerate function
!
X
i
i
i
λA (xA\a , ya )
λ̃A (xA ) ≡ λA (xA ) −
ya
and not affect the result. By repeating the argument we can assume that each
λiA is degenerate in every a ∈ A, and each λiB degenerate in every b ∈ B.
Lemma B.2. Let λ be a discrete (A△B)-degenerate function. Then λ can be
written as a finite sum
X j j
λA λB
λ=
j
of A-degenerate functions λjA , and B-degenerate functions λjB .
Proof. Let A′ = A \ B and B ′ = B \ A and D = A ∩ B, so that A△B = A′ ∪ B ′ ,
A = A′ ∪ D and B = B ′ ∪ D; note that A′ , B ′ and D are all disjoint.
For each yD ∈ XD , define a degenerate function ηD (·; yD ) : XD → R by
Y
ηD (xD ; yD ) = α−1
|Xd |1{xd =yd } − 1 .
d∈D
where α = d∈D |Xd | · (|Xd | − 1) and 1 denotes an indicator function. One can
verify easily that
X
ηD (xD ; yD ) = 0
Q
xd ∈Xd
for any yD and xD\d , and that
X
ηD (xD ; yD )2 = 1;
yD ∈XD
37
in particular the last expression is independent of xD .
Now, let λ be a discrete (A△B)-degenerate function, and using Lemma B.1
write it as
λ=
j
X
λjA′ λjB ′
i=1
where λjA′ and λjB ′ are respectively A′ and B ′ degenerate. Then for each k ∈
j
jk
j
XD , define λjk
A = λA′ ηD (·; k) and λB = λB ′ ηD (·; k). Clearly each of these is
degenerate in A = A′ ∪ D and B = B ′ ∪ D respectively. Further,
j
X
X
jk
λjk
A λB =
j
X
X
λjA′ λjB ′ ηD (·; k)2
i=1 k∈XD
i=1 k∈XD
=
j
X
λjA′ λjB ′
i=1
=
j
X
X
ηD (·; k)2
k∈XD
λjA′ λjB ′
i=1
= λ.
a
Lemma B.3. Let λ : XA → R be an A-degenerate function, and let A = i∈I Ai
for some finite collection of sets {Ai : i ∈ I}. Then there exists a finite collection
of Ai -degenerate functions λji : XAi → R for i ∈ I, j ∈ J, such that
XY j
λ=
λi .
j∈J i∈I
Proof. This just follows from repeatedly applying Lemma B.2.
B.2
Proof of Lemma 5.6
Lemma B.4. Let X and Y be finite sets, define F = {f : X → Y}, and take
λ : Y → R. Then for any A ⊆ Y and x ∈ X ,
X
X
λ(f (x)) = |Y||X|−1
λ(y),
f ∈F
f (x)∈A
y∈A
and if x1 6= x2 ,
X
λ(f (x2 )) = |A||Y||X|−2 ·
f ∈F
f (x1 )∈A
X
λ(y).
y∈Y
In particular note that if λ is degenerate, the last expression is zero.
38
Proof. Clearly if A = Y, then
X
X
λ(f (x)) =
λ(f (x))
f ∈F
f ∈F
f (x)∈Y
= |Y||X|−1
X
λ(y),
y∈Y
since there are exactly |Y||X|−1 functions in F such that f (x) = y for each
y ∈ Y. The first result follows in general by applying the result for A = Y to
the function λ′ (y) = λ(y)1{y∈A} .
The second result follows by similar combinatorical methods.
Proof of Lemma 5.6. It is clear that we only need prove the result for E = ∅,
since we can just incorporate fE as though they were observable parents of C,
and the result is the same.
First consider the case C = {v}; let L = paG (v) and take any set K ⊆ L. Let
λ : Xv × XK → R be a degenerate function, and for each f : XL × Fπ(x) → Xv ,
define
X
δ(f ) =
λ(f (yL , gπ(v) ), yK ).
yL ∈XL
gπ(v) ∈Fπ(v)
Then for fixed xv , xL , fπ(v) ,
X
δ(f ) =
f ∈Fv
f (xL ,fπ(v) )=xv
X
X
λ(f (yL , gπ(v) ), yK )
f ∈Fv
yL ∈XL
f (xL ,fπ(v) )=xv gπ(v) ∈Fπ(v)
=
X
X
λ(f (yL , gπ(v) ), yK ).
yL ∈XL
f ∈Fv
gπ(v) ∈Fπ(v) f (xL ,fπ(v) )=xv
But since λ is degenerate, the inner sum is zero unless both xL = yL and
fπ(v) = gπ(v) by Lemma B.4. This leaves
=
X
λ(f (xL , fπ(v) ), xK )
f ∈Fv
f (xL ,fπ(v) )=xv
= |Xv ||XL ||Fπ(v) |−1 · λ(xv , xK )
again by Lemma B.4, where the constant represents the number of distinct
functions f ∈ Fv such that f (xL , fπ(v) ) = xv . Hence the result holds for
C = {v}.
Now consider a general C ⊆ Ri ; we prove the result by induction on the size
of C. Given any sterileG (C) ⊆ A ⊆ C ∪ paG (C), we first claim that we can write
A = A1 △A2 where sterileG (Ci ) ⊆ Ai ⊆ Ci ∪ paG (Ci ) for i = 1, 2 and disjoint
non-empty C1 , C2 with C1 ∪ C2 = C.
To see this pick C2 = {w}, C1 = C \ {w} for some w ∈ sterileG (C), and
then set A1 = (A ∪ sterileG (C1 )) ∩ (C1 ∪ paG (C1 )) and A2 = A \ A1 . Clearly
A1 satisfies the required conditions. Since w was chosen to be sterile in C we
39
have w ∈
/ A1 and therefore w ∈ A2 ; in addition, the only elements of A not
contained in A1 are those which are neither in C1 nor paG (C1 ); but since they
are in C ∪ paG (C), they must instead be in {w} ∪ paG (w). Hence the claim
holds.
Now first suppose that λ = λ1 · λ2 for degenerate functions λi : XAi → R.
By the induction hypothesis, we can find degenerate δ1 , δ2 such that
X
δ1 (fC1 ) = c1 · λ1 (xA1 )
fv (x,f )=xv
v∈C1
X
δ2 (fC2 ) = c2 · λ2 (xA2 ).
fv (x,f )=xv
v∈C2
(Here we have written fv (x, f ) for fv (xL , fπv ) to reduce notational clutter.)
Then letting E = Ri \ C,
X
X
X
X
δ1 (fC1 ) · δ2 (fC2 )
δ1 (fC1 ) · δ2 (fC2 ) =
fv (x,f )=xv
v∈Ri
fv (x,f )=xv fv (x,f )=xv fv (x,f )=xv
v∈C2
v∈C1
v∈E
= c0
X
X
δ1 (fC1 ) · δ2 (fC2 )
fv (x,f )=xv fv (x,f )=xv
v∈C2
v∈C1
= c0
X
δ1 (fC1 )
fC1 ∈FC1
fC1 (x)=xC1
X
δ2 (fC2 )
fC2 ∈FC2
fC2 (x)=xC2
= c0 c1 c2 · λ1 (xA1 ) · λ2 (xA2 ).
However a general degenerate function λ : XA → R can be written as a finite
linear combination
X j j
λ1 · λ2
λ=
j
of degenerate functions λji : XAi → R, so the result follows by linearity of
summations.
For the final part, note that if v ∈ C \ B, then the summation over ΦB
i will
include every function fv ∈ Fv . Since δ is degenerate and a function of fv , the
sum is 0. On the other hand, if v ∈ (Ri ∩ B) \ C, then δ is not a function of fv
and summing over all Fv just involves |Xv | identical terms.
Lemma B.5. Let G be a single-district, geared mDAG, and C a bidirectedconnected set of vertices. There exists a rooted tree ΠC with vertex set
IC = {i | Ri ∩ C 6= ∅},
and edges i → j only if there exist vj ∈ Rj ∩ C and vi ∈ Ri ∩ C such that
vj ∈ π(vi ).
Proof of Lemma B.5. First construct a directed graph Π∗ on IC in which i → j
precisely when there exist vj ∈ Rj ∩ C and vi ∈ Ri ∩ C such that vj ∈ π(vi ).
Note that vj ∈ π(vi ) implies r(vj ) > r(vi ), so Π∗ is acyclic.
40
Let j be the minimal element of IC ; we claim that for any other i ∈ IC ,
there is always a directed path in ΠC from j to i. To see this, note that since C
is bidirected-connected, there is a bidirected path in G from some vj ∈ C ∩ Rj
to vi ∈ C ∩ Ri ; given such a path, ρ, trim it so that only the end-points are in
C ∩ Rj and C ∩ Ri respectively.
If ρ is just vj ↔ vi , then we are done, since vj ∈ π(vi ) by definition of π.
Otherwise, ρ begins vi ↔ vk ↔ · · · for some vk ∈ Rk ∩ C, where i > k > j. So
we can apply an inductive argument to find a path from j to k in Π∗C , and the
edge vi ↔ vk implies that k → i in Π∗C .
Now, Π∗C is a connected DAG with a unique root node j, so we can simply
take any singly connected subgraph ΠC to fulfil the conditions of the lemma.
41
| 10math.ST
|
Deep Convolutional Neural Networks for Massive
MIMO Fingerprint-Based Positioning
Joao Vieira, Erik Leitinger, Muris Sarajlic, Xuhong Li, Fredrik Tufvesson
arXiv:1708.06235v1 [stat.ML] 21 Aug 2017
Dept. of Electrical and Information Technology, Lund University, Sweden
[email protected]
Abstract—This paper provides an initial investigation on
the application of convolutional neural networks (CNNs) for
fingerprint-based positioning using measured massive MIMO
channels. When represented in appropriate domains, massive
MIMO channels have a sparse structure which can be efficiently learned by CNNs for positioning purposes. We evaluate
the positioning accuracy of state-of-the-art CNNs with channel
fingerprints generated from a channel model with a rich clustered
structure: the COST 2100 channel model. We find that moderately deep CNNs can achieve fractional-wavelength positioning
accuracies, provided that an enough representative data set is
available for training.
Index Terms—Deep Learning, Convolutional Neural Networks,
Massive MIMO, Fingerprint, Positioning, Localization.
I. I NTRODUCTION
Massive MIMO is a candidate technology to integrate next
generation cellular systems, such as 5G systems, and deliver
manifold enhancements in the communications link [1]. In its
conceived form [2], massive MIMO uses a large number of
base station (BS) antennas together with measured channel
state information (CSI) to multiplex user terminals spatially.
Measured CSI is essential to yield spectrally efficient communications, but it can also be a key enabler to achieve highlyaccurate terminal positioning, where down to centimeter order
accuracy may be required in some 5G applications, e.g.,
autonomous driving [3]. Explained briefly, since positioning
is a spatial inference problem, it makes sense to use large
antenna arrays that oversample the spatial dimension of a
wireless channel (thus benefiting from, e.g., increased angular
resolution, resilience to small-scale fading, and array gain
effects) to aid the positioning task.
Several approaches that make use of measured massive
MIMO channels for positioning exist. For example, the approach proposed in [4] detects a terminal’s position (from a
grid of candidate positions) using line-of-sight (LOS) based
triangulation from a terminal to several distributed massive
MIMO BSs. In [5], positioning is performed using the phases
of multipath components estimated from massive MIMO channels. Another positioning approach was proposed in [6], where
received signal strength (RSS) based fingerprinting from one
single-antenna terminal to NBS M -antenna BSs is employed.
Here, the challenge was to learn the inverse map
−1
fRSS
: {|Yi |2 } → {xi }
(1)
from a set of training observations, i.e. the training set
Train
. Here the label xi ∈ R2x1 is the 2{Yi , xi }N
i=1
dimensional terminal coordinate of training observation i, and
Yi ∈ CM xNBS is its associated channel fingerprint. Gaussian
−1
Process-based Regression was used to learn fRSS
(·).
What the previous two mentioned positioning proposals and
other proposals have in common (e.g., [7] and [8]), is that
the structure of their solutions is typically composed by 2
distinct steps. In the first step, empirical feature extraction
(from the measured channel snapshots) is performed (e.g.,
RSS), and in the second step, positioning of the terminal is
done using the extracted features and a suitable algorithm-the
algorithm typically being is the main contribution of the work.
Although such 2 step solutions simplify the entire positioning
task, they are inherently sub-optimal since they are constrained
to use only partial–and typically not statistically sufficient [9]–
channel statistics to solve the problem at hand. Thus, is it of
interest to explore positioning frameworks that jointly extract
and process channel features from measurements–under some
joint optimality criterion.
In this work, fingerprinting-based positioning is performed
using a framework that jointly extracts and processes channel
features for positioning. More specifically, we are interested
to learn
f −1 : {s (Yi )} → {xi },
(2)
i.e., the inverse of the underlying function f (·) that maps
a set of single-antenna terminal coordinate vectors {xi } to
their respective measured but transformed channel snapshots
s (Yi ) ∈ Cd1 x...xdD ∀i, where D is the dimensionality of
the transformed snapshot. We note that the main point of the
transformation s(·) is to obtain a sparse representation for
s (Yi ). This is motivated in detail in Sec. II-B. For now, we
remark that the sparse transformations considered in this work
are bijective, and thus yield no information loss.1
Our proposal to learn (2) is by means of deep convolutional
neural networks (CNNs). Deep neural networks provide stateof-the-art learning machines that yield the most learning
capacity from all machine learning approaches [10], and lately
1 We remark that this positioning approach is inherently designed for the
single-user case. This fits well within a massive MIMO context since mutually
orthogonal pilot sequences, which are seen as sounding sequences in the
context of this work, are typically used by different users during uplink
training [2]. The extention of this approach to a multi-user case is thus
straightforward.
have been very successful in image classification tasks. Just
like most relevant information for an image classification task
is sparsely distributed at some locations of the image [10],
measured channel snapshots Yi have, when represented in
appropriate domains, a sparse structure which–from a learning
perspective–resemble that of images. This sparse channel
structure can be learned by CNNs and therefore used for
positioning purposes. To the best of the authors’ knowledge
there is no prior work on this matter.
The main contributions of this paper are summarized below.
•
•
We investigate the feasibility of deep CNNs for
fingerprint-based positioning with massive MIMO channels, and provide insights on how to design such networks
based on machine learning and wireless propagation
theory.
As a proof-of-concept, we demonstrate the accuracy of
our approach by performing fractional-wavelength positioning using channel realizations generated from a
widely accepted cellular channel model: the COST 2100
MIMO channel model [11].
II. C HANNEL F INGERPRINTING AND P RE -P ROCESSING
In this section, we explain the fingerprinting scenario addressed in this work, and motivate why CNNs are learning
machines suitable to perform positioning under such scenarios.
To maximize insights, we focus most motivational remarks on
the current case-study, but also provide several generalization
remarks at the end of the section.
A. Channel Fingerprinting
In this work, we assume a BS equipped with a linear M antenna array made of omnidirectional λ/2-spaced elements,
and that narrowband channels sampled at NF equidistant
frequency points are used for positioning. With that, the
dimensionality of each channel fingerprint Yi (and, as it will
be seen later, of each transformed fingerprint s(Yi )) is D = 3
and
in this work, f (·) is implemented as a bijective deterministic
map, i.e., there is only one unique fingerprint per position.2
B. Motivation for CNNs and Sparse Input Structures
Applying standard feed-forward neural networks to learn
the structure of {Yi } may be computationally intractable,
specially when M grows very large. However, the structure of
neural networks can be enhanced, both from a computational
complexity and a learning point-of-view, if designed with
sparse interaction and parameter sharing properties [10]. This
is a widely used architecture for CNNs, suitable to process
inputs with grid-like structures (e.g., an image can be thought
as a two-dimensional grid of pixels) with minimal amounts of
pre-processing.
CNNs are efficient learning machines given that their inputs
meet the following two structural assumptions:
1) most relevant information features are sparsely distributed
in the input space;
2) the shape of most relevant information features is invariant to their location in the input space, and are well
captured by a finite number of kernels.
From a wireless channel point-of-view, these assumptions
apply well when channels snapshots (i.e., the CNN inputs)
are represented in domains that yield a sparse structure [13].
For example, in the current case study, sparsity is achieved by
representing Yi in its, so-called, angular-delay domains, see
Fig. 1. Trivially, s(·) can take the form of a two-dimensional
discrete Fourier transform, i.e.,
s(Yi ) = F Yi FH .
(3)
If specular components of the channel, which are typically
modeled by Dirac delta functions [13], are seen as the information basis for positioning, then the two structural assumptions
of the CNNs inputs listed above are met. The same applies,
if instead, clusters of multipath components are seen as the
information features for positioning.
C. Generalizing the Current Case Study
d1 = M, d2 = NF , and d3 = 2.
The current case study can be extended to more
generic/higher-dimensionality fingerprints. For example, if
{Yi } is comprised by snapshots measured from a multiantenna terminal to a BS array, both with arbitrary array
structures and non-omnidirectional antenna elements, the effective aperture distribution function (EADF) [14] may be
accounted in s(·) in order to obtain a sparse fingerprint
Given a terminal position, its associated fingerprint is generated through f (·), i.e. the inverse of the function we wish
to learn. We implement f (·) using the COST 2100 channel
model, the structure of which is illustrated by Fig. 1, under
the parametrization proposed in [12]. This parametrization was
performed for outdoor environments and is further detailed in
Sec. IV. However, we note that our method is not restricted
to work only in outdoor channels–we remark on the required
channel properties in Sec. II-B. It is important to note that,
2 Bijectiveness of f (·) applies in most practical propagation scenarios with
high probability (the probability typically approaches one as M increases).
This is an important aspect to consider as it addresses the conditions needed
to be able to use CNNs (or more generally, to solve the inverse problem). On
a different note, regarding the deterministic structure for f (·), this is done
by generating both training and test sets from the same given realization
of the COST 2100 channel model stochastic parameters. This makes each
fingerprint to be completely determined solely by the geometry of the
propagation channel itself. Stochastic effects in the fingerprinting process,
such as measurement and labeling noise, or even time-variant channel fading
are interesting impairments to be considered in the design of CNN in future
work. For now, we focus on the case of having unique fingerprints per position,
due to simplicity.
r
aA
y
ra
Magnitude of a (Sparse) Channel Snapshot
n
en
t
An
ar
3
Delay [µs]
e
Lin
25
λ
×10 -4
1
2.5
0.8
2
0.6
1.5
0.4
1
0.2
0.5
0
25
50
75
100
125
150
175
Angle [ °]
Fig. 1. Left–Link setup considered in this work: an M -linear BS array positioning one single-antenna terminal in a confined square area. Channel realizations
are generated through the COST 2100 MIMO channel model. This geometry-based stochastic channel model is composed by different types of clusters of
multipath components (MPCs) that illuminate certain visibility regions (VRs) of an area. Right–Example of the magnitude of a channel snapshot represented
in a sparse domain. Such channel channel has a rich structure that can be learned by a CNN for positioning purposes.
representation. However, we emphasize that, in contrast to
most propagation studies, one does not need to necessarily deembed the measurement system from the propagation channel
in order to obtain valid fingerprints for positioning. Also,
multi-antenna channels yielding phenomena such as violation
of the plane wave assumption, or even the existence of cluster
visibility regions [15], can be made sparse by means of proper
transformations, e.g., generalized Fourier transforms. In any
case, the key is the ability to obtain a sparse representation
for s (Yi ).
III. D EEP CNN A RCHITECTURE
In this section, we describe the network architecture used
for learning f −1 (·), and discuss some design aspects.
For notation convenience, we drop the dependence of the
training sample index i, and write Y , Yi and x , xi until
explicitly stated otherwise.
A. Convolutional-Activation-Pooling Layers
After the input layer, which takes the transformed snapshots, a typical structure of CNNs employs a cascade of
L convolutional-activation-pooling (CAP) layers. Each CAP
layer is composed by: i) a convolutional operation of its input
with K convolutional Kernels, ii) a non-linear transformation,
i.e., activation function, and iii) a pooling layer, respectively.
A detailed description of the CAP layer structure used in this
work follows below.
Let the tensor H`−1 ∈ RM xNF xS3 be the input of the `th
CAP layer, with 1 ≤ ` ≤ L. Also, let the jth convolutional
Kernel of the `th layer be denoted by wj` ∈ RS1 xS2 xS3 , with
1 ≤ j ≤ K, and S1 and S2 denoting the sizes of the Kernels
S1 xS2 xS3
(which are CNN hyper-parameters). With h`−1
r,c ∈ R
3
being a sub-tensor of a zero-padded version of H`−1 , an
zero-padded version of H`−1 is obtained by padding the borders of
the volume of H`−1 with zeros, such that, when convolved with any Kernel
wj` , the input and output volumes are the same [10].
3 The
output entry of the `th convolutional layer can be written as
c`r,c,j = b`j + 1T wj` ◦ h`−1
1.
(4)
r,c
Here ◦ denotes the Hadamard product, b`j is a bias term, 1
denotes the all-ones column vector, and r and c are indices in
the convolution output volume which are implicitly defined.
In the first CAP layer, the input H0 is a tensor made out of
the complex-valued entries of (3), and thus we have S3 = 2
(real dimensions). This is because, although channel snapshots
are inherently complex-valued, we pursue an implementation
of a real-valued CNN with real-value inputs (thus we have
S3 = 2 in the first CAP layer). We motivate why we do so in
Sec. III-D. In the remaining CAP layers, we have S3 = K.
Each convolutional output entry (4) is fed to an activation
function. We use the current default choice for activation
functions in CNNs, namely, the rectified linear unit (RELU)
[10], where the output can be written as
`
gr,c,j
= max c`r,c,j , 0 .
(5)
Finally, after each activation function follows a pooling operation which down samples the outputs of the activation
functions. A standard option, also used here, is to forward
propagate the maximum value within group of N1 × N2
activation functions outputs. The pooling result can be written
as
N1
N2
`
h`r,c,j =max max g(r−1)N
.
(6)
1 +m, (c−1)N2 +n, j
m=1
n=1
B. Fully-Connected Layer
A fully-connected layer, following the L CAP layers, finalizes the CNN. With that, the position estimate of the network,
t ∈ R2x1 , is given by
t = W vec HL + bL ,
(7)
L T
where vec {·} vectorizes its argument, bL = [bL
is a
1 b2 ]
vector of biases, and W is a weight matrix whose structure is
implicitly defined.
C. Network Optimization
wj`
The CNN network learns the weights W and
,
and biases b`j , in order to make t the best approximation of x. Since we address positioning as a regression problem, we use the squared residuals averaged over
the training set as the optimizing metric. Re-introducing
the dependence on the sample index i, and defining the
column
θ by stacking all network parameters
as
h vector
iT
1
T 1
K T
L
θ = vec [w1 . . . wL ] vec {W} [b1 . . . b2 ] , the optimum parameters are given by
θ̂ = argmin J(θ),
(8)
θ
with
J(θ) =
NX
train
1
β T
(xi − ti (θ))2 .
θ θ+
2
Ntrain i=1
be statistically dependent, but different clusters are typically
not.
E. Complexity Aspects
The most computationally challenging aspect of the entire
approach is the optimization stage of the network. This is due
to the large size of the training sets, network dimensionality,
non-convexity of J(θ), etc. However, once the optimization
stage is finalized, real-time positioning can easily be achieved
due to the feed-forward structure of the network. This can
be easily observed (in the current case-study) by looking
at the overall complexity order of a CNN point-estimate:
O(K 2 M LNF S1 S2 ). The fact that the complexity does not
depend on the training set size Ntrain is one main advantage
of using CNNs for positioning.
(9)
IV. P OSITIONING R ESULTS
A Tikhonov penalty term is added to harvest the benefits of
regularization in CNNs–β is its associate hyper-parameter. On
a practical note, we minimize (9) using stochastic gradientdescent and back propagation [10].
Next, we address the positioning capabilities of CNNs by
means of numerical results. We omit showing optimization
aspects of the network (e.g., convergence across epochs) as
the main point of the paper is to analyse the positioning
capabilities of an optimized CNN.
D. Network Design Considerations
A. Simulation Setup
We now motivate our choice for the implementation of
real-valued CNNs. From our experience, the main challenge
of generalizing CNNs to the complex-valued case appears to
be in finding a suitable generalization of (5), the RELU, in
order to ”activate” complex inputs. For example, the complexvalued CNN generalizations presented in [16] apply (5) to
both real and imaginary parts separately, which from our
experience appears not to perform well from a network optimization point-of-view. Explained briefly, such RELUs are
non-continuous functions in C, as opposed to only being nondifferentiable in R as in real-valued network. This makes the
network’s optimization unstable with the current optimization
method (i.e. gradient descent). Therefore, as real-valued CNNs
have been very successful in image classification tasks, we
conservatively choose to pursue this design option, and let the
extension to the complex-valued case to be a matter of future
work.
A practical remark regarding the choice of the number of
CAP Kernels, K, follows. In practice, information features,
addressed in Sec. II-B, are subject to a number of variability
factors. For example, the shape of a ”measured Dirac delta
function” can vary with its location due to discretization. Also,
it is clear that, there is a higher variability in the informational
features if they are seen as clusters of MPCs, rather than single
MPCs themselves. In both cases, a practical approach to deal
with such variability factors is to set a high number for K,
and let the CNNs learn a set of kernels that span most possible
variations.
To finalize, if the information features are seen as the
clusters of MPCs, the sizes of the Kernels, S1 and S2 ,
should cover their range in the angular and delay channel
representations. This is because MPCs within a cluster may
The setup used in our experiments is illustrated by Fig. 1:
the terminal is constrained to be in a square area A of 25 × 25
wavelengths. Channel fingerprints are obtained in this area
through the COST 2100 channel model under the 300 MHz
parameterization (e.g., for path-loss and cluster-based parameters) established in [12]. The remaining parameters are shown
in Table I, and the other CNNs hyper-parameters, i.e. L and
K, are varied during the simulations.
TABLE I
C HANNEL AND CNN S PARAMETERS
Parameter
Variable
Value
Carrier frequency
Bandwidth
# Frequency points
# BS antennas
First BS antenna coordinate
Last BS antenna coordinate
fc
W
NF
M
B1
BM
300 MHz
20 MHz
128
128
[−200λ − 200λ]T
[−200λ − 200λ +
Tikhonov hyper-parameter
Kernel angular length [ ◦ ]
Kernel delay length [µs]
Pooling windows length
β
S1
S2
N1 and N2
10−3
9.8
0.175
2
(M −1)λ T
]
2
The closest and furthest coordinate points of A with respect
to the first BS antenna are:
uc = [−12.5λ − 12.5λ]T and uf = [12.5λ 12.5λ]T ,
respectively (i.e., the user is at least ||uc −B1 ||/λ wavelengths
away from the first BS antenna). The coordinates of these
two spatial points implicitly define the relative orientation of
the linear array with the area A. Similarly, the upcoming
performance analysis is done by means of the normalized root
mean-squared error (NRMSE) where the mean is calculated as
the average over the test sets samples. Thus, we have
v
u
Ntest
1u
1 X
NRMSE = t
(xi − ti (θ))2 .
λ Ntest i=1
This error metric has an understandable physical intuition as
it shows how the error distance relates to the wavelength.
The CNN training and testing is described as follow:
1) First, the training set is obtained by fingerprinting a 2dimensional uniformly-spaced (thus, deterministic) grid
of positions spanning the totality of A. The impact of
the sampling density is discussed in Sec. IV-C.
2) For the test set, each fingerprint’s position is obtained by
sampling a random variable with a uniform distribution
with support A.
Note that, if the CNN cannot use the available fingerprints
for training, then the position estimator is E{x} = 0, see (9).
Its NRMSE, for the current case study, is given by
s
Z
1
1
2
ref
R
(d − E{x}) ∂d ≈ 10.2.
NRMSE (A) =
λ
∂d A
A
(10)
Obviously, this reference value increases when A is larger.
Since an optimized CNN with a non-zero number of fingerprints should be able to do better or equal than (10), we use
(10) as a reference level in the analysis.
For benchmarking purposes, we also contrast our CNN
results against a standard non-parametric fingerprinting approach [17]. Seeing a training fingerprint as a function of its
position, i.e. Yi (xi ), this approach computes the position from
a new fingerprint Ynew through a grid-search over normalized
correlations as
we see that for the same parameterizations, a network fed
with untransformed inputs (i.e., s (Yi ) = Yi ) cannot effectively learn the channel structure for positioning purposes–
the order of magnitude of the positioning error is similar to
(10). However, with transformed inputs, fractional-wavelength
positioning can be achieved in both network settings, with the
lowest achieved test NRMSE being of about −6dB ≈ 1/2
of a wavelength. This showcases the capabilities of CNNs to
learn the structure of the channel for positioning purposes. We
remind that such positioning accuracies are attained with only
20 MHz of signaling bandwidth, see Table I, which suggests
that CNNs can efficiently trade-off signal bandwidth by BS
antennas and still achieve very good practical performance.
Decreasing the error further than fractional-wavelength ranges
becomes increasingly harder due to the increased similarities
of nearby fingerprints–such range approaches the coherence
distance of the channel. Also, as an illustrative example of the
positioning accuracy, Fig. 2 (right) shows the point-estimates
of the CNN under the parameterization that attained the lowest
test NRMSE.
C. Accuracy for Different Training Grids
To finalize, we analyze the impact of spatial sampling during
training. For benchmarking, we contrast the CNN performance
with the performance of the correlation-based classifier (11).
We use the CNN hyper-parameters that attained the lowest
MSE in Fig. 2, namely, the model with L = 4 and K = 20.4
Fig. 3 contrasts the NRMSEs obtained from a CNN and the
correlation-based classifier (11), against spatial sampling in
the training set. Overall, both approaches are able to attain
fractional-wavelength accuracies at smaller training densities.
Noticeably, the CNN tend to behave better than (11) for
less dense training sampling. Given that (11) does not have
interpolation abilities, this result is closely connected with the
| Tr{Yi (xi )H Ynew }|
p
x̂i = argmax
. inherent interpolation abilities of the CNNs. The fact that the
H Y
Ntrain
| Tr{Yi (xi )H Yi (xi )} Tr{Ynew
CNNs achieve similar, or even superior performance compared
new }|
xi ∈{xi }i=1
(11) to standard non-parametric approaches while having attractive
The following remarks can be made about this non-parametric implementation complexity further corroborates their use in
approach:
fingerprint-based localization systems.
1) Compared to the use of CNNs, a main drawback is its
V. TAKEAWAYS AND F URTHER W ORK
computational complexity order, O(M NF2 Ntrain ), as it
We have investigated a novel approach for massive MIMO
depends on the size of the training set;
fingerprint-based
positioning by means of CNNs and measured
2) it has no inherent interpolation abilities, and thus its error
channel
snapshots.
CNNs have a feedforward structure that is
can be lower bounded given the spatial density of the
able
to
compactly
summarize
relevant positioning information
training set;
in large channel data sets. The positioning capabilities of
B. Proof-of-Concept and Accuracy for Different CNN CNNs tend to generalize well, e.g. in highly-clustered propaParametrizations
gation scenarios with or without LOS, thanks to their inherent
Here, we report the positioning results when the spacing feature learning abilities. Proper design allows fractionalbetween neighbor training fingerprints is λ/4. We consider wavelength positioning to be obtained under real-time requirethis extreme case for now, in order to mitigate the impact of ments, and with low signal bandwidths.
The current investigation showcased some of the potentials
spatial undersampling from the results-here the focus is solely
of
CNNs for positioning using channels with a complex
on the positioning capabilities of the network. The impact of
the training fingerprints spacing is addressed later in Sec. IV-C.
4 Ideally, the CNN hyper-parameters should be tuned according to the
Fig. 2 (left) illustrates the positioning accuracy for different current training set. However, we keep the same hyper-parameterization
cases of CNN parameterizations. First, and as a sanity check, throughout this analysis, for simplicity.
225
15
True Positions
CNN Estimates
10
5
eq. (10)
K=20
0
Y coordinate [λ]
20 log10 NRMSE
Untransformed
K=10
Transformed
-5
-10
3
4
5
6
7
8
Number of CAP layers L
220
RMSE = 0.6 [λ]
215
210
205
200
200
205
210
215
220
225
X coordinate [λ]
Fig. 2. Left–NRMSE obtained by CNNs under different parameterizations. The upper horizontal line corresponds to the reference level (10). Here we only
report the test error, since a similar error value was obtained during training (i.e., no overfitting exists). Right–An illustrative example of the point-estimates
from a pre-defined set of positions by the optimized CNN (from the left figure) with K = 20 and L = 4.
5
CNN
Correlation-based
NRMSE
4
3
2
1
0
0
1
2
3
4
5
Spacing between Training Samples [λ]
Fig. 3. NRMSE obtained by different positioning approaches for different
spacings between samples of the uniform training grid.
structure. However, the design of CNNs in this contexts should
be a matter of further investigation, in order to be able to deal
with real-world impairments during the fingerprinting process.
In this vein, some questions raised during this study are, for
example, i) how to achieve a robust CNN design that is able
to deal with impairments such as measurement and labeling
noise, or channel variations that are not represented in the
training set, or ii) how to design complex-valued CNNs that
perform well and are robust during optimization.
R EFERENCES
[1] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Communications
Magazine, vol. 52, no. 2, pp. 186–195, 2014.
[2] T. L. Marzetta, “Noncooperative cellular wireless with unlimited numbers of base station antennas,” IEEE Transactions on Wireless Communications, vol. 9, no. 11, pp. 3590–3600, Nov. 2010.
[3] Z. Chaloupka, “Technology and Standardization Gaps for High Accuracy
Positioning in 5G,” IEEE Communications Standards Magazine, vol. 1,
no. 1, pp. 59–65, March 2017.
[4] N. Garcia, H. Wymeersch, E. G. Larsson, A. M. Haimovich, and
M. Coulon, “Direct Localization for Massive MIMO,” IEEE Transactions on Signal Processing, vol. 65, no. 10, pp. 2475–2487, May 2017.
[5] X. Li et al., “Robust Phase-Based Positioning Using Massive MIMO
with Limited Bandwidth,” in 2017 IEEE International Symposium on
Personal, Indoor and Mobile Radio Communications (PIMRC), Oct
2017.
[6] V. Savic and E. G. Larsson, “Fingerprinting-Based Positioning in
Distributed Massive MIMO Systems,” in 2015 IEEE 82nd Vehicular
Technology Conference (VTC2015-Fall), Sept 2015, pp. 1–5.
[7] S. A. Shaikh and A. M. Tonello, “Localization based on angle of arrival
in EM lens-focusing massive MIMO,” in 2016 IEEE 6th International
Conference on Consumer Electronics - Berlin (ICCE-Berlin), Sept 2016,
pp. 124–128.
[8] M. Zhu, J. Vieira, Y. Kuang, K.Astrom, A. F. Molisch, and F. Tufvesson,
“Tracking and positioning using phase information from estimated
multi-path components,” in 2015 IEEE International Conference on
Communication Workshop (ICCW), June 2015, pp. 712–717.
[9] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Prentice-Hall, Inc., 1993.
[10] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press,
2016.
[11] L. Liu et al., “The COST 2100 MIMO channel model,” IEEE Wireless
Communications, vol. 19, no. 6, pp. 92–99, December 2012.
[12] M. Zhu, G. Eriksson, and F. Tufvesson, “The COST 2100 Channel
Model: Parameterization and Validation Based on Outdoor MIMO
Measurements at 300 MHz,” IEEE Transactions on Wireless Communications, vol. 12, no. 2, pp. 888–897, February 2013.
[13] A. Molisch, Wireless Communications, ser. Wiley - IEEE. Wiley, 2010.
[14] T. Kaiser, H. B. Andre Bourdoux, J. B. A. Javier Rodriguez Fonollosa,
and W. Utschick, Smart Antennas - State of the Art. Hindawi Publishing
Corporation, 2005.
[15] X. Gao, F. Tufvesson, and O. Edfors, “Massive MIMO channels Measurements and models,” in 2013 Asilomar Conference on Signals,
Systems and Computers, Nov 2013, pp. 280–284.
[16] N. Guberman, “On complex valued convolutional neural networks,”
CoRR, vol. abs/1602.09046, 2016. [Online]. Available: http://arxiv.org/
abs/1602.09046
[17] Z. H. Wu, Y. Han, Y. Chen, and K. J. R. Liu, “A time-reversal paradigm
for indoor positioning system,” IEEE Transactions on Vehicular Technology, vol. 64, no. 4, pp. 1331–1339, April 2015.
| 7cs.IT
|
On the Ergodic Rate of OFDMA Cognitive Radios
under Imperfect Cross-Link Information
H. Saki, A. Shojaeifard, M. M. Mahyari, and M. Shikh-Bahaei
arXiv:1711.07371v1 [cs.IT] 17 Nov 2017
Institute of Telecommunications, King’s College London, London, WC2R 2LS, United Kingdom
Email: {hadi.saki,a.shojaeifard,m.sbahaei}@kcl.ac.uk
Abstract—The ergodic rate performance and limits of orthogonal frequency-division multiple access (OFDMA) cognitive
radios (CRs) is studied under imperfect cross-link knowledge. We
propose a novel stochastic interference management and exploitation technique to mitigate and control the imposed interference
by CRs on licensed users in underlay spectrum sharing. The
optimum downlink channel-adaptive resource allocation (RA)
algorithm is designed to maximize the CRs functionality subject
to satisfying average transmit and probabilistic peak interference
power constraints. An expression for the cumulative density
function (cdf) of CRs’ received signal-to-interference-plus-noise
ratio (SINR) is developed to evaluate the resultant ergodic rate.
Simulation studies are conducted to examine the proposed RA
algorithm and investigate the impact of parameters uncertainty
on the overall system performance.
I. I NTRODUCTION
Spectrum sharing has attracted a lot of attention recently
as measurements performed by regulators have indicated that
the allocated radio spectrum experiences low utilization [1].
In fact, according to the Office of Communications (Ofcom)
report [2], the spectrum is severely underutilized in both
spacial and temporal dimensions as a result of static management policies, and that 90% of locations have around 100
MHz of spectrum available for other services. Cognitive radio
(CR), defined as an intelligent system capable of observing
the surrounding environment and adapting accordingly [3],
is a promising solution to overcome the spectrum crunch in
modern wireless systems by exploiting spectral opportunities.
Three main paradigms have been proposed in regards to
CRs access to primary spectrum [4], [5]: (i) underlay approach
where CRs coexist with primary users (PUs), provided they
satisfy an interference limit set by a regulatory authority (ii)
overlay approach in which CRs access the vacant parts of the
licensed spectrum, and (iii) hybrid approach, a combination
of the two former strategies where CRs can dynamically
select between the underlay and overlay modes depending on
the traffic and interference characteristics. Here, we focus on
underlay spectrum access, where interference management and
exploitation is essential for achieving desirable performance
whilst tackling any harmful cross-service interference.
Orthogonal frequency-division multiple-access (OFDMA)
has emerged as a prominent air interface technology for
new generation of wireless communication systems including
long term evolution (LTE)-advanced [6, 7] and [8]. Further,
OFDMA is considered a de facto standard for CR networks
due to its inherent advantages in terms of flexibility and
adaptability in allocating radio resources in shared-spectrum
environments [9]. Resource allocation (RA) plays a significant role in improving the spectral efficiency of conventional
OFDMA systems [10, 11, 12]. In addition, RA is an active
area of research in OFDMA CR networks with the aim of
achieving a balance between maximizing the cognitive system
performance and minimizing the inflicted interference on the
licensed users.
Analysing the performance and limits of CR is essential
towards identifying the viable business models and initiating
standardization. A substantial amount of study on various
performance metrics of CR systems has recently emerged,
see, e.g., [13], [14], and the references herein. A central
assumption for controlling the cross-service interference in
many of the works in the literature is perfect knowledge of
the interfering links between the CRs and licensed receivers.
Imposing transmit and interference power constraints under the
perfect channel state information (CSI) assumption guarantees
a limited inflicted interference on the PUs. However, obtaining
cross-link information requires a cooperative signalling protocol between the primary and cognitive networks [15]. This
may be infeasible in practice, and even if a cross-connection
mechanism is implemented, the inherent nature of wireless
environment and estimation errors result in noisy CSI.
A number of works on CR have considered the impact of
imperfect cross-link knowledge: [13] studies the rate under
an average interference constraint, the capacity under peak
interference constraint is derived in [16], and optimum power
allocation policy and ergodic capacity are derived under two
different interference outage and signal-to-interference-plusnoise ratio (SINR) outage constraints in [17]. However, all of
these studies consider a single CR scenario, and further, [13]
and [16] adopt deterministic interference constraints.
The motivation of this work is to analyze the impact of
noisy cross-link knowledge on ergodic transmission rate of
multi-user multi-band OFDMA CR networks under average
transmit and probabilistic peak interference power constraints.
We propose a novel stochastic approach for mitigating the CSI
imperfections and design the optimum downlink RA algorithm
in the cognitive network. Further, in order to compute the
ergodic rate, we develop an approximated expression for the
cumulative density function (cdf) of CRs’ received SINR.
Notations: E {x} and P(x) denote the expectation and
probability of x, respectively. var(x) denotes the variance of
x and cov(y, z) is defined as the covariance of y and z. [x]+
signifies max(0, x), and x∗ is the optimal value of x.
II. C OGNITIVE S YSTEM M ODEL AND P RELIMINARIES
Consider a shared-spectrum environment as shown in Fig.
1, where a secondary network of N CRs (indexed by n)
coexist with a primary network over K non-overlapping bands
(indexed by k) subject to satisfying the power constraints set
by a regulatory authority. The bandwidth of each band is assumed to be much smaller than the coherence bandwidth of the
wireless channel, thus, each subcarrier experiences frequencyps
ss
flat fading. Let Hn,k
(t), Hn,k
(t), and Hksp (t), at time t, denote
the complex channel gains over band k from the cognitive
transmitter (CTx) to nth cognitive receiver (CRx), primary
transmitter (PTx) to nth CRx, and CTx to primary receiver
ss
(PRx), respectively. The channel power gains |Hn,k
(t)|2 ,
ps
sp
2
2
|Hn,k (t)| , and |Hk (t)| , are assumed to be ergodic and
stationary with continuous probability density functions (pdfs)
ps 2 (.), and f
ss |2 (.), f
f|Hn,k
|Hn,k
|
|Hksp |2 (.), respectively. In addition, the instantaneous values and distribution information of
secondary-secondary channel gains are assumed to be available at the CTx [17]. Due to the impact of several systematic
factors, such as channel estimation error, feedback delay, and
mobility, perfect cross-link information is not available. We
consider imperfect cross-link knowledge at the CTx obtained
from maximum likelihood (ML) estimation and model the
associated uncertainty in cross-link in the following form
Hksp (t) = Ĥksp (t) + ∆Hksp (t)
(1)
Fig. 1: Schematic diagram of the underlay spectrum sharing system.
For simplicity, the corresponding channels of a single CR are drawn.
sp
sp
2
δ∆H
sp
sp = var(∆Hk |Ĥk )
k |Ĥk
cov 2 (∆Hksp , Ĥksp )
2
2
sp 1 −
sp .
= δ∆H
= (1 − ρ2 )δ∆H
2
2
k
k
δ∆H
sp δ sp
Ĥ
k
k
Proposition 2: The posterior distribution of the true crosslink Hksp given the estimation Ĥksp is a complex Gaussian
random variable with respective mean and variance of
µH sp |Ĥ sp = E (Ĥksp + ∆Hksp |Ĥksp )
k
k
= E (Ĥksp |Ĥksp ) + E (∆Hksp |Ĥksp ) = (1 + ρ2 )Ĥksp
where over subcarrier (band) k, at time t,
is the
sp
true cross-link gain, Ĥk (t) is the estimated channel gain
considered to be known, and ∆Hksp (t) denotes the estimation
error. Hksp (t), Ĥksp (t), and ∆Hksp (t) are assumed to be zeromean complex Gaussian random variables with respective vari2
2
2
ances δH
sp , δ sp , and δ
∆H sp [16]. For robust receiver design,
Ĥ
k
k
k
we consider the estimation Ĥksp (t) and error ∆Hksp (t) to
be statistically correlated random variables with a correlation
factor ρ, which determines the accuracy of channel estimation
in relation to true channel states, defined as [18]
r
.
ρ=
2
δ∆H
sp
k
2
2
δ∆H
sp + δ sp .
H
k
(2)
k
We proceed by deriving posteriori distributions of the true
cross-link given the estimation and hence the estimation error
given the estimation. For notational brevity, we omit the time
reference t where the context is clear.
Proposition 1: The posterior distribution of the estimation
error ∆Hksp given the estimation Ĥksp is a complex Gaussian
random variable with respective mean and variance of
µ∆H sp |Ĥ sp = E (∆Hksp |Ĥksp ) = E (∆Hksp )+
k
k
cov(∆Hksp , Ĥksp ) sp
Ĥk − E (Ĥksp ) = ρ2 Ĥksp
2
2
δ∆H sp + δH sp
k
k
(3)
(5)
2
δH
= var(Ĥksp + ∆Hksp |Ĥksp ) = var(Ĥksp |Ĥksp )
sp
|Ĥ sp
k
k
+var(∆Hksp |Ĥksp ) + 2cov(∆Hksp |Ĥksp , Ĥksp |Ĥksp )
2
sp .
= (1 − ρ2 )δ∆H
k
Hksp (t)
(4)
(6)
In the OFDMA CR network, the downlink RA algorithm
allocates power and subcarrier to CRs based on the fadinginduced fluctuations in the true secondary-secondary and estimated interfering secondary-primary channels. Each subcarrier
is assigned exclusively to at most one CR at any given time,
hence, there is no mutual interference between different CRxs
[8]. It should also be noted that by utilizing an appropriate
cyclic prefix (CP), the inter-symbol-interference (ICI) can be
ignored [15]. The instantaneous received SINR of CR n over
band k with a fixed transmit power Pn,k is given by
γn,k =
ss 2
Pn,k |Hn,k
|
2
2
σn + σps
(7)
where σn2 is the power of circularly symmetric zero-mean
2
complex Gaussian noise and σps
is the received interference
power from the primary network taken as white Gaussian noise
2
[15], [19], Without loss of generality, σn2 and σps
are assumed
to be the same across all users and subcarriers [19], [20].
III. I NTERFERENCE M ANAGEMENT
AND
E XPLOITATION
In a spectrum sharing paradigm, and particularly for delaysensitive licensed services, the PUs’ quality of service (QoS)
is highly dependent on the instantaneous received interference power from the CRs. To protect the licensed spectrum
from harmful interference, extra constraints are needed to
facilitate interference control. However, due to the underlying
uncertainties about the shared-spectrum environment and PUs’
activity, it is unrealistic to assume that the CRs can always
satisfy deterministic interference constraints. In practice, the
tolerable interference level is confined by a maximum collision
probability that guarantees a certain grade of QoS for PUs.
In this paper, we consider an underlay spectrum sharing
system where the primary network tolerates a maximum
collision probability of ε. Collision is considered to occur
when the interference level imposed by CRs is higher than
I th . Adopting a probabilistic interference constraint is crucial
for robust interference management given noisy cross-link
knowledge. To improve overall system performance and to
mitigate the impact of channel estimation errors, the following
probabilistic peak interference constraint is considered
P
N X
K
X
ϕn,k (γn,k )Pn,k (γn,k )
Hksp
2
Ĥksp > I th
n=1 k=1
!
≤ε
(8)
where for cognitive user n over band k, ϕn,k (γn,k ) is the timesharing factor (subcarrier allocation policy), and Pn,k (γn,k )
is the allocated transmit power. Assuming equal variance
2
δH
across all users and subcarriers, the collision probasp
|Ĥksp
k
bility constraint in (8) can be expressed as
2
P δH
sp
|Ĥ sp
k
k
K
N X
X
n=1 k=1
!
ϕn,k (γn,k )Pn,k (γn,k )|Ξk |2 > I th ≤ ε
(9)
where |Ξk |2 is a non-central Chi-Square random variable
with two degrees of freedom and non-centrality parameter
2
µΞk = µH sp |Ĥ sp /δH sp |Ĥ sp . Note that (9) includes a sum of
k
k
k
k
non-equal-weighted Chi-Square random variables. In general,
obtaining the exact distribution of the linear combination
of weighted Chi-Square random variables is rather complex.
Although several approximations have been proposed, e.g.,
[21], they mostly rely on prior knowledge of the weights,
in this case prior knowledge of power and subcarrier assignments. Hence, conventional approaches cannot be applied to
the resource allocation problem under consideration. In this
work, we propose a novel
PN simple
PK approximation based on the
2
2
moments of δH
sp
sp
n=1
k=1 ϕn,k (γn,k )Pn,k (γn,k )|Ξk | .
PN k |Ĥ2 k
Let βk = n=1 δH sp |Ĥ sp ϕn,k (γn,k )Pn,k (γn,k ), hence
k
2
δH
sp
|Ĥ sp
k
k
N X
K
X
k
ϕn,k (γn,k )Pn,k (γn,k )|Ξk |2 =
n=1 k=1
K
X
k=1
βk |Ξk |2 .
(10)
Proposition 3: The distribution of the sum of nonequal-weighted
non-central Chi-Square random variables,
PK
2
non-central
k=1 βk |Ξk | , is similar to that of a weighted
′
′
Chi-Square-distributed random variable ξχ2D (δ ), where δ , D,
and ξ are respectively the non-centrality parameter, degrees of
freedom, and weight of the new random variable:
′
δ =
K
X
µΞ k
(11)
D = 2K
(12)
k=1
ξ=
PK
βk (2 + µΞk )
.
P
2K + K
k=1 µΞk
k=1
(13)
To investigate the above similarity, we compare the cdf of
the proposed Chi-Square distribution with that of (10), using
Monte-Carlo simulations. The results in Fig. 2 illustrate that
the approximation is accurate over a wide range of practical
values for K over randomly-distributed - e.g., Chi-Square or
Gamma - weights βk . Now (9) can be simplified to:
!
N X
K
X
2
ϕn,k (γn,k )Pn,k (γn,k )|Ξk |2 > I th
P δH
sp
|Ĥ sp
k
≈P
k
n=1 k=1
′
ξχ2D (δ )
> I th .
(14)
Given the non-centrality parameter is small relative to the
degrees of freedom, we can approximate the non-central ChiSquare distribution with a central one using the following [21]
′
I th /ξ
2
th
2
. (15)
P ξχD (δ ) > I
≈ P χD (0) >
1 + δ ′ /D
The right hand side of (15) can be formulated using the upper
Gamma function as
th
P χ2D (0) >
I th /ξ
1 + δ ′ /D
=
I /ξ
G(K, 2(1+δ
)
′
/D)
G(K)
(16)
where G(., .) is the upper incomplete Gamma function, and
G(.) is the complete Gamma function.
Proposition 4: For all integer values K 6= 1, and all positive
′
I th /ξ
- this condition is always true because, I th , δ βk , and
1+δ ′ /D
′
K are positive, consequently, ξ, δ , and D are also positive the deterministic inequality
2
δH
sp
|Ĥ sp
k
k
K
X
k=1
≤
(2 + µΞk )
N
X
ϕn,k (γn,k )Pn,k (γn,k )
n=1
th
KI
(K!)1/K ln 1 − (1 − ε)1/K
(17)
satisfies the probabilistic inequality (9). Therefore, the constraint (9) can be replaced by (17).
Additionally, mitigating the interference between neighbouring cells is a vital issue due to the increasing frequency reuse
aggressiveness in modern wireless communication systems
[22]. As a remedy to inter-cell interference, and to maintain
effective and efficient power consumption, we impose a total
average transmit power limit Pt on the cognitive network:
N X
K
X
n=1 k=1
n
o
Eγn,k ϕn,k (γn,k )Pn,k (γn,k ) ≤ Pt .
(18)
1
1
0.9
0.9
K = 64
K = 128
Cumulative Probability
Cumulative Probability
K = 32
0.8
0.7
0.6
0.5
0.4
0.3
0.8
K = 64
K = 32
K = 128
0.7
0.6
0.5
0.4
0.3
0.2
0.2
Approximated Model
Empirical Data
0.1
0
0
200
400
600
800
1000
1200
1400
1600
1800
2000
Approximated Model
Empirical Data
0.1
0
200
400
600
Data
800
1000
1200
Data
2
(a) βk ∼ Chi-Square(2, 2), δH
sp
sp = 1.
|Ĥ
k
k
2
(b) βk ∼ Gamma(2, 0.5, 4), δH
sp
sp = 0.5.
|Ĥ
k
k
Fig. 2: Approximated Model and Empirical Data cdfs, obtained from Monte-Carlo simulations.
IV. E RGODIC R ATE
The ergodic transmission rate (bps/Hz) of the OFDMA
CR network operating under interference and transmit power
constraints can be expressed as
!)
(
N X
K
X
γn,k Pn,k (γn,k )
ϕn,k (γn,k )
R=
Eγn,k log2 1+
t
min PKt , NIsp
n=1 k=1
(19)
P
K
2
where N sp = δH
=
sp
sp
k=1 (2 + µΞk ) and It
k |Ĥk
th
1/K
1/K
K I (K!)
ln 1 − (1 − ε)
. In order to evaluate the
ergodic rate, the distribution of CRs’ received SINR, dependent on secondary-secondary and secondary-primary channel
gains, must be developed. Limited by the constraints (17) and
(18), the cdf of γn,k can be written as
ss 2
ss 2
It |Hn,k
|
Pt |Hn,k
|
≤
Γ,
≤
Γ
.
Fγn,k (Γ) = P
2 )
2 )
K(σn2 + σps
N sp (σn2 + σps
(20)
The expression in (20) can be simplified by considering the
ss 2
ss 2
Pt |Hn,k
|
It |Hn,k
|
sp
cases K(σ2 +σ
:
2 ) S N sp (σ 2 +σ 2 ) and conditioning on N
n
ps
n
ps
ss 2
It |Hn,k
|
1−P
> Γ, sp 2
>
Γ
=
2 )
N (σn + σps
2
KΓ(σn2 + σps
)
It K
ss 2
N sp ≤
P |Hn,k | >
Pt
Pt
1−
sp
2
2
N Γ(σn + σps )
It K
ss 2
N sp >
.
| >
P |Hn,k
It
Pt
(21)
ss 2
Pt |Hn,k
|
2
2 )
K(σn + σps
Invoking central limit theorem for large values of K, N sp
can be approximated as a Gaussian
random variable
with
2 2
2
2 2
mean µN sp = 2KδĤ
+ 2K 1 − ρ2 δ∆H
sp and
sp 1 + ρ
k
k
2
4
2 4
ss 2
variance δN
.
Suppose
that
|H
sp = 4Kδ sp 1 + ρ
n,k |
Ĥk
ss |2 , hence,
follows an Exponential distribution with mean µ|Hn,k
with further manipulation, a closed-form expression for cdf
of γn,k is developed in (22). Trivially, through respective
differentiation of (22), the pdf of γn,k can be obtained.
V. R ADIO R ESOURCE A LLOCATION A LGORITHM
In this section, we design the optimal downlink power and
subcarrier allocation algorithm to maximize the ergodic rate
of the OFDMA CR network subject to satisfying the transmit
and interference power constraints under noisy cross-link CSI.
The optimization problem can be formulated as
max
ϕn,k (γn,k ),Pn,k (γn,k )
R
(23a)
s. t.: constraints in (17), (18)
N
X
n=1
ϕn,k (γn,k ) = 1, ∀k ∈ {1, ..., K}
(23b)
ϕn,k (γn,k ) ∈ {0, 1}, ∀n ∈ {1, ..., N }, ∀k ∈ {1, ..., K} (23c)
where constraints (23b) and (23c) ensure that every subcarrier
is allocated to at most one CR. It can be observed that the
optimization problem is convex with respect to Pn,k (γn,k ),
however, it is non-convex with respect to ϕn,k (γn,k ). By
applying Lagrangian dual decomposition, the non-convex optimization problem can be decomposed into independent subproblems each corresponding to a given CR. By solving
the Lagrangian optimization problem the following potential
optimal power allocation solution can be obtained for user n
over subcarrier k with Lagrangian multipliers µ and η
∗
Pn,k
(γn,k ) =
"
1
−
2
ln(2) µ+ηδH
sp
sp (2 + µΞk )
|Ĥ
k
min
Pt
It
K , N sp
γn,k
#+
. (24)
k
The solution in (24) can be considered as a multi-level waterfilling algorithm. Using (24) and by applying the KarushKuhn-Tucker (KKT) conditions, the optimal subcarrier allocation problem is formulated as
n∗ = arg max Λ(γn,k ) , ∀n ∈ {1, ..., N } , ∀k ∈ {1, ..., K}
(25)
Fγn,k (Γ) ≈
!
!"
It K
#
2
2
2
2
2
ss |2 It + δ sp Γ(σ
Γ(σn2 + σps
) −2µN sp µ|Hn,k
+ σps
)
sp
−
µ
n
−KΓ(σn2 + σps
)
N
N
1
1
t
1 + erf Pp
− exp
1 − exp
2
ss |2
2
Pt µ|Hn,k
2
2µ2H ss It 2
2δN
sp
n,k
"
µ ss 2 I −µ sp + It K + δ 2 sp Γ(σ 2 + σ 2 ) #
N
|Hn,k | t
n
ps
N
Pt
√
.
(22)
1 − erf
ss
sp
2µ|Hn,k |2 It δN
where n∗ is the optimal CR index, and
∗
γn,k Pn,k
(γn,k )
It
P
min( Kt , N sp
)
+
∗
γn,k Pn,k (γn,k )
ln(2) 1 +
It
P
min( Kt , N sp
)
Λ(γn,k ) =
ln 1 +
∗
γn,k Pn,k
(γn,k )
P
I
t
min( Kt , N sp
)
ln(2)
.
(26)
The optimal subcarrier allocation policy is achieved by assigning the k th subcarrier to the CR with the highest value
of Λ(γn,k ), i.e., placing a Lagrangian multiplier λk between
the first and second maximas of Λ(γn,k ). If there are multiple
equal maximas, the time-slot can be identically shared among
the respective users. Here, we use the subgradient-based
method to update the values of the multipliers µ and η:
N X
K
X
∗
ϕ∗n,k (γn,k )Pn,k
(γn,k ) (27)
µi+1 = µi − τ1i Pt −
n=1 k=1
η i+1 (γn,k ) = η i (γn,k ) − τ2i ×
It −
2
δH
sp
sp
k |Ĥk
K
X
(2+µΞk )
k=1
N
X
∗
ϕ
!
∗
(γn,k )
(γn,k )Pn,k
n=1
(28)
where for the iteration number i, τ1i and τ2i are the step sizes.
The initial values of dual multipliers and step size selection
are important towards obtaining the optimal solution, and can
greatly affect the optimization problem convergence.
VI. P ERFORMANCE E VALUATION
AND
D ISCUSSION
The performance of the OFDMA CR network using the
proposed RA algorithm, subject to satisfying average transmit and probabilistic peak interference power limits under
imperfect cross-channel estimation, is studied. Perfect CSI is
assumed between CTx and CRxs through an error-free feedss
back channel. |Hn,k
|, ∀{n, k}, are assumed to be Rayleighdistributed and the secondary-secondary power gain mean valss |2 , ∀{n, k}, are taken as Uniformly-distributed ranues, µ|Hn,k
dom variables within 0 to 2. Interfering cross-channel values,
Hksp , ∀{k}, are distributed according to a complex Gaussian
distribution with mean 0.05 and variance 0.1. The channel
estimation and error for all sub-channels are taken as independent and identically distributed (i.i.d.) zero-mean Normallydistributed random variables. The noise and primary-secondary
interference power spectral densities are set to N0 = −110
dBm and 2500N0, respectively. All results correspond to a
CR network with 64 subcarriers and 3 CRxs.
The approximated pdfs for the received SINR of the three
CRs, dependent on noise floor and true secondary-secondary
and estimated secondary-primary channel gains, in a randomly
taken subcarrier, i.e., here k = 55, is plotted in Fig. 3. The
results are obtained using Monte-Carlo simulations.
Fig. 4 shows the ergodic rate of the OFDMA CRs versus
the collision threshold. As expected, a higher collision limit
results in improved R as I th limits the CRs’ transmit power
functionality. The gain in ergodic rate, however, approaches a
plateau in high I th region as Pt becomes the dominant power
constraint. Given Pt is dominant, imposing a higher average
power setting enhances the CRs’ uppermost performance, e.g.,
where I th = 10 Watts, a 7.50% increase in ergodic rate is
realized as Pt is increased from 20 to 25 Watts.
The ergodic rate against the probabilistic interference constraint for different collision thresholds is illustrated in Fig. 5.
Increasing the maximum probability of violating the collision
threshold improves R, the tradeoff, however, is the greater
possibility of degrading the primary service operation which
is deemed undesirable in practical scenarios. For more stringent collision limits, i.e., small I th , increasing the maximum
collision probability provides higher gains in ergodic rate. E.g.,
varying ε from 0% to 5%, increases R by 26.2% and 15.4%
with I th = 5 and I th = 7, respectively; the gain diminishes
quickly however as increasing ε in high I th region implies a
faster rate towards reaching a saturated performance limit.
System performance versus the correlation coefficient between the estimation and error random variables is depicted
in Fig. 6. Higher ρ correspond to greater accuracy of the
estimation technique, with the ideal case of ρ = 1 where error
variance is zero. Precise cross-link estimation is essential towards robust interference control, however, it typically implies
more training symbols and thus increased signalling overhead.
VII. C ONCLUSIONS
The ergodic transmission rate performance and limits of
multi-user multi-band OFDMA CRs with underlay spectrum
settings under imperfect cross-link knowledge was studied.
We designed the downlink RA algorithm to maximize the
cognitive network performance subject to satisfying average
transmit and probabilistic peak interference power constraints.
To compute the quantified loss in ergodic rate due to the added
constraints, an approximated expression of the CRs’ received
0.4
270
CRx #1
CRx #2
CRx #3
260
0.3
Ergodic Rate, R (bps/Hz)
Probability Density Function
0.35
0.25
0.2
0.15
0.1
250
240
230
220
th
I = 11 Watts
210
th
I = 9 Watts
0.05
th
I = 7 Watts
200
th
I = 5 Watts
0
5
10
15
Received SINR, γ
20
n,k
190
25
0
2
(dB)
Fig. 3: Probability density functions of the received SINR for
OFDMA CRs in a given subcarrier k. System parameters are: k = 55,
Pt = 25 Watts, I th = 5 Watts, ρ = 0.25, ε = 30%.
4
6
8
Maximum Probability of Collision, ε (%)
10
Fig. 5: Ergodic rate against the probabilistic interference constraint
with different values of collision threshold. System parameters are:
Pt = 40 Watts, ρ = 0.75.
242
260
250
240
Ergodic Rate, R (bps/Hz)
Ergodic Rate, R (bps/Hz)
240
230
220
210
200
190
180
236
234
232
Pt = 35 Watts
170
Pt = 30 Watts
160
150
238
ε = 25%
ε = 22.5%
ε = 20%
230
Pt = 25 Watts
228
2
4
6
8
10
12
14
th
Collision Threshold, I (Watts)
0
0.2
0.4
0.6
Correlation Coefficient, ρ
0.8
1
Fig. 4: Ergodic rate versus the collision threshold with different values
of average transmit power limit. System parameters are: ρ = 0.5,
ε = 15%.
Fig. 6: Ergodic rate against the correlation factor between estimation
and error random variables for different probabilistic interference
constraints. System parameters are: Pt = 25 Watts, I th = 20 Watts.
SINR was developed. As a remedy to the imperfections associated with CSI and varying shared-spectrum environment, a
stochastic interference management and exploitation technique
was employed to confine the probability of collision to a predefined limit. The impact of parameters uncertainty on overall
system performance was investigated through simulation results. Our proposed framework is an improvement in terms of
practical feasibility and flexibility over the existing literature
with perfect cross-link availability assumption, deterministic
interference management policies, and single-user scenarios.
[5] S. Boyd, J. Frye, M. Pursley, and T. Royster, “Spectrum monitoring during reception
in dynamic spectrum access cognitive radio networks,” IEEE Trans. Commun.,
vol. 60, no. 2, pp. 547–558, Feb. 2012.
[6] I. Koffman and V. Roman, “Broadband wireless access solutions based on ofdm
access in ieee 802.16,” IEEE Commun. Mag., vol. 40, no. 4, pp. 96–103, Apr. 2002.
[7] K. Nehra, A. Shadmand, and M. Shikh-Bahaei, “Cross-layer design for
interference-limited spectrum sharing systems,” in 2010 IEEE Global Telecommunications Conference GLOBECOM 2010, Dec 2010, pp. 1–5.
[8] G. Li and H. Liu, “Downlink radio resource allocation for multi-cell ofdma system,”
IEEE Trans. Wireless Commun., vol. 5, no. 12, pp. 3451–3459, Dec. 2006.
[9] X. Zhou, G. Li, D. Li, D. Wang, and A. Soong, “Probabilistic resource allocation
for opportunistic spectrum access,” IEEE Trans. Wireless Commun., vol. 9, no. 9,
pp. 2870–2879, Sep. 2010.
[10] A. Kobravi and M. Shikh-Bahaei, “Cross-layer adaptive arq and modulation
tradeoffs,” in 2007 IEEE 18th International Symposium on Personal, Indoor and
Mobile Radio Communications, Sept 2007, pp. 1–5.
[11] H. Saki and M. Shikh-Bahaei, “Cross-layer resource allocation for
video streaming over OFDMA cognitive radio networks,” IEEE Trans.
Multimedia, vol. 17, no. 3, pp. 333–345, 2015. [Online]. Available:
https://doi.org/10.1109/TMM.2015.2389032
[12] A. Olfat and M. Shikh-Bahaei, “Optimum power and rate adaptation for mqam
in rayleigh flat fading with imperfect channel estimation,” IEEE Transactions on
Vehicular Technology, vol. 57, no. 4, pp. 2622–2627, July 2008.
[13] L. Musavian and S. Aissa, “Fundamental capacity limits of cognitive radio in fading
environments with imperfect channel information,” IEEE Trans. Commun., vol. 57,
no. 11, pp. 3472–3480, Nov. 2009.
[14] X. Kang, Y.-C. Liang, A. Nallanathan, H. Garg, and R. Zhang, “Optimal power
allocation for fading channels in cognitive radio networks: Ergodic capacity and
outage capacity,” IEEE Trans. Wireless Commun., vol. 8, no. 2, pp. 940–950, Feb.
R EFERENCES
[1] A. MacKenzie, J. Reed, P. Athanas, C. Bostian, R. M. Buehrer, L. DaSilva,
S. Ellingson, Y. Hou, M. Hsiao, J.-M. Park, C. Patterson, S. Raman, and C. da Silva,
“Cognitive radio and networking research at virginia tech,” Proc. IEEE, vol. 97,
no. 4, pp. 660–688, Apr. 2009.
[2] OfCom, “OfCom report. Digital dividend: cognitive access.” Jul. 2009.
[3] J. Mitola and J. Maguire, G.Q., “Cognitive radio: making software radios more
personal,” IEEE Pers. Commun., vol. 6, no. 4, pp. 13–18, May 1999.
[4] I. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “A survey on spectrum
management in cognitive radio networks,” IEEE Commun. Mag., vol. 46, no. 4,
pp. 40–48, Apr. 2008.
2009.
[15] N. Mokari, K. Navaie, and M. Khoshkholgh, “Downlink radio resource allocation
in ofdma spectrum sharing environment with partial channel state information,”
IEEE Trans. Wireless Commun., vol. 10, no. 10, pp. 3482–3495, Oct. 2011.
[16] H. Suraweera, P. Smith, and M. Shafi, “Capacity limits and performance analysis
of cognitive radio with imperfect channel knowledge,” IEEE Trans. Veh. Technol.,
vol. 59, no. 4, pp. 1811–1822, May 2010.
[17] Z. Rezki and M.-S. Alouini, “Ergodic capacity of cognitive radio under imperfect
channel-state information,” IEEE Trans. Veh. Technol., vol. 61, no. 5, pp. 2108–
2119, Jun. 2012.
[18] S.-S. Sadough and P. Duhamel, “Improved iterative detection and achieved throughputs of ofdm systems under imperfect channel estimation,” IEEE Trans. Wireless
Commun., vol. 7, no. 12, pp. 5039–5050, Dec. 2008.
[19] R. Zhang, S. Cui, and Y.-C. Liang, “On ergodic sum capacity of fading cognitive
multiple-access and broadcast channels,” IEEE Trans. Inf. Theory, vol. 55, no. 11,
pp. 5161–5178, Nov. 2009.
[20] X. Gong, S. Vorobyov, and C. Tellambura, “Optimal bandwidth and power
allocation for sum ergodic capacity under fading channels in cognitive radio
networks,” IEEE Trans. Signal Process., vol. 59, no. 4, pp. 1814–1826, Apr. 2011.
[21] D. R. Cox and N. Reid, “Approximations to noncentral distributions,” The Canadian
Journal of Statistics, vol. 15, no. 2, pp. 105–114, 1987.
[22] H. Zhang, L. Venturino, N. Prasad, P. Li, S. Rangarajan, and X. Wang, “Weighted
sum-rate maximization in multi-cell networks via coordinated scheduling and
discrete power control,” IEEE J. Sel. Areas Commun., vol. 29, no. 6, pp. 1214–1224,
Jun. 2011.
| 7cs.IT
|
ULB-TH/08-10
AEI-2008-013
arXiv:0805.3018v2 [math.RT] 14 Jul 2017
Hyperbolic Weyl groups and
the four normed division algebras
Alex J. Feingold 1 , Axel Kleinschmidt 2 and Hermann Nicolai 3
1
Department of Mathematical Sciences, The State University of New York
Binghamton, New York 13902–6000, U.S.A.
2
Physique Théorique et Mathématique,
Université Libre de Bruxelles & International Solvay Institutes,
Boulevard du Triomphe, ULB – CP 231, B-1050 Bruxelles, Belgium
3
Max-Planck-Institut für Gravitationsphysik, Albert-Einstein-Institut
Am Mühlenberg 1, D-14476 Potsdam, Germany
Abstract: We study the Weyl groups of hyperbolic Kac–
Moody algebras of ‘over-extended’ type and ranks 3, 4, 6 and
10, which are intimately linked with the four normed division algebras K = R, C, H, O, respectively. A crucial role is played by
integral lattices of the division algebras and associated discrete
matrix groups. Our findings can be summarized by saying that
the even subgroups, W + , of the Kac–Moody Weyl groups, W , are
isomorphic to generalized modular groups over K for the simply
laced algebras, and to certain finite extensions thereof for the nonsimply laced algebras. This hints at an extended theory of modular
forms and functions.
“The mathematical universe is inhabited not only by important species but
also by interesting individuals.” – C. L. Siegel
1
Introduction
In [20] Feingold and Frenkel gained significant new insight into the structure
of a particularly interesting rank 3 hyperbolic Kac–Moody algebra which they
called F , along with some connections to the theory of Siegel modular forms
of genus 2. The first vital step in their work was the discovery that the Weyl
group of that hyperbolic algebra is W (F ) ∼
= P GL2 (Z), the projective group
of (2 ×2) integral matrices with determinant ±1, isomorphic to the hyperbolic
triangle group T (2, 3, ∞). They showed that the root system of that algebra
could be realized as the set of (2 × 2) symmetric integral matrices X with
det(X) ≥ −1, and that the action of M ∈ W (F ) on X is given by MXM T . In
notation used commonly by physicists today, the hyperbolic algebra studied
in [20] is designated as A++
since it is obtained from the finite-dimensional
1
Lie algebra sl2 of type A1 by a process of double extension. The first step of
(1)
the extension gives the affine algebra A1 ≡ A+
1 , and the second step, often
referred to as over-extension, gives the hyperbolic algebra. In the realization of
the root lattice via symmetric matrices X, the real roots consist of the integral
points X with det(X) = −1 on a single-sheeted hyperboloid, and the imaginary
roots consist of the integral points X on the light-cone det(X) = 0 and on
the two-sheeted hyperboloids det(X) > 0. In [20] it was also mentioned that
these results could be extended to two other (dual) rank 4 hyperbolic algebras
whose Weyl groups were both the Klein–Fricke group Ψ∗1 containing as an
index 4 subgroup the Picard group Ψ1 = P SL2 (Z(i)) whose entries are from
the Gaussian integers. The results expected to hold for these rank 4 hyperbolic
algebras included connections to the theory of Hermitian modular forms, but
that line of research was not pursued in later work. A paper by Kac, Moody
and Wakimoto [26] generalized the structural results of [20] to the hyperbolic
algebra E10 = E8++ , but until now there has not been any new insight into
the structure of the Weyl groups of hyperbolic algebras which are usually just
given as Coxeter groups defined by generators and relations.
In the present work, we take up this line of development again, and present a
coherent picture for many higher rank hyperbolic Kac–Moody algebras which
is based on the relation to generalized modular groups associated with lattices
and subrings of the four normed division algebras. More specifically, we shall
show that the Weyl groups of all hyperbolic algebras of ranks 4, 6 and 10 which
can be obtained by the process of double extension described above, admit realizations in terms of generalized modular groups over the complex numbers
C, the quaternions H, and the octonions O, respectively. We are encouraged
to find that these Weyl groups are amenable to explicit matrix descriptions,
but understand that the hyperbolic algebras themselves have still eluded any
effective characterization. For K = C, the present work on these hyperbolic
Weyl groups is thus a very natural extension of [20], showing in particular that
++
the rank 4 hyperbolics A++
and G++
are naturally connected with cer2 , C2
2
1
tain subrings in the normed division algebra C. Analogous results are obtained
++
++
++
for all the ‘over-extended’ rank 6 hyperbolics A++
and F4++ ,
4 , B4 , C4 , D4
whose (even) Weyl groups can be described in terms of quaternionic modular
groups. Finally, K = O, the largest and non-associative division algebra of
octonions, is associated with rank 10 hyperbolics, and in particular, the maximally extended hyperbolic Kac–Moody algebra E10 . The other two hyperbolic
over-extended algebras B8++ and D8++ can also be described using octonions.
In this paper we present the rich structure which we found in the complex and
quaternionic cases, as well as partial (and intriguing) results for the octonionic
case. As explained in more detail in Section 3, a new feature for the division
algebras beyond R is that simple reflections involve complex conjugation of
all entries of X. For that reason only the even part of a given Weyl group will
act by matrix conjugation of X and is therefore the main focus of our study.
Throughout the paper we will denote by W + the even part of a Weyl group
W . In the case of F one has W + (F ) ∼
= P SL2 (Z), the modular group, and for
the other division algebras we discover a number of apparently new modular
groups. An announcement of our work appears in [21], based on a talk presented by A.J.F. at the conference on “Vertex Operator Algebras and Related
Areas” in honor of Geoffrey Mason, July 2008, at Illinois State University.
In table 1 we summarize our findings for the (even) Weyl groups of the finite
and hyperbolic Kac–Moody algebras studied in relation to the various division
algebras. By g++ we mean the over-extension of the finite-dimensional Lie
algebra g to a hyperbolic Kac–Moody algebra, where the first extension g+
is an untwisted affine algebra. In Appendix A we discuss other cases where
the first extension is twisted affine, so a different notation is required. In the
table, we use the standard group theory notation C = A . B to mean a group
C which contains group A as a normal subgroup, with quotient group C/A
isomorphic to B. Such a group C is called an extension of A by B. If B
has order |B|, A is said to be of index |B| in A . B. It can happen that the
extension is a semi-direct product, so that B is a subgroup of C which acts on
A via conjugation as automorphisms, and in this case the product is denoted
by A ⋊ B. By Sn we denote the symmetric group on n letters.
The various rings appearing in the table are as follows. For K = C these
are the Gaussian integers G ≡ Z(i) and the Eisenstein integers E; for the
quaternions H we use the terminology of [8], referring to the maximal order
of quaternions having all coefficients in Z or all in Z+ 12 as Hurwitz integers H.
To understand W + (A++
4 ) we will need the icosian quaternions I [7,35]. The
octonionic integers are called the octavians O [8]. The generalized modular
groups that we find are all discrete as matrix groups.
Our results are complete for the cases K = C and K = H. For all of these
we can reformulate the even Weyl groups as new kinds of modular groups,
most of which have not yet appeared in the literature so far. We note that
2
W + (g++ )
K
Root system g
W (g)
R
A1
2 ≡ Z2
P SL2 (Z)
C
A2
Z3 ⋊ 2
P SL2 (E)
C
B2 ≡ C2
Z4 ⋊ 2
P SL2 (G) ⋊ 2
C
G2
Z6 ⋊ 2
P SL2 (E) ⋊ 2
H
A4
S5
P SL2 (I)
H
B4
24 ⋊ S4
P SL2 (H) ⋊ 2
H
C4
24 ⋊ S4
]
P
SL2 (H) ⋊ 2
H
D4
23 ⋊ S4
P SL2 (H)
H
F4
25 ⋊ (S3 × S3 )
P SL2 (H) ⋊ 2
O
D8
27 ⋊ S8
P SL2 (O)
O
B8
28 ⋊ S8
P SL2 (O) ⋊ 2
O
E8
2 . O8+ (2) . 2
P SL2 (O)
(0)
(0)
(0)
(0)
(0)
(0)
Table 1
Root systems, number systems and Weyl groups within the normed division algebras.
The groups in the right column are defined in detail in the relevant sections.
A. Krieg [30,31] has developed a theory of modular forms over the quaternions,
however the modular groups defined there do not appear to coincide with the
ones found here. Nevertheless, we will make extensive use of some results of
[30] in our analysis, in particular Theorem 2.2 on p. 16.
The most interesting Weyl group, however, is the one of the maximally extended hyperbolic Kac–Moody algebra E10 ≡ E8++ , where our results are still
incomplete, and which provided the chief motivation for the present work. In
this case, the even Weyl group is given by
W + (E10 ) ∼
= P SL2 (O) .
(1.1)
Because the (2 × 2) matrices over the octonions obviously do not form a
group, the ‘modular group’ over the octavians O appearing on the r.h.s. of this
equation can so far only be defined recursively, by nested sequences of matrix
conjugations with the generating (octonionic) matrices introduced in Section 3
of this paper. This is in analogy with the generators and relations description
of the E10 Weyl group as products of fundamental reflections (and also in
analogy with the description of the continuous Lorentz group SO(1, 9; R) via
octonionic (2 × 2) matrices in [41,34]). It remains an outstanding problem to
find a more manageable realization of this modular group directly in terms of
(2×2) matrices with octavian entries and the G2 (2) automorphism group of the
octavian integers. In the final section we present some results in this direction
3
which we believe are new, and which may be of use in future investigations.
The case E10 is also the most important because it has recently been shown
that all simply laced hyperbolic Kac–Moody algebras can be embedded into
E10 [42]. Because the associated Weyl groups are then subgroups of W (E10 ),
all these Weyl subgroups should admit an octonionic realization. Conversely,
the structure and explicit realizations of modular groups found here for the
quaternionic case should help in understanding the group P SL2 (O), because
all rank 6 algebras occur as subalgebras inside E10 (and A++
and D4++ as
4
regular subalgebras, in particular). The remarks above also apply to the Weyl
groups of B8++ and D8++ .
Ref. [20] also highlighted a possible link with the theory of Siegel modular
forms. In the present context, one would thus start from the Jordan algebra
H2 (KC ) over the complexified division algebra KC , and consider a generalized
Siegel upper half-plane for Z ∈ H2 (KC ). This can be done not only for the
quaternions [31], but also for K = O, where the outlines of a corresponding
theory of ‘Siegel modular forms’ have been described and developed in [16].
However, it remains to be seen what role P SL2 (O) as defined here has to play
in this context, and whether results in this direction can substantially advance
our understanding of the hyperbolic algebra E10 .
We believe that our results are also interesting in the light of recent developments in the study of black hole microstates for particular classes of black
holes. It was found already in [14] that the degeneracy formula of dyonic quarter BPS black holes of N = 4 is controlled by the denominator formula of a
specific Borcherds completion of a hyperbolic subalgebra of F , whose Weyl
group is index 6 in W (F ). In more recent work [5] it was argued that the Weyl
group has a more immediate interpretation in this setup as realizing the socalled attractor flow of a given dyonic solution to an ‘immortal dyon’ by crossing walls of marginal stability in moduli space [3,40,10]. The correspondence to
the Weyl group is such that a point X inside the forward light-cone corresponds
to dyonic solutions with the given moduli and charges. As the light-cone is
tessellated by the action of the Weyl group one can move such a point to a
standard fundamental domain by a finite number of Weyl reflections and the
endpoint of this motion is the immortal dyon. The walls between the different
cells crossed during the motion in moduli space are the walls of marginal stability. It would be interesting to study the degeneracy and the attractor flow of
less supersymmetric solutions and their relation to the higher rank hyperbolic
algebras and modular groups discussed in this work. The possible relevance of
modular forms is also mentioned in recent work on E10 unification and quantum gravity [4,12]. Weyl groups of the so-called hidden symmetry groups also
appear in the context of U-duality groups [23,17,36,19]. 1
1
We stress, however, that the Weyl groups discussed here must not be confused
with the arithmetic duality groups conjectured to be symmetries of (compactified)
4
The organization of the paper is as follows. In Section 2.1 we fix the common notation for the four division algebras and describe the general structure
of the hyperbolic root systems. Section 3 contains the central general results
about the realization of the simple reflections, the even hyperbolic Weyl group
and the affine and finite Weyl groups for all division algebras. The commutative division algebras R and C and the associated hyperbolic Weyl groups
are treated in Section 4, the algebras admitting a quaternionic realization in
Section 5 and the case E8++ which requires the octonions in Section 6. In appendix A we also present two cases involving complex numbers which are not
of over-extended type but involve a twisted affine extension.
Acknowledgements: We would like to thank M.C.N. Cheng, F. Englert,
H.R.P. Ferguson, R.L. Griess, M. Mazur, W. Nahm, P. Ramond and D. Zagier
for discussions and correspondence. A.J.F. and A.K. gratefully acknowledge
the hospitality of the Albert Einstein Institute on various visits, while H.N.
thanks U.L.B. Brussels for hospitality. A.K. is a Research Associate of the
Fonds de la Recherche Scientifique–FNRS, Belgium.
2
Division algebras and hyperbolic root lattices
We wish to describe the Weyl groups of the various hyperbolic Kac–Moody
algebras as matrix groups which are to be interpreted as modular groups. In
order to accomplish this, we will describe in this section a lattice (integral
linear combinations of explicit elements) in the Lorentzian space of Hermitian
(2 × 2) matrices over one of the normed division algebras K. This lattice will
be shown to be isometric to the root lattice of the hyperbolic algebra. In the
following section we will show how the Weyl group acts as on this lattice by
matrix conjugation, leading to interesting matrix groups in the associative
cases.
2.1 Subrings and integers in division algebras
To begin we review some facts about division algebras, their subrings and
orders in certain algebraic extensions of the rational division algebras.
M theory. For instance, W (E10 ) can be realized as an arithmetic subgroup of O(1, 9),
whereas the hypothetical arithmetic duality group E10 (Z) is infinitely larger. Independently of its possible physical significance, a proper definition of this object at
the very least would presuppose properly understanding the continuous group E10 ,
a goal still beyond reach.
5
Let K be one of the four normed division algebras (over R): R (the real numbers), C (the complex numbers), H (the Hamilton quaternions) or O (the
Cayley octonions). K will also be a topological metric space with respect to
the norm topology. Each algebra K has an involution, sending a ∈ K to ā,
generalizing the complex conjugation in C, such that the norm of a ∈ K,
N(a) = aā, is a non-negative real number, and satisfies the composition law
N(ab) = N(a)N(b) for all a, b ∈ K. For a ∈ R, we have a = ā, and for
a ∈ C, ā is the usual complex conjugation in C. One also defines the real part
a + ā = 2Re(a) ∈ R (also sometimes called trace), and the imaginary part of
a is defined by a − ā = 2Im(a). The real-valued symmetric bilinear form on
K
(a, b) = N(a + b) − N(a) − N(b) = ab̄ + bā
(2.1)
is positive definite, giving K the structure of a real Euclidean space.
For each choice of K there is a standard basis such that the structure constants
are all in {−1, 0, 1}, so if one takes only rational linear combinations of these
basis vectors, one gets four rational normed division algebras KQ (over Q).
If F is any subfield of R one has the normed division algebra KF (over F),
consisting of all F-linear combinations of the standard basis elements of K,
such that K = KF ⊗F√R. We √
will actually√only be using a few specific choices
for F, namely, Q, Q( 2), Q( 3) and Q( 5).
The consideration of KF is necessary to describe the various finite root systems which can have angles with cosines involving the listed square roots.
The matrices acting on the root lattices will have entries from certain orders
O within KF . We briefly review this concept from algebraic number theory
(see e.g. [38,8]). The subfield F of R, assumed to be a finite extension of Q,
determines
a√
ring S of algebraic
in F, which for our cases are simply
√
√ integers
2
Z, Z( 2), Z( 3) and Z((1 + 5)/2). Note that, as a subring of a field, S
is always commutative. An order O in KF over S is a subring of KF (nonassociative for K = O) containing 1 and which is finitely generated as an
S-module (under addition). Furthermore it has to contain a basis of KF over
F, such that O ⊗F F = KF . An order is called a maximal order if it is an order
not properly contained in another order. Orders in rings are a generalization of
the integers in the rationals, and occupy an important role in number theory.
We stress that it is important for an order O to be finitely generated and that
this implies that elements x ∈ O satisfy a monic polynomial equation
√
These examples follow the general pattern of algebraic integers in Q( D) which
are those elements
that satisfy
a monic minimal polynomial with coefficients in Z,
√
√
that is, Z( D) or Z((1 + D)/2), depending on D(mod 4).
2
6
xn + cn−1 xn−1 + · · · + c0 = 0
with cn−1 , . . . , c0 ∈ Z ,
(2.2)
which also can be used for defining orders for the non-associative octonions
[8]. The polynomial (2.2) arises as follows. Let B = {x1 , · · · , xm } be a finite
set of generators for O over S, so that
O=
(
m
X
si xi
i=1
)
si ∈ S
.
(2.3)
Left multiplication by x ∈ O, Lx , can be represented by a matrix [lij ] such
that
xxj =
m
X
i=1
lij xi
and all lij ∈ S
(2.4)
since O is a subring and an S-module. The characteristic polynomial det(λI −
Lx ) of the linear transformation Lx is a monic polynomial of degree m with
coefficients in the commutative ring S and which is satisfied by Lx . Since K
is a division algebra, the polynomial is also satisfied by x, and so x satisfies a
monic polynomial equation with coefficients
in S. Multiplying this polynomial
√
√
by the conjugate polynomial, where D has been replaced by − D in the
coefficients (where D = 2, 3, 5), will give a polynomial of degree 2m for x with
coefficients in Z as claimed. 3 It can happen that the minimal polynomial
satisfied by x is of lower degree but for our purposes it is sufficient to know
that some such polynomial exists. Since in all cases we consider in this paper
there is finite group of units EO we know that the ring spanned by them under
addition and multiplication will form an order in the corresponding KF , which
we interpret as integers.
Although one usually thinks of integers as being ‘discrete’ in some topology,
note that orders need
For example,
√ not be discrete in the norm topology.
√
the order O = Z( 2) = S is a dense subring of F = Q( 2) where K = R
(and thus also dense in R). Although we will use in some cases such dense
subrings (cf. in particular Section 5.2), the Weyl groups we define will be
discrete groups (in the sense that there exists a fixed ǫ > 0 such that the
ǫ-neighborhoods of different matrices do not intersect). As we will explain in
Section 3.4, this discreteness can be traced back to the discreteness of the
hyperbolic root lattices. The discreteness is also important with regard to the
3
This can also be stated as follows: Since S is the ring of algebraic integers over F
the norm from F to Q sends S to Z. Applying the norm to the equation satisfied by
x gives another equation satisfied by x, but with coefficients in Z.
7
properly discontinuous action of the matrix group on a suitable ‘upper halfplane’ and the existence of fundamental domains with non-empty interior. 4
For the division algebra K = R we take F = Q, S = Z and O = Z. For
K = C we use different choices of F and order O. Either F = Q, S = Z,
such that KF = Q(i) is the ‘standard’ rational form of C, and the order
is
√
the ring of Gaussian integers O = G ≡ Z(i); or we choose F = Q( 3) so
that KF = F(i) is an F-form of C containing √the order of Eisenstein integers
√
√
3
. Each of these is a maximal
O = E ≡ Z( 3 1), where 3 1 = e2πi/3 = −1+i
2
order in its algebra KF . Matters get a little more complicated for K = H, as
is to be expected in order to get the different rank 6 hyperbolic Weyl groups.
In one case we choose F = Q, S = Z and the (non-maximal) order is the ring
of Lipschitz integers O = L which are generated over Z by the standard basis
elements {1, i, j, k}. The same rational form contains the maximal order of
Hurwitz integers O = H which is generated over Z by {i, j, k, 1+i+j+k
}, or by
2
the elements given below in (5.40). In another case we add to H the octahedral
√
units (see Section 5.2), producing an order R which is generated over Z( 2)
√ and b = 1+j
√ . In yet another case we will need
by {a, b, ab, ba} where a = 1+i
2
2
√
to make use of icosian units, and appropriate choices of field F = Q( 5),
algebraic integers S and order O, which are described in detail in Section 5.
For the octonions K = O, and to get the Weyl group of E10 we will need the
octavians which are both discrete and a maximal order in the (non-associative)
ring of rational octonions.
Finally, recall that a unit ε ∈ O is an element having a multiplicative inverse
ν ∈ O in the order, but since 1 = N(εν) = N(ε)N(ν) and for all the orders
we use, N(ε), N(ν) ∈ Z, we find that N(ε) = ε̄ε = εε̄ = 1. We will use the
symbol EO to denote the finite group (unless K = O, where the product is not
associative) of units in O.
2.2 Root lattices
The Kac–Moody algebras we are concerned with are hyperbolic extensions
of finite-dimensional algebras associated with discrete subsets of the normed
division algebras K = R, C, H, O and, therefore, are of ranks 3, 4, 6, and
10, respectively. In the main text we consider algebras g++ of over-extended
type [22,11] which arise from simple finite-dimensional Lie algebras g by constructing the non-twisted affine extension g+ (also often denoted by g(1) ), and
4
This can also be traced back to the fact that the matrices making up the Weyl
group, though formally taking entries in some dense order, always are constructed
in such a way that an underlying discrete order governs the structure, see table 1,
consistent with the group multiplication laws of the matrices.
8
connecting the hyperbolic node of the Dynkin diagram by a single line to
the affine node. Our methods also apply to hyperbolic algebras which are extensions of twisted affine algebras, and we discuss examples in Appendix A.
But to make the required formulas more uniform we restrict ourselves to the
non-twisted cases in the main text.
We begin with (the Jordan algebra of) all Hermitian (2 × 2) matrices over K,
X = X† ,
+
x
H2 (K) = X =
z
−
z̄ x
T
x+ , x− ∈ R, z ∈ K
(2.5)
where X† = X is the conjugate transpose of X. 5 H2 (K) is equipped with a
quadratic form
||X||2 = −2 det(X) = −2 x+ x− − zz̄
(2.6)
and a corresponding symmetric bilinear form (X, Y) := 12 (||X+Y||2 −||X||2 −||Y||2 )
which is Lorentzian. The subspace of matrices X ∈ H2 (K) with x+ = x− = 0 is
isomorphic to K, and the restriction of the bilinear form to this subspace agrees
with the positive definite form on K, making the isomorphism an isometry. It
is in that Euclidean subspace that we will find root systems of finite type, and
extend them to hyperbolic root systems in H2 (K).
Suppose that we have been able to find a set of simple roots ai ∈ K, i =
1, . . . , ℓ, where ℓ = dimR (K), with Cartan matrix
"
#
2(ai , aj )
.
C = [Cij ] =
(aj , aj )
(2.7)
of finite type (the bilinear product is defined in (2.1)). Let Q = i Zai ⊂ K
be the finite type root lattice additively generated by those (finite) simple
roots. If necessary, we will denote the Lie algebra type by a subscript on Q.
We also find the highest root θ of the finite type root system, and normalize
the lengths of the simple roots ai so that θθ = 1 always. In some twisted affine
cases described in Appendix A, we may instead choose θ to be the highest
short root, but we still assume that θθ = 1.
P
In the Jordan algebra H2 (K), depending on the choice of Q, we define the
lattice
5
We note that all Jordan algebras employed in this paper can also be realized via
Lorentzian signature Clifford algebras.
9
+
z
x
Λ = Λ(Q) := X =
∈ H2 (K)
z̄ x−
x+ , x− ∈ Z, z ∈ Q
.
(2.8)
The restriction of the bilinear form to Λ makes it a Lorentzian lattice, which
we will identify as the root lattice of a hyperbolic Kac–Moody Lie algebra.
Our first step is to identify the hyperbolic simple roots in the lattice Λ as
follows:
α−1 =
1 0
,
0 −1
α0 =
−1 −θ
,
−θ 0
αi =
0 ai
āi 0
,
1 ≤ i ≤ ℓ.
(2.9)
It follows immediately that
(α−1 , α−1 ) = 2 ,
(α−1 , α0 ) = −1 ,
(α−1 , αi ) = 0 for 1 ≤ i ≤ ℓ ,(2.10)
as well as
(α0 , α0 ) = 2θθ = 2 and
(αi , αj ) = ai āj + aj āi = (ai , aj ) for 1 ≤ i, j ≤ ℓ.
(2.11)
We will give specific choices for the ai so that the matrix C in (2.7) is the
Cartan matrix of a finite type root system, consistent with the embedding of
the corresponding Euclidean root lattice Q into Λ. We will see that choosing
θ to be the highest root of the finite root system (or the highest short root)
makes {α0 , α1 , · · · , αℓ } the simple roots of an untwisted (or a twisted) affine
root system. Furthermore we will see that including α−1 to get all the simple
roots in (2.9) gives the simple roots of a hyperbolic root system with Cartan
matrix
"
2(αI , αJ )
[AIJ ] :=
(αJ , αJ )
#
for I, J = −1, 0, 1, · · · , ℓ.
(2.12)
The unit norm condition on θ is necessary to obtain the single line between
the hyperbolic and affine node. For simply laced algebras all the ai can also
be chosen as units whereas this is no longer true for cases where the Dynkin
diagram has arrows.
The hyperbolic Kac–Moody algebras (of over-extended type) of rank 3,4,6 and
10, respectively, which we obtain by this construction are as follows: F ≡ A++
1
++
++
++
++
++
++
++
(for K = R), A++
,
C
and
G
(for
K
=
C),
A
,
B
,
C
,
D
and
F
2
2
2
4
4
4
4
4
10
(for K = H), and E10 ≡ E8++ , D8++ and B8++ (for K = O). 6 In all cases, real
roots are characterized by det X < 0 and imaginary roots by det X ≥ 0, with
null roots obeying det X = 0.
3
Weyl groups
In this section we study abstractly the Weyl groups of the hyperbolic Kac–
Moody algebras identified in the preceding section. As we define the root
lattice using the Jordan algebra H2 (K) we aim to find a description of the
Weyl group acting on this space. Except where explicitly stated otherwise,
the results of this section apply to all four division algebras, including K =
O. Specific features of the four individual division algebras will be studied
separately in the following sections.
3.1 The simple reflections
The Weyl group associated with the hyperbolic Cartan matrix (2.12) is the
Coxeter group generated by the simple reflections
wI (X) = X −
2(X, αI )
αI ,
(αI , αI )
I = −1, 0, 1, · · · , ℓ .
(3.1)
These generators are known to satisfy the Coxeter relations, which give a
complete presentation for the Weyl group. Recall [24] that a Coxeter group is
defined by the presentation
D
E
RI | RI2 = 11, (RI RJ )mIJ = 11 for I 6= J .
(3.2)
For Weyl groups only special values may occur for the entries of the matrix
M = [mIJ ], namely mIJ ∈ {2, 3, 4, 6, ∞}. More specifically, the relations (3.2)
say that |wI | = 2, that is, each generator has order 2, and they give the orders
of the products of pairs of distinct generators, determined by the entries of
the Cartan matrix as follows:
|wI wJ | = mIJ ,
I 6= J ∈ {−1, 0, 1, · · · , ℓ},
(3.3)
where
The over-extensions of the other finite simple rank 8 root systems C8++ and A++
8
do not give rise to hyperbolic reflection groups.
6
11
mIJ =
2 if AIJ AJI = 0
3 if AIJ AJI = 1
4
6
∞
(3.4)
if AIJ AJI = 2
if AIJ AJI = 3
if AIJ AJI ≥ 4.
For other values of mIJ , the Coxeter group defined by (3.2) cannot be the Weyl
group of a Kac–Moody algebra. Nevertheless, Coxeter groups with other values
of mIJ besides 2, 3, 4, 6, ∞ may be of interest in the present context, as such
groups may occur as subgroups of Weyl groups. A prominent example of such
a non-crystallographic Coxeter group is the group H4 , which is a subgroup of
the Weyl group of E8 [18,35,28].
Of central importance for our analysis is that the action of the simple hyperbolic Weyl reflections wI can be rewritten as a matrix action on X.
Theorem 1 Denote by X = (X† )T the matrix X with each entry conjugated but
the matrix not transposed, let
q
εi = ai / N(ai )
for i = 1, · · · , ℓ
(3.5)
be the unit norm versions of the simple roots ai ∈ Q, and let θ ∈ Q be the
highest root of the finite root system with simple roots ai , always normalized
so that θθ = 1. (For simply laced finite root systems we have εi = ai .) Define
the (2 × 2) matrices MI , I = −1, 0, 1, · · · , ℓ, by
0
M−1 =
1
10
,
−θ
M0 =
1
0 θ
,
εi
Mi =
0
0 −ε¯i
,
(3.6)
i = 1, · · · , ℓ. Then the simple reflections (3.1) can be written as
wI (X) = MI X̄MI†
, I = −1, 0, 1, . . . , ℓ.
(3.7)
Proof: In the non-associative octonionic case one needs to check that the
expression on the right side of (3.7) is well defined for the matrices in (3.6)
without placing parentheses. Since each matrix MI involves only one non-real
octonion, this follows immediately from the alternativity property of O. (See
the section on octonions below for more details.)
12
Now we will check directly that the two formulas (3.1) and (3.7) for wI agree.
First, note that
(X, α−1 ) = x+ − x− ,
(X, α0 ) = −zθ − θz̄ + x− ,
(X, αi ) = zāi + ai z̄(3.8)
for 1 ≤ i ≤ ℓ, so
w−1 (X) = X −
w0 (X) = X −
=
=
−
z
2(X, α−1 )
x
α−1 = X − (x+ − x− )α−1 =
,
(α−1 , α−1 )
z̄ x+
2(X, α0 )
α0 = X − (−zθ − θz̄ + x− )α0
(α0 , α0 )
+
−
−
(x − zθ − θz̄ + x ) (z − zθθ − θz̄θ + x θ)
(z̄ − θzθ − θθz̄ + θx− )
wi (X) = X −
x−
zāi + ai z̄
2(X, αi )
αi = X −
αi
(αi , αi )
N(ai )
+
x
−ε̄i z ε̄i
−εi z̄εi
.
x−
,
(3.9)
where we used the definition (3.5) in the last equation. Compare these with
0
†
M−1 XM−1
=
+
1 x
10
z̄ 0 1
z x−
10
−
x
=
z
z̄ x+
+
z̄ −θ 0
−θ 1 x
M0 XM0† =
0 θ
z x−
1 θ
,
+
−
−
(θθx − zθ − θz̄ + x ) (−θz̄θ + x θ)
=
(−θzθ + θx− )
θθx−
+
z̄ ε¯i 0
εi 0 x
Mi XMi† =
=
0 −ε¯i
+
z x−
εi ε̄i x −εi z̄εi
−ε¯i z ε¯i ε¯i εi x−
.
0 −εi
,
(3.10)
13
†
Hence w−1 (X) = M−1 XM−1
, w0 (X) = M0 XM0† , since θθ = 1, and wi (X) =
Mi XMi† since εi ε̄i = 1. The explicit expressions (3.8) and (3.10) are manifestly
well defined for octonions without placing parentheses.
Our aim in this paper is to study the group generated by the simple reflections
(3.7) as an extension of a matrix group. This extension may always be realized
in the associative cases (K 6= O) as an extension of a matrix group by a small
finite group. Since the full Weyl group W is a semi-direct product of the even
part W + with Z2 = hw−1 i, we will be satisfied to understand just the even
Weyl group W + as an extension of a matrix group.
3.2 Even part of the Weyl group
The formula (3.7) for the simple reflections involves complex conjugation of X.
Therefore all even elements s ∈ W + ⊂ W can be represented without complex
conjugation of X, and it turns out to be simpler to study the even Weyl group
W + in many cases.
The even Weyl group W + is an index 2 normal subgroup of W and consists of
those elements which can be expressed as the product of an even number of
simple reflections. It is generated by the following list of ℓ+1 double reflections:
s0 = w−1 w0 ,
si = w−1 wi
(i = 1, . . . , ℓ) .
(3.11)
Of course, this is not a unique set of generating elements. From the Coxeter
relations (3.2)–(3.4) satisfied by the simple reflections, wI , these even elements
satisfy the relations
s30 = 11 ,
s2i = 11 for i 6= 0
(3.12)
and
mij
(s−1
= 11 for i 6= j and i, j = 0, 1, . . . , ℓ,
i sj )
where mij is given as before in (3.4).
In analogy with Theorem 1 we have
Theorem 2 Define the matrices
14
(3.13)
0 θ
S0 =
,
−θ 1
0 −εi
Si =
.
ε̄i 0
(3.14)
Then the generating double reflections (3.11) acting on X ∈ Λ can be written
for all K as
sI (X) = SI XSI†
, I = 0, 1, . . . , ℓ .
(3.15)
Proof: Follows by direct computation as in the proof of Theorem 1.
An important corollary is
Corollary 3 For the associative division algebras K = R, C, H, all elements
s ∈ W + can be realized by a matrix action according to
s(X) = SXS † ,
(3.16)
where S = Si1 · · · Sin if s = si1 · · · sin ∈ W + in terms of the generating elements (3.11).
Proof: The iterated action of two even Weyl transformations is given by the
associative product of matrices (for K = R, C, H)
(s1 s2 )(X) = S1 (S2 XS2† )S1† = (S1 S2 )X(S1 S2 )†
(3.17)
and has an obvious extension to arbitrary words in the even Weyl group by
associativity.
For K = O, the formula (3.16) no longer holds (even though we have for
any octonionic matrices (S1 S2 )† = S2† S1† ). This can be seen most easily in the
continuous case by a dimension count as will be discussed in Section 6 where
we collect our results specific to the octonionic case.
Let us remark that in the commutative case (3.15) follows immediately by
acting with two successive simple Weyl reflections, say wJ and wI ; the effect
on X is
†
,
wI (wJ (X)) = MI (MJ X̄MJ† )MI† = SIJ XSIJ
(3.18)
where SIJ ≡ MI M J . In this notation the matrices (3.14) are Si ≡ S−1 i for
i = 0, 1, . . . , ℓ. However, in the non-commutative cases one has to be more
15
careful because quaternionic and octonionic conjugation also reverses the order
of factors inside a product, such that, in general
MJ X̄MJ† 6= M̄J XMJT
(3.19)
and MI and M J do not obviously combine into a matrix SIJ which acts by
conjugation as in (3.16). Therefore it is crucial that Theorem 2 applies to all
division algebras.
In the associative cases we obtain from Corollary 3 that W + is isomorphic to
a matrix group generated by (3.14). All elements s ∈ W + act by invertible
matrices and we therefore obtain subgroups of GL2 (K). Furthermore, they all
act by matrix conjugation and therefore a matrix and its negative have the
same action. Other scalar matrices ε11 ∈ GL2 (K), for ε a central unit in K,
would act trivially, so we should be finding W + isomorphic to a subgroup of
P GL2 (K). It is important to note that for general (2 × 2) matrices over K the
determinant is not well defined unless K is commutative. In the cases K = R
and K = C, the matrices MI satisfy det MI = −1, so products of two such
fundamental Weyl reflections have determinant +1 and hence W + is a subgroup of P SL2 (K) for commutative K. With a suitable definition of P SL2 (K)
this statement is also true for non-commutative K. The necessary refinements
required for P SL2 (H) and P SL2 (O) will be presented in the relevant sections.
For all K the P in P SL means that only the quotient by {11, −11} has been
taken. We will argue that the even Weyl groups constitute interesting discrete
‘modular’ subgroups of P SL2 (K).
3.3 Finite and affine Weyl subgroups
By specialization, the construction given above also yields matrix representations of the finite and the affine Weyl subgroups contained in the respective
hyperbolic algebras. We first note that the action of the finite Weyl group
Wfin ≡ W (g) on the root lattice of the finite subalgebra g is obtained as a special case of (3.7) by setting x± = 0 and restricting indices to I ≡ i = 1, · · · , ℓ;
as follows immediately from the last formula in (3.10), the simple reflections
are thus realized on any lattice vector z via
wi (z) = −εi z̄εi
(for z ∈ Q ⊂ K).
(3.20)
The same transformation on z is obtained by matrix conjugation (3.16) with
the even element Si with i = 1, . . . , ℓ, i.e. without complex conjugation of X.
Similarly, it follows from (3.11) that
16
si sj = w i w j
for i, j = 1, . . . , ℓ .
(3.21)
so that, in terms of the matrix representation (3.14) for K 6= O we obtain
−εi ε̄j
Si Sj =
0
0
(3.22)
−ε̄i εj
whence the even part W + (g) of the finite Weyl group acts by diagonal matrices. Defining uij = εi ε̄j and vij = ε̄i εj we deduce si (sj (z)) = uij zv̄ij = uij zvji
for z ∈ Q. To summarize: the even and odd parts of the finite Weyl group,
respectively, act by purely diagonal or purely off-diagonal matrices for K =
R, C, H. 7 Because of the extra w−1 contained in the definition of Si the action
of the odd part is only correct on the subspace with x± = 0; on the full H2 (K)
the odd parts act with an additional interchange of x+ and x− .
The affine subalgebra g+ is characterized by all roots with det X ≤ 0 which
are of the form
m z
X=
z̄ 0
for z ∈ Q and m ∈ Z .
(3.23)
Its associated affine Weyl group Waff is well known to be isomorphic to a
semi-direct product of the finite Weyl group and an abelian group of (affine)
translations T , such that Waff = T ⋊ Wfin . The latter is generated by elements
of the form wtθ w −1 = tw(θ) where w ∈ Wfin and the relevant affine translation
is
1 θ
tθ = w0 wθ =
01
.
(3.24)
Here, wθ is the reflection about the highest root θ and this is the correct expression for the affine translation for all the algebras we consider in this paper.
It is straightforward to check that this matrix indeed generates translations
since
7
For K = H the same action can be alternatively written in terms of pairs [l, r]
and ∗[l, r] of unit quaternions, cf. [8] p. 42. The relation to our notation is as follows:
[l, r] ←→ S =
¯
l 0
0 r̄
,
∗[l, r] ←→ S =
17
0 ¯l
r̄ 0
.
m z † m + θz̄ + zθ z
tθ
tθ =
z̄ 0
z̄
0
(3.25)
This statement also holds for the octonionic case. We note that the interesting S-type transformations in the Weyl groups are then solely due to the
hyperbolic extension.
If K is associative, and once the finite Weyl group has been identified in terms
of diagonal and off-diagonal matrices, the full even hyperbolic Weyl group is
obtained by adjoining the affine Weyl transformation (3.24) to the set of both
diagonal and off-diagonal matrices as generating set.
3.4 Lattice symmetries and Weyl groups
In the associative cases, there are general necessary constraints on the structure of the matrices
a b
S=
cd
,
(3.26)
which represent transformations of the even Weyl group acting on the hyperbolic root lattice Λ of (2.8). These follow from the fact that the transformed
X′ = SXS † again has to lie in the root lattice Λ and should have the same norm
as X. Working out the matrix product one finds the following set of conditions
for the transformed X′ to lie in the root lattice:
aā, bb̄, cc̄, dd¯∈ Z ,
cā, db̄ ∈ Q ,
¯
aai b̄ + bāi ā, cai d + dāi c̄ ∈ Z for i = 1, . . . , ℓ ,
aai d¯ + bāi c̄ ∈ Q for i = 1, . . . , ℓ .
(3.27)
Here, Q is the finite root lattice with simple basis vectors ai . It is the collection
of the conditions (3.27) that will turn the matrices S into a discrete subgroup
of the group of (2 × 2) matrices over K = R, C, H (even if the relevant rings
inside K to which the matrix entries belong are not discrete). The norm preservation requirement will lead to additional determinant-type constraints. The
conditions (3.27) are only necessary but not sufficient since they also allow for
solutions which correspond to lattice symmetries of Q which are not elements
of the Weyl group. This happens when the finite Dynkin diagram admits outer
automorphisms, and we will be concerned with finding manageable conditions
which eliminate these.
18
4
Commutative cases
We first discuss the commutative cases K = R and K = C where one has the
usual definition of the determinant.
4.1 K = R, type A1
For this case we recover the results of [20] for the rank 3 hyperbolic algebra
F = A++
1 , where W (F ) = P GL2 (Z). The root system of type A1 is shown in
fig. 1, and the Dynkin diagram of F is shown in fig. 2. A simplification here
is that we do not have to worry about conjugation; it is for this reason that
an isomorphism with a matrix group exists for the full Weyl group W , rather
than only its even subgroup W + . We have the simple root a1 = 1, which is
identical to the highest root θ = 1, so that the simple roots of the hyperbolic
algebra are represented by the three matrices
1
α−1 =
0
0 −1
,
−1
α0 =
−1
−1 0
,
0
α1 =
1
10
.
(4.1)
The corresponding matrices MI representing the simple Weyl group generators
are
0 1
M−1 =
10
−1 1
,
M0 =
0 1
,
1 0
M1 =
0 −1
.
(4.2)
The even part of the Weyl group is thus generated by the following two matrices (conjugation can be omitted)
0
S := M−1 M1 =
−1
1 0
,
−1
T := M0 M1 =
−1
0 −1
∼
=
1 1
01
(4.3)
implying that W + (F ) ∼
= P SL2 (Z). The full Weyl group in this case can be
obtained by adjoining the matrix M−1 of determinant −1 from which one
recovers W (F ) ∼
= P GL2 (Z) [20].
19
t
✲ta1
r
Fig. 1. The A1 root system with simple root labeled. The root lattice is the lattice
of (rational) integers Z.
✇
−1
✇
✇
0
1
Fig. 2. Dynkin diagram of F ≡
A++
1
with numbering of nodes.
4.2 K = C, simple reflections
For K = C there are different choices of simple finite root systems which we
discuss separately in the following sections, corresponding to the root lattices
A2 , B2 ∼
= C2 , G2 . Here, we collect some common features of all cases. However,
the overextension of the B2 ∼
= C2 root system leads to the hyperbolic algebra
++
C2 if θ is the highest root.
In all cases in this subsection, the matrices giving rise to the simple reflections
are
0
M−1 =
1
10
0 −1
−θ
,
1 0
M1 =
1
M0 =
0 θ
M2 =
0 −ε̄2
,
ε2 0
,
.
(4.4)
where we used the fact that for K = C it is always possible to choose ε1 = 1
(as for K = R). The even Weyl group W + is thus generated by the elements
0 θ
S0 = M−1 M 0 =
−θ 1
S2 = M−1 M 2 =
ε̄2 0
,
0 −ε2
0 −1
S1 = M−1 M 1 =
1 0
,
(4.5)
and we note that S22 = −11 which acts as the identity on X in agreement
with the Coxeter relation s22 = 11. We repeat that we will always normalize
θθ = 1. In the above form the main difference between the three algebras is
20
encoded in θ and ε2 . The three algebras are then distinguished simply by the
multiplicative order of these numbers, which will be specified below for each
case.
We note that the even parts of the finite Weyl groups W + (A2 ), W + (B2 ) ∼
=
W + (C2 ) and W + (G2 ) are cyclic groups of orders 3, 4, and 6, respectively. This
follows also since their generating elements are rotations in the plane and so
one obtains finite subgroups of the abelian group SO(2). From this point of
view the non-abelian nature of the full finite Weyl group arises because of a
single reflection realized as complex conjugation, which shows that the full
finite Weyl group is a dihedral group.
In order to determine the even Weyl group W + for the hyperbolic algebras
A++
and G++
the following result will be useful.
2
2
Proposition 4 Let O be a discrete Euclidean ring in C, that is, a discrete
additive group, closed under multiplication, satisfying the Euclidean algorithm.
Furthermore, assume that O = {m + nθ2 ∈ C | m, n ∈ Z} with θ ∈ O a unit,
and that all units in O have norm 1. Then SL2 (O), the group of (2 × 2)
matrices with entries from O and with determinant 1, is generated by
1 1 ε
,
0
01
0 0
,
−1
ε̄
1
0
,
(4.6)
where ε ∈ EO runs through all units of O.
Proof: We prove the proposition by using arguments from [30]. Let ∆ denote
the group generated by the matrices in (4.6). These matrices belong to SL2 (O)
and we want to show that ∆ = SL2 (O). First, we claim that ∆ contains all
‘translation’ matrices
1
a
01
for a ∈ O.
(4.7)
To see this, note that
θ 01 1θ 0
0θ
01
0θ
2
1 θ
=
0 1
.
(4.8)
We used that the unit θ ∈ O satisfies θθ = 1. Since the translation matrices
form an abelian group with addition of the upper right corner entry, and 1
and θ2 are an integral basis of O, for any m, n ∈ Z, we have
21
m
n
2
1 1 1 θ
01
0 1
2
1 m + nθ
=
0
1
,
(4.9)
so all matrices (4.7) are in ∆. We also get that all matrices
0 1 1 −a 0 −1 1
=
a
1 0
0 1
−1 0
0
1
(4.10)
are in ∆. To complete the proof, choose any matrix A ∈ SL2 (O) and consider
the set of norms
n
o
N = N(bij ) | [bij ] = B = UAV for some U, V ∈ ∆ \{0}.
(4.11)
Because the ring O is discrete, and the norm is positive definite, any set of
non-zero norms has a least element. Then N contains a non-zero minimum
element N(b), and suppose
b11
B=
b12
b21 b22
(4.12)
is a matrix for which this occurs. Multiplying B on the left and/or right by the
third generating matrix in (4.6) (the rotation matrix), we can move any of its
entries into the upper left corner, so we may assume that the minimum occurs
with N(b) = N(b11 ). Now, using the Euclidean algorithm, we may write
with 0 ≤ N(r1 ) < N(b11 )
with 0 ≤ N(r2 ) < N(b11 )
b12 = q1 b11 + r1
b21 = q2 b11 + r2
(4.13)
for q1 , q2 , r1 , r2 ∈ O. Then we have
1
B
−q1
0 1
b11
=
r1
b21 b22 − q1 b21
(4.14)
is a matrix of the form UAV for U, V ∈ ∆, so the norms of its entries cannot
be less than the minimal value N(b11 ). This forces r1 = 0, so b12 = q1 b11 .
Similarly, we have
22
b12
b11
1 0
B =
r2 b22 − q2 b12
−q2 1
(4.15)
is a matrix of the form UAV for U, V ∈ ∆, so the norms of its entries cannot
be less than the minimal value N(b11 ). This forces r2 = 0, so b21 = q2 b11 .
Finally, we see that
1 0 1 −q1 b11 0
′
=B ,
=
B
′
0 b22
0 1
−q2 1
(4.16)
where b′22 = b22 − q1 q2 b11 , is a diagonal matrix of the form UAV for U, V ∈ ∆,
with determinant b11 b′22 = 1. Therefore, b11 must be a unit in O and b′22 =
′
′
b−1
11 = b11 so that B ∈ ∆. All operations transforming A into B ∈ ∆ were
performed using matrices from ∆, so we conclude that A ∈ ∆, completing the
proof.
We remark that this proposition does not apply to the Gaussian integers since
all Gaussian units square to real numbers and one therefore cannot generate
the whole ring from 1 and θ2 for any unit θ.
4.3 K = C, type A2
The first choice of integers we consider is the case of type A2 , which is simply
laced. The simple roots can therefore be chosen as units ai = εi and we take
them to be
√
√
1+i 3
−1 + i 3
, θ = −ε̄2 =
,
ε1 = a1 = 1 , ε2 = a2 =
2
2
where also the highest root θ has been given. The A2 root lattice is spanned
by integral linear combinations of the simple roots, and they form the order
of ‘Eisenstein integers’ E. The A2 root system is depicted in fig. 3.
23
a2 t
❪
❏
t
t
✡
✡
θ
❏
✲ta1
r
❏✡
✡❏
✡ ❏
t
✡
❏t
Fig. 3. The A2 root system with simple roots labeled. The root lattice is the ring of
Eisenstein integers.
✇
−1
✇
2
✇
0❅
❅
❅✇
1
Fig. 4. Dynkin diagram of A++
with numbering of nodes.
2
The hyperbolic algebra A++
has the Dynkin diagram shown in fig. 4. The
2
following choice of simple roots in Λ provides us with that diagram:
1
α−1 =
0
0 −1
0 1
α1 =
10
,
,
−1
,
,
−θ
α0 =
−θ̄ 0
α2 =
−θ 0
0 −θ̄
(4.17)
where we used ε2 = −θ̄. All roots have equal length in this case, so the
hyperbolic extension A++
is also simply laced. We have
2
Proposition 5 The even part of the Weyl group in this case is
∼
W + (A++
2 ) = P SL2 (E)
(4.18)
where P SL2 (E) denotes the ‘Eisenstein modular subgroup’ of P SL2 (C) obtained by restricting all entries to be Eisenstein integers.
Proof: The statement (4.18) is an immediate corollary of Proposition 4, given
that E satisfies the Euclidean algorithm, and all generating matrices in (4.6)
can be obtained within the even part of the Weyl group of A++
2 . This is true
by inspection of the matrices (4.5): the rotation matrix is just S1−1 , and all
diagonal matrices are obtained from powers of
24
t
t
a2 ❅
■
❅
❅
t
θ
❅
❅
❅
t
❅
✲t
❅r
a1
❅
❅
t
t
❅t
Fig. 5. The root system of type C2 , with simple roots labeled and indicated by
arrows. The lattice they generate is a scaled version of the ring of Gaussian integers.
There are both long and short roots, and the same system gives type B2 .
B := S1 S2 =
θ 0
.
0θ
(4.19)
Finally the translation matrix is obtained from
−θ
(S2 S0 )(S1 S2 )−1 =
2
θ θ
·
0 −θ2
0
0
θ
1
=
1
01
(4.20)
where we used θ3 = −1. Thus S0 , S1 and S2 generate a group isomorphic to
SL2 (E). Since in the action, the normal subgroup {11, −11} acts trivially, the
action of the even Weyl group on X is that of the quotient P SL2 (E).
4.4 K = C, type C2
The root system of type C2 is shown in fig. 5. It is the same as the system
of type B2 but the over-extension constructed from this root system is C2++
with Dynkin diagram shown in fig. 6. The dual diagram corresponds to the
(2)
extension of the twisted affine Lie algebra D2 (not B2++ !) and will be discussed
in Appendix A.
The C2 root system is not simply laced, having simple roots whose squared
lengths are in the ratio 2 to 1:
√
−1 + i
1+i
ε1 = a1 2 = 1 , ε2 = a2 = √
, θ = −ε̄2 = √ .
2
2
(4.21)
We obtain the hyperbolic C2++ Dynkin diagram since our simple roots satisfy
(α−1 , α−1 ) = (α0 , α0 ) = (α2 , α2 ) = 2 , (α1 , α1 ) = 1,
25
(4.22)
(α−1 , α0 ) = −1 , (α0 , α1 ) = −1 , (α1 , α2 ) = −1
(4.23)
and all other inner products are zero. From (4.21) we see that θ is a primitive
eighth root of unity with θ2 = −θ̄2 = i. The hyperbolic simple roots of C2++
from (4.21) are
1 0
α−1 =
α1 =
0 −1
0
√1
2
√1
2
0
−1
,
α0 =
−1+i
√
2
0
α2 =
0
−1+i
√
2
−1−i
√
2
0
,
−1−i
√
2
,
.
(4.24)
The long root α2 has a2 as a unit whereas a1 , entering the short root α1 , is
not a unit.
To determine W + (C2++ ) it proves convenient to bring the matrices {S0 , Si } to
another form by means of a similarity transformation
θ
S̃ = USU −1 ,
U =
1/2
0
0 θ−1/2
(4.25)
which gives
2
0 θ
S̃0 =
−θ̄2 1
0 −θ
,
S̃1 =
θ 0
,
0 1
S̃2 =
−1 0
.
(4.26)
Here, we have used θ = −ε̄2 . From these we can build the matrices
A = S̃2 =
0 1
,
−1 0
B = S̃1 S̃2 =
2
0 −θ
C = S̃1 S̃2 S̃1 =
,
θ̄2 0
θ 0
,
0 θ̄
1 −1
D = S̃1 S̃0 S̃1 S̃2 =
,
0 1
1 −θ2
E = S̃1 S̃2 S̃1 S̃0 =
.
0 1
(4.27)
The group generated by these matrices is isomorphic to the even part of the hyperbolic Weyl group. Hence, these matrices contain inversions and rotations
(generated by A, B and C), and translations along some lattice directions
26
✇❅ ✇
✇
Fig. 6.
C2++
−1
0
1
❅
✇
2
Dynkin diagram with numbering of nodes.
(generated by D and E). The similarity transformation (4.25) is useful for explicitly exhibiting the correct lattice translations along two independent basis
vectors 1 and θ2 of the chosen integers via the matrices D and E, respectively.
We note that for all θ the relation C · A = B 2 is valid, showing that the group
generated by A, B, C, D and E is an index 2 extension of the group generated
by A, C, D and E.
With this we can easily recover the link with the so-called Klein–Fricke group
which was first noticed in [20]. Because θ2 = i the matrices C and E of (4.27)
become
0 −i
C=
−i 0
,
1 −i
E=
0 1
.
(4.28)
Together with A and D, these matrices generate the Picard group P SL2 (G)
(where G ≡ Z(i) are the Gaussian integers), see e.g. [32]. We thus recover the
result of [20] (where, however, the explicit form of the embedding was not
given).
Proposition 6 The even Weyl group W + (C2++ ) is an index 2 extension of
P SL2 (G), that is,
W + (C2++ ) ∼
= P SL2 (G) ⋊ 2 ≡ P SL2 (Z(i)) ⋊ 2 .
(4.29)
Semi-directness follows since conjugation of P SL2 (G) by the matrix B is an
automorphism.
4.5 K = C, type G2
The root system of type G2 is depicted in fig. 7. It is not simply laced, having
simple roots whose squared lengths are in the ratio 3 to 1.
ε1 = a1 = 1 , ε2 =
√
√
√
1 + 3i
− 3+i
, θ=
.
3a2 =
2
2
(4.30)
The G2 root system is thus
√ the superposition◦ of two A2 root systems which
are scaled by a factor of 3 and rotated by 30 degrees relative to each other.
27
Note that in this case θ 6= −ε̄2 ; rather, we have ε22 = θ, and that θ is identical
to the highest root of A2 .
The Dynkin diagram of the hyperbolic algebra G++
is shown in fig. 8. We will
2
get this diagram if our simple roots satisfy
2
(α−1 , α−1 ) = (α0 , α0 ) = (α1 , α1 ) = 2 , (α2 , α2 ) = ,
3
(4.31)
(α−1 , α0 ) = −1 , (α0 , α1 ) = −1 , (α1 , α2 ) = −1
(4.32)
and all others zero. The hyperbolic simple roots from (4.30) of G++
are
2
α−1 =
1 0
,
0 −1
0 1
α1 =
,
10
√
−1− 3i
2
−1
α0 =
−1+ 3i
2
0
α2 =
0
√
−1− 3i
2
√
−1+ 3i
2
√
0
,
.
(4.33)
(NOTE: This formula for α2 had a typographical error in the published version
which is corrected above.) We label α2 as a short root, so a2 is not a unit.
The finite Weyl group W (A2 ) is a subgroup of index 2 in W (G2 ). The same
is true for their hyperbolic extensions: we have
Proposition 7 The even Weyl group W + (G++
2 ) is an index 2 extension of
++
W (A2 ), that is,
+
++
W + (G++
2 ) = W (A2 ) ⋊ 2 = P SL2 (E) ⋊ 2 .
(4.34)
Proof: Because θ 6= −ε̄2 we must proceed slightly differently than before.
First we notice that
−ε̄2 0
S1 S2 =
⇒
0 −ε2
S1 S2
2
θ 0
=
0θ
(4.35)
Since θ6 = 1 (as for A2 ), the set of matrices S0 , S1 and (S1 S2 )2 S1 coincides with
the set (4.5) for A2 , hence these matrices generate the group P SL2 (E), again
by Proposition 4. To get the full even Weyl group we must adjoin the matrix
S2 obeying S22 = −11, generating a Z2 . The semi-directness of the product
28
a2
t
θt
t
❏
✡
❏
✡
t
❏
✡
✡
t
t ❏
❨❍ ❏ ✡ ✟✟
❍
❍❍
❏✡
✟✟
✟
❍❍
✟ ✡❏
✟
t ✟ ✡ ❏ ❍❍t
❏
✡
✡
t ❏
❏
✡
❏t
✡
t
✲t
a1
Fig. 7. The root system of type G2 . The long and short roots are labeled and
indicated by arrows.
✇
Fig. 8.
G++
2
−1
✇
0
✇❅ ✇
1
2
Dynkin diagram with numbering of nodes.
will follow from the action S2 SS2−1 ∼
= S2 SS2 of the extending matrix S2 on
S ∈ P SL2 (E), if the resulting matrix is in P SL2 (E) so that this action gives
an automorphism of P SL2 (E). By expanding the product, commutativity and
the fact that ε22 ∈ E the result follows.
5
Quaternions K = H
Within the four-dimensional quaternion algebra one can find the root systems
of types A4 , B4 , C4 , D4 and F4 . The associated hyperbolic rank 6 Kac–Moody
++
++
++
++
algebras are now A++
4 , B4 , C4 , D4 , and F4 , and we will give explicit
descriptions of their even Weyl groups in terms of matrix groups below. 8
The quaternionic case is more subtle than the commutative cases because the
criterion for selecting the matrix group to which the even Weyl group belongs
cannot be so easily done via a determinant. Before turning to the issue of
how to define determinants and matrix groups with quaternionic entries we
first discuss different types of quaternionic integers and numbers required for
exhibiting the root systems.
5.1 Hurwitz and Lipschitz integers
The standard basis for H, {1, i, j, k}, has the famous products
8
A description of the finite root systems of types B4 , D4 and F4 , and their Weyl
groups in terms of quaternions has been given in [29].
29
i2 = j 2 = k 2 = −1, ij = −ji = k, jk = −kj = i, ki = −ik = j, (5.1)
so an obvious subring of integers is formed by the Lipschitz integers
L = {n0 + n1 i + n2 j + n3 k n0 , n1 , n2 , n3 ∈ Z} .
(5.2)
They form an order (a subring of the rational quaternions, HQ , finitely generated as a Z-module, containing a Q-basis of HQ ) but are not a maximal order
in HQ since they are contained in the ring of Hurwitz integers
H = n0 + n1 i + n2 j + n3 k n0 , n1 , n2 , n3 ∈ Z or n0 , n1 , n2 , n3 ∈ Z +
1
2
,
(5.3)
which constitute a maximal order in HQ . We note that these two rings, L and
H, are generated by the Lipschitz and Hurwitz units given below in (5.9) and
(5.10), respectively, and as Z-modules, they are discrete lattices. Also note that
the non-commutative ring H of Hurwitz integers satisfies the division with small
remainder property required for the Euclidean algorithm (cf. Proposition 4),
whereas the Lipschitz integers L do not [8].
For the determination of some of the even Weyl groups we rely on the following
definition and lemma.
Define C to be the two-sided ideal in the ring H generated by the commutators
[a, b] = ab − ba
for all a, b ∈ H .
(5.4)
To understand this we use the integral basis for H used in [30],
1
e0 = (1 + i + j + k),
2
i,
j,
k,
(5.5)
which has the nice property that an integral combination m0 e0 + m1 i + m2 j +
m3 k is in L when m0 ∈ 2Z. It is easy to check the commutators
[e0 , i] = j − k,
[i, j] = 2k,
[e0 , j] = k − i,
[j, k] = 2i,
[e0 , k] = i − j,
[k, i] = 2j,
(5.6)
which are all purely imaginary Lipschitz integers. [H, H] is all integral linear
combinations of the commutators above, but that is not even a subring of H,
as we can see, for example, from the fact that (2i)(j − i) − 2k = 2 is not a
commutator. The ideal C = H[H, H]H consists of finite sums of the form a[b, c]d
for any a, b, c, d ∈ H, and it is enough to compute these for a and d from the
above integral basis and [b, c] from (5.6). The only products we need to know
30
are
e0 i = −e0 + i + j,
ie0 = −e0 + i + k,
e0 j = −e0 + j + k,
je0 = −e0 + i + j,
e0 e0 = −e0 + i + j + k.
e0 k = −e0 + i + k,
ke0 = −e0 + j + k,
(5.7)
Using these, it is clear that a[b, c]d is in L, and consists of all integral linear
combinations of the elements
±1 ± i, ±1 ± j, ±1 ± k, ±i ± j, ±i ± k, ±j ± k,
(5.8)
which is clearly an index 2 integral lattice in L. Since L is an index 2 sublattice
in H, we have that H/C is a ring of four elements, and we can take as coset
representatives {0, −e0 , (−e0 )2 = e0 − 1, (−e0 )3 = 1} which form the field F4
of order 4 whose nonzero elements form the cyclic group of order 3. Also note
that there are no units in C since its nonzero elements have minimal length 2.
Lemma 1 Let a1 , . . . , an ∈ H be any n Hurwitz numbers. Then the product
a1 · · · an is commutative modulo C.
Proof: Since the quotient ring H/C is a field, where the product is commutative, the projection of the product a1 · · · an is equal to the projection of
aσ(1) · · · aσ(n) for any permutation σ of {1, . . . , n}.
5.2 Quaternionic units and rings
For the root lattices of A4 , B4 , C4 and F4 (but not D4 ) we also need quaternionic units which are neither Lipschitz nor Hurwitz numbers. These other
units parametrize finite subgroups of SU(2), and are, respectively, related to
the octahedral (for B4 , C4 and F4 ) and icosahedral (for A4 ) groups, as explained e.g. in [8]. Using diagonal and off-diagonal (2 × 2) matrices of such
units (or alternatively pairs of units [8], cf. footnote 7 in Section 3.3) we can
then reconstruct the Weyl groups of all the finite simple rank 4 algebras, as
we shall explain below. Besides the eight Lipschitz units
EL = {±1, ±i, ±j, ±k}
(5.9)
which form the quaternionic group, often denoted by Q8 , we have the 24
Hurwitz units
1
EH = ±1 , ±i , ±j , ±k , (±1 ± i ± j ± k)
2
31
(5.10)
which form a subgroup in the unit quaternions. According to [7], p. 55, EH ∼
=
2 · A4 is an index two extension of the alternating group on four letters. It
can also be seen that EH ∼
= EL ⋊ Z3 is a semi-direct product of the group of
Lipschitz units with a cyclic group of order 3 related to triality and explained
below in Section 5.4. Any quaternion z = n0 + n1 i + n2 j + n3 k is a root of the
real polynomial
pz (t) = (t − z)(t − z̄) = t2 − 2n0 t + (n20 + n21 + n22 + n23 )
(5.11)
whose coefficients will be integers if z is in L or H, and in that case, unless
z ∈ Z, this is the minimal polynomial satisfied by z ∈ H.
We will also need the following set of 24 octahedral units
(
±1 ± i ±1 ± j ±1 ± k ±i ± j ±i ± k ±j ± k
√
, √
, √
, √ , √
, √
2
2
2
2
2
2
)
(5.12)
which do not form a group by themselves, but the product of any two of them
is a Hurwitz unit, and the product of any Hurwitz unit and an octahedral
unit is an octahedral unit, so their union forms a group of order 48 which we
call the octahedral subgroup, ER . We can write it as the disjoint union of two
cosets of its normal subgroup, EH ,
ER ≡ EH ∪ iO · EH
j−k
where iO ≡ √ .
2
(5.13)
These are, in fact, all the units in the ring, R, generated by integral linear
combinations of units in ER . Contrary to the Lipschitz and Hurwitz numbers,
the ring R is not discrete but dense in H√(using the usual topology). Nevertheless R is an order of HF (with F = Q( 2)) as defined in Section 2.1, since
it is finitely generated by√its 48 units over Z. It is not hard to show that it
is generated over S = Z[ 2] just by the four elements, {a, b, ab, ba}, where
√ and b = 1+j
√ . Note also that the integral span of the coset of ‘purely
a = 1+i
2
2
octahedral numbers’ iO · H constitutes a lattice in H which can be regarded
as a ‘rotated’ version of the lattice H (but which, unlike H, is not closed under
multiplication). 9
Finally, we have 96 icosian units [7], p. 207, defined in terms of
9
In fact, we will only encounter matrices S whose entries belong either to H (and
have minimal polynomial of degree two) or to iO H (and have minimal polynomial
of degree four). This will be distributed over the matrix in such a way that this
structure is preserved under multiplication.
32
√
√
1
1
τ = (1 + 5) and σ = (1 − 5) ,
2
2
(5.14)
to be all elements obtained from the following eight basic elements
1
± i ± σj ± τ k
2
(5.15)
by the 12 even permutations of the Lipschitz units (1, i, j, k). Together with
the 24 Hurwitz units they form the icosian subgroup EI , which are the 120
units in the ring I defined to be their integral span. According to [7], the
structure of the icosian group EI ∼
= 2 · A5 is an index two extension√ of the
alternating group on five letters. I is an order of HF where F = Q( 5). As
in the order R, the ring I is dense in H, but it is finitely generated over the
integers by its 120 units. In fact, it is generated over S = Z(τ ) by the four
elements
i,
j,
1
ω = (−1 + i + j + k) ,
2
1
iI = (i + σj + τ k) .
2
(5.16)
It is straightforward to check that all their products can be expressed as linear
combinations of those four elements with coefficients from S, for example:
i2 = −1 = −(σ + 1)i − (σ + 2)j + 2ω + 2σiI ,
(5.17)
jiI = −2σi − (2σ + 1)j + σω + (2σ + 1)iI ,
ωiI = ω − i − k
iI ω = −i − (σ + 1)j + ω + σiI ,
iI j = −i − σj + σω + iI .
These can be used to compute the commutators of pairs of generators, e.g.
[i, iI ] = −τ j + σk,
[i, j] = 2k,
(5.18)
If we try to imitate what we did for the Hurwitz integers, and let CI = I[I, I]I
be the ideal of I generated by the commutators [I, I], we find that CI = I
is the whole icosian ring. This follows since the two elements in (5.18) have
norms 2 and 3, respectively, so that the difference of their norms is 1 and
hence 1 ∈ CI . This means that the quotient ring, I/CI is trivial, not a field as
was H/C, and we do not have an analog of Lemma 1 for I.
Even though most rank four root systems contain units from one of the dense
rings R or I we will explain how one can nevertheless construct discrete matrix
groups from them which are related to the even Weyl groups.
33
5.3 Quaternionic determinants
Useful introductory references for quaternionic determinants are [15,1]. When
K = H, the matrices in (3.16) are (2 × 2) matrices over the quaternions, that
is, matrices of the form
a b
S=
cd
with a, b, c, d ∈ H .
(5.19)
To describe the hyperbolic Weyl groups, we will only need to take these entries from the rings introduced above, but the following observations apply
to all quaternionic matrices. If one tries to define the determinant of the
quaternionic matrix S to be ad − bc, the non-commutativity of H spoils the
usual multiplication law for determinants (see [1]). A quantity which does have
good properties for K = H is obtained by observing that the determinant is
well defined for Hermitian quaternionic matrices (for which a and d are real,
and b = c̄). Given any quaternionic matrix S, we therefore associate with
it the Hermitian matrix SS † which has a well-defined determinant given by
det(SS † ) = det(S † S). 10 The multiplication law is then obeyed provided that
det (S1 S2 ) · (S2† S1† ) = det S1 S1† det S2 S2† .
(5.20)
This relation (which is obviously satisfied for K = R, C) can be verified for
K = H by direct calculation (however, it will fail for K = O because of nonassociativity). Any quaternionic matrix S satisfying det(SS † ) 6= 0 has left and
right inverses given by
S †S
−1
S † = S † SS †
−1
.
(5.21)
Here, (SS † )−1 denotes the well-defined inverse of a Hermitian matrix. That
the two expressions
can be shown by explicit computation. 11 Hence,
coincide
†
the condition det SS = 1 for (2 × 2) matrices with quaternionic entries
indeed defines a continuous group which we denote by
10
a
SL2 (H) := S =
b
cd
a, b, c, d ∈ H , det SS † = 1 .
(5.22)
These expressions also make sense for K = O but in general are no longer equal.
For K = O, the two expressions in (5.21) are generally different because of nonassociativity.
11
34
Writing 11 for the (2×2) identity matrix, the subgroup {±11} is normal and we
define P SL2 (H) = SL2 (H)/{±11} to be the quotient group. We write elements
in the quotient as matrices S but we identify S with −S. 12 This group has
real dimension 15 = 4 · 4 − 1, and is known to be isomorphic to the Lorentz
group in six dimensions [41,2], i.e.
P SL2 (H) ∼
= P SO(1, 5; R) .
(5.23)
We will soon see that the even Weyl groups of all rank six Kac–Moody algebras
under consideration are discrete subgroups of P SL2 (H).
The easiest of these discrete groups to describe is P SL2 (H), which is obtained
from P SL2 (H) by restricting all matrix entries to be Hurwitz integers. We
will also need its subgroup
(0)
a
P SL2 (H) := S =
b
cd
∈ P SL2 (H) ad − bc ≡ 1(mod C)
(5.24)
which can be understood in the following way. The composition of the group
homomorphism
P SL2 (H) → P SL2 (H/C)
(5.25)
with the usual determinant of a (2 × 2) matrix over the field H/C ∼
= F4 , yields
a group homomorphism
Det : P SL2 (H) → F∗4
(5.26)
onto F∗4 ∼
= Z3 the cyclic group of order 3. 13 Since this composition is a group
homomorphism, we have Det (S1 S2 ) = Det S1 · Det S2 . Its kernel is a normal
subgroup of index 3 in P SL2 (H), giving an alternative form of (5.24) as
(0)
o
n
P SL2 (H) = S ∈ P SL2 (H) Det S = 1 .
(5.27)
Since L = C ∪ (1 + C) is the coset decomposition of L, the condition ad − bc ≡
1(mod C) says that ad − bc ∈ (1 + C) ⊂ L and Lemma 1 says that the order of
the products in the expression ad − bc does not matter. Since ad − bc ∈
/ C, the
(0)
condition for S ∈ P SL2 (H) is just ad − bc ∈ L. We have proved the following.
12
In fact, if S ∈ SL2 (H) then so is εS for any unit ε. Because of non-commutativity,
only ε = ±1 are in the center and quotiented out.
13 This ‘determinant’ is different from the Dieudonné determinant [1].
35
(0)
Lemma 2 P SL2 (H) is index 3 in P SL2 (H), and
(0)
P SL2 (H)/P SL2 (H) = A3
(5.28)
where A3 ∼
= Z3 is the alternating group on three letters.
(0)
The modular group P SL2 (H) contains the modular group P SL2 (L), but is
strictly larger, so that we arrive at the following chain of subgroup relations
(0)
P SL2 (L) ⊂ P SL2 (H) ⊂ P SL2 (H) .
(5.29)
The map Det extends to the ring of all (2 × 2) Hurwitz matrices, first reducing entries modulo C, and then taking the usual determinant. It is still a
multiplicative map so that the invertible matrices have non-zero Det. But it is
possible for a ‘non-invertible’ matrix in that ring to also have a non-zero Det.
For example, the diagonal matrix diag(1, 3) does not have an inverse over the
Hurwitz integers, but since 2 ∈ C, it reduces to the identity matrix modulo C,
whose usual determinant is 1 + C ∈ F∗4 .
(0)
Proposition 8 The modular group P SL2 (H) is generated by the matrices
1 1
01
,
0 1
10
,
a 0
0b
,
(5.30)
where a, b ∈ EH are Hurwitz units and ab ≡ 1(mod C).
Proof: This claim can be proved by adapting Theorem 2.2 on page 16 of [30],
and arguments very similar to those of Proposition 4. According to Krieg’s
theorem, the set of all uni-modular (= invertible) (2×2) Hurwitz quaternionic
matrices, here denoted by SL2 (H) ⊂ SL2 (H) in accordance with definition
(5.22) (but designated as GL2 (H) in [30]) is generated by taking products of
the matrices in (5.30) without any restriction on the product ab, so that a, b
run through all Hurwitz units. This is also true for the quotient P SL2 (H). So
we can write any element S ∈ P SL2 (H) as
S = G1 · · · Gn
(5.31)
where each Gi , i = 1, . . . , n, is one of the three types of generating matrices
in (5.30), but with no restriction on the product ab for the third type in
accordance with Krieg’s theorem. We will show that we can rewrite S in the
form
36
c 0 ′
′
S=
G1 · · · Gn ′ ,
01
(5.32)
where c ∈ H and all G′i matrices are from (5.30) with the third type satisfying
the ab ≡ 1(mod C) condition. In order to arrive at this new presentation
we show explicitly how to convert the presentation (5.31) to one of the form
(5.32), and how the element c represents Det(S).
Examine the matrices in the expression G1 · · · Gn from the right to the left. We
don’t need to change a matrix of the first or second kind from (5.30), nor do
we change one of the third kind if it satisfies the ab ≡ 1(mod C) condition. Let
i0 be the largest index for which Gi0 = diag(ai0 , bi0 ) with ai0 bi0 6≡ 1(mod C).
Then we can factorize it as
ai0 0
0 bi0
ε̄ 0 εai0 0
=
01
0 bi0
(5.33)
where the unit ε ∈ EH \ EL is chosen such that εai0 bi0 = 1 ∈ L. We will show
below how the new matrix diag(ε̄, 1) can be moved to the left by passing
through the other generators Gi with i < i0 , and leaving only acceptable
generators on its right.
We first note that from (5.30) (with the restriction on the product) we can
generate all translation matrices
1 a
01
,
for a ∈ H .
(5.34)
For a ∈ {±i, ±j, ±k} we have
a
0 1 1 −a 0
01
01
0 1
1
=
a
01
(5.35)
and a ≡ 1 (mod C) so each matrix on the left side is from (5.30). Products
of these provide all translation matrices (5.34) with a ∈ L. To extend this
to all a ∈ H it is sufficient to find one translation matrix with a ∈ H \ L.
The computation in (4.8) where θ = 12 (1 − i − j − k) satisfies θ2 = −θ̄ =
1
(−1 − i − j − k), yields this, so that we indeed obtain all matrices (5.34)
2
from (5.30). From this and
37
01 ε
ε̄
1 1 ε̄ 0
=
0
01
01
1
01
(5.36)
we see that we can pass the matrix diag(ε̄, 1) across a matrix Gi of the first
kind in (5.30) by replacing Gi by a translation matrix which is a product of
the allowed generators.
Similarly we have
0 1 ε̄ 0
10
01
ε̄ 0 ε 0 0 1
=
01
0 ε̄
10
,
(5.37)
so that we can also pass through the second type of generating matrices by
replacing it by a product of allowed matrices.
Finally, for any units a, b ∈ H, we can also write
a
0 ε̄ 0
0b
01
aε̄b
=
0 b̄ 0
0 1
0b
,
(5.38)
so the matrix on the far right is an allowed generator, but the matrix being
moved to the left has been changed to another of the form diag(c, 1) with
c = aε̄b ∈ H a unit. Using (5.36)–(5.38), for some unit c ∈ H, we finally arrive
at the expression
c 0
S=
01
′
′
G1 · · · Gn ′
(5.39)
with all G′i belonging to the list of generating matrices (5.30) and those of
the third kind have ab ≡ 1(mod C), so all Det(G′i ) = 1 ∈ F∗4 . Then Det(S) =
Det(diag(c, 1))Det(G′1 ) · · · Det(G′n′ ) = c + C ∈ H/C = F4 . If we assume that
(0)
S ∈ P SL2 (H), so Det(S) = 1, then we must have c ≡ 1(mod C) so that
diag(c, 1) is an allowed generator from (5.30) and the proof is complete.
We close this section by counting the number of diagonal matrices diag(a, b)
such that a, b ∈ H are units and ab ∈ L. There are 64 = 8 × 8 such matrices
when both a, b ∈ L. If a ∈ H\L is one of the 16 pure Hurwitz units there are
8 choices for b such that ab ∈ L, which gives 128 = 16 × 8 such matrices.
Descending to the quotient by {11, −11} we are left with 96 = 12 × 192 diagonal
38
(0)
matrices in P SL2 (H). The full P SL2 (H) has 288 =
matrices. This gives another way to see Lemma 2.
1
2
× 24 × 24 diagonal
5.4 Even Weyl group W + (D4++ )
We first study the Weyl group of the rank 6 hyperbolic algebra D4++ . Among
the Weyl groups associated to the rank 6 algebras, this is the ‘easiest’ to
understand because it is the only one whose root system can be expressed
solely in terms of Hurwitz numbers. As we will see, the Weyl groups in all the
other cases, except for A++
4 , are finite extensions of this one.
In the lattice H equipped with the inner product (2.1), we can recognize that
EH forms the D4 root system, and among these we choose as simple roots
ε1 = a1 = 1 ,
ε3 = a3 =
1
−1+i−j−k ,
2
1
ε4 = a4 =
−1−i−j+k .
2
ε2 = a2 =
1
−1−i+j −k ,
2
(5.40)
We can choose ai = εi for all i because D4 is simply laced. This system of
simple roots exhibits S3 ‘triality’ symmetry as outer automorphisms, that is,
it is symmetric under cyclic permutation of the three imaginary units (i, j, k),
and under the exchange of any two imaginary units, e.g. (j ↔ k). The former
is concretely realized by the map z 7→ θz θ̄, whereas the latter corresponds to
z 7→ iO z̄ı̄O . Here, θ is the highest D4 root
θ = 2a1 + a2 + a3 + a4 =
1
1 − i − j − k = jε2 = kε3 = iε4
2
(5.41)
and iO is the specific octahedral unit of order four defined in (5.13). The 16
units in EH \ EL are of order three or six, viz.
ε32 = ε33 = ε34 = 1 ,
θ6 = 1 .
(5.42)
With these choices we build the simple roots (2.9) of the hyperbolic algebra
D4++ in its root lattice Λ(Q) where Q = H is the D4 root lattice. The corresponding Dynkin diagram of D4++ , the hyperbolic over-extension of the D4
Dynkin diagram, is depicted in fig. 9. Recall that the Weyl group of D4 is
W (D4 ) = 23 ⋊ S4 .
Our central result is
(0)
Proposition 9 W + (D4++ ) ∼
= P SL2 (H).
39
✇
✇
✇
−1
2
✇
0
✇
1
3
✇
4
Fig. 9. Dynkin diagram of D4++ with numbering of nodes.
Proof: Substituting (5.40) into the expressions of the generating matrices Si ,
0 ≤ i ≤ 4, given in (3.14), one sees that all these matrices have det(Si Si† ) = 1
(0)
and Det(Si ) = 1 so they belong to P SL2 (H) ⊂ P SL2 (H) as defined in (5.24).
(0)
Therefore, we get W + (D4++ ) ⊂ P SL2 (H). To prove the converse we show
that all the generating matrices of Proposition 8 are in W + (D4++ ). To do this
we compute
εi 0
Si S1−1 =
0 ε̄i
i 0
S2 S3 =
0j
for i = 2, 3, 4 ,
j 0
, S3 S4 =
0k
k 0
, S4 S2 =
0 i
.
(5.43)
It is straightforward to see that by further multiplication and permutation we
can obtain from these matrices and S1 any matrix of the form
a 0
0b
and
0 a
b0
(5.44)
where (a, b) is any pair of Hurwitz units whose product is a Lipschitz unit, i.e.
N(a) = N(b) = 1 and ab ≡ 1 (mod C). We note that up to (a, b) ∼
= (−a, −b)
there are 96 pairs with this property so that there are 96 matrices of each
type in (5.44) in P SL2 (H) corresponding to the even and odd part of the finite
Weyl group W (D4 ) of order 192.
Given the matrices (5.44) we can also reconstruct the translation matrix from
S0 (using θ3 = −1) and
2
0 θ 0 θ −θ 0
1 −1
2
·
·
=
.
0 θ
θ 0
−θ 1
0 1
(5.45)
The translation matrix is the inverse of this, giving the last of the required
generators (5.30).
40
(0)
The appearance of the group P SL2 (H) for D4++ (rather than P SL2 (H)) may
be viewed as a manifestation of the triality symmetry of the D4 Dynkin diagram. Since the even cyclic permutation of (i, j, k) is realized by conjugation
by the Hurwitz unit θ, the associated diagonal matrix diag(θ, θ) ∈ P SL2 (H)
solves the necessary conditions (3.27), but itself is not part of the Weyl group,
and therefore has to be removed from P SL2 (H). This is what the DetS = 1
(0)
condition achieves, and since θ3 = −1 we find that the subgroup P SL2 (H) is
index three in P SL2 (H), in accordance with Lemma 2. In this way, the group of
(even) outer automorphism of D4 can also be realized by matrix conjugation.
For later reference we denote the alternative generating set of the group
(0)
P SL2 (H) furnished by the even D4++ Weyl group by
gi = SiD4
for i = 0, . . . , 4 .
(5.46)
An alternative argument leading to the statement of the proposition can be
based on the fact that every Hermitian matrix X ∈ Λ(D4 ) obeying det X = −1
is a real root, 14 and that the (full) Weyl group acts transitively on the set of
real roots. In the case of D4++ they form a single Weyl orbit, since the Dynkin
diagram has only single lines. This implies that we can generate all real roots
from the hyperbolic simple root α−1 by acting with the full Weyl group, viz.
n
1 0
o
∆re = X ∈ Λ(D4 ) | det X = −1 = W ·
0 −1
.
(5.47)
In particular, considering all even Weyl images of the form s(α−1 ) = Sα−1 S † ,
and using the fact that Weyl transformations preserve the norm, we get
−1 = det(α−1 ) = det(Sα−1 S † ) = − det(SS † )
(5.48)
where the third equality can be verified by direct computation for the quaternions. This is the same condition as (5.22). Conversely, to any S ∈ P SL2 (H)
satisfying (5.22) we can associate the real root Sα−1 S † , whence S becomes
associated to some symmetry of Λ(D4 ). This seems to give all of P SL2 (H) but
the reasoning does not distinguish between inner and outer transformations.
As argued above, taking the outer (diagram) automorphisms into account is
the same as demanding the extra condition (5.24).
14
It is here that the hyperbolicity of the Kac–Moody algebra is essential; this statement is no longer true if the indefinite Kac–Moody algebra is not hyperbolic.
41
✇
✇
−1
✇
2
✇❅ ✇
✇
0
1
3
4
Fig. 10. Dynkin diagram of B4++ with numbering of nodes.
5.5 Even Weyl group W + (B4++ )
The B4 root lattice is isomorphic to the hypercubic lattice of the Lipschitz
integers (in the same way that B2 was associated with the cubic lattice of
Gaussian integers). The simple roots can be chosen as follows
1
−1+i−j−k ,
ε2 = a2 =
2
√
−j + k
ε4 = 2a4 = √
.
(5.49)
2
ε1 = a1 = 1 ,
ε3 = a3 =
1
−1−i+j−k ,
2
The highest (long) root is
θ = a2 + 2a1 + 2a3 + 2a4 =
1
1−i−j−k .
2
(5.50)
Hence, a4 is the short simple root, while the other (long) roots, including the
highest root θ, are normalized to unity. These simple roots were chosen to be
close to the D4 simple roots: a1 , a2 and a3 agree and only a4 is changed; the
highest roots agree. This somewhat obscures the fact that the B4 lattice is a
scaled and rotated version of the lattice of Lipschitz numbers but makes the
determination of the Weyl group simpler. The finite Weyl group is W (B4 ) =
24 ⋊ S4 , and is concretely realized via permutations and reflections of the four
basis elements 1, i, j, k. Observe also that the above set of simple roots is not
invariant under cyclic (triality) transformations of the three imaginary units,
as was to be expected.
The hyperbolic extension B4++ can be obtained by following the procedure
described before and leads to the Dynkin diagram of fig. 10. Turning to its even
hyperbolic Weyl group W + (B4++ ) we notice that among the generators Si ∈
P SL2 (H) (for i = 0, . . . , 4) the first four are identical to those of W + (D4++ ).
By explicit computation one finds furthermore that
S4 S3 S4 =
0
1
(−1
2
+ i + j − k)
1
(1
2
+ i + j − k)
0
42
= g4 ,
(5.51)
(0)
where g4 was defined in (5.46) in the alternative generating set of P SL2 (H).
(0)
This shows that W + (B4++ ) has a subgroup isomorphic to P SL2 (H) generated
by S0 , S1 , S2 , S3 and S4 S3 S4 . Adjoining to these the element S4 which satisfies
(0)
S42 = −11 ∼
= 11 gives an extension in which P SL2 (H) is a normal subgroup of
index 2, and we arrive at
Proposition 10 W + (B4++ ) ∼
= P SL2 (H) ⋊ 2.
(0)
(0)
The semi-directness follows from expanding the action S4 SS4−1 for S ∈ P SL2 (H)
and using the properties of√ε4 . As the index 2 extension is realized via the matrix S4 which contains 1/ 2 in all entries we see that W + (B4++ ) consists of
all matrices (5.19) in P SL2 (H) such that either all a, b, c, d are Hurwitz numbers, or all are pure octahedral numbers ∈ iO · H (it is easy to see that this
property is preserved under matrix multiplication because iO H · H = iO H and
iO H · iO H = H). This is the ‘trick’ by which the discreteness of the matrix group
is achieved despite the fact that R is dense in H.
To conclude the analysis of B4++ we relate our results, restricted to the finite
Weyl group W + (B4 ), to the classification of [8] (see their Table 4.2) where
W + (B4 ) is denoted by ± 61 [O×O]. It is not hard to check that all the generating
elements given in [8] can be obtained from products of matrices of the type
(3.22) when the B4 units (5.49) are plugged in.
5.6 Even Weyl group W + (C4++ )
For C4 the direction of the arrow in the finite Dynkin diagram is reversed
compared to B4 . A convenient choice of simple roots requires again quaternions
which are not Hurwitz. In order to facilitate the comparison with D4 we choose
the first three simple roots of C4 to coincide with the simple roots a3 , a1 , a4
of D4 up to a re-scaling required to maintain the unit normalization of the
highest root θ. Thus
ε1 =
√
2a1 = 1 ,
ε3 =
√
2a3 =
1
−1−i+j−k ,
2
√
1
−1+i−j−k
2
−j + k
ε4 = a4 = √
(5.52)
2
ε2 =
2a2 =
with the (long) highest root
θ = 2a1 + 2a2 + 2a3 + a4 =
−j − k
√
,
2
43
(5.53)
✇❅ ✇
✇
−1
0
2
✇
✇
1
3
❅
✇
4
Fig. 11. Dynkin diagram of C4++ with numbering of nodes.
For the determination of the Weyl group we again choose a convenient basis.
The hyperbolic extension C4++ has Dynkin diagram given in fig. 11. Although
the B4 and C4 Weyl groups are isomorphic (cf. table 1), the Weyl groups of
C4++ and B4++ are nevertheless different. This difference in our basis is reflected
only in the difference between the highest roots; for C4 we have an octahedral
unit of order four whereas for B4 the highest root was a Hurwitz number of
order six.
Proceeding as for W + (B4++ ) we first observe that all Si for i = 1, . . . , 4 are
identical to those of B4++ (since the finite Weyl groups are isomorphic), and
hence again
S4 S3 S4 =
1
(1
2
0
1
(−1
2
+ i + j − k)
+ i + j − k)
0
= g4
(5.54)
in terms of (5.46). Therefore we obtain all diagonal and off-diagonal matrices (5.44) reflecting the fact that the finite Weyl groups of B4 and C4 are
isomorphic. Using the C4 highest root (5.53) we now define
θ
g̃i =
θ̄
gi
0
01
0
01
for i = 1, 2, 3, 4 ,
(5.55)
(0)
which is a unitary transformation of all the generators of P SL2 (H) defined in
(5.46) except for g0 . The matrices g̃i (i = 1, 2, 3, 4) also belong to W + (C4++ )
since θaθ̄b ≡ ab (mod C). Furthermore, since
θ 0 1 1 θ̄ 0
01
01
01
1 θ
=
01
.
(5.56)
we also have the unitarily transformed translation matrix. From this we conclude that the group generated by S0 , S1 , S2 , and S3 and S4 S3 S4 consists of
matrices of the type
44
θ 0 a b θ̄ 0
01
cd
01
a b
(0)
for
∈ P SL2 (H) .
cd
(5.57)
(0)
(0)
]
We denote this transformed (or ‘twisted’) P SL2 (H) by P
SL2 (H). In order
++
+
to describe the full W (C4 ) we still need to adjoin the matrix S4 which still
satisfies S42 = −11 ∼
= 11 so that we find
(0)
]
Proposition 11 W + (C4++ ) ∼
SL2 (H) ⋊ 2 .
=P
The semi-directness of the product can be verified by expanding the action of
(0)
(0)
]
]
SL2 (H).
S4 on P
SL2 (H) which can be seen to give an automorphism of P
(0)
(0)
]
Although P
SL2 (H) ∼
= P SL2 (H), it must be stressed that the Weyl groups of
C4++ and B4++ are nevertheless different, since for C4++ the extension S4 matrix
is not unitarily transformed, so that the group extension is different. 15 The
P SL2 (H) matrices (5.19) belonging to W + (C4++ ) can be described by saying
that either a, d ∈ H and b, c ∈ iO ·H, or conversely a, d ∈ iO ·H and b, c ∈ H. Again,
one can easily verify that this structure is preserved under multiplication.
Since the finite Weyl group satisfies W (C4) ∼
= W (B4 ), the relation to the
1
chiral group ± 6 [O × O] of the classification in [8] holds analogously to the last
section.
5.7 Even Weyl group W + (F4++ )
The root system of F4 consists of two copies of the D4 system, one of which
is rescaled. There are two short simple roots and two long simple roots. We
choose the simple roots as follows
ε1 = a1 = 1 ,
√
−i + j
ε3 = 2a3 = √
,
2
1
ε2 = a2 =
−1+i−j−k ,
2
√
−j + k
ε4 = 2a4 = √
;
(5.58)
2
the highest (long) root is
θ = 2a1 + 3a2 + 4a3 + 2a4 =
1
1−i−j−k .
2
(5.59)
This is similar to the affine Weyl groups W + (B4+ ) and W + (C4+ ) although both
are semi-direct products of isomorphic finite groups with abelian groups of the
same rank. However, the automorphism involved in the definition of the semi-direct
product is different.
15
45
✇
−1
✇
0
✇❅ ✇
✇
1
2
✇
3
4
Fig. 12. Dynkin diagram of F4++ with numbering of nodes.
The hyperbolic extension F4++ has the Dynkin diagram shown in fig. 12.
(0)
The even Weyl group W + (F4++ ) can be related to P SL2 (H) in the following
way: The generators g0 , g1 and g2 are identical to S0 , S1 and S2 constructed
from the F4 simple roots (5.58). Now consider
S3 S2 S3 =
1
(−1
2
0
1
(1
2
− i + j − k)
− i + j − k)
0
≡ g3
(5.60)
and
S4 S3 S2 S3 S4 =
0
1
(1
2
1
(−1
2
− i − j + k)
− i − j + k)
0
≡ g4 ,
(5.61)
(0)
in terms of (5.46), so that we find all P SL2 (H) generators in W + (F4++ ).
(0)
Adjoining to P SL2 (H) the generators S3 and S4 one obtains all of W + (F4++ ).
Since S3 and S4 together generate the symmetric group on three letters we
arrive at
(0)
Proposition 12 W + (F4++ ) ∼
= P SL2 (H) ⋊ S3 ∼
= P SL2 (H) ⋊ 2.
The last equality follows because the (diagonal) entries of S3 S4 are pure Hurwitz units, and both of order three, generating the cyclic group Z3 which
(0)
extends P SL2 (H) to P SL2 (H). The semi-directness of the product can be
verified explicitly by using the properties of ε3 and ε4 .
The finite Weyl group W + (F4 ) is now the chiral group ± 21 [O × O] in the
classification of [8]. Again it is not hard to see that the units (5.58) give rise
to the generating elements given there.
46
4
✇
−1
✇
0
✇
❅
❅
❅
✇
✇ ❅
❅✇
1
2
3
Fig. 13. Dynkin diagram of A++
with numbering of nodes.
4
5.8 Even Weyl group W + (A++
4 )
We discuss the algebra A++
last, because, somewhat surprisingly, it is the
4
most cumbersome to describe in the present framework among the rank 6
algebras. In order to embed its root system into the quaternions one needs to
make use of the icosians I introduced at the end of Section 5.2. The algebra
is simply laced, and we choose the following four units
A++
4
1
ε2 = a2 = (−1 + i − j − k) ,
2
1
ε4 = a4 = (i + σj + τ k) ,
2
ε1 = a1 = 1 ,
ε3 = a3 = −i ,
(5.62)
which have inner products corresponding to the A4 Cartan matrix. The highest
root is
1
θ = a1 + a2 + a3 + a4 = (1 − τ j − σk) .
2
(5.63)
The Dynkin diagram of A++
is displayed in fig. 13.
4
As a first step we identify the finite Weyl group W (A4 ) of order 120 in terms
of diagonal and off-diagonal matrices. The diagonal matrices can be generated from diag (εi , ε̄i ) for i = 2, 3, 4 and, modulo −11, one obtains all 60 even
elements of W (A4 ) in this way. This shows that the root system of A4 is
not multiplicatively closed. Closer inspection reveals that the diagonal and
off-diagonal matrices are of the form
a 0
0 a∗
and
0 a
,
−a∗ 0
(5.64)
where
a∗ = a′0 − a′1 i − a′3 j − a′2 k
for a = a0 + a1 i + a2 j + a3 k
47
(5.65)
√
√
with (p + 5q)′ := p − 5q for p, q ∈ Q. For icosians the conjugation in
√
5 amounts to interchanging τ and σ everywhere. The operation (5.65) is an
involutive automorphism of the ring I of icosians and satisfies in particular
√
(ab)∗ = a∗ b∗ . The automorphism is outer due to the ‘conjugation’ in 5;
however, the exchange of j and k together with the complex conjugation alone
is an inner automorphism of the quaternion algebra H, and is explicitly realized
by conjugation with the octahedral unit iO from (5.13).
Since there are 120 icosian units there are 60 matrices of either type in (5.64)
after quotienting out by {11, −11} in agreement with the order of the Weyl
group W (A4 ) ∼
= S5 . Moreover, the A4 root lattice is defined by the condition
QA4 = {z ∈ I z ∗ = z̄} .
(5.66)
Because conjugation reverses the order of factors but the ∗-automorphism
does not, we see from (5.66) again that QA4 is not multiplicatively closed and
therefore does not form a subring of I. QA4 is closed under complex conjugation. 16 The action of the even finite Weyl group on z ∈ QA4 is computed to
be z 7→ azā∗ , which manifestly preserves the root lattice.
The full W + (A++
4 ) is obtained by adjoining to (5.64) the matrix
1
θ
01
(5.67)
and the result is
(0)
∼
Proposition 13 W + (A++
4 ) = P SL2 (I).
(0)
Here, P SL2 (I) is defined as the discrete subgroup of P SL2 (I) satisfying the
conditions (3.27) with the four units ai from (5.62); the (0)-superscript indicates that additionally the Z2 outer automorphism of the A4 Dynkin diagram
has to be removed. This outer automorphism acts concretely as z 7→ −z on
z ∈ QA4 and the corresponding matrix is diag(1, −1) which has to be quotiented out. The explicit result of solving the conditions (3.27) and taking the
quotient for diagonal and off-diagonal matrices leaves the ones listed in (5.64).
For the general matrices S it leaves in particular the matrix (5.67).
1
¯
The finite Weyl group W + (A4 ) in terms of the classification of [8] is + 60
[I × I]
and again we see that there the generators given there can be reproduced from
the icosian units (5.62).
16
The fixed point set of ∗ on the set of icosian units is {1}.
48
(0)
One might hope to give a more natural description of P SL2 (I) as the kernel
(0)
of a group homomorphism analogous to Det used to define P SL2 (H), but
since the ideal CI generated by the commutators in I is not proper, this idea
fails.
5.9 Icosians and E8
Using icosians it is also possible to give an embedding of the E8 root system
into H at the cost of introducing a new inner product. This inner product is
defined by
(a, b)τ := x ,
if (a, b) = ab̄ + bā = x + τ y
(5.68)
for a, b, x, y being rational quaternions. The E8 simple roots are then given by
(see e.g. [35])
1
a1 = (−σ − τ i − k) ,
2
1
a3 = (+i − σj − τ k) ,
2
1
a5 = (+τ i + j − τ 2 k) ,
2
1
a7 = (+1 − τ 2 i − τ k) ,
2
1
a2 = (−σi − τ j + k) ,
2
1
a4 = (−τ i + j + τ 2 k) ,
2
1
a6 = (+i − τ 2 j + τ k) ,
2
1
a8 = (−i − σj + τ k) .
2
(5.69)
Note that τ 2 = τ + 1. Here, a4 , a5 , a6 and a7 belong to τ I constituting the
second half of the 240 E8 roots. The subspace spanned by integer linear combinations of the basis vectors (5.69) is dense in H (as one would expect, since
we are projecting an 8-dimensional lattice onto a 4-dimensional hyperplane in
such a way that distincts points remain distinct), and the same is, of course,
true for the ring generated by the above elements in H. In the following section
we will present and discuss an octonionic realization of the E8 root system,
and we will see that the non-associativity leads to drastic changes in the matrix realization of the Weyl group. For this reason, one might have hoped to
be able to avoid problems with non-associativity by working with the above
quaternionic realization, but unfortunately the difficulties remain, mainly due
to the modified inner product (5.68). Namely, the projection orthogonal to τ ,
which enters (5.68), invalidates our Theorem 1, and we therefore cannot use
the above representation to find a matrix representation of the E8 Weyl group.
For this reason we do not pursue this quaternionic realization of E8 further in
this paper.
49
6
Octonions K = O
In this section we turn to the non-associative octonions and their relation to
hyperbolic algebras of maximal rank 10. We restrict our attention mostly to
the case E8++ ≡ E10 and exhibit in this case the kind of new complications
arising due to the non-associativity of O. For the other two hyperbolic overextended algebras B8++ and D8++ we content ourselves with briefly presenting
the octonionic realizations of their root lattices and stating their relation to
that of E10 .
6.1 Octavians and the E8 lattice
We use the octonionic multiplication conventions of Coxeter [9]. Hence, the
seven imaginary units ei appearing in the expansion of an octonion z via
(e0 = 1)
z = n0 +
7
X
ni ei
(6.1)
i=1
multiply (in quaternionic subalgebras) according to
ei ei+1 ei+3 = −1 ,
(6.2)
where the indices are to be taken modulo seven and the relation is totally
antisymmetric. The multiplication table for the octonions can be neatly summarized by using the Fano plane, as shown in Fig. 14. The following notation
for the structure constants of the imaginary units can also be used
ei ej = −δij + fijk ek
(6.3)
where we adopt the usual summation convention from now on. The structure
constants fijk are totally antisymmetric, and satisfy (see e.g. [13] where many
further identities involving these structure constants can be found)
1
fijk f kmn = 2δijmn − ǫij mnrst frst ,
6
(6.4)
with δijmn := 21 δim δjn − 12 δin δjm .
We introduce a few concepts which are useful for studying the E8 and E8++
Weyl groups. The associator of octonions is defined by
50
1
1
0
0
1
11
00
4
00
11
00
11
11
00
6
00
11
00
11
7
1
0
0
1
0
1
1
0
0
1
3
0
21
1
0
05
1
Fig. 14. The Fano plane illustrating associative triples and octonionic multiplication.
[a, b, c] = a(bc) − (ab)c
(6.5)
and is totally antisymmetric in a, b, c by virtue of the alternativity of O. Put
differently, any subalgebra generated by two elements is associative. In components it reads
[a, b, c] = [a, b, c]i ei
1
with [a, b, c]i = − ǫi jklrstaj bk cl frst .
3
(6.6)
The Jacobi identity gets modified by an associator term to
[[a, b] , c] + [[b, c] , a] + [[c, a] , b] + 6 [a, b, c] = 0 .
(6.7)
The Moufang laws can be written as
x(yz) = (xyx)(x−1 z) ,
(yz)x = (yx−1)(xzx) .
(6.8)
They imply for example that a(xy)a = (ax)(ya). Similarly one can transform
axa−1 to show that this is an automorphism whenever a ∈ O satisfies a3 = ±1.
Another consequence of alternativity is that the following expressions are well
defined without brackets
axa = (ax)a = a(xa) ,
axa−1 = (ax)a−1 = a(xa−1 ) .
51
(6.9)
✇
✇
−1
✇
0
✇
1
✇
2
✇
3
✇
4
8
✇
5
✇
6
✇
7
Fig. 15. Dynkin diagram of E10 ≡ E8++ with numbering of nodes.
The first one implies the second one since a−1 = ā/N(a). These relations are
the basic reason why our formulas (3.7) with the generating matrices (3.6)
also make sense for octonions.
Coxeter (following Bruck) has established that there is a maximal order of
integers within O [9]. We briefly recall the construction from [8]. One starts
from the sets of all Hurwitz numbers for all associative triples (1, ei , ei+1 , ei+3 )
of fig. 14 and their complements, arriving at the ‘Kirmse numbers’. This is
not a maximal order but will be so after exchanging 1 with any of the seven
imaginary units; for definiteness we choose this unit as e3 . In addition, the lattice of integers is a scaled copy of the E8 root lattice. The following octonionic
units correspond to the simple roots of E8 (labeled as in fig. 15)
a1 = e3
,
a3 = e1
,
a5 = 1
,
a7 = e6
,
1
− e1 − e2 − e3 + e4
2
1
− 1 − e1 − e4 + e5
a4 =
2
1
a6 =
− 1 − e5 − e6 − e7
2
1
− 1 + e2 + e4 + e7 .
a8 =
2
a2 =
(6.10)
Their integer span gives all octonionic integers and, following [8], we call these
integers octavians and denote them by O. The highest E8 root is represented
by the unit
θ = 2a1 + 3a2 + 4a3 + 5a4 + 6a5 + 4a6 + 2a7 + 3a8
1
= e3 + e4 + e5 − e7 .
2
(6.11)
There is a total of 240 unit octavians corresponding to the 240 roots of E8 .
These definitions imply that no octavian contains only a quaternionic triple
or its complement in its expansion.
52
✇
✇
−1
✇
0
7
✇
1
✇
✇
2
✇
3
✇
4
✇
5
8
✇
6
Fig. 16. Dynkin diagram of DE10 ≡ D8++ with numbering of nodes.
6.2 D8 and B8 root systems
Before studying the E8 and E10 Weyl groups we give the octonionic realizations
of the root lattices of D8 and B8 which can be extended to the hyperbolic overextensions in the usual manner.
For D8 the simple roots are (as labeled in fig. 16)
D8 :
a7 =
a1 = e3
,
a3 = e1
,
a5 = 1
,
1
e2 − e3 + e6 − e7
2
,
1
− e1 − e2 − e3 + e4
2
1
a4 =
− 1 − e1 − e4 + e5
2
1
− 1 − e5 − e6 − e7
a6 =
2
1
a8 =
− 1 + e2 + e4 + e7 .
2
a2 =
(6.12)
Compared to the E8 simple roots of (6.10) the only difference is in the simple
root a7 which is a specific linear combination describing an embedding of D8
into E8 [39,27]. The highest root of D8 is given by
θD = 2a1 + 2a2 + 2a3 + 2a4 + 2a5 + a6 + a7 + a8
1
= e3 + e4 + e5 − e7 .
2
(6.13)
which is identical to the highest E8 root. This will make it easy to relate the
two hyperbolic Weyl groups.
For B8 the simple roots are (as labeled in fig. 17)
53
✇
✇
−1
✇
7
✇
0
1
✇
✇
2
3
✇
4
✇
5
✇❅ ✇
6
8
Fig. 17. Dynkin diagram of BE10 ≡ B8++ with numbering of nodes.
B8 :
a7 =
a1 = e3
,
a3 = e1
,
a5 = 1
,
1
e2 − e3 + e6 − e7
2
,
1
− e1 − e2 − e3 + e4
2
1
a4 =
− 1 − e1 − e4 + e5
2
1
− 1 − e5 − e6 − e7
a6 =
2
1
a8 = e2 + e4 + e5 + e6 + 2e7 .
4
a2 =
(6.14)
Compared to the D8 simple roots of (6.12) the only difference is in the simple
root a8 , which is no longer an octavian, in agreement with the fact that B8
is not a subalgebra of E8 . A specific linear combination of the B8 simple root
gives the D8 expressions, describing an embedding of D8 into B8 . The highest
root of B8 is given by
θB = 2a1 + 2a2 + 2a3 + 2a4 + 2a5 + 2a6 + a7 + 2a8
1
= e3 + e4 + e5 − e7 .
2
(6.15)
which is identical to the highest E8 and D8 roots.
6.3 Weyl group W (E8 )
It is known that the Weyl group of E8 is of order 696 729 600 and has the
structure [6]
W (E8 ) = 2 . O8+ (2) . 2 .
(6.16)
Since W (E8 ) naturally is a discrete subgroup of the continuous O(8; R) of
real dimension 28, we first record some facts about O(8; R). The space of unit
octonions is isomorphic to the seven sphere S 7 and hence of dimension 7. At
the level of Lie algebras one has [41,2]
so(8; R) ∼
= Im(O) ⊕ G2 ⊕ Im(O)
(6.17)
54
as a direct sum of vector spaces but not as Lie algebras. Here, Im(O) denotes
the purely imaginary octonions which constitute the tangent space to S 7 at the
octonion unit 1. The exceptional 14-dimensional Lie algebra G2 is isomorphic
to the algebra of derivations of the octonion algebra; the associated Lie group,
which we also denote by G2 , is the group of automorphism of the octonionic
multiplication table. The formula (6.17) suggests a representation of an arbitrary finite SO(8; R) transformation as constructed from two unit octonions
(aL , aR ) and a G2 automorphism γ. The following formula was proposed in
[37]
z 7→ (aL γ(z))aR
(6.18)
where
γ(z) = ḡ1 ḡ2 (g1 (g2 zḡ2 )ḡ1 )g2 g1
(6.19)
is also expressed in terms of unit octonions g1 and g2 . This formula reproduces
the correct infinitesimal so(8, R) transformations when all unit octonions are
close to the identity and expressed as aL = exp(αL ), etc., for some αL ∈ Im(O).
However, it is not true that any finite transformation is of this form as we will
show below, see also [33]. It is known that any SO(8; R) transformation can
be written as seven times iterated (left) multiplication of z by unit octonions
[8].
Turning to the finite E8 Weyl group we observe that a statement similar to
(6.17) is true for the integral octavians O. For this we first need some facts
about the discrete automorphism group of O.
The group Aut(O) is a finite group of order 12 096, usually denoted by G2 (2)
[8]. This group is not simple but has as simple part U3 (3) of order 6 048. One
obtains an explicit description of a generating set of U3 (3) in the following way.
As above, the transformation z 7→ azā (for a, z ∈ O) gives an automorphism
iff a3 ∈ R, i.e. a3 = ±1. Such a ∈ O were called Brandt transformers in [43].
Among the simple roots of E8 given in (6.10) this is true for the simple roots
a4 , a5 , a6 , a8 which constitute a D4 tree in the E8 diagram. The automorphisms
of O generated by z 7→ ai zāi (for i = 4, 5, 6, 8) generate in fact all of U3 (3).
In order to describe the index two extension needed for the full G2 (2) it is
instructive to consider the following four ‘diagonal’ automorphisms of z =
P7
i=0 ni ei (where e0 = 1)
(n0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 ) 7→ (n0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 ) ,
(n0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 ) 7→ (n0 , −n1 , −n2 , n3 , n4 , −n5 , n6 , −n7 ) ,
(n0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 ) 7→ (n0 , n1 , −n2 , n3 , −n4 , −n5 , −n6 , n7 ) ,
(n0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 ) 7→ (n0 , −n1 , n2 , n3 , −n4 , n5 , −n6 , −n7 ) ,
55
corresponding to choosing the three lines passing through unit e3 as associative
triples for the Dickson doubling process. 17 This suggests that one can obtain
another automorphism by choosing a line not passing through e3 , for example
(n0 , n1 , n2 , n3 , n4 , n5 , n6 , n7 ) 7→ (n0 , −n1 , −n2 , −n3 , n4 , n5 , −n6 , n7 ) .
This automorphism is a Z2 and one can check that it combines with U3 (3)
above to give all automorphisms of O. In this way one obtains eight diagonal automorphisms: one is the identity, while the other seven correspond to
treating any of the seven lines of the Fano plane as the quaternion subalgebra
entering the Dickson doubling process. 18
Combining Aut(O) with two sets of octavian units (of order 240), one finds
that there is a total of
240 × |Aut(O)| × 240 = 696 729 600 = |W (E8 )|
(6.20)
elements, coinciding with the order of the E8 Weyl group. This is the promised
discrete (and finite) version of (6.17). It is again tempting to associate the
action of this group on z ∈ O with the expression z 7→ (aL γ(z))aR , similar
to (6.18), in particular since (6.17) is invariant under (aL , aR ) ↔ (−aL , −aR )
thereby reducing the tentative number of transformation described in this way
to the order of the even Weyl group W + (E8 ), consistent with the fact that the
odd Weyl transformations will involve an additional conjugation of z. However,
the map (6.18) has a kernel in the discrete case in the sense that there are
non-trivial triples (aL , γ, aR ) which act trivially on z ∈ O: Consider a γ given
by a Brandt transformer a ∈ O with a 6= 1. Choosing aL = ā and aR = a leads
to z 7→ z by alternativity of O. There are 56 Brandt transformers among the
240 units of O and one can check that the kernel of the action (6.18) comes
solely from situations of the type just described. Since we do not have the
correct formula for any w ∈ W (E8 ) expressed in terms of unit octavians and
a G2 (2) transformation (also expressed in terms of unit octavians) we cannot
give a closed description of W (E8 ) using the set (6.20). This will also impede
our giving a fully explicit description of W (E8++ ).
We close this section by explaining why the formula for finite G2 automorphisms given in (6.19) is not correct for arbitrary unit octonions g1 and g2 .
Using the Moufang identities one can show that z 7→ g1 (g2 zḡ2 )ḡ1 is an automorphism iff g12 g23g1 ∈ R, so in particular whenever g1 and g2 are Brandt
17
Dickson doubled pairs (a, b) of quaternions manifestly have the automorphism
(a, b) 7→ (a, −b).
18 The Fano plane itself has the automorphism group P SL(2, 7) of order 168 and
there are several copies contained in Aut(O). None is generated as a subgroup of the
permutation group acting on the seven imaginary units.
56
transformers. Conjugating with ḡ1 ḡ2 to obtain (6.19) leads to the necessary
and sufficient criterion
{{ḡ1 , ḡ2 } , g2 } ∈ R
(6.21)
for (6.19) to be an automorphism. Here, {x, y} = xyx−1 y −1 is the (group)
commutator which is well defined by virtue of alternativity of O. The criterion
(6.21) is clearly not satisfied for arbitrary octavians; one example is given by
g1 = a2 and g2 = a1 in terms of the E8 simple roots. We note as a curiosity
that W + (E8 ) has a subgroup of type H4 which is non-crystallographic and
plays a role in the theory of quasi-crystals [35].
6.4 W + (E10 ) ≡ W + (E8++ )
As already remarked after equation (3.16) the formula SXS † for general even
Weyl transformations of the even hyperbolic Weyl group acting on Hermitian
(2 × 2) matrices ceases to be valid in the octonionic case. The even Weyl
transformations form a discrete subgroup of SO(9, 1; R) and similar to (6.17)
it is known that at the level of Lie algebras [41,2]
so(1, 9; R) ∼
= sl2 (O) ) ,
= L′2 (O) ⊕ Im(O) ⊕ G2 (∼
(6.22)
again as a sum vector spaces. Here, L′2 (O) denotes the 24-dimensional vector
space of all octonionic traceless (2 × 2) matrices. Combining this with Im(O)
one obtains the 31-dimensional vector space of all octonionic (2 × 2) matrices with vanishing real part of the trace. This is analogous to (5.23) in the
quaternionic case; the additional complication of non-associativity is that one
also requires an ‘intertwining’ G2 automorphism of O.
Similar to the suggestive, but incorrect, formula (6.18) one could envisage the
action of P SO(1, 9; R) on H2 (O) to be given by
X 7→ Sγ(X)S † .
(6.23)
However, this ‘definition’ is ambiguous since it requires a prescription for how
to place the parentheses in case there are more than two independent octonionic entries involved. Furthermore, we expect that in analogy with the finite
E8 case discussed above that (6.23) is not general enough to describe all transformations. Presumably both shortcomings of (6.23) are related and can be
resolved together. We stress nevertheless that Theorem 2 is still applicable
and any even Weyl transformation can be described by an iterated action of
(2 × 2) matrices.
57
Though we cannot offer a complete resolution to this problem we make a
comment on the issue of placing parentheses. In [34] the approach was taken
that S has to be such that there is no ambiguity when placing parentheses
in the matrix expressions. This leads to very restrictive conditions on S, allowing essentially only one octonionic entry, which is the situation covered
by Theorem 1. However, we are here interested in the case with more than
one independent octonionic entry. In order to make sense of (6.23), one must
presumably define the triple product by putting parentheses inside the matrix
elements in such a way that the resulting matrix is again Hermitian (simply
bracketing whole matrices will not do). This is for example required to reduce
(6.23) to (6.18) for the embedding of W (E8 ) in W (E10 ) for the cases when
(6.18) is correct.
Defining the Weyl group of E10 by iterated action of the basic matrices given
in Theorem 1 with the simple roots (6.10) (for which there arise no ambiguities
due to non-associativity) we arrive at
W + (E8++ ) ∼
= P SL2 (O) ,
(6.24)
where the group P SL2 (O) on the r.h.s. is defined by the iterated action. Since,
unlike the D4 Dynkin diagram, the E8 diagram has no outer automorphism no
additional quotients of P SL2 (O) are necessary for describing the Weyl group.
It is an outstanding problem to find a better and more ‘intrinsic’ definition of
the group P SL2 (O), and to explore its implications for an associated theory of
modular forms. We point out that P SL2 (O) has a rich structure of subgroups,
(0)
of types P SL2 (H), P SL2 (E), W (D9 ), and others, which remains to be exploited. We hope that the information on these subgroups we have obtained
in the preceding sections will help to find a better description of (6.24).
6.5 W + (D8++ ) and W + (B8++ )
We close by giving the relations between the Weyl groups W + (D8++ ) and
W + (B8++ ), and W + (D8++ ) and W + (E8++ ), respectively.
One can check that
S8D = S8B ◦ S6B ◦ S8B
and (S8D )2 = 11
(6.25)
by using either the abstract relations or iterated matrix action of octonionic
matrices on X; the ◦ is meant to indicate that the product is to be understood
as an iterated action. All other generators are identical and since S8B is acting
by a conjugation automorphism this shows that
58
Proposition 14 W + (B8++ ) ∼
= W + (D8++ ) ⋊ 2.
Turning to W + (D8++ ) one finds that S7D is the only generator which is not
common to W + (D8++ ) and W + (E8++ ) since all other simple roots and the
highest roots are identical. A calculation reveals that S7D can be expressed as
the result of a conjugation action of the W + (E8++ ) generators as
S7D = S7 S6 S5 S8 S4 S3 S5 S4 S6 S7 S5 S8 S6 S5 S4 S3 S2
◦S3 S4 S5 S6 S8 S5 S7 S6 S4 S5 S3 S4 S8 S5 S6 S7 ,
(6.26)
where we omitted a superscript on SiE and the ◦ symbols on the r.h.s. in order
not to clutter notation. The Weyl transformation (6.26) maps the simple root
α7E to α7D . This shows that W + (D8++ ) is a subgroup of W + (E10 ) but one can
check that it is not normal in W + (E10 ) since conjugation by S7E does not
preserve the subgroup. For the finite Weyl groups W (D8 ) has index 135 in
W (E8 ). By contrast, for the hyperbolic Weyl groups one has the following
result from [44,45]. 19
Proposition 15 W + (D8++ ) is an index 527 subgroup of W + (E8++ ) ∼
= P SL2 (O).
Without a direct definition of P SL2 (O) we cannot give a more detailed octonionic description of W + (D8++ ) that would be analogous to (5.24). There is
no direct relation between W + (B8++ ) and W + (E8++ ).
A
Examples with twisted affine algebras
In this appendix we show that our techniques are also suited for treating
hyperbolic algebras which do not arise as over-extensions but whose Dynkin
diagrams instead involve subdiagrams of twisted affine type. For simplicity we
will exemplify this for two cases constructed over the complex numbers with
(2)
(3)
twisted affine algebras D2 and D4 in the terminology of [25], both of which
have rank three. The hyperbolic node is attached to the ‘twisted affine’ node
with the single node for the cases we consider and we therefore denote the
(2)+
(3)+
associated hyperbolic algebras by D2
and D4 . However, we anticipate
that there are also other, more general hyperbolic cases which fit into the
picture we develop in this paper.
19
We thank N. Johnson and P. Tumarkin for bringing to our attention a mistake
in the published version of the following proposition.
59
θs
t
t
t
a2 ❅
■
❅
❅
❅
❅
❅
t
❅
✲t
❅r
a1
❅
❅
t
t
❅t
Fig. A.1. The root system of type B2 , with simple roots labeled and indicated by
arrows. The lattice they generate is the ring of Gaussian integers. There are both
long and short roots and we indicate the highest short root θs .
A.1
(2)+
K = C, type D2
(2)+
The hyperbolic Kac–Moody algebra D2
can be constructed from the root
system of type B2 . Although the root system is identical to that of C2 already
depicted in fig. 5 we draw it again in fig. A.1 where we highlight the short
(2)
highest root which enters the twisted affine extension D2 .
The root system is not simply laced, having simple roots whose squared lengths
are in the ratio 2 to 1. We choose the following units and simple roots for B2
in the complex plane
a2
−1 + i
ε1 = a1 = 1 , ε2 = √ = √
, θs = i .
2
2
(A.1)
These formulas are identical to
√ those of (4.21) except that the whole lattice
has been rescaled by a factor 2.
(2)+
The hyperbolic simple roots of D2
1
0
α−1 =
0 −1
α1 =
0
1
10
now take the form
,
−θ̄s 0
α2 =
−1 − i
,
−1 −θs
α0 =
0
−1
=
−1 + i
0
i
.
−i
0
,
(A.2)
(2)+
The Dynkin diagram of the hyperbolic algebra D2 is shown in fig. A.2. The
non-zero inner products between the simple roots are
(α−1 , α−1 ) = (α0 , α0 ) = (α1 , α1 ) = 2 , (α2 , α2 ) = 4 ,
(α−1 , α0 ) = −1 , (α0 , α2 ) = −2 , (α1 , α2 ) = −2
(2)+
and these give rise to the D2
diagram.
60
(A.3)
✇
Fig. A.2.
(2)+
D2
✇
−1
0
Fig. A.3.
✇❅ ✇
2
1
Dynkin diagram with numbering of nodes.
✇
(3)+
D4
❅
✇
−1
0
✇
2
❅
✇
1
Dynkin diagram with numbering of nodes.
a2
t
t
t
❏
✡
θs
❏
✡
t
❏
✡
✡
t
t ❏
❍❍
❨
❍
❍ ❏ ✡ ✟✟
✟
❍❍
❍
❏✡
✟
❍
❍❍
✟✟
✡❏
✟
✟
✡ ❏ ❍❍t
t
❏
✡
t ❏
✡
❏
✡
❏t
✡
t
✲t
a1
Fig. A.4. The root system of type G2 with highest short root indicated.
(2)+
Turning to the Weyl group of D2 , we realize that the Weyl group is isomorphic to that of C2++ since the Coxeter group is insensitive to the direction of
the arrows.
(2)+
Proposition 16 W + (D2
A.2
)∼
= W + (B2++ ) ∼
= P SL2 (G) ⋊ 2.
(3)+
K = C, type D4
The final hyperbolic algebra we consider is the hyperbolic extension of the
(3)
twisted affine algebra of type D4 (the only one twisted with an order three
(3)+
automorphism). We denote the hyperbolic extension by D4 and its diagram
is shown in fig. A.3. Its construction employs the G2 root system, shown again
in fig. A.4, and we make the following choices of units and simple roots in the
complex plane:
√
a1
− 3+i
ε1 = √ = 1 , ε2 = a2 =
, θs = i .
2
3
(A.4)
√
Again, these formulas are those of (4.30) except for a rescaling by a factor 3
and the use of the highest short root θs instead of the highest (long) root θ.
61
(3)+
The associated hyperbolic simple roots of D4
1 0
,
0 −1
√
0
3
α1 =
√
,
3 0
α−1 =
are then
−1 −θs
α0 =
−θ̄s 0
α2 =
0
√
− 3−i
2
=
0
√
− 3+i
2
−1 −i
i
0
,
.
(A.5)
The non-zero inner products between these simple roots are
(α−1 , α−1 ) = (α0 , α0 ) = (α2 , α2 ) = 2 , (α1 , α1 ) = 6 ,
(α−1 , α0 ) = −1 , (α0 , α2 ) = −1 , (α1 , α2 ) = −3 .
(A.6)
The Weyl group is again easy to determine in this case since it is isomorphic
to that of the standard G++
case.
2
(3)+
Proposition 17 W + (D4
∼
)∼
= W + (G++
2 ) = P SL2 (E) ⋊ 2.
References
[1] H. Aslaksen, Quaternionic determinants, Math. Int. 18 (1996), 57-65.
[2] J.C. Baez, The Octonions, Bull. Am. Math. Soc. 39 (2002) 145, Bull. Am.
Math. Soc. 42 (2005), 213 (Erratum), arXiv:math/0105155.
[3] O. Bergman, Three-pronged strings and 1/4 BPS states in N = 4 super-YangMills theory, Nucl. Phys. B 525 (1998), 104, arXiv:hep-th/9712211.
[4] J. Brown, O. Ganor and C. Helfgott, M theory and E10 : Billiards, branes and
imaginary roots, JHEP 0408: 063 (2004), arXiv:hep-th/0401053.
[5] M.C.N. Cheng and E. Verlinde, Wall crossing, discrete attractor flow and
Borcherds algebra, SIGMA 4 (2008), 068, 33 pages.
[6] J.H. Conway, R.T. Curtis, S.P. Norton, R.A. Parker and R.A. Wilson, Atlas
of finite groups, Clarendon Press, Oxford, 1985.
[7] J.H. Conway and N.J.A.Sloane, Sphere Packings, Lattices and Groups,
Springer-Verlag, New York, 1988.
[8] J.H. Conway and D.A. Smith, On quaternions and octonions, A K Peters,
Natick, MA, 2003.
[9] H.M.S. Coxeter, Integral Cayley numbers, Duke Math. J. 13 (1946), 561-578.
62
[10] A. Dabholkar, D. Gaiotto and S. Nampuri, Comments on the spectrum of CHL
dyons, JHEP 0801 (2008) 023, arXiv:hep-th/0702150.
[11] T. Damour, S. de Buyl, M. Henneaux and C. Schomblond, Einstein billiards
and overextensions of finite-dimensional simple Lie algebras, JHEP 0208 (2002)
030, arXiv:hep-th/0206125.
[12] T. Damour and H. Nicolai, Symmetries, singularities and the de-emergence of
space, Int. J. Mod. Phys. D 17 (34) (2008) 525531, arXiv:0705.2643.
[13] B. de Wit and H. Nicolai, The Parallelizing S 7 Torsion In Gauged N = 8
Supergravity, Nucl. Phys. B 231 (1984), 506.
[14] R. Dijkgraaf, E.P. Verlinde and H.L. Verlinde, Counting dyons in N = 4 string
theory, Nucl. Phys. B 484 (1997), 543, arXiv:hep-th/9607026.
[15] F.J. Dyson, Quaternion determinants, Helv. Phys. Acta 45 (1972), 289–302.
[16] M. Eie and A. Krieg, The Maass-space on the half-plane of Cayley numbers of
degree two, Math. Z. 210 (1992), 113.
[17] S. Elitzur, A. Giveon, D. Kutasov and E. Rabinovici, Algebraic aspects of matrix
theory on T d , Nucl. Phys. B 509 (1998), 122, arXiv:hep-th/9707217.
[18] V. Elser and N.J.A. Sloane, A highly symmetric four-dimensional crystal, J.
Phys. A 20 (1987), 6161–6168.
[19] F. Englert, L. Houart, A. Kleinschmidt, H. Nicolai and N. Tabti, An E9
multiplet of BPS states, JHEP 0705 (2007), 065, arXiv:hep-th/0703285.
[20] A.J. Feingold and I.B. Frenkel, A hyperbolic Kac–Moody algebra and the theory
of Siegel modular forms of genus 2, Math. Ann. 263 (1983), 87-144.
[21] A.J. Feingold, A. Kleinschmidt and H. Nicolai, Hyperbolic Weyl groups and the
four normed division algebras, in Vertex operator algebras and related
areas, Contemporary Mathematics, vol. 497, Amer. Math. Soc., Providence,
RI (2009), 53–64.
[22] P. Goddard and D. Olive, Algebras, Lattices, and Strings, in: Vertex
Operators in mathematics and physics, eds. J. Lepowsky et al., Springer,
Berlin, 1985.
[23] C. M. Hull and P. K. Townsend, Unity of superstring dualities, Nucl. Phys. B
438 (1995), 109, arXiv:hep-th/9410167.
[24] J.E. Humphreys, Reflection groups and Coxeter groups, Cambridge
University Press, Cambridge, 1997.
[25] V. Kac, Infinite dimensional Lie algebras, Cambridge University Press,
Cambridge, 1990.
[26] V. Kac, R.V. Moody and M. Wakimoto, On E10 , in Differential geometrical
methods in theoretical physics, eds. K. Bleuler and M. Werner, Kluwer,
1988.
63
[27] A. Kleinschmidt and H. Nicolai, E10 and SO(9, 9) invariant supergravity JHEP
0407 (2004), 041, arXiv:hep-th/0407101.
[28] M. Koca, R. Koc and M. Al-Barwani, Noncrystallographic Coxeter group H4
in E8 , J. Phys. A 34 (2001), 11201–11213.
[29] M. Koca, R. Koc and M. Al-Barwani, Quaternionic roots of SO(8), SO(9), F 4
and the related Weyl groups, J. Math. Phys. 44 (2003), 3123–3140.
[30] A. Krieg, Modular forms on half-spaces of quaternions, Lecture Notes in
Mathematics 1143, Springer, Berlin, 1985.
[31] A. Krieg, The Maass-space on the half-space of quaternions of degree 2, Math.
Ann. 276 (1987), 675.
[32] D. Lanphier and J. Rosenhouse, A decomposition theorem for Cayley graphs of
Picard group quotients, J. Comb. Math. Comb. Comput. 50 (2004), 95-104.
[33] C.A. Manogue and T. Dray, Octonionic Moebius transformations, Mod. Phys.
Lett. A 14 (1999), 1243, arXiv:math-ph/9905024.
[34] C.A. Manogue and J. Schray, Finite Lorentz transformations, automorphisms,
and division algebras, J. Math. Phys. 34 (1993), 3746, arXiv:hep-th/9302044.
[35] R.V. Moody and J. Patera, Quasicrystals and icosians, J. Phys. A: Math. Gen.
26 (1993), 2829–2853.
[36] N. A. Obers and B. Pioline, U-duality and M-theory, Phys. Rept. 318 (1999),
113, arXiv:hep-th/9809039.
[37] P. Ramond, Introduction to exceptional Lie groups and algebras, Caltech
preprint, CALT 68-577, 1977.
[38] I. Reiner, Maximal Orders, London Mathematical Society Monographs, Vol.
5, Academic Press, London 1975. Oxford University Press, reprinted 2003.
[39] I. Schnakenburg and P. West, Kac-Moody symmetries of ten-dimensional nonmaximal supergravity theories, JHEP 0405 (2004), 019, arXiv:hep-th/0401196.
[40] A. Sen, Walls of Marginal Stability and Dyon Spectrum in N=4 Supersymmetric
String Theories, JHEP 0705 (2007), 039, arXiv:hep-th/0702141.
[41] A. Sudbery, Division algebras, (pseudo)orthogonal groups and spinors, J. Phys.
A 17 (1984), 939-955.
[42] S. Viswanath, Embeddings of Hyperbolic Kac–Moody Algebras into E10 , Lett.
Math. Phys. 83 (2008), 139–148, arXiv:0801.2586 [math.QA].
[43] M. Zorn, The automorphisms of Cayley’s non-associative algebra, Proc. Nat.
Acad. Sc. 21 (1935), 355–358.
[44] N. W. Johnson, J. G. Ratcliffe, R. Kellerhals and S. T. Tschantz, The size of a
hyperbolic Coxeter simplex, Transform. Groups 4 (1999) 329–353.
[45] P. V. Tumarkin, Root subsystems of full rank in hyperbolic root systems, Sb.
Math. 195 (2004) 121–134.
64
| 4math.GR
|
A Genetic Algorithm for solving Quadratic
Assignment Problem (QAP)
H. Azarbonyada, R. Babazadehb
a
b
Department of Electrical and computers engineering, University of Tehran, Tehran, Iran, [email protected]
Department of industrial engineering, University of Tehran, Tehran, Iran, [email protected]
Abstract— The Quadratic Assignment Problem (QAP) is
one of the models used for the multi-row layout problem
with facilities of equal area. There are a set of n facilities
and a set of n locations. For each pair of locations,
a distance is specified and for each pair of facilities a
weight or flow is specified (e.g., the amount of supplies
transported between the two facilities). The problem is to
assign all facilities to different locations with the aim of
minimizing the sum of the distances multiplied by the
corresponding flows. The QAP is among the most difficult
NP-hard combinatorial optimization problems. Because of
this, this paper presents an efficient Genetic algorithm
(GA) to solve this problem in reasonable time. For
validation the proposed GA some examples are selected
from QAP library. The obtained results in reasonable time
show the efficiency of proposed GA.
Key words- Genetic Algorithm, QAP, Multi-row layout problems
optimization problems. The QAP is NP-hard optimization
problem [7], so, to practically solve the QAP one has to apply
heuristic algorithms which find very high quality solutions in
short
computation
time.
And
also
there
is
no
known algorithm for solving this problem in polynomial time,
and even small instances may require long computation time
[2].
The location of facilities with material flow between them was
first modeled as a QAP by Koopmans and Beckmann [3]. In a
facility layout problem in which there are n facilities to be
assigned to n given location. The QAP formulation requires an
equal number of facilities and locations. If there are fewer than
n, say m<n, facilities to be assigned to n locations, then to use
the QAP formulation, n-m dummy facilities should be created
and a zero flow between each of these and all others must be
1. INTRODUCTION
assigned ( including the other dummy facilities). If there are
Some multi-row layout problems are the control layout
problem, the machine layout problem in an automated
manufacturing system and office layout problem also
the Travelling Salesman Problem (TSP) may be seen as a
fewer locations than facilities, then the problem is infeasible
[4]. Considering previous works in QAP ([2],[8] and [9]) this
paper is developed to solve this problem with GA in short
computation time.
special case of QAP if one assumes that the flows connect all
facilities only along a single ring, all flows have the same non-
In the next section the QAP is described and formulated. After
zero
of
introducing GA in Section 3, the encoding scheme, solution
standard combinatorial optimization problems may be written
representation and GA operators are described in this section.
in this form (see [1] and [2]).
The computational results are reported in Section 4. Finally,
The Quadratic Assignment Problem (QAP) is one of the
Section 5 concludes this paper and offers some directions for
classical combinatorial optimization problems and is known
further research.
for its diverse applications and is widely regarded as one of
2. Mathematical model
(constant)
value.
Many
other
problems
the most difficult problem in classical combinatorial
The following notation is used in formulation of QAP [4]:
parameters
total number of facilities and locations
n
flow of material from facility I to facility k
fik
distance from location j to location l
djl
Variable
1 If facility I is assigned to location j
=
xij
0 Otherwise
individuals and happens in generations. In each generation, the
fitness of every individual in the population is evaluated,
multiple individuals are stochastically selected from the
current population (based on their fitness), and modified
(recombined and possibly randomly mutated) to form a new
population. The new population is then used in the next
iteration of the algorithm. Commonly, the algorithm
n
n
n
i
j
k
(1)
n
∑∑∑∑
Min
f ik d jl x ij x kl
terminates when either a maximum number of generations has
been produced, or a satisfactory fitness level has been reached
l
for the population. If the algorithm has terminated due to a
n
∑x
ij
=1
(2)
∀i
maximum number of generations, a satisfactory solution may
or may not have been reached [5].
j
n
∑x
ij
=1
(3)
∀j
i
(4)
xij is binary
3.1. Encoding scheme
In this section we describe the encoding scheme for different
component of proposed GA when n=5. For creating initial
The objective function (1) minimizes the total distances and
flows between facilities. Constraints (2) and (4) ensure that
each facility I is assigned to exactly one location. As well as
Constraints (3) and (4) assure that each location j has exactly
one facility which assigned to it. The term quadratic stems
population we consider a chromosome with n gene, as any
gene is depictive of assignment each facility to exactly one
location. For example the chromosome in Figure 1 show that
facilities 2,4,3,1 and 5 is assigned to locations 1,2,3,4 and 5
respectively that it is a feasible solution.
from the formulation of the QAP as an integer optimization
problem with a quadratic objective function.
3. Genetic Algorithm
2
4
3
1
5
1
2
3
4
5
Fig.1. chromosome representation
Genetic Algorithms (Gas) are routinely used to generate useful
For making above chromosome (feasible solution) to create
solutions to optimization and search problems. Genetic
initial population, firstly we assume that all content of
algorithms belong to the larger class of evolutionary
chromosome are placed equal to 1, then random integer
algorithms (EA), which generate solutions to optimization
number from 1 to 5 is produced, say that 3, and then it is
problems using techniques inspired by natural evolution. In a
compared with the content of corresponding square, so if the
genetic
(called
corresponding content is equal to 1, facility 3 is assigned to
chromosomes or the genotype of the genome), which encode
location 1 and then this content is set equal to zero (see figure
candidate solutions (called individuals, creatures) to an
2). Again a random integer number from 1 to 5 is produced,
optimization problem, evolves toward better solutions.
then If it is equal to 3 because of the corresponding content is
Traditionally, solutions are represented in binary as strings of
zero (it means that facility 3 is assigned) so, another number
0s and 1s, but other encodings are also possible. The evolution
should be produced. This way continues until all facilities are
algorithm,
a
population
of
strings
usually starts from a population of randomly generated
assigned to locations and also the determined number of initial
2
5
3
1
4
2
4
3
1
5
population of chromosomes is satisfied.
1
1
1
1
1
1
1
0
1
1
Fig.4. Mutation representation
3.1.3. Selection
Fig.2. making chromosome (feasible solution)
In selection operator the number of chromosomes required to
3.1.1. Crossover
complete next generation from crossover, mutation and
We use two point crossovers for implementation of GA. For
making child 1 and child 2, two points are randomly selected
from 1 to 5 and between these points is hold fixed. Then, child
1 is created according to orders of parent 2 and fixed
components of parent 1, as well as child 2 is produced
previous generation are determined special mechanism that we
use roulette wheel mechanism. In roulette wheel selection,
candidate solutions are given a probability of being selected
that is directly proportionate to their fitness. The selection
probability of a candidate solution I is
Fi
, where Fi is
P
∑
according to order of parent 1 and fixed components of parent
Fj
j
2 (see Figure 3). After tuning the crossover probability is
the fitness of chromosome I (objective function in QAP
determined equal to 0.8. This means that 80% of selected
formulation) and P is the population size.
parents participate in crossover operation.
3
5
2
4
1
5
3
4. Computational Results
PARENT 1
2
1
4
PARENT 2
For validation the proposed GA some examples are selected
from QAP library. The proposed GA was programmed in Java
net beans 6.9.1 on a Pentium dual-core 2.66 GHZ computer
with 4 GB RAM. The objective function values of the best
1
5
2
4
3
As it is illustrated in Table 1 the proposed GA results in
CHILD 1
3
2
4
known solutions are given by Burkard et al. [6].
5
1
the short computational time with acceptable GAP.
CHILD 2
Table 1: Computational results of proposed GA
FIG.3. CROSSOVER REPRESENTATION
3.1.2. Mutation
Mutation operator changes the value of each gene in a
chromosome with probability 0.2. For mutation two genes are
randomly selected from selected chromosome and then their
locations are substituted (see Figure 4).
Test Problems
Nug 12
Nug 17
Nug 20
Nug 24
Nug 28
chr12a.dat
chr12b.dat
chr15a.dat
Global/local
optimum
global
local
local
local
local
local
local
local
Gap%
0
.0034
0
.0034
.012
0
0
1
Run time
(second)
1
3
3
4
5
1
2
4
5. Conclusion
In this paper at first QAP as a special type of multi-row layout
problem with facilities of equal area is introduced and its
various applications in different fields are described. Then,
due to NP-hard nature of QAP an efficient GA algorithm is
developed to solve it in reasonable time. Finally, some
computational results from data set of QAP library are
provided to show the efficiency and capability of proposed
GA. Comparing outcome results with those works available in
literature justify the proficiency of proposed GA in achieving
acceptable solutions in reasonable time. For future research
direction, developing other heuristic and/or meta-heuristic
such as simulated annealing (SA) algorithm and PSO
algorithm and then comparison results of them with results of
this paper can be attractive work.
References
[1] Burkard R.E.(1984). Location with spatial interaction – Quadratic
assignment problem in discrete location theory edited by R.L.
Francis and P.B. Mirchandani. New York: Academic.
[2] Ravindra K. Ahuja, James B. Orlin, Ashish Tiwari (2000). A
greedy genetic algorithm for the quadratic assignment Problem.
Computers & Operations Research 27. 917-934.
[3] Koopmans T.C. and Beckmann M. (1957). Assignment problems
and the location of economic activities. Econometrica 25, 1:53-76.
[4] Heragu S. (1997). Facilities Design. Boston, PWS Publishing
company.
[5] Eiben, A. E. et al (1994). Genetic algorithms with multi-parent
recombination. PPSN III: Proceedings
of the International
Conference on Evolutionary Computation. The Third Conference on
Parallel Problem Solving from Nature: 78–87.
[6] Burkard RE, Karisch SE, Rendl F. (1997). QAPLIB – A quadratic
assignment program library. J. Global Optim.10:391- 403.
[7] S. Sahni and T. Gonzales, P-Complete Approximation Problems,
J. ACM, vol. 23, pp. 555-565, 1976.
[8] V. Maniezzo and A. Colorni. The Ant System Applied to the
Quadratic Assignment Problem. Accepted for publication in IEEE
Transactions on Knowledge and Data Engineering, 1999.
[9] P. Merz and B. Freisleben. A Genetic Local Search Approach to
the Quadratic Assignment Problem. In T. B¨ack, editor, Proceedings
of the Seventh International Conference on Genetic Algorithms
(ICGA’97), pages 465–472. Morgan Kaufmann, 1997.
| 9cs.NE
|
From output regulation theory to flatness based tracking:
a bridge for linear systems
arXiv:1612.01406v1 [cs.SY] 5 Dec 2016
Saman Khodaverdian
Institute of Automatic Control and Mechatronics, Control Methods and Robotics Lab, Technische
Universität Darmstadt, Landgraf-Georg-Str. 4, 64283 Darmstadt, Germany.
Email: [email protected]
The trajectory tracking problem for multivariable linear systems is considered. Two different techniques
are examined: the output regulation theory (ORT) and the flatness based tracking (FBT). ORT and FBT
are two different approaches to solve the tracking problem, and both methods have different restrictions.
The tracking controller of the ORT furthermore depends on the solution of the so-called regulator equations. In this paper, a special analytic solution of the regulator equations is presented. Additionally, based
on this analytic solution, a link from the ORT to the FBT approach is provided, and the connection of
both tracking controllers is highlighted. It is shown how the ORT controller can be converted to the FBT
controller and that both methods lead to identical control laws for a certain class of systems.
Keywords: trajectory tracking; output regulation theory; regulator equations; flat outputs
1.
Introduction
This paper considers the tracking problem for multivariable linear systems. Tracking means that
the system is controlled in a way such that its output follows a given reference trajectory. There is a
considerable amount of literature dealing with this problem, where the tracking problem is usually
solved by means of system inverses (Brockett & Mesarovic, 1965; Hirschorn, 1979; Sain & Massey,
1969; Silverman, 1969). Based on these results, there have been developed different methods to
solve the tracking problem.
A proven way is the flatness based tracking (FBT) approach (Fliess, Levine, Martin, & Rouchon,
1995, 1999), where a set of (virtual) outputs is searched such that the system states and inputs
can be completely characterized by this so-called flat output and its time derivatives. The tracking
problem can easily be solved for such flat outputs. However, the conversion to the real system
output might be problematic since for non-minimum phase systems the relation is determined
by an unstable differential equation.1 Nevertheless, the FBT is a common and widespread technique in control literature (see for instance (Fliess & Marquez, 2000; Hagenmeyer & Delaleau, 2003;
Hagenmeyer & Zeitz, 2004; Martin, Devasia, & Paden, 1996; Rothfuss, Rudolph, & Zeitz, 1996)
and the references therein).
Another approach to solve the tracking problem is provided by the output regulation theory
(ORT) which has been developed in (Francis, 1977; Francis & Wonham, 1976) for linear systems,
and was extended to the nonlinear case in (Isidori & Byrnes, 1990).2 The ORT achieves stable
tracking under mild assumptions (especially in the linear case), and bypasses the stability prob1
It
is
possible
to
overcome
this
problem,
for
instance,
with
approximate
tracking
techniques
(Benvenuti, Di Benedetto, & Grizzle, 1994; Hauser, Sastry, & Meyer, 1992) where non-minimum phase systems are approximated by minimum phase systems. However, this results in inexact tracking due to neglected terms. Another technique
leading to exact tracking is the stable inversion theory (Chen & Paden, 1996; Devasia, Chen, & Paden, 1996), but the solution
needs non-causal and possibly unbounded control inputs (Tomlin, Lygeros, Benvenuti, & Sastry, 1995).
2 An alternative name for the ORT is the robust servomechanism problem (Davison, 1976).
lems of non-minimum phase systems. The solution depends on the solvability of a set of equations,
the regulator equations, which is usually solved numerically. The drawback of the ORT is that the
reference trajectory must be given by a so-called exosystem, while in the FBT approach the reference trajectory can be any function that is sufficiently often differentiable. Thus, the ORT method is
restricted in the treatable class of reference trajectories. Like the FBT, the ORT is a common technique for the design of tracking controllers (the textbooks (Saberi, Stoorvogel, & Sannuti, 2000)
and (Huang, 2004) are recommendable for a detailed overview).
FBT and ORT are different approaches to solves the tracking problem, but in this paper it is
shown that under certain conditions both methods lead to identical control laws. First, a special
analytic solution of the regulator equations is presented which allows to interpret the tracking
controller of the ORT in a new light. Based on this analytic solution, it is shown that ORT and
FBT result in identical tracking controllers if the system output is flat, which also implies that in
case of flat system outputs no exosystem is needed. Additionally, it is shown that in case of non-flat
outputs, the ORT tracking controller can be converted to an FBT-like controller for a special choice
of the free parameters. As expected, this controller parametrization leads to an unstable closedloop behavior if the system is non-minimum phase. Note that the ORT approach has the ability
to achieve stable tracking even in case of non-minimum phase systems. But then, the reference
trajectory must be restricted to the class of exosystem trajectories, which can also be seen from
the derived solution of the regulator equations.
In the following, some notations are summarized, before the problem setup and basics are stated
in Section 2. The main result is presented in Section 3, where the single-input single-output (SISO)
case is considered first (Section 3.1), before the general multiple-input multiple-output (MIMO)
case is examined (Section 3.2). Finally, the results are summarized in Section 4.
Notation. Boldface letters denote vectors and matrices and italic letters represent scalar values.
The identity matrix of dimension n is written as I n , while 0 denotes a zero matrix or vector of
appropriate size. For a linear time-invariant system ẋ = Ax + Bu, y = Cx, the invariant zeros
are all numbers λ for which the matrix
λI n − A −B
R(λ) =
C
0
has a rank deficit (MacFarlane & Karcanias, 1976). R(λ) is known as the Rosenbrock’s system
matrix. The system is called minimum phase if the real part of all invariant zeros is negative,
otherwise it is called non-minimum phase. The relative degree of the k-th output states how often
yk must be differentiated with respect to time until the input u appears, i.e. δk = min{ν ≥ 1 :
ν
C k Aν−1 B 6= P
0} where C k is the k-th row vector of C. The relative degree of the system is then
given by δ = m
k=1 δk .
2.
Problem statement and preliminaries
The trajectory tracking problem for multivariable linear systems is considered (as depicted in
Figure 1) where the plant is described by
ẋ = Ax + Bu
y = Cx,
(1a)
(1b)
with state, input and output vector x ∈ Rn , u ∈ Rm and y ∈ Rm , respectively. It is assumed that
the plant is fully controllable and rank (B) = rank (C) = m. To avoid vastness, square systems
are considered (dim (u) = dim (y)), but the results hold also for systems with more inputs than
2
Trajectory
generator
Tracking
controller
u
y
Linear
plant
x
Figure 1. Basic setup for trajectory tracking.
outputs. The goal is to find a feedback controller such that for any initial condition the output of
the plant tracks a desired reference trajectory y d , i.e.
lim (y(t) − y d (t)) = 0.
(2)
t→∞
The desired output vector y d can be given by a trajectory generator as illustrated in Figure 1. Two
different approaches for solving the tracking problem (2) are summarized subsequently: the output
regulation theory (ORT) and the flatness based tracking (FBT).
2.1
Review of FBT
A system is called flat if there exists a virtual output z such that all input and state variables can
be expressed as functions of this output vector and its time derivatives, i.e. x = f x (z, ż, z̈, . . .) and
u = f u (z, ż, z̈, . . .) (Fliess et al., 1995, 1999). The virtual output z is called a flat output and does
not necessarily need to coincide with the real output y. Some basics regarding the FBT of linear
systems are summarized in the following. More details can be found in the mentioned references.
Proposition 1: A linear system (1) is flat if and only if it is controllable (Fliess et al., 1995).
Thus, for controllable linear systems there exists always a flat output which (in contrast to the
nonlinear case) can be found very easily: since (1) is assumed to be controllable, there exists a linear
e x which brings the system into controllable canonical form (Luenberger,
e = T
transformation x
e B,
e C)
e = (T
e AT
e −1 , T
e B, C T
e −1 ) are given
1967), meaning that the transformed system matrices (A,
by
e 11
A
A
e 21
e =
A
.
..
e m1
A
0 1
.. ..
e ii =
A
. .
0 0
⋆ ⋆
e ii = 0 · · ·
B
e 12 · · · A
e 1m
e 11
A
B
0
e 22 · · · A
e 2m
A
e =
,
B
.
..
.
..
..
..
.
.
e m2 · · · A
e mm
A
0
··· 0
0
..
..
..
. .
e ij =
A
.
∈ Rκi ×κi ,
0
· · · 1
i6=j
··· ⋆
⋆
T
e ij = 0
0 1 ∈ R1×κi ,
B
e 12 · · ·
B
e 22 · · ·
B
..
..
.
.
0
···
0
..
.
∈ Rκi ×κj ,
0
⋆
T
0 ⋆ ∈ R1×κi .
0 ···
.. . .
.
.
0 ···
⋆ ···
···
j>i
e 1m
B
e 2m
B
,
..
.
e mm
B
(3a)
(3b)
(3c)
P
Herein, the κi ’s are the controllablity (or Kronecker ) indices, with m
i=1 κi = n, and ⋆ denotes elee
ments that are not necessarily zero. The transformed output matrix C does not have any specified
structure. Obviously, a flat output is
z = z1 z2 · · ·
zm
T
e1 x
eκ1 +1 · · ·
= x
3
x
en−κm +1
T
(4)
since
(k)
with zi
h
= z1 ż1 · · ·
T h
x
eκ1 +κ2 = z2 ż2 · · ·
x
e1 x
e2 · · ·
x
eκ1 +1 x
eκ1 +2 · · ·
x
eκ1
T
(κ1 −1)
z1
(κ2 −1)
z2
..
.
x
en−κm +1 x
en−κm +2 · · ·
x
en
h
= zm żm · · ·
T
iT
(5a)
iT
(κ −1)
zm m
iT
(5b)
,
(5c)
= dk zi /dtk . By a further derivation, and with the relations (5), it can simply be seen that
(κ1 )
u = f u (z1 , ż1 , . . . , z1
(κm )
, . . . , zm , żm , . . . , zm
).
(6)
thatPfor the flat output z, we have κi = δi which follows from (3) and (4). Thus, δ =
PNote
m
δ
= m
i
i=1
i=1 κi = n which implies that with the flat output there is no zero dynamics (since
δ = n). Given a sufficiently often differentiable reference trajectory z d , a tracking controller for
the flat output can be designed such that lim (z(t) − z d (t)) = 0. Considering the error differential
t→∞
equations
(κ1 )
(z1
d (κ1 )
− z1
)+
κX
1 −1
(k)
d (k)
p1k (z1 − z1
)=0
(7a)
k=0
(κm )
(zm
−
d (κm )
zm
)+
κX
m −1
..
.
(k)
d (k)
pmk (zm
− zm
)=0
(7b)
k=0
and the relations (5) and (6), a state feedback controller of the form
e + F z d∆ ,
u = −K x
(8)
iT
h
d (κ )
d
d
with z d∆ = z1d ż1d · · · z1d (κ1 ) · · · zm
żm
· · · zm m , can be found which regulates the
flat output z to its desired trajectory z d . Herein, the controller parameters pik must be chosen
such that the dynamics (7) is stable.
Remark 1: Flat outputs are not unique since every virtual output for which δ = n holds is a flat
output (Martin et al., 1996).
With the described approach, we get a tracking controller for z, but the goal is to design a
tracking controller for the real output y. Note that the relationship from the virtual flat output to
the real system output is given by
h
e z ż · · ·
y=C
1
1
(κ1 −1)
z1
···
zm żm · · ·
(κ −1)
zm m
iT
.
(9)
The differential equations (9) can be used to calculate the virtual reference output z d from a
desired reference output y d . Then, the feedback controller (8) regulates the flat output z to z d ,
which implies that y will be regulated to y d .
4
The problem of the FBT approach is that (9) describes the zero dynamics which is unstable in
case of non-minimum phase systems. Thus, the solution of (9) might lead to unbounded reference
signals z d and, therefore, needs unbounded control inputs u. A common way to bypass this problem
is to approximate the virtual reference trajectory by a polynomial which fits the starting and ending
point of the desired trajectory (Fliess, Sira-Ramirez, & Marquez, 1998). However, this method is
only suitable for the change of the operating point along a specified trajectory, but not for the
tracking of, e.g., sinusoidal signals. In addition, this approximation results in inexact tracking.
Remark 2: An alternative way to handle non-minimum phase systems is provided by the stable
inversion theory (Chen & Paden, 1996; Devasia et al., 1996), but the obtained control input is
non-causal and can lead to large input signals (Tomlin et al., 1995).
For minimum phase systems, (9) describes stable zero dynamics and the calculation of z d is
unproblematic. In this case, the FBT approach provides an intuitive way to design a trajectory
tracking controller.
2.2
Review of ORT
For the ORT approach, it is assumed that the trajectory to be tracked is described by an autonomous linear system (the exosystem)
ω̇ = Sω
d
y = Qω,
(10a)
(10b)
with ω ∈ Rr , y d ∈ Rm . Like in the FBT approach, the goal is to design a feedback controller
u = −Kx + F ω
(11)
which leads to a stable closed-loop system and achieves output regulation in the sense of (2).
Without loss of generality, it is assumed that all eigenvalues of S have nonnegative real part
since eigenmodes with negative real part vanish asymptotically. According to (Francis, 1977;
Francis & Wonham, 1976), the following proposition is formulated.
Proposition 2: The output regulation problem is solvable if and only if there exists a matrix pair
(Π, Γ) solving the so-called regulator equations (sometimes called Francis equations)
ΠS = AΠ + BΓ
Q = CΠ.
(12a)
(12b)
Then, control law (11) achieves output regulation (or tracking) if K is chosen such that A − BK
is stable and F = KΠ + Γ.
Since Proposition 2 is a well known result, the proof is omitted and can be found for instance in
(Saberi et al., 2000).
Remark 3: The feedback matrix K in (11) must be determined such that A − BK is stable.
Thus, for the ORT approach it is sufficient to have a stabilizable system instead of a controllable
system.
The next proposition gives information about the solvability of the matrix equations (12).
5
Proposition 3: The regulator equations (12) are solvable if and only if
rank (R(λ)) = rank
λI n − A −B
C
0
=n+m
for every λ which is an eigenvalue of S.
This follows from a result presented in (Hautus, 1983). In this paper, an alternative proof will be
derived in Section 3. Note that R(λ) is the Rosenbrock’s system matrix. Therefore, the rank condition in Proposition 3 implies that none of the systems invariant zeros coincide with an eigenvalue
of S.
Summing up, we conclude that if the reference trajectory is described by an exosystem of the
form (10), tracking can be achieved by a feedback controller (11) if and only if the rank condition
in Proposition 3 holds.
It should be noted that the classical ORT can be used not only to track a desired reference
trajectory, but also to reject disturbances. This will be achieved if the exosystem is extended by
the disturbance dynamics that need to be rejected. However, since we are interested in output
tracking, the disturbance rejection problem is not considered in the following.
The problem of the ORT approach is that y d must be determined by an exosystem of the
form (10). This restricts the class of treatable reference trajectories. For the FBT approach, the
reference trajectory can be any function which is sufficiently often differentiable, however, with the
drawback that – in contrast to the ORT approach – it is not directly applicable to non-minimum
phase systems.
Remark 4: The important class of Bohl functions (combinations of polynomial, exponential and
sinusoidal signals) can be represented by exosystems (Trentelman, Stoorvogel, & Hautus, 2001).
3.
From ORT to FBT
In the following, an analytic solution of the regulator equations (12) is presented. Based on this
analytic solution, it is shown that the tracking controller of the ORT is identical to the tracking
controller of the FBT for systems with δ = n. Furthermore, the difficulties in case of δ < n
are highlighted, and it is shown how the controller parameters given from ORT and FBT are
correlated. For reasons of clarity, SISO systems are considered at first, before the general MIMO
case is examined.
3.1
SISO systems
Let system (1) be a SISO system with transfer function
g(s) =
b0 + b1 · s + . . . + b τ · s τ
,
a0 + a1 · s + . . . + an−1 · sn−1 + sn
τ < n.
(13)
Without loss of generality, it is assumed that the system matrices are given in controllable canonical
form, i.e.
0
1
···
..
..
..
.
.
A= .
0
0
···
−a0 −a1 · · ·
0
..
.
1
−an−1
,
6
0
..
B = . ,
0
1
b0
..
CT = . .
bτ
0
(14)
Then, the regulator equations (12) read
π1
0
1
···
π2
..
..
..
.
.
.. S = .
.
0
0
···
−a0 −a1 · · ·
πn
Q = b0 · · ·
bτ
0
..
.
π1
0
π 2 ..
.
.. + Γ
. 0
1
−an−1
πn
π1
π2
0 · · · 0 .. ,
.
(15a)
1
(15b)
πn
where the matrix Π is split into its row vectors. From the first n − 1 equations in (15a), it follows
that
π k = π 1 S k−1 ,
k = 2, . . . , n.
(16)
Taking this into account, the last equation in (15a) can be written as
Γ = π 1 (a0 I r + a1 S + . . . + an−1 S n−1 + S n ) = π 1 N .
{z
}
|
(17)
Q = π 1 (b0 I r + b1 S + . . . + bτ S τ ) = π 1 M .
|
{z
}
(18)
π 1 = QM −1
(19)
N
Furthermore, from (15b) and (16), we get
M
Note that the coefficients b0 , . . . , bτ are the coefficients of the zero polynomial of the plant (cf.
(13)). From the Cayley-Hamilton theorem (Gantmacher, 1959), it is clear that M is regular if and
only if no plant zero coincides with an eigenvalue of the exosystem. In this case, (18) yields
and, finally, the solution of the regulator equations reads
Π=
QM −1
QM −1 S
..
.
QM −1 S n−1
,
Γ = QM −1 N .
(20)
The above derivation is also a proof for Proposition 3.
Theorem 1: Given system (13) or (14) has no zero dynamics, i.e. δ = n, ORT and FBT lead to
the same control law.
Proof. First note that τ = 0 for δ = n, and y = b0 · x1 is a flat output. Thus, as described for the
FBT approach, taking the error dynamics
(y (n) − y d (n) ) + . . . + p1 (ẏ − ẏ d ) + p0 (y − y d ) = 0
7
(21)
and considering (14), we get
1
(p0 · y d + p1 · ẏ d + . . . + y d (n) )
b0
p1 · · · 1 y d∆ ,
{z
}
u = −k1 · x1 − k2 · x2 − . . . − kn · xn +
= − k1 k2 · · ·
|
{z
1
kn x +
p0
b0
}
|
K
(22)
F
T
with ki = pi−1 − ai−1 and y d∆ = y d ẏ d · · · y d (n) . The controller gain K and pre-filter F are
determined by the parameters pk which are the coefficients of the specified error dynamics (21).
In the ORT approach, the tracking controller is given by
u = −Kx + (KΠ + Γ)ω.
Furthermore, with (20) and M −1 =
Πω =
1
b0
1
b0 I r ,
Qω
QSω
..
.
(23)
we have
yd
ẏ d
..
.
1
=
b0
n−1
d
(n−1)
QS
ω
y
(24a)
1
Q(a0 I r + a1 S + . . . + an−1 S n−1 + S n )ω
b0
1
= (a0 · y d + a1 · ẏ d + . . . + y d (n) ).
b0
Γω =
(24b)
(24c)
Therefore, the tracking controller can be written as
u = − k1 k2 · · ·
{z
|
K
1
k1 + a0 k2 + a1 · · ·
kn x +
b
}
{z
|0
F
Substituting ki + ai−1 = pi−1 , (25) is exactly the same as (22).
1 y d∆ .
}
(25)
It should be noted that the coefficients of the closed-loop characteristic polynomial are ki + ai−1
(since the system is in controllable canonical form). Thus, choosing ki = pi−1 − ai−1 , the closedloop characteristic polynomial and the tracking error dynamics (21) are both determined by the
coefficients pi−1 .
The above result shows that for systems without zero dynamics, ORT and FBT lead to the same
control law. Thus, the following is a logical consequence.
Corollary 1: In the ORT approach, the reference trajectory does not need to be determined by an
exosystem of the form (10) if the plant has no zero dynamics.
For systems without zero dynamics, we only need to know y d and its time derivatives up to the
order y d (n) , like in the FBT approach. If there are zero dynamics, i.e. τ > 0, reformulating the
feedforward part of the tracking controller as in (24) is not possible and the exosystem matrices
S and Q appear in the control law (25). The next result presents how the ORT controller can be
converted into a FBT-like control law even if the plant is determined by zero dynamics.
Corollary 2: An exosystem formulation can be omitted if the plant zeros are canceled by closedloop poles.
8
Proof. Given that τ > 0 and taking (20) into account, the tracking controller (23) is written as
k1 I r + . . . + e
kn S n−1 + S n ω,
u = −Kx + QM −1 e
{z
}
|
(26)
KS
with e
ki = ki + ai−1 . Furthermore, the controller gains can be chosen such that
K S = k1 I r + . . . + kτ +1 S τ b
k1 I r + . . . + b
kδ S δ−1 +
1
k τ +1
Sδ ,
(27)
kj are the free parameters and δ = n − τ . Including (27) into (26), we see that for
where ki and b
ki = bi−1 the matrix M −1 cancels out, leading to
1
u = −Kx + Q b
k1 I r + . . . + b
kδ S δ−1 + S δ ω
bτ
1
d (δ)
d
d
(δ−1)
b
b
.
= −Kx + k1 · y + . . . + kδ · y
·y
+
bτ
(28a)
(28b)
Note that the choice ki = bi−1 implies pole-zero cancellations in the closed-loop system (A −
BK, B, C). However, (28) depends only on y d , ẏ d , . . . , y d (δ) and the exosystem can be omitted.
Corollary 2 is not surprising since the transfer function of the closed-loop system is given by
g(s) =
1
b0 + . . . + bτ · s τ
b0 + . . . + bτ · sτ
·
=
e
k1 + . . . + kτ +1 · sτ b
k1 + e
k2 · s + . . . + sn
k1 + . . . +
1
kτ +1
sδ
,
and choosing ki = bi−1 yields
g(s) =
1
b1 + . . . + k
bδ sδ−1 +
k
1 δ
bτ s
.
(29)
Transfer function (29) describes a zero free system with reduced dynamical order δ = n − τ for
which Theorem 1 is applicable.
It is clear that Corollary 2 is only useful for minimum phase systems since for non-minimum
phase systems the cancellation of zeros leads to an unstable behavior. Summarizing the results of
Corollary 1 and 2, we conclude that the ORT tracking controller can be converted into a FBT-like
control law. This also implies that for the ORT approach, the exosystem can be omitted in case of
minimum phase systems. Only y d and its time derivatives up to order y d (δ) are needed for output
tracking as in the FBT approach. It should be noted that for the ORT approach, the controller
gain does not need to be determined as in Corollary 2, but then the reference trajectory must be
provided by an exosystem. Thus, in contrast to the FBT approach, the ORT approach can handle
unstable zero dynamics, albeit at the expense of the treatable class of reference trajectories.
The results presented for SISO systems are extended to the MIMO case in the following.
9
3.2
MIMO systems
Since system (1) is controllable, we can transform its system matrices into the MIMO controllable
fx
eu
e+F
e such that
canonical form (3). In addition, it is always possible to choose u = −K
eF
eu
e −B
e K)
f x
e
e + |{z}
ė = (A
B
x
{z
}
|
′
B
A′
with
0 ···
A′11
0 A′
22 · · ·
A′ = ..
..
..
.
.
.
0
0 ···
0
..
A′ii = .
0
0
(30)
0 ···
B ′11
′
0
B
22 · · ·
′
, B = ..
..
..
.
.
.
0
0 ···
A′mm
0
1 ··· 0
..
..
.. . .
. .
.
B ′ii = . .
,
0
0 ··· 1
1
0 ··· 0
0
0
..
.
0
0
..
.
B ′mm
,
(31a)
(31b)
e has no
This means that system (30) is described by integrator chains of length κi . Note that C
specific structure. Since such input and state transformations are always possible for controllable
systems, in the following it is assumed that the MIMO system (1) is determined by
A11 · · ·
0
B 11 · · ·
..
.
.
..
.. x + ...
..
ẋ = .
.
0 · · · Amm
0 ···
C 11 · · · C 1m
.. x,
..
y = ...
.
.
C m1 · · · C mm
0
.. u
.
(32a)
B mm
(32b)
where Aii and B ii are given as in (31). (The symbols f and ′ are omitted for simplicity.)
Furthermore, we assume that the exosystem matrices are specified by
S1 · · ·
..
..
S= .
.
0 ···
0
..
.
Sm
and
Q1 · · ·
..
..
Q= .
.
0 ···
0
.. ,
.
Qm
(33)
i.e. every reference output is described by its own exosystem S k ∈ Rrk ×rk , Qk ∈ R1×rk . This
assumption is not restrictive since every output has (usually) its own reference trajectory.3
Partitioning the matrices Π and Γ in (12) as
Π11
..
.
Πm1
3 The
···
..
.
···
Π1m
..
.
and
Πmm
Γ11
..
.
Γm1
···
..
.
···
Γ1m
.. ,
.
Γmm
assumption can be relaxed for the subsequent steps, but it considerably simplifies the derivations.
10
(34)
it is not hard to see that
Πij S j = Aii Πij + B ii Γij ,
i, j ∈ {1, . . . , m}.
(35)
Note that Aii and B ii are given as in (31b). With the result presented in Section 3.1, it follows
immediately that
π 1ij
π 1ij S j
..
.
Πij =
κi −1
1
π ij S j
and
Γij = π 1ij S κj i .
(36)
Hence, the only unknowns are the row vectors π 1ij (i, j ∈ {1, . . . , m}), which can be determined
from the second regulator equation (12b):
h
Let C ij = c0ij
Q1 · · ·
..
..
.
.
0 ···
c1ij
···
Pm
···
0
l=1 C 1l Πl1
..
.. =
..
.
.
. P
m
Qm
l=1 C ml Πl1 · · ·
κ −1
cijj
Pm
l=1 C 1l Πlm
..
.
.
Pm
l=1 C ml Πlm
(37)
i
, the entries in the k-th column block of (37) are given by
m
X
π 1lk
κX
l −1
ν=0
l=1
|
cνil S νk
{z
=
}
M kil
(
Qk , i = k,
0,
i=
6 k.
(38)
The index in M kil ∈ Rrk ×rk denotes the l-th summand of the i-th block entry in the k-th column
block. Now, (38) can be rearranged to
with the square matrix
π 1mk M k = 0 · · ·
π 11k · · ·
M k11 · · ·
..
..
Mk = .
.
M k1m · · ·
Qk · · ·
0
M km1
.. ∈ Rmrk ×mrk .
.
(39a)
(39b)
M kmm
Thus, given that M k is regular, the solution reads
1
π 1k · · ·
π 1mk = 0 · · ·
Remark 5: A simple construction rule for M k is
M k = S k (C T ⊗ I rk )
Qk · · ·
0 M −1
k .
(40)
(41a)
11
where S k ∈ Rmrk ×nrk is defined as
I rk
..
Sk = .
0
S kκ1 −1 · · ·
..
..
.
.
Sk · · ·
..
.
0
···
0
..
.
···
0
I rk
0
..
.
···
Sk · · ·
0
..
.
S κk m −1
.
(41b)
From (40), it is clear that the solvability of the regulator equations depends on the invertibility
of the matrices M k , k = 1, . . . , m. Hence, we have to check the regularity of M k .
Proposition 4: M k is regular if and only if no invariant system zero coincides with an exosystem
eigenvalue, i.e. if rank (R(λ)) = n + m for every λ which is an eigenvalue of S.
The proof of Proposition 4 is given in Appendix A. The following theorem summarizes the results.
Theorem 2: Given a plant in canonical form (31)–(32) and let the exosystem be described as in
(33). Then, provided that M k (defined as in (41)) is regular, the solution of the regulator equations
is determined by (36) and (40).
e B,
e C)
e = (T
e AT
e −1 , T
e B, C T
e −1 ) into account, the solution of the
Remark 6: Taking (30) and (A,
regulator equations for the original system (1) is given by
e −1 Π,
Πorig = T
e Γ − KΠ.
f
Γorig = F
Theorem 2 provides an analytic solution of the regulator equations for MIMO systems in controllable canonical form. However, the canonical form (32) is not well suited to compare the ORT and
FBT methods. Therefore, given the system output is flat, an alternative solution will be presented
which is better suited to compare both methods. Let us consider the transformation matrix
C1
..
.
C 1 Aδ1 −1
..
T =
.
C
m
..
.
C m Aδm −1
∈ Rδ×n ,
(42)
where C k is the k-th row vector of C. Note that rank (T ) = δ (cf. Appendix B). Furthermore, assume that y is a flat output such that δ = n and T is a square matrix. Applying the transformation
(42), we get A′ = T AT −1 as in (31) and
C ′ = CT −1
′
C 11 · · ·
..
..
= .
.
0
···
0
.. ,
.
C ′mm
with C ′kk = 1 0 · · ·
0 .
(43)
Due to the definition of the relative degrees, the input matrix changes to
B′ = T B = 0 · · ·
(d⋆1 )T · · ·
0 ···
(d⋆m )T
T
,
(44)
with d⋆k = C k Aδk −1 B. It can easily be checked that A′ has the same integrator chain form as
12
before. The transformed system (A′ , B ′ , C ′ ) is suitable to prove the following claim.
Theorem 3: ORT and FBT approach lead to the same control law for systems with δ = n, i.e.
Theroem 1 also holds for MIMO systems.
Proof. First note that the output of (A′ , B ′ , C ′ ) is a flat output (since δ = n). Thus, as described
in Section 2.1, for the FBT controller we consider the error differential equations
(δ )
d (δ1 )
(y1 1 − y1
)+
δX
1 −1
(k)
d (k)
p1k (y1 − y1
)=0
(45a)
k=0
(δm )
(ym
−
d (δm )
ym
)
+
δX
m −1
..
.
(k)
d (k)
pmk (ym
− ym
) = 0,
(45b)
k=0
leading to
D ⋆ u − y dδ + P (x − y d∇ ) = 0,
(46)
with
h
iT
T
d (δm )
D ⋆ = (d⋆1 )T · · · (d⋆m )T , y dδ = y1d (δ1 ) · · · ym
h
iT
d (δm −1)
d
d
y d∇ = y1d ẏ1d · · · y1d (δ1 −1) · · · ym
,
ẏm · · · ym
p10 · · · p1δ1 −1 · · ·
0 ···
0
..
.
.
.. .
.
..
..
..
P = .
.
0 ···
0
· · · pm0 · · · pmδm −1
Therefore, the control law
u = −(D ⋆ )−1 P (x − y d∇ ) − y dδ
(47)
achieves the desired tracking behavior. Note that the inverse (D⋆ )−1 exists, which follows from
(44). Since T is a regular matrix and rank (B) = m, it is clear that rank (B ′ ) = m and, therefore,
rank (D⋆ ) = m.
Next, we consider the ORT approach. Apparently, for u = (D ⋆ )−1 u′ we get completely decoupled
SISO systems where the k-th SISO system is zero free with dynamical order δk (the k-th transfer
function reads yk (s)/u′k (s) = 1/sδk ). Thus, we can use the results of Section 3.1 and the ORT
tracking controller for the k-th SISO system is given by (cf. (25))
u′k = − pk0 · · ·
or after some rearrangements
u′k = −
pkδk −1 xk + pk0 · · ·
pk0 · · ·
pkδk −1 1 y d∆,k
d (δ )
pkδk −1 (xk − y d∇,k ) − yk k ,
13
(48)
(49)
with
iT
h
y d∆,k = (y d∇,k )T ykd (δk ) ,
iT
h
y d∇,k = ykd ẏkd · · · ykd (δk −1) ,
h
iT
Pk
·
·
·
x
xk = xPk−1
.
δ
+1
δ
j=1 j
j=1 j
Stacking the vectors together for all k = 1, . . . , m, i.e.
T
y d∇ = (y d∇,1 )T · · · (y d∇,m )T ,
T
x = xT
,
· · · xT
m
1
u′ is given by
u′ = − P (x − y d∇ ) − y dδ ,
and
u = (D ⋆ )−1 u′ = −(D ⋆ )−1 P (x − y d∇ ) − y dδ
(50)
which is identical to (47).
From the above prove, it is clear that Corollary 1 is also true in the MIMO case. Furthermore,
similar to Corollary 2, we have the following result for MIMO systems with δ < n.
Corollary 3: Given that D⋆ is invertible, the invariant zeros can be compensated by closed-loop
poles and an exosystem formulation can be omitted.
Proof. Let us define the matrix
T
tδ+1
T n = .. ∈ Rn×n ,
.
tn
(51)
where T is determined as in (42) and tδ+1 , . . . , tn are chosen such that T n is regular. For the
−1
transformed system An = T n AT −1
n , B n = T n B, C n = CT n , we get
A′ Aδ
An =
,
Aφ Aη
B′
,
Bn =
Bη
Cn = C′ 0 ,
with A′ ∈ Rδ×δ , B ′ ∈ Rδ×m , C ′ ∈ Rm×δ as before, and
Aδ,1
..
Aδ = . ∈ Rδ×(n−δ) ,
Aδ,m
Aδ,k
14
0 ···
..
= .
0 · · ·
⋆ ···
0
..
.
∈ Rδk ×(n−δ) .
0
⋆
(52)
The matrices Aφ ∈ R(n−δ)×δ , Aη ∈ R(n−δ)×(n−δ) and B η ∈ R(n−δ)×m do not have any specific
T
structure. Let us denote the last row of Aδ,k as A⋆δ,k and A⋆δ = (A⋆δ,1 )T · · · (A⋆δ,m )T . Then,
provided that D⋆ is invertible, we can design a feedback controller with gain K n = (D ⋆ )−1 0 A⋆δ
such that the closed-loop system matrix reads
A′ 0
An − B n K n =
,
Aφ A′η
with
A′η = Aη − B η (D ⋆ )−1 A⋆δ .
(53)
The dynamics A′η are not observable and represent the internal dynamics of the system. The
reduced system (A′ , B ′ , C ′ ) – with dynamical order δ – has no zero dynamics and Theorem 3
is directly applicable for it. This means that, like in the SISO case, we can compensate the zero
dynamics by an appropriate feedback and design the tracking controller without an exosystem.
Note that the eigenvalues of the internal dynamics A′η are the invariant zeros of the plant. If
there are invariant zeros in the closed right half of the complex plane, the controller K n leads to
an unstable closed-loop behavior.
Additionally to the minimum phase characteristics of the plant, there is another restriction on
the zero compensation approach in case of MIMO systems. For the compensation gain K n , it is
required that D⋆ is invertible. Thus, the compensation approach is only applicable for systems
with regular D⋆ .
Remark 7: For δ = n, D⋆ is regular as shown in the proof of Theorem 3, but for δ < n this must
not be true.
It should be noted that if the FBT approach is directly applied to systems with δ < n, this
results into the zero compensation method described above. Then, the error dynamics (45)–(46) or
the tracking controller (47) leads to an input-output decoupled system for which the internal zero
dynamics are made unobservable by pole-zero compensations. In this case, the regularity of D⋆
is a necessary (and sufficient) condition for the zero compensation or input-output decoupling (cf.
e.g. (Falb & Wolovich, 1967)), and leads to unstable closed-loop dynamics for non-minimum phase
systems. However, the ORT approach enables tracking even if the system is non-minimum phase
or if D ⋆ is singular. The only requirement is that the plant is stabilizable and that no invariant
zero coincides with an eigenvalue of the exosystem. It is not necessary to compensate the zero
dynamics, but then the reference trajectory must be provided by an exosystem.
4.
Conclusion
Two different approaches for the trajectory tracking problem of multivariable linear systems have
been considered: the output regulation theory (ORT) and the flatness based tracking (FBT). It has
been shown that both methods lead to identical tracking controllers if the output of the plant
is flat. Moreover, an analytic solution of the regulator equations has been derived which sheds
new light into the tracking controller of the ORT. Based on this solution, the connection to the
FBT approach has been discussed and it has been shown that an exosystem formulation for the
reference trajectory can be omitted if the controller parameters are chosen appropriately. In case of
non-minimum phase systems, however, this special choice leads to an unstable closed-loop behavior.
An interesting point for future research would be to extend the results of this paper to nonlinear
systems. One of the main difficulties is to find an analytic solution of the regulator equations which
in the nonlinear case are partial differential equations. Presumably, a solution in closed form may
be impossible to find.
15
References
Benvenuti, L., Di Benedetto, M. D., & Grizzle, J. W. (1994). Approximate output tracking for nonlinear
non-minimum phase systems with an application to flight control. International Journal of Robust and
Nonlinear Control, 4 (3), 397–414.
Brockett, R., & Mesarovic, M. D. (1965). The reproducibility of multivariable systems. Journal of Mathematical Analysis and Applications, 11, 548–563.
Chen, D., & Paden, B. (1996). Stable inversion of nonlinear non-minimum phase systems. International
Journal of Control, 64 (1), 81–97.
Davison, E. J. (1976). The robust control of a servomechanism problem for linear time-invariant multivariable
systems. IEEE Transactions on Automatic Control, AC-21 (1), 25–34.
Devasia, S., Chen, D., & Paden, B. (1996). Nonlinear inversion-based output tracking. IEEE Transactions
on Automatic Control, 41 (7), 930–942.
Falb, P. L., & Wolovich, W. (1967). Decoupling in the design and synthesis of multivariable control systems.
IEEE Transactions on Automatic Control, AC-12 (6), 651–659.
Fliess, M., Levine, J., Martin, P., & Rouchon, P. (1995). Flatness and defect of non-linear systems: introductory theory and examples. International Journal of Control, 61 (6), 1327–1361.
Fliess, M., Levine, J., Martin, P., & Rouchon, P. (1999). A Lie-Bäcklund approach to equivalence and flatness
of nonlinear systems. IEEE Transactions on Automatic Control, 44 (5), 922–937.
Fliess, M., & Marquez, R. (2000). Continuous-time linear predictive control and flatness: a module-theoretic
setting with examples. International Journal of Control, 73 (7), 606–623.
Fliess, M., Sira-Ramirez, H., & Marquez, R. (1998). Requlation of non-minimum phase outputs: a flatness
based approach. In D. Normand-Cyrot (Ed.), Perspectives in Control (pp. 143–163). London: Springer.
Francis, B. A. (1977). The linear multivariable regulator problem. SIAM Journal on Control and Optimization, 15 (3), 486–505.
Francis, B. A., & Wonham, W. M. (1976). The internal model principle of control theory. Automatica, 12 (5),
457–465.
Gantmacher, F. R. (1959). The theory of matrices. New York, NY: Chelsea.
Hagenmeyer, V., & Delaleau, E. (2003). Exact feedforward linearization based on differential flatness. International Journal of Control, 76 (6), 537–556.
Hagenmeyer, V., & Zeitz, M. (2004). Internal dynamics of flat nonlinear SISO systems with respect to a
non-flat output. Systems & Control Letters, 52 (3–4), 323–327.
Hauser, J., Sastry, S., & Meyer, G. (1992). Nonlinear control design for slightly non-minimum phase systems:
application to V/STOL aircraft. Automatica, 28 (4), 665–679.
Hautus, M. L. J. (1983). Linear matrix equations with applications to the regulator problem. In I. D. Landau
(Ed.), Outils et Modéles Mathèmatique pour l’Automatique (pp. 339–412). Paris: C.N.R.S.
Hirschorn, R. M. (1979). Invertibility of multivariable nonlinear control systems. IEEE Transactions on
Automatic Control, AC-24 (6), 855–865.
Huang, J. (2004). Nonlinear Output Regulation: Theory and Applications. Philadelphia: SIAM.
Isidori, A., & Byrnes, C. I. (1990). Output regulation of nonlinear systems. IEEE Transactions on Automatic
Control, 35 (2), 131–140.
Lohmann, B. (1991). Vollständige und teilweise Führungsentkopplung im Zustandsraum. PhD thesis, Universität Karlsruhe, Germany.
Luenberger, D. G. (1976). Canonical forms for linear multivariable systems. IEEE Transactions on Automatic
Control, 12 (3), 290–293.
MacFarlane, A. G. J., & Karcanias, N. (1976). Poles and zeros of linear multivariable systems: a survey of
the algebraic, geometric and complex-variable theory. International Journal of Control, 24 (1), 33–74.
Martin, P., Devasia, S., & Paden, B. (1996). A different look at output tracking: control of a VTOL aircraft.
Automatica, 32 (1), 101–107.
Rothfuss, R., Rudolph, J., & Zeitz, M. (1996). Flatness based control of a nonlinear chemical reactor model.
Automatica, 32 (10), 1433–1439.
Saberi, A., Stoorvogel, A. A., & Sannuti, P. (2000). Control of Linear Systems with Regulation and Input
Constraints. London: Springer.
Sain, M. K., Massey, J. L. (1969). Invertibility of linear time-invariant dynamical systems. IEEE Transactions
on Automatic Control, AC-14 (2), 141–149.
16
Silvermann, L. M. (1969). Inversion of multivariable linear systems. IEEE Transactions on Automatic Control, AC-14 (3), 270–276.
Tomlin, C., Lygeros, J., Benvenuti, L., & Sastry, S. (1995). Output tracking for a non-minimum phase
dynamic CTOL aircraft model. Proceedings of the 34th IEEE Conference on Decision and Control, New
Orleans, USA (pp. 1867–1872).
Trentelman, H. L., Stoorvogel, A. A., & Hautus, M. L. J. (2001). Control Theory for Linear Systems. London:
Springer.
Appendix A. Proof of Proposition 4
Let us define the matrices Ξk ∈ R(n−m)rk ×nrk and Ψk ∈ Rnrk ×(n−m)rk :
0 −I rk · · ·
0
..
..
..
..
(κ1 − 1) · rk rows
.
.
.
.
0
0
· · · −I rk
.
.
Ξk =
.
0
−I
·
·
·
0
r
k
.
.
.
..
..
..
..
(κm − 1) · rk rows
.
0
0
· · · −I rk
Sk
0
···
0
−I rk S k · · ·
0
.
.
..
..
−I rk
κ1 · rk rows
0
.
..
..
.
.
Sk
.
.
0
0
· · · −I rk
..
Ψk =
.
S
0
·
·
·
0
k
−I
S
·
·
·
0
rk
k
.
.
..
..
0
−I rk
κm · rk rows
.
.
.
..
..
..
S
k
0
0
· · · −I rk
where the empty entries are all zero. Then, the following equation holds
i
T
Ξk h
T
C
Φ
Ξ
k
k
k
=
S k |Ψk {zC k }
0
Mk
| {z }
Θk
with
Φ k = Ξk Ψ k =
I rk
0
..
.
0
Zk
−S k · · ·
..
.
I rk
..
..
.
.
···
···
17
0
..
.
∈ R(n−m)rk ×(n−m)rk .
−S k
I rk
(A1)
Note that Θk on the left-hand side of (A1) is regular since there is exactly one identity matrix in
every column and row block. Therefore, the regularity of the matrix on the right-hand side of (A1)
depends on the regularity of Z k . Obviously, Φk is a regular matrix, thus, M k is regular/singular
if Z k is regular/singular. However, Z k is regular if and only if no invariant system zero coincides
with an exosystem eigenvalue. To see this, consider the Rosenbrock’s system matrix
λ −1 · · ·
..
.
0 λ
..
..
..
.
.
.
0
R(λ) =
0
···
0
..
.
0
−1
λ
0
−1
0
..
..
.
λ −1 · · ·
..
.
0 λ
..
..
..
.
.
.
0
0
···
.
0
..
.
0
0
..
.
−1
λ
C
−1
0
.
(A2)
Expanding from the last m columns, it is clear that rank (R(λ)) = rank R(λ) , where R(λ) is the
P
matrix which results from deleting the last m columns and the rows ji=1 κi for all j = 1, . . . , m:
R(λ) =
λ −1 0 · · ·
0 λ −1 · · ·
..
..
..
..
.
.
.
.
0 0 ··· λ
0
0
..
.
−1
..
.
λ −1 0 · · ·
0 λ −1 · · ·
..
..
..
..
.
.
.
.
0 0 ··· λ
C
0
0
..
.
−1
.
(A3)
Note that R(λ) is singular for every λ which is an invariant zero of the system or, equivalently, the
invariant zeros are the roots of the polynomial P (λ) = det R(λ) . Comparing (A3) with Z k , it
T
can be seen that Z T
k is equal to (R(λ) ⊗ I rk ) if λ is replaced by S k , i.e. Z k = (R(λ) ⊗ I rk )|λ=S k .
Using determinant
and Kronecker
product rules, we know that the following identity holds:
det R(λ) ⊗ I rk = det R(λ) det (I rk ) = det R(λ) = P (λ). Thus, by the Cayley-Hamilton
theorem, the matrix P (S k ) = (P (λ) ⊗ I rk )|λ=S k is regular if and only if no invariant zero of the
plant coincides with an eigenvalue of S k . This implies that Z k is regular and since regularity of
M k is equivalent to regularity of Z k , Proposition 4 is proved.
18
Appendix B. Rank of transformation matrix T
It is shown that the rank of the matrix T in (42) is δ. The proof is inspired by (Lohmann, 1991)
and given as follows. Consider the equation
λij I n − A −B
C
0
v ij
0
=
,
wij
ei
i = 1, . . . , m,
j = 1, . . . , δi ,
(B1)
where ei is the unit vector with the i-th element equal to 1 and all other elements 0. The matrix on
the left-hand side of equation (B1) is the Rosenbrock’s system matrix. Thus, choosing λij different
from the invariant zeros, (B1) can be solved for v ij and wij . In this case, the second row block of
(B1) yields
(
1,
C k vij =
0,
k = i,
k 6= i.
(B2)
Multiplying the first row block of (B1) with C k , . . . , C k Aδk −2 leads to
(
λij , k = i,
C k Av ij + C k B wij =
| {z }
0,
k=
6 i,
(B3a)
0
..
.
(
δi −1
λij
, k = i,
C k Aδk −1 v ij + C k Aδk −2 B wij =
{z
}
|
0,
k=
6 i.
(B3b)
0
Eqs. (B2) and (B3) are written in a more compact way as
Ck
..
.
C k Aδk −1
v ij =
(
Λij , k = i,
0,
k=
6 i,
with
1
λij
..
.
Λij =
.
δi −1
λij
This procedure is repeated for all j = 1, . . . , δi , leading to
with Λi = Λi1 · · ·
that
Ck
..
.
CkA
v i1 · · · v iδi =
{z
}
δk −1 |
Vi
(
Λi ,
0,
k = i,
k 6= i,
(B4)
Λiδi . Writing (B4) for all k, i ∈ {1, . . . , m} in a matrix, it can easily be seen
Λ1 · · ·
.
..
T V 1 · · · V m = ..
.
{z
}
|
0 ···
V ∈Rn×δ
19
0
.. ∈ Rδ×δ ,
.
Λm
(B5)
where T is given as in (42). Note that the matrices
Λi =
1
λi1
..
.
···
···
λδi1i −1 · · ·
1
λiδi
..
.
δi −1
λiδ
i
∈ Rδi ×δi
are Vandermonde matrices which are regular if all the numbers λij are distinct. Thus, for distinct
λij , the matrix on the right-hand side of (B5) is regular. But this is only possible if the rank of
both matrices T and V on the left-hand side of (B5) is δ.
20
| 3cs.SY
|
Fast and Scalable Inference of Multi-Sample Cancer Lineages
Victoria Popic1 , Raheleh Salari1 , Iman Hajirasouliha1 , Dorna Kashef-Haghighi1 , Robert B West2 and
Serafim Batzoglou1*
1
2
Department of Computer Science, Stanford University, Stanford, CA
Department of Pathology, Stanford University School of Medicine, Stanford, CA
arXiv:1412.8574v1 [cs.CE] 30 Dec 2014
Abstract
Somatic variants can be used as lineage markers for the phylogenetic reconstruction of cancer evolution. Since
somatic phylogenetics is complicated by sample heterogeneity, novel specialized tree-building methods are required
for cancer phylogeny reconstruction. We present LICHeE (Lineage Inference for Cancer Heterogeneity and
Evolution), a novel method that automates the phylogenetic inference of cancer progression from multiple somatic
samples. LICHeE uses variant allele frequencies of SSNVs obtained by deep sequencing to reconstruct multi-sample
cell lineage trees and infer the subclonal composition of the samples. LICHeE is open-sourced and available at
http://viq854.github.io/lichee.
Background
Cancer is driven by the accumulation of somatic mutations that confer fitness advantages to the tumor cells.
Numerous studies have shown tumors to be highly heterogeneous, consisting of mixtures of cell subpopulations with distinct sets of somatic variants (for example see review papers [1, 2]). With the advent of nextgeneration sequencing technologies, many large-scale
efforts are underway to catalog the somatic mutational
events driving the progression of cancer [3, 4] and infer the phylogenetic relationships of tumor subclones.
Characterizing the heterogeneity and inferring tumor
phylogenies are key steps for developing targeted cancer therapies [5] and understanding the biology and
progression of cancer.
In order to reconstruct tumor phylogenies, studies
have utilized variant allele frequency (VAF) data of
somatic single nucleotide variants (SSNVs) obtained
by whole-genome [6, 7], exome [8], and targeted deep
sequencing [6, 9]. Clustering of SSNVs based on VAF
similarity [10–12] and detection of copy number aberrations, while accounting for variable sample purity
[8, 13, 14], have been used to differentiate and order
groups of mutational events. While many evolutionary studies of cancer have focused on single-sample
intra-tumor heterogeneity [15], several studies have
also compared multiple tumor samples extracted from
a single patient either at different points in time during cancer progression [16–18] or from different regions of the same tumor or its metastases [7,19–23]. In
multi-sample approaches, the patterns of SSNV sharing (i.e., distinguishing somatic mutations that are omnipresent, partially shared, or private among the samples) can serve as phylogenetic markers from which
lineage trees are reconstructed [24]. On the basis of
*
Correspondence: [email protected]
the lineage trees, the evolutionary timing of each mutational event can then be inferred with high confidence [7, 17, 19, 25].
Most existing multi-sample studies with a relatively
small number of SSNVs infer the tumor phylogenies
manually by analyzing SSNV VAFs and presence patterns across samples [7, 22, 26]. Several other studies used implementations of traditional phylogeny reconstruction methods, such as neighbor joining with
Pearson correlation distances [27], or maximum parsimony [21] on patterns of somatic mutational sharing
across samples. However, in order to scale to datasets
comprised of large numbers of samples per patient and
extract fine-grained SSNV timing information, as well
as handle sample heterogeneity, which traditional treebuilding techniques are not designed to do, specialized
computational approaches need to be developed for tumor cell lineage reconstruction. Two recent computational methods, SubcloneSeeker [28] and PhyloSub [29],
have been developed to address this need. The method
SubcloneSeeker requires as input clusters of variant cell
prevalence (CP) estimates that need to be obtained by
processing the data using other existing tools. Given
the CP clusters, the method generates all possible subclone structures in each sample separately. The persample solutions are then trimmed by checking their
compatibilities during the merge step. However, the
merge step is designed to check compatibilities of two
tumor samples only (e.g. relapse/primary tumor sample pairs that are common in clinical studies) and it
cannot merge the subclone structures of more than
two samples. Currently it only reports which sample
trees are compatible across the given pair of samples.
The method PhyloSub performs reasonably on samples with very few mutations that form simple (chain)
topologies; however, it produces unsatisfactory results
on larger multi-sample datasets, such as [21] (see Appendix A for details).
Page 2 of 18
In this work we introduce LICHeE (Lineage Inference
for Cancer Heterogeneity and Evolution), a novel
computational method for the reconstruction of multisample tumor phylogenies and tumor subclone decomposition from targeted deep sequencing SSNV
datasets. Given SSNV VAFs from multiple samples,
LICHeE finds the set of lineage trees that are consistent with the SSNV presence patterns and VAFs
within each sample and are valid under the cell division process. Given each such tree, LICHeE provides
estimates of the subclonal mixtures of the samples
by inferring sample heterogeneity simultaneously with
phylogenetic cell lineage tree reconstruction. LICHeE
is able to search for lineage trees very efficiently by incorporating the SSNVs into an evolutionary constraint
network that embeds all such trees and applying VAF
constraints to reduce the search space. LICHeE runs
in only a few seconds given hundreds of input SSNVs
and does not require data preprocessing.
We demonstrate that LICHeE is highly effective in
reconstructing the lineage trees and sample heterogeneity by evaluating it on simulated trees of heterogeneous cancer cell lineage evolution, as well as on three
recently published ultra-deep sequencing multi-sample
datasets of clear cell renal cell carcinoma (ccRCC) by
Gerlinger et. al [21], high-grade serous ovarian cancer
(HGSC) by Bashashati et. al [27], and breast cancer
xenoengraftment in immunodeficient mice by Eirew et.
al [30], for which single-cell validation results are also
available. LICHeE found unique trees for each ccRCC
patient and for all except one patient (for which multiple valid trees were found) of the HGSC study. For the
ccRCC dataset, LICHeE trees were nearly identical to
the published trees, which are the result of a multistep thorough analysis of the data, involving SSNV
calling, clustering, and tree-building using maximum
parsimony. For the HGSC dataset, LICHeE improved
on the results reported by the study, producing trees
with better support from the data. LICHeE also revealed additional heterogeneity in the samples of both
studies. In particular, LICHeE identified subclones in
one more sample of the ccRCC study (in addition to
the reported six samples) and three samples of the
HGSC study, all supported by the data. Finally, the
trees reconstructed by LICHeE on the xenoengraftment dataset can be highly-validated by the singlecell analysis presented in the paper. LICHeE is opensource and freely distributed at [31], and includes an
intuitive graphical user interface (GUI) that may aid
users to perform quality control on the output trees as
well as interpret the trees biologically.
Results and Discussion
Overview of the Multi-Sample Cancer Phylogeny
Inference Method, LICHeE
LICHeE is a method designed to reconstruct cancer
cell lineages using SSNVs from multiple related normal and tumor samples of individual cancer patients,
allowing for heterogeneity within each sample. Given
a set of validated deeply sequenced SSNVs, LICHeE
uses the presence patterns of SSNVs across samples
and their VAFs as lineage markers by relying on the
perfect phylogeny model [32]. This model assumes that
mutations do not recur independently in different cells;
hence, cells sharing the same mutation must have inherited it from a common ancestral cell. This assumption can be used to derive the following SSNV ordering
constraints. Firstly, (1) a mutation present in a given
set of samples cannot be a successor of a mutation
that is present in a smaller subset of these samples,
since it could not have arisen independently by chance
in the additional samples. Similarly, (2) a given mutation cannot have a VAF higher than that of its predecessor mutation (except due to CNVs), since all cells
containing this mutation will also contain the predecessor. Finally, (3) the sum of the VAFs of mutations
disjointly present in distinct subclones cannot exceed
the VAF of a common predecessor mutation present
in these subclones, since the subclones with the descendent mutations must contain the parent mutations
(this constraint is formally defined in the Methods section). These constraints provide key information about
the topology of the true cell lineage tree and are leveraged by LICHeE to define the search space of the possible underlying lineage trees and evaluate the validity
of the resulting topologies. The final goal of LICHeE
is to find phylogenetic trees encoding an evolutionary
ordering of the input SSNVs that does not violate any
of these three constraints.
At a high level, the LICHeE algorithm can be broken
down into the following main steps (Figure 1). First,
LICHeE partitions SSNVs into groups based on their
occurrence in each sample, such that each group stores
all the mutations that were called in the same subset
of samples. To separate subclone lineages, the SSNV
members of each resulting group are then further clustered based on their VAFs, such that SSNVs with similar VAFs across samples are clustered together. The
final lineage tree needs to provide a valid ordering of
these resulting SSNV clusters (i.e. an ordering that
does not violate the three constrains defined above).
In order to find such a tree, we construct an evolutionary constraint network, which encodes whether a given
cluster of SSNVs could have preceded another, for each
pair of clusters. More specifically, this network is an
acyclic directed graph (DAG) that has SSNV clusters
Page 3 of 18
✢✣✤ ✤❀❁
✦r✩ ✚❡s
✫❛r✩ss
s✫♠✦✚❡s
✢❂ ❃❃❃ ✢❄
❫❴
●❡r♠✚✛✜❡ ✢✣✤
▲②
✁✂
❙✄❙☛ ❙✑ ❙✙
✵✯✰
✢✢✣✤s ✥✦r❡s❡✜t ✛✜ s✧★s❡ts ✩✪ s✫♠✦✚❡s✱ ✜✩t ✛✜ ✚✬♠✦✭✮
✴✳
✲
✹✺✻✼③♦✽✻
❇✾✽✿✺✼
P✺♦ ❧✻
✄ ✄ ✄ ✄ ✄
✵ ✄ ✄ ✄ ✵
✵ ✄ ✵ ✄ ✄
✵ ✵ ✄ ✄ ✵
✵ ✵ ✵ ✵ ✄
❜❴
✢✢✣✤
●r✩✧✦s
❙✄ ❙☛ ❙✑
✢✧★✦✩✦✧❅
✚✫t✛✩✜
❛✚✧st❡rs
❵❴
✵✯✰
✞✝
✆☎
❱
❙✄
▼✻✿✽ ❆❈❉❑■❋ ❈❉❑❋ ❈❉❊❏
✵✯✰
❆❈❉❍❋❈❉❊❋❈❉❍❏
✎✏
☞✌✍
✵✯✰
✟✠✡ ❙☛
❙☛ ❙✑
❙✑ ❙✙
✵✯✰
✵✯✰
❆❈❉❊❋ ❈❉❊❋ ❈❉❍■❏
✞✝
✆☎
❱
❙✙
✵✯✰
✕✖
✒✓✔
✵✯✰
✟✠✡ ❙✑
✵✯✰
❆❈❉❊❋ ❈❉❍■❏
✆☎
✘✝
❱
✽♦✾♥✻
✶✷✸
✟✠✡ ❙✑
✗✝
✆☎
❱
❆❈❉❍❏
✵✄✄✄✵
❞❴
❖✈✩✚✧t✛✩✜✫r✬
❛✩✜str✫✛✜t
✜❡t✇✩r❦
✵✵✄✄✵
❣✻✺◆❧✾✽✻
❆❈❉■❋❈❉■❋❈❉■❋❈❉■❋❈❉■❏ ✄✄✄✄✄
✵✄✄✄✵
❚❯❲ ❝♦✽♥❳✺✿✾✽❳
❨✾♦❧✿❳✾♦✽
✵✵✵✵✄
✵✄✵✄✄
✢✫♠✦✚❡ ❛✩♠✦✩s✛t✛✩✜
❩✬♠✦✭
❢❴
❖✈✩✚✧t✛✩✜✫r✬
✚✛✜❡✫◗❡ tr❡❡
✫✜❘ s✫♠✦✚❡
❛✩♠✦✩s✛t✛✩✜
✢✫♠✦✚❡ ❬
✵✄✄✄✵
✄✄✄✄✄
✵✄✄✄✵
✢✫♠✦✚❡ ❭
✵✵✄✄✵
✢✫♠✦✚❡ ❪
✵✄✵✄✄
✵✵✵✵✄
✢✫♠✦✚❡ ❄
Figure 1 (a) Toy example of five samples (Lymph control and four tumor samples S1– S4) with one germline SNV and four SSNVs,
each associated with a binary sample profile. (b) SSNVs are partitioned into groups based on their binary profile (displayed groups
contain other SSNVs with varying VAFs). (c) Group SSNV are clustered based on their VAFs. (d) The clusters of each SSNV group
are incorporated into an evolutionary constraint network. An edge is placed between a parent and child node if, for each sample, the
VAF of the parent is greater than or equal that of the child node. For example, this constraint is violated for the green 01110 node
and node 00110. (e) The lineage tree is constructed from the constraint network. The leaves of the tree are the individual samples.
The horizontal line subdivisions in a sample indicate mixed lineages, separating the different subpopulations of cells in the sample.
The colors in each subdivision describe the mutation groups that the cells in this subpopulation have. Consider the orange node
01110, which denotes SSNVs of class 01110. Those SSNVs are found in samples 1, 2, and 3. After an ancestral cell division, the
daughter cells’ lineages accumulated SSNVs too (01110, green; 00110) that are now present in their descendant samples or
subclones. About 20% of sample 1 are cells that come from the orange and green lineage, and about 40% come from the blue
lineage. Samples 1, 2, and 3 grew from two or more subclones, whereas sample 4 only grew from one subclone.
as its nodes and whose edges encode possible predecessor relationships among the nodes’ mutations (i.e.
an edge denotes that the mutations of a given pair of
clusters satisfy ordering constraints (1) and (2)). This
network greatly reduces the search space of possible
valid trees and allows us to formulate the task of inferring such trees as a search for spanning trees of the
network that satisfy constraint (3) (within a given error margin), which ensures that the reconstructed trees
are composed of parent-daughter edges that exhibit
somatic VAF consistency with the cell lineage expansion. If multiple valid lineage trees are found during
the search, the trees are ranked based on how well
they support the cluster VAF data, the top-ranking
Page 4 of 18
tree minimizing the use of the permitted error margin
(see Methods for details). Finally, as shown in Figure
1, the leaves of the resulting trees are the individual
samples (added post-search), whose composition can
be reconstructed by tracing back their respective subclone cell lineages in the tree. We detail each of these
steps in Methods.
Since finding true SSNV groups is a crucial step of
the algorithm, it is important to minimize false positive and false negative SSNV calls across samples.
However, accurately detecting SSNVs in each sample
is a challenging task due to high levels of noise in
the data, which can come from various sources, such
as sequencing errors, systematic amplification bias,
mapping errors, and sample impurities. Multiple techniques have been developed to address this problem
to date [33–37], many employing a Bayesian approach
to model the distributions of noise and true genotypes
in matched-normal samples. LICHeE can work with
variant calls produced for each sample by any specialized existing method. However, it does not require the
users to pre-process the data using these tools, providing its own heuristic mechanism to call SSNVs using
the multi-sample VAF data. At a high level, it first
finds SSNVs that can be called reliably in each sample
using two hard thresholds Tpresent and Tabsent , above
and below which, respectively, the SSNV are considered robustly present or absent. Then, assuming that
the presence patterns of such SSNVs capture most of
the topology of the true underlying evolutionary tree,
it uses this inferred tree to inform the group assignment of the SSNVs whose VAF falls in between the
thresholds (the ’greyzone’).
Currently LICHeE does not automatically detect or
incorporate the CNVs explicitly into the model, although the method can still find valid phylogenies even
in the presence of SSNVs within such regions (for example, each patient in the ccRCC dataset had numerous variants from CNV regions). In order to address this limitation, LICHeE also accepts cell prevalence (CP) values instead of VAFs, which can be computed by several recently-developed tools, such as PyClone [12], ABSOLUTE [8], and ASCAT [38], and account for CNVs, LOH status, and sample purity. The
same algorithmic steps can then be directly applied to
CP values of each variant. It can be easily seen that
the three perfect phylogeny ordering constraints still
hold for CP values and can be used to search for the
underlying lineage tree. Furthermore, in order to support outputs of specialized clustering approaches (e.g.
PyClone [12]), LICHeE also accepts already computed
clusters of mutations (with given CP and VAF-based
centroid values) and uses these clusters as nodes of the
phylogenetic constraint network.
We evaluated LICHeE on three recently published
ultra-deep sequencing multi-sample datasets of clear
cell renal cell carcinoma (ccRCC) by Gerlinger et.
al [21], high-grade serous ovarian cancer (HGSC) by
Bashashati et. al [27], and breast cancer xenoengraftment by Eirew et. al [30], as well as on simulated
trees of heterogeneous cancer cell lineage evolution. On
the ccRCC dataset, LICHeE constructs near-identical
trees to the trees published in the study. We show
that the interesting difference in the topology of one
tree arises due to potential heterogeneity of a sample
that cannot be discovered using the traditional maximum parsimony approach and analyze when this approach can fail to detect existing sample subclones. For
each patient LICHeE finds a unique valid tree. On the
HGSC datasets we show that the trees generated by
LICHeE are better supported by the data and demonstrate why applying neighbor joining with Pearson correlation distance metric, used by the study, might not
be suitable for cancer datasets. LICHeE finds a unique
valid tree for all patients except Case5. Finally, we
show that the trees inferred by LICHeE on the xenoengraftment dataset are consistent with the single-cell
analysis done by the study. On each patient dataset
LICHeE takes only a few seconds to run.
Lineage Tree Reconstruction on ccRCC Data
The ccRCC study by Gerlinger et. al [21] validated
602 nonsynonymous nucleotide substitutions and indels from multiple samples of 8 individuals. It used
VAF-based clustering of each sample to detect subclones prior to determining the variant presence patterns used in tree reconstruction. The phylogenetic
trees were then reconstructed using maximum parsimony, revealing a branched pattern of ccRCC evolution in all tumors. Figure 2 juxtaposes the trees produced by LICHeE with the trees presented by Gerlinger et. al [21] (for details about the parameters
used by LICHeE and the ccRCC dataset see Appendix
B). The trees presented in the figure have been redrawn using AI. We can see that the trees generated
by LICHeE are topologically identical (consisting of
the same branches) to the published trees for patients
EV005, EV007, RMH002, RMH008, and RK26. Furthermore, LICHeE identified subclones in all the samples reported to be heterogeneous by the study. In particular, it identified the following regions as a mixture
of two subclones: R6 in EV005 (with frequencies of
0.29 and 0.04), R3 and R9 in EV007 (with frequencies
of 0.15 and 0.03 and 0.21 and 0.03, respectively), R4
and R6 in RMH008 (with frequencies of 0.19 and 0.14
and 0.21 and 0.15, respectively), and R5 in RK26 (with
frequencies of 0.15 and 0.03). These subclones correspond to the dominant (dom) and minor (min) shown
Page 5 of 18
GL
GL
6
27
4
21
3
2
1
2
1
R1 R2
R5
R3 R6
R4
1
1
R9
R7
RK26
1
2
6
R9 R10 R5 R8 R7 R6 R11 R2 R3 R1 R4
GL
GL
18
32
2
18
10
2
4
3
2
EV003
10
8
14
2
3
13
10
2
6
2
1
6
17
29
EV005
R1 R5
RMH004
R7 R6 R4 R3 R2
VT R10 R2 R4
GL
12
R8 R3
GL
50
22
10
8
5
2
1
EV006
4
4
R1 R2 R3 R4 R6 R7 R15 LN1a LN1b
RMH008
R1 R2 R3 R6
GL
R4 R5 R7
GL
19
Tumor region
2
2
8
Mixed lineage
9
8
2
5
3
Terminal branch
1
4
8
1
4
RMH002
R7 R5
R6
R8
Number of SSNVs
12
3
EV007
6
7
3
3
6
2
1
9
R1
R9 R2
R3
R4
R1 R2
R3
8
4
R7 R6
Internal branch
Trunk
Contribution link
Figure 2 Comparison of lineage trees for ccRCC patients 001-008. Phylogenetic trees obtained with LICHeE (on the left) are
contrasted with the trees published in [21] for each patient in the study.
in the published trees. Evidence supporting each subclone can be analyzed using the presented trees.
The trees generated for patients EV003 and EV006
also highly match the published results. For EV006,
the LICHeE-generated tree does not contain the following two partially shared groups: (R2, LN1a, and
LN1b) and (R3, R1, R4, R7, R15). For the first group
we find no evidence in the data – no SSNVs are shared
in these three samples and absent from the others. We
do find one mutation that supports the second group.
Because, by default, LICHeE eliminates nodes that
have evidence from only one SSNV, this group is not
shown in our tree.
The RMH004 dataset contains three partially overlapping groups (R3, VT, R10, R4, R2), (R3, VT, R10,
R4, R8), and (R10, R4, R2, R8). These groups represent separate branches where their lowest common
ancestor is the group with mutations present across
all samples (R3, VT, R10, R4, R2, R8). However, the
VAF of this parent group in sample R10 is 0.34, while
the average VAF across each of the three groups is 0.32,
0.27, and 0.21, respectively. Therefore, no more than
one of these groups can be a descendant of the parent
group, without violating the VAF phylogenetic constraint. In order to generate a valid tree, the two least
populated conflicting branches among the three are
removed from the dataset. Similarly, these two groups
are ignored by the maximum parsimony algorithm and
are not present in the published tree. The difference between our tree and the published one comes from the
mixed lineage we observe in sample R4. Due to the
VAF of 0.17 of the group of R4 private mutations, the
private group is not a descendant of the group shared
by R4 and R2 since the average VAF of that node is
0.06, which is too small to be a parent to 0.17. Therefore, our method suggests two different subclones in
R4. Since the group (R2, R4) is small (total of three
mutations) and has a low frequency in R4, the evidence of the two subclones cannot be considered very
strong; however, it does constitute a signal in the data
for the R4 mixed lineage possibility.
While the ccRCC study does perform VAF clustering of each sample to find subclones at the onset, it
runs the phylogenetic reconstruction using the presence pattern profiles only. On the other hand, our
tree reconstruction with LICHeE applies the VAF constraint to the resulting tree topologies and clusters
Page 6 of 18
SSNVs in each group based on their VAFs across all
the samples. For patient RMH004, applying the VAF
constraint, reveals additional sample heterogeneity, for
which we produced a potentially improved tree compared to the tree reported using maximum parsimony.
It can also be shown that clustering across all the samples rather than each sample individually, is a more appropriate approach for revealing the heterogeneity in
the data. For example, if two subclones occur with high
and low VAFs in one sample but are uniform in another sample, single sample clustering will detect them
only in the sample where they differ. For example, if
the two subclones in samples R3 and R9 of patient
EV007 (discussed above) had highly similar VAFs in
these samples, clustering would not be able to differentiate the subclones. On the other hand, LICHeE would
still be able to detect the mixed lineages in the two
samples due to the presence of groups (R1, R2, R3,
R5, R6, R7, R9) and (R3, R4, R9) in the data, which
must form divergent branches in the lineage tree.
Lineage Tree Reconstruction on HGSC Data
We further evaluated LICHeE on the HGSC dataset
from the study by Bashashati et. al. [27]. This study
validated 340 somatic mutations from 19 tumor samples of six patients. It used neighbor joining with
Pearson correlation distances (computed on the binary sample presence patterns) to infer lineage trees.
The trees generated by LICHeE juxtaposed with those
presented in the paper are shown in Figure 3 (for details about the parameters used by LICHeE and the
HGSC dataset see Appendix B). In four out of six
cases (Case2, Case3, Case 4, and Case6) the possible
tree topologies are very simple and the trees produced
by the two methods are unsurprisingly highly similar.
Below we discuss the difference between the two remaining trees.
For Case1 the tree reported in the paper suggests
that sample d diverged first, followed by c. This suggests that there should be a group of mutations shared
exclusively between samples a and b and another group
shared between samples a, b, and c. However, examining the dataset VAF values, as well as the results of the
binomial test determining SSNV presence in the samples used by this study, we found no evidence for these
two groups. On the other hand, the tree produced by
LICHeE suggests the presence of mutations shared by
samples a, b, and d only, which we found both in the
results of the study’s binomial test and by applying
the LICHeE hard threshold caller.
For Case5, the study reports an early divergence of
sample c. This suggests that there should be mutations
shared between all the samples except c. We have confirmed that no such mutational profile exists in the
data. On the other hand, the data shows the presence
of mutations that exist in samples a, b, c, e, and f but
not in d that cannot be supported by the reported tree.
The reason why the neighbor-joining algorithm of the
study chose sample c as the first diverging branch of
the tree must be because of the large presence of private mutations in sample c, which led to a low Pearson correlation (and hence a greater distance) with
the profiles of other samples. This shows that using
the Pearson correlation metric is not suitable for this
data. Furthermore, directly applying traditional phylogeny reconstruction techniques (e.g. neighbor joining) cannot reveal sample heterogeneity.
The tree produced by LICHeE for Case5, reveals
mixed lineages in samples b, e, and f. Interestingly,
LICHeE detects two clusters in group (a, b, c, e, f)
with mean VAFs of [0.29, 0.34 0.3, 0.19, 0.4] and
[0.17, 0.14, 0.13, 0.1, 0.13], respectively. The two subclones in each sample are then produced by the divergence of the higher VAF cluster into two branches:
one branch containing the lower VAF cluster and the
other branch containing the group of mutations (e, f,
b). While the presence of group (e, f, b) is weak (supported by three mutations only), the presence of group
(b, f) is substantial (supported by 14 mutations) and
provides strong evidence for the mixed lineage of b
and f. Furthermore, due to the presence of group (a,
b), the group (b, f) cannot be assigned as a child of
the low VAF cluster (a, b, c, e, f) without violating
the VAF phylogenetic constraint. However, since there
are only three mutations in group (a, b), eliminating this group may be a reasonable alternative reconstruction. Therefore, multiple topologies are possible
in this scenario, each supported by varying degrees of
evidence. LICHeE presents the tree that best fits the
data given input parameters of expected noise level
and minimum mutational support needed for a node.
Using LICHeE with different parameters and interacting with the trees can allow users to explore other evidence existing in the data. It is important to point
out that, as opposed to the results of the neighborjoining algorithm that produces trees with lack of evidential support in some branches, each branch in a
tree reported by LICHeE reflects the presence of the
corresponding SSNV groups in the data.
Lineage Tree Reconstruction on Xenoengraftment Data
The study by Eirew et. al. [30] used deep-genome and
single-cell sequencing to evaluate the clonal dynamics
of xenoengraftment of breast cancer tissue into immunodeficient mice. The study applied PyClone [12] to infer mutational clusters representing clonal genotypes
and then validated these results using single-cell analysis of two cases, SA494 and SA501. The study used
Page 7 of 18
a.
GL
Case 1
GL
Case 2
Case 3
GL
37
GL
Case 4
15
21
11
37
5
4
5
5
15
1
c
d
a
5
b
d
b
Case 5
5
9
1
c
a
a
c
d
e
f
g
h
i
j
Case 6
GL
8
13
Number of SSNVs
5
27
3
Tumor region
7
Mixed lineage
3
Terminal branch
1
6
b
b.
GL
14
a
c
b
Internal branch
Trunk
63
4
Contribution link
d
b
f
e
a
c
a
b
Figure 3 Comparison of lineage trees for HGSC cases 1-6. Phylogenetic trees obtained with LICHeE (Panel a) are contrasted with
the trees published in [27] (Panel b) for each patient in the study.
a.
Ancestral
B
27
A
16
C
D
b.
E
10
16
501X1
501X2
501X4
46
494T
5
494X4
Figure 4 LICHeE reconstruction on breast cancer xenoengraftment data [30]. a, Lineage tree generated by LICHeE (on left) is
juxtaposed with the phylogeny derived from single-cell data in the study (on right) for SA501. b, Lineage tree generated by LICHeE
for SA494.
Bayesian phylogenetic inference to reconstruct the evolutionary relationships between the single-cell nuclei.
We applied LICHeE to the deep-sequencing VAF data
of this study. We then compared the results of LICHeE
to the single-cell phylogenetic trees and clonal genotypes for both SA494 and SA501.
Single-cell phylogenetic inference of SA501 passages
X1, X2, and X4 reveals an ancestral genotype that
branched into two sibling clades A and B. Sequential acquisition of additional mutations in clade B then
gave rise to genotypes C, D, and E. Samples X1 and X2
were found to be a mixture of clones with genotypes A,
B, C, and D; while, sample X4 was found to be dominated by genotype E and did not contain clones with
genotype A (see Figure 2 in [30]). As can be seen in Figure 4a, the tree reconstructed by LICHeE mirrors the
phylogeny and sample compositions revealed by the
study. In particular, it presents the same two sibling
clades derived from the ancestral genotype. Samples
X1 and X2 both contain subclones with genotype A
Page 8 of 18
(with a higher percentage of this genotype in sample
X2, as confirmed by the single-cell analysis), while this
genotype is missing from sample X4. Similarly, genotype E is found to be private to and dominates sample
X4. The only difference in the LICHeE tree is the collapse of genotypes B and C into one cluster; however,
examining the bulk CP values reported by the study,
we can see that the CP values of these genotypes are
highly similar in the three given samples (X1, X2, and
X4) and, as a result, cannot get subdivided into two
separate clusters during the clustering step (note: the
PyClone analysis of the study was simultaneously performed on three additional samples T, X3, and X5,
which showed higher CP value differences).
The single-cell analysis of SA494 samples T and X4
reveals two clades, with mutually exclusive sets of mutations, emerging from an ancestral clone present in
both samples (see Extended Data Figure 3 in [30]).
LICHeE reconstructs the exact same topology (Figure 4b), showing two groups of mutations private to T
and X4, respectively. In accordance with the results reported by PyClone, LICHeE also finds two clusters of
mutations private to X4 (with the descendant cluster
present in about 20% of the sample).
Simulations
We developed a cancer cell lineage simulator to better asses the performance of LICHeE. Our simulator
models cancer evolution from normal tissue producing
a branching hierarchy of monoclonal cell populations
in accordance with the branched-tree cancer evolution
model [26,39,40]. Starting with the normal cell population, the simulator iteratively expands the cell lineage
tree by introducing (with some given probability) new
daughter cell populations corresponding to newly acquired SSNV or CNV events. In particular, in every
iteration, each cell population present in the tree can
give rise to a new population of cells (with a given
randomly generated size) representing a new SSNV
with probability PSSN V or a CNV event with probability PCN V . Each cell population can also undergo
a cell death event with probability PDeath . Each simulated SSNV is randomly associated with a genome location (chromosome and position) and haplotype; the
CNVs are associated with a chromosome arm and haplotype and correspond to a duplication of this chromosome arm. For the evaluation, we generated 100 lineage
trees, each expanded over 50 iterations, with the following parameters: PSSN V = 0.15; PCN V = 0, 0.1, and
0.18; and PDeath = 0.06. This process results in lineage
trees with an arbitrary number of branches and nodes
(several hundred to thousands of nodes on average).
Figure 5 illustrates one such lineage tree.
Multiple samples are then collected from each lineage tree. Each sample consists of several cell populations (nodes) of the tree, where each such cell population represents a subclone in the sample. We implemented two sampling schemes: randomized sampling and localized sampling. The randomized sampling process selects a random subset of nodes from
the tree for each sample; on the other hand, the localized sampling process is meant to mimic biopsies from
spatially distinct sites and selects nodes such that samples mostly contain subclones from distinct branches
of the simulated tree. Localized sampling for n samples is achieved by selecting n disjoint subtrees in the
simulated tree using an approach based on breadthfirst search, which finds disjoint subtree roots as high
on the tree as possible (if n disjoint subtrees do not
exist in the tree, the maximum number of subtrees is
used and some samples are obtained from the same
subtree in a round-robin fashion). Figure 5 illustrates
the localized sampling procedure for 10 samples.
Given a selected subset of cell populations, the sample is then created by obtaining a fraction of the cells
from each population by sampling from a multinomial distribution with probabilities corresponding to
the cell population sizes. For randomized sampling,
we select up to 5 subclones for each sample. For localized sampling, there can be up to 5 subclones from
the same distinct subtree, exactly one subclone from
a neighboring subtree, and some fraction of normal
cells to represent normal contamination (the fraction
is randomly selected to be from 0 up to 20% of the
sample). In order to determine how the performance
of LICHeE degrades as the number of SSNVs located
in CNV regions grows, we ran several experiments with
increasing numbers of CNV events. In particular, we
set PCN V = [0, 0.1, 0.18], which resulted in a total of
0%, 65%, and 80% of SSNVs to be affected by CNVs
in the collected samples, respectively. Given the sample cell population counts, we can compute the VAF
of each SSNV in this sample. In the absence of CNVs,
the VAF of a given SSNV is simply the fraction of
the cells containing the SSNV out of all the cells in
the sample. When CNVs are present, we count the
number of haplotypes containing the SSNV and the
reference allele across all cells of the sample. Given an
SSNV, we consider the chain of mutations present in
each cell population affecting its genome position. Let
HPv be the number of haplotypes containing the SSNV
in population P and HPr be the number of haplotypes
containing the reference allele. The VAF of a given
SSNV M is then:
PN
v
P ·HP
V AF (M ) = PN PP=1·(Hn v +H
r ,
P =1 n
P
P)
where N is the total number of selected populations
and Pn is the number of cells selected from population
P.
Page 9 of 18
ROOT
M10
M15
M35
CNV
M13
M23
M89
M79
M39
M101
M61
M33
M45
M83
M65
M95
M40
M106
M50 M54
M49
Sample 2
Sample 3
Sample 4
M100
M62 M118
M92
M103
M56
M67
M102
M81 M104
Sample 5
M73
M51
M60
M113
Sample 1
M42
M119
M8
M22
M25
M34
M93 M66
M110
Sample 7
M19
M38
CNV
M29
M71
M31
M37
M44
M107
M41
M75
M55
M82
Sample 6
M11
M17
CNV
CNV
M68
M7
M28
M74
M0
M1
M69
M26
M2
CNV
M5
M3
M30 M47 M59 M72
M9
M12
M16
M63
M6
M86
M14
M20
CNV
M36
M48
M85
M57
M32
M70 M18
CNV
M64
M80
M90
M91
M99
M108
M115
M117
M105
M88
CNV
M43
M27
M84
M24
M53
M46
M98
M78
M97
M76 M112 M116
M96
M94 M111
Sample 8
Sample 9
Figure 5 Example of a small simulated lineage tree (with SSNV and CNV events) with localized sampling (10 samples total
indicated by different colors). Each node represents a cell population that contains every mutation in its lineage. Gray indicates dead
cell populations.
Finally, given the true VAFs of each SSNV, we add
sampling and sequencing noise to each value. In particular, in order to simulate reads covering each SSNV
position, we sample the VAFs from the Binomial distribution B(n, p), where p is the true VAF of each SSNV
and n is the total simulated read coverage (100X,
1,000X, and 10,000X); the generated frequencies have
. We simulate a Q30 (1
a mean p and variance p(1−p)
n
in 1000) base call sequencing accuracy.
Given the simulated VAFs, we first assessed the performance of LICHeE in SSNV calling. Table 1 presents
LICHeE’s sensitivity in calling mutations across the
samples (i.e. the number of SSNVs with correctly identified sample presence patterns out of all the simulated
SSNVs) given true VAF values and for coverages of
100x, 1,000X, and 10,000X. As expected, higher coverage results in higher sensitivity. The method achieves
94 − 99% sensitivity across all the experiments.
Next we compared the topology of the reported
lineage trees and the simulated trees and measured
LICHeE’s accuracy in reconstructing the ancestordescendant and sibling relationships. For every pair
of mutations in the simulated tumor hierarchy, we
checked if LICHeE preserved the relationship between
them in the generated top tree. More specifically, we
checked for the following two types of violations: 1)
if an ancestor-descendant relationship was inverted or
became a sibling relationship, and 2) if a sibling relationship became an ancestor-descendant relationship.
Tables 2 and 3 summarize the accuracy of these metrics. Since LICHeE may remove nodes from the network representing non-robust SSNV groups (as described in Methods) if no valid trees are found during
the search, not all of the simulated mutations will be
present in the final tree. We report the percentage of
the SSNVs and of simulated ancestor-descendant (AD)
and sibling (Sib) mutation pairs that are present in the
tree. As expected, the number of mutations present in
the trees decreases with higher numbers of input samples, lower coverage, and the presence of CNVs (which
can significantly alter the VAF values, causing the violation of the VAF ordering constraints). In particular,
with a low coverage of 100X and 15 samples, only 81%
of SSNVs are preserved in the tree. Therefore, LICHeE
is best applied to data with higher coverage when the
number of input samples is high. For instance, with a
higher coverage of 1,000X, 91-94% of SSNVs and 8388% of mutation pairs are present in the trees with 15
samples (although this metric drops with the presence
of CNVs to 91-92% SSNVs and 66-70% SSNV pairs).
Since LICHeE groups mutations with the same presence patterns across samples and similar VAFs, only
the mutations occurring in a different set of samples
or with significantly different VAFs will be placed in
distinct nodes of the tree. We report the percentage of
such ancestor-descendant pairs (AD-Ord) in Tables 2
and 3. We then evaluate how many ordered mutations
preserved the correct ancestor-descendant relationship
(AD-Corr). Across all the experiments without CNVs,
we get 99-100% correctness (with less than 1% of pairs
being in reversed order). We see 92-96% correctness in
experiments with 80% of SSNVs located in CNV regions and 1000X coverage. AD mutations that were not
ordered or grouped in the same node, will be siblings
in the reconstructed tree (AD →Sib). We find that
Page 10 of 18
the vast majority of such mutation pairs involve private mutations, whose placement in the tree is usually
under-constrained (i.e. multiple tree nodes can serve
as ancestors to such mutation groups). Finally, we can
see that the reverse violation of sibling mutations being placed into ancestor-descendant nodes (Sib→AD)
is also very rare (up to 7% across all the experiments).
Therefore, we conclude that the trees reconstructed by
LICHeE provide highly accurate ordering of the mutations in its nodes.
Conclusion
LICHeE has been designed to automatically infer cell
lineages of multiple tumor samples and the sample subclone decomposition. Our analysis shows that LICHeE
is highly effective in reconstructing the phylogenies
and uncovering the heterogeneity of previously published datasets and in simulations, improving not only
upon traditional tree-building methods, but also on
recent developments specialized for cancer data. Currently LICHeE works with deep sequencing data that
provides VAF estimates with low variance (as well as
on CP values that can be obtained from existing tools
and can correct for SCNAs, LOH, and sample purity). SSNV data obtained from deep whole-genome
sequencing, targeted resequencing of informative SSNVs, or exome sequencing in tumors with a high degree of somatic SSNVs present in exomes, should be
appropriate inputs to LICHeE. Several directions for
future work are open: extension of this method to lower
coverage whole-genome sequencing data and incorporation of aneuploidies and large CNVs directly into the
model.
Methods
Grouping and Clustering SSNVs
When partitioning SSNVs into groups based on sample occurrence, each SSNV is first associated with a
binary sample profile denoting its presence or absence
in each sample. Given S samples from an individual,
the binary profile is defined as a binary sequence of
length S where the ith bit is set to 1 if this SSNV is
called in the ith sample, and is 0 otherwise (e.g. given
5 samples, an SSNV with the profile 01011 is called in
samples 2, 4, and 5). SSNVs with the same profile are
assigned to the same SSNV group (e.g. a group with
the profile 01101 will contain all the SSNVs occurring
in samples 2, 3, and 5). The group with the profile
consisting entirely of 1s will contain SSNVs that occur
in all the samples and are, therefore, germline variants (assuming that the sample set includes a normal
control sample).
The SSNV binary profiles can be passed as input or
computed from SSNV VAFs as follows. First two hard
VAF thresholds, Tpresent and Tabsent are used to determine if an SSNV is robustly present or absent from
a sample. An SSNV profile is classified as robust if its
VAF is above or below the two thresholds, respectively,
and if at least a minimum number of other robust SSNVs (default set to one) have the same binary profile. All other SSNVs are considered non-robust and
are assigned to a group as follows. Given a non-robust
SSNV m, its VAF across samples can either fall below
(marked 0), above (marked 1), or in between (marked
*) the thresholds Tpresent and Tabsent , resulting in a
profile such as 01*11. The candidate groups, to which
m can be assigned, must have an identical profile in
all the samples that are marked 0 or 1 (e.g. for profile 01*11 two valid candidate groups are 01111 and
01011). Since m can be assigned to more than one
target group, we consider that the group containing a
robust SSNV that is most similar in VAF to m is the
best candidate. The following metric is used to compute the similarity between two SSNVs m and n:
simmn =
X
i∈samples
min(m.VAFi , n.VAFi )
max(m.VAFi , n.VAFi )
,
(1)
where m.VAFi is the VAF of m in sample i. If the maximum similarity is higher than a given threshold, m
is assigned to the group of argmaxn∈Candidates simmn .
Unassigned non-robust SSNVs will form new profile
groups. We minimize the number of such new groups
by formulating this task as a set cover problem. In particular, let X be the set of all unassigned SNVs. Denote Y as the set of subsets of X, where each subset
represents mutations that can be assigned to the same
potential target group. The target groups are a list
of all possible binary profiles that the SSNVs can be
assigned to, obtained by substituting all *s by 0 or 1.
We want to find the minimum number of target groups
(i.e., smallest subset of Y ), which cover all mutations
in X. This problem is known to be NP-complete and
searching for the exact solution is not feasible. Instead
we apply the standard greedy algorithm by choosing
(at each stage) the target group that covers the largest
number of non-robust SSNVs. For SSNVs whose targets are not supported by any other mutations, we
convert each * to 1 or 0 depending on whether the
VAF is closest to Tpresent or Tabsent , respectively.
Once the SSNVs are partitioned into groups, the SSNVs of each group are further clustered based on their
sample VAFs. Each SSNV group is associated with a
matrix M of VAFs of size n × s, where n is the number
of SSNVs in this group and s is the number of represented samples (e.g. s = 3 for a group with the profile
0110001). The EM clustering algorithm for GMMs is
Page 11 of 18
GL
19 011111111 [-, .08, .29, .08, .17, .14, .18, .2, .2]
3
[-, -, -, -, -, -, .03, .21, .2]
8
000000111
001000010
[-, -, 0.33, -, -, -, -, .01, -]
1
000000010
[.14]
011111110 [-, .07, .27, .09, .18, .14, .15, .03, -]
000011100
8
3
000000100
[.1]
[-, -, -, -, .18, .15, .14, -, -]
[-, -, -, -, -, .11, .12, -, -]
000001100
1
8
001000000
[.22]
2
2
000010000
[.21]
Figure 6 Evolutionary constraint network for patient EV007 with 8 tumor samples and 55 SSNVs. Each node is associated with the
binary profile of its corresponding SSNV group, the VAF centroid, and the number of SSNVs assigned to it. The edges represent the
potential precedence relationships between the node SSNVs. The spanning tree reported for patient EV007 is highlighted (see
Results).
run on the resulting VAF matrix M using [41]. The result is a set of SSNV clusters with an associated VAF
~ . To handle the high variance in
centroid vector, V AF
the VAF data due to noise, some of the resulting clusters are eliminated (based on size) or collapsed with
neighboring clusters, based on the distance between
their centroid vectors.
Evolutionary Constraint Network Construction
Given the clusters of each SSNV group, we construct
an evolutionary constraint network to capture valid
evolutionary timing relationships between the mutations of each cluster pair. The network is a DAG, where
each node corresponds to an SSNV cluster (except the
root, which represents the germline) and each edge between two nodes, (u → v) denotes that parent node u
could be an evolutionary predecessor of child node v
(i.e. that SSNVs in cluster u could have ”happened
before” SSNVs in cluster v). In particular, an edge
(u → v) is added only if the nodes satisfy the following two constraints ∀i ∈ samples (which guarantee
that the network will be acyclic):
(1) u.VAFi ≥ v.VAFi − uv and
(2) if u.VAFi = 0, v.VAFi = 0 ,
(2)
where uv is the VAF noise error margin (note: the
error margin is the maximum of the sum of the standard errors for sample i of the two clusters and a configurable parameter). In the resulting network, each
node will have at least one parent (since all nodes can
be connected to the root). We avoid checking all node
pairs, by organizing the nodes into levels according to
the Hamming weight (i.e. number of 1s) of the binary
group profile to which they belong. Nodes that are in
the same level and have conflicting binary profiles cannot satisfy the above constraints. Nodes from different
levels can only be connected such that the node in
the higher level is the parent of the node in the lower
level. Finally, for nodes that are in the same level and
have the same binary profile, the edge is added in the
direction that minimizes VAFERR , where:
X
VAFERRu→v =
~1ERR · (v.VAFi − u.VAFi )2
i∈samples
(3)
with the indicator function:
1, v.VAFi > u.VAFi
~1ERR =
0, otherwise
.
(4)
Figure 6 illustrates the constraint network produced
for the dataset of ccRCC Patient EV007 (described in
the Results section).
Phylogenetic Tree Search
By the constraint network construction, a valid lineage
tree T of the SSNV clusters (i.e. a tree that does not
violate the three constraints of the perfect phylogeny
model) must be a spanning tree of the network that
satisfies the following requirement ∀ nodes u ∈ T :
X
∀i ∈ samples :
v
s.t.
v.VAFi ≤ u.VAFi + .
(u→v)∈T
(5)
That is, the sum of the VAF centroids of all the children must not exceed the centroid of the parent. We
Page 12 of 18
use inequality here since our method does not require
all the true lineage branches to have been observed.
To tolerate noise in the VAF data, we relax the constraint by allowing the sum of children VAFs to exceed
the parent by an error margin . Given such a valid lineage tree, each sample can then be decomposed into
subpopulations by enumerating all the paths in the
tree starting with the germline root and ending in the
last node containing mutations in that sample.
The problem of finding all such trees is equivalent to
the problem of finding all spanning trees of the constraint network DAG for which Eqn. (5) holds. We
have extended the Gabow and Myers spanning tree
search algorithm [42] to generate all such spanning
trees. The original algorithm generates all spanning
trees of a directed graph using backtracking and an
efficient bridge edge detection method based on DFS.
Our extension consists of enforcing Eqn. (5) during the
tree search by terminating the expansion of a given tree
and backtracking as soon as an edge violating Eqn. (5)
is added (see Algorithm 1). Same as the original algorithm, this search runs in O(|V | + |E| + |E|N ) time,
where |V | is the number of nodes, |E| is the number
of edges, and N is the number of spanning trees in the
network. While the runtime of the program depends
on the number of spanning trees in the constraint network, in practice the search is very fast (taking on the
order of a few seconds). However, since, in theory, it is
possible for the algorithm to take longer on datasets
that result in constraint networks with many spanning
trees, we provide a bound on the maximum number of
lineage trees to generate in order to avoid searches that
are too long. Similarly, we also have a high bound on
the number of calls to the GROW procedure in Algorithm 1. We expect this scenarios to be very rare in
typical validation datasets. To reduce search space we
also optionally constrain the placement of private mutation nodes in the constraint network to their closest
valid predecessors.
The above search algorithm will find all spanning
trees for which Eqn. (5) holds locally at each individual node; however, it is possible that in order to satisfy this constraint, the centroid values are deviated
(within ) in a globally inconsistent way. Therefore, in
order to enforce consistency, we apply one additional
requirement to the trees returned by the search. In
particular, we formulate the following quadratic programming (QP) problem to find a set of ev,i such that
for every sample i and node v:
minimize
subject
to
P
2
i ev,i
P P
v
s.t. (u→v)∈T v.VAFi + ev,i ≤ u.VAFi + eu,i and
|ev,i | ≤ , ev,i ≤ v.VAFi
v
Algorithm 1 Finding All Lineage Trees
1: Initialization: f ← emtpy list, L ← null // stores the last tree
output
2: procedure Lineage Tree Search(N ) // N is a constraint
network rooted at r
3:
Tree t ← new empty Tree
4:
t.addNode(r)
5:
add all edges (r → v) ∈ N to f
6:
grow(t)
7: procedure grow(t, N )
8:
if t contains all the nodes in N then
9:
L←t
10:
output L
11:
else
12:
s ← emtpy list
13:
b ← false
14:
while (not b and f not empty) do
15:
// e defined as (e.From → e.To)
16:
Edge e ← f.removeLast()
17:
Node v ← e.To
18:
t.addNode(v)
19:
t.addEdge(e.From → v)
20:
// ret. true if Eqn. (5) is satisfied for node e.From
21:
if t.checkConstraint(e.From) then
22:
add all edges (v → w), w 6∈ t to f
23:
remove all edges (w → v), w ∈ t from f
24:
grow(t)
25:
if number of returned trees > max trees return
26:
remove all edges (v → w), w 6∈ t from f
27:
add all edges (w → v), w ∈ t to f
28:
t.removeEdge(e.From → e.To)
29:
N.removeEdge(e.From → e.To)
30:
s.add(e)
31:
if ∃ an edge (w → v) s.t. w not a descendent of v
in L then
32:
b ← false
33:
else b ← true
34:
for all edges e starting from the end of s do
35:
remove e from s, add e to f , add e to N
(6)
If no solution exists, the tree is considered invalid.
Since multiple valid lineage trees can be generated, we
rank them using the resulting ev,i solution, which corresponds to how well each tree fits the VAF data. The
top-ranking tree will be thePtree
P with the minimum
sum of squared deviations: v i e2v,i . Since for certain networks the total number of valid lineage trees
can be large, it is impractical to run a QP program
on each tree. Therefore, we first rank the trees based
on the sum of squared deviations computed locally at
each node, and then run QP on the resulting top k
trees (k is 5 by default).
Multiple lineage trees can support a dataset equally
well since the placement of certain nodes in the tree
may be ambiguous using perfect phylogeny SSNV ordering constraints only (especially when not all lineage
branches have been observed). In these cases, more
sophisticated custom evaluation criteria would be required to rank the trees. We do not address defining
such criteria in this work. Instead we report all the
Page 13 of 18
SSNV node selected
ID: Binary Profile (Size)
Subclone B
Subclone A
Selected Sample
Selected Decomposition
(a) SSNV group selected
(b) Sample selected
Figure 7 LICHeE graphical user interface output for top lineage tree of ccRCC patient RK26. (a) Information is displayed about the
SSNV members of the selected node. (b) Information is displayed about the sample composition.
produced trees to the user ranked by their associated
score (the number of trees reported is configurable)
and provide a GUI to allow the user to easily explore
the differences in the topologies (see the Visualization
for details). We expect that under any optimality criteria, the sub-optimal trees can also represent signals
in the data of potential biological significance.
It is also possible that no valid lineage trees are found
during the search. This can happen if the noise error
margin is too narrow or the network contains nodes
corresponding to SSNV groups with a misclassified binary profile. In this case the network is adjusted and
searched again. Currently the network adjustment procedure will remove one-by-one all the nodes that belong to non-robust SSNV groups (smallest nodes first).
Other adjustments, such as increasing the noise error margin, are also possible but not currently implemented.
Visualization
The constraint network and the phylogenetic trees can
be visualized and interacted with in GUI form. The
JUNG graph library [43] is used to generate the resulting graphs. When visualizing lineage trees, each input
sample appears as a leaf in the tree and is connected
to the nodes that contain SSNVs present in the sample. By clicking on the nodes of a tree, it is possible
to obtain additional information about each node. The
information displayed about an internal (non-sample)
tree node consists of its binary group profile, its cluster
centroid and standard deviation vectors, and the list
of the SSNVs in its cluster (these SSNVs can be annotated with information from public databases such as
COSMIC, TCGA, etc). If a sample leaf node is clicked,
the information displayed consists of the lineage of this
specific sample obtained by doing a DFS traversal of
the tree starting with the germline root. Finally, the
user can rearrange the nodes in the tree, as well as
remove nodes and collapse nodes (provided they are
clusters of the same group). See Figure 7 for several
examples. In addition to the GUI, the program reports
the number of trees found and the score of the highestranking tree. The user can control how many trees to
display.
Implementation
The LICHeE algorithm was implemented in Java. It is
open-sourced and freely available online at [31].
Competing interests
SB is a founding advisor of DNAnexus Inc, and member of the Scientific
Advisory Boards of 23andMe and Eve Biomedical.
Author’s contributions
VP, RS, and SB developed the method. VP implemented the LICHeE
program and simulator and drafted the manuscript. RS, SB, DK, IH
contributed to the manuscript. VP, RS, SB performed the evaluation and
analysis. All authors were involved in discussions.
Acknowledgements
Authors would like to thank Arend Sidow for valuable discussions and
Aaron C. Abajian for contributing to the simulation studies.
VP was supported by the Stanford-KAUST grant. RS and IH were also
supported by Natural Sciences and Engineering Research Council of Canada
(NSERC) Postdoctoral Fellowships. DK was supported by an
STMicroelectronics Stanford Graduate Fellowship. This work was funded by
a grant from KAUST to SB.
References
1. Yates, L.R., Campbell, P.J.: Evolution of the cancer genome. Nature
Reviews Genetics 13(11), 795–806 (2012)
2. Marusyk, A., Almendro, V., Polyak, K.: Intra-tumour heterogeneity: a
looking glass for cancer? Nature Reviews Cancer 12(5), 323–334
(2012)
3. Bamford, S., Dawson, E., Forbes, S., Clements, J., Pettett, R., Dogan,
A., Flanagan, A., Teague, J., Futreal, P.A., Stratton, M., et al.: The
cosmic (catalogue of somatic mutations in cancer) database and
website. British journal of cancer 91(2), 355–358 (2004)
Coverage
# Sim SSNVs
True VAF (+CNV)
10,000X (+CNV)
1,000X (+CNV)
100X (+CNV)
26.6 (42.2)
99.2 (96.9)
98.9 (96.7)
99.1 (96.7)
97.2 (95.5)
5
25 (44.5)
99.5 (97.4)
99.1 (97.3)
99.3 (97.4)
97.4(95.9)
43.9 (89.7)
99.2 (97.8)
98.5 (97.5)
98.5 (97.7)
95 (96)
10
40.7 (88.6)
99.5 (97.5)
99.1 (97.3)
99 (97.3)
96.3 (95.9)
53.6 (131.1)
98.8 (96.9)
98.2 (96.7)
97.9 (96.5)
94.2 (95.4)
15
51 (136.1)
99.1(97.6)
98.7 (97.4)
98.5 (97.2)
94.5 (95.8)
Table 1 SSNV group assignment sensitivity on simulated data. Values indicate the number of correctly assigned SSNVs out of the total number of SSNVs collected in each experiment
(# Sim SSNVs). Results are shown for 5, 10, 15 samples given true VAFs, 10000X, 1000X, and 100X coverage data obtained with localized (top row) and randomized (bottom row)
sampling, from trees without CNVs and with 80% of SSNVs in CNV regions (in parentheses). All values are averaged over 100 simulated trees.
# Samples
Samples
Cov
Page 14 of 18
Trees
% SSNVs
% AD
% AD-Ord
% AD-Corr
% AD→Sib (-priv)
% Sib
% Sib-Corr
% Sib→AD (-priv)
94
98.9
99
41.5
100
30.2 (2.6)
98.2
83.2
7.3 (1.5)
t-VAF
98
99.5
99.5
40.4
100
26.3
99.2
84.8
6.6 (0.9)
95
98.8
98.7
40.4
99.9
30.5 (2.5)
97.8
82.5
7.7 (1.4)
10K
98
99.2
98.7
39
100
27.4
98.8
84.7
7 (1.3)
5
95
98.6
97.7
39.7
100.0
30.5 (2.3)
97.1
82.8
7.2 (1.3)
1K
98
99.4
99.5
38.1
99.7
29.2
98.9
83.9
7.5 (2.2)
97
96.8
93.2
38.7
99.8
31.4 (3.5)
93.1
81.8
7.9 (1.7)
100
97
97.3
94.8
35.2
99.7
30.1
95.3
83.7
7.2 (2.2)
97
97.5
96.2
58.8
100
43.1 (2.2)
96.9
94.4
3.3 (0.2)
t-VAF
95
97.7
96.5
58.2
100
38.1 (1.9)
96.7
93.5
4.2 (1.4)
98
96.5
94.4
57.3
100
44.9 (4)
95.2
94
3.6 (0.4)
10K
96
97
96.1
56.5
100
39.8 (2.3)
96.2
93.3
4.3 (1.2)
10
98
97.4
96.4
57.5
99.9
45.2 (3.9)
96.8
94.1
3.7 (0.4)
1K
96
97
95.7
57.4
99.9
38.7 (2.4)
96
92.7
4.8 (1.5)
93
90.8
75.7
52.3
99.7
38.6 (4.4)
78.0
93.9
3.4(0.7)
100
91
90.5
80.4
47.9
99.5
40.3 (5.7)
83.3
92.8
4.4 (1.5)
99
91.7
85
63.1
100
40.8 (2.2)
88
96.6
2 (0.2)
t-VAF
96
92.1
87.5
61.7
100
41.5 (2)
89.1
96.9
2 (0.3)
98
92.3
85.3
61.2
100
44.3 (3.5)
87.6
96.2
2.4 (0.3)
10K
100
90.6
83.6
59.1
100
41.8 (2.6)
85.8
96.6
2.1 (0.3)
15
93
94.7
83.5
61.5
99.9
42.7 (4.3)
85.4
96.1
2.6 (0.6)
1K
100
91.8
85.4
59.1
100
43.5 (2.9)
87.8
96.7
2.1 (0.3)
92
81.1
55.2
48.4
99.4
37.4 (5)
61.3
96.1
2.1 (0.3)
100
98
81.8
57.1
46.1
99.8
36.5 (2.2)
63.1
96.2
2.2 (0.4)
Table 2 Topological ancestry-descendant and sibling relationship reconstruction on simulated data. Results are shown for 5, 10, 15 samples given true VAFs, 10000X, 1000X, and
100X coverage data (without CNVs) obtained with localized (top row) and randomized (bottom row) sampling. All values are averaged over the number of reconstructed trees (Trees)
out of 100 simulated trees. The following metrics are presented: SSNVs present in the tree (% SSNVs), ancestor-descendant pairs of mutations in the tree (AD), ordered AD pairs
(AD-Ord), correctly ordered AD pairs (AD-Corr), unordered AD pairs reconstructed as siblings (AD→Sib) (with and without private mutation nodes), sibling pairs of mutations in the
tree (Sib), correctly reconstructed sibling pairs (Sib-Corr), sibling pairs reconstructed as AD (Sib→AD) (with and without private mutation nodes).
Samples
Trees
% SSNVs
% AD
% AD-Ord
% AD-Corr
% AD→Sib (-priv)
% Sib
% Sib-Corr
% Sib→AD (-priv)
95
96.2
89.1
14.4
93.7
20 (0.4)
92.6
80.8
2.4(0)
98
97.42
91.96
10.74
96.1
0.20
94.85
80.46
2 (0.02)
91
94.8
77.1
14.6
93.5
24.9 (0.6)
90.9
91.8
1.3(0)
10
92
94.55
78.97
11.95
93.1
0.41
90.5
91.7
1.1 (0.01)
97
91.8
65.9
11.1
92.5
23.3 (0.4)
85.5
94.6
0.7 (0)
15
94
92.9
69.6
10.8
94.5
24.2 (0.35)
87.5
94.8
0.6 (0.01)
Table 3 Topological ancestry-descendant and sibling relationship reconstruction on simulated data in the presence of CNVs. Results are shown for 5, 10, 15 samples given 1000X
coverage data obtained with localized (top row) and randomized (bottom row) sampling with 80% of SSNVs in CNV regions. All values are averaged over the number of
reconstructed trees (Trees).
5
Page 15 of 18
Page 16 of 18
4. Collins, F.S., Barker, A.D.: Mapping the cancer genome. Scientific
American 296(3), 50–57 (2007)
5. Fisher, R., Pusztai, L., Swanton, C.: Cancer heterogeneity: implications
for targeted therapeutics. British journal of cancer 108(3), 479–485
(2013)
6. Schuh, A., Becq, J., Humphray, S., Alexa, A., Burns, A., Clifford, R.,
Feller, S.M., Grocock, R., Henderson, S., Khrebtukova, I., et al.:
Monitoring chronic lymphocytic leukemia progression by whole genome
sequencing reveals heterogeneous clonal evolution patterns. Blood
120(20), 4191–4196 (2012)
7. Newburger, D.E., Kashef-Haghighi, D., Weng, Z., Salari, R., Sweeney,
R.T., Brunner, A.L., Zhu, S.X., Guo, X., Varma, S., Troxell, M.L., et
al.: Genome evolution during progression to breast cancer. Genome
research 23(7), 1097–1108 (2013)
8. Carter, S.L., Cibulskis, K., Helman, E., McKenna, A., Shen, H., Zack,
T., Laird, P.W., Onofrio, R.C., Winckler, W., Weir, B.A., et al.:
Absolute quantification of somatic dna alterations in human cancer.
Nature biotechnology 30(5), 413–421 (2012)
9. Campbell, P.J., Pleasance, E.D., Stephens, P.J., Dicks, E., Rance, R.,
Goodhead, I., Follows, G.A., Green, A.R., Futreal, P.A., Stratton,
M.R.: Subclonal phylogenetic structures in cancer revealed by
ultra-deep sequencing. Proceedings of the National Academy of
Sciences 105(35), 13081–13086 (2008)
10. Nik-Zainal, S., Van Loo, P., Wedge, D.C., Alexandrov, L.B.,
Greenman, C.D., Lau, K.W., Raine, K., Jones, D., Marshall, J.,
Ramakrishna, M., et al.: The life history of 21 breast cancers. Cell
149(5), 994–1007 (2012)
11. Shah, S.P., Roth, A., Goya, R., Oloumi, A., Ha, G., Zhao, Y.,
Turashvili, G., Ding, J., Tse, K., Haffari, G., et al.: The clonal and
mutational evolution spectrum of primary triple-negative breast
cancers. Nature 486(7403), 395–399 (2012)
12. Roth, A., Khattra, J., Yap, D., Wan, A., Laks, E., Biele, J., Ha, G.,
Aparicio, S., Bouchard-Côté, A., Shah, S.P.: Pyclone: statistical
inference of clonal population structure in cancer. Nature methods
(2014)
13. Oesper, L., Mahmoody, A., Raphael, B.J.: Inferring intra-tumor
heterogeneity from high-throughput dna sequencing data. In: Research
in Computational Molecular Biology, pp. 171–172 (2013). Springer
14. Ha, G., Roth, A., Khattra, J., Ho, J., Yap, D., Prentice, L.M., Melnyk,
N., McPherson, A., Bashashati, A., Laks, E., et al.: Titan: inference of
copy number architectures in clonal cell populations from tumor
whole-genome sequence data. Genome research, 180281 (2014)
15. Hajirasouliha, I., Mahmoody, A., Raphael, B.J.: A combinatorial
approach for analyzing intra-tumor heterogeneity from high-throughput
sequencing data. Bioinformatics 30(12), 78–86 (2014)
16. Ding, L., Ley, T.J., Larson, D.E., Miller, C.A., Koboldt, D.C., Welch,
J.S., Ritchey, J.K., Young, M.A., Lamprecht, T., McLellan, M.D., et
al.: Clonal evolution in relapsed acute myeloid leukaemia revealed by
whole-genome sequencing. Nature 481(7382), 506–510 (2012)
17. Landau, D.A., Carter, S.L., Stojanov, P., McKenna, A., Stevenson, K.,
Lawrence, M.S., Sougnez, C., Stewart, C., Sivachenko, A., Wang, L.,
et al.: Evolution and impact of subclonal mutations in chronic
lymphocytic leukemia. Cell 152(4), 714–726 (2013)
18. McFadden, D.G., Papagiannakopoulos, T., Taylor-Weiner, A., Stewart,
C., Carter, S.L., Cibulskis, K., Bhutkar, A., McKenna, A., Dooley, A.,
Vernon, A., et al.: Genetic and clonal dissection of murine small cell
lung carcinoma progression by genome sequencing. Cell 156(6),
1298–1311 (2014)
19. Campbell, P.J., Yachida, S., Mudie, L.J., Stephens, P.J., Pleasance,
E.D., Stebbings, L.A., Morsberger, L.A., Latimer, C., McLaren, S., Lin,
M.-L., et al.: The patterns and dynamics of genomic instability in
metastatic pancreatic cancer. Nature 467(7319), 1109–1113 (2010)
20. Yachida, S., Jones, S., Bozic, I., Antal, T., Leary, R., Fu, B.,
Kamiyama, M., Hruban, R.H., Eshleman, J.R., Nowak, M.A., et al.:
Distant metastasis occurs late during the genetic evolution of
pancreatic cancer. Nature 467(7319), 1114–1117 (2010)
21. Gerlinger, M., Horswell, S., Larkin, J., Rowan, A.J., Salm, M.P.,
Varela, I., Fisher, R., McGranahan, N., Matthews, N., Santos, C.R., et
al.: Genomic architecture and evolution of clear cell renal cell
carcinomas defined by multiregion sequencing. Nature genetics 46(3),
225–233 (2014)
22. Green, M.R., Gentles, A.J., Nair, R.V., Irish, J.M., Kihira, S., Liu,
C.L., Kela, I., Hopmans, E.S., Myklebust, J.H., Ji, H., et al.: Hierarchy
in somatic mutations arising during genomic evolution and progression
of follicular lymphoma. Blood 121(9), 1604–1611 (2013)
23. de Bruin, E.C., McGranahan, N., Mitter, R., Salm, M., Wedge, D.C.,
Yates, L., Jamal-Hanjani, M., Shafi, S., Murugaesu, N., Rowan, A.J.,
et al.: Spatial and temporal diversity in genomic instability processes
defines lung cancer evolution. Science 346(6206), 251–256 (2014)
24. Hajirasouliha, I., Raphael, B.J.: Reconstructing mutational history in
multiply sampled tumors using perfect phylogeny mixtures. In:
Algorithms in Bioinformatics, pp. 354–367. Springer, ??? (2014)
25. Salari, R., Saleh, S.S., Kashef-Haghighi, D., Khavari, D., Newburger,
D.E., West, R.B., Sidow, A., Batzoglou, S.: Inference of tumor
phylogenies with improved somatic mutation discovery. Journal of
Computational Biology 20(11), 933–944 (2013)
26. Gerlinger, M., Rowan, A.J., Horswell, S., Larkin, J., Endesfelder, D.,
Gronroos, E., Martinez, P., Matthews, N., Stewart, A., Tarpey, P., et
al.: Intratumor heterogeneity and branched evolution revealed by
multiregion sequencing. New England Journal of Medicine 366(10),
883–892 (2012)
27. Bashashati, A., Ha, G., Tone, A., Ding, J., Prentice, L.M., Roth, A.,
Rosner, J., Shumansky, K., Kalloger, S., Senz, J., et al.: Distinct
evolutionary trajectories of primary high-grade serous ovarian cancers
revealed through spatial mutational profiling. The Journal of pathology
231(1), 21–34 (2013)
28. Qiao, Y., Quinlan, A.R., Jazaeri, A.A., Verhaak, R.G., Wheeler, D.A.,
Marth, G.T.: Subcloneseeker: a computational framework for
reconstructing tumor clone structure for cancer variant interpretation
and prioritization. Genome Biol 15, 443 (2014)
29. Jiao, W., Vembu, S., Deshwar, A.G., Stein, L., Morris, Q.: Inferring
clonal evolution of tumors from single nucleotide somatic mutations.
BMC bioinformatics 15(1), 35 (2014)
30. Eirew, P., Steif, A., Khattra, J., Ha, G., Yap, D., Farahani, H.,
Gelmon, K., Chia, S., Mar, C., Wan, A., et al.: Dynamics of genomic
clones in breast cancer patient xenografts at single-cell resolution.
Nature (2014)
31. Http://viq854.github.io/lichee/
32. Gusfield, D.: Efficient algorithms for inferring evolutionary trees.
Networks 21(1), 19–28 (1991)
33. Gerstung, M., Beisel, C., Rechsteiner, M., Wild, P., Schraml, P.,
Moch, H., Beerenwinkel, N.: Reliable detection of subclonal
single-nucleotide variants in tumour cell populations. Nature
communications 3, 811 (2012)
34. Yost, S., Alakus, H., Matsui, H., Jepsen, K., Schwab, R., Frazer, K.,
Harismendy, O.: Mutascope: a high-sensitivity variant caller dedicated
to high-throughput pcr amplicons sequencing. In: BMC Proceedings,
vol. 6, p. 22 (2012). BioMed Central Ltd
35. Larson, D.E., Harris, C.C., Chen, K., Koboldt, D.C., Abbott, T.E.,
Dooling, D.J., Ley, T.J., Mardis, E.R., Wilson, R.K., Ding, L.:
Somaticsniper: identification of somatic point mutations in whole
genome sequencing data. Bioinformatics 28(3), 311–317 (2012)
36. Koboldt, D.C., Zhang, Q., Larson, D.E., Shen, D., McLellan, M.D.,
Lin, L., Miller, C.A., Mardis, E.R., Ding, L., Wilson, R.K.: Varscan 2:
somatic mutation and copy number alteration discovery in cancer by
exome sequencing. Genome research 22(3), 568–576 (2012)
37. Saunders, C.T., Wong, W.S., Swamy, S., Becq, J., Murray, L.J.,
Cheetham, R.K.: Strelka: accurate somatic small-variant calling from
sequenced tumor–normal sample pairs. Bioinformatics 28(14),
1811–1817 (2012)
38. Van Loo, P., Nordgard, S.H., Lingjærde, O.C., Russnes, H.G., Rye,
I.H., Sun, W., Weigman, V.J., Marynen, P., Zetterberg, A., Naume,
B., et al.: Allele-specific copy number analysis of tumors. Proceedings
of the National Academy of Sciences 107(39), 16910–16915 (2010)
39. Greaves, M., Maley, C.C.: Clonal evolution in cancer. Nature
481(7381), 306–313 (2012)
40. Yap, T.A., Gerlinger, M., Futreal, P.A., Pusztai, L., Swanton, C.:
Intratumor heterogeneity: seeing the wood for the trees. Science
translational medicine 4(127), 127–1012710 (2012)
41. Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P.,
Witten, I.H.: The weka data mining software: an update. ACM
SIGKDD explorations newsletter 11(1), 10–18 (2009)
Page 17 of 18
42. Gabow, H.N., Myers, E.W.: Finding all spanning trees of directed and
undirected graphs. SIAM Journal on Computing 7(3), 280–287 (1978)
43. O’Madadhain, J., Fisher, D., Smyth, P., White, S., Boey, Y.-B.:
Analysis and visualization of network data using jung. Journal of
Statistical Software 10(2), 1–35 (2005)
Appendix A — Evaluation of Phylosub
We ran PhyloSub [29] on the Gerlinger et al dataset
[21] using default parameters. For each patient in the
dataset we compared the top tree structures (by default, three) with the published tree. It is hard to
find similarities between the topology of reported trees
and the published trees. Therefore, we analyzed the
performance of PhyloSub in finding the correct clusters of mutations, which is a preliminary requirement
for building the cancer progression pathway. Gerlinger
et. al have categorized mutation groups as shared,
heterogeneous (partially shared), and private. Generally speaking, we observed that shared mutations
were clustered correctly but partially shared mutations were often clustered all together as one group,
which violates their presence pattern. As an example, PhyloSub reported a tree for patient EV003 with
three mutation groups (see Figure 8 below). The trunk
branch (colored blue) includes all shared mutations
as well as one private and two partially shared mutations. Second branch (colored green) contains several partially shared mutations and the rest of private mutations with very different presence patterns,
and the third branch (colored red) contains three
partially shared mutations. Case EV003 has a heterogenous group with 8 mutations shared by R9 and
R6 (LONRF2, RHOB, BHLHE40, CMYA5, NOD1,
GCC1, SLC5A12, YLPM1), whereas PhyloSub puts
two of them in the trunk branch, four in the second,
and two in the third brach. Note that, as discussed in
the Results section, for EV003 LICHeE was able to
find all mutation groups and build the lineage matching the reported tree in Gerlinger et. al dataset. For
other cases PhyloSub showed similar performance. We
also observed that PhyloSub does not deal with private
mutations properly. Often private mutations are either
clustered into singleton groups or are distributed into
shared or partially shared groups. For example, for six
out of eight patients in at least one of the top-trees
there are private mutations clustered with shared mutations and located in the trunk branch of the corresponding tree. Note that PhyloSub shows acceptable
performance in analyzing samples with very few mutations and simple (chain) topology. However, when
tested on more complex multi-sample cancer datasets,
the performance of LICHeE is far more superior.
Figure 8 PhyloSub results for patient EV003.
Appendix B — Experimental Details of
the ccRCC, HGSC, and Xenoengraftment
Comparison
ccRCC. The clear cell renal cell carcinoma (ccRCC)
study by Gerlinger et. al [21] validated 602 nonsynonymous nucleotide substitutions and indels from multiple
samples of 8 individuals using ultra-deep amplicon sequencing with an average depth of >400x. The analysis
was performed on 587 out of these mutations guaranteeing a minimum coverage higher than 100x. Based
on the expected sequencing platform error rate, mutations were called in a sample if their VAF was ≥ 0.5%
for substitutions and 1% for indels. In accordance with
the study, we set the cutoffs to Tpresent = Tabsent =
0.005 for calling SSNVs across samples for all patients
except RMH004 and required at least two mutations
as evidence of a node in the tree. For RMH004 we
set Tpresent = Tabsent = 0.01 to allow editing, observing that some mutations with the frequencies slightly
higher that 0.5% were considered absent in some samples of the study.
HGSC. In the high-grade ovarian cancer (HGSC)
dataset from the study by Bashashati et. al. [27], 19
tumor samples from six patients were used to validate
340 somatic mutations using deep amplicon sequencing (with a median coverage of > 5000x). When run-
Page 18 of 18
ning LICHeE, we used the hard thresholds of Tabsent
= 0.005 and Tpresent = 0.01 for all the patients except
Case5, where we used Tabsent = 0.01 and Tpresent =
0.04 due to a higher level of noise in the data, and required at least three mutations as evidence of a node in
the tree. For Case5 we also removed the samples g and
h from consideration since they had multiple inconclusive validation results as stated in the manuscript [27].
Xenoengraftment. The study by Eirew et. al. [30]
used deep-genome and single-cell sequencing to evaluate the clonal dynamics of xenoengraftment of breast
cancer tissue into immunodeficient mice. Single-cell
analysis was done on passages SA501 (samples X1,
X2, and X4) and SA494 (samples T and X4). When
running LICHeE on both passages, we used the hard
thresholds of Tabsent = 0.03 and Tpresent = 0.05 (the
value 0.05 was indicated in the study’s supplementary
materials); a minimum cluster size of 6 for non-private
mutations and 5 for private mutations; and a maximum cluster collapse distance of 0.085.
| 5cs.CE
|
New Integrality Gap Results for the Firefighters Problem on
Trees
Parinya Chalermsook1 and Daniel Vaz1,2
1
arXiv:1601.02388v3 [cs.DM] 29 Jul 2016
2
Max-Planck-Institut für Informatik, Saarbrücken, Germany
Graduate School of Computer Science, Saarland University, Saarbrücken, Germany
{parinya, ramosvaz}@mpi-inf.mpg.de
1st August 2016
Abstract
In the firefighter problem on trees, we are given a tree G = (V, E) together with a vertex
s ∈ V where the fire starts spreading. At each time step, the firefighters can pick one vertex
while the fire spreads from burning vertices to all their neighbors that have not been picked.
The process stops when the fire can no longer spread. The objective is to find a strategy
that maximizes the total number of vertices that do not burn. This is a simple mathematical
model, introduced in 1995, that abstracts the spreading nature of, for instance, fire, viruses,
and ideas. The firefighter problem is NP-hard and admits a (1 − 1/e) approximation based
on rounding the canonical LP. Recently, a PTAS was announced[1]. 1
The goal of this paper is to develop better understanding on the power of LP relaxations
for the firefighter problem. We first show a matching lower bound of (1 − 1/e + ) on the
integrality gap of the canonical LP. This result relies on a powerful combinatorial gadget that
can be used to derive integrality gap results in other related settings. Next, we consider the
canonical LP augmented with simple additional constraints (as suggested by Hartke). We
provide several evidences that these constraints improve the integrality gap of the canonical
LP: (i) Extreme points of the new LP are integral for some known tractable instances and
(ii) A natural family of instances that are bad for the canonical LP admits an improved
approximation algorithm via the new LP. We conclude by presenting a 5/6 integrality gap
instance for the new LP.
1
Introduction
Consider the following graph-theoretic model that abstracts the fire spreading process: We are
given graph G = (V, E) together with the source vertex s where the fire starts. At each time
step, we are allowed to pick some vertices in the graph to be saved, and the fire spreads from
burning vertices to their neighbors that have not been saved so far. The process terminates
when the fire cannot spread any further. This model was introduced in 1995 [13] and has been
used extensively by researchers in several fields as an abstraction of epidemic propagation.
There are two important variants of the firefighters problem. (i) In the maximization variant
(Max-FF), we are given graph G and source s, and we are allowed to pick one vertex per time
step. The objective is to maximize the number of vertices that do not burn. And (ii) In
the minimization variant (Min-FF), we are given a graph G, a source s, and a terminal set
1
The (1 − 1/e) approximation remained the best until very recently when Adjiashvili et al. [1] showed a PTAS.
Their PTAS does not bound the LP gap.
1
X ⊆ V (G), and we are allowed to pick b vertices per time step. The goal is to save all terminals
in X , while minimizing the budget b.
In this paper, we focus on the Max-FF problem. The problem is n1− hard to approximate
in general graphs [2], so there is no hope to obtain any reasonable approximation guarantee.
Past research, however, has focused on sparse graphs such as trees or grids. Much better
approximation algorithms are known on trees: The problem is NP-hard [15] even on trees of
degree at most three, but it admits a (1 − 1/e) approximation algorithm. For more than a
decade [2, 6, 5, 10, 14, 15], there was no progress on this approximability status of this problem,
until a PTAS was recently discovered [1].
Besides the motivation of studying epidemic propagation, the firefighter problem and its
variants are interesting due to their connections to other classical optimization problems:
• (Set cover) The firefighter problem is a special case of the maximum coverage problem with
group budget constraint (MCG) [7]: Given a collection of sets S = {S1 , . . . , Sm } : Si ⊆ X,
together with group constraints, i.e. a partition of S into groups G1 , . . . , G` , we are
interested in choosing one set from each group in a way that maximizes the total number
of elements covered, i.e. a feasible solution is a subset S 0 ⊆ S where |S 0 ∩ Gj | ≤ 1 for all j,
S
and | Si ∈S 0 Si | is maximized. It is not hard to see that Max-FF is a special case of MCG.
We refer the readers to the discussion by Chekuri and Kumar [7] for more applications of
MCG.
• (Cut) In a standard minimum node-cut problem, we are given a graph G together with
a source-sink pair s, t ∈ V (G). Our goal is to find a collection of nodes V 0 ⊆ V (G) such
that G \ V 0 has s and t in distinct connected components. Anshelevich et al. [2] discussed
that the firefighters’ solution can be seen as a “cut-over-time” in which the cut must be
produced gradually over many timesteps. That is, in each time step t, the algorithm is
allowed to choose vertex set Vt0 to remove from the graph G, and again the final goal
is to “disconnect” s from t 2 . This cut-over-time problem is exactly equivalent to the
minimization variant of the firefighter problem. We refer to [2] for more details about this
equivalence.
1.1
Our Contributions
In this paper, we are interested in developing a better understanding of the Max-FF problem
from the perspective of LP relaxation. The canonical LP relaxation has been used to obtain
the known (1 − 1/e) approximation algorithm via straightforward independent LP rounding
(each node is picked independently with the probability proportional to its LP-value). So far, it
was not clear whether an improvement was possible via this LP, for instance, via sophisticated
dependent rounding schemes 3 . Indeed, for the corresponding minimization variant, Min-FF,
Chalermsook and Chuzhoy designed a dependent rounding scheme for the canonical LP in
order to obtain O(log∗ n) approximation algorithm, improving upon an O(log n) approximation
obtained via independent LP rounding. In this paper, we are interested in studying this potential
improvement for Max-FF.
Our first result refutes such possibility for Max-FF. In particular, we show that the integrality gap of the standard LP relaxation can be arbitrarily close to (1 − 1/e).
Theorem 1. For any > 0, there is an instance (G, s) (whose size depends on ) such that the
ratio between optimal integral solution and fractional one is at most (1 − 1/e + ).
2
The notion of disconnecting the vertices here is slightly non-standard.
Cai, Verbin, and Yang [5] claimed an LP-respecting integrality gap of (1 − 1/e), but many natural rounding
algorithms in the context of this problem are not LP respecting, e.g. in [6]
3
2
Our techniques rely on a powerful combinatorial gadget that can be used to prove integrality
gap results in some other settings studied in the literature. In particular, in the b-Max-FF
problem, the firefighters can pick up to b vertices per time step, and the goal is to maximize the
number of saved vertices. We provide an integrality gap of (1 − 1/e) for the b-Max-FF problem
for all constant b ∈ N, thus matching the algorithmic result of [9]. In the setting where
√ an input
tree has degree at most d ∈ [4, ∞), we show an integrality gap result of (1−1/e+O(1/ d)). The
best known algorithmic result in this setting was previously a (1 − 1/e + Ω(1/d)) approximation
due to [14].
Motivated by the aforementioned negative results, we search for a stronger LP relaxation for
the problem. We consider adding a set of valid linear inequalities, as suggested by Hartke [12].
We show the following evidences that the new LP is a stronger relaxation than the canonical
LP.
• Any extreme point of the new LP is integral for the tractable instances studied by Finbow
et al. [11]. In contrast, we argue that the canonical LP does not satisfy this integrality
property of extreme points.
• A family of instances which captures the bad integrality gap instances given in Theorem 1,
admits a better than (1 − 1/e) approximation algorithm via the new LP.
• When the LP solution is near-integral, e.g. for half-integral solutions, the new LP is
provably better than the old one.
Our results are the first rigorous evidences that Hartke’s constraints lead to improvements
upon the canonical LP. All the aforementioned algorithmic results exploit the new LP constraints
in dependent LP rounding procedures. In particular, we propose a two-phase dependent rounding algorithm, which can be used in deriving the second and third results. We believe the new
LP has an integrality gap strictly better than (1 − 1/e), but we are unable to analyze it.
Finally, we show a limitation of the new LP by presenting a family of instances, whose
integrality gap can be arbitrarily close to 5/6. This improves the known integrality gap ratio [12],
and puts the integrality gap answer somewhere between (1 − 1/e) and 5/6. Closing this gap is,
in our opinion, an interesting open question.
Organization: In Section 2, we formally define the problem and present the LP relaxation.
In Section 3, we present the bad integrality gap instances. We present the LP augmented with
Hartke’s constraints in Section 4 and discuss the relevant evidences of its power in comparison to
the canonical LP. Some proofs are omitted for space constraint, and are presented in Appendix.
Related results: King and MacGillivray showed that the firefighter problem on trees is
solvable in polynomial time if the input tree has degree at most three, with the fire starting
at
√
a degree-2 vertex. From exponential time algorithm’s perspective, Cai et al. showed 2O( n log n)
time, exact algorithm. The discrete mathematics community pays particularly high attention
to the firefighter problem on grids [16, 10], and there has also been some work on infinite
graphs [13].
The problem also received a lot of attention from the parameterized complexity perspectives [8, 3, 5] and on many special cases, e.g., when the tree has bounded pathwidth [8] and on
bounded degree graphs [8, 4].
Recent update: Very recently, Adjiashvili et al. [1] showed a polynomial time approximation scheme (PTAS) for the Max-FF problem, therefore settling the approximability status.
Their results, however, do not bound the LP integrality gap. We believe that the integrality
gap questions are interesting despite the known approximation guarantees.
3
2
Preliminaries
A formal definition of the problem is as follows. We are given a graph G and a source vertex s
where the fire starts spreading. A strategy is described by a collection of vertices U = {ut }nt=1
where ut ∈ V (G) is the vertex picked by firefighters at time t. We say that a vertex u ∈ V (G)
is saved by the strategy U if for all path P = (s = v0 , . . . , vz = u) from s to u, we have
vi ∈ {u1 , . . . , ui } for some i = 1, . . . , z. A vertex v not saved by U is said to be a burning vertex.
The objective of the problem is to compute U so as to maximize the total number of saved
vertices. Denote by OPT(G, s) the number of vertices saved by an optimal solution.
When G is a tree, we think of G as being partitioned into layers L1 , . . . , Lλ where λ is the
height of the tree, and Li contains vertices whose distance is exactly i from s. Every strategy
has the following structure.
Proposition 2. Consider the firefighters problem’s instance (G, s) where G is a tree. Let
U = {u1 , . . . , un } be any strategy. Then there is another strategy U 0 = {u0t } where u0t belongs to
layer t in G, and U 0 saves at least as many vertices as U does.
We remark that this structural result only holds when an input graph G is a tree.
LP Relaxation: This paper focuses on the linear programming aspect of the problem. For
any vertex v, let Pv denote the (unique) path from s to v, and let Tv denote the subtree rooted
at v. A natural LP relaxation is denoted by (LP-1): We have variable xv indicating whether v
is picked by the solution, and yv indicating whether v is saved.
(LP-2)
(LP-1)
max
X
max
yv
xv ≤ 1 for all layer j
X
xv ≤ 1 for all layer j
v∈Lj
v∈Lj
yv ≤
yv
v∈X
v∈V
X
X
X
xu for all v ∈ V
yv ≤
X
xu for all v ∈ X
u∈Pv
u∈Pv
xv , yv ∈ [0, 1] for all v
xv , yv ∈ [0, 1] for all v
Let LP(T, s) denote the optimal fractional LP value for an instance (T, s). The integrality
gap gap(T, s) of the instance (T, s) is defined as gap(T, s) = OPT(T, s)/LP(T, s). The integrality
gap of the LP is defined as inf T gap(T, s).
Firefighters with terminals: We consider a more general variant of the problem, where
we are only interested in saving a subset X of vertices, which we call terminals. The goal is
now to maximize the number of saved terminals. An LP formulation of this problem, given an
instance (T, v, X ), is denoted by (LP-2). The following lemma argues that these two variants
are “equivalent” from the perspectives of LP relaxation.
Lemma 3. Let (T, X , s), with |X | > 0, be an input for the terminal firefighters problem that
gives an integrality gap of γ for (LP-2), and that the value of the fractional optimal solution is
at least 1. Then, for any > 0, there is an instance (T 0 , s0 ) that gives an integrality gap of γ +
for (LP-1).
Proof. Let M = 2|V (T )|/. Starting from (T, X , s), we construct an instance (T 0 , s0 ) by adding
M children to each vertex in X , so the number of vertices in T 0 is |V (T 0 )| = |V (T )| + M |X |.
We denote the copies of X in T 0 by X 0 and the set of their added children by X 00 . The root of
4
the new tree, s0 , is the same as s (the root of T .) Now we argue that the instance (T 0 , s0 ) has
the desired integrality gap, i.e. we argue that OPT(T 0 , s0 ) ≤ (γ + )LP(T 0 , s0 ).
Let (x0 , y 0 ) be an integral solution to the instance (T 0 , s0 ). We upper bound the number of
P
vertices saved by this solution, i.e. upper bounding v∈V (T 0 ) yv0 . We analyze three cases:
• For a vertex v ∈ V (T 0 ) \ X 00 , we upper bound the term yv0 by 1, and so the sum
P
0
v∈V (T 0 )\X 00 yv by |V (T )|.
• Now define X̃ ⊆ X 00 as the set of vertices v for which yv0 = 1 but yu0 = 0 for the parent u of
P
0 ≤ |X |: We break the
v. This means
that x0v = 1 for all vertices
in X̃ . Notice that v∈X̃ yvo
n o
n
set X̃ into X̃u
where X̃u = v ∈ X̃ : u is the parent of v . The LP constraint
u∈V (T 0 )
P
P
guarantees that v∈X̃u yv0 = v∈X̃u x0v ≤ 1 (all vertices in X̃u belong to the same layer.)
Summing over all such u ∈ X 0 , we get the desired bound.
• Finally, consider the term v∈X 0 \X̃ yv0 . Let (x∗ , y ∗ ) be an optimal fractional solution
to (T, X , s) for (LP-2). We only need to care about all vertices v such that yu0 = 1
P
for the parent u of v. This term is upper bounded by M u∈X 0 yu0 , which is at most
P
M γ ( v∈X yv∗ ), due to the fact that the solution (x0 , y 0 ) induces an integral solution on
instance (T, X , s).
P
Combining the three cases, we get |V (T )| + |X | + M γ ( v∈X yv∗ ) ≤ M + γM ( v∈X yv∗ ) ≤
P
P
M (γ + ) ( v∈X yv∗ ), if v∈X yv∗ ≥ 1. Now, notice that the fractional solution (x∗ , y ∗ ) for (LP-2)
on instance (T, X , s) is also feasible for (LP-1) on (T 0 , s0 ) with at least a multiplicative factor
of M times the objective value: Each fractional saving of u ∈ X 0 contributes to save all M
P
children of u. Therefore, M ( v∈X yv∗ ) ≤ M · LP(T 0 , s0 ), thus concluding the proof.
P
P
We will, from now on, focus on studying the integrality gap of (LP-2).
3
Integrality Gap of (LP-2)
We first discuss the integrality gap of (LP-2) for a general tree. We use the following combinatorial gadget.
Gadget: A (M, k, δ)-good gadget is a collection of trees T = {T1 , . . . , TM }, with roots
S
r1 , . . . , rM where ri is a root of Ti , and a subset S ⊆ V (Ti ) that satisfy the following properties:
• (Uniform depth) We think of these trees as having layers L0 , L1 , . . . , Lh , where Lj is the
union over all trees of all vertices at layer j and L0 = {r1 , . . . , rm }. All leaves are in the
same layer Lh .
• (LP-friendly) For any layer Lj , j ≥ 1, we have |S ∩ Lj | ≤ k. Moreover, for any tree Ti
and a leaf v ∈ V (Ti ), the unique path from ri to v must contain at least one vertex in S.
• (Integrally adversarial) Let B ⊆ {r1 , . . . , rM } be any subset of roots. Consider a subset of
vertices U = {uj }hj=1 such that uj ∈ Lj . For ri ∈ B and a leaf v ∈ Lh ∩ V (Ti ), we say that
v is (U, B)-risky if the unique path from ri to v does not contain any vertex in U. There
must be at least (1 − 1/k − δ) |B|
M |Lh | vertices in Lh that are (U, B)-risky, for all choices of
B and U.
We say that vertices in S are special and all other vertices are regular.
5
Lemma 4. For any integers k ≥ 2, M ≥ 1, and any real number δ > 0, a (M, k, δ)-good gadget
exists. Moreover, the gadget contains at most (k/δ)O(M ) vertices.
We first show how to use this lemma to derive our final construction. The proof of the
lemma follows later.
Construction: Our construction proceeds in k phases, and we will define it inductively.
The first phase of the construction is simply a (1, k, δ)-good gadget. Now, assume that we
have constructed the instance up to phase q. Let l1 , . . . , lMq ∈ Lαp be the leaves after the
construction of phase q that all lie in layer αq . In phase q + 1, we take the (Mq , k, δ)-good
gadget (Tq , {rq }, Sq ); recall that such a gadget consists of Mq trees. For each i = 1, . . . , Mq , we
unify each root ri with the leaf li . This completes the description of the construction.
S
Denote by S̄q = q0 ≤q Sq0 the set of all special vertices in the first q phases. After phase q,
we argue that our construction satisfies the following properties:
• All leaves are in the same layer αq .
• For every layer Lj , |Lj ∩ S̄q | ≤ k. For every path P from the root to v ∈ Lαi , |P ∩ S̄q | = q.
• For any integral solution U, at least |Lαq | ((1 − 1/k)q − qδ) vertices of Lαq burn.
It is clear from the construction that the leaves after phase q are all in the same layer. As
to the second property, the properties of the gadget ensure that there are at most k special
vertices per layer. Moreover, consider each path P from the root to some vertex v ∈ Lαq+1 . We
can split this path into two parts P = P 0 ∪ P 00 where P 0 starts from the root and ends at some
v 0 ∈ Lαq , and P 00 starts at v 0 and ends at v. By the induction hypothesis, |P 0 ∩ S̄q | = q and the
second property of the gadget guarantees that |P 00 ∩ Sq+1 | = 1.
To prove the final property, consider a solution U = {u1 , . . . , uαq+1 }, which can be seen as
0
U ∪ U 00 where U 0 = {u1 , . . . , uαq } and U 00 = {uαq +1 , . . . , uαq+1 }. By the induction hypothesis,
we have that at least ((1 − 1/k)q − qδ) |Lαq | vertices in Lαq burn; denote these burning vertices
|B|
by B. The third property of the gadget will ensure that at least (1 − 1/k − δ) M
|Lαq+1 | vertices
q
00
in Lαq+1 must be (U , B)-risky. For each risky vertex v ∈ Lαq+1 , a unique path from the root
to v 0 ∈ B does not contain any vertex in U 0 , and also the path from v 0 to v does not contain a
vertex in U 00 (due to the fact that it is (U 00 , B)-risky.) This implies that such vertex v must burn.
Therefore, the fraction of burning vertices in layer Lαq+1 is at least (1 − 1/k − δ)|B|/Mq ≥ (1 −
1/k −δ)((1−1/k)q −qδ), by induction hypothesis. This number is at least (1−1/k)q+1 −(q +1)δ,
maintaining the invariant.
After the construction of all k phases, the leaves are designated as the terminals X . Also,
Mq+1 ≤ (k/δ)2Mq , which means that, after k phases, Mk is at most a tower function of (k/δ)2 ,
···
that is, (k/δ)2(k/δ) with k − 1 such exponentiations. The total size of the construction is
P
2Mq ≤ (k/δ)2Mk = O(M
k+1 ).
q (k/δ)
An example construction, when k = 2, is presented in Figure 3 (in Appendix).
Theorem 5. A fractional solution, that assigns xv = 1/k to each special vertex v, saves every
terminal. On the other hand, any integral solution can save at most a fraction of 1−(1−1/k)k +.
Proof. We assign the LP solution xv = 1/k to all special vertices (those vertices in S̄k ), and
xv = 0 to regular vertices. Since the construction ensures that there are at most k special vertices
P
per layer, we have v∈Lj xv ≤ 1 for every layer Lj . Moreover, every terminal is fractionally
P
saved: For any t ∈ X , the path |Pt ∩ S̄k | = k, so we have v∈Pt xv = 1.
For the integral solution analysis, set δ = /k. The proof follows immediately from the
properties of the instance.
6
3.1
Proof of Lemma 4
We now show that the (M, k, δ)-good gadget exists for any value of M ∈ N, k ∈ N, k ≥ 2 and
δ ∈ R>0 . We first describe the construction and then show that it has the desired properties.
Construction: Throughout the construction, we use a structure which we call spider. A
spider is a tree in which every node except the root has at most one child. If a node has no
children (i. e. a leaf), we call it a foot of the spider. We call the paths from the root to each
foot the legs of the spider.
Let D = d4/δe. For each i = 1, . . . , M , the tree Ti is constructed as follows. We have a
spider rooted at ri that contains kDi−1 legs. Its feet are in Di−1 consecutive layers, starting at
P
layer αi = 1 + j<i Dj−1 ; each such layer has k feet. Denote by S (i) the feet of these spiders.
Next, for each vertex v ∈ S (i) , we have a spider rooted at v, having D2M −i+1 feet, all of which
P
S
(i)
belong to layer α = 1 + j≤M Dj−1 . The set S is defined as S = M
i=1 S . This concludes the
construction. We will use the following observation:
Observation 6. For each root ri , the number of leaves of Ti is kD2M .
Analysis: We now prove that the above gadget is (M, k, δ)-good. The construction ensures
that all leaves are in the same layer Lα .
The second property also follows obviously from the construction: For i 6= i0 , we have that
0
(i)
S ∩ S (i ) = ∅, and that each layer contains exactly k vertices from S (i) . Moreover, any path
from ri to the leaf of Ti must go through a vertex in S (i) .
The third and final property is established by the following two lemmas.
Lemma 7. For any ri ∈ B and any subset of vertices U = {uj }hj=1 such that uj ∈ Lj , a fraction
of at least (1 − 1/k − 2/D) of S (i) are (U, B)-risky.
Proof. Notice that a vertex v is (U, B)-risky if U is not a vertex cut separating v from B. There
are kDi−1 vertex-disjoint paths from ri ∈ B to vertices in S (i) . But the cut U induced on these
P
0
paths contains at most i0 ≤i Di −1 vertices (because all vertices in S (i) are contained in the
P
0
first i0 ≤i Di −1 ≤ Di−1 + 2Di−2 layers.) Therefore, at most (1/k + 2/D) fraction of vertices in
S (i) can be disconnected by U, and those that are not cut remain (U, B)-risky.
Lemma 8. Let v ∈ S (i) that is (U, B)-risky. Then at least (1 − 2/D) fraction of descendants
of v in Lα must be (U, B)-risky.
Proof. Consider each v ∈ S (i) that is (U, B)-risky and a collection of leaves Lv that are descendants of vertex v. Notice that a leaf u ∈ Lv is (U, B)-risky if removing U does not disconnect
vertex v from u.
There are D2M −i+1 ≥ DM +1 vertex disjoint paths connecting vertex v with leaves in Lv ,
while the cut set U contains at most 2DM vertices. Therefore, removing U can disconnect at
most 2/D fraction of vertices in Lv from v.
Combining the above two lemmas, for each ri ∈ B, the fraction of leaves of Ti that are
(U, B)-risky are at least (1 − 1/k − 2/D)(1 − 2/D) ≥ (1 − 1/k − 4/D). Therefore, the total
number of such leaves, over all trees in T , are (1 − 1/k − δ)|B||Lα |/M .
We extend the cosntruction to other settings in Section 3.1 (in Appendix).
Arbitrary number of firefighters: Let b ∈ N. In the b-firefighter problem, at each time
step, the firefighters may choose up to b vertices, and the fire spreads from the burning vertices
to vertices that have not been chosen so far. The goal is to maximize the number of saved
vertices. In this section, we show the following:
7
Theorem 9. For any integer b ∈ N (independent of |V (G)|), the integrality gap of the canonical
LP can be arbitrarily close to (1 − 1/e).
Proof. To prove this theorem, one only need to design a slightly different good gadget. That
is, an (M, k, δ)-good gadget is now a collection of trees T with roots r1 , . . . , rM together with
S
S ⊆ V (Ti ) that satisfy the following properties:
• All leaves of Ti are in the same layer Lh .
• For each layer Lj , we have |S ∩ Lj | ≤ kb. Moreover, for any tree Ti and a leaf v ∈ V (Ti ),
the unique path from ri to v must contain at least one vertex in S.
h|
• For any subset B ⊆ {r1 , . . . , rM } of roots and for any strategy U, at least (1−1/k−δ) |B||L
M
vertices in Lh are (U, B)-risky.
It is not hard to see that these gadgets can be used to construct the integrality gap in the
same way as in the previous section. Details are omitted.
Bounded degrees: Iwakawa et al. showed a (1 − 1/e + Ω(1/d)) approximation algorithm for
the instance that has degree at most d. We show an instance where this dependence on 1/d is
almost the best possible that can be obtained from this LP.
Theorem
10. For all d ≥ 4, the integrality gap of (LP-1) on degree-d graphs is (1 − 1/e +
√
O(1/ d)).
Proof. To prove this theorem, we construct a “bounded degree” analogue of our good gadgets.
That is, the (M, k)-good gadget in this setting guarantees that
• All leaves of Ti are in the same layer Lh .
• For each layer Lj , we have |S ∩ Lj | ≤ k. For each tree Ti , for each leaf v ∈ V (Ti ), the
unique path from ri to v contains one vertex in S.
h|
• For any subset B ⊆ {r1 , . . . , rM }, for any strategy U, at least (1 − 1/k − O(1/d)) |B||L
M
vertices in Lh are (U, B)-risky.
This gadget can be used to recursively construct the instance in k phases.
The final instance
√
k
guarantees the integrality
√ gap of 1−(1−1/k) +O(k/d). By setting k = d, we get the integrality
gap of (1 − 1/e + O(1/ d)) as desired4 .
4
Hartke’s Constraints
Due to the integrality gap result in the previous section, there is no hope to improve the best
known algorithms via the canonical LP relaxation. Hartke [12] suggested adding the following
constraints to narrow down the integrality gap of the LP.
X
xu ≤ 1 for all vertex v ∈ V (T ) and layer Lj below the layer of v
u∈Pv ∪(Tv ∩Lj )
We write the new LP with these constraints below:
4
By analyzing the Taylor’s series expansion of 1/e − (1 − 1/k)k , we get the term
8
1
2ek
+ O(1/k2 )
(LP’)
max
X
yv
v∈V
xu ≤ 1 for all layer j below vertex v
X
u∈Pv ∪(Tv ∩Lj )
yv ≤
X
xu for all v ∈ V
u∈Pv
xv , yv ∈ [0, 1] for all v
Proposition 11. Given the values {xv }v∈V (T ) that satisfy the first set of constraints, then the
P
solution (x, y) defined by yv = u∈Pv xv is feasible for (LP’) and at least as good as any other
feasible (x, y 0 ).
In this section, we study the power of this LP and provide three evidences that it may be
stronger than (LP-1).
4.1
New properties of extreme points
In this section, we show that Finbow et al. tractable instances [11] admit a polynomial time
exact algorithm via (LP’) (in fact, any optimal extreme point for (LP’) is integral.) In contrast,
we show that (LP-1) contains an extreme point that is not integral.
We first present the following structural lemma.
Lemma 12. Let (x, y) be an optimal extreme point for (LP’) on instance T rooted at s. Suppose
s has two children, denoted by a and b. Then xa , xb ∈ {0, 1}.
Proof. Suppose that xa , xb ∈ (0, 1). We will define two solutions (x0 , y0 ) and (x00 , y00 ) and derive
that (x, y) can be written as a convex combination of (x0 , y0 ) and (x00 , y00 ), a contradiction.
First, we define (x0 , y0 ) by setting x0b = 1, x0a = 0. For each vertex v ∈ Tb , we set x0v = 0. For
each vertex v ∈ Ta , we define x0v = xv /(1 − xa ). We verify that x0 is feasible
for (LP’):
For each
P
v ∈ Ta and any layer Lj below v,
0
u∈Pv xu
P
+
0
u∈Tv ∩Lj xu
P
=
P
xu −xa
(1−xa )
u∈Pv
+
u∈Tv ∩Lj
xu
(1−xa )
≤
P
u∈Pv ∪(Tv ∩Lj )
xu −xa
≤ 1 (the last inequality is due to the fact that x is feasible). The
P
constraint is obviously satisfied for all v ∈ Tb . For the root node v = s, we have u∈Lj x0u =
(1−xa )
P
u ∈(Lj ∩Ta
x −xa
) u
≤ 1.
We define
analogously: x00b = 0, x00a = 1. For each vertex v ∈ Ta , we set x00v = 0,
and for each v ∈ Tb , we define x00v = xv /(1 − xb ). It can be checked similarly that (x00 , y00 ) is a
feasible solution.
(1−xa )
(x00 , y00 )
Claim 13. If x is an optimal extreme point, then xa + xb = 1.
Proof. Observe that, for each v ∈ Tb , yv0 = 1 and for each v ∈ Ta , yv0 =
value of x0 is |Tb | +
P
v∈Ta
yv0 = |Tb | +
1
(1−xa )
v∈Ta (yv
P
Similarly, the objective value of solution x00 is |Ta | +
xb
(1−xb ) |Tb |.
9
− xa ) = |Tb | +
1
(1−xb )
v∈Tb (yv
P
yv −xa
P1−xa
. The objective
y
v∈Ta v
(1−xa )
−
− xb ) = |Ta | +
xa
(1−xa )
P
v∈Tb
|Ta |.
yv
(1−xb )
−
Consider the convex combination
has the objective value of
1−xa
0
(2−xa −xb ) x
+
1−xb
00
(2−xa −xb ) x .
This solution is feasible and
1
· (1 − xa − xb ) (|Ta | + |Tb |) +
yv
(2 − xa − xb )
v∈V (T )
X
If xa + xb < 1, we apply the fact that |Ta | + |Tb | > v∈V (T ) yv to get the objective of strictly
P
more than v∈V (T ) yv , contradicting the fact that (x, y) is optimal.
P
Finally, we define the convex combination by z = (1 − xa )x0 + xa x00 . It can be verified easily
that zv = xv for all v ∈ V (T ).
Finbow et al. Instances: In this instance, the tree has degree at most 3 and the root has
degree 2. Finbow et al. [11] showed that this is polynomial time solvable.
Theorem 14. Let (T, s) be an input instance where T has degree at most 3 and s has degree
two. Let (x, y) be a feasible fractional solution for (LP-3). Then there is a polynomial time
P
algorithm that saves at least v∈V (T ) yv vertices.
Proof. We prove this by induction on the number of nodes in the tree that, for any tree (T 0 , s0 )
that is a Finbow et al. instance, for any fractional solution (x, y) for (LP’), there is an integral
P
P
solution (x0 , y 0 ) such that v∈T 0 \{s0 } yv0 = v∈T 0 \{s0 } yv . Let a and b be the children of the
P
root s. From Lemma 12, assume w.l.o.g. that xa = 1, so we have v∈Ta yv = |Ta |. By
the induction hypothesis, there is an integral solution (x0 , y 0 ) for the subtree Tb such that
P
P
P
0 0
0
0
extended to the instance T
v∈Tb yv . The solution (x , y ) can be P
v∈Tb \{b} yv =
v∈Tb yv =
P
0
by defining xa = 1. This solution has the objective value of |Ta | + v∈Tb yb0 = |Ta | + v∈Tb yb ,
completing the proof.
Bad instance for (LP-1): We show in Figure 1 a Finbow et al.
instance as well as a solution for (LP-1) that is optimal and an extreme
point, but not integral.
a
b
Claim 15. The solution (x, y) represented in Figure 1, with y defined
c
d
according to Proposition 11, is an extreme point of this instance for
(LP-1).
Figure 1 Instance with
Proof. Suppose (for contradiction) that (x, y) is not an extreme point. a non-integral extreme
Then, there are distinct solutions (x0 , y 0 ), (x00 , y 00 ) and α ∈ (0, 1) such point for (LP-1). Gray
vertices: xv = 1/2; oththat (x, y) = α(x0 , y 0 ) + (1 − α)(x00 , y 00 ). Since yc = 1 and yc0 , yc00 ≤ 1, erwise: x =
0.
v
then yc0 = yc00 = 1, and likewise, yd0 = yd00 = 1. Combining that
x0a + x0c = yc0 = 1 with x0a + x0d = yd0 = 1 and x0c + x0d ≤ 1, we conclude
that x0a ≥ 1/2. Similarly, we get that x00a ≥ 1/2, which implies that x0a = x00a = 1/2.
Similar reasoning using that x0a + x0b ≤ 1 allows us to conclude that x0b = x00b = 1/2, and
thus, (x0 , y 0 ) = (x00 , y 00 ) = (x, y), which contradicts our assumption.
4.2
Rounding 1/2-integral Solutions
We say that the LP solution (x, y) is (1/k)-integral if, for all v, we have xv = rv /k for some
integer rv ∈ {0, . . . , k}. By standard LP theory, one can assume that the LP solution is (1/k)integral for some polynomially large integer k.
In this section, we consider the case when k = 2 (1/2-integral LP solutions). From Theorem 5, (LP-1) is not strong enough to obtain a 3/4 + approximation algorithm, for any > 0.
Here, we show a 5/6 approximation algorithm based on rounding (LP’).
10
Theorem 16. Given a solution (x, y) for (LP’) that is 1/2-integral, there is a polynomial time
P
algorithm that produces a solution of cost 5/6 v∈V (T ) yv .
We believe that the extreme points in some interesting special cases will be 1/2-integral.
Algorithm’s Description: Initially, U = ∅. Our algorithm considers the layers L1 , . . . , Ln
in this order. When the algorithm looks at layer Lj , it picks a vertex uj and adds it to U, as
follows. Consider Aj ⊆ Lj , where Aj = {v ∈ Lj : xv > 0}. Let A0j ⊆ Aj contain vertices v such
that there is no ancestor of v that belongs to Aj 0 for some j 0 < j, and A00j = Aj \ A0j , i.e. for
each v ∈ A00j , there is another vertex u ∈ Aj 0 for some j 0 < j such that u is an ancestor of v.
We choose the vertex uj based on the following rules:
• If there is only one v ∈ Aj , such that v is not saved by U so far, choose uj = v.
• Otherwise, if |A0j | = 2, pick uj at random from A0j with uniform probability. Similarly, if
|A00j | = 2, pick uj at random from A00j .
• Otherwise, we have the case |A0j | = |A00j | = 1. In this case, we pick vertex uj from A0j with
probability 1/3; otherwise, we take from A00j .
Analysis: Below, we argue that each vertex v ∈ V (T ) : xv > 0 is saved with probability
at least (5/6)yv . It is clear that this implies the theorem: Consider a vertex v 0 : xv0 = 0. If
yv0 = 0, we are immediately done. Otherwise, consider the bottommost ancestor v of v 0 such
that xv > 0. Since yv = yv0 , the probability that v 0 is saved is the same as that of v, which is
at least (5/6)yv .
We analyze a number of cases. Consider a layer Lj such that |Aj | = 1. Such a vertex v ∈ Aj
is saved with probability 1.
Next, consider a layer Lj such that |A0j | = 2. Each vertex v ∈ A0j is saved with probability
1/2 and yv = 1/2. So, in this case, the probability of saving v is more than (5/6)yv .
Lemma 17. Let Lj be the layer such that |A0j | = |A00j | = 1. Then the vertex u ∈ A0j is saved
with probability 2/3 ≥ (5/6)yu and vertex v ∈ A00j is saved with probability 5/6.
Proof. Let v 0 ∈ Aj 0 be the ancestor of v in some layer above Aj . The fact that v has not been
saved means that v 0 is not picked by the algorithm, when it processed Aj 0 .
We prove the lemma by induction on the value of j. For the base case, let Lj be the first
layer such that |A0j | = |A00j | = 1. This means that the layer Lj 0 must have |A0j 0 | = 2, and
therefore the probability of v 0 being saved is at least 1/2. Vertex u is not saved only if both
v 0 and u are not picked, and this happens with probability 1/2 · 2/3 = 1/3. Hence, vertex u is
saved with probability 2/3 as desired. Consider now the base case for vertex v, which is not
saved only if v 0 is not saved and u is picked by the algorithm among {u, v}. This happens with
probability 1/2 · 1/3 = 1/6, thus completing the proof of the base case.
For the purpose of induction, we now assume that, for all layer Li above Lj such that
|A0i | = |A00i | = 1, the probability that the algorithm saves the vertex in A0i is at least 2/3. Since
the vertex u is not saved only if v 0 is not saved, this probability is either 1/2 or 1/3 depending
on the layer to which v 0 belongs. If it is 1/3, we are done; otherwise, the probability is at most
1/2 · 2/3 = 1/3. Now consider vertex v, which is not saved only if v 0 is not saved and u is picked
at Lj . This happens with probability at most 1/2 · 1/3 = 1/6.
Lemma 18. Let Lj be a layer such that A00j = {u, v} (containing two vertices). Then each such
vertex is saved with probability at least 5/6.
11
Proof. Let u0 and v 0 be the ancestors of u and v in some sets A0i and A0k above the layer Lj .
There are the two possibilities: either both u0 and v 0 are in layers with |A0i | = |A0k | = 2 (maybe
i = k); or u0 is in the layer with |A0i | = |A00i | = 1. We remark that u0 6= v 0 : otherwise, the LP
constraint for v 0 and Lj would not be satisfied.
For u or v to be unsaved, we need that both u0 and v 0 are not saved by the algorithm.
Otherwise, if, say, u0 is saved, u is also saved, and the algorithm would have picked v.
P [u is not saved] = P [u not picked ∧ u0 is not saved ∧ v 0 is not saved]
= P [u not picked] · P [u0 is not saved ∧ v 0 is not saved]
1
1 1
= · =
2 4
8
7
5
P [u is saved] = ≥
8
6
It must be that P [u0 burns ∧ v 0 burns] ≤ 1/4, since either u0 and v 0 are in different layers or
they are in the same layer. If they are in different layers, picking each of them is independent,
and the probability of neither being saved is at most 1/4. If they are in the same layer, one
of them is necessarily picked, which implies that the probability of neither being saved is 0. In
any case, the probability is at most 1/4.
In the second case, at least one of the vertices u0 , v 0 is in a layer with one 2-special vertex.
W. l. o. g. let u0 be in such a layer. By Lemma 17, we know that the probability that u0 is not
saved is at most 1/3. Therefore,
P [u burns] = P [u not picked ∧ u0 burns ∧ v 0 burns]
= P [u not picked] · P [u0 burns ∧ v 0 burns]
≤ P [u not picked] · P [u0 burns]
1 1
1
≤ · =
2 3
6
5
P [u is saved] ≥
6
The proof for both cases works analogously for v.
4.3
Ruling Out the Gap Instances in Section 3
In this section, we show that the integrality gap instances for (LP-1) presented in the previous
section admit a better than (1 − 1/e) approximation via (LP’). To this end, we introduce the
concept of well-separable LP solutions and show an improved rounding algorithm for solutions
in this class.
Let η ∈ (0, 1). Given an LP solution (x, y) for (LP-1) or (LP’), we say that a vertex v is
P
η-light if u∈Pv \{v} xu < η; if a vertex v is not η-light, we say that it is η-heavy. A fractional
solution is said to be η-separable if for all layer j, either all vertices in Lj are η-light, or they
are all η-heavy. For an η-separable LP solution (x, y), each layer Lj is either an η-light layer
that contains only η-light vertices, or η-heavy layer that contains only η-heavy vertices.
Observation 19. The LP solution in Section 3 is η-separable for all values of η ∈ {1/k, 2/k, . . . , 1}.
Theorem 20. If the LP solution (x, y) is η-separable for some η, then there is an efficient
P
algorithm that produces an integral solution of cost (1 − 1/e + f (η)) v yv , where f (η) is some
function depending only on η.
12
Algorithm: Let T be an input tree, and (x, y) be a solution for (LP’) on T that is ηseparable for some constant η ∈ (0, 1). Our algorithm proceeds in two phases. In the first
phase, it performs randomized rounding independently for each η-light layer. Denote by V1 the
(random) collection of vertices selected in this phase. Then, in the second phase, our algorithm
performs randomized rounding conditioned on the solutions in the first phase. In particular,
when we process each η-heavy layer Lj , let L̃j be the collection of vertices
that
have not yet been
saved by V1 . We sample one vertex v ∈ L̃j from the distribution
xv
x(L̃j )
. Let V2 be the
v∈L̃j
set of vertices chosen from the second phase. This completes the description of our algorithm.
For notational simplification, we present the proof when η = 1/2. It will be relatively
obvious that the proof can be generalized to work for any η. Now we argue that each terminal
t ∈ X is saved with probability at least (1 − 1/e + δ)yt for some universal constant δ > 0
that depends only on η. We will need the following simple observation that follows directly by
standard probabilistic analysis.
Proposition 21. For each vertex v ∈ V (T ), the probability that v is not saved is at most
Q
−yv .
u∈Pv (1 − xu ) ≥ 1 − e
We start by analyzing two easy cases.
Lemma 22. Consider t ∈ X . If yt < 0.9 or there is some ancestor v ∈ Pt such that xv > 0.2,
then the probability that v is saved by the algorithm is at least (1 − 1/e + δ)yt .
Proof. First, let us consider the case where yt < 0.9. The probability of t being saved is at
least 1 − e−yv , according to the straightforward analysis. If yt < 0.9, we have 1 − e−yt /yt >
1.04(1 − 1/e)yt as desired.
Consider now the second case when xv > 0.2 for some ancestor v ∈ Pt . The bound used
typically in the analysis is only tight when the values are all small, and, therefore, we get an
advantage when one of the values is relatively big. In particular,
Pr [t is saved] ≥ 1 −
Y
(1 − xu )
u∈Pt
≥ 1 − (1 − xv )e−(yt −xv )
≥ 1 − (1 − 0.2)e−(yt −0.2)
≥ 1.01(1 − 1/e)yv
From now on, we only consider those terminals t ∈ X such that yt ≥ 0.9 and xv < 0.2, for
all v ∈ Pt . We remark here that if the value of η is not 1/2, we can easily pick other suitable
thresholds instead of 0.9 and 0.2.
For each vertex v ∈ V , let X1 ⊆ X be the set of terminals that are saved by V1 , i.e. a vertex
t ∈ X1 if and only if t is a descendant of some vertex in V1 . Let X2 ⊆ X \ X1 contain the set of
terminals that are not saved by the first phase, but are saved by the second phase, i.e. t ∈ X2
if and only if t has some ancestor in V2 .
PrV1 ,V2 [t 6∈ X1 ∪ X2 ] = PrV1 ,V2 [t 6∈ X1 ] PrV1 ,V2 [t 6∈ X2 : t 6∈ X1 ]
For any terminal t, let St0 and St00 be the sets of ancestors of t that are η-light and η-heavy
respectively, i.e. ancestors in St0 and St00 are considered by the algorithm in Phase 1 and 2
0
respectively. By Proposition 21, we can upper bound the first term by e−x(St ) . In the rest of
13
00
this section, we show that the second term is upper bounded by e−x(St ) c for some c < 1, and
0
00
therefore Pr [t 6∈ X1 ∪ X2 ] ≤ ce−x(St )−x(St ) ≤ ce−yt , as desired.
The following lemma is the main technical tool we need in the analysis. We remark that
this lemma is the main difference between (LP’) and (LP-2).
Lemma 23. Let t ∈ X and Lj be a layer containing some η-heavy ancestor of t. Then
EV1 [x(L̃j ) | t 6∈ X1 ] ≤ α
1
+ (1 − e−1/2 ) ≤ 0.9
2
Intuitively, this lemma says that any terminal that is still not saved by the result of the first
phase will have a relatively “sparse” layer above it. We defer the proof of this lemma to the
next subsection. Now we proceed to complete the analysis.
For each vertex v, denote by `(v) the layer to which vertex v belongs. For a fixed choice of
P
V1 , we say that terminal t is partially protected by V1 if v∈St00 xv x(L̃`(v) ) ≤ Cx(St00 ) (we will
choose the value of C ∈ (α, 1) later). Let X 0 ⊆ X \ X1 denote the subset of terminals that are
partially protected by V1 .
for α =
Claim 24. For any t ∈ X , PrV1 [t ∈ X 0 | t 6∈ X1 ] ≥ 1 − α/C.
Proof. By linearity of expectation and Lemma 23,
EV1
X
xv x(L̃`(v) ) | t 6∈ X1 =
v∈St00
X
h
i
xv EV1 x(L̃`(v) ) | t 6∈ X1 ≤ αx(St00 )
v∈St00
Using Markov’s inequality,
PrV1
X
xv x(L̃`(v) ) ≤ Cx(St00 ) | t 6∈ X1
v∈St00
= 1 − Pr
≥1−
=1−
X
xv x(L̃`(v) ) > Cx(St00 ) | t 6∈ X1
v∈St00
αx(St00 )
Cx(St00 )
α
C
We can now rewrite the probability of a terminal t ∈ X not being saved by the solution
after the second phase.
PrV1 ,V2 [t 6∈ X2 | t 6∈ X1 ]
= Pr t ∈ X 0 | t 6∈ X1 Pr t 6∈ X2 | t ∈ X 0 + Pr t 6∈ X 0 | t 6∈ X1 Pr t 6∈ X2 | t 6∈ X 0
α
00
≤ (1 − α/C)PrV1 ,V2 t 6∈ X2 | t ∈ X 0 + · e−x(St )
C
00
The last inequality holds because PrV1 ,V2 [t 6∈ X2 | t 6∈ X 0 ] is at most e−x(St ) from Proposition 21.
It remains to provide a better upper bound for Pr [t 6∈ X2 | t ∈ X 0 ]. Consider a vertex
v ∈ St00 that is involved in the second phase rounding. We say that vertex v is good for t and
14
V1 if x(L̃`(v) ) ≤ C 0 (we will choose the value C 0 ∈ (C, 1) later.) Denote by Stgood ⊆ St00 the set
of good ancestors of t. The following claim ensures that good ancestors have large LP-weight
in total.
Claim 25. For any node t ∈ X 0 , x(Stgood ) =
P
xv ≥ (1 − C/C 0 )x(St00 ).
v∈Stgood
Proof. Suppose (for contradiction) that the fraction of good layers was less than 1 − C/C 0 . This
means that x(St00 \ Stgood ) ≥ C/C 0 . For each such v ∈ St00 \ Stgood , we have x(L̃(v)) > C 0 . Then,
P
P
x C 0 ≥ C. This contradicts the assumption that t is partially
v∈St00 xv x(L̃`(v) ) >
v∈St00 \Stgood v
protected, and concludes our proof.
Now the following lemma follows.
Lemma 26. PrV1 ,V2 [t 6∈ X2 | t ∈ X 0 ] ≤ e−x(St ) e−(1− C 0 )x(St )( C 0 −1)
C
00
00
1
Proof.
PrV1 ,V2 t 6∈ X2 | t ∈ X 0
=
PrV1 V1 = V10 PrV2 t 6∈ X2 | V1 = V10
X
V10 :t∈X 0
≤
PrV1 V1 = V10
X
V10 :t∈X 0
≤
V10 :t∈X 0
≤
X
bad v∈St00
PrV1
1−
good v∈St00
e−xv
Y
Y
bad v∈St00
0
PrV1 V1 = V1
X
(1 − xv )
Y
Y
e−xv /C
xv
C0
0
good v∈St00
0
00
0
00 C
V1 = V10 e−x(St ) C 0 −(1−C/C )x(St )/C
V10 :t∈X 0
00
C
0
00
≤ e−x(St ) C 0 −(1−C/C )x(St )/C
0
X
PrV1 V1 = V10
V10 :t∈X 0
00
C
0
00
0
≤ e−x(St ) C 0 −(1−C/C )x(St )/C
1
0
00
00
≤ e−x(St ) e−(1−C/C )x(St )( C 0 −1)
Now we choose the parameters C and C 0 such that C = (1 + δ)α, C 0 = (1 + δ)C, and
(1 + δ)C 0 = 1, where (1 + δ)3 = 1/α. Notice that this choice of parameters satisfy our previous requirements that α < C < C 0 < 1. The above lemma then gives the upper bound of
00
δ2
00
00
e−x(St ) e− 1+δ x(St ) , which is at most e−(1+δ /2)x(St ) . Since δ > 0 is a constant, notice that we
do have an advantage over the standard LP rounding in this case. Now we plug in all the
parameters to obtain the final result.
2
PrV1 ,V2 [t 6∈ X1 ∪ X2 ] = PrV1 ,V2 [t 6∈ X1 ] PrV1 ,V2 [t 6∈ X2 | t 6∈ X1 ]
0
(1 − α/C) PrV1 ,V2 t 6∈ X2 | t ∈ X 0 +
0
(1 − α/C) e−x(St ) e− 2 x(St ) +
≤ e−x(St )
≤ e−x(St )
≤ e−yt
≤ e−yt
00
δ2
00
δ2
(1 − α/C) e− 2 x(St ) + α/C
00
δ −(1+δ2 /2)x(St00 )
1
e
+
1+δ
1+δ
15
α −x(St00 )
e
C
α −x(St00 )
e
C
Since we assume that yt > 0.9 and xv ≤ 0.2, we must have x(St00 ) ≥ 0.2, and therefore the
above term can be seen as e−yt · δ 0 for some δ 0 < 1. Overall, the approximation factor we get is
(1 − δ 0 /e) for some universal constant δ 0 ∈ (0, 1).
4.3.1
Proof of Lemma 23
For each u, let Eu denote the event that u is not saved by V1 . First we break the expectation
P
term into u∈Lj xu Pr [Eu | t 6∈ X1 ]. Let v ∈ L be the ancestor of t in layer Lj . We break down
the sum further based on the “LP coverage” of the least common ancestor between u and v, as
follows:
k/2
X
xu Pr [Eu | t 6∈ X1 ]
X
i=0 u∈Lj :q 0 (lca(u,v))=i
Here, q 0 (u) denotes k · x(Pu ); this term is integral since we consider the 1/k-integral solution
(x, y). The rest of this section is devoted to upper bounding the term Pr [Eu | t 6∈ X1 ]. The
following claim gives the bound based on the level i to which the least common ancestor belongs.
Claim 27. For each u ∈ Lj such that q 0 (lca(u, v)) = i,
Pr [Eu | t 6∈ X1 ] ≤ e−(1/2−i/k)
Proof. First, we recall that yu ≥ 1/2 and q 0 (u) ≥ k/2, since u is in the 1/2-heavy layer Lj . Let
w = lca(u, v) and P 0 be the path that connects w to u. Moreover, denote by S ⊆ P 0 the set of
P
P
light vertices on the path P 0 , i.e. S = St0 ∩ P 0 . Notice that x(S) ≥ a∈St0 ∩Pu xa − a∈Pw xa ≥
(1/2 − i/k).
For each w0 ∈ S, Pr [w0 6∈ V1 | t 6∈ X1 ] ∈ {1 − xw0 , 1 − xw0 /(1 − xv0 )} depending on whether
there is a vertex v 0 in Pv that shares a layer with w0 . In any case, it holds that Pr [w0 6∈ V1 | t 6∈ X1 ] ≤
(1 − xw0 ). This implies that
Pr [Eu | t 6∈ X1 ] ≤
Y
Pr w0 6∈ V1 | t 6∈ X1
w0 ∈S
≤
Y
(1 − xw0 )
w0 ∈S
≤
Y
e−xw0
w0 ∈S
−(1/2−i/k)
≤e
Claim 28. Let i be an integer and L0 ⊆ Lj be the set of vertices u such that q 0 (lca(u, v)) is at
least i. Then x(L0 ) ≤ (k − i)/k.
Proof. This claim is a consequence of Hartke’s constraints. Let v 0 be the topmost ancestor of v
such that q 0 (v 0 ) ≥ i. We remark that all vertices in L0 must be descendants of v 0 , so it must be
P
that w∈Pv0 xw + x(L0 ) ≤ 1. The first term is i/k, implying that x(L0 ) ≤ (k − i)/k.
Let Lij ⊆ Lj denote the set of vertices u whose least common ancestor lca(u, v) satisfies
P
0
q 0 (lca(u, v)) = i. As a consequence of Claim 28, i0 ≥i x(Lij ) ≤ (k − i)/k. Combining this
inequality with Claim 27, we get that
h
i
E x(L̃j ) | t 6∈ X1 ≤
k/2
X
x(Lij )e−1/2+i/k
i=0
16
k/2
This term is maximized when x(Lj ) = 1/2 and x(Lij ) = 1/k for all other i = 0, 1, . . . , k/2−
1. This implies that
k/2−1
h
i
E x(L̃j ) | t 6∈ X1 ≤ 1/2 +
X
e−1/2+i/k /k
i=0
Finally, using some algebraic manipulation and the fact that 1 + x ≤ ex , we get
k/2−1
h
i
E x(L̃j ) | t 6∈ X1 ≤ 1/2 +
X
e−1/2+i/k /k
i=0
1 −1/k 1 − e−1/2
e
k
1 − e−1/k
1/k
1
= 1/2 + (1 − e−1/2 ) 1/k
e
1 − e−1/k
1/k
= 1/2 + (1 − e−1/2 ) 1/k
e −1
≤ 1/2 + (1 − e−1/2 )
= 1/2 +
4.4
Integrality Gap for (LP’)
In this section, we present an instance where (LP’) has an integrality gap of 5/6 + , for any
> 0. Interestingly, this instance admits an optimal 21 -integral LP solution.
Gadget: The motivation of our construction is a simple gadget represented in Figure 2.
In this instance, vertices are either special (colored gray) or regular. This gadget has three
properties of our interest:
• If we assign an LP-value of xv = 1/2 to every special vertex, then this is a feasible LP
solution that ensures yu = 1 for all leaf u.
• For any integral solution U that does not pick any vertex in the first layer of this gadget,
at most 2 out of 3 leaves of the gadget are saved.
• Any pair of special vertices in the same layer do not have a common ancestor inside this
gadget.
Our integrality gap instance is constructed by creating partially
overlapping copies of this gadget. We describe it formally below.
Construction: The first layer of this instance, L1 , contains 4 nodes:
two special nodes, which we name a(1) and a(2), and two regular
nodes, which we name b(1) and b(2). We recall the definition of spider
from Section 3.1.
Let α = 5 d1/e. The nodes b(1) and b(2) are the roots of two Figure 2 Gadget used
spiders. Specifically, the spider Z1 rooted at b(1) has α feet, with one to get 5/6 integrality
foot per layer, in consecutive layers L2 , . . . , Lα+1 . For each j ∈ [α], gap. Special vertices are
denote by b0 (1, j), the j th foot of spider Z1 . The spider Z2 , rooted at colored gray.
b(2), has α2 feet, with one foot per layer, in layers Lα+2 , . . . , Lα2 +α+1 .
For each j ∈ [α2 ], denote by b0 (2, j), the j th foot of spider Z2 . All the
feet of spiders Z1 and Z2 are special vertices.
0 , with α2 feet, lying in the α2
For each j ∈ [α], the node b0 (1, j) is also the root of spider Z1,j
consecutive layers L2+α+jα2 , . . . , L1+α+(j+1)α2 (one foot per layer). For j 0 ∈ [α2 ], let b00 (1, j, j 0 )
17
0 that lies in layer L
3
denote the j 0 -th foot of spider Z1,j
1+α+jα2 +j 0 . Notice that we have α such
feet of these spiders
[α2 ],
n
0
Z1,j
b0 (2, j)
oα
j=1
lying in layers L2+α+α2 , . . . , L1+α+α2 +α3 . Similarly, for each
0
j ∈
the node
is the root of spider Z2,j
with α2 feet, lying in consecutive layers
L2+α+α3 +jα2 , . . . , L1+α+α3 +(j+1)α2 . We denote by b00 (2, j, j 0 ) the j 0 -th foot of this spider.
The special node a(1) is also the root of spider W1 which has α + α3 feet: The first α feet,
denoted by a0 (1, j) for j ∈ [α], are aligned with the nodes b0 (1, j), i.e. for each j ∈ [α], the foot
a0 (1, j) of spider W1 is in the same layer as the foot b0 (1, j) of Z1 . For each j ∈ [α], j 0 ∈ [α2 ], we
also have a foot a00 (1, j, j 0 ) which is placed in the same layer as b00 (1, j, j 0 ). Similarly, the special
node a(2) is the root of spider W2 having α2 + α4 feet. For j ∈ [α2 ], spider W2 has a foot a0 (2, j)
placed in the same layer as b0 (2, j). For j ∈ [α2 ], j 0 ∈ [α2 ], W2 also has a foot a00 (2, j, j 0 ) in the
layer of b00 (2, j, j 0 ). All the feet of both W1 and W2 are special vertices.
Finally, for i ∈ {1, 2}, and j ∈ [αi ], each node a0 (i, j) has α5−i children, which are leaves
of the instance. For j ∈ [α], j 0 ∈ [α2 ], the nodes b00 (i, j, j 0 ), a00 (i, j, j 0 ) have α3−i children each
which are also leaves of the instance. The set of terminals X is simply the set of leaves.
Proposition 29. We have |X | = 6α5 . Moreover, (i) the number of terminals in subtrees
Ta(1) ∪ Tb(1) is 3α5 , and (ii) the number of terminals in subtrees Ta(2) ∪ Tb(2) is 3α5 .
Proof. Each node a0 (1, j) has α4 children, and there are α such nodes. Similarly, each node
a0 (2, j) has α3 children. There are α2 such nodes. This accounts for 2α5 terminals.
For i ∈ {1, 2}, each node a00 (i, j, j 0 ) has α3−i children. There are αi+2 such nodes. This
accounts for another 2α5 terminals. Finally, there are α3−i children of each b00 (i, j, j 0 ), and
there are α2+i such nodes.
Fractional Solution: Our construction guarantees that any path from root to leaf contains 2
special vertices: For a leaf child of a0 (i, j), its path towards the root must contain a0 (i, j) and
a(i). For a leaf child of a00 (i, j, j 0 ), its path towards the root contains a00 (i, j, j 0 ) and a(i). For a
leaf child of b00 (i, j, j 0 ), the path towards the root contains b00 (i, j, j 0 ) and b0 (i, j).
Lemma 30. For each special vertex v, for each layer Lj below v, the set Lj ∩ Tv contains at
most one special vertex.
Proof. Each layer contains two special vertices of the form {a0 (i, j), b0 (i0 , j 0 )} or {a00 (i, j), b00 (i0 , j 0 )}.
In any case, the least common ancestor of such two special vertices in the same layer is always
the root s (since one vertex is in Ta(i) , while the other is in Tb(i) ) This implies that, for any
non-root vertex v, the set Lj ∩ Tv can contain at most one special vertex.
Notice that, there are at most two special vertices per layer. We define the LP solution x,
with xv = 1/2 for all special vertices v and xv = 0 for all other vertices. It is easy to verify that
this is a feasible solution.
P
We now check the constraint at v and layer Lj below v: If the sum u∈Pv xu = 0, then the
P
P
constraint is immediately satisfied, because u∈Lj ∩Tv xu ≤ 1. If u∈Pv xu = 1/2, let v 0 be the
P
P
special vertex ancestor of v. Lemma 30 guarantees that u∈Lj ∩Tv xu ≤ u∈Lj ∩Tv0 xu ≤ 1/2,
P
and therefore the constraint at v and Lj is satisfied. Finally, if u∈Pv xu = 1, there can be no
P
special vertex below v and therefore u∈Lj ∩Tv xu = 0.
Integral Solution: We argue that any integral solution cannot save more than (1 + 5/α)5α5
terminals. The following lemma is the key to our analysis.
Lemma 31. Any integral solution U : U ∩{a(1), b(1)} = ∅ saves at most (1+5/α)5α5 terminals.
18
Proof. Consider the set Q = {a0 (1, j)}αj=1 ∪{b0 (1, j)}αj=1 , and a collection of paths from {a(1), b(1)}
to vertices in set Q. These paths are contained in the layers L1 , . . . , Lα+1 , so the strategy U
induces a cut of size at most α + 1 on them. This implies that at most α + 1 vertices (out of
2α vertices in Q) can be saved by U. Let Q̃ ⊆ Q denote the set of vertices that have not been
saved by U. We remark that |Q̃| ≥ α − 1. We write Q̃ = Q̃a ∪ Q̃b where Q̃a contains the set of
vertices a0 (1, j) that are not saved, and Q̃b = Q̃ \ Q̃a . For each vertex in Q̃a , at least α4 − 1 of
its children cannot be saved, so we have at least (α4 − 1)|Q̃a | ≥ α4 |Q̃a | − α unsaved terminals
that are descendants of Q̃a . If |Q̃b | ≤ 3, we are immediately done: We have |Q̃a | ≥ α − 4, so
(α4 − 1)(α − 4) ≥ α5 − 5α4 unsaved terminals.
Consider the set
R=
[
00
a (1, j, j 0 ) ∪
j∈[α],j 0 ∈[α2 ]
[
[
b00 (1, j, j 0 )
j:b0 (1,j)∈Q̃b j 0 ∈[α2 ]
This set satisfies |R| = α3 + |Q̃b |α2 , and the paths connecting vertices in R to Q̃b ∪ {a(1)} lie
in layers L1 , . . . , Lα3 +α2 +α+1 . So the strategy U induced on these paths disconnects at most
α3 + α2 + α + 1 vertices. Let R̃ ⊆ R contain the vertices in R that are not saved by U, so we
have |R̃| ≥ (|Q̃b | − 1)α2 − α − 1, which is at least (|Q̃b | − 2)α2 . Each vertex in R̃ has α2 children.
We will have (α2 − 1) unsaved terminals for each such vertex, resulting in a total of at least
(α2 − 1)(|Q̃b | − 2)α2 ≥ α4 |Q̃b | − 4α4 terminals that are descendants of b(1).
In total, by summing the two cases, at least (α4 |Q̃a |−α)+(α4 |Q̃b |−4α4 ) ≥ (|Q̃a |+|Q̃b |)α4 −
4
5α ≥ α5 − 5α4 terminals are not saved by U, thus concluding the proof.
Lemma 32. Any integral solution U : U ∩{a(2), b(2)} = ∅ saves at most (1+5/α)5α5 terminals.
Since nodes a(1), a(2), b(1), b(2) are in the first layer, it is only possible to save one of them.
Therefore, either Lemma 31 or Lemma 32 apply, which concludes the analysis.
5
Conclusion and Open Problems
In this paper, we settled the integrality gap question for the standard LP relaxation. Our results
ruled out the hope to use the canonical LP to obtain better approximation results. While a
recent paper settled the approximability status of the problem [1], the question whether an
improvement over (1−1/e) can be done via LP relaxation is of independent interest. We provide
some evidences that Hartke’s LP is a promising candidate for doing so. Another interesting
question is to find a more general graph class that admits a constant approximation algorithm.
We believe that this is possible for bounded treewidth graphs.
References
[1] Adjiashvili, D., Baggio, A., Zenklusen, R.: Firefighting on Trees Beyond Integrality Gaps.
ArXiv e-prints (Jan 2016)
[2] Anshelevich, E., Chakrabarty, D., Hate, A., Swamy, C.: Approximability of the firefighter
problem - computing cuts over time. Algorithmica 62(1-2), 520–536 (2012)
[3] Bazgan, C., Chopin, M., Cygan, M., Fellows, M.R., Fomin, F.V., van Leeuwen, E.J.:
Parameterized complexity of firefighting. JCSS 80(7) (2014)
19
[4] Bazgan, C., Chopin, M., Ries, B.: The firefighter problem with more than one firefighter
on trees. Discrete Applied Mathematics 161(7-8), 899–908 (2013)
[5] Cai, L., Verbin, E., Yang, L.: Firefighting on trees: (1-1/e)-approximation, fixed parameter
tractability and a subexponential algorithm. In: ISAAC (2008)
[6] Chalermsook, P., Chuzhoy, J.: Resource minimization for fire containment. In: SODA
(2010)
[7] Chekuri, C., Kumar, A.: Maximum coverage problem with group budget constraints and
applications. In: APPROX 2004. pp. 72–83 (2004)
[8] Chlebíková, J., Chopin, M.: The firefighter problem: A structural analysis. In: Cygan, M.,
Heggernes, P. (eds.) IPEC. Lecture Notes in Computer Science, vol. 8894, pp. 172–183.
Springer (2014)
[9] Costa, V., Dantas, S., Dourado, M.C., Penso, L., Rautenbach, D.: More fires and more
fighters. Discrete Applied Mathematics 161(16–17), 2410 – 2419 (2013)
[10] Develin, M., Hartke, S.G.: Fire containment in grids of dimension three and higher. Discrete
Applied Mathematics 155(17), 2257–2268 (2007)
[11] Finbow, S., MacGillivray, G.: The firefighter problem: a survey of results, directions and
questions. Australas. J. Combin 43, 57–77 (2009)
[12] Hartke, S.G.: Attempting to narrow the integrality gap for the firefighter problem on trees.
Discrete Methods in Epidemiology 70, 179–185 (2006)
[13] Hartnell, B.: Firefighter! an application of domination. In: Manitoba Conference on Combinatorial Mathematics and Computing (1995)
[14] Iwaikawa, Y., Kamiyama, N., Matsui, T.: Improved approximation algorithms for firefighter problem on trees. IEICE Transactions 94-D(2), 196–199 (2011)
[15] King, A., MacGillivray, G.: The firefighter problem for cubic graphs. Discrete Mathematics
310(3), 614–621 (2010)
[16] Wang, P., Moeller, S.A.: Fire control on graphs. Journal of Combinatorial Mathematics
and Combinatorial Computing 41, 19–34 (2002)
20
A
Omitted Figures
21
k
kD
D2
D
D
. . . . . . . . . . . . . . . .
Figure 3 Simplified example of the instance used to achieve integrality gap of 1 − 1/e, when k = 2 and
D = 2. The labels in the figure indicate, in general, the number of edges in that location, in terms of k
and D. Special vertices are colored gray.
y1
y2
x2
x1
... ...
...
...
...
...
...
...
Figure 4 Simplified example of the instance with low integrality gap for 1/2-integral solutions. Special
vertices are colored gray.
22
| 8cs.DS
|
The Simulation and Mapping of Building Performance Indicators based
on European Weather Stations
A.W.M. (Jos) van Schijndel
Department of the Built Environment, Eindhoven University of Technology, Netherlands
Keywords: mapping, modeling, building, performance
ABSTRACT
Due to the climate change debate, a lot of research and maps of external climate parameters are available. However, maps of
indoor climate performance parameters are still lacking. This paper presents a methodology for obtaining maps of
performances of similar buildings that are virtually spread over whole Europe. The produced maps are useful for analyzing
regional climate influence on building performance indicators such as energy use and indoor climate. This is shown using the
Bestest building as a reference benchmark. An important application of the mapping tool is the visualization of potential
building measures over the EU. Also the performances of single building components can be simulated and mapped. It is
concluded that the presented method is efficient as it takes less than 15 minutes to simulate and produce the maps on a
2.6GHz/4GB computer. Moreover, the approach is applicable for any type of building.
1. Introduction
Buildings and their systems can be considered as complex
dynamic systems. The control and operation of building
systems are of eminent importance regarding durability,
health and costs. Energy use can be drastically reduced if
building systems are properly controlled and optimal
operated. At the same time comfort can be improved and
costs can significantly be lowered. The goal of
Computational Building Physics is to search for physics
based models behind the empirical facts in order to control
the built environment. Furthermore Computational Building
Physics studies the built environment on several scales (see
Figure 1)
Figure 1. Scale levels involved within building physics
research
Building physics research applications are so far almost
always limited to the scale levels of material, building and
urban area. In this paper we present a building physics
research application of the scale level of the EU.
1.1
Related work
In this Section we will focus on two important building
related research areas where EU mappings are already
common techniques. First, we start with cultural heritage and
climate change. Grossi et al. (2007) are using maps to
visualize the prediction of the evolution in frost patterns due
to climate change during the 21st century and the potential
damage to historic structures and archeological remains in
Europe. Figure 2 shows an exemplarily result of the
application of a freezing event map.
Figure 2. Pan-European maps of average yearly freezing
events in 30 years period 1961–1990 (top) and far future
2070–2099 (bottom) by Grossi et al. (2007).
Similar maps as presented in Figure 2 are used to show the
expected reduction of freezing and lowering the potential for
frost shattering of porous building stone. The underlying data
for these maps are based on regional climate models. This is
the second research area where EU maps are commonly used.
There is an enormous amount of literature on climate change
and mapping. Therefore we illustrate the use of these maps
by one state of the art regional climate model: REMO (Jacob
1997, Larsen 2010). Figure 3 shows the twenty-five-year
mean modeled wind at 10 m height over Central Europe
using REMO with a 10 km resolution.
Figure 4. The distributions of the locations of the external
climates in Europe.
Figure 3. Twenty-five-year mean modeled wind at 10 m
height over the entire domain REMO 10 km resolution
(Larsen 2010)
Maps like figure 3 are suited for wind energy assessment
application in Northern Europe. Moreover, literature of the
related work shows that a lot of EU maps of external climate
parameters are available.
Each climate file includes hourly based values for the
common used external climate parameters: Horizontal global
solar radiation [W/m2] (ISGH), Diffuse solar radiation
Air temperature
[W/m2] (ISD), Cloud cover [0-1] (CI),
[oC] (TA), Relative humidity [%] (HREL), Wind speed [m/s]
(WD), Rain intensity
(WS), Wind direction [0-360o]
[mm/h] (RN), Long wave radiation [W/m2] (ILAH).
1.2
2.2
Goal and Outline
The maps presented in the previous section are all based on
external climate parameters. However, the goal of this work
is produce maps that include also indoor climate performance
parameters.
The outline of the paper is as follows: Section 2 presents the
methodology of the application for obtaining maps of
performances of similar buildings that are virtual spread over
whole Europe. The produced maps are useful for analyzing
regional climate influence on building performance indicators
such as energy use and indoor climate. Section 3 shows a
benchmark of the EU mapping of the Bestest building.
Section 4 comprehends a case study on the simulation and
mapping of the effect of improved glazing using the Bestest
building as a reference.
2. Methodology
The methodology used for obtaining the required simulation
results and maps can be divided into three steps. These are
presented in the following sections.
2.1
External climate files
Over 130 external hourly-based climate files were produced
using commercially available software (Meteonorm 2011)
using so-called wac format. Figure 4 presents the distribution
of the locations over Europe.
Whole building simulation model
The whole building model originates from the thermal indoor
climate model ELAN which was already published in 1987
(de Wit et al. 1988). Separately a model for simulating the
indoor air humidity (AHUM) was developed. In 1992 the two
models were combined and programmed in the MatLab
environment. Since that time, the model has constantly been
improved using newest techniques provided by recent
MatLab versions. The current hourly-based model
HAMBase, is part of the Heat, Air and Moisture Laboratory
(HAMLab 2011), and is capable of simulating the indoor
temperature, the indoor air humidity and energy use for
heating and cooling of a multi-zone building. The physics of
this model is extensively described by de Wit (2006). The
main modeling considerations are summarized below. The
HAMBase model uses an integrated sphere approach. It
reduces the radiant temperatures to only one node. This has
the advantage that complicated geometries can easily be
modeled. Figure 5 shows the thermal network.
ΣΦxy
ΣΦab
Tx
Ta
Lxa
-h Φ /h
Φcv
cv r r
Φr+hcvΦr/hr
Ca
Figure 5. The room radiative model as a thermal network
Where Ta is the air temperature and Tx is a combination of air
and radiant temperature. Tx is needed to calculate
transmission heat losses with a combined surface coefficient.
hr and hcv are the surface weighted mean surface heat transfer
coefficients for convection and radiation. Φr and Φcv are
respectively the radiant and convective part of the total heat
input consisting of heating or cooling, casual gains and solar
gains.
For each heat source a convection factor can be given
between 0 and 1 by the user. For example, for air heating this
factor is close to 1 and for radiative heat sources a factor
ranging from 0.4 – 0.6 can be used. The factor for solar
radiation depends on the window system and the amount of
radiation falling on furniture. Ca is the heat capacity of the
air. Lxa is a coupling coefficient (de Wit et al. 1988):
h
Lxa = At hcv 1 + cv
hr
(1)
∑Φab is the heat loss by air entering the zone with an air
temperature Tb. At is the total area. In case of ventilation Tb
is the outdoor air temperature. ∑Φ xy is transmission heat loss
through the envelope part y. For external envelope parts Ty is
the sol-air temperature for the particular construction
including the effect of atmospheric radiation.
The admittance for a particular frequency can be represented
by a network of a thermal resistance (1/Lx) and capacitance
(Cx) because the phase shift of Yx can never be larger than
π/2. To cover the relevant set of frequencies (period 1 to 24
hours) two parallel branches of such a network are used
giving the correct admittance's for cyclic variations with a
period of 24 hours and of 1 hour. This means that the heat
flow Φxx(tot) is modeled with a second order differential
equation. For air from outside the room with temperature Tb a
loss coefficient Lv is introduced. This model is summarized in
Figure 6.
Ty
Τb
Lyx
∆Φxy
Lv
Tx
Ta
Lxa
The HAMbase model is adapted in such a way that all
climate (.wac) files in a directory are automatically
processed. For each climate file a corresponding output file is
produced containing hourly based values for the indoor
climate and heating and cooling power.
A separate Matlab mfile is developed for calculating annual
means and peak values for each location (i.e. wac file) and
together with the longitude and latitude stored in a single file
suitable for mapping purposes.
2.3
Mapping of the results
A MatLab mfile was developed for the visualization of the
just mentioned mapping file. For the exact details of this
mfile, we refer to the HAMLab website (HAMLab 2012).
3. Benchmark: EU Mapping of the Bestest
building
3.1
Bestest using HAMBase
The Bestest (ASHRAE, (2001)) is a structured approach to
evaluate the performance of building performance simulation
tools. The evaluation is performed by comparing results of
the tested tool relative to results by reference tools. The
procedure requires simulating a number of predefined and
hierarchal ordered cases. Firstly, a set of qualification cases
have to be modeled and simulated. If the tool passes all
qualification cases the tool is considered to perform Bestest
compliant. In case of compliance failure the procedure
suggests considering diagnostic cases to isolate its cause.
Diagnostic cases are directly associated with the qualification
cases (Judkoff and Neymark 1995). The first qualification
case, case 600 (see Figure 7) was used for the performance
comparison.
Φp1
Φg1
Φg2
Φp2
b) the moisture storage in walls and furniture, carpets etc is
dependent on the relative humidity and temperature.
Lx2
L x1
Figure 7. Bestest case 600 geometry
Cx1
Cx2
Ca
Figure 6. The thermal ventilation model for one zone
In a similar way a model for the air humidity is made. Only
vapour transport is modeled, the hygroscopic curve is
linearized between RH 20% and 80%. The vapour
permeability is assumed to be constant. The main differences
are: a) there is only one room node (the vapour pressure) and
The thermal part of HAMBase has been subjected to a
standard method of test (Bestest ASHRAE, (2001)), with
satisfactory results. The accompanying climate file of the
Bestest is based on weather station near Denver (USA). For
further details, see Table I.
Table I Comparison of the HAMBase model with some cases
of the standard test.
Case
600ff
600ff
600ff
900ff
900ff
Nr. Simulation of
model
o
mean indoor temperature [ C] 25.1
o
min. indoor temperature [ C] -17.9
o
max. indoor temperature [ C] 64.0
o
mean indoor temperature [ C] 25.1
test min ..max
24.2 ..
25.9
-18.8 ..
-15.6
64.9 ..
69.5
24.5..
25.9
o
-5.1
-6.4..
-1.6
o
min. indoor temperature [ C]
900ff
max. indoor temperature [ C]
43.5
41.8..
44.8
600
annual sensible heating [MWh] 4.9
4.3..
5.7
600
annual sensible cooling [MWh] 6.5
6.1..
8.0
600
peak heating [kW]
4.0
3.4..
4.4
600
peak sensible cooling [kW]
5.9
6.0..
6.6
900
annual sensible heating [MWh] 1.7
1.2..
2.0
900
annual sensible cooling [MWh] 2.6
2.1 ..
3.4
900
peak heating [kW]
2.9 ..
3.9
3.2
3.5
EU mapping
Now we apply the method of Section 2 in order to produce
EU maps of certain (performance) parameters.
3.2.1
External climate
We start with the outdoor climate. For these maps simulation
is not required because the information is already present in
the climate (.wac) files. Figure 8 shows the expected northsouth temperature distribution. The influence of higher
altitude weather stations near the Alps is also visible.
Figure 9. Mean annual external relative humidity [%] (This
figure relies on color, see digital version of the paper)
Quite remarkably there is a peak in relative humidity at the
east part of the Alps corresponding with the low temperature
(see Figure 8) which is not present at the west side of the
Alps.
3.2.2
Indoor Climate
The simulation of all 130 weather stations of the Bestest case
600 building takes less than 15 minutes using HAMBase on a
2.6GHz/4GB computer. Figure 10 shows the mean annual
indoor air temperature distribution.
Figure 10. Mean annual indoor air temperature [°C] of the
Bestest case 600 building (This figure relies on color, see
digital version of the paper)
Figure 8. Mean annual external air temperature [°C] (This
figure relies on color, see digital version of the paper)
The relative humidity distribution is shown in Figure 9.
As expected the mean annual indoor temperature is within the
20 to 27 oC range (the upper left part of the map near
Greenland is not accurate due to the absence of weather
stations). Although it is not within the Bestest, it could be
interesting to simulate the mean annual indoor relative
humidity (in case that there are no internal moisture gains).
This is presented in Figure 11.
Figure 11. Mean annual indoor relative humidity [%] of the
Bestest case 600 building (This figure relies on color, see
digital version of the paper)
Figure 13. Mean annual heating power [W] of the Bestest
case 600 building (This figure relies on color, see digital
version of the paper)
As expected at the cold regions (see Figure 8) the indoor
relative humidity is low.
Because the Bestest case 600 building has 12 m2 of window
area facing south it is interesting to visualize the mean annual
solar gain [W] entering the room. This is shown in Figure 12.
Figure 14. Mean annual cooling power [W] of the Bestest
case 600 building (see also digital version of the paper for
color)
Figure 12. Mean annual solar gain [W] of the Bestest case
600 building (This figure relies on color, see digital version
of the paper)
Figure 12 is in line with what we expect.
3.2.3
Heating & Cooling
The energy use for heating and cooling the Bestest case 600
building is shown in figure 13 through 16. The mean annual
heating and cooling power maps are quite as expected. Also
the peak heating power (Figure 15) has an intuitive map.
Remarkably, the peak cooling power map (Figure 16) shows
a quite uniform distribution ranging from 5.5-7 kW.
Figure 15. Peak heating power [W] the of Bestest case 600
building (see also digital version of the paper for color)
Figure 16. Peak cooling power [W] of the Bestest case 600
building (see also digital version of the paper for color)
4. EU Mapping of single building components
effects
In the previous Section the main performances of a reference
building model subjected to EU weather stations data, were
visualized using EU maps. This leads to a new interesting
application. If we change a single building component we can
visualize the single effect of this change by subtracting these
(new) results with the reference case. In order to demonstrate
this idea we study the effect of replacing the original glazing
(U value = 3 W/m2K; Solar gain factor = 0.787) of the case
600 building with improved glazing (U value = 1.5 W/m2K;
Solar gain factor = 0.40) called variant.
Figure 17. Reduction of mean annual solar gain [W] of the
Bestest case 600 building due to the improved glazing (see
also digital version of the paper for color)
4.2
4.2.1
Heating & Cooling
Mean annual heating reduction
Figures 18 through 20 show the three types of maps for the
mean annual heating. Figure 18 provides the mean annual
heating of the building with improved glazing:
The method to obtain the results below was as follows: (1)
The building was simulated and mapped again using the new
glazing parameters. (2) A net change effect map was
obtained by subtracting two maps, i.e. reference - variant
map. (3) A percentage change effect map was obtained by
using the formula: (1- variant/reference) * 100% map.
4.1
Indoor climate change
The results show only minor changes in the mean indoor
climates of respectively 1 oC and 1%. Because the solar gain
factor is reduced from 0.787 to 0.40 we expect substantial
differences between the reference and the variant. Figure 17
shows the result:
Figure 18. Mean annual heating power [W] of the Bestest
case 600 building with improved glazing (see also digital
version of the paper for color)
Using the map of mean annual heating power of the reference
building (i.e. Figure 13), a net change effect map was
obtained by subtracting two maps, i.e. reference map minus
variant map. This is shown in figure 19.
4.2.2
Mean annual cooling reduction
Figure 21 provides the mean annual cooling reduction.
Figure 19. Change in mean annual heating power [W] of the
Bestest case 600 building with improved glazing (see also
digital version of the paper for color)
The highest reduction in heating amount can be observed in
the Nordic countries (up to 250Wy x 8.76 kh/y = 2190 kWh).
The lowest reduction can be found in the Mediterranean as
expected. The reduction can also be expressed as a
percentage change effect map, using the formula: (1variant/reference) * 100%. This is presented in figure 20.
Figure 20. Percent change in mean annual heating power [%]
of the Bestest case 600 building with improved glazing (see
also digital version of the paper for color)
Please note that although the Mediterranean has a relatively
high percent change, the absolute values for the mean annual
heating power are low, so almost no heating energy savings
are expected in this region. Secondly in areas where
improved glazing do have impact, the relative improvement
is quite uniform between 10 and 20%.
Figure 21. Change in mean cooling power [W] of the Bestest
case 600 building with improved glazing (see also digital
version of the paper for color)
The highest reduction in cooling amount can be observed in
the Mediterranean (up to -500Wy x 8.76 kh/y = 4400 kWh).
The reduction can also be expressed as a percentage change
effect map. This is presented in figure 22.
Figure 22. Percent change in mean annual cooling power [%]
of the Bestest case 600 building with improved glazing (This
figure relies on color, see digital version of the paper)
Although the percent change at the Alps is relatively high, the
absolute values for the mean annual cooling power are low,
so almost no cooling energy savings are expected in this
region.
4.2.3
Peak heating reduction
Figure 23 shows the relative the peak heating power
reduction.
concluded that the presented method efficient as it takes less
than 15 minutes to simulate and produce the maps on a
2.6GHz/4GB computer. Moreover, the approach is
applicable for any type of building
Future research
Within the mentioned EU FP7 project ‘Climate for Culture’,
detailed EU external climate files are currently under
development for the period 1960 – 2100 using the REMO
model (Jacob et al. 1997) and a moderate climate scenario.
With these future external climate files we will be able to
predict future building performance indicators. Together with
the EU mapping tool this could be helpful to locate EU
regions with the highest impact on the specific building
performances.
Figure 23. Reduction of the peak heating power [%] of the
Bestest case 600 building (see also digital version of the
paper for color)
Remarkably, figure 23 shows a quite uniform relative change
in peak heating power reduction.
4.2.4
Peak cooling reduction
Figure 24 shows the absolute peak cooling power reduction.
6. Acknowledgement
This work was inspired and sponsored by the EU FP7 project
‘Climate for Culture’.
7. References
ASHRAE (2001). Standard method of test for the evaluation
of building energy analysis computer programs, standard
140-2001.
Judkoff, R. and Neymark, J. (1995). “International energy
agency building energy simulation test (Bestest) and
diagnostic method”, National Renewable Energy
Laboratory, Golden, CO.
Schijndel, A.W.M. van & Wit, M.H. de, (1999). A building
physics toolbox in MatLab, 7TH Symposium on Building
Physics in the Nordic Countries Goteborg, pp 81-88
Grossi, C.M., Brimblecombe, P., Harris, I. (2007). Predicting
long term freeze-thaw risks on Europe built heritage and
archaeological sites in a changing climate. Science of the
Total Environment 377, pp 273-281
HAMLab
(2012), http://archbps1.campus.tue.nl/bpswiki/index.php/Ha
mlab
Jacob, D. and Podzun, R. (1997). Sensitivity studies with the
Regional Climate Model REMO. Meteorology and
Atmospheric Physics, 63, pp 119-129.
Larsen X.G., Mann, J., Berg, J., Gottel, H., Jacob, D. (2010).
Wind climate from the regional climate model REMO.
Wind Energy 13, pp 279-206
Figure 24. Reduction of the peak cooling power [W] of the
Bestest case 600 building (see also digital version of the
paper for color)
Meteonorm (2011).http:/meteonorm.com/
Figure 24 shows a uniform relative change in cooling power
reduction (3 – 4 kW).
Schijndel A.W.M. van (2007). Integrated Heat Air and
Moisture Modeling and Simulation, PhD Dissertation,
Eindhoven University of Technology
5. Conclusions
The produced maps are useful for analyzing regional climate
influence on building performance indicators such as energy
use and indoor climate. This is shown using the Bestest
building as a reference benchmark. An important application
of the mapping tool is the visualization of potential building
measures over the EU. Also the performances of single
building components can be simulated and mapped. It is
Wit M.H. de ,H.H. Driessen (1988). ELAN A Computer
Model for Building Energy Design. Building and
Environment 23, pp 285-289
Wit, M.H. de, (2006). HAMBase, Heat, Air and Moisture
Model for Building and Systems Evaluation, Bouwstenen
100, Eindhoven University of Technology
| 5cs.CE
|
Electronic Journal of Statistics
Vol. 0 (2017) 1–99
ISSN: 1935-7524
Errors-in-variables models with dependent
measurements
arXiv:1611.04701v2 [stat.ML] 1 Apr 2017
Mark Rudelson∗ and Shuheng Zhou†
Department of Mathematics, Department of Statistics
University of Michigan, Ann Arbor, MI 48109
e-mail: [email protected]; [email protected]
Abstract:
Suppose that we observe y ∈ Rn and X ∈ Rn×m in the following errors-invariables model:
y
=
X0 β ∗ +
X
=
X0 + W
where X0 is an n × m design matrix with independent subgaussian row vectors,
∈ Rn is a noise vector and W is a mean zero n × m random noise matrix with
independent subgaussian column vectors, independent of X0 and . This model is significantly different from those analyzed in the literature in the sense that we allow
the measurement error for each covariate to be a dependent vector across its n observations. Such error structures appear in the science literature when modeling the
trial-to-trial fluctuations in response strength shared across a set of neurons.
Under sparsity and restrictive eigenvalue type of conditions, we show that one is
able to recover a sparse vector β ∗ ∈ Rm from the model given a single observation matrix X and the response vector y. We establish consistency in estimating β ∗ and obtain
the rates of convergence in the `q norm, where q = 1, 2 for the Lasso-type estimator,
and for q ∈ [1, 2] for a Dantzig-type Conic programming estimator. We show error
bounds which approach that of the regular Lasso and the Dantzig selector in case the
errors in W are tending to 0. We analyze the convergence rates of the gradient descent
methods for solving the nonconvex programs and show that the composite gradient descent algorithm is guaranteed to converge at a geometric rate to a neighborhood of the
global minimizers: the size of the neighborhood is bounded by the statistical error in
the `2 norm. Our analysis reveals interesting connections between computational and
statistical efficiency and the concentration of measure phenomenon in random matrix
theory. We provide simulation evidence illuminating the theoretical predictions.
MSC 2010 subject classifications: Primary 60K35, 60K35; secondary 60K35.
Keywords and phrases: Errors-in-variable models, measurement error data, subgaussian concentration, matrix variate distributions, nonconvexity.
Contents
1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
∗
Mark Rudelson is partially supported by NSF grant DMS 1161372 and USAF Grant FA9550-14-10009.
† Shuheng Zhou is supported in part by NSF under Grant DMS-1316731 and Elizabeth Caroline Crosby
Research Award from the Advance Program at the University of Michigan.
1
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
2
Rudelson and Zhou/Errors-in-variables
1.1 The model and the method . . . . . . . . . . . . . . . . . . .
1.2 Gradient descent algorithms . . . . . . . . . . . . . . . . . .
1.3 Our contributions . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Assumptions and preliminary results . . . . . . . . . . . . . . . . .
3 Main results on the statistical error . . . . . . . . . . . . . . . . . .
3.1 Regarding the MA constant . . . . . . . . . . . . . . . . . . .
3.2 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Improved bounds when the measurement errors are small . . . . . .
4.1 Oracle results on the Lasso-type estimator . . . . . . . . . . .
4.2 Oracle results on the Conic programming estimator . . . . . .
5 Optimization error on the gradient descent algorithm . . . . . . . .
5.1 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Proof of theorems . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1 Stochastic error terms . . . . . . . . . . . . . . . . . . . . . .
6.2 Outline for proof of Theorem 3 . . . . . . . . . . . . . . . . .
6.3 Improved bounds for the corrected Lasso estimator . . . . . .
6.4 Outline for proof of Theorem 4 . . . . . . . . . . . . . . . .
6.5 Improved bounds for the DS-type estimator . . . . . . . . . .
7 Lower and Upper RE conditions . . . . . . . . . . . . . . . . . . .
8 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1 Relative error . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Corrected Lasso via GD versus Conic programming estimator
8.3 Sensitivity to tuning parameters . . . . . . . . . . . . . . . .
8.4 Statistical and optimization error in Gradient Descent . . . . .
9 Proof of Lemma 1 . . . . . . . . . . . . . . . . . . . . . . . . . . .
10 Proof of Theorem 3 . . . . . . . . . . . . . . . . . . . . . . . . . .
11 Proof of Theorem 4 . . . . . . . . . . . . . . . . . . . . . . . . . .
12 Proof of Theorem 6 . . . . . . . . . . . . . . . . . . . . . . . . . .
13 Proof of Theorem 7 . . . . . . . . . . . . . . . . . . . . . . . . . .
14 Proof of Theorem 9 . . . . . . . . . . . . . . . . . . . . . . . . . .
14.1 Proof of Corollary 10 . . . . . . . . . . . . . . . . . . . . . .
15 Proof of Theorem 12 . . . . . . . . . . . . . . . . . . . . . . . . .
16 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
B Proof of Theorem 2 . . . . . . . . . . . . . . . . . . . . . . . . . .
C Some auxiliary results . . . . . . . . . . . . . . . . . . . . . . . . .
C.1 Proof of Lemma 5 . . . . . . . . . . . . . . . . . . . . . . . .
C.2 Proof of Lemma 11 . . . . . . . . . . . . . . . . . . . . . . .
D Proof of Corollary 13 . . . . . . . . . . . . . . . . . . . . . . . . .
E Proof of Corollary 14 . . . . . . . . . . . . . . . . . . . . . . . . .
F Proof of Lemma 15 . . . . . . . . . . . . . . . . . . . . . . . . . .
F.1 Comparing the two type of RE conditions in Theorems 3 and 4
G Proof of Theorem 16 . . . . . . . . . . . . . . . . . . . . . . . . .
H Proofs for the Lasso-type estimator . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
6
7
9
11
16
19
20
22
24
27
28
29
31
31
33
35
35
37
37
38
40
43
44
49
50
52
53
54
55
56
57
60
61
62
62
62
67
68
71
73
73
74
75
76
78
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
3
Rudelson and Zhou/Errors-in-variables
H.1 Proof of Lemma 17 . . . . . . . . . . .
H.2 Proof of Lemma 18 . . . . . . . . . . .
I Proofs for the Conic Programming estimator .
I.1 Proof of Lemma 19 . . . . . . . . . . .
I.2 Proof of Lemma 20 . . . . . . . . . . .
I.3 Proof of Lemma 21 . . . . . . . . . . .
J Proof for Theorem 7 . . . . . . . . . . . . . .
J.1 Proof of Lemma 22 . . . . . . . . . . .
J.2 Proof of Lemma 23 . . . . . . . . . . .
J.3 Proof of Lemma 24 . . . . . . . . . . .
K Some geometric analysis results . . . . . . .
L Proof of Corollary 25 . . . . . . . . . . . . .
M Proof of Theorem 26 . . . . . . . . . . . . .
N Proof of Lemma 32 . . . . . . . . . . . . . .
O Proof of Lemmas 40 and 41 and Corollary 42
References . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
79
80
81
81
81
82
83
84
84
85
86
90
90
93
94
96
1. Introduction
The matrix variate normal model has a long history in psychology and social sciences.
In recent years, it is becoming increasingly popular in biology and genomics, neuroscience, econometric theory, image and signal processing, wireless communication,
and machine learning; see for example [15, 22, 17, 52, 5, 54, 18, 2, 26] and references
therein. We call the random matrix X, which contains n rows and m columns a single data matrix, or one instance from the matrix variate normal distribution. We say
that an n × m random matrix X follows a matrix normal distribution with a separable covariance matrix ΣX = A ⊗ B and mean M ∈ Rn×m , which we write
Xn×m ∼ Nn,m (M, Am×m ⊗ Bn×n ). This is equivalent to say vec { X } follows a
multivariate normal distribution with mean vec { M } and covariance ΣX = A ⊗ B.
Here, vec { X } is formed by stacking the columns of X into a vector in Rmn . Intuitively, A describes the covariance between columns of X, while B describes the
covariance between rows of X. See [15, 22] for more characterization and examples.
In this paper, we introduce the related sum of Kronecker product models to encode the
covariance structure of a matrix variate distribution. The proposed models and methods
incorporate ideas from recent advances in graphical models, high-dimensional regression model with observation errors, and matrix decomposition. Let Am×m , Bn×n be
symmetric positive definite covariance matrices. Denote the Kronecker sum of A =
(aij ) and B = (bij ) by
Σ
=
=
A ⊕ B := A ⊗ In + Im ⊗ B
a11 In + B
a12 In
...
a1m In
a
I
a
I
+
B
.
.
.
a
21
n
22
n
2m In
...
am1 In
am2 In
. . . amm In + B
(mn)×(mn)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
4
where In is an n × n identity matrix. This covariance model arises naturally from the
context of errors-in-variables regression model defined as follows.
Suppose that we observe y ∈ Rn and X ∈ Rn×m in the following model:
y
= X0 β ∗ +
(1.1a)
X
= X0 + W
(1.1b)
where X0 is an n × m design matrix with independent row vectors, ∈ Rn is a noise
vector and W is a mean zero n × m random noise matrix, independent of X0 and ,
with independent column vectors ω 1 , . . . , ω m .
In particular, we are interested in the additive model of X = X0 + W such that
vec { X } ∼ N (0, Σ) where Σ = A ⊕ B := A ⊗ In + Im ⊗ B
(1.2)
where we use one covariance component A ⊗ In to describe the covariance of matrix
X0 ∈ Rn×m , which is considered as the signal matrix, and the other component Im ⊗B
to describe that of the noise matrix W ∈ Rn×m , where Eω j ⊗ ω j = B for all j,
where ω j denotes the j th column vector of W . Our focus is on deriving the statistical
properties of two estimators for estimating β ∗ in (1.1a) and (1.1b) despite the presence
of the additive error W in the observation matrix X. We will show that our theory and
analysis works with a model much more general than that in (1.2), which we will define
in Section 1.1.
Before we go on to define our estimators, we now use an example to motiviate (1.2) and
its subgaussian generalization in (1.4). Suppose that there are n patients in a particular
study, for which we use X0 to model the “systolic blood pressure” and W to model the
seasonal effects. In this case, X models the fact that among the n patients we measure,
each patient has its own row vector of observed set of blood pressures across time,
and each column vector in W models the seasonal variation on top of the true signal
at a particular day/time. Thus we consider X as measurement of X0 with W being
the observation error. That is, we model the seasonal effects on blood pressures across
a set of patients in a particular study with a vector of dependent entries. Thus W is a
matrix which consists of repeated independent sampling of spatially dependent vectors,
if we regard the individuals as having spatial coordinates, for example, through their
geographic locations. We will come back to discuss this example in Section 1.4.
1.1. The model and the method
We first need to define an independent isotropic vector with subgaussian q
marginals as
P 2
in Definition 1.1. For a vector y = (y1 , . . . , yp ) in Rp , denote by kyk2 =
j yj the
length of y.
Definition 1.1. Let Y be a random vector in Rp
2
2
1. Y is called isotropic if for every y ∈ Rp , E | h Y, y i | = kyk2 .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
5
Rudelson and Zhou/Errors-in-variables
2. Y is ψ2 with a constant α if for every y ∈ Rp ,
k h Y, y i kψ2 := inf{t : E exp( h Y, y i 2 /t2 ) ≤ 2} ≤ α kyk2 .
(1.3)
The ψ2 condition on a scalar random variable V is equivalent to the subgaussian tail
decay of V , which means P (|V | > t) ≤ 2 exp(−t2 /c2 ), for all t > 0.
Throughout this paper, we use ψ2 vector, a vector with subgaussian marginals and
subgaussian vector interchangeably.
The model. Let Z be an n × m random matrix with independent entries Zij satisfying
2
EZij = 0, 1 = EZij
≤ kZij kψ2 ≤ K. Let Z1 , Z2 be independent copies of Z. Let
X = X0 + W
(1.4)
such that X0 = Z1 A1/2 is the design matrix with independent subgaussian row vectors,
and W = B 1/2 Z2 is a random noise matrix with independent subgaussian column
vectors.
Assumption (A1) allows the covariance model in (1.2) and its subgaussian variant
in (1.4) to be identifiable.
(A1) We assume tr(A) = m is a known parameter, where tr(A) denotes the trace of
matrix A.
In the Kronecker sum model, we could assume we know tr(B), in order not to assume
knowing tr(A). Assuming one or the other is known is unavoidable as the covariance
model is not identifiable otherwise. Moreover, by knowing tr(A), we can construct an
estimator for tr(B):
2
1
b
b
kXkF − ntr(A) + and define τbB := n1 tr(B)
≥ 0 (1.5)
tr(B)
= m
P P 2
2
where (a)+ = a ∨ 0 and kXkF := i j Xij
. We first introduce the corrected Lasso
estimator, adapted from those as considered in [30].
b
Suppose that tr(B)
is an estimator for tr(B); for example, as constructed in (1.5). Let
b
Γ
=
1 T
1b
X X − tr(B)I
m
n
n
and γ
b =
1
T
n X y.
(1.6)
For a chosen penalization parameter λ ≥ 0, and parameters b0 and d, we consider the
following regularized estimation with the `1 -norm penalty,
βb =
arg min√
β:kβk1 ≤b0
1 Tb
β Γβ − h γ
b, β i + λkβk1 ,
2
d
(1.7)
which is a variation of the Lasso [48] or the Basis Pursuit [12] estimator. Although in
our analysis, we set b0 ≥ kβ ∗ k2 and d = |supp(β ∗ )| := {j : βj∗ 6= 0} for simplicity,
in practice, both b0 and d are understood to be parameters chosen to provide an upper
bound on the `2 norm and the sparsity of the true β ∗ .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
6
Rudelson and Zhou/Errors-in-variables
For a vector β ∈ Rm , denote by kβk∞ := maxj |βj |. Recently, [3] discussed the
following conic programming compensated matrix uncertainly (MU) selector , which
is a variant of the Dantzig selector [6, 35, 36]. Adapted to our setting, it is defined as
follows. Let λ, µ, τ > 0,
βb = arg min kβk1 + λt : (β, t) ∈ Υ where
(1.8)
o
n
m
b
≤ µt + ω, kβk2 ≤ t
b − Γβ
Υ =
(β, t) : β ∈ R , γ
∞
q
b are as defined in (1.6) with µ ∼ log m , ω ∼
where γ
b and Γ
n
estimator as the Conic programming estimator from now on.
q
log m
n .
We refer to this
1.2. Gradient descent algorithms
In order to obtain fast, approximate solutions to the optimization goal as in (1.10), we
adopt the computational framework of [1, 30], namely, the composite gradient descent
method due to Nesterov [34] to analyze our computational and statistical errors in an
integrated manner. First we denote the population and empirical loss functions by
L(β) =
1 T
β Σx β − β ∗T Σx β
2
and
Ln (β) =
1 Tb
β Γβ − γ
bT β
2
(1.9)
respectively. We consider regularizers that are separable across all coordinates and
write
ρλ (β) =
m
X
ρλ (βi ).
i=1
Throughout this paper, we denote by
φ(β) =
1 Tb
β Γβ − γ
bT β + ρλ (β).
2
From the formulation (1.7), the corrected linear regression estimator is given by minimizing the penalized loss function φ(β) subject to the constraint that g(β) ≤ R:
1 Tb
T
b
β Γβ − γ
b β + ρλ (β)
(1.10)
β ∈ arg min
β∈Rm ,g(β)≤R 2
where g(β) is a convex function, which is allowed to be identical to kβk1 and R is a
second tuning parameter that is chosen to confine the solution βb within the `1 ball of
radius R, while at the same time ensuring that β ∗ is a feasible solution. The gradient
descent method generates a sequence {β t }∞
t=0 of iterates by first initializing to some
parameter β 0 ∈ Rm , and then for t = 0, 1, 2, . . ., applying the recursive updates:
ζ
2
β t+1 = arg min
Ln (β t ) + h ∇Ln (β t ), β − β t i +
β − β t 2 + ρλ (β) (1.11)
2
β∈Rm ,g(β)≤R
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
7
where ζ is the step size parameter.
More generally, we consider loss function Ln : Rm → R and ρλ which are possibly
nonconvex and consider the regularized M-estimator of the form
βb ∈
arg min
{Ln (β; X) + ρλ (β)}
(1.12)
β∈Rm ,g(β)≤R
where ρλ : Rm → R is a regularizer depending on a tuning parameter λ > 0. Because
of this potential nonconvexity, we also include a side constraint in the form of g(β) ≤
R, where
o
µ
1n
2
(1.13)
ρλ (β) + kβk2
g(β) :=
λ
2
so that this choice of g is convex for properly chosen parameter µ ≥ 0 for a class of
weakly convex penalty functions ρ [51]; See Assumption 1 in [31] where properties of
g and ρλ are stated in terms of the univariate function ρλ : R → R and the parameter
µ ≥ 0. While our results hold for the general nonconvex penalty ρλ that is weakly convex in the sense that (1.13) holds for some parameter µ > 0, we focus our discussion
to the choice of ρλ (β) = λ kβk1 and µ = 0 in the present paper.
1.3. Our contributions
We provide a unified analysis of the rates of convergence for both the corrected Lasso
estimator (1.7) and the Conic programming estimator (1.8), which is a Dantzig selectortype, although under slightly different conditions. We will show the rates of convergence in the `q norm for q = 1, 2 for estimating a sparse vector β ∗ ∈ Rm in the
model (1.1a) and (1.1b) using the corrected Lasso estimator (1.7) in Theorems 3 and 6,
and the Conic programming estimator (1.8) in Theorems 4 and 7 for 1 ≤ q ≤ 2. We
also show bounds on the predictive errors for the Conic programming estimator. The
bounds we derive in Theorems 3 and 4 focus on cases where the errors in W are not too
small in their magnitudes in the sense that τB := tr(B)/n is bounded from below. For
the extreme case when τB approaches 0, one hopes to recover bounds close to those
for the regular Lasso or the Dantzig selector since the effect of the noise in matrix W
on the procedure becomes negligible. We show in Theorems 6 and 7 that this is indeed
the case. These results are new to the best of our knowledge.
Let Z1 , Z2 be independent subgaussian random matrices with independent entries (cf. (1.4)).
In Theorems 3 to 7, we consider the regression model in (1.1a) and (1.1b) with subgaussian random design, where X0 = Z1 A1/2 is a subgaussian random matrix with
independent row vectors, and W = B 1/2 Z2 is an n × m random noise matrix with
independent column vectors, This model is significantly different from those analyzed
in the literature. For example, unlike the present work, the authors in [30] apply Theorem 16 which states a general result on statistical convergence properties of the estimator (1.7) to cases where W is composed of independent subgaussian row vectors,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
8
when the row vectors of X0 are either independent or follow a Gaussian vector autoregressive model. See also [35, 36, 3] for the corresponding results on the compensated
MU selectors, variations on the Conic programming estimator (1.8).
The second key difference between our framework and the existing work is that we assume that only one observation matrix X with the single measurement error matrix W
is available. Assuming (A1) allows us to estimate EW T W as required in the estimation procedure (1.6) directly, given the knowledge that W is composed of independent
column vectors. In contrast, existing work needs to assume that the covariance matrix
ΣW := n1 EW T W of the independent row vectors of W or its functionals are either
known a priori, or can be estimated from a dataset independent of X, or from replicated
X measuring the same X0 ; see for example [35, 36, 3, 30, 10]. Although the model we
consider is different from those in the literature, the identifiability issue, which arises
from the fact that we observe the data under an additive error model, is common. Such
repeated measurements are not always available or costly to obtain in practice [10]. We
will explore such tradeoffs in future work.
A noticeable exception is the work of [11], which deals with the scenario when the
noise covariance is not assumed to be known. We now elaborate on their result, which
is a variant of the orthogonal matching pursuit (OMP) algorithm [49, 50]. Their support recovery result, that is, recovering the support set of β ∗ , applies only to the case
when both signal matrix and the measurement error matrix have isotropic subgaussian
row vectors. In other words, they assume independence among both rows and columns
in X (X0 and W ). Moreover, their algorithm requires the knowledge of the sparsity
∗
parameter d, which is the number
in β , as well as a βmin conqof non-zero entries
log m
∗
dition: minj∈supp β ∗ βj∗ = Ω
n (kβ k2 + 1) . Under these conditions, they
recover essentially the same `2 -error bounds as in the current work, and [30], where
the covariance ΣW is assumed to be known.
Finally, we present in Theorems 2 and 9 the optimization error for the gradient descent
algorithms in solving (1.12) and more specifically (1.7). Let βb be a global optimizer
of (1.12). Let λmax (A) and λmin (A) be the largest and smallest eigenvalues, and κ(A)
be the condition number for matrix A. Let 0 < κ < 1 be a contraction factor to be defined in (2.11). Similar to the work of [1, 30], we show that the geometric convergence
is not guaranteed to an arbitrary precision, but only to an accuracy related to statistical
precision of the problem, measured by the `2 error: kβb − β ∗ k22 =: ε2stat between the
global optimizer βb and the true parameter β ∗ .
More precisely, our analysis guarantees geometric convergence of the sequence {β t }∞
t=0
to a parameter β ∗ up to a neighborhood of radius defined through the statistical error
bound ε2stat
δ2
ε2stat d log m
,
1−κ n
where κ is a contraction coefficient to be defined (2.11), so that for all t ≥ T ∗ (δ) as
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
9
in (2.17), α` λmin (A) and αu λmax (A),
4δ 2
α` ε2stat
4δ 4
+
+ 2
= O(ε2stat )
α`
4
b0 α` λmax (A)
2
q
n
e
e
for λ, ζ ≥ αu appropriately chosen, R = O(
log m ) and n = Ω (d log m), where the
β t − βb
2
≤
e and Ω(·)
e symbols hide spectral parameters regarding A and B. To quantify such
O(·)
results, we first need to introduce some conditions in Section 2. See Theorem 2 and
Corollary 10 for the precise conditions and statements.
1.4. Discussion
The theory on matrix variate normal data show that having replicates will allow one to
estimate more complicated graphical structures and achieve faster rates of convergence
under less restrictive assumptions [56]. Our consistency results in the present work
deal with only a single random matrix following the model (1.4), assuming that tr(A)
is known. With replicates, this assumption can be lifted off immediately. Assume there
exists a replicate
e = X0 + W
f,
X
(1.14)
e −X = W
f − W to estimate B using existing methods. The rathen we can use X
tionale for considering such an option is one may have a repeated measurement of X0
f follow the same error distribution. Such external data
for which the errors W and W
or knowledge of the noise distribution is needed in order to do inference under such
additive measurement error model [10].
The second key modeling question is: would each row vector in W for a particular
patient across all time points be a correlated normal or subgaussian vector as well? It is
our conjecture that combining the newly developed techniques, namely, the concentration of measure inequalities we have derived in the current framework with techniques
from existing work [56], we can handle the case when W follows a matrix normal distribution with a separable covariance matrix ΣW = C ⊗ B, where C is an m × m
positive semi-definite covariance matrix. Moreover, for this type of ”seasonal effects”
as the measurement errors, the time varying covariance model would make more sense
to model W , which we elaborate in the second example.
In neuroscience applications, population encoding refers to the information contained
in the combined activity of multiple neurons [27]. The relationship between population
encoding and correlations is complicated and is an area of active investigation, see for
example [40, 13]. It becomes more often that repeated measurements (trials) simultaneously recorded across a set of neurons and over an ensemble of stimuli are available.
In this context, one can use a random matrix X0 ∼ Nn,m (µ, A ⊗ B) which follows a
matrix-variate normal distribution, or its subgaussian correspondent, to model the ensemble of mean response variables, e.g., the membrane potential, corresponding to the
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
10
cross-trial average over a set of experiments. Here we use A to model the task correlations and B to model the baseline correlation structure among all pairs of neurons at
the signal level. It has been observed that the onset of stimulus and task events not only
change the cross-trial mean response in µ, but also alter the structure and correlation of
the noise for a set of neurons, which correspond to the trial-to-trial fluctuations of the
neuron responses. We use W to model such task-specific trial-to-trial fluctuations of a
set of neurons recorded over the time-course of a variety of tasks. Models as in (1.1a)
and (1.1b) are useful in predicting the response of set of neurons based on the current and past mean responses of all neurons. Moreover, we could incorporate non-i.i.d.
non-Gaussian W = [w1 , . . . , wm ] with wt = B 1/2 (t)z(t), where z(1), . . . , z(m) are
independent isotropic subgaussian random vectors and B(t) 0 for all t, to model the
time-varying correlated noise as observed in the trial-to-trial fluctuations. It is possible to combine the techniques developed in the present paper with those in [57, 56] to
develop estimators for A, B and the time varying B(t), which is itself an interesting
topic, however, beyond the scope of the current work.
In summary, oblivion in ΣW and a general dependency condition in the data matrix
X are not simultaneously allowed in existing work. In contrast, while we assume that
X0 is composed of independent subgaussian row vectors, we allow rows of W to be
dependent, which brings dependency to the row vectors of the observation matrix X.
In the current paper, we focus on the proof-of-the-concept on using the Kronecker
sum covariance and additive model to model two way dependency in data matrix X,
and derive bounds in statistical and computational convergence for (1.7) and (1.8). In
some sense, we are considering a parsimonious model for fitting observation data with
two-way dependencies: we use the signal matrix to encode column-wise dependency
among covariates in X, and error matrix W to explain its row-wise dependency. When
replicates of X or W are available, we are able to study more sophisticated models and
inference problems, some of which are described earlier in this section.
We leave the investigation of this more general modeling framework and relevant
statistical questions to future work. We refer to [10] for an excellent survey of the
classical as well as modern developments in measurement error models. In future
work, we will also extend the estimation methods to the settings where the covariates are measured with multiplicative errors which are shown to be reducible to the
additive error problem as studied in the present work [36, 30]. Moreover, we are interested in applying the analysis and concentration of measure results developed in
the current paper and in our ongoing work to the more general contexts and settings where measurement error models are introduced and investigated; see for example [16, 8, 44, 24, 20, 45, 9, 7, 14, 46, 25, 28, 47, 53, 23, 29, 32, 2, 43, 41, 42] and
references therein.
Notation. Let e1 , . . . , ep be the canonical basis of Rp . For a set J ⊂ {1, . . . , p}, denote
EJ = span{ej : j ∈ J}. For a matrix A, we use kAk2 to denote its operator norm. For
a set V ⊂ Rp , we let conv V denote the convex hull of V . For a finite set Y , the cardinality is denoted by |Y |. Let B1p , B2p and S p−1 be the unit `1 ball, the unit Euclidean
ball and the unit sphere respectively. For a matrix A = (aij )1≤i,j≤m
Pm , let kAkmax =
maxi,j |aij | denote the entry-wise max norm. Let kAk1 = maxj i=1 |aij | denote the
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
11
Rudelson and Zhou/Errors-in-variables
P P
2
matrix `1 norm. The Frobenius norm is given by kAkF = i j a2ij . Let |A| denote
2
the determinant and tr(A) be the trace of A. The operator or `2 norm kAk2 is given
T
by λmax (AA ). For a matrix A, denote by r(A) the effective rank tr(A)/ kAk2 . Let
2
2
kAkF /kAk2 denote the stable rank for matrix A. We write diag(A) for a diagonal matrix with the same diagonal as A. For a symmetric matrix A, let Υ(A) = (υij ) where
υij = I(aij 6= 0), where I(·) is the indicator function. Let I be the identity matrix.
For two numbers a, b, a ∧ b := min(a, b) and a ∨ b := max(a, b). For a function
g : Rm → R, we write ∇g to denote a gradient or subgradient, if it exists. We write
a b if ca ≤ b ≤ Ca for some positive absolute constants c, C which are independent
of n, m or sparsity parameters. Let (a)+ := a ∨ 0. We write a = O(b) if a ≤ Cb for
some positive absolute constants C which are independent of n, m or sparsity parameters. The absolute constants C, C1 , c, c1 , . . . may change line by line.
2. Assumptions and preliminary results
We will now define some parameters related to the restricted and sparse eigenvalue
conditions that are needed to state our main results. We also state a preliminary result
in Lemma 1 regarding the relationships between the two conditions in Definitions 2.1
and 2.2.
Definition 2.1. (Restricted eigenvalue condition RE(s0 , k0 , A)). Let 1 ≤ s0 ≤ p,
and let k0 be a positive number. We say that a q × p matrix A satisfies RE(s0 , k0 , A)
condition with parameter K(s0 , k0 , A) if for any υ 6= 0,
kAυk2
1
min
:= min
> 0.
J⊆{1,...,p}, kυJ c k ≤k0 kυJ k
K(s0 , k0 , A)
kυJ k2
1
1
(2.1)
|J|≤s0
where υJ represents the subvector of υ ∈ Rp confined to a subset J of {1, . . . , p}.
It is clear that when s0 and k0 become smaller, this condition is easier to satisfy. We
also consider the following variation of the baseline RE condition.
Definition 2.2. (Lower-RE condition) [30] The matrix Γ satisfies a Lower-RE condition with curvature α > 0 and tolerance τ > 0 if
2
2
θT Γθ ≥ α kθk2 − τ kθk1 ∀θ ∈ Rm .
P
where kθk1 := j |θj |. As α becomes smaller, or as τ becomes larger, the Lower-RE
condition is easier to be satisfied.
Lemma 1. Suppose that the Lower-RE condition holds for Γ := AT A with α, τ > 0
such that τ (1 + k0 )2 s0 ≤ α/2. Then the RE(s0 , k0 , A) condition holds for A with
r
α
1
≥
> 0.
K(s0 , k0 , A)
2
Assume that RE((k0 + 1)2 , k0 , A) holds. Then the Lower-RE condition holds for Γ =
AT A with
1
α=
>0
(k0 + 1)K 2 (s0 , k0 , A)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
12
Rudelson and Zhou/Errors-in-variables
where s0 = (k0 + 1)2 , and τ > 0 which satisfies
λmin (Γ) ≥ α − τ s0 /4.
The condition above holds for any τ ≥
4
(k0 +1)3 K 2 (s0 ,k0 ,A)
(2.2)
−
4λmin (Γ)
(k0 +1)2 .
The first part of Lemma 1 means that, if k0 is fixed, then smaller values of τ guarantee
RE(s0 , k0 , A) holds with larger s0 , that is, a stronger RE condition. The second part
of the Lemma implies that a weak RE condition implies that the Lower-RE (LRE)
holds with a large τ . On the other hand, if one assumes RE((k0 + 1)2 , k0 , A) holds
with a large value of k0 (in other words, a strong RE condition), this would imply
LRE with a small τ . In short, the two conditions are similar but require tweaking the
parameters. Weaker RE condition implies LRE condition holds with a larger τ , and
Lower-RE condition with a smaller τ , that is, stronger LRE implies stronger RE. We
prove Lemma 1 in Section 9.
Definition 2.3. (Upper-RE condition) [30] The matrix Γ satisfies an upper-RE condition with smoothness α
e > 0 and tolerance τ > 0 if
2
2
θT Γθ ≤ α
e kθk2 + τ kθk1 ∀θ ∈ Rm .
Definition 2.4. Define the largest and smallest d-sparse eigenvalue of a p × q matrix
A to be
ρmax (d, A)
:=
and ρmin (d, A)
:=
2
(2.3)
2
(2.4)
max
kAtk22 / ktk2 , where d < p,
min
kAtk22 / ktk2 .
t6=0;d−sparse
t6=0;d−sparse
Before stating some general result for the optimization program (1.12) and its implications for the Lasso-type estimator (1.7) in terms of statistical and optimization errors, we need to introduce some more notation and the following assumptions. Let
amax = maxi aii and bmax = maxi bii be the maximum diagonal entries of A and B
respectively. In general, under (A1), one can think of λmin (A) ≤ 1 and for s ≥ 1,
1 ≤ amax ≤ ρmax (s, A) ≤ λmax (A),
(2.5)
where λmax (A) denotes the maximum eigenvalue of A.
(A2) The minimal eigenvalue λmin (A) of the covariance matrix A is bounded: 1 ≥
λmin (A) > 0.
(A3) Moreover,
we
assume that the condition number κ(A) is upper bounded by
q
n
O
log m and τB = O(λmax (A)).
Throughout the rest of the paper, s0 ≥ 32 is understood to be the largest integer chosen
such that the following inequality still holds:
r
√
λmin (A)
n
where $(s0 ) := ρmax (s0 , A) + τB
(2.6)
s0 $(s0 ) ≤
32C
log m
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
13
where we denote by τB = tr(B)/n and C is to be defined. Denote by
MA =
64C$(s0 )
≥ 64C.
λmin (A)
(2.7)
b as
Throughout this paper, we denote by A0 the event that the modified gram matrix Γ
defined in (1.6) satisfies the Lower as well as Upper RE conditions with
curvature
and tolerance
11
5
λmin (A), smoothness α
e=
λmax (A)
8
8
384C 2 $(s0 )2 log m
λmin (A) − α
396C 2 $2 (s0 + 1) log m
≤ τ :=
≤
λmin (A)
n
s0
λmin (A)
n
α=
for α, α
e and τ as defined in Definitions 2.2 and 2.3, and C, s0 , $(s0 ) in (2.6).
To bound the optimization errors, we show that the corrected linear regression loss
function (1.9) satisfies the following Restricted Strong Convexity (RSC) and Restricted
Smoothness (RSM) conditions when the sample size and effective rank of matrix B
satisfy certain lower bounds (cf. Theorem 3); namely, for all vectors β0 , β1 ∈ Rm and
T (β1 , β0 )
:= Ln (β1 ) − Ln (β0 ) − h ∇Ln (β0 ), β1 − β0 i ,
we show that for some parameters (α` , τ` (Ln )) and (αu , τu (Ln )),
α`
2
2
kβ1 − β0 k2 − τ` (Ln ) kβ1 − β0 k1 and
2
αu
2
2
kβ1 − β0 k2 + τu (Ln ) kβ1 − β0 k1 .
2
T (β1 , β0 ) ≥
T (β1 , β0 ) ≤
(2.8)
(2.9)
Applied to (1.12), the composite gradient descent procedure of [34] produces a sequence of iterates {β t }∞
t=0 via the updates
)
(
2
∇Ln (β t )
ρλ (β)
1
t
t+1
β− β −
+
(2.10)
β
= arg min
2
ζ
ζ
β∈Rm ,g(β)≤R
2
where ζ1 is the step size. Let ν` = 64dτ` (Ln ) and ᾱ` := α` − ν` . We show that the
composite gradient updates exhibit a type of globally geometric convergence in terms
of the compound contraction coefficient
κ =
1−
ᾱ`
4ζ
+%
1−%
, where % :=
2ν(d, m, n)
128dτu (Ln )
:=
α` − ν`
ᾱ`
(2.11)
where ν` < α` /C for some C > 1 to be specified. Let τ (Ln ) = τ` (Ln ) ∨ τu (Ln ).
Define
2τ (Ln ) ᾱ`
ξ :=
+ 2% + 5 > 10τ (Ln ).
(2.12)
1−%
4ζ
For simplicity, we present in Theorem 2 the case for ρλ (β) = λ kβk1 only.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
14
Rudelson and Zhou/Errors-in-variables
Theorem 2. Consider the optimization program (1.12) for a radius R such that β ∗ is
feasible. Let g(β) = λ1 ρλ (β) where ρλ (β) = λ kβk1 . Suppose that the loss function
Ln satisfies the RSC/RSM conditions (2.8) and (2.9) with parameters (α` , τ` (Ln )) and
(αu , τu (Ln )) respectively. Let %, κ and ξ be defined as in (2.11) and (2.12) respectively.
Suppose that the regularization parameter is chosen such that for ζ ≥ αu
16Rξ
.
(2.13)
λ ≥ max 12 k∇Ln (β ∗ )kmax ,
(1 − κ)
Suppose that κ < 1. Suppose that βb is a global minimizer of (1.12). Then for any step
size parameter ζ ≥ αu and tolerance parameter
δ2
≥
cε2stat d log m
=: δ̄ 2 ,
1−κ n
where ε2stat = βb − β ∗
2
,
2
(2.14)
the following hold for all t ≥ T ∗ (δ)
b ≤ δ2 ,
φ(β t ) − φ(β)
2
β t − βb
≤
2
where ν = 64dτ (Ln ), τ (Ln )
0
2
ᾱ`
log m
n ,
16δ 4
∧ 4R2 ,
λ2
δ 2 + 4νε2stat + 4τ (Ln )2 ,
and for 2 =
b
(2.16)
and
β)
2 log( φ(β )−φ(
)
δ2
T (δ) =
+ log log
log(1/κ)
∗
(2.15)
λR
δ2
1+
log 2
log(1/κ)
.
(2.17)
We prove Theorem 2 in Section B. Theorem 2 is similar in spirit to the main result
Theorem 2 in [1] that deals with a convex loss function, and Theorem 3 in [31] on a
similar setting to the present work. Compared to [31], we simplified the condition on λ
by not imposing an upper bound. Moreover, we present refined analysis on the sample
requirement and illuminate its dependence upon the condition number κ(A) and the
tolerance parameter τ when applied to the corrected linear regression problem (1.10).
It is understood throughout the paper that for the same C as in (2.7),
τ τ0
log m
400C 2 $(s0 + 1)2
, where τ0
≈ MA2 λmin (A)/10
n
λmin (A)
(2.18)
and it is helpful to consider MA as being upper bounded by O(κ(A)) in view of (2.5)
and (A3). Toward this end, we prove in Section 5 that under event A0 ∩B0 , the RSC and
RSM conditions as stated in Theorem 2 hold with α` λmin (A) and αu λmax (A)
and τ` (Ln ) = τu (Ln ) τ ; then we have for all t ≥ T ∗ (δ) as defined in (2.17) and for
ε2stat d log m
δ 2 1−κ
,
n
β t − βb
2
≤
2
4 2 α` 2
δ + εstat + O
α`
4
δ 2 ε2stat
b20
,
(2.19)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
15
Rudelson and Zhou/Errors-in-variables
where 0 < κ < 1 so long as ζ λmax (A) and n = Ω(κ(A)MA2 d log m).
We now check the conditions on λ in Theorem 2. First, we note that both types of
conditions on λ are also required in the present paper for the statistical error bounds
shown in Theorems 3 and 6. We state in Theorem 16 a deterministic result from [30]
on the statistical error for the corrected linear model, which requires that
λ ≥ 2 k∇Ln (β ∗ )kmax and
√
log m
λ ≥ 4b0 dτ 4Rτ for τ := τ0
n
(2.20)
as defined in (2.18) and dτ ≤ α32` in order to obtain the statistical error bound for the
corrected linear model at the order of
ε2stat = βb − β ∗
2
2
400 2
λ d.
α`2
(2.21)
Under suitable conditions on the sample size n and the effective rank of matrix B to
be stated in Theorem 3, we show that for the loss function (1.9), the RSC and RSM
conditions hold under event A0 (cf. Lemma 15) following the Lower and Upper-RE
conditions as derived in Lemma 15,
ᾱ` ≈ α`
λmin (A)
,
2
αu
3λmax (A)
,
2
and
τ (Ln ) τ.
Compared with the lower bound imposed on λ as in (2.20) that we use to derive statisξ
that crucially depends on the
tical error bounds, the penalty now involves a term 1−κ
condition number κ(A) in (2.13); Assuming that ζ ≥ αu , then the second condition in
(2.13) on λ implies that
λ
ξ
1−κ
=
Ω(Rτ (Ln )κ(A)) given
ζ
≥ 40τ (Ln ) + 2τ (Ln ) τ κ(A),
ᾱ`
(2.22)
which now
√ depends explicitly on the condition number κ(A) in addition to the radius
R b0 d and the tolerance parameter τ . This is expected given that both RSC and
RSM conditions are needed in order to derive the computational convergence bounds,
while for the statistical error, we only require the RSC (Lower RE) condition to hold.
Remarks. Consider the regression model in (1.1a) and (1.1b) with independent random
matrices X0 , W as in (1.4), and an error vector ∈ Rn independent of X0 , W , with
independent entries j satisfying Ej = 0 and kj kψ2 ≤ M . Theorem 12 and its
b ∗−
corollaries provide an upper bound on the `∞ norm of the gradient ∇Ln (β ∗ ) = Γβ
b
γ
b of the loss function in the corrected linear model, where Γ and γ
b are as defined
in (1.6). Let
1/2
D00 = kBk2
+ a1/2
max , and
1/2
Doracle = 2(kAk2
1/2
+ kBk2 ).
(2.23)
Specializing to the case of corrected linear models, we have by Corollary 14, on event
B0 as defined therein,
r
log m
∗
∗
b −γ
k∇Ln (β )k∞ = Γβ
b
≤ψ
n
∞
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
16
+/2
+/2
1/2
oracle
where ψ := C0 D00 K M + τB K kβ ∗ k2 and τB = τB + D√
for D00 , Doracle
m
as defined in (2.23).
b for solving (1.7) using the
The bound (2.15) characterizes the excess loss φ(β t ) − φ(β)
composite gradient algorithm; moreover, for any iterate β t such that (2.15) holds, the
following bound on the optimization error β t − βb follows immediately:
2
2
64τ` (Ln )δ 4
≤
β t − βb
,
δ 2 + 4νε2stat +
ᾱ`
λ2
2
4
where ν = 64dτ (Ln ) and 4τ (Ln )2 = 64τ (Ln ) λδ 2 by definition of 2 in view of
(2.21). Finally, we note that Theorem 2 holds for a class of weakly convex penalties
as considered in [31] with suitable adaptation of RSC and parameters and conditions
to involve µ, following exactly the same sequence of arguments. Notable examples of
such weakly convex penalty functions are SCAD [19] and MCP [55].
The rest of the paper is organized as follows. In Section 3, we present two main results in Theorems 3 and 4. In Section 4, we state more precise results which improve
upon Theorems 3 and 4; these results are more precise in the sense that our bounds and
penalty parameters now take tr(B), the parameter that measures the magnitudes of errors in W , into consideration. In Section 5, we show that the RSC and RSM conditions
hold for the corrected linear loss function and present our computational convergence
bounds with regard to (1.7) in Theorem 9 and Corollary 10. In Section 6, we outline the
proof of the main theorems. In particular, we outline the proofs for Theorems 3, 4, 6
and 7 in Section 6, 6.3 and 6.5 respectively. In Section 7, we show a deterministic result
b − A for Γ
b as in (1.6) with regards to
as well as its application to the random matrix Γ
the upper and Lower RE conditions. In Section 8, we present results from numerical
simulations designed to validate the theoretical predictions in previous sections. The
technical details of proofs are collected at the end of the paper. We prove Theorem 3
in Section 10. We prove Theorem 4 in Section 11. We prove Theorems 6 and 7 in Section 12 and Section 13 respectively. We defer the proof of Theorem 2 to Section B. The
paper concludes with a discussion of the results in Section 16. We list a set of symbols
we use throughout the paper in Table 1. Additional proofs and theoretical results are
collected in the Appendix.
3. Main results on the statistical error
In this section, we will state our main results in Theorems 3 and 4 where we consider
the regression model in (1.1a) and (1.1b) with random matrices X0 , W ∈ Rn×m as
defined in (1.4). For the corrected Lasso estimator, we are interested in the case where
the smallest eigenvalue of the column-wise covariance matrix A does not approach 0
too quickly and the effective rank of the row-wise covariance matrix B is bounded
from below (cf. (3.2)). More precisely, (A2) thus ensures that the Lower-RE condition
as in Definition 2.2 is not vacuous. (A3) ensures that (2.6) holds for some s0 ≥ 1.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
17
Rudelson and Zhou/Errors-in-variables
TABLE 1
symbols we used throughout the proof
Symbol
α
α`
αu
¯stat
Definition
curvature: α := 85 λmin (A)
Lower RE/ RSC curvature parameter: α` = α
Upper RE/ RSM parameter αu 23 λmax (A)
√
¯stat = 8 dεstat where εstat = βb − β ∗
τ0
τ0
λ
(A)−α
400C 2 $(s0 +1)2
λmin (A)
2
τ = mins
0
τB
s0 ≥ 1
tolerance parameter τ = τ0 lognm in Lower/Upper RE conditions
τB = tr(B)/n
the largest integer chosenqsuch that the following inequality still holds:
√
λ
(A)
n
s0 $(s0 ) ≤ min
32C
log m
$(s0 )
τ` (Ln )
τu (Ln )
ν`
ν(d, m, n)
ᾱ`
φ(β)
∇Ln (β)
ρmax (s0 , A) + τB
tolerance parameter in RSC condition: τ` (Ln ) τ0 log m/n
tolerance parameters in RSM condition: τu (Ln ) τ0 log m/n
`
ν` = 64dτ` (Ln ) < α
60
ν(d, m, n) = 64dτu (Ln )
effective RSC coefficient ᾱ` = α` − ν`
b −γ
bT + ρλ (β)
loss function: φ(β) = 12 β T Γβ
b −γ
Gradient of the
loss
function
Γβ
b
q
log m
n
q
m
2C0 K 2 log
mn
ρn = C0 K
ρn
rm,n
ζ
%
κ
rm,n =
step size parameter: ζ ≥ αu = 11λmax /8
128dτu (Ln )
2ν(d,m,n)
`
=
< ᾱ
contraction parameter % :=
ᾱ`
ᾱ`
8ζ
ᾱ`
−1
contraction coefficient as κ := (1 − 4ζ + %)(1 − %)
<1
δ2
tolerance parameter in computational errors δ 2 ≥
MA
MA =
cε2
stat d log p
1−κ
n
64C$(s0 )
where $(s0 ) = ρmax (s0 , A) + τB .
λmin (A)
32C$(s0 +1)
where $(s0 + 1) = ρmax (s0 + 1, A)
M+ = λ (A)
min
`
+ 2% + 5 (1 − %)−1
ξ = 2(τ` (Ln ) ∨ τu (Ln )) ᾱ
4ζ
3 /2
V = 3eMA
M+
ξ
V
+ τB .
Throughout this paper, for the corrected Lasso estimator, we will use the expression
τ :=
λmin (A) − α
4n
5
, where α = λmin (A) and s0 2
s0
8
MA log m
where MA is as defined in (2.7). Let
√
D0 = τB + a1/2
max and D2 = 2(kAk2 + kBk2 ).
(3.1)
Theorem 3. (Estimation for the corrected Lasso estimator) Consider the regression
model in (1.1a) and (1.1b) with independent random matrices X0 , W as in (1.4), and
an error vector ∈ Rn independent of X0 , W , with independent entries j satisfying
Ej = 0 and kj kψ2 ≤ M . Set n = Ω(log m). Suppose n ≤ (V/e)m log m, where
V is a constant which depends on λmin (A), ρmax (s0 , A) and tr(B)/n. Suppose m is
sufficiently large.
Suppose (A1), (A2) and (A3) hold. Let C0 , c0 , c2 , c3 > 0 be some absolute constants.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
18
Rudelson and Zhou/Errors-in-variables
2
2
Suppose that kBkF / kBk2 ≥ log m. Suppose that c0 K 4 ≤ 1 and
r(B) :=
tr(B)
kBk2
≥
16c0 K 4
Vm log m
n
log
.
log m
n
(3.2)
Let b0 , φ be numbers which satisfy
M2
≤ φ ≤ 1.
K 2 b20
(3.3)
Assume that the sparsity of β ∗ satisfies for some 0 < φ ≤ 1
c0 φK 4 n
2 log m < n/2,
40M+
32C$(s0 + 1)
M+ =
λmin (A)
d := |supp(β ∗ )| ≤
where
(3.4)
(3.5)
for $(s0 + 1) = ρmax (s0 + 1, A) + τB .
Let βb be an optimal solution to the corrected Lasso estimator as in (1.7) with
r
log m
λ ≥ 4ψ
where ψ := C0 D2 K (K kβ ∗ k2 + M ) .
n
(3.6)
Then for any d-sparse vectors β ∗ ∈ Rm , such that
2
φb20 ≤ kβ ∗ k2 ≤ b20 ,
(3.7)
Vm log m
n
4c2 n
we have with probability at least 1−4 exp − M 2c3log
log
−
−2
exp
−
2
4
n
m
M K
A
A
22/m3 ,
βb − β ∗
≤
2
20 √
λ d and
α
βb − β ∗
≤
1
80
λd.
α
We give an outline of the proof of Theorem 3 in Section 6.2. We prove Theorem 3 in
Section 10. We defer discussions on conditions appearing Theorem 3 in Section 3.2.
For the Conic programming estimator, we impose a restricted eigenvalue condition
as
p
formulated in [4, 38] on A and assume that the sparsity of β ∗ is bounded by o( n/ log m).
These conditions will be relaxed in Section 4 where we allow τB to approach 0.
Theorem 4. Suppose (A1) holds. Set 0 < δ < 1. Suppose that n < m exp(n)
and 1 ≤ d0 < n. Let λ > 0 be the same parameter as in (1.8). Suppose that
2
2
kBkF / kBk2 ≥ log m. Suppose that the sparsity of β ∗ is bounded by
p
(3.8)
d0 := |supp(β ∗ )| ≤ c0 n/ log m
for some constant c0 > 0. Suppose
2000dK 4
60em
where
log
n ≥
δ2
dδ
d =
2d0 + 2d0 amax
16K 2 (2d0 , 3k0 , A1/2 )(3k0 )2 (3k0 + 1)
.
δ2
(3.9)
(3.10)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
19
Consider the regression model in (1.1a) and (1.1b) with X0 , W as in (1.4) and an error
vector ∈ Rn , independent of X0 , W , with independent entries j satisfying Ej = 0
and kj kψ2 ≤ M . Let βb be an optimal solution to the Conic programming estimator
b as defined in (1.6). Recall τB := tr(B)/n. Choose for
as in (1.8) with input (b
γ , Γ)
D0 , D2 as in (3.1) and
r
r
log m
log m
2
and ω D0 KM
.
µ D2 K
n
n
βb − β ∗
q
c0
m2
− 2 exp(−δ 2 n/2000K 4 ),
r
log m
M
2 1/q
≤ CD2 K d0
kβ ∗ k2 +
n
K
Then with probability at least 1 −
(3.11)
for 2 ≥ q ≥ 1. Under the same assumptions, the predictive risk admits the following
bounds with the same probability as above,
2
2
1
log m
M
∗
0 2 4
∗
b
X(β − β ) ≤ C D2 K d0
kβ k2 +
n
n
K
2
where c0 , C0 , C, C 0 > 0 are some absolute constants.
We give an outline of the proof of Theorem 4 in Section 6 while leaving the detailed
proof in Section 11.
3.1. Regarding the MA constant
Denote by
MA =
64C$(s0 )
ρmax (s0 , A) + τB
λmin (A)
λmin (A)
• (A3) ensures that MA and
qM+ are
upper bounded by the condition number of A:
λmax (A)
n
κ(A) := λmin (A) = O
log m given that τB := tr(B)/n = O(λmax (A)).
• So the condition (3.4) in Theorem 3 allows d n/ log m in the optimal setting when the condition number κ(A) is understood to be a constant. As κ(A)
increases, the conservative worst case upper bound on d needs to be adjusted
correspondingly. Moreover, this adjustment is also crucial in order to ensure the
composite gradient algorithm to converge in the sense of Theorem 2. We will
illustrate such dependencies on κ(A) in numerical examples in Section 8.
• The condition τB = O(λmax (A)) puts an upper bound on how large the measurement error in W can be. We do not allow the measurement error to overwhelm the signal entirely. When τB → 0, we recover the ordinary Lasso bound
in [4], which we elaborate in the next two sections.
Throughout this paper, we assume that MA M+ , where recall M+ =
32C$(s0 +1)
λmin (A) .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
20
Rudelson and Zhou/Errors-in-variables
3.2. Discussions
Throughout our analysis, we set the parameter b0 ≥ kβ ∗ k2 and d = |supp(β ∗ )| :=
{j : βj∗ 6= 0} for the corrected Lasso estimator. In practice, both b0 and d are understood to be parameters chosen to provide an upper bound on the `2 norm and the
sparsity of the true β ∗ . The parameter 0 < φ < 1 is a parameter that we use to describe
2
the gap between kβ ∗ k2 and its upper bound b20 . Denote the Signal-to-noise ratio by
2
2
S/N := K 2 kβ ∗ k2 /M2 , where N := M2 and φK 2 b20 ≤ S := K 2 kβ ∗ k2 ≤ K 2 b20 .
The two conditions (3.3) and (3.7) on b0 and φ imply that N ≤ K 2 φb20 ≤ S. Notice that
this could be restrictive if φ is small. We will show in Section 6.2 that condition (3.3)
is not needed in order for the `p , p = 1, 2 errors as stated in the Theorem 3 to hold.
It was indeed introduced so as to further simplify the expression for the condition on
d as shown in (3.4). Therefore we provide slightly more general conditions on d in
(6.9) in Lemma 17, where (3.3) is not required. We introduce the parameter φ so that
the conditions on d depend on φ and b20 rather than the true signal kβ ∗ k2 (cf. Proof of
Lemmas 17 and 18). It will also become clear in the sequel from the proof of Lemma 17
(cf. (H.4)) that we could use kβ ∗ k2 rather than its the lower bound b20 φ in the expression
for d. However, we choose to state the condition on d as in Theorem 3 for clarity of our
exposition. See also Theorem 6 and Lemma 18.
In fact, we prove that Theorem 3 holds with N = M2 and S = φK 2 b20 in arbitrary
orders, so long as conditions (3.2) and
q (3.4) or (6.9) hold. For both cases, we require
√
that λ (kAk2 + kBk2 )K S + N lognm as expressed in (3.6). That is, when either
the noise level M or the signal strength K kβ ∗ k increases, we need to increase λ
2
correspondingly; moreover, when N dominates the signal K 2 kβ ∗ k2 , we have for d
1
n
as in (3.4),
M 2 log m
A
r
βb − β
∗
∗
2
/ kβ k2 = OP
D2 K
2
N
1
S $(s0 + 1)
!
,
which eventually becomes a vacuous bound when N S. This bound appears a bit
crude as it does not entirely discriminate between the noise, measurement error, and the
signal strength. We further elaborate on the relationships among these three elements
in Section 4. We will then present an improved bound in Theorem 6.
1. The choice of λ for the Lasso estimator and parameters µ, ω for the DS-type
estimator satisfy
λ µ kβ ∗ k2 + ω.
This relationship is made clear through Theorem 16 regarding the corrected
Lasso estimator, which follows from Theorem 1 by [30], and Lemmas 19 and 22
for the Conic programming estimator. The penalty parameter λ is chosen to
b ∗
bound γ
b − Γβ
from above, which is in turn bounded in Theorem 12. See
∞
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
21
Corollaries 13 and 14, which are the key results in proving Theorems 3, 4, 6,
and 7.
2. Throughout our analysis of Theorems 3 and 4, our error bounds are stated in a
way assuming the errors in W are sufficiently large in the sense that these bounds
are optimal only when τB is bounded from below by some absolute constant. For
example, when kBk2 is bounded away from 0, the lower bound on the effective
rank r(B) = tr(B)/ kBk2 implies that τB must also be bounded away from 0.
More precisely, by the condition on the effective rank as in (3.2), we have
τB =
tr(B)
n
≥ 16c0 K 4
kBk2
Vm log m
log
where V = 3eMA3 /2.
log m
n
Later, we will state our results with τB = tr(B)/n > 0 being explicitly included in the error bounds as well as the penalization parameters and sparsity
constraints.
3. In view of the main Theorems 3 and 4, at this point, we do not really think one
estimator is preferable to the other. While the `q error bounds we obtain for the
two estimators are at the same order for q = 1, 2, the conditions under which
these error bounds are obtained are somewhat different. In Theorem 4, we only
require that RE(2d0 , 3k0 , A1/2 ) holds for k0 = 1 + λ where λ 1, while in
Theorem 3 we need the minimal eigenvalue of A to be bounded from below,
namely, we need to assume that (A2) holds. As mentioned earlier, (A2) ensures
that the Lower-RE condition as in Definition 2.2 is not vacuous while (A3) ensures that (2.6) holds for some s0 ≥ 1. Th condition (3.2) on the effective rank
of the row-wise covariance matrix B is also needed to establish the Lower and
Upper RE conditions in Lemma 15 for the corrected Lasso estimator. Moreover,
for the sparsity parameter d0 in (3.8), we show in Lemma 34 that (A2) is a sufficient condition for a type of RE(2d0 , 3k0 ) condition to hold on non positive
b as defined in (1.6). See also Theorem 26.
definite Γ
4. In some sense, the assumptions in Theorem 3 appear to be slightly stronger, while
at the same time yielding correspondingly stronger results in the following sense:
The corrected Lasso procedure can recover a sparse model using O(log m) number of measurements per nonzero component despite the measurement error in
X and the p
stochastic noise , while the Conic programming estimator allows
only d
n/ log m to achieve the error rate at the same order as the corrected Lasso estimator. Hence, while Conic programming estimator is conceptually more adaptive by not fixing an upper bound on kβ ∗ k2 a priori, the price we
pay seems to be a more stringent upper bound on the sparsity level.
5. We note that following Theorem 2 as in [3], one can show that without the relatively restrictive sparsity condition (3.8), a bound similar to that in (3.11) holds,
however, with kβ ∗ k2 being replaced by kβ ∗ k1 , so long as the sample size satisfies the condition as in (4.9). However, we show in Theorem 7 in Section 6.5
that this restriction on the sparsity can be relaxed for the Conic programming
estimator (1.8), when we make a different choice for the parameter µ based on a
more refined analysis.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
22
Results similar to Theorems 3 and 4 have been derived in [30, 3], however, under
different assumptions on the distribution of the noise matrix W . When W is a random
matrix with i.i.d. subgaussian noise, our results in Theorems 3 and 4 will essentially
recover the results in [30] and [3]. We compare with their results in Section 4 in case
B = τB I after we present our improved bounds in Theorems 6 and 7. We refer to the
paper of [3] for a concise summary of these and some earlier results.
Finally, one reviewer asked about the dependence of the tuning parameter on properties
√
1/2
1/2
1/2
of A and B, namely parameters D0 = τB + amax , D00 = kBk2 + amax and
D2 = kAk2 + kBk2 . We now state in Lemma 5 a sharp bound on estimating τB using
τbB as in (1.5), which will provide a natural plug-in estimate for parameters such as D0
that involve τB .
Lemma 5. Let m ≥ 2. Let X be defined as in (1.4) and τbB be as defined in (1.5).
Denote by τB = tr(B)/n and τA = tr(A)/m. Suppose that n ∨ (r(A)r(B)) > log m.
Denote by B6 the event such that
r
kBk
log m kAkF
2
√
+ √ F =: D1 rm,m ,
|b
τB − τB | ≤ 2C0 K
mn
m
n
q
kAk
kBk
m
3
where D1 = √mF + √nF and rm,m = 2C0 K 2 log
mn . Then P (B6 ) ≥ 1 − m3 .
√
If we replace log m with log m in the definition of event B6 , then we can drop the
tr(B)
condition on n or r(A)r(B) = tr(A)
kAk kBk to achieve the same bound on event B6 .
2
2
In an earlier version of the present work by the same authors [39], we presented the
b := 1 XX T − tr(A) Im to
rate of convergence for using the corrected gram matrix B
m
m
estimate B and proved isometry properties in the operator norm once the effective rank
b and its
of A is sufficiently large compared to n; one can then use such estimated B
0
operator norm in D2 and D0 . See Theorem 21 and Corollary 22 therein. As mentioned,
we use the estimated τbB (cf. Lemma 5) in D0 . The dependencies on A, kβ ∗ k2 and are
known problems in the Lasso and corrected Lasso literature; see [4, 30]. For example,
the RE condition as stated in Definition 2.1 and its subgaussian concentration properties
as shown [38] clearly depend on unknown parameter amax related to covariance matrix
A. See Theorem 27 in the present paper. We prove Lemma 5 in Section C.1. Lemma 5
provides the powerful technical insight and one of the key ingredients leading to the
tight analysis in Theorems 6 and 7 for the corrected Lasso estimator (1.7) as well as
the Conic programming estimator (1.8) in Section 4, where we also present theory for
which the dependency on kAk2 becomes extremely mild.
4. Improved bounds when the measurement errors are small
Although the conclusions of Theorems 3 and 4 apply to cases when kBk2 → 0, the
error bounds are not as tight as the bounds we are about to derive in this section. So far,
we have used more crude approximations on the error bounds in terms of estimating
b ∗
γ
b − Γβ
for the sake of reducing the amount of unknown parameters we need to
∞
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
23
Rudelson and Zhou/Errors-in-variables
consider. The bounds we derive in this section take the magnitudes of the measurement
errors in W into consideration. As such, we allow the error bounds to depend on the
parameter τB explicitly, which become much tighter as τB becomes smaller. For the
extreme case when τB approaches 0, one hopes to recover a bound close to the regular
Lasso or the Dantzig selector as the effect of the noise on the procedure should become
negligible. We show in Theorems 6 and 7 that this is indeed the case. Denote by
+/2
τB
:=
√
Doracle
1/2
1/2
τB + √
, where Doracle = 2(kAk2 + kBk2 ). (4.1)
m
We first state a more refined result for the Lasso-type estimator, for which we now only
require that
r
p
log m
1/2
1/2
λ (amax + kBk2 )K N + τB S
.
n
√
√
That is, we replace N + S in λ (3.6) now with N + τB S, which leads to significant
improvement on the rates of convergence for estimating β ∗ when τB → 0.
Theorem 6. Suppose all conditions in Theorem 3 hold, except that we drop (3.3) and
replace (3.6) with
r
log m
+/2
λ ≥ 4ψ
, where ψ := C0 D00 K M + τB K kβ ∗ k2
(4.2)
n
+/2
for D00 and τB as defined in (2.23) and (4.1) respectively. Let c0 , φ, b0 , M , K and
+/2
M+ be as defined in Theorem 3. Let τB+ = (τB )2 .
Suppose that for 0 < φ ≤ 1 and CA :=
1
2 ,
160M+
n
{c0 c00 Dφ ∧ 8} =: d¯0 , where
log m
kBk2 + amax
K 2 M2
and Dφ =
+ τB+ K 4 φ
c00 =
2
$(s0 + 1)
b20
d := |supp(β ∗ )| ≤ CA
(4.3)
(4.4)
2
Then for any d-sparse vectors β ∗ ∈ Rm , such that φb20 ≤ kβ ∗ k2 ≤ b20 , we have
80
20 √
λ d and βb − β ∗ ≤
λd
(4.5)
α
α
1
2
Vm log m
n
4c2 n
with probability at least 1 − 4 exp − M 2c3log
log
−
2
exp
−
−
2
4
n
M K
m
βb − β ∗
≤
A
A
22/m3 .
We give an outline for the proof of Theorem 6 in Section 6.3, and show the actual proof
in Section 12.
We next state in Theorem 7 an improved bounds for the Conic programming estimator (1.8), which dramatically improve upon those in Theorem 4 when τB is small,
where an “oracle” rate for estimating β ∗ with the Conic programming estimator βb (1.8)
2
is defined and the predictive error kXvk2 when τB = o(1) is derived.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
24
Let C0 satisfy (H.6) for c as defined in Theorem 31. Throughout the rest of the paper,
we denote by:
r
r
log m
log m
2
ρn = C0 K
and rm,m = 2C0 K
;
(4.6)
n
mn
3
†/2
1/2
1/2
τB
= (τB + C6 rm,m
) and τB‡ 2τB + 3C62 rm,m .
(4.7)
2
√
1/2
Theorem 7. Let D0 = τB + amax , and D00 , Doracle be as defined in (2.23). Let C6 ≥
Doracle . Let ρn and rm,m be as defined in (4.6). Suppose all conditions in Theorem 4
hold, except that we replace the condition on d as in (3.8) with the following.
Suppose that the sample size n and the size of the support of β ∗ satisfy the following
requirements:
r
1
n
−
, where τB− ≤ 1/2
,
(4.8)
d0 = O τB
1/2
log m
τB + 2C6 rm,m
2000dK 4
60em
and n ≥
log
, where
(4.9)
δ2
dδ
d =
2d0 + 2d0 amax
16K 2 (2d0 , 3k0 , A1/2 )(3k0 )2 (3k0 + 1)
. (4.10)
δ2
Let τbB be as defined in defined in (1.5). Let βb be an optimal solution to the Conic
b as defined in (1.6). Suppose
programming estimator as in (1.8) with input (b
γ , Γ)
ω
D0 M ρn
where
βb − β ∗
q
1/2
τeB
:=
1/2
µ D00 τeB Kρn ,
1/2
τbB
(4.11)
1/2
+ C6 rm,m
.
c00
m2
− 2 exp(−δ 2 n/2000K 4 ),
r
log m
M
†/2
0 0
2 1/q
≤ C D0 K d0
τB kβ ∗ k2 +
;(4.12)
n
K
Then with probability at least 1 −
for 2 ≥ q ≥ 1,
and
Under the same assumptions, the predictive risk admits the following bound
2
1
b − β ∗ ) ≤ C 00 (kBk + amax )K 2 d0 log m τ ‡ K 2 kβ ∗ k2 + M 2 ,
X(
β
B
2
2
n
n
2
with the same probability as above, where c00 , C 0 , C 00 > 0 are some absolute constants.
We give an outline for the proof of Theorem 7 in Section 6.5, and show the actual proof
in Section 13.
4.1. Oracle results on the Lasso-type estimator
We now discuss the improvement being made in Theorem 6 and Theorem 7.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
25
Rudelson and Zhou/Errors-in-variables
The Signal-to-noise ratio. Let us redefine the Signal-to-noise ratio by
2
S/M :=
K 2 kβ ∗ k2
2
τB+ K 2 kβ ∗ k2 + M2
, where
2
2
:= K 2 kβ ∗ k2 and M := M2 + τB+ K 2 kβ ∗ k2 .
S
+/2
When either the noise level M or the measurement error strength in terms of τB K kβ ∗ k2
increases, we need to increase the penalty parameter λ correspondingly; moreover,
when d M12 lognm , we have
A
βb − β ∗
kβ ∗ k2
r
2
= OP
D00 K 2
M
1
S $(s0 + 1)
!
,
which eventually becomes a vacuous bound when M S.
1/2
2
2
. In this setting,
I, we have kBk2 = σw and τB = σw
Finally, suppose B = σw
we recover essentially the same `2 error bound as that in Corollary 1 of [30] in case
kβ ∗ k2 1, as we have on event A0 ∩ B0 ,
1/2
βb − β
∗
2
C(σw + amax )
≤
λmin (A)
r
q
d log m
2
2 kβ ∗ k
σ2 + σw
2
n
(4.13)
where σ2 M2 and K 2 1. However, when kβ ∗ k2 = Ω(1), our statistical precision
appears to be sharper as we allow the term kβ ∗ k2 to be removed entirely from the RHS
when σw → 0 and hence recover the regular Lasso rate of convergence.
The penalization parameter. We focus now on the penalization parameter λ in (1.7).
The effective rank condition in (3.2) implies that for n = O(m log m)
kBk2 ≤
τB
log m
≤ CB τB log m
16c0 K 4 log(3eMA3 /2) + log(m log m) − log n
(4.14)
1
where CB = 16c0 K 4 log(3eM
given that log(m log m) − log n > 0. This bound is
3
A /2)
very crude given that in practice, we focus on cases where n m log m. Note that
under (A1) (A2) and (A3), we have for n = O(m log m),
τB+
kAk2 + kBk2
m
1
λmin (A)
√
≤ τB + (κ(A)λmin (A) + CB τB log m) τB + O
.
m
m
τB +
Without knowing τB , we will use τbB as defined in (1.5). Notice that we know neither
2
D00 nor Doracle in the definition of λ, where Doracle
D2 ; Indeed,
2
2D2 ≤ Doracle
≤ 4D2 .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
26
Rudelson and Zhou/Errors-in-variables
However, assuming that we normalize the column norms of the design matrix X to be
roughly at the same scale, we have for τB = O(1) and m sufficiently large,
√
D00 1 while Doracle / m = o(1) in case kAk2 , kBk2 ≤ M
for some large enough constant M . In summary, compared to Theorem 3, in ψ, we
1/2
1/2
replace D2 = 2(kAk2 + kBk2 ) with D00 := kBk2
+ amax so that the dependency
on kAk2 becomes much weaker. As mentioned in Section 3.2, we may use the plugb 2 in D0 , where B
b is the corrected gram matrix 1 XX T − tr(A) Im .
in estimate kBk
0
m
m
Finally, the concentration of measure bound for the estimator τbB as in (1.5) is stated in
Lemma 5, which ensures that τbB is indeed a good proxy for τB (cf. Lemma 23).
The sparsity parameter. The condition on d (and Dφ ) for the Lasso estimator as
defined in (4.3) suggests that as τB → 0, and thus τB+ → 0, the constraint on the
sparsity parameter d becomes slightly more stringent when K 2 M2 /b20 1 and much
more restrictive when K 2 M2 /b20 = o(1). Moreover, suppose we require
2
M2 = Ω(τB+ K 2 kβ ∗ k2 ),
that is, the stochastic error in the response variable y as in (1.1a) does not converge
to 0 as quickly as the measurement error W in (1.1b) does, then the sparsity constraint
becomes essentially unchanged as τB+ → 0 as we show now.
+/2
Case 1. Suppose τB → 0 and M = Ω(τB K kβ ∗ k2 ). In this case, essentially, we
require that
0 00 2 2
c c K M
c0 λ2min (A) n
∧1
(4.15)
d≤ 2
$ (s0 + 1) log m
b20
2
where Dφ
τB+ K 4 kβ ∗ k2
K 2 M2
K 2 M2
+ 4
given
that
τ
K
φ
≤
B
2
2
b0
b0
b20
kBk +a
max
2
where c0 , c0 are absolute constants and c00 := $2 (s
1 where $(s0 +1) =
0 +1)
ρmax (s0 + 1, A) + τB . In this case, the sparsity constraint becomes essentially
unchanged as τB+ → 0.
Case 2. Analogous to (3.4), when M2 ≤ τB+ φK 2 b20 , we could represent the condition on
d as follows:
n
n
≤ CA c0 c00 Dφ
d ≤ CA c0 c00 τB+ K 4 φ
log m
log m
which is sufficient for (4.3) to hold for τB → 0; Indeed, by assumption that
c0 K 4 ≤ 1 and M2 ≤ τB+ φK 2 b20 , we have
8 > 2c0 K 4 τB+ φ ≥ c0 Dφ c0 τB+ K 4 φ.
Hence, for c0 τB+ K 4 ≤ 1, we have
n
n
CA c00 (c0 τB+ K 4 φ ∧ 8)
log m
log m
n
n
00 0 + 4
00 0
≤ CA c c τB K φ
CA c c Dφ
log m
log m
d ≤ CA (c0 c00 τB+ K 4 φ ∧ 8)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
27
Rudelson and Zhou/Errors-in-variables
This condition, however, seems to be unnecessarily strong, when τB → 0 (and
M → 0 simultaneously). We focus on the following Case 2 in the present work.
For both cases, it is clear that sample size needs to satisfy
(ρmax (s0 + 1, A) + τB )4
e
n = Ω d log m
,
λmin (A)2 (kBk2 + amax )
(4.16)
e
where Ω(·)
notation hides parameters K, M , φ and b0 , which we treat as absolute
constants that do not change as τB → 0. These tradeoffs are somehow different from
the behavior of the Conic programming estimator (cf (4.17)). We will provide a more
detailed analysis in Sections 6.1 and 6.3.
4.2. Oracle results on the Conic programming estimator
b ∗
In order to exploit the oracle bound as stated in Theorem 12 regarding γ
b − Γβ
∞
,
we need to know the noise level τB := tr(B)/n in W and then we can set
µ
Doracle
1/2
)Kρn while retaining ω D0 M ρn ,
D00 (τB + √
m
r
√
√
log m
where recall ρn = C0 K
and D0 = τB + amax .
n
This will in turn lead to improved bounds in Theorems 6 and 7.
The penalization parameter. Without knowing the parameter τB , we rely on the estimate from τbB as in (1.5), as discussed in Section 3. For a chosen parameter C6
√
+/2
1/2
1/2
1/2
Doracle , we use τbB + C6 rm,m to replace τB := τB + Doracle / m and set
r
log m
0
2 1/2
1/2
µ C0 D0 K (b
τB + Doracle rm,m )
n
b ∗
in view of Corollary 14, where an improved error bound over γ
b − Γβ
is stated.
∞
Without knowing Doracle , we could replace
it with an upper bound; for example, asq
n
2
suming that Doracle
kAk2 + kBk2 = O
log m , we could set
µ
1/2
C0 D00 K 2 (b
τB
r
−1/4
+ O(m
))
log m
.
n
The sparsity parameter. Roughly speaking, for the Conic programming estimator (1.8),
one can think of d0 as being bounded:
r
n ^
n
−1/2
d0 = O τB−
where τB− τB
(4.17)
log m
log(m/d0 )
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
28
That is, when τB decreases, we allow larger values of d0 ; however, when τB → 0,
the sparsity level of d = O (n/ log(m/d)) starts to dominate, which enables the Conic
programming estimator to achieve results similar to the Dantzig Selector when the
design matrix X0 is a subgaussian random matrix satisfying the Restricted Eigenvalue
conditions; See for example [6, 4, 38].
In particular, when τB → 0, Theorem 7 allows us to recover a rate close to that of the
Dantzig selector with an exact recovery if τB = 0 is known a priori; see Section 16.
Moreover the constraint (3.8) on the sparsity parameter d0 appearing in Theorem 4 can
now be relaxed as in (4.8). In summary, our results in Theorem 7 are stronger than
those in [3] (cf. Corollary 1) as their rates as stated therein are at the same order as ours
in Theorem 4. We illustrate this dependency on τB in Section 8 with numerical examples, where we clearly show an advantage by taking the noise level into consideration
when choosing the penalty parameters for both the Lasso and the Conic programming
estimators.
5. Optimization error on the gradient descent algorithm
We now present our computational convergence bounds. First we present Lemma 8
regarding the RSC and RSM conditions on the loss function (1.7). Lemma 8 follows
from Lemma 15 immediately.
Lemma 8. Suppose all conditions as stated in Theorem 3 hold. Suppose event A0
holds. Then (2.8) and (2.9) hold with α` = 85 λmin (A), αu = 11
8 λmax (A) and
τ` (Ln ) = τu (Ln ) = τ0
log m
,
n
where
τ0
400C 2 $(s0 + 1)2
.
λmin (A)
(5.1)
Theorem 9. Suppose all conditions in Theorem 6 hold and let ψ be defined therein.
Let g(β) = λ1 ρλ (β) where ρλ (β) = λ kβk1 . Consider the optimization program (1.10)
for a radius R such that β ∗ is feasible and a regularization parameter chosen such that
!
r
16Rξ _
log m
λ ≥
12ψ
.
(5.2)
1−κ
n
Suppose that the step size parameter ζ ≥ αu 23 λmax (A). Suppose that the sparsity
parameter and sample size further satisfy the following relationship:
n
λmin (A)2 ^ (α` )2
¯
d <
=: d.
(5.3)
512τ0 log m 12λmax (A)
5ζ
Then on event A0 ∩ B0 , the conclusions in Theorem 2 hold, where
c3 n
Vm log m
4c2 n
P (A0 ∩ B0 ) ≥ 1−4 exp − 2
log
−2 exp − 2 4 −22/m3 .
MA log m
n
MA K
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
29
Rudelson and Zhou/Errors-in-variables
Corollary 10. Suppose all conditions as stated in Theorem 9 hold and event A0 ∩ B0
defined therein holds. Consider for some constant M ≤ 400τ0 and δ̄ 2 as defined in
Theorem 2,
cε2stat d log m
=: δ̄ 2 and δ 2 ≤ M δ̄ 2 ≤ 400τ0 δ̄ 2 .
1−κ n
√
Then for all t ≥ T ∗ (δ) as in (2.17) and R = Ω(b0 d),
2 2
2
3 2 α` 2
δ εstat
≤
β t − βb
.
δ + εstat + O
α`
4
b20
2
δ2
Finally, suppose we fix for M+ =
R
p
¯0
db
(5.4)
32C$(s0 +1)
λmin (A) ,
20M+
b0
p
r
6κ(A)
n
,
log m
in view of the upper bound d¯ (5.3). Then for all t ≥ T ∗ (δ) as in (2.17),
β t − βb
2
≤
2
3 2 α` 2
2
δ4
δ + εstat +
.
2
α`
4
α` b0 kAk2
(5.5)
We prove Theorem 9 and Corollary 10 in Section 14.
5.1. Discussions
Throughout this section, we assume ψ (4.2) is as defined in√Theorem 6. Assume that
ζ ≥ αu ≥ ᾱ` . In addition, suppose that the radius R b0 d as we set in (1.7). Let
d¯0 ≤ 160M n2 log m be as defined in (4.3), where recall that we require the following
+
condition on d:
d ≤ CA {c0 Cφ ∧ 8}
Cφ
=
n
1
=: d¯0 , where CA =
2,
log m
160M+
kBk2 + amax
2
Dφ and b20 ≥ kβ ∗ k2 ≥ φb20 .
$(s0 + 1)2
Then by the proof of Lemma 18,
r
r
p
5s0 log m
ψ log m
¯
ψ =:
,
b0 d 0 ≤
3α
n
τ
n
where
τ=
3α
.
5s0
(5.6)
In contrast, under (5.3), the following upper bound holds on d, which is slightly more
restrictive in the sense that the maximum level of sparsity allowed on β ∗ has decreased
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
30
Rudelson and Zhou/Errors-in-variables
by a factor proportional to κ(A) compared to the upper bound d¯0 (4.3) in Theorem 6;
¯ where for CA = 1 2 ,
Now we require that |supp(β ∗ )| ≤ d,
160M
+
2
nλmin (A)
1
1024C 2 $(s0 + 1)2 log m 2400κ(A)
n
λmin (A)
1
≈ CA
d¯0
.
log m 15λmax (A)
κ(A)
d¯
(5.7)
To consider the general cases as stated in Theorem 6, we consider the ideal case when
we set
11
λmax (A)
ζ = αu =
8
such that
59
5
ζ
≈ αu /( α` ) κ(A), where α` = λmin (A).
ᾱ`
60
8
Following the derivation in Remark 14.1, we have
80ζ
ξ
≤ 6τ (Ln ) +
τ (Ln ) ≈ 200κ(A)τ (Ln ).
1−κ
ᾱ`
Combining (5.6) and (5.8), it is clear that one can set
!
r
log m
λ = Ω κ(A)ψ
n
in order to satisfy the condition (5.2) on λ in Theorem 2 when we set
!
r
p
log
m
ψ
R b0 d¯0 = O
τ
n
!
r
log m
and hence Rτ κ(A) = O κ(A)ψ
.
n
(5.8)
(5.9)
(5.10)
This choice is potentially too conservative because we are setting R in (5.10) with respect to the upper sparsity level d¯0 chosen to guarantee statistical convergence, leading
to a larger than necessary penalty parameter as in (5.9). Similarly, when we choose
step size parameter ζ to be too large, we need to increase
the penalty parameter λ
Rξ
correspondingly given the following lower bound: λ = Ω 1−κ
where
!!
ᾱ`
Rξ
5
4ζ + 2%
= R 2τ (Ln ) ᾱ`
+ ᾱ`
1−κ
4ζ − 2%
4ζ − 2%
≥ 40Rτ (Ln )
ζ
+ 2Rτ (Ln ) Rτ (Ln )κ(A).
ᾱ`
Suppose we set ζ = 23 λmax (A) and ᾱζ` ≈ 3κ(A) as in Theorem 9. It turns out that the
less conservative choice of λ as in (5.11)
r
p
_ log m
λ
b0 κ(A)$(s0 ) ψ
(5.11)
n
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
31
is sufficient, for example when τB = Ω(1), for which we now set
r
√
b0
1
n
p
R b0 d
20M+ 6κ(A) log m
as in Corollary 10. We will discuss the two scenarios as considered in Section 4. See
the detailed discussions in Section 14.
6. Proof of theorems
b ∗ .
b − Γβ
In Section 6.1, we develop in Theorem 12 the crucial large deviation bound on γ
This entity appears in the constraint set in the Conic programming estimator (1.8), and
is directly related to the choice of λ for the corrected Lasso estimator in view of Theorem 16. Its corollaries are stated in Corollary 13 and Corollary 14. In section 6.2,
we provide an outline and additional Lemmas 15 and 17 to prove Theorem 3. The full
proof of Theorem 3 appears in Section 10. In Section 6.3, we give an outline illustrating
the improvement for the Lasso error bounds as stated in Theorem 6. We emphasize the
impact of this improvement over sparsity parameter d, which we restate in Lemma 18.
In Section 6.4, we provide an outline as well as technical results for Theorem 4. In
Section 6.5, we give an outline illuminating the improvement in error bounds for the
Conic programming estimator as stated in Theorem 7.
6.1. Stochastic error terms
In this section, we first develop stochastic error bounds in Lemma 11, where we also
define some events B4 , B5 , B10 . Recall that B6 was defined in Lemma 5. Putting the
bounds in Lemma 11 together with that in Lemma 5 yields Theorem 12.
2
2
Lemma 11. Assume that the stable rank of B, kBkF / kBk2 ≥ log m. Let Z, X0 and
T
W as defined in Theorem 3. Let Z0 , Z1 and Z2 be independent copies of Z. Let
q ∼
Y M /K where Y := eT1 Z0T . Denote by B4 the event such that for ρn := C0 K
and
1
1
n
A 2 Z1T
1
n
Z2T B 2
∞
1
∞
log m
n ,
≤ ρn M a1/2
max
√
tr(B)
≤ ρn M τB where τB =
.
n
Then P (B4 ) ≥ 1 − 4/m3 . Moreover, denote by B5 the event such that
1
n
(Z T BZ − tr(B)Im )β ∗
∞
≤
X0T W β ∗
∞
≤
and
1
n
kBk
ρn K kβ ∗ k2 √ F
n
√
∗
ρn K kβ k2 τB a1/2
max .
Then P (B5 ) ≥ 1 − 4/m3 .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
32
Finally, denote by B10 the event such that
1
n
(Z T BZ − tr(B)Im )
and
1
n
X0T W
max
≤
max
≤
kBk
ρn K √ F
n
√
ρn K τB a1/2
max .
Then P (B10 ) ≥ 1 − 4/m2 .
We prove Lemma 11 in Section C.2. Denote by B0 := B4 ∩ B5 ∩ B6 , which we use
throughout this paper.
q
Theorem 12. Suppose (A1) holds. Let ρn = C0 K lognm . Suppose that
2
2
kBkF / kBk2 ≥ log m where m ≥ 16.
√
√
b and γ
b be as in (1.6). Let D0 = τB + amax and D00 be as defined in (2.23).
Let Γ
kAkF
kBkF
Let D1 = √m + √n . On event B0 , for which P (B0 ) ≥ 1 − 16/m3 ,
2D1 K ∗
1/2
∗
0
∗
b
γ
b − Γβ
≤
D0 KτB kβ k2 + √
kβ k∞ + D0 M ρn . (6.1)
m
∞
We next state the first Corollary 13 of Theorem 12, which we use in proving Theob ∗
rems 3 and 4. Here we state a somewhat simplified bound on γ
b − Γβ
for the sake
∞
of reducing the number of unknown parameters involved with a slight worsening of the
statistical error bounds when τB 1. On the other hand, the bound in (6.1) provides a
significant improvement over the error bound in Corollary 13 in case τB = o(1).
b and γ
Corollary 13. Suppose all conditions in Theorem 12 hold. Let Γ
b be as in (1.6).
On event B0 , we have for D2 = 2(kAk2 + kBk2 ) and some absolute constant C0
r
log m
∗
b
, where ψ = C0 D2 K (K kβ ∗ k2 + M )
γ
b − Γβ
≤ψ
n
∞
is as defined in Theorem 3.
In particular, Corollary 13 ensures that for the corrected Lasso estimator, (6.7) holds
with high probability for λ chosen as in (3.6). We prove Corollary 13 in Section D.
p
√
1/2
1/2
What happens when τB → 0? Recall D0 = τB + amax and D00 := kBk2 + amax .
When τB → 0, we have by Theorem 12
r
log m
1
∗
∗
b
γ
b − Γβ
= O D1 K √ kβ k∞ + D0 KM K
n
m
∞
1/2
kAk
kBk
1/2
where D0 → amax and D1 = √mF + √nF → kAk2 under (A1), given that
√
1/2
1/2
kBkF / n ≤ τB kBk2 → 0. In this case, the error term involving kβ ∗ k2 in (4.2)
vanishes, and we only need to set (cf. Theorem 16)
r
log m
1/2
2
∗
1/2
, (6.2)
λ ≥ 2ψ
for ψ a1/2
max KM + kAk2 K kβ k∞ /m
n
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
33
Rudelson and Zhou/Errors-in-variables
where the second term in ψ defined immediately above comes from the estimation error
in Lemma 5; this term vanishes if we were to assume that (1) tr(B) is also
qknown or
1/2
(2) kβ ∗ k∞ = o(M m1/2 /K). For both cases, by setting λ 4amax KM
can recover the regular Lasso rate of
βb − β ∗
q
= Op (λd1/q ),
log m
n ,
we
for q = 1, 2,
when the design matrix X is almost free of measurement errors.
Finally, we state a second Corollary 14 of Theorem 12. Corollary 14 is essentially a
restatement of the bound in (6.1).
Corollary 14. Suppose all conditions in Theorem 12 hold. Let D0 , D00 , Doracle , and
+/2
1/2
oracle
τB := τB + D√
be as defined in (2.23) and (4.1). On event B0 ,
m
r
b ∗
γ
b − Γβ
≤ ψ
∞
log m
,
n
where
+/2
ψ := C0 K D00 τB K kβ ∗ k2 + D0 M .
(6.3)
Then P (B0 ) ≥ 1 − 16/m3 .
We mention in passing that Corollaries 13 and 14 are crucial in proving Theorems 3, 4, 6
and 7.
6.2. Outline for proof of Theorem 3
In this section, we state Theorem 16, and two Lemmas 15 and 17. Theorem 3 follows from Theorem 16 in view of Corollary 13, Lemmas 15 and 17. In more details,
Lemma 15 checks the Lower and the Upper RE conditions on the modified gram matrix,
b A :=
Γ
1
T
n (X X
b
− tr(B)I
m ),
(6.4)
while Lemma 17 checks condition (6.6) as stated in Theorem 16 for curvature α and
tolerance τ regarding the lower RE condition as derived in Lemma 15.
First, we replace (A3) with (A3’) which reveals some additional information regarding
the constant hidden inside the O(·) notation.
(A3’) Suppose (A3) holds; moreover, mn ≥ 4096C02 D22 K 4 log m/λ2min (A) for D2 =
2(kAk2 + kBk2 ), or equivalently,
r
λmin (A)
log m
> CK
for some large enough contant CK .
kAk2 + kBk2
mn
Lemma 15. (Lower and Upper-RE conditions) Suppose (A1), (A2) and (A3’) hold.
Denote by V := 3eMA3 /2, where MA is as defined in (2.7). Let s0 ≥ 32 be as defined
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
34
Rudelson and Zhou/Errors-in-variables
b as
in (2.6). Recall that we denote by A0 the event that the modified gram matrix Γ
defined in (1.6) satisfies the Lower as well as Upper RE conditions with
curvature
and tolerance
11
5
λmin (A), smoothness α
e=
λmax (A)
8
8
384C 2 $(s0 )2 log m
λmin (A) − α
396C 2 $2 (s0 + 1) log m
≤ τ :=
≤
λmin (A)
n
s0
λmin (A)
n
α=
for α, α
e and τ as defined in Definitions 2.2 and 2.3, and C, s0 , $(s0 ) in (2.6). Suppose
that for some c0 > 0 and c0 K 4 < 1,
3em
s0
1
tr(B)
≥ c0 K 4 2 log
where ε =
.
(6.5)
kBk2
ε
s0 ε
2MA
Vm log m
4c2 n
n
log
−
2
exp
−
− 6/m3 .
Then P (A0 ) ≥ 1 − 4 exp − M 2c3log
n
m
M 2 K4
A
A
The main focus of the current section is then to apply Theorem 16 to show Theorem 3.
Theorem 16 follows from Theorem 1 by [30].
b
Theorem 16. Consider the regression model in (1.1a) and (1.1b). Let d ≤ n/2. Let γ
b, Γ
b
be as constructed in (1.6). Suppose that the matrix Γ satisfies the Lower-RE condition
with curvature α > 0 and tolerance τ > 0,
√
α
λ
√ ,
,
(6.6)
dτ ≤ min
32 d 4b0
where d, b0 and λ are as defined in (1.7). Then for any d-sparse vectors β ∗ ∈ Rm , such
that kβ ∗ k2 ≤ b0 and
b ∗
γ
b − Γβ
∞
≤
1
λ,
2
(6.7)
the following bounds hold:
βb − β ∗
≤
2
20 √
λ d and
α
βb − β ∗
≤
1
80
λd,
α
(6.8)
where βb is an optimal solution to the corrected Lasso estimator as in (1.7).
We include the proof of Theorem 16 for the sake of self-containment and defer it to
Section G for clarity of presentation.
Lemma 17. Let c0 , φ, b0 , M , M+ and K be as defined in Theorem 3, where we assume
2
that b20 ≥ kβ ∗ k2 ≥ φb20 for some 0 < φ ≤ 1. Suppose all conditions in Lemma 15
hold. Suppose that s0 ≥ 32 and
n
1
d := |supp(β ∗ )| ≤ CA
{c0 Dφ ∧ 2} where CA :=
2 (6.9)
log m
40M+
M2
and Dφ = K 4
+ φ ≥ K 4 φ ≥ φ.
K 2 b20
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
35
Rudelson and Zhou/Errors-in-variables
Then the following condition holds
2
α ^ 1 log m ψ
,
d≤
32τ
τ2 n
b0
(6.10)
where ψ = C0 D2 K(K kβ ∗ k2 + M ) is as defined in (3.6), α = 5λmin (A)/8, and τ is
as defined in Lemma 15.
We prove Lemmas 15 and 17 in Sections F and H.1 respectively. Lemma 15 follows
immediately from Corollary 25. We prove Lemmas 15 and Corollary 25 in Sections F
and L respectively.
Remark 6.1. Clearly for d, b0 , φ as bounded in Theorem 3, we have by assumption (3.3) the following upper and lower bound on Dφ :
2 2
M K
4
4
+ K φ ≥ K 4 φ.
2K φ ≥ Dφ :=
b20
In this regime, the conditions on d as in (6.9) can be conveniently expressed as that in
(3.4) instead.
6.3. Improved bounds for the corrected Lasso estimator
The proof of Theorem 6 follows exactly the same line of arguments as in Theorem 3,
b ∗
except that we now use the improved bound on the error term γ
b − Γβ
given
∞
in Corollary 14, instead of that in Corollary 13. Moreover, we replace Lemma 17 with
Lemma 18, the proof of which follows from Lemma 17 with d now being bounded as in
(4.3) and ψ being redefined as in (6.3). The proof of Lemma 18 appears in Section H.2.
See Section 12 for the proof of Theorem 6.
Lemma 18. Let c0 , φ, b0 , M , M+ and K be as defined in Theorem 3. Suppose all
conditions in Lemma 15 hold. Suppose that (4.3) holds:
1
n
{c0 c00 Dφ ∧ 8} , where CA :=
2 , (6.11)
log m
160M+
2
D00
K 2 M2
≤
and Dφ =
+ τB+ K 4 φ.
$(s0 + 1)
b20
d := |supp(β ∗ )| ≤ CA
c00 :=
kBk2 + amax
$(s0 + 1)2
Then (6.10) holds with ψ as defined in Theorem 6 and α = 58 λmin (A).
6.4. Outline for proof of Theorem 4
We provide an outline and state the technical lemmas needed for proving Theorem 4.
Our first goal is to show that the following holds with high probability,
b ∗
γ
b − Γβ
∞
=
1
T
n X (y
∗
b
− Xβ ∗ ) + n1 tr(B)β
∞
≤ µ kβ ∗ k2 + ω,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
36
where µ, ω are chosen as in (6.12). This forms the basis for proving the `q convergence,
where q ∈ [1, 2], for the Conic programming estimator (1.8). This follows immediately
from Theorem 12 and Corollary 13. More explicitly, we will state it in Lemma 19.
√
√
Lemma 19. Let D0 = τB + amax and D2 = 2(kAk2 + kBk2 ) be as in Theorem 4.
Suppose all conditions in Theorem 12 hold. Then on event B0 as defined therein, the
pair (β, t) = (β ∗ , kβ ∗ k2 ) belongs to the feasible set of the minimization problem (1.8)
with
r
log m
.
(6.12)
µ 2D2 Kρn and ω D0 M ρn ,
where ρn := C0 K
n
Before we proceed, we first need to introduce some notation and definitions. Let X0 =
Z1 A1/2 be defined as in (1.4). Let k0 = 1 + λ. First we need to define the `q -sensitivity
parameter for Ψ := n1 X0T X0 following [3]:
κq (d0 , k0 )
=
ConeJ (k0 )
=
kΨ∆k∞
, where
k∆kq
(6.13)
{x ∈ Rm | s.t. kxJ c k1 ≤ k0 kxJ k1 } .
(6.14)
min
min
J:|J|≤d0 ∆∈ConeJ (k0 )
bb
See also [21]. Let (β,
t) be the optimal solution to (1.8) and denote by v = βb − β ∗ . We
will state the following auxiliary lemmas, the first of which is deterministic in nature.
The two lemmas reflect the two geometrical constraints on the optimal solution to (1.8).
The optimal solution βb satisfies:
1. The vector v obeys the following cone constraint: kvS c k1 ≤ k0 kvS k1 , and b
t≤
1
∗
λ kvk1 + kβ k2 .
2. kΨvk∞ is upper bounded by a quantity at the order of O (µ(kβ ∗ k2 + kvk1 ) + ω).
Lemma 20. Let µ, ω > 0 be set. Suppose that the pair (β, t) = (β ∗ , kβ ∗ k2 ) belongs
bb
to the feasible set of the minimization problem (1.8), for which (β,
t) is an optimal
∗
b
solution. Denote by v = β − β . Then
kvS c k1
≤
1
(1 + λ) kvS k1 and b
t ≤
kvk1 + kβ ∗ k2 .
λ
Lemma 21. On event B0 ∩ B10 ,
kΨvk∞ ≤ µ1 kβ ∗ k2 + µ2 kvk1 + ω 0 ,
where µ1 = 2µ, µ2 = µ( λ1 + 1) and ω 0 = 2ω for µ, ω as defined in (6.12).
Now combining Lemma 6 of [3] and an earlier result of the two authors (cf. Theorem 27 [38]), we can show that the RE(2d0 , 3(1 + λ), A1/2 ) condition and the sample
requirement as in (4.9) are enough to ensure that the `q -sensitivity parameter satisfies
the following lower bound for all 1 ≤ q ≤ 2: for some contant c,
κq (d0 , k0 ) ≥
−1/q
cd0
,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
37
Rudelson and Zhou/Errors-in-variables
which ensures that for v = βb − β ∗ and Ψ =
kΨvk∞
1
T
n X0 X0 ,
−1/q
≥
κq (d0 , k0 )kvkq ≥ cd0
kvkq .
(6.15)
Combining (6.15) with Lemmas 19, 20 and 21 gives us both the lower and upper
bounds on kΨvk∞ , with the lower bound being κq (d0 , k0 ) kvkq and the upper bound as
specified in Lemma 21. Following some algebraic manipulation, this yields the bound
on the kvkq for all 1 ≤ q ≤ 2. We prove Theorem 4 in Section 11 and Lemmas 19, 20
and 21 in Section I. The proof of Lemma 20 follows the same line of arguments in [3]
in view of Lemma 19.
6.5. Improved bounds for the DS-type estimator
Lemma 22 follows directly from Corollary 14.
√
√
Lemma 22. Suppose all conditions in Corollary 14 hold. Let D0 = τB + amax 1
under (A1). Then on event B0 , the pair (β, t) = (β ∗ , kβ ∗ k2 ) belongs to the feasible set
Υ of the minimization problem (1.8) with
+/2
µ ≥ D00 τB Kρn and ω ≥ D0 M ρn ,
+/2
where τB
1/2
:= τB +
D√
oracle
m
is as defined in (4.1).
1/2
Lemma 23. On event B6 andq
(A1), the choice of τeB
where recall rm,m = 2C0 K 2
+/2
τB
τeB
(6.16)
log m
mn ,
1/2
1/2
:= τbB + C6 rm,m as in (4.11),
satisfies for m ≥ 16 and C0 ≥ 1,
3
1/2
†/2
1/2
≤ τB + C6 rm,m
=: τB ,
2
1/2
≤ 2τB + 3C62 rm,m τB‡ , and moreover τeB τB− ≤ 1.
1/2
≤ τeB
(6.17)
(6.18)
We next state an updated result in Lemma 24.
Lemma 24. On event B0 ∩ B10 , the solution βb to (1.8) with µ, ω as in (4.11) satisfies
for v := βb − β ∗
1
T
n X0 X0 v ∞
where µ1 = 2µ, µ2 = 2µ(1 +
≤ µ1 kβ ∗ k2 + µ2 kvk1 + ω 0 ,
1
2λ )
and ω 0 = 2ω.
7. Lower and Upper RE conditions
The goal of this section is to show that for ∆ defined in (7.4), the presumption in
Lemmas 37 and 39 as restated in (7.1) holds with high probability (cf Theorem 26).
We first state a deterministic result showing that the Lower and Upper RE conditions
b A under condition (7.1) in Corollary 25. This allows us to prove Lemma 15 in
hold for Γ
Section F. See Sections K and L, where we show that Corollary 25 follows immediately
from the geometric analysis result as stated in Lemma 39.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
38
Rudelson and Zhou/Errors-in-variables
Corollary 25. Let 1/8 > δ > 0. Let 1 ≤ k < m/2. Let Am×m be a symmetric
b A be an m × m symmetric matrix and
positive semidefinite covariance matrice. Let Γ
b
∆ = ΓA − A. Let E = ∪|J|≤k EJ , where EJ = span{ej : j ∈ J}. Suppose that
∀u, v ∈ E ∩ S m−1
uT ∆v ≤ δ ≤
3
λmin (A).
32
(7.1)
Then the Lower and Upper RE conditions hold: for all υ ∈ Rm ,
and
bA υ
υT Γ
≥
bA υ
υT Γ
≤
3λmin (A)
5
2
2
λmin (A) kυk2 −
kυk1
8
8k
11
3λmin (A)
2
2
λmax (A) kυk2 +
kυk1 .
8
8k
(7.2)
(7.3)
Theorem 26. Let Am×m , Bn×n be symmetric positive definite covariance matrices.
Let E = ∪|J|≤k EJ for 1 ≤ k < m/2. Let Z, X be n × m random matrices defined as
in Theorem 3. Let τbB be defined as in (1.5). Let
b A − A :=
∆ := Γ
1
T
nX X
− τbB Im − A.
(7.4)
Suppose that for some absolute constant c0 > 0 and 0 < ε ≤ C1 ,
_
3em
tr(B)
0 4 k
≥
c K 2 log
log m,
(7.5)
kBk2
ε
kε
√
where C = C0 / c0 for C0 as chosen to satisfy (H.6).
Then with probability at least 1 − 4 exp −c2 ε2 Ktr(B)
− 2 exp −c2 ε2 Kn4 − 6/m3
4 kBk
2
for c2 ≥ 2, we have for all u, v ∈ E ∩ S m−1 ,
r
T
u ∆v ≤ 8C$(k)ε + 4C0 D1 K
where $(k) = τB + ρmax (k, A), and D1 ≤
kAkF
√
m
2
+
log m
,
mn
kBkF
√
n
,
We prove Theorem 26 in Section M.
8. Numerical results
In this section, we present results from numerical simulations designed to validate the
theoretical predictions as presented in previous sections. We implemented the composite gradient descent algorithm as described in [1, 30, 31] for solving the corrected Lasso
b γ
objective function (1.7) with (Γ,
b) as defined in (1.6). For the Conic programming estimator, we use the implementation provided by the authors [3] with the same input
b γ
(Γ,
b) (1.6). Throughout our experiments, A is a correlation matrix with amax = 1. We
√
1/2
set the following as our default parameters: D00 = kBk2 + 1, D0 = τB + 1 and
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
39
√
R = kβ ∗ k2 d, where d is the sparsity parameter, the number of non-zero entries in
β ∗ . In one set of simulations, we also vary R.
In our simulations, we look at three different models from which A and B will be
chosen. Let Ω = A−1 = (ωij ) and Π = B −1 = (πij ). Let E denote edges in Ω, and
F denote edges in Π. We choose A from one of these two models:
• AR(1) model. In this model, the covariance matrix is of the form A = {ρ|i−j| }i,j .
The graph corresponding to the precision matrix A−1 is a chain.
• Star-Block model. In this model the covariance matrix is block-diagonal with
equal-sized blocks whose inverses correspond to star structured graphs, where
Aii = 1, for all i. We have 32 subgraphs, where in each subgraph, 16 nodes
are connected to a central hub node with no other connections. The rest of the
nodes in the graph are singletons. The covariance matrix for each block S in A
is generated by setting Sij = ρA if (i, j) ∈ E, and Sij = ρ2A otherwise.
We choose B from one of the following models. Recall that τB = tr(B)/n.
• For B and B ∗ = B/τB = ρ(B), we consider the AR(1) model with two parameters. First we choose the AR(1) parameter ρB ∗ ∈ {0.3, 0.7} for the correlation
matrix B ∗ . We then set B = τB B ∗ , where τB ∈ {0.3, 0.7, 0.9} depending on
the experimental setup.
• We also consider a second model based on Π = B −1 , where we use the random
concentration matrix model in [57]. The graph is generated according to a type
of Erdős–Rényi random graph model. Initially, we set Π = cIn×n , and c is
a constant. Then we randomly select n log n edges and update Π as follows:
for each new edge (i, j), a weight w > 0 is chosen uniformly at random from
[wmin , wmax ] where wmax > wmin > 0; we subtract w from πij and πji , and
increase πii and πjj by w. This keeps Π positive definite. We then rescale B to
have a certain desired trace parameter τB .
For a given β ∗ , we first generate matrices A and B, where A is m × m and B is n × n.
For the given covariance matrices A and B, we repeat the following steps to estimate
β ∗ in the errors-in-variables model as in (1.1a) and (1.1b),
1. We first generate random matrices X0 ∼ Nf,m (0, A ⊗ I) and W ∼ Nf,m (0, I ⊗
B) independently from the matrix variate normal distribution as follows. Let Z ∈
Rn×m be a Gaussian random ensemble with independent entries Zij satisfying
2
EZij = 0, EZij
= 1. Let Z1 , Z2 be independent copies of Z. Let X0 = Z1 A1/2
and W = B 1/2 Z2 , where A1/2 and B 1/2 are the unique square root of the
positive definite matrix A and B = τB B ∗ respectively.
2. We then generate X = X0 + W and y = X0 β ∗ + , where i i.i.d. ∼ N (0, 1).
b according to (1.5) and (1.6) using X, y, where by (1.5),
We compute τbB , γ
b and Γ
2
1
b
= mn
kXkF − ntr(A) + .
τbB := n1 tr(B)
3. Finally, we feed X and y to the Composite Gradient Descent algorithm as described in [1, 30] to solve the Lasso program (1.7) to recover β ∗ , where we set
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
40
Rudelson and Zhou/Errors-in-variables
b the estithe step size parameter to be ζ. The output of this step is denoted by β,
∗
∗
∗
b
b
mated β . We then compute the relative error of β: kβ − β k/kβ k, where k·k
denotes either the `1 or the `2 norm.
The final relative error is the average of 100 runs for each set of tuning and step-size
parameters; for the Conic programming estimator, we solve (1.8) instead of (1.7) to
recover β ∗ .
8.1. Relative error
In the first experiment, A and B are generated using the AR(1) model with parameters
ρA , ρB ∗ ∈ {0.3, 0.7} and trace parameter τB ∈ {0.3, 0.7, 0.9}. We see in Figures 1
and 2 that a larger sample size is required when ρA , ρB ∗ or τB increases. To explain
these results, we first recall the following definition of the Signal-to-noise ratio, where
we take K = Mε 1
2
S/M
kβ ∗ k2
2
τB kβ ∗ k2 +
2
kβ ∗ k2 and
S :=
1
=
1
2
τB + (1/ kβ ∗ k2 )
, where
2
M := 1 + τB kβ ∗ k2 ,
2
which clearly increases as kβ ∗ k2 increases or as the measurement error metric τB decreases. We keep kβ ∗ k2 = 5 throughout our simulations. The corrected Lasso recovery
problem thus becomes more difficult as τB increases. Indeed, we observe that a larger
sample size n is needed when τB increases from 0.3 to 0.9 in order to control the relative `2 error to stay at the same level. Moreover, in view of Theorem 6, we can express
the relative error as follows: for α λmin (A) and K 1,
βb − β ∗
kβ ∗ k2
1/2
2
= OP
(kBk2 + 1)
λmin (A)
r
M
S
r
d log m
n
!
.
Note that when kβ ∗ k2 is large enough and τB = Ω(1), the factor preceding
on the RHS of (8.1) is proportional to
√
1/2
(kBk2 +1) τB
λmin (A)
(8.1)
q
d log m
n
.
When we plot the relative `2 error kβb − β ∗ k2 /kβ ∗ k2 versus the rescaled sample size
n
d log m under the same S/M ratio, the two sets of curves corresponding to ρA = 0.3
and ρA = 0.7 indeed line up in Figure 1(b), as predicted by (8.1). We observe in
Figure 1(b), the rescaled curves overlap well for different values of (m, d) for each
ρA when we keep (ρB ∗ , τB ) and the length kβ ∗ k2 = 5 invariant. Moreover, the upper
bound on the relative `2 error (8.1) characterizes the relative positions of these two
sets of curves in that the ratio between the `2 error corresponding to ρA = 0.7 and
that for ρA = 0.3 along the y-axis roughly falls within the interval (2, 3) for each n,
while λmin (AR(1), 0.3)/λmin (AR(1), 0.7) = 3. These results are consistent with the
theoretical predictions in Theorems 3 and 6.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
41
Rudelson and Zhou/Errors-in-variables
0.8
ρB*=0.3 τB=0.3
0.8
ρB*=0.3 τB=0.3
0.5
0.6
0.7
ρA=0.3 m = 256
ρA=0.3 m = 512
ρA=0.3 m = 1024
ρA=0.7 m = 256
ρA=0.7 m = 512
ρA=0.7 m = 1024
0.1
0.2
0.3
0.4
relative l2 error
0.5
0.4
0.1
0.2
0.3
relative l2 error
0.6
0.7
ρA=0.3 m = 256
ρA=0.3 m = 512
ρA=0.3 m = 1024
ρA=0.7 m = 256
ρA=0.7 m = 512
ρA=0.7 m = 1024
500
1000
1500
2000
2500
2
4
n
6
(a)
0.8
0.6
relative l2 error
0.2
0.4
0.2
τB=0.3 m = 256
τB=0.3 m = 512
τB=0.3 m = 1024
τB=0.7 m = 256
τB=0.7 m = 512
τB=0.7 m = 1024
τB=0.9 m = 256
τB=0.9 m = 512
τB=0.9 m = 1024
0.4
0.8
12
ρA=0.3 ρB* = 0.3
τB=0.3 m = 256
τB=0.3 m = 512
τB=0.3 m = 1024
τB=0.7 m = 256
τB=0.7 m = 512
τB=0.7 m = 1024
τB=0.9 m = 256
τB=0.9 m = 512
τB=0.9 m = 1024
0.6
10
(b)
ρA=0.3 ρB* = 0.3
relative l2 error
8
n / d log(m)
500
1000
1500
n / d log(m)
(c)
2000
2500
2
4
6
8
10
12
n / d log(m)
(d)
F IG 1. Plots of the relative `2 error after running composite gradient descent √
algorithm on recovering β ∗
using the corrected Lasso objective function with sparsity parameter d = b mc, where we vary m ∈
{256, 512, 1024}. We generate A and B using the AR(1) model with ρA , ρB ∗ ∈ {0.3, 0.7} and τB =
{0.3, 0.7, 0.9}. In the left and right column, we plot the relative `2 error with respect to sample size n as
well as the rescaled sample size n/(d log m). As n increases, we see that the statistical error decreases. In
the top row, we vary the AR(1) parameter ρA ∈ {0.3, 0.7}, while holding (τB , ρ∗B ) and kβ ∗ k2 invariant.
Plot (a) shows the relative `2 error versus n for m = 256, 512, 1024. In Plots (c) and (d), we vary the trace
parameter τB ∈ {0.3, 0.7, 0.9}, while fixing the AR(1) parameters ρA , ρ∗B = 0.3. Plot (b) and (d) show
the relative `2 error versus the rescaled sample size n/(d log m). The curves now align for different values
of m in the rescaled plots, consistent with the theoretical prediction in Theorem 6.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
42
Rudelson and Zhou/Errors-in-variables
0.8
τB= 0.3
ρB*=0.3
ρB*=0.3
ρB*=0.3
ρB*=0.7
ρB*=0.7
ρB*=0.7
0.6
m=256
m=512
m=1024
m=256
m=512
m=1024
τB= 0.7
ρB*=0.3
ρB*=0.3
ρB*=0.3
ρB*=0.7
ρB*=0.7
ρB*=0.7
0.2
relative l2 error
τB= 0.7
ρB*=0.3
ρB*=0.3
ρB*=0.3
ρB*=0.7
ρB*=0.7
ρB*=0.7
0.4
τB= 0.3
ρB*=0.3
ρB*=0.3
ρB*=0.3
ρB*=0.7
ρB*=0.7
ρB*=0.7
0.4
0.6
m=256
m=512
m=1024
m=256
m=512
m=1024
ρA = 0.3
0.2
relative l2 error
0.8
ρA = 0.3
500
1000
1500
n
(a)
2000
2500
2
4
6
8
10
12
n / d log(m)
(b)
F IG 2. Plots of the relative `2 error after running composite gradient descent √
algorithm on recovering β ∗
using the corrected Lasso objective function with sparsity parameter d = b mc, where we vary m ∈
{256, 512, 1024}. We generate A and B using the AR(1) model with ρA = 0.3 and ρ∗B ∈ {0.3, 0.7}.
We set B = τB B ∗ and vary the trace parameter τB ∈ {0.3, 0.7} for each value of ρB ∗ . The parameters
1/2
1/2
1/2
τB and ρ∗B affect the rate of convergence through D00 = kBk2 + amax and τB . Plot (b) shows the
relative `2 error versus the rescaled sample size n/(d log m). We observe that as τB increases from 0.3 to
0.7, the two sets of curves corresponding to ρB ∗ = 0.3, 0.7 become visibly more separated. As n increases,
all curves converge to 0.
In Figure 1(c) and (d), we also show the effect of τB when τB is chosen from {0.3, 0.7, 0.9},
while fixing the AR(1) parameters ρA = 0.3 and ρB ∗ = 0.3. As predicted by our theory, as the measurement error magnitude τB increases, M increases, resulting in a larger
relative `2 error for a fixed sample size n.
While the effect of ρA as shown in (8.1) through the minimal eigenvalue of A is directly
visible in Figure 1(b), the effect of ρB ∗ is more subtle, as it is modulated by τB as
shown in Figure 2(a) and (b). When τB is fixed, our theory predicts that kBk2 plays
a role in determining the `p error, p = 1, 2, through the penalty parameter λ in view
1/2
1/2
of (8.1). The effect of ρB ∗ , which goes into the parameter D00 = kBk2 +amax 1, is
not changing the sample requirement or the rate of convergence as significantly as that
of ρA when τB = 0.3. This is shown in the bottom set of curves in Figure 2(a) and (b).
On the other hand, the trace parameter τB plays a dominating role in determining the
sample size as well as the `p error for p = 1, 2, especially when the length of the signal
β ∗ is large: kβ ∗ k2 = Ω(1). In particular, the separation between the two sets of curves
in Figure 2(b), which correspond to the two choices of ρB ∗ , is clearly modulated by
τB and becomes more visible when τB = 0.7.
These findings are also consistent with our theoretical prediction that in order to guarantee statistical and computational convergence, the sample size needs to grow according to the following relationship to be specified in (8.2). We will show in the proof of
Theorem 9 that the condition on sparsity d as stated in (5.3) implies that as ρA , or τB ,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
43
or the step size parameter ζ increases, we need to increase the sample size in order
to guarantee computational convergence for the composite gradient descent algorithm
given the following lower bound:
ζ
λmax (A) _
, where
(8.2)
n = Ω dτ0 log m
λmin (A)2
(ᾱ` )2
(ρmax (s0 , A) + τB )2
τ0
.
λmin (A)
We illustrate the effect of the penalty and step size parameters in Section 8.2.
8.2. Corrected Lasso via GD versus Conic programming estimator
In the second experiment, both A and B are generated using the AR(1) model with
parameters ρA = 0.3, ρB ∗ = 0.3, and τB ∈ {0.3, 0.7}. We set m = 1024, d = 10 and
kβ ∗ k2 = 5. We then compare the performance of the corrected Lasso estimator (1.7)
using the composite gradient descent algorithmic with the Conic programming estimator, which is a convex program designed and implemented by authors of [3].
We consider three choices for the step size parameter for the composite gradient descent
algorithm: ζ1 = λmax (A) + 12 λmin (A), ζ2 = 32 λmax (A) and ζ3 = 2λmax (A). We
observe that the gradient descent algorithm consistently produces an output such that
its statistical error in `2 norm is lower than the best solution produced by the Conic
programming estimator,
√ when both methods are subject to optimal tuning after we fix
upon the radius R = d kβ ∗ k2 for (1.10) and (ω, λ) in (1.8) as follows. As illustrated
in our theory, one can think of the parameter λ in (1.7) and parameters µ, ω in (1.8)
satisfying
λ µ kβ ∗ k2 + ω,
q
where we set ω = 0.1D0 lognm , where the factor 0.1 is chosen without loss of generality, as we will sweep over f ∈ (0, 0.8] to run through a sufficiently large range of
values of the tuning parameters:
q
1/2
• For the corrected Lasso estimator, we set λ = f D00 τbB kβ ∗ k2 lognm + ω;
q
1/2
log m
• For the Conic programming estimator, we use µ = f D00 τbB
n . We set
λ = 1 in (1.8), which is independent of the Lasso penalty.
The factor f is chosen to reflect the fact that in practice, we do not know the exact value
of kβ ∗ k2 or kβ ∗ k1 , D0 or D00 , or other parameters related to the spectrum properties of
A, B; moreover, in practice, we wish to understand the whole-path behavior for both
estimators.
In Figures 3 and 4, we plot the relative error in `1 and `2 norm as n increases from
100 to 2500, while sweeping over penalty factor f ∈ [0.05, 0.8] for τB = 0.3 and
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
44
τB = 0.7 respectively. For both estimators, the relative `2 and `1 error versus the scaled
sample size n/(d log m) are also plotted. In these figures, green dashed lines are for
the corrected Lasso estimator via gradient descent algorithm, and blue dotted lines are
for the Conic programming estimator. These plots allow us to observe the behaviors of
the two estimators across a set of tuning parameters. Overall, we see that both methods
are able to achieve low relative error `p , p = 1, 2 norm when λ and µ are chosen from
a suitable range.
For the corrected Lasso estimator, we display results where the step size parameter ζ is
set to ζ2 = 23 λmax (A) and ζ3 = 2λmax (A) in the left and right column respectively. We
mention in passing that the algorithm starts to converge even when we set ζ = ζ1 =
λmax (A) + 21 λmin (A) as we observe quantitively similar behavior as the displayed
cases. For both estimators, we observe that we need a larger sample size n in case
τB = 0.7 in order to control the error at the same level as in case τB = 0.3.
In Figure 5, we plot the `2 and `1 error versus the penalty factor f ∈ [0.05, 0.8] for
sample size n ∈ {300, 600, 1200}. We plot results for τB = 0.3 and τB = 0.7
in the left and right column respectively. For these plots, we focus on cases when
n > dκ(A) log m, by choosing n ∈ {300, 600, 1200}; Otherwise, the gradient descent
algorithm does not yet reach the sample requirement (8.2) that guarantees computational convergence. In Figure 5, we observe that the Conic programming estimator is
relatively stable over the choices of µ once f ≥ 0.2. The composite gradient algorithm
favors smaller penalties such as f ∈ [0.05, 0.2], leading to smaller relative error in the
`1 and `2 norm, consistent with our theoretical predictions. These results also confirm
our theoretical prediction that the Lasso and Conic programming penalty parameters λ
and µ need to be adaptively chosen based on the noise level τB , because a larger than
necessary amount of penalty will cause larger relative error in both `1 and `2 norm.
8.3. Sensitivity to tuning parameters
∗
∗
∗
In the third experiment,
√ we change the `1 -ball radius R ∈ {R , 5R , 9R } in (1.10),
∗
∗
where R = kβ k2 d, while running through different penalties for the composite
gradient descent algorithm. In the left column in Figure 6, A and B are generated
using the AR(1) model with ρA = 0.3, ρB ∗ = 0.3 and τB = 0.7. In the right column,
we set τB = 0.3, while keeping other parameters invariant.
As predicted by our theory, a larger radius demands correspondingly larger penalty
to ensure consistent estimation using the composite gradient descent algorithm; this
in q
turn will increase the relative error when R is too large, for example, when R =
e
e notation hides parameters involving τB and κ(A). This is
Ω( lognm ), where the Ω(·)
observed in Figure 6. When n is sufficiently large relative to τB and κ(A), the optimal
`1 and
q `2 error become less sensitive with regard to the choice of R, so long as R =
n
e
e
O(
log m ), where O(·) hides parameters involving τB and κ(A), as shown in Figure 6.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
45
Rudelson and Zhou/Errors-in-variables
ρA = 0.3 ρB* = 0.3 τB = 0.3 ζ = ζ2
1.2
1.0
0.8
0.6
relative L2 error
0.2
0.2
4
6
8
10
12
2
4
6
8
10
n / d log(m)
n / d log(m)
ρA = 0.3 ρB* = 0.3 τB = 0.3 ζ = ζ2
ρA = 0.3 ρB* = 0.3 τB = 0.3 ζ = ζ3
12
0.4
0.6
0.8
1.0
1.2
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
0.2
0.2
0.4
0.6
0.8
1.0
1.2
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
relative L1 error
2
relative L1 error
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
0.4
0.6
0.8
1.0
1.2
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
0.4
relative L2 error
ρA = 0.3 ρB* = 0.3 τB = 0.3 ζ = ζ3
2
4
6
8
n / d log(m)
10
12
2
4
6
8
10
12
n / d log(m)
F IG 3. Plots of the relative `1 and `2 error βb − β ∗ / kβ ∗ k for the Conic programming estimator and the
corrected Lasso estimator obtained via running the composite gradient descent algorithm on (approximately)
recovering β ∗ . Set parameters d = 10 and m = 1024 while varying n. Generate A and B using the AR(1)
model with parameters ρA = 0.3, ρB ∗ = 0.3 and τB = 0.3. Set ζ ∈ {ζ1 , ζ2 , ζ3 }. We compare the
performance of the corrected Lasso and the Conic programming estimators over choices of λ and µ while
sweeping through f ∈ (0, 0.8]. In the top row, we plot the relative `2 error for the Conic programming
estimator (blue dotted lines) and the corrected Lasso (green dashed lines) via the composite gradient descent
algorithm with step size parameter set to be ζ2 = 32 λmax (A) and ζ3 = 2λmax (A); in the bottom row,
we plot the relative `1 error under the same settings. We note that the composite gradient descent algorithm
starts to converge even when we set the step size parameter to be ζ1 = λmax (A) + 21 λmin (A).
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
46
Rudelson and Zhou/Errors-in-variables
1.4
ρA = 0.3 ρB* = 0.3 τB = 0.7 ζ = ζ3
1.0
0.8
relative L2 error
0.4
0.2
4
6
8
10
12
2
4
6
8
10
n / d log(m)
ρA = 0.3 ρB* = 0.3 τB = 0.7 ζ = ζ2
ρA = 0.3 ρB* = 0.3 τB = 0.7 ζ = ζ3
12
1.4
n / d log(m)
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
1.0
0.8
relative L1 error
0.6
0.4
0.2
0.2
0.4
0.6
0.8
1.0
1.2
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
1.2
1.4
2
relative L1 error
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
0.6
0.8
0.2
0.4
0.6
relative L2 error
1.0
1.2
f = 0.05
f = 0.10
f = 0.20
f = 0.30
f = 0.40
f = 0.50
f = 0.60
f = 0.80
1.2
1.4
ρA = 0.3 ρB* = 0.3 τB = 0.7 ζ = ζ2
2
4
6
8
n / d log(m)
10
12
2
4
6
8
10
12
n / d log(m)
F IG 4. Plots of the relative `1 and `2 error βb − β ∗ / kβ ∗ k after running the Conic programming estimator and composite gradient descent algorithm on recovering β ∗ using the corrected Lasso objective function
with sparsity parameter d = 10 and m = 1024 while varying n. Both A and B are generated using the
AR(1) model with parameters ρA = 0.3, ρB ∗ = 0.3 and τB = 0.7. We compare the performance of
the corrected Lasso (green dashed lines) and the Conic programming estimators (blue dotted lines) over
choices of λ and µ while sweeping through f ∈ (0, 0.8]. For the composite gradient descent algorithm, we
choose ζ from {ζ1 , ζ2 , ζ3 }. In the top row, we plot the `2 error for the Conic and the corrected Lasso with
ζ2 = 23 λmax (A) and ζ3 = 2λmax (A), while in the bottom row, we plot the `1 error corresponding to the
two step size parameters.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
47
Rudelson and Zhou/Errors-in-variables
0.8
0.6
0.7
0.4
0.6
0.4
0.3
0.8
0.0
0.2
0.4
0.6
f
ρA = 0.3 ρB* = 0.3 τB = 0.3 n = 600
ρA = 0.3 ρB* = 0.3 τB = 0.7 n = 600
0.8
0.9
f
0.5
0.6
0.7
0.8
L2 GD
L2 Conic
L1 GD
L1 Conic
0.3
0.2
0.1
0.2
0.4
0.4
0.5
relative error
0.6
0.7
L2 GD
L2 Conic
L1 GD
L1 Conic
0.3
0.2
0.4
0.6
0.8
0.0
0.2
0.4
0.6
f
ρA = 0.3 ρB* = 0.3 τB = 0.3 n = 1200
ρA = 0.3 ρB* = 0.3 τB = 0.7 n = 1200
0.8
0.9
f
0.8
0.0
0.6
0.7
0.8
L2 GD
L2 Conic
L1 GD
L1 Conic
0.4
0.3
0.2
0.1
0.2
0.3
0.4
0.5
relative error
0.6
0.7
L2 GD
L2 Conic
L1 GD
L1 Conic
0.5
relative error
L2 GD
L2 Conic
L1 GD
L1 Conic
0.2
0.2
0.1
0.2
0.8
0.0
relative error
0.5
relative error
0.5
0.4
L2 GD
L2 Conic
L1 GD
L1 Conic
0.3
relative error
0.6
0.7
0.9
ρA = 0.3 ρB* = 0.3 τB = 0.7 n = 300
0.8
ρA = 0.3 ρB* = 0.3 τB = 0.3 n = 300
0.0
0.2
0.4
f
0.6
0.8
0.0
0.2
0.4
0.6
f
F IG 5. Plot of the relative error in `2 and `1 norm versus the penalty factor f ∈ (0, 0.8] as we change
the sample size n. Set m = 1024 and d = 10. Both A and B are generated using the AR(1) model
with parameters ρA = 0.3 and ρB ∗ = 0.3. We plot the relative error in `1 and `2 norm versus the penalty
parameter factor f ∈ (0, 0.8] for n = 300, 600, 1200 when ζ = 23 λmax (A). In the left column, τB = 0.3.
In the right column, we set τB = 0.7.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
0.8
48
Rudelson and Zhou/Errors-in-variables
B
* = 0.3
B
= 0.7 n = 600
ρA = 0.3 ρB * = 0.3 τB = 0.3 n = 600
0.8
0.4
0.6
0.8
0.2
0.4
0.6
ρA = 0.3 ρB * = 0.3 τB = 0.7 n = 1200
ρA = 0.3 ρB * = 0.3 τB = 0.3 n = 1200
0.8
1.0
f
1.0
f
0.6
0.8
R* L2
5R* L2
9R* L2
R* L1
5R* L1
9R* L1
0.2
0.2
0.4
0.6
relative error
0.8
R* L2
5R* L2
9R* L2
R* L1
5R* L1
9R* L1
0.4
0.2
0.4
0.6
0.8
0.2
0.4
0.6
ρA = 0.3 ρB * = 0.3 τB = 0.7 n = 2500
ρA = 0.3 ρB * = 0.3 τB = 0.3 n = 2500
0.8
1.0
f
1.0
f
0.8
R* L2
5R* L2
9R* L2
R* L1
5R* L1
9R* L1
0.4
0.2
0.2
0.4
0.6
relative error
0.8
R* L2
5R* L2
9R* L2
R* L1
5R* L1
9R* L1
0.6
relative error
0.6
relative error
0.2
0.2
0.2
relative error
R* L2
5R* L2
9R* L2
R* L1
5R* L1
9R* L1
0.4
0.6
R* L2
5R* L2
9R* L2
R* L1
5R* L1
9R* L1
0.4
relative error
0.8
1.0
= 0.3
1.0
A
0.2
0.4
0.6
f
0.8
0.2
0.4
0.6
f
F IG 6. Plot of the relative error in `2 and `1 norm versus the penalty factor f ∈ (0, 0.8] as we change
the radius R. Set m = 1024, d = 10 and
√ n ∈ {600, 1200, 2500}. We change the `1 -ball radius R ∈
{R∗ , 5R∗ , 9R∗ }, where R∗ = kβ ∗ k2 d, while running through different penalties for the composite
gradient descent algorithm. In the left column, A and B are generated using the AR(1) model with ρA =
0.3, ρB ∗ = 0.3 and τB = 0.7. In the right column, we set τB = 0.3, while keeping other parameters
invariant.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
0.8
49
Rudelson and Zhou/Errors-in-variables
−5
log(L2 error)
−5
1
2
3
6
12
25
−15
−15
−10
1
2
3
6
12
25
−10
log(L2 error)
0
ρA = 0.3 ρB * = 0.3 τB = 0.7 f = 0.2
0
ρA = 0.3 ρB * = 0.3 τB = 0.3 f = 0.2
0
20
40
60
80
100
0
Iteration count
(a)
20
40
60
80
Iteration count
(b)
b 2 ) versus
F IG 7. Plots of the statistical error log(kβ t − β ∗ k2 ), and the optimization error log(kβ t − βk
iteration number t, generated by running the composite gradient descent algorithm on the corrected Lasso
objective function. Each curve represents an average over 10 random trials, each with a different initialization point of β 0 . In Plots (a) and (b), B is generated using the AR(1) model with ρB ∗ = 0.3 and A is
generated using the AR(1) model with ρA = 0.3. We set τB = 0.3, 0.7 in Plot (a) and (b) respectively. We
set n = dρd log me, where we vary ρ ∈ {1, 2, 3, 6, 12, 25}.
8.4. Statistical and optimization error in Gradient Descent
In the last set of experiments, we study the statistical error and optimization error for
each iteration within the composite gradient descent algorithm. We observe a geometric
b 2.
convergence of the optimization error kβ t − βk
For each experiment, we repeat the following procedure 10 times: we start with a random initialization point β 0 and apply the composite gradient descent algorithm to comb we compute the optimization error log(kβ t − βk
b 2 ), which records
pute an estimate β;
t
b
b
the difference between β and β, where β is the final solution. In all simulations, we
b 2 ) between the iterate β t at time t versus the final soluplot the log error log(kβ t − βk
b
tion β, as well as the statistical error log(kβ t − β ∗ k2 ), which is the difference between
β t and β ∗ at time t. Each curve plots the results averaged over ten random instances.
In the first experiment, both A and B are generated using the AR(1) model with parameters ρA = 0.3 and ρB ∗ = 0.3. We set m = 1024, d = 10 and τB ∈ {0.3, 0.7}.
These results are shown in Figure 7. Within each plot, the red curves show the statistical error and the blue curves show the optimization error. We can see the optimization
b 2 decreases exponentially for each iteration, obeying a geometric conerror kβ t − βk
vergence. To illuminate the dependence of convergence rate on the sample size n, we
b 2 ) when n = dρd log me, where we vary
study the optimization error log(kβ t − βk
ρ ∈ {1, 2, 3, 6, 12, 25}. When n = d log m, the composite gradient algorithm fails to
converge since the sample size is too small for the RSC/RSM conditions to hold, resultimsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
100
50
Rudelson and Zhou/Errors-in-variables
ing in the oscillatory behavior of the algorithm for a constant step size. As the factor
ρ increases, the lower and upper RE curvature α and smoothness parameter α
e become
more concentrated around λmin (A) and λmax (A) respectively, and the tolerance parameter τ decreases at the rate of lognm . Hence we observe faster rates of convergence
for ρ = 25, 12, 6 compared to ρ = 2, 3. This is well aligned with our theoretical
τ0
prediction that once n = Ω(κ(A) λmin
(A) d log m) (cf. (8.2)), we expect to observe a
b 2.
geometric convergence of the computational error kβ t − βk
For the statistical error, we first observe the geometric contraction, and then the curves
flatten out after a certain number of iterations, confirming the claim that β t converges
to β ∗ only up to a neighborhood of radius defined through the statistical error bound
ε2stat ; that is, the geometric convergence is not guaranteed to an arbitrary precision, but
only to an accuracy related to statistical precision of the problem measured by `2 error:
kβb − β ∗ k22 =: ε2stat between the global optimizer βb and the true parameter β ∗ .
In the second experiment, A is generated from the Star-Block model, where we have
32 subgraphs and each subgraph has 16 edges; B is generated using the random graph
model with n log n edges and adjusted to have τB = 0.3. We set m = 1024, n =
2500 and d = 10. We then choose ρA ∈ {0.3, 0.5, 0.7, 0.9}. The results are shown
in Figure 8(b). As we increase ρA , we need larger sample size to control the statistical
error. Hence for a fixed n, the statistical error is bigger for ρA = 0.7, compared to cases
where ρA = 0.5 or ρA = 0.3, for which we have κ(A) = 42.06 and κ(A) = 10.2 (for
ρA = 0.3) respectively; Moreover, the rates of convergence are faster for the latter two
compared to ρA = 0.7, where κ(A) = 169.4. When ρA = 0.9, the composite gradient
descent algorithm fails to converge as ρ(A) is too large (hence not plotted here) with
respect to the sample size we fix upon. In Figure 8(a), we show results of A being
generated using the AR(1) model with four choices of ρA ∈ {0.3, 0.5, 0.7, 0.9} and B
being generated using the AR(1) model with ρB ∗ = 0.7 and τB = 0.3. We observe
quantitively similar behavior as in Figure 8(b).
9. Proof of Lemma 1
Proof of Lemma 1. Part I: Suppose that the Lower-RE condition holds for Γ :=
AT A. Let x ∈ Cone(s0 , k0 ). Then
√
kxk1 ≤ (1 + k0 ) kxT0 k1 ≤ (1 + k0 ) s0 kxT0 k2 .
Thus for x ∈ Cone(s0 , k0 ) ∩ S p−1 and τ (1 + k0 )2 s0 ≤ α/2, we have
kAxk2 = (xT AT Ax)1/2
≥
≥
≥
2
2
α kxk2 − τ kxk1
2
1/2
2
α kxk2 − τ (1 + k0 )2 s0 kxT0 k2
r
1/2
α
α − τ (1 + k0 )2 s0
≥
.
2
1/2
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
51
Rudelson and Zhou/Errors-in-variables
ρB * = 0.7 τB = 0.3 n = 2500
0
−5
log(L2 error)
−5
ρA= 0.3
ρA= 0.5
ρA= 0.7
ρA= 0.9
−15
−15
−10
ρA= 0.3
ρA= 0.5
ρA= 0.7
ρA= 0.9
−10
log(L2 error)
0
A: Star−Block; B: Random τB = 0.3 n = 2500
0
20
40
60
80
100
0
20
40
Iteration count
60
80
Iteration count
(a)
(b)
F IG 8. Plots of the statistical error log(kβ t − β ∗ k2 ), and the optimization error when we change the
topology. In the last experiment, we have m = 1024, d = 10 and n = 2500. In (a), A is generated using
the AR(1) model with four choices of ρA ∈ {0.3, 0.5, 0.7, 0.9} and B is generated using AR(1) model
with ρB ∗ = 0.7 and τB = 0.3. In (b), A follows the Star-Block model and B follows the random graph
model. We show four choices of ρA ∈ {0.3, 0.5, 0.7, 0.9}
Thus the RE(s0 , k0 , A) condition holds with
1
K(s0 , k0 , A)
kAxk2
:=
min
≥
x∈Cone(s0 ,k0 ) kxT0 k2
r
α
2
where we use the fact that for any J ∈ {1, . . . , p} such that |J| ≤ s0 , kxJ k2 ≤ kxT0 k2 .
We now show the other direction.
Part II. Assume that RE(4R2 , 2R − 1, A) holds for some integer R > 1. Assume that
for some R > 1
kxk1 ≤ R kxk2 .
Let (x∗i )pi=1 be non-increasing arrangement of (|xi |)pi=1 . Then
2 1/2
s
∞
X
X
kxk1
≤ R (x∗j )2 +
j
j=1
j=s+1
kxk1
1/2
1
2
2 1
≤ R kx∗J k2 + kxk1
≤ R kx∗J k2 + kxk1 √
s
s
where J := {1, . . . , s}. Choose s = 4R2 . Then
kxk1 ≤ R kx∗J k2 +
1
kxk1 .
2
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
100
Rudelson and Zhou/Errors-in-variables
52
Thus we have
kxk1
kx∗J c k1
≤ 2R kx∗J k2 ≤ 2R kx∗J k1
and hence
(2R − 1) kx∗J k1 .
≤
(9.1)
(9.2)
Then x ∈ Cone(4R2 , 2R − 1). Then for all x ∈ S p−1 such that kxk1 ≤ R kxk2 , we
have for k0 = 2R − 1 and s0 := 4R2 ,
2
xT Γx ≥
2
kxT0 k2
kxk2
2
≥√
=: α kxk2
K 2 (s0 , k0 , A)
s0 K 2 (s0 , k0 , A)
2
2
where we use the fact that (1 + k0 ) kxT0 k2 ≥ kxk2 by Lemma 33 with xT0 as defined
therein. Otherwise, suppose that kxk1 ≥ R kxk2 . Then for a given τ > 0,
2
2
α kxk2 − τ kxk1 ≤ ( √
1
2
− τ R2 ) kxk2 .
s0 K 2 (s0 , k0 , A)
(9.3)
Thus we have by the choice of τ as in (2.2) and (9.3)
2
xT Γx ≥ λmin (Γ) kxk2
≥ (√
1
2
− τ R2 ) kxk2
s0 K 2 (s0 , k0 , A)
2
2
≥ α kxk2 − τ kxk1 .
The Lemma thus holds.
10. Proof of Theorem 3
Throughout this proof, we assume that A0 ∩ B0 holds. First we note that it is sufficient to have (3.2) in order for (6.5) to hold. Condition (3.2) guarantees that for
V = 3eMA3 /2,
r(B) :=
tr(B)
kBk2
n
Vm log m
log
log m
n
3emMA3 log m
0 4 n
≥ 16c K
log
log m
2n
≥ 16c0 K 4
!
6emMA
1 4
n
log
= cK 2 2
4
ε MA log m
2 (n/ log m)
MA
1
6emMA
s0
3em
≥ c0 K 4 2 s0 log
= c0 K 4 2 log
(10.1)
ε
s0
ε
s0 ε
0
4
1
, and the last inequality holds given that k log(cm/k) on the
where ε = 2M1 A ≤ 128C
RHS of (10.1) is a monotonically increasing function of k,
s0
≤
4n
64C(ρmax (s0 , A) + τB )
and MA =
≥ 64C.
MA2 log m
λmin (A)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
53
Rudelson and Zhou/Errors-in-variables
Next we check that the choice of d as in (3.4) ensures that (6.9) holds for Dφ defined
there. Indeed, for c0 K 4 ≤ 1, we have
d ≤ CA (c0 K 4 ∧ 1)
φn
n
≤ CA (c0 Dφ ∧ 1)
.
log m
log m
b A := 1 (X T X −
By Lemma 15, we have on event A0 , the modified gram matrix Γ
n
b
tr(B)I
m ) satisfies the Lower RE conditions with α and τ as in (10.2). Theorem 3
follows
q from Theorem 16, so long as we can show that condition (6.6) holds for λ ≥
4ψ
log m
n ,
where the parameter ψ is as defined (3.6),
curvature α =
5
λmin (A) − α
3α
λmin (A) and tolerance τ =
=
.
8
s0
5s0
(10.2)
Combining (10.2) and (6.6), we need to show (6.10) holds. This is precisely the content
of Lemma 17. This is the end of the proof for Theorem 3
11. Proof of Theorem 4
For the set ConeJ (k0 ) as in (F.3),
κRE (d0 , k0 )
:=
min
∆T Ψ∆
min
J:|J|≤d0 ∆∈ConeJ (k0 )
2
k∆J k2
=
1
√
K(d0 , k0 , (1/ n)Z1 A1/2 )
2
.
Recall the following Theorem 27 from [38].
Theorem 27. ([38]) Set 0 < δ < 1, k0 > 0, and 0 < d0 < p. Let A1/2 be an m × m
matrix satisfying RE(d0 , 3k0 , A1/2 ) condition as in Definition 2.1. Set
d =
d0 + d0 max A1/2 ej
j
2
2
16K 2 (d0 , 3k0 , A1/2 )(3k0 )2 (3k0 + 1)
.
δ2
Let Ψ be an n × m matrix whose rows are independent isotropic ψ2 random vectors in
Rm with constant α. Suppose the sample size satisfies
2000dα4
60em
n≥
log
.
(11.1)
δ2
dδ
√
Then with probability at least √
1 − 2 exp(−δ 2 n/2000α4 ), RE(d0 , k0 , (1/ n)ΨA1/2 )
condition holds for matrix (1/ n)ΨA with
√
K(d0 , k0 , A1/2 )
0 < K(d0 , k0 , (1/ n)ΨA1/2 ) ≤
.
1−δ
(11.2)
Proof of Theorem 4. Suppose RE(2d0 , 3k0 , A1/2 ) holds. Then for d as defined
in (3.10) and n = Ω(dK 4 log(m/d)), we have with probability at least 1−2 exp(δ 2 n/2000K 4 ),
RE(2d0 , k0 , √1n Z1 A1/2 ) condition holds with
κRE (2d0 , k0 )
=
1
√
K(2d0 , k0 , (1/ n)Z1 A1/2 )
2
≥
1
2K(2d0 , k0 , A1/2 )
2
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
54
Rudelson and Zhou/Errors-in-variables
by Theorem 27.
The rest of the proof follows from [3] Theorem 1 and thus we only provide a sketch. In
more details, in view of the lemmas shown in Section 6, we need
−1/q
κq (d0 , k0 ) ≥ cd0
to hold for some constant c for Ψ := n1 X0T X0 . It is shown in Appendix C in [3] that
under the RE(2d0 , k0 , √1n Z1 A1/2 ) condition, for any d0 ≤ m/2 and 1 ≤ q ≤ 2,
κ1 (d0 , k0 ) ≥
cd−1
0 κRE (d0 , k0 )
κq (d0 , k0 ) ≥
−1/q
c(q)d0 κRE (2d0 , k0 ),
and
(11.3)
where c(q) > 0 depends on k0 and q. The theorem is thus proved following exactly the
same line of arguments as in the proof of Theorem 1 in [3] in view of the `q sensitivity
condition derived immediately above, in view of Lemmas 19, 20 and 21. Indeed, for
v := βb − β ∗ , we have by definition of `q sensitivity as in (6.13),
−1/q
c(q)d0
κRE (2d0 , k0 ) kvkq
≤
κq (d0 , k0 ) kvkq ≤
1
T
n X0 X0 v ∞
∗
≤
µ1 kβ k2 + µ2 kvk1 + ω
≤
µ1 kβ ∗ k2 + µ2 (2 + λ) kvS k1 + ω
≤
µ1 kβ ∗ k2 + µ2 (2 + λ)d0
≤
µ1 kβ ∗ k2 + µ2 (2 + λ)d0
1−1/q
kvS kq + ω
1−1/q
kvkq + ω.(11.4)
p
Thus we have for d0 = c0 n/ log m, where c0 is sufficiently small,
−1/q
d0
and hence
(c(q)κRE (2d0 , k0 ) − µ2 (2 + λ)d0 ) kvkq ≤ µ1 kβ ∗ k2 + ω
1/q
kvkq ≤ C(4D2 ρn K kβ ∗ k2 + 2D0 M ρn )d0
1/q
≤ 4CD2 ρn (K kβ ∗ k2 + M )d0
for some constant C = 1/ (c(q)κRE (2d0 , k0 ) − µ2 (2 + λ)d0 ) ≥ 1/ (2c(q)κRE (2d0 , k0 )),
where
p
1
µ2 (2 + λ)d0 = 2D2 Kρn ( + 1)(2 + λ)c0 n/ log m
λ
1
= 2c0 C0 D2 K 2 (2 + λ)( + 1)
λ
is sufficiently small and thus (3.11) holds. The prediction error bound follows exactly
the same line of arguments as in [3] which we omit here. See proof of Theorem 7 in
Section 6.5 for details.
12. Proof of Theorem 6
Throughout this proof, we assume that A0 ∩ B0 holds. The proof is also identical to
the proof of Theorem 3 up till (10.2), except that we replace the condition on d as in
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
55
the theorem statement by (4.3). Theorem 6 follows from Theorem 16, so long as we
as defined in (10.2), and
can show that condition (6.6) holds for α and τ = λmins(A)−α
0
q
log m
λ ≥ 2ψ
n , where the parameter ψ is as defined (6.3). Combining (10.2) and (6.6),
we need to show (6.10) holds. This is precisely the content of Lemma 18. This is the
end of the proof for Theorem 6.
13. Proof of Theorem 7
Throughout this proof, we assume that B0 ∩ B10 holds. The rest of the proof follows
that of Theorem 4, except for the last part. Let µ1 , µ2 , ω be as defined in Lemma 21.
p
1/2
1
), where µ = D00 Kρn τeB and d0 = c0 τB− n/ log m,
We have for µ2 := 2µ(1 + 2λ
µ2 (2 + λ)d0
1
+ 1)(2 + λ)c0 τB−
(13.1)
2λ
1
1
≤ 2c0 C0 D00 K 2 (2 + λ)(
+ 1) ≤ c(q)κRE (2d0 , k0 ),
2λ
2
=
1/2
2C0 D00 K 2 τeB (
1/2
which holds when c0 is sufficiently small, where τB− τeB
µ2 d0 ≤
≤ 1 by (6.18). Hence
c(q)κRE (2d0 , k0 )
.
2(2 + λ)
Thus for c0 sufficiently small, µ1 = 2µ, we have by (11.3), (13.1), (11.4) and (6.17),
−1/q 1
d0
(c(q)κRE (2d0 , k0 )) kvkq
2
−1/q
= d0 (c(q)κRE (2d0 , k0 ) − µ2 (2 + λ)d0 ) kvkq
1−1/q
≤ (κq (d0 , k0 ) − µ2 (2 + λ)d0
≤
) kvkq ≤ µ1 kβ ∗ k2 + ω
1/2
1/2
2D00 ρn K 2 ((τB + (3/2)C6 rm,m
) kβ ∗ k2 + M /K)
(13.2)
and thus (4.12) holds, following the proof in Theorem 4. The prediction error bound
follows exactly the same line of arguments as in [3], which we now include for the sake
completeness. Following (4.12), we have by (13.2),
kvk1
and hence µ2 kvk1
≤ C11 d0 (µ1 kβ ∗ k2 + ω), where C11 = 2/ (c(q)κRE (2d0 , k0 )) ,
≤ C11 µ2 d0 (µ1 kβ ∗ k2 + ω)
1
≤ C11
(c(q)κRE (2d0 , k0 )) (µ1 kβ ∗ k2 + ω)
2(2 + λ)
1
=
(µ1 kβ ∗ k2 + ω).
2+λ
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
56
Rudelson and Zhou/Errors-in-variables
Thus we have by (13.2), (6.18) and the bounds immediately above,
1
n
X(βb − β ∗ )
2
≤
2
1
T
n X0 X0 v ∞
C11 d0 (µ1 kβ ∗ k2 + ω) (µ1
kvk1
kβ ∗ k2 + µ2 kvk1 + 2ω)
1
) (µ1 kβ ∗ k2 + 2ω)
≤ C11 d0 (µ1 kβ ∗ k2 + ω)(1 +
2+λ
2
log m
M
1/2
0
0 2 4
∗
= C (D0 ) K d0
τeB kβ k2 +
n
K
log m
2
≤ C 00 (kBk2 + amax )K 2 d0
(2τB + 3C62 rm,m )K 2 kβ ∗ k2 + M2 ,
n
≤
where (D00 )2 ≤ 2 kBk2 + 2amax . The theorem is thus proved.
14. Proof of Theorem 9
Suppose that event A0 ∩ B0 holds. The condition on d in (5.3) implies that
4ζ
12λmax (A) _
, where
n > 512dτ0 log m
λmin (A)2
(ᾱ` )2
400C 2 $(s0 + 1)2
τ0
.
λmin (A)
(14.1)
(14.2)
To see this, note that the following holds by the first bound in (5.3):
ν` =
64dτ0 log m
λmin (A)
λmin (A)
α`
≤ 64dτ0 log m
=
≤
, (14.3)
n
256dτ0 log m ∗ 24κ(A)
96κ(A)
60
where α` = 85 λmin (A) by Lemma 8, and hence ᾱ` ≥
ᾱ2
α`2
≤ `,
5ζ
4ζ
59α`
60 .
Thus we have
where ζ ≥ αu > λmax (A) ≥ κ(A)ᾱ` .
Now, by definition of ν(d, m, n) and the second bound on n in (14.1),
2ν(d, m, n) = 128dτu (Ln )
:=
128dτ0 log m
(ᾱ` )2
≤
n
16ζ
Then
2% :=
4ν(d, m, n)
256dτu (Ln )
ᾱ`
=
≤
.
ᾱ`
ᾱ`
8ζ
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
57
Rudelson and Zhou/Errors-in-variables
That is, we actually need to have for 2% ≤ ᾱ8ζ`
ξ
1
ᾱ`
1−%
=
2τ (Ln )
+ 2% + 5 ᾱ`
1−κ
1−%
4ζ
4ζ − 2%
ᾱ`
1
= 2τ (Ln )
+ 2% + 5 ᾱ`
4ζ
4ζ − 2%
!
ᾱ`
40ζ
40ζ
4ζ + 2%
= 2τ (Ln ) ᾱ`
+
≤ 2τ (Ln ) 3 +
ᾱ`
ᾱ`
4ζ − 2%
≤
6τ (Ln ) +
80ζ
τ (Ln ),
ᾱ`
where we use the second bound in (14.1), and hence
ᾱ`
+ 2% ≤
4ζ
ᾱ`
(1 − κ)(1 − %) =
− 2% ≥
4ζ
3 ᾱ`
and
2 4ζ
1 ᾱ`
.
2 4ζ
Finally, putting all bounds in (2.11), we have 0 < κ < 1. Thus the conclusion of
Theorem 2 hold.
14.1. Proof of Corollary 10
Suppose that event A0 ∩ B0 holds. We first show that
δ4
in case δ 2 ≤ M δ̄ 2 .
4νε2stat + 4τ (Ln )2 64τ (Ln ) 4dε2stat + 2
λ
Recall that ξ ≥ 10τ` (Ln ) by definition of ξ in (2.12). The condition (5.2) on λ as stated
in Theorem 2 indicates that
√
√
160b0 dτ` (Ln )
λ≥
where R b0 d.
(14.4)
1−κ
We first show that for the choice of λ and R as in (14.4),
2
δ4
2
bt
∆
δ 2 + 64τ 4dε2stat + 2
≤
ᾱ`
λ
2
2
2
3 2 α` εstat
2
δ M ε2stat
≤
δ +
+ O
.
α`
4
α`
τ0 400b20
Then (5.4) holds.
For the second term on the RHS of (2.16), we have by (14.1),
n
≥
256dτ0 log m 8ζ
,
ᾱ`
ᾱ`
where
τ0
400C 2 $(s0 + 1)2
.
λmin (A)
(14.5)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
58
Rudelson and Zhou/Errors-in-variables
Thus
4ν`
=
256dτ0
log m
ᾱ`
≤ ᾱ`
n
8ζ
2
ᾱ` 2
α` ε2stat
4νε2stat ≤
εstat ≤
.
ᾱ`
4ζ
4ζ
and
Consider the choice of η̄ = δ 2 , where M δ̄ 2 ≥ η̄ = δ 2 ≥
cε2stat d log m
1−κ
n
=: δ̄ 2 .
Thus we have for (14.4),
2δ 2
λ
≤
and hence =
4δ 2
λ .
√
M cε2stat d
M cε2stat d log m
1−κ
√
=
<R
1−κ
n 160b0 dτ` (Ln )
160b0 τ0
Then for the last term on the RHS of (2.16), we have for τ` (Ln ) τ ,
=
=
≤
=
=
where δ 2 ≤
2
2δ 2
,R
λ
4
4
δ
δ (1 − κ)2
n
64τ 2 ≤
λ
400b20 τ0 d log m
cM ε2stat (1 − κ)2
δ2
1 − κ 400b20 τ0
cδ 2 M ε2stat (1 − κ)
τ0
400b20
2
cδ M ε2stat
O
τ0 400b20
4τ` (Ln )2
16τ` (Ln ) min
M cε2stat d log m
.
1−κ
n
Finally, suppose we fix
R
20M+
b0
p
r
6κ(A)
n
log m
in view of the upper bound d¯ (5.7). Then in order for
λ ≥ 16R
ξ
1−κ
to hold, we need to set
λ
≥
because of the following lower bound
640Rτ (Ln )κ(A),
ξ
1−κ
≥ 40τ (Ln )κ(A) as shown in (2.22).
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
59
Rudelson and Zhou/Errors-in-variables
Then (5.5) holds given that the last term on the RHS of (2.16) is now bounded by
2
4τ` (Ln )2
ᾱ`
2δ 2
,R
λ
2
60 2
64δ 4
59 α` 6402 R2 κ(A)2 τ` (Ln )
√ !2
20M+ 6
b0
=
2
16τ` (Ln ) min
ᾱ`
≤
60 2
δ4
59 α` 6400κ(A)τ0
=
2
60 12 δ 4 M+
60 3δ 4
2δ 4
1024
≤
≤ 2
.
2
2
59 α` 16b0 κ(A)τ0
59 4α` 400b0 kAk2
b0 α` kAk2
≤
Remark 14.1. First we obtain an upper bound on
59 5
60 8 λmin (A) ≤ ᾱ`
ξ
1−κ
ξ
1−κ
for ζ = αu
3λmax (A)
2
and
80ζ
τ (Ln )
ᾱ`
80ζ
τ (Ln )
≤ 6τ (Ln ) + 59 5
60 8 λmin (A)
≈ 200τ (Ln )κ(A).
≤ 6τ (Ln ) +
√
Now we obtain an upper bound using R ≤ b0 d for d ≤ d¯ as in (5.7),
r
√
ξ
log m
b0
p
R
≤ 200κ(A)τ b0 d ≤ 200κ(A)τ0
1−κ
n 20M+ 6κ(A)
r
λmin (A)
log m
b
p0
= 200κ(A)τ0
$(s0 + 1)
n 640C 6κ(A)
r
τ0 p
log m
√
≤ 10b0
κ(A)
n
M+ 6
r
p
125b0
log m
√ C$(s0 + 1) κ(A)
,
=
n
6
where we use (5.3) and the fact that Mτ0+ = 12.5C$(s0 + 1). We now discuss the
implications of this bound on the choice of λ in Section 5.1. We consider two cases.
• When τB = Ω(1). It is sufficient to have for kβ ∗ k2 ≤ b0 and τB 1,
λ
≥
16Cb0
r
_
p
M
log m
1/2
0
50 κ(A)$(s0 + 1)
D0 K(KτB +
)
b0
n
following the discussions in Section 4, where the first and the second term on
the RHS are at the same order except that the new lower bound involves the
condition number κ(A), while the original bound in Theorem 6 involves only
1/2
1/2
D00 = kBk2 + amax .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
60
Rudelson and Zhou/Errors-in-variables
+/2
• When τB = o(1) and M = Ω(τB K kβ ∗ k2 ). Now d satisfies (4.15) and hence
(√
)
r
√
n
c0 D00 KM
1
b0 d ≤ √
∧ b0 .
$(s0 + 1)
4 5M+ log m
Now combining this with the condition on d as in (5.3) implies that it is sufficient
to set R such that
!r
n
ξ
κ(A)τ
b0 ^ D00 KM
p
R
1−κ
M+
$(s
+
1)
log
m
κ(A)
0
!r
b0 ^ D00 KM
log m
= κ(A)$(s0 + 1) p
$(s0 + 1)
n
κ(A)
r
p
log m
=: Ū .
≈
b0 $(s0 + 1) κ(A) ∧ κ(A)D00 KM
n
+/2
Hence it is sufficient to have for ψ D00 K M + KτB kβ ∗ k2 as in (4.2),
r
≥
λ
Ū ∨ ψ
log m
.
n
15. Proof of Theorem 12
Clearly the condition on the stable rank of B guarantees that
n ≥ r(B) =
tr(B) kBk2
tr(B)
2
2
=
≥ kBkF / kBk2 ≥ log m.
2
kBk2
kBk2
Thus the conditions in Lemmas 11 and 5 hold. First notice that
( n1 X T X −
X0T X0 β ∗ + W T X0 β ∗ + X0T + W T
γ
b
=
1
n
b
tr(B)
∗
n Im )β
=
1
T
n (X0 X0
+ W T X0 + X0T W + W T W −
b
tr(B)
∗
n Im )β .
Thus
b ∗
γ
b − Γβ
∞
≤
=
≤
≤
γ
b−
1
n
1
n
1
n
1
n
X0T
X0T
X0T
∗
b
X T X − tr(B)I
m β
∞
∗
b
+ W − W W + X0T W − tr(B)I
m β
T
T
T
1
n
1
n(
∗
+W
∞
+
+ WT
∞
+
(Z T BZ −
∞
1
T
∗
n X0 W β
∞
tr(B)Im )β ∗ ∞ ) + n1 X0T W β ∗ ∞
∗
b
(W W − tr(B)I
m )β
T
∞
+
b
+ n1 tr(B)
− tr(B) kβ k∞ =: U1 + U2 + U3 + U4 .
By Lemma 11 we have on B4 for D0 :=
U1 =
1
n
X0T + W T
∞
=
1
n
√
1
1/2
τB + amax ,
1
A 2 Z1T + Z2T B 2
∞
≤ ρn M D0 ,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
61
Rudelson and Zhou/Errors-in-variables
and on event B5 for D00 :=
p
1/2
kBk2 + amax ,
(Z T BZ − tr(B)Im )β ∗ ∞ + n1 X0T W β ∗ ∞
kBkF √
∗
1/2
√ + τB amax ≤ Kρn kβ ∗ k2 τB1/2 D00 ,
≤ ρn K kβ k2
n
p
1/2
where recall kBkF ≤ tr(B) kBk2 . Denote by B0 := B4 ∩ B5 ∩ B6 . We have on
B0 and under (A1), by Lemmas 11 and 5 and D1 defined therein,
U2 + U3 =
b ∗
γ
b − Γβ
∞
1
n
≤ U1 + U2 + U3 + U4
1/2
≤ ρn M D0 + D00 τB Kρn kβ ∗ k2 +
1
n
b
tr(B)
− tr(B) kβ ∗ k∞
1/2
≤ D0 M ρn + D00 KτB kβ ∗ k2 ρn + D1 kβ ∗ k∞ rm,m
1
1/2
≤ D0 M ρn + D00 KτB kβ ∗ k2 ρn + 2D1 K √ ρn .
m
Finally, we have by the union bound, P (B0 ) ≥ 1 − 16/m3 . This is the end of the proof
of Theorem 12.
16. Conclusion
In this paper, we provide a unified analysis on the rates of convergence for both the corrected Lasso estimator (1.7) and the Conic programming estimator (1.8). As n increases
or as the measurement error metric τB decreases, we see performance gains over the
entire paths for both `1 and `2 error for both estimators as expected. When we focus
on the lowest `2 error along the paths as we vary the penalty factor f ∈ [0.05, 0.8], the
corrected Lasso via the composite gradient descent algorithm performs slightly better
than the Conic programming estimator as shown in Figure 5.
For the Lasso estimator, when we require that the stochastic error in the response
variable y as in (1.1a) does not approach 0 as quickly as the measurement error W in
(1.1b) does, then the sparsity constraint becomes essentially unchanged as τB → 0.
These tradeoffs are somehow different from the behavior of the Conic programming
estimator versus the Lasso estimator; however, we believe the differences are minor.
Eventually, as τB → 0, the relaxation on d as in (4.17) enables the Conic programming
estimator to achieve bounds which are essentially identical to the Dantzig Selector
when the design matrix X0 is a subgaussian random matrix satisfying the Restricted
Eigenvalue conditions; See for example [6, 4, 38].
When τB → 0 and M = Ω(τB+ K kβ ∗ k2 ), we set
r
log m
λ ≥ 2ψ
, where ψ := 4C0 D00 KM ,
n
(16.1)
so as to recover the regular lasso bounds in `q loss for q = 1, 2 in (4.5) in Theorem 6.
Moreover, suppose that tr(B) is given, then one can drop the second term in ψ as in
(4.2) involving kβ ∗ k2 entirely and hence recover the lasso bound as well.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
62
Finally, we note that the bounds corresponding to the Upper RE condition as stated in
Corollary 25, Theorem 26 and Lemma 15 are not needed for Theorems 3 and 6. They
are useful to ensure algorithmic convergence and to bound the optimization error for
the gradient descent-type of algorithms as considered in [1, 30], when one is interested
in approximately solving the nonconvex optimization function (1.7). Our Theorem 9
illustrates this result. Our theory in Theorem 9 predicts the dependencies of the computational and statistical rates of convergence for the corrected Lasso via gradient descent
algorithm on the condition number κ(A), the trace parameter τB and the radius R as
λ
R log m
(ρmax (s0 , A) + τB )2
Rξ
τ0 κ(A)
, where τ0
1−κ
n
λmin (A)
depends on τB , sparse and minimal eigenvalues of A. Therefore, we need to increase
the penalty when we increase the `1 -ball radius R in (1.10) in order to ensure algorithmic and statistical convergence as predicted in Theorem 9. This is well-aligned with the
observation in Figure 6. Our numerical results validate such algorithmic and statistical
convergence properties.
Acknowledgements
The authors thank the Editor, the Associate editor and two referees for their constructive comments. We thank Rob Kass, Po-Ling Loh, Seyoung Park, Kerby Shedden and
Martin Wainwright for helpful discussions. We thank Professor Alexander Belloni for
providing us the code implementing the Conic programming estimator.
Appendix A: Outline
We prove Theorem 2 in Section B. In Sections C, we present variations of the HansonWright inequality as recently derived in [37] (cf. Lemma 32). We prove Lemma 11
in Section C.2. In Sections H and I, we prove the technical lemmas for Theorems 3
and 4 respectively. In Section J, we prove the Lemmas needed for Proof of Theorem 7.
In order to prove Corollary 25, we need to first state some geometric analysis results
Section K. We prove Corollary 25 in Section L and Theorem 26 in Section M.
Appendix B: Proof of Theorem 2
Let us first define the following shorthand notation
bt
∆
=
b
β t − βb and δ t = φ(β t ) − φ(β).
The proof of the theorem requires two technical Lemmas 28 and 30. Both are stated
under assumption (B.1), which is stated in terms of a given tolerance η̄ > 0 and integer
T > 0 such that
b ≤ η̄, ∀t ≥ T,
φ(β t ) − φ(β)
(B.1)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
63
where the distance between β t and the global optimizer βb is measured in terms of the
b
objective function φ, namely, δ t = φ(β t ) − φ(β).
b t := β t − βb satisfies a certain
We first show Lemma 28, which ensures that the vector ∆
cone-type condition. The proof is omitted, as it is a shortened proof of Lemma 1 of [31].
Lemma 28. (Iterated Cone Bound) Under the conditions of Theorem 2, suppose there
exists a pair (η̄, T ) such that (B.1) holds. Then for any iteration t ≥ T , we have
√
√
2η̄
β t − βb ≤ 4 d β t − βb + 8 d βb − β ∗ + 2 · min
,R .
λ
2
1
2
We next state the following auxiliary result on the loss function. We use Lemma 29 in
the proof of Lemma 28 and Corollary 10.
√
Lemma 29. Denote by τ` (Ln ) := τ0 lognm and ν` = 64dτ` (Ln ). Let ¯stat = 8 dεstat ,
where εstat = βb − β ∗ and = 2·min 2η̄ , R . Under the assumptions of Lemma 28,
λ
2
b t := β t − βb and t > T ,
we have for ∆
b βt) ≥
T (β,
b ≥
φ(β t ) − φ(β)
α` − ν` b t 2
∆
− 2τ` (Ln )(¯
stat + )2 and
(B.2)
2
2
2
b ≥ α` − ν` ∆
b t − 2τ` (Ln )(¯
stat + )2 . (B.3)
T (β t , β)
2
2
Lemma 30. (Lemma 3 of Loh-Wainwright (2015)) Suppose the RSC and RSM conditions as stated in (2.8) and (2.9) hold with parameters (α` , τ` (Ln )) and (αu , τu (Ln ))
respectively. Under the conditions of Theorem 2, suppose there exists a pair (η̄, T ) such
that (B.1) holds. Then for any iteration t ≥ T , we have for 0 < κ < 1,
ξ
(¯
2 + 2 ) for
1 − κ stat
2η̄
,R ,
and = 2 · min
λ
b
φ(β t ) − φ(β)
≤
b +
κt−T (φ(β T ) − φ(β))
¯stat
:=
√
8 dεstat
where the quantities κ and ψ are as defined in Theorem 2 (cf. (2.11) and (2.12)).
Proof of Theorem 2. We are now ready to put together the final argument for
the theorem. First notice that (2.16) follows from (2.15) directlyin viewof (B.3) and
√
2
Lemma 28, where we set η̄ = δ 2 , ¯stat = 8 dεstat and = 2 min 2δλ , R .
Following (B.3), we have for ν` = 64dτ` (Ln ),
α` − ν` b t
∆
2
2
2
b + 2τ` (Ln )(¯
≤ φ(β t ) − φ(β)
stat + )2 ,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
64
Rudelson and Zhou/Errors-in-variables
and thus
bt
∆
2
≤
2
≤
≤
≤
2
b + 4 τ` (Ln )(¯
(φ(β t ) − φ(β)
stat + )2
ᾱ`
ᾱ`
2 2
δ + 2τ` (Ln )(2¯
2stat + 22 )
ᾱ`
2 2
δ + 2τ` (Ln )(128dε2stat + 22 )
ᾱ`
2 2
δ + 4ν` ε2stat + 4τ` (Ln )2 .
ᾱ`
(B.4)
The remainder of the proof follows an argument in [1]. We first prove the following
inequality:
b ≤ δ2 ,
φ(β t ) − φ(β)
∀t ≥ T ∗ (δ).
We divide the iterations t ≥ 0 into a series of epochs [T` , T`+1 ] and defend the tolerances η̄0 > η̄1 > . . . such that
b ≤ η̄` ,
φ(β t ) − φ(β)
∀t ≥ T` .
b to obtain
In the first iteration, we apply Lemma 30 with η̄0 := φ(β 0 ) − φ(β)
b ≤ κt (φ(β 0 ) − φ(β))
b +
φ(β t ) − φ(β)
ξ
(¯
2 + 4R2 )
1 − κ stat
for any iteration t ≥ 0.
Set
2ξ
η̄1 :=
(¯
2 + 4R2 ) and
1 − κ stat
log(2η̄0 /η̄1 )
T1 :=
.
log(1/κ)
Then we have for any iteration t ≥ T1
b ≤ η̄1 :=
φ(β t ) − φ(β)
4ξ
max ¯2stat , 4R2 .
1−κ
The same argument can be now be applied in a recursive manner. Suppose that for
some ` ≥ 1, we are given a pair (η̄` , T` ) such that
b ≤ η̄` , ∀t ≥ T` .
φ(β t ) − φ(β)
(B.5)
We now define
η̄`+1 :=
2ξ
(¯
2 + 2` ) and
1 − κ stat
T`+1 :=
log(2η̄` /η̄`+1 )
+ T` .
log(1/κ)
We can apply Lemma 30 to obtain for any iteration t ≥ T` and ε` := 2 min{ η̄λ` , R},
b ≤ κt−T` (φ(β T` ) − φ(β))
b +
φ(β t ) − φ(β)
ξ
(¯
2 + ε2` ),
1 − κ stat
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
65
which implies that for all t ≥ T`+1 ,
b ≤ η̄`+1 ≤
φ(β t ) − φ(β)
4ξ
max{¯
2stat , ε2` }
1−κ
by our choice of {η` , T` }`≥1 . Finally, we use the recursion
η̄`+1 ≤
log(2` η̄0 /η̄` )
log(1/κ)
(B.6)
η̄`+1
R
≤ 2` ∀` = 1, 2, . . . .
λ
4
(B.7)
4ξ
max(¯
2stat , ε2` ) and
1−κ
T` ≤ ` +
to establish the recursion that
η̄`+1 ≤
η̄`
2
4 `−1
and
ε`+1 :=
Taking these statements as given, we need to have
η̄` ≤ δ 2 .
It is sufficient to establish that
λR
≤ δ2 .
42`−1
Thus we find that the error drops below δ 2 after at most
`δ ≥ log log(Rλ/δ 2 )/ log(4) / log 2 + 1 = log log(Rλ/δ 2 )
epochs. Combining the above bound on `δ with the recursion (B.6)
T` ≤ ` +
log(2` η̄0 /η̄` )
,
log(1/κ)
we conclude that
b ≤ δ2
φ(β t ) − φ(β)
is guaranteed to hold for all iterations
log 2
log(η̄0 /δ 2 )
t > `δ 1 +
+
.
log(1/κ)
log(1/κ)
√
To√establish (B.7), we start with ` = 0 and establish that for ¯stat = 8 dstat =
o( d) = o(R)
η̄1
λ
:=
and thus ε1
:=
4ξ
16Rξ
R
max(¯
2stat , 4R2 ) =
R≤
(1 − κ)λ
(1 − κ)λ
4
η̄1
2 min{ , R} = R/2 ≤ ε0 = R.
λ
(B.8)
(B.9)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
66
Rudelson and Zhou/Errors-in-variables
Assume that ¯stat ≤ ε1 (otherwise, we are done at the first iteration). First, we obtain
for ` = 1,
2
4ξ
4ξ 2
4ξ
2η̄1
2
2
η̄2 ≤
max(¯
stat , ε1 ) =
ε =
1−κ
1−κ 1
1−κ
λ
2
16ξR η̄1
η̄1
16ξ η̄1
≤
≤ ,
≤
1 − κ λ2
1 − κ 4λ
4
η̄2
η̄1
R
and
≤
≤
,
λ
4λ
16
where in the last three steps, we use the fact that λ ≥
for ` = 1.
16Rξ
(1−κ)
and (B.8). Thus (B.6) holds
Now assume that (B.7) holds for d ≤ `. In the induction step, we again use the assumption that ε` := 2 η̄λ` ≥ ¯stat and (B.6) to obtain
η̄`+1
≤
≤
≤
4ξ
16ξ η̄`2
max(¯
2stat , ε2` ) =
1−κ
1 − κ λ2
16ξ
R η̄`
16Rξ 1 η̄`
=
1 − κ 42(`−1) λ
1 − κ λ 42(`−1)
η̄`
.
42(`−1)
Finally, by the induction assumption
R
η̄`
≤ 2`−1 ,
λ
4
we use the bound immediately above to obtain
η̄`+1
λ
≤
η̄`
(`−1)
2
4
1
R
1
R
≤ 2`−1 2(`−1) ≤ 2` .
λ
4
4
4
The rest of the proof follows from that of Corollary 10. This is the end of the proof for
Theorem 2.
It remains to prove Lemma 29.
Proof of Lemma 29. Using the RSC condition, we have for τ` (Ln ) := τ0 lognm
and ν` = 64dτ` (Ln ) ≤ α48` ,
b βt) ≥
T (β,
≥
≥
2
α` b t 2
bt
∆
− τ` (Ln ) ∆
2
2
1
α` b t 2
bt
∆
− τ` (Ln ) 2 ∗ 16d ∆
2
2
2
1
b t − 2τ` (Ln )(¯
ᾱ` ∆
stat + )2
2
2
(B.10)
2
2
+ 2(¯
stat + )2
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
67
Rudelson and Zhou/Errors-in-variables
and by Lemma 28, for any iteration t ≥ T ,
b t − βb
∆
1
√
≤ 4 d β t − βb
√
bt
4 d ∆
≤
2
√
+ 8 d βb − β ∗
+ 2 · min
2
2η̄
,R
λ
+ (¯
stat + ).
2
By convexity of function g, we have
b − h ∇g(β),
b β t − βb i ≥ 0.
g(β t ) − g(β)
(B.11)
Thus
b − h ∇φ(β),
b β t − βb i
φ(β t ) − φ(β)
b − h ∇Ln (β),
b β t − βb i + λ(g(β t ) − g(β)
b − h ∇g(β),
b β t − βb i ).
= Ln (β t ) − Ln (β)
b we have for all feasible β t ∈ Ω
Moreover, by the first order optimality condition for β,
b β t − βb i ≥ 0,
h ∇φ(β),
and thus
b ≥
φ(β t ) − φ(β)
b − h ∇Ln (β),
b β t − βb i = T (β t , β),
b
Ln (β t ) − Ln (β)
where similar to (B.10), we have
b ≥
T (β t , β)
≥
=
bt
α1 ∆
2
2
bt
− τ` (Ln ) ∆
bt
(α1 − 32dτ` (Ln )) ∆
1
bt
ᾱ` ∆
2
2
2
2
2
2
1
− 2τ` (Ln )(¯
stat + )2
− 2τ` (Ln )(¯
stat + )2 ,
and by Lemma 28,
bt
∆
2
≤
1
bt
32d ∆
bt
≤ 32d ∆
2
√
2η̄
+ 2 8 dεstat + 2 · min
,R
λ
2
2
2
2
+ 2(¯
stat + )2 .
Appendix C: Some auxiliary results
We first need to state the following form of the Hanson-Wright inequality as recently
derived in Rudelson and Vershynin [37], and an auxiliary result in Lemma 32 which
may be of independent interests.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
68
Rudelson and Zhou/Errors-in-variables
Theorem 31. Let X = (X1 , . . . , Xm ) ∈ Rm be a random vector with independent
components Xi which satisfy E (Xi ) = 0 and kXi kψ2 ≤ K. Let A be an m × m
matrix. Then, for every t > 0,
"
!#
t2
t
T
T
P X AX − E X AX > t ≤ 2 exp −c min
.
2 ,
K 4 kAkF K 2 kAk2
We note that following the proof of Theorem 31, it is clear that the following holds: Let
X = (X1 , . . . , Xm ) ∈ Rm be a random vector as defined in Theorem 31. Let Y, Y 0 be
independent copies of X. Let A be an m × m matrix. Then, for every t > 0,
!#
"
t
t2
T
0
.
(C.1)
P Y AY > t ≤ 2 exp −c min
2 ,
K 4 kAkF K 2 kAk2
We next need to state Lemma 32, which we prove in Section N.
Lemma 32. Let u, w ∈ S n−1 . Let A 0 be an m × m symmetric positive definite
matrix. Let Z be an n × m random matrix with independent entries Zij satisfying
EZij = 0 and kZij kψ2 ≤ K. Let Z1 , Z2 be independent copies of Z. Then for every
t > 0,
!!
t2
t
T
1/2 T
P u Z1 A Z2 w > t
≤ 2 exp −c min
,
,
K 4 tr(A) K 2 kAk1/2
2
!!
t
t2
T
T
T
T
,
P u ZAZ w − Eu ZAZ w > t ≤ 2 exp −c min
2 ,
K 4 kAkF K 2 kAk2
where c is the same constant as defined in Theorem 31.
C.1. Proof of Lemma 5
First we write
XX T − tr(A)In
=
=
=
ˇ
Thus we have for tr(B)
:=
Z1 A1/2 + B 1/2 Z2 ) Z1 A1/2 + B 1/2 Z2
1/2
Z1 A + B Z2 ) Z2T B 1/2 + A1/2 Z1T
Z1 A1/2 Z2T B 1/2 + B 1/2 Z2 Z2T B 1/2
+B 1/2 Z2 A1/2 Z1T + Z1 AZ1T − tr(A)In .
1
m
1/2
T
− tr(A)In
− tr(A)In
2
kXkF − ntr(A) ,
2
1
− tr(B)) := mn
kXkF − ntr(A) − mtr(B)
1
(tr(XX T ) − ntr(A) − mtr(B))
mn
2
tr(B 1/2 Z2 Z2T B 1/2 ) tr(B)
tr(Z1 A1/2 Z2T B 1/2 ) +
−
mn
mn
n
1 ˇ
n (tr(B)
=
=
+
tr(Z1 AZ1T ) tr(A)
−
.
mn
m
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
69
Rudelson and Zhou/Errors-in-variables
By constructing a new matrix An = In ⊗ A, which is block diagonal with n identical
submatrices A along
√ its diagonal, we prove the following large deviation bound: for
t1 = C0 K 2 kAkF n log m and n > log m,
T
P tr(Z1 AZ1T ) − ntr(A) ≥ t1 = P vec { Z1 } (I ⊗ A)vec { Z1 } − ntr(A) ≥ t1
!!
t21
t1
≤ exp −c min
2 ,
K 4 kAn kF K 2 kAn k2
!!
√
√
(C0 K 2 n log m kAkF )2 C0 K 2 n log m kAkF
≤ 2 exp −c min
,
2
K 2 kAk2
nK 4 kAkF
≤ 2 exp (−4 log m) ,
where the first inequality holds by Theorem 31 and the second inequality holds given
2
2
that kAn kF = n kAkF and kAn k2 = kAk2 .
Similarly, by constructing a new matrix Bm = Im ⊗ B, which is block diagonal with
m identical submatrices B along
√ its diagonal, we prove the following large deviation
bound: for t2 = C0 K 2 kBkF m log m and m ≥ 2,
T
P tr(Z2T BZ2 ) − mtr(B) ≥ t2 = P vec { Z2 } (Im ⊗ B)vec { Z2 } − mtr(B) ≥ t2
!!
t22
t2
≤ exp −c min
2 ,
K 4 m kBkF K 2 kBk2
!!
√
√
(C0 K 2 m log m kBkF )2 C0 K 2 m log m kBkF
≤ 2 exp −c min
,
2
K 2 kBk2
K 4 m kBkF
≤ 2 exp (−4 log m) .
p
Finally, we have by (C.1) for t0 = C0 K 2 tr(A)tr(B) log m,
T
P vec { Z1 } B1/2 ⊗ A1/2 vec { Z2 } > t0
≤ 2 exp −c min
=
2 exp −c min
t20
2 ,
K 4 B 1/2 ⊗ A1/2 F K 2
p
(C0 tr(A)tr(B) log m)2
,
tr(A)tr(B)
!!
t0
B 1/2 ⊗ A1/2 2
!!
p
C0 tr(A)tr(B) log m
1/2
kBk2
1/2
kAk2
≤ 2 exp(−4 log m),
where we use the fact that r(A)r(B) ≥ log m, B 1/2 ⊗ A1/2
and
B 1/2 ⊗ A1/2
2
=
1/2
2
= kBk2
1/2
kAk2
tr((B 1/2 ⊗ A1/2 )(B 1/2 ⊗ A1/2 )) = tr(B ⊗ A) = tr(A)tr(B).
F
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
70
Rudelson and Zhou/Errors-in-variables
Thus we have with probability 1 − 6/m4 ,
1
n
1
ˇ
tr(XX T ) − f tr(A) − mtr(B)
tr(B)
− tr(B) = mn
2
T
vec { Z1 } (B1/2 ⊗ A1/2 )vec { Z2 }
≤
mn
tr(Z1 AZ1T ) tr(A)
tr(Z2T BZ2 ) tr(B)
+
+
−
−
mn
n
mn
m
√
√
kBkF
kAkF
1
log m
2
√
≤
(2t0 + t1 + t2 ) = √
C0 K
+ 2 τA τB + √
mn
mn
m
n
√
log m 2
≤ 2C0 √
K D1 =: D1 rm,m ,
mn
where recall rm,m = 2C0 K 2
√
√log m ,
mn
kAkF
√
m
D1 =
+
kBkF
√
n
, and
√
kAk
kBk
2 τA τB ≤ τA + τB ≤ √ F + √ F .
m
n
To see this, recall
mτA
=
nτB
=
m
X
i=1
n
X
λi (A) ≤
λi (B) ≤
√
m(
√
n(
i=1
m
X
i=1
n
X
λ2i (A))1/2 =
λ2i (B))1/2 =
√
√
m kAkF
and
(C.2)
n kBkF
i=1
where λi (A), i = 1, . . . , m and λi (B), i = 1, . . . , n denote the eigenvalues of positive
semidefinite covariance matrices A and B respectively.
Denote by B6 the following event
1
ˇ
n tr(B) − tr(B) ≤ D1 rm,m .
b
ˇ
b
Clearly tr(B)
:= (tr(B))
=
+ by definition (1.5). As a consequence, on B6 , tr(B)
ˇ
tr(B)
> 0 when τB > D1 rm,m ; hence
1
n
b
tr(B)
− tr(B) =
1
n
ˇ
tr(B)
− tr(B) ≤ D1 rm,m .
ˇ
Otherwise, it is possible that tr(B)
< 0. However, suppose we set
τbB :=
1 b
n tr(B)
:=
1 ˇ
n (tr(B)
∨ 0),
then we can also guarantee that
|b
τB − τB | = |τB | ≤ D1 rm,m in case τB ≤ D1 rm,m .
The lemma is thus proved.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
71
Rudelson and Zhou/Errors-in-variables
C.2. Proof of Lemma 11
Following Lemma 32, we have for all t > 0, B 0 being an n × n symmetric positive
definite matrix, and v, w ∈ Rm
"
!#
t2
t
T T 1/2
,
P v Z1 B Z2 w > t ≤ 2 exp −c min
(C.3)
K 4 tr(B) K 2 kBk1/2
2
and
T
T
T
t2
t
, 2
2
K 4 kBkF K kBk2
T
P v Z BZw − Ev Z BZw > t ≤ 2 exp −c min
!!
Proof of Lemma 11. Let e1 , . . . , em ∈ Rm be the canonical basis spanning
R . Let x1 , . . . , xm , x01 , . . . , x0m ∈ Rn be the column vectors Z1 , Z2 respectively. Let
1/2
Y ∼ eT1 Z0T . Let wi = AA1/2 eei for all i. Clearly the condition on the stable rank of
k
ik
2
B guarantees that
m
tr(B) kBk2
tr(B)
2
2
=
≥ kBkF / kBk2 ≥ log m.
2
kBk2
kBk2
p
By (C.1), we obtain for t0 = C0 M K tr(B) log m
P ∃j, T B 1/2 Z2 ej > t0 =
p
1
M T T 1/2
2
P ∃j,
e Z B Z2 ej > C0 M K log mtr(B)
K 1 0
p
1
≤ exp(log m)P Y T B 1/2 x0j > C0 K 2 log mtr(B) 2 ≤ 2/m3
n ≥ r(B) =
where the last inequality holds by the union bound, given that
√
for all j and t = C0 K 2 log mtr(B)1/2 ,
P
Y
T
B 1/2 x0j
>t
≤
2 exp −c min
tr(B)
kBk2
≥ log m; Similarly,
t2
t
,
4
K tr(B) K 2 kBk1/2
!!
,
2
≤
≤
2 exp −c min
C02
log m,
C0 log1/2 m
!!
p
tr(B)
1/2
kBk2
2
2 exp −c min(C0 , C0 ) log m ≤ 2 exp (−4 log m) .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
.(C.4)
72
Rudelson and Zhou/Errors-in-variables
√
Let v, w ∈ S m−1 . Thus we have by Lemma 32, for t0 = C0 M K n log m, τ =
1/2
√
A
e
C0 K 2 n log m, wj = A1/2 e j and n ≥ log m,
k
jk
2
p
M T
T
P ∃j, Z1 wj > t0 ≤ P ∃j,
Y Z1 wj > C0 M K n log m
K
p
≤ mP Y T Z1 wj > C0 K 2 n log m
2
τ
τ
,
=: V
= exp(log m)P eT1 Z0T Z1 wj > τ ≤ 2 exp −c min
nK 4 K 2
where
≤
V
≤
≤
√
√
(C0 K 2 n log m)2 C0 K 2 n log m
2 exp −c min
+ log m
nK 4
K2
√
2m exp −c min C02 log m, C0 log1/2 m n
2m exp −c min(C02 , C0 ) log m ≤ 2 exp (−3 log m) .
Therefore we have with probability at least 1 − 4/m3 ,
1
Z2T B 2
∞
1
A 2 Z1T
∞
:=
:=
≤
≤
max h T B 1/2 Z2 , ej i ≤ t0 = C0 M K
p
j=1,...,m
tr(B) log m
max h A1/2 ej , Z1T i
j=1,...,m
max h wj , Z1T i
p
1/2
n log m.
a1/2
max t0 = amax C0 M K
max
j=1,...,m
A1/2 ej
2 j=1,...,m
The “moreover” part follows exactly the same arguments as above. Denote by β̄ ∗ :=
β ∗ / kβ ∗ k2 ∈ E ∩ S m−1 and wi := A1/2 ei / A1/2 ei 2 . By (C.3)
p
P ∃i, h wi , Z1T B 1/2 Z2 β̄ ∗ i ≥ C0 K 2 log mtr(B)1/2
≤
m
X
p
P h wi , Z1T B 1/2 Z2 β̄ ∗ i ≥ C0 K 2 log mtr(B)
i=1
≤ 2 exp −c min C02 log m, C0 log m + log m ≤ 2/m3 .
√
√
Now for t = C0 K 2 log m kBkF and kBkF / kBk2 ≥ log m,
p
P ∃ei : h ei , (Z T BZ − tr(B)Im )β̄ ∗ i ≥ C0 K 2 log m kBkF
"
!#
t
t2
≤ 2m exp −c min
≤ 2/m3 .
2 ,
K 4 kBkF K 2 kBk2
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
73
Rudelson and Zhou/Errors-in-variables
By the two inequalities immediately above, we have with probability at least 1 − 4/m3 ,
= A1/2 Z1T B 1/2 Z2 β ∗
∞
sup h wi , Z1T B 1/2 Z2 β̄ ∗ i
≤ kβ ∗ k2 max A1/2 ei
ei
2
wi
p
p
2
∗
1/2
≤ C0 K kβ k2 log mamax tr(B)
X0T W β ∗
∞
and
(Z T BZ − tr(B)Im )β ∗ ∞ = (Z T BZ − tr(B)Im )β̄ ∗
∗
T
∗
= kβ k2 sup h ei , (Z BZ − tr(B)Im )β̄ i
ei
p
2
∗
≤ C0 K kβ k2 log m kBkF .
∞
kβ ∗ k2
The last two bounds follow exactly the same arguments as above, except that we replace
β ∗ with ej , j = 1, . . . , m and apply the union bounds to m2 instead of m events, and
thus P (B10 ) ≥ 1 − 4/m2 .
Appendix D: Proof of Corollary 13
Now following (6.1), we have on event B0 ,
3
1
b ∗
γ
b − Γβ
≤ ρn
D2 + D2 √
K kβ ∗ k2 + D0 M
4
m
∞
where 2D1 ≤ 2 kAk2 + 2 kBk2 = D2 , and for (D00 )2 ≤ 2 kBk2 + 2amax ,
q
2(kBk2 + amax ) ≤ 2(amax + kBk2 ) = D2 ,
D0 ≤ D00 ≤
1/2
and D00 τB
1/2
≤ (kBk2
1/2
+ a1/2
max )τB
3
1
≤ τB + (kBk2 + amax ) ≤ D2
2
4
1/2
given that under (A1) : τA = 1, kAk2 ≥ amax ≥ amax ≥ 1. Hence the lemma holds
for m ≥ 16 and ψ = C0 D2 K (K kβ ∗ k2 + M ).
Appendix E: Proof of Corollary 14
1/2
1/2
Suppose
that event B0 holds. Recall D00 = kBk2 + amax . Denote by ρn :=
q
kAk
kBk
1/2
C0 K lognm . By (6.1) and the fact that 2D1 := 2( √mF + √nF ) ≤ 2(kAk2 +
√
√
1/2
kBk2 )( τA + τB ) ≤ Doracle D00 ,
b ∗
γ
b − Γβ
∞
1
1/2
≤ D00 KτB kβ ∗ k2 ρn + 2D1 K √ kβ ∗ k∞ ρn + D0 M ρn
m
D
oracle
1/2
0
∗
≤ D0 K kβ k2 ρn τB + √
+ D0 M ρn
m
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
74
Rudelson and Zhou/Errors-in-variables
The corollary is thus proved.
Appendix F: Proof of Lemma 15
In view of Remark F.1, Condition (6.5) implies that (7.5) in Theorem 26 holds
for k = s0 and ε = 2M1 A . Now, by Theorem 26, we have ∀u, v ∈ E ∩ S m−1 ,
under (A1) and (A3), condition (7.1) holds under event A0 , and so long as mn ≥
4096C02 D22 K 4 log m/λmin (A)2 ,
r
log m
T
2
=: δ with
u ∆v ≤ 8C$(s0 )ε + 2C0 D2 K
mn
λmin (A) λmin (A)
3
1
δ ≤
+
=
λmin (A) ≤ ,
16
32
32
8
which holds for all
1
1
1 λmin (A)
:=
≤
2 64C$(s0 )
2MA
128C
with P (A0 ) ≥ 1 − 4 exp −c2 ε2 Ktr(B)
− 2 exp −c2 ε2 Kn4 − 6/m3 .
4 kBk
ε≤
2
Hence, by Corollary 25, ∀θ ∈ Rm ,
2
2
b A θ ≥ α kθk − τ kθk
θT Γ
2
1
where α = 85 λmin (A) and α
e=
2
2
bA θ ≤ α
and θT Γ
e kθk2 + τ kθk1 ,
11
8 λmax (A)
and τ =
3 λmin (A)
.
8
s0
Now for s0 ≥ 32 as defined in (2.6), we have
and
s0
≤
s0 + 1
≥
λ2min (A)
n
log m 1024C 2 $(s0 )2
n
λ2min (A)
log m 1024C 2 $2 (s0 + 1)
(F.1)
(F.2)
given that τB + ρmax (s0 + 1, A) = O(λmax (A)) in view of (2.5) and (A3). Thus
384C 2 $(s0 )2 log m
≤τ
λmin (A)
n
=
≤
≤
3 λmin (A)
8
s0
33
3 λmin (A)
32(s0 + 1) 8
s0
396C 2 $2 (s0 + 1) log m
.
λmin (A)
n
The lemma is thus proved in view of Remark F.1.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
75
Rudelson and Zhou/Errors-in-variables
Remark F.1. Clearly the condition on tr(B)/ kBk2 as stated in Lemma 15 ensures
,
that we have for ε = 2M1 A and s0 M 2 4n
log m
A
ε2
tr(B)
K 4 kBk2
and hence
2 tr(B)
exp −c2 ε 4
≤
K kBk2
3em
ε2 0 4 s0
c
log
K
K4
ε2
s0 ε
6emM
A
≥ c0 s0 log
,
s0
≥
6emMA
exp −c c2 s0 log
s0
4n
3eMA3 m log m
exp −c3 2
log
.
MA log m
2n
0
F.1. Comparing the two type of RE conditions in Theorems 3 and 4
We define Cone(d0 , k0 ), where 0 < d0 < m and k0 is a positive number, as the set of
vectors in Rm which satisfy the following cone constraint:
Cone(d0 , k0 ) = {x ∈ Rm | ∃I ∈ {1, . . . , m}, |I| = d0 s.t. kxI c k1 ≤ k0 kxI k1 } .
For each vector x ∈ Rm , let T0 denote the locations of the d0 largest coefficients of x
in absolute values. The following elementary estimate [38] will be used in conjunction
with the RE condition.
Lemma 33. For each vector x ∈ Cone(d0 , k0 ), let T0 denotes the locations of the d0
largest coefficients of x in absolute values. Then
kxk
kxT0 k2 ≥ √ 2 .
1 + k0
(F.3)
Lemma 34. Suppose all conditions
in Lemma 15 hold. Let k0 := 1 + λ. Suppose that
d0 = o s0 /64(1 + 3λ/4)2 . Now suppose that
τ (1 + 3k0 )2 2d0 = 2τ (4 + 3λ)2 d0 ≤ α/2.
b A ) condition holds on Γ
b A in the sense that
Then on event A0 , we have RE2 (2d0 , 3k0 , Γ
min
bA x
xT Γ
2
x∈Cone(2d0 ,3k0 ) kxT k
0 2
≥
α
.
2
(F.4)
Under (A2) and (A3), we could set d0 such that for some large enough constant CA ,
2
n
λmin (A)
n
d0 ≤
=O
(F.5)
CA κ(A)2 log m
$2 (s0 + 1) log m
where κ(A) :=
λmax (A)
λmin (A) ,
so that d0 = O(s0 ) and (F.4) holds.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
76
Rudelson and Zhou/Errors-in-variables
Proof. Now following the proof Lemma 1, Part I. We have on A0 , the Lower-RE condition holds for ΓA . Thus for x ∈ Cone(2d0 , 3k0 ) ∩ S m−1 and τ (1 + 3k0 )2 2d0 ≤ α/2,
2
2
2
kxk1 ≤ (1 + 3k0 )2 kxT0 k1 ≤ (1 + 3k0 )2 2d0 kxT0 k2 .
Thus
bA x
xT Γ
≥
≥
≥
2
2
α kxk2 − τ kxk1
2
2
α kxk2 − τ (1 + 3k0 )2 2d0 kxT0 k2
α
2
2
α − τ (1 + 3k0 )2 2d0 kxT0 k2 ≥ kxT0 k2 .
2
Thus (F.4) holds. Now (F.5) follows from (F.1), which holds by definition of s0 as in
(2.6), where s0 is tightly bounded in the sense that both (F.1) and (F.2) need to hold.
Remark F.2. We note that (F.4) can be understood topbe the RE(2d0 , 3k0 ) condition
b A . In view of Lemma 15, it is clear that for d0 n/ log m, it holds that
on Γ
4d0 (4 + 3λ)2 = o(s0 )
given that τ s0 = O(α) on event A0 ; indeed, we have by Lemma 15 the Lower-RE
b A := AT A − tr(B)I
b
condition holds for Γ
m , with α, τ > 0 such that
curvature α =
3 λmin (A)
5
λmin (A) and tolerance τ :=
,
8
8
s0
where recall s0 ≥ 32 is as defined in (2.6); moreover, we replaced the parameter
0 ,A)+τB
MA ρmaxλ(s
with κ(A) in view of (2.5) and (A3).
min (A)
Appendix G: Proof of Theorem 16
Denote by β = β ∗ . Let S := supp β, d = |S| and
υ = βb − β,
where βb is as defined in (1.7).
We first show Lemma 35, followed by the proof of Theorem 16.
Lemma 35. [4, 30] Suppose that (6.7) holds. Suppose that there exists a parameter ψ
such that
r
r
√
ψ log m
log m
dτ ≤
and λ ≥ 4ψ
,
b0
n
n
where b0 , λ are as defined in (1.7). Then
kυS c k1 ≤ 3 kυS k1 .
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
77
Rudelson and Zhou/Errors-in-variables
b we have
Proof. By the optimality of β,
λ kβk1 − λ βb
≥
1
=
=
Hence, we have for λ ≥ 4ψ
1 b
υ Γυ
2
≤
≤
q
1 bb b
β Γβ −
2
1 b
υ Γυ +
2
1 b
υ Γυ −
2
1 b
β Γβ − h γ
b, v i
2
b i − h υ, γ
h υ, Γβ
bi
b i.
h υ, γ
b − Γβ
log m
n ,
b i + λ kβk − βb
h υ, γ
b − Γβ
1
1
b
kυk1 .
b − Γβ
+ γ
λ kβk1 − βb
(G.1)
∞
1
Hence
b
υ Γυ
≤ λ 2 kβk1 − 2 βb
≤ λ 2 kβk1 − 2 βb
r
+ 2ψ
1
log m
kυk1
n
(G.2)
1
kυk1
2
+
1
1
≤ λ (5 kυS k1 − 3 kυS c k1 ) ,
2
(G.3)
where by the triangle inequality, and βS c = 0, we have
2 kβk1 − 2 βb
+
1
1
kυk1
2
=
≤
≤
1
1
kυS k1 + kυS c k1
2
2
1
1
1
2 kυS k1 − 2 kυS c k1 + kυS k1 + kυS c k1
2
2
1
(5 kυS k1 − 3 kυS c k1 ) .
(G.4)
2
2 kβS k1 − 2 βbS
− 2 kυS c k1 +
We now give a lower bound on the LHS of (G.1), applying the lower-RE condition as
in Definition 2.2,
b
υ T Γυ
Tb
and hence − υ Γυ
2
2
2
≥ α kυk2 − τ kυk1 ≥ −τ kυk1
√
2
≤ kυk1 τ ≤ kυk1 2b0 dτ
r
r
ψ log m
log m
≤ kυk1 2b0
= kυk1 2ψ
b0
n
n
1
≤
λ(kυS k1 + kυS c k1 ),
2
where we use the assumption that
r
√
ψ log m
dτ ≤
and kυk1 ≤ βb
b0
n
1
(G.5)
√
+ kβk1 ≤ 2b0 d,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
78
Rudelson and Zhou/Errors-in-variables
which holds by the triangle
inequality and the fact that both βb and β have `1 norm
√
being bounded by b0 d. Hence by (G.3) and (G.5)
0
≤
≤
≤
b + 5 λ kυS k − 3 λ kυS c k
−υ Γυ
1
1
2
2
1
3
1
5
λ kυS k1 + λ kυS c k1 + λ kυS k1 − λ kυS c k1
2
2
2
2
3λ kυS k1 − λ kυS c k1 .
(G.6)
(G.7)
Thus we have
kυS c k1 ≤ 3 kυS k1 ,
and the lemma holds.
Proof of Theorem 16.
Following the conclusion of Lemma 35, we have
√
kυk1 ≤ 4 kυS k1 ≤ 4 d kυk2 .
(G.8)
Moreover, we have by the lower-RE condition as in Definition 2.2
b
υ T Γυ
2
2
1
2
α kυk2 ,
2
2
≥ α kυk2 − τ kυk1 ≥ (α − 16dτ ) kυk2 ≥
(G.9)
where the last inequality follows from the assumption that 16dτ ≤ α/2.
Combining the bounds in (G.9), (G.8) and (G.2), we have
1
2
α kυk2
2
Tb
≤ υ Γυ ≤ λ 2 kβk1 − 2 βb
r
+ 2ψ
1
√
5
λ kυS k1 ≤ 10λ d kυk2 .
2
√
And thus we have kυk2 ≤ 20λ d. The theorem is thus proved.
log m
kυk1
n
≤
Appendix H: Proofs for the Lasso-type estimator
Let
M+
=
32C$(s0 + 1)
and $(s0 + 1) = ρmax (s0 + 1, A) + τB =: D.
λmin (A)
By definition of s0 , we have s0 MA2 ≤
(s0 + 1) ≥
given that
√
4n
log m
and
n
2 log m
M+
s0 + 1$(s0 + 1)
≥
λmin (A)
32C
r
n
.
log m
(H.1)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
79
Rudelson and Zhou/Errors-in-variables
To prove the first inequality in (6.6) and (6.10), we need to show that
d≤
α
α
s0
5s0
=
=
.
32τ
32 λmin (A) − α
96
The first inequality in (6.6) holds so long as
n
1 1
s0 + 1
5(s0 + 1)
5s0
≤
≤
,
2 log m ≤
20 M+
20
100
96
d ≤
(H.2)
where the last inequality holds so long as s0 ≥ 24. To prove the second inequality
in (6.10), we need to show that
2
1 log m ψ
3α
5
, where τ =
for
α = λmin (A),
d≤ 2
τ
n
b0
5 s0
8
which in turn ensures that the second inequality in (6.6) holds for λ ≥ 4ψ, for ψ
appropriately chosen. We use the following inequality in the proof of Lemma 17 and
Lemma 18:
2
s0 + 1
64
1
8
1
n
n
≥
≥
2
2
2
α
25λmin (A) M+ log m
5 32C$(s0 + 1)
log m
2
2
1
1
n
n
=
≥
,
(H.3)
20CD
log m
10CD2
log m
where we use the fact that D = $(s0 +1) = ρmax (s0 +1, A)+τB ≤ kAk2 +kBk2 :=
D2 /2.
H.1. Proof of Lemma 17
1
2 . The first inequality in (6.10) holds in view of (H.2). Recall that
40M+
√
∗ 2
2
kβ k2 ≥ φb0 by definition of 0 < φ ≤ 1. Let C = C0 / c0 . By (6.9) and (H.3),
Let CA =
b20 ≥
2
1
C0 D2
n
n
≤
Dφ
2
log m
40M+
CD2
log m
2
25 32 32
n
1
C02 D22 Dφ
2 log m
9 33 33 M+
10CD2
2
25 32 32(s0 + 1) (s0 + 1) log m ψ
9 33
33
α2
n
b0
2
25 (s0 )2 log m ψ
,
9 α2
n
b0
d ≤ CA c0 Dφ
≤
≤
≤
where
C02 D22 Dφ
= C02 D22
≤
K 2 M2
+ K 4φ
b20
K2
C02 D22 2 (M
b0
∗
2
+ K kβ k2 ) =
ψ
b0
2
,
(H.4)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
80
Rudelson and Zhou/Errors-in-variables
for ψ = C0 D2 K(K kβ ∗ k2 +M ) as defined in (3.6). We have shown that (6.10) indeed
holds, and the lemma is thus proved.
H.2. Proof of Lemma 18
5s0
α
1
Let CA = 160M
2 . The proof for d ≤ 32τ = 96 follows from (H.2). In order to
+
show the second inequality, we follow the same line of arguments except that we need
to replace one inequality (H.4) with (H.5). By definition of D00 , we have kBk2 +amax ≤
(D00 )2 ≤ 2(kBk2 + amax ). Let D = $(s0 + 1).
0 2
D
By (6.11), (H.1) and (H.3), we have for c00 ≤ D0 ,
2
n
1
n
C0 D00
d ≤ CA c c Dφ
≤
Dφ
2 log m
log m
160M+
CD
2
n
1
25 322
C02 (D00 )2 Dφ
≤
2 log m
2
9 33
20CD
M+
2
2
25 322 (s0 + 1)2 log m ψ
25 (s0 )2 log m ψ
≤
≤
,
9 332
α2
n
b0
9 α2
n
b0
0 00
where assuming that s0 ≥ 32, we have the following inequality by definition of s0 and
α = 58 λmin (A),
s0 + 1 log m
α2
n
≥
8
1
5 32C$(s0 + 1)
2
≥
1
20CD
2
.
We now replace (H.4) with
C02 (D00 )2 Dφ
=
≤
where Dφ
:=
C02 (D00 )2
K4
b20
M2
+ τB+ φb20
K2
2 ψ 2
2
+/2
∗
2
0 2K
C0 (D0 ) 2 M + τB K kβ k2 ≤
b0
b0
2
4
K 2 M2
K
M
+ 4
+
∗ 2
+
τ
K
φ
≤
+
τ
kβ
k
B
B
2
b20
b20
K2
,
(H.5)
+/2
and ψ = C0 D00 K KτB kβ ∗ k2 + M K is now as defined in (4.2). The lemma is
thus proved.
Remark H.1. Throughout this paper, we assume that C0 is a large enough constant
such that for c as defined in Theorem 31,
c min{C02 , C0 } ≥ 4.
(H.6)
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
81
Rudelson and Zhou/Errors-in-variables
By definition of s0 , we have for $2 (s0 ) ≥ 1,
c0 λ2min (A) n
, and hence
1024C02 log m
λ2 (A) n
c0 λ2min (A) n
≤ min 2
=: š0 .
s0 ≤
2
1024C0 log m
1024C0 log m
√
Remark H.2. The proof shows that one can take C = C0 / c0 , and take
s0 $2 (s0 ) ≤
V = 3eMA3 /2 =
3e643 C 3 $3 (s0 )
3e643 C03 $3 (š0 )
.
≤
2λ3min (A)
2(c0 )3/2 λ3min (A)
Hence a sufficient condition on r(B) is:
n
64C0 $(š0 )
3em log m
r(B) ≥ 16c0 K 4
3 log √
+ log
.
log m
2n
c0 λmin (A)
(H.7)
Appendix I: Proofs for the Conic Programming estimator
We next provide proof for Lemmas 19 to 21 in this section.
I.1. Proof of Lemma 19
Suppose event B0 holds. Then by the proof of Corollary 13,
1
T
n X (y
∗
b
− Xβ ∗ ) + n1 tr(B)β
∞
b ∗
γ
b − Γβ
=
∞
r
≤
2
∗
2C0 D2 K kβ k2
log m
+ C0 D0 KM
n
r
log m
n
=: µ kβ ∗ k2 + ω.
The lemma follows immediately for the chosen µ, ω as in (6.12) given that (β ∗ , kβ ∗ k2 ) ∈
Υ.
I.2. Proof of Lemma 20
bb
By optimality of (β,
t), we have
βb
1
+ λ βb
2
≤ βb
1
+ λb
t ≤ kβ ∗ k1 + λ kβ ∗ k2 .
Thus we have for S := supp(β ∗ ),
βb
1
= βbS c
1
+ βbS
≤
1
kβ ∗ k1 + λ(kβ ∗ k2 − βb ).
2
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
82
Rudelson and Zhou/Errors-in-variables
Now by the triangle inequality,
βbS c
1
= kvS c k1
≤
kβS∗ k1 − βbS
1
+ λ(kβ ∗ k2 − βb )
2
∗
≤
kvS k1 + λ(kβ k2 − βb )
≤
kvS k1 + λ(kβ ∗ k2 − βbS
2
=
)
2
kvS k1 + λ kvS k2 ≤ (1 + λ) kvS k1 .
The lemma thus holds given
b
t ≤
1
1
(kβ ∗ k1 − βb ) + kβ ∗ k2 ≤ kvk1 + kβ ∗ k2 .
λ
λ
1
I.3. Proof of Lemma 21
Recall the following shorthand notation:
√
√
D0 = ( τB + amax ) and D2 = 2(kAk2 + kBk2 ).
b = tr(B),
b
First we rewrite an upper bound for v = βb − β ∗ , D = tr(B) and D
X0T X0 v
∞
=
(X − W )T X0 (βb − β ∗ )
≤
b βb
X T (X βb − y) − D
+
b − D)βb
(D
∞
∞
∞
≤ X T X0 (βb − β ∗ )
+ XT
+ W T X0 v
∞
+ W T X0 v
+ (X T W − D)βb
∞
∞
,
∞
where
X T X0 (βb − β ∗ )
∞
≤
X T (X0 βb − y + )
∞
T
=
X ((X − W )βb − y)
≤
b βb
X T (X βb − y) − D
T
+ (X W − D)βb
∞
∞
∞
+ XT
+ XT
∞
∞
b − D)βb
+ (D
∞
.
On event B0 , we have by Lemma 20 and the fact that βb ∈ Υ,
b βb
I := γ
b−Γ
∞
=
1
T
n X (y
b + 1D
bb
− X β)
n β
∞
≤ µb
t+ω
1
≤ µ( kvk1 + kβ ∗ k2 ) + ω
λ
1
= 2D2 Kρn ( kvk1 + kβ ∗ k2 ) + D0 ρn M ;
λ
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
∞
83
Rudelson and Zhou/Errors-in-variables
and on event B4 ,
II
1
n
:=
XT
∞
1/2
ρn M (amax
≤
≤
+
1
n(
√
X0T
∞
+ WT
∞
)
τB ) = D0 ρn M .
Thus on event B0 , we have
1
1
I + II ≤ 2D2 Kρn ( kvk1 + kβ ∗ k2 ) + 2D0 ρn M = µ( kvk1 + kβ ∗ k2 ) + 2ω.
λ
λ
Now on event B6 , we have for 2D1 ≤ D2
b − D)βb
IV := (D
≤
∞
b −D
D
βb
∞
1
≤ 2D1 K √ ρn (kβ ∗ k∞ + kvk∞ )
m
1
≤ D2 K √ ρn (kβ ∗ k2 + kvk1 ).
m
On event B5 ∩ B10 , we have
III :=
1
n
(X T W − D)βb
∞
≤
1
n
(X T W − D)β ∗
≤
1
n
1
n
X0T W β ∗ ∞
T
1
n
W T X0 v
∞
+
1
n
T
(X T W − D)v
(W W − D)β
∞
∗
∞
(Z BZ − tr(B)Im ) max + X0T W max kvk1
kBkF √
1/2
√ + τB amax (kvk1 + kβ ∗ k2 ),
≤ ρn K
n
√
1/2
≤ n1 W T X0 max kvk1 ≤ ρn K τB amax kvk1 .
+
and V =
+
1
n
∞
Thus we have on B0 ∩ B10 ,
III + IV + V
2
≤ ρn K kBk2 + τB + amax + √ (kAk2 + kBk2 ) (kvk1 + kβ ∗ k2 )
m
≤ ρn K (4 kBk2 + 3 kAk2 ) (kvk1 + kβ ∗ k2 )
≤ 2D2 Kρn (kvk1 + kβ ∗ k2 )
≤ µ(kvk1 + kβ ∗ k2 ),
where D0 ≤ D2 and τA = 1, and
1
T
n X0 X0 v ∞
≤
I + II + III + IV + V
1
µ( kvk1 + kβ ∗ k2 ) + 2D0 M ρn + µ(kvk1 + kβ ∗ k2 )
λ
1
2µ kβ ∗ k2 + µ( + 1) kvk1 + 2ω.
λ
≤
≤
The lemma thus holds.
Appendix J: Proof for Theorem 7
We prove Lemmas 22 to 24 in this section.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
84
Rudelson and Zhou/Errors-in-variables
J.1. Proof of Lemma 22
Suppose event B0 holds. Then by the proof of Corollary 14, we have for D00 =
1/2
1/2
kBk2 + amax ,
b ∗
γ
b − Γβ
≤
∞
+/2
D00 τB Kρn kβ ∗ k2 + D0 M ρn ,
√
+/2
1/2
1/2
oracle
where τB = τB + D√
and Doracle = 2(kBk2 + kAk2 ). The lemma follows
m
immediately for µ, ω as chosen in (6.16).
J.2. Proof of Lemma 23
Suppose event B6 holds. We first show (6.17) and (6.18). Recall rm,m := 2C0 K 2
1/2
2C0 K 2 log m
m
q
log m
mn
. By Lemma 5, we have on event B6 ,
|b
τB − τB |
≤ D1 rm,m .
Moreover, we have under (A1),
1 = τA ≤ D1 :=
kBk
kAkF
Doracle 2
+ 1/2F ≤ kAk2 + kBk2 ≤ (
) ,
2
m1/2
n
in view of (C.2). Hence
p
D1 ≤
Doracle
1/2
1/2
= kBk2 + kAk2 .
2
By definition and construction, we have τB , τbB ≥ 0,
1/2
1/2
τbB − τB
and
1/2
1/2
≤
2
τbB − τB
≤
1/2
1/2
τbB + τB ,
1/2
1/2
1/2
1/2
(b
τB + τB )(b
τB − τB ) = |b
τB − τB | .
Thus,
1/2
1/2
τbB − τB
≤
p
p
Doracle 1/2
1/2
|b
τB − τB | ≤ D1 rm,m
≤
rm,m
2
√
1/2
1/2
and for C6 ≥ Doracle ≥ 2 D1 and Doracle = 2(kAk2 + kBk2 ),
1/2
τbB −
Doracle 1/2
Doracle 1/2
1/2
1/2
rm,m ≤ τB ≤ τbB +
rm,m .
2
2
+/2
Thus we have for τB
1/2
rm,m
≥
p
(J.1)
as defined in (4.1), (J.1) and the fact that
2C0 K
√
(log m)1/4
√
≥ 2/ m for m ≥ 16 and C0 ≥ 1,
m
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
≥
85
Rudelson and Zhou/Errors-in-variables
the following inequalities hold: for K ≥ 1,
+/2
1/2
:= τB + Doracle m−1/2
Doracle 1/2
Doracle 1/2
1/2
≤ τbB +
rm,m +
rm,m
2
2
1/2
1/2
1/2
≤ τbB + Doracle rm,m
≤ τeB ,
τB
1/2
1/2
1/2
1/2
1/2
≥ τbB +Doracle rm,m as in (4.11).
where the last inequality holds by the choice of τeB
Moreover, by (J.1),
τeB
(J.2)
1/2
1/2
:= τbB + C6 rm,m
≤ τB +
Doracle 1/2
1/2
rm,m + C6 rm,m
2
3
1/2
1/2
,
τB + C6 rm,m
2
1/2
1/2 2
:= (b
τB + C6 rm,m
) ≤ 2b
τB + 2C62 rm,m
≤
and
τeB
≤
2τB + 2D1 rm,m + 2C62 rm,m
≤
2τB +
2
Doracle
rm,m + 2C62 rm,m ≤ 2τB + 3C62 rm,m .
2
2
Thus (6.17) and (6.18) hold given that 2D1 ≤ Doracle
/2 ≤ C62 /2.
Finally, we have for τB− as defined in (4.8),
1/2
1/2
τ
+ 3 C6 rm,m
3
1/2
1/2
1/2
)τB− ≤ B1/2 2
≤ 1.
τeB τB− ≤ (τB + C6 rm,m
1/2
2
τB + 2C6 rm,m
Remark J.1. The set Υ in our setting is equivalent to the following: for µ, ω as defined
in (4.11) and β ∈ Rm ,
b
Υ = (β, t) : n1 X T (y − Xβ) + n1 tr(B)β
≤ µt + ω, kβk2 ≤ t . (J.3)
∞
J.3. Proof of Lemma 24
For the rest of the proof, we will follow the notation in the proof for Lemma 21.
Notice that the bounds as stated in Lemma 20 remain true with ω, µ chosen as in (6.16),
so long as (β ∗ , kβ ∗ k2 ) ∈ Υ. This indeed holds by Lemma 22: for ω and µ (4.11) as
chosen in Theorem 7, we have by (J.2),
1/2
+/2
+/2
µ D00 τeB Kρn ≥ D00 Kρn τB , where τB
√
Doracle
= ( τB + √
),
m
which ensures that (β ∗ , kβ ∗ k2 ) ∈ Υ by Lemma 22.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
86
Rudelson and Zhou/Errors-in-variables
On event B0 , we have by Lemma 20 and the fact that βb ∈ Υ as in (J.3)
I + II
1
n
XT
:=
b βb
γ
b−Γ
∞
≤
1
T
n X (y
b + 1D
bb
− X β)
n β
≤
+
∞
∞
+ ω ≤ µb
t + 2ω
1
µ( kvk1 + kβ ∗ k2 ) + 2ω,
λ
for ω, µ as chosen in (4.11). Now on event B6 , we have under (A1),
b − D)βb
IV := (D
∞
≤
b −D
D
βb
1
≤ 2D1 K √ ρn (kβ ∗ k∞ + kvk∞ )
m
∞
Doracle
≤ D00 √
Kρn (kβ ∗ k2 + kvk1 ),
m
1/2
where 2D1
≤ Doracle D00 for 1 ≤ D00 := kBk2
1/2
1/2
Doracle = 2 kBk2 + kAk2 . Hence
1/2
+ amax , for amax ≥ τA = 1 and
√
1/2
∗
III + IV + V ≤ ρn K τB kBk2 + a1/2
max (kvk1 + kβ k2 )
≤
≤
≤
≤
√
1
+2D1 K √ ρn (kβ ∗ k2 + kvk1 ) + ρn K τB a1/2
max kvk1
m
√
√
Doracle
D00 Kρn (kvk1 + kβ ∗ k2 )( τB + √
) + ρn K τB a1/2
max kvk1
m
√
+/2
D00 Kρn τB (kvk1 + kβ ∗ k2 ) + D00 Kρn τB kvk1
r
log m 1/2 Doracle
C0 D00 K 2
)(2 kvk1 + kβ ∗ k2 )
(τB + √
n
m
µ(2 kvk1 + kβ ∗ k2 ),
for µ as defined in (4.11) in view of (J.2).
Thus we have
I + II + III + IV + V
1
≤ µ( kvk1 + kβ ∗ k2 ) + 2ω + µ(2 kvk1 + kβ ∗ k2 )
λ
1
= 2µ((1 +
) kvk1 + kβ ∗ k2 ) + 2ω,
2λ
and the improved bound as stated in the Lemma thus holds.
Appendix K: Some geometric analysis results
Let us define the following set of vectors in Rm :
Cone(s) := {υ : kυk1 ≤
√
s kυk2 }
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
87
Rudelson and Zhou/Errors-in-variables
For each vector x ∈ Rm , let T0 denote the locations of the s largest coefficients of x in
absolute values. Any vector x ∈ S m−1 satisfies:
xT0c
∞
≤ kxT0 k1 /s ≤
kxT0 k2
√ .
s
(K.1)
We need to state the following result from [33]. Let S m−1 be the unit sphere in Rm ,
for 1 ≤ s ≤ m,
Us := {x ∈ Rm : | supp(x)| ≤ s}.
(K.2)
The sets Us is an union of the s-sparse vectors. The following three lemmas are wellknown and mostly standard; See [33] and [30].
Lemma 36. For every 1 ≤ s ≤ m and every I ⊂ {1, . . . , m} with |I| ≤ s,
[
p
|I|B1m ∩ S m−1 ⊂ 2 conv(Us ∩ S m−1 ) =: 2 conv
EJ ∩ S m−1
|J|≤s
and moreover, for ρ ∈ (0, 1],
p
|I|B1m ∩ ρB2m ⊂ (1 + ρ) conv(Us ∩ B2m ) =: (1 + ρ) conv
[
EJ ∩ S m−1 .
|J|≤s
Proof. Fix x ∈ Rm . Let xT0 denote the subvector of x confined to the locations of its
s largest coefficients in absolute values; moreover, we use it to represent its 0-extended
version x0 ∈ Rm such that x0T c = 0 and x0T0 = xT0 . Throughout this proof, T0 is
understood to be the locations of the s largest coefficients in absolute values in x.
m
Moreover, let (x∗i )m
i=1 be non-increasing rearrangement of (|xi |)i=1 . Denote by
√ m
L =
sB1 ∩ ρB2m and
[
R = 2 conv
EJ ∩ B2m = 2 conv E ∩ B2m .
|J|≤s
Any vector x ∈ Rm satisfies:
xT0c
∞
≤ kxT0 k1 /s ≤
kxT0 k2
√ .
s
(K.3)
It follows that for any ρ > 0, s ≥√1 and for all z ∈ L, we have the ith largest coordinate
in absolute value in z is at most s/i, and
sup h x, z i
≤
z∈L
max h xT0 , z i + max√ h xT0c , z i
kzk2 ≤ρ
≤
ρ kxT0 k2 + xT0c
≤
kxT0 k2 (ρ + 1) ,
kzk1 ≤ s
√
∞
s
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
88
Rudelson and Zhou/Errors-in-variables
where clearly maxkzk2 ≤ρ h xT0 , z i = ρ
EJ ,
sup h x, z i
Ps
∗2 1/2
. And denote by S J
i=1 (xi )
=
(1 + ρ) max max h x, z i
=
(1 + ρ) kxT0 k2 ,
:= S m−1 ∩
J:|J|≤s z∈S J
z∈R
given that for a convex function h x, z i , the maximum happens at an extreme point;
x
and in this case, it happens for z such that z is supported on T0 , such that zT0 = x T0
k T0 k 2
and zT0c = 0.
Lemma 37. Let 1/5 > δ > 0. Let E = ∪|J|≤s EJ for 0 < s < m/2 and k0 > 0. Let
∆ be a m × m matrix such that
uT ∆v ≤ δ,
√ m
Then for all v ∈
sB1 ∩ B2m ,
∀u, v ∈ E ∩ S m−1
υ T ∆υ
(K.4)
≤ 4δ.
(K.5)
Proof. First notice that
max
υ∈
√
sB1m ∩B2m
T
υ ∆υ
≤
max
w,u∈
√
sB1m ∩B2m
T
w ∆u .
(K.6)
Now that we have decoupled u and w on the RHS of (K.6), we first fix u.
T
Then for any fixed u ∈ S m−1 and matrix ∆ ∈ Rm×m , f (w)
S= w ∆u is a convex
√ m
m−1
function of w, and hence for w ∈
sB1 ∩ B2m ⊂ 2 conv
E
∩
S
,
J
|J|≤s
max
w∈
√
sB1m ∩B2m
T
w ∆u
≤
2
=
2
max
w∈conv (E∩S m−1 )
max
w∈E∩S m−1
wT ∆u
wT ∆u ,
where the maximum occurs at an extreme point of the set conv(E ∩ S m−1 ) because of
the convexity of the function f (w).
Clearly the RHS of (K.6) is bounded by
max
u,w∈
√
sB1m ∩B2m
T
w ∆u
=
max
u∈
√
≤ 2
2
u∈
where the function g of u ∈
√
max
u∈
=
sB1m ∩B2m
√
max
√ m
w∈
sB1 ∩B2m
max
sB1m ∩B2m
max
g(u),
√
sB1m ∩B2m
w∈
E∩S m−1
wT ∆u
T
w ∆u
sB1m ∩ B2m is defined as
g(u) =
max
w∈ E∩S m−1
T
w ∆u ;
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
89
Rudelson and Zhou/Errors-in-variables
g(u) is convex since it is the maximum of a function fw (u) := wT ∆u which is
convex in u for each w ∈ (E ∩ S m−1 ).
S
√
m−1
=: 2 conv E ∩ S m−1 ,
Thus we have for u ∈ ( sB1m ∩B2m ) ⊂ 2 conv
|J|≤s EJ ∩ S
max
u∈
√ m
sB1 ∩B2m
g(u) ≤ 2
=
2
=
2
max
u∈conv (E∩S m−1 )
max
u∈E∩S m−1
g(u)
g(u)
max
(K.7)
max
u∈E∩S m−1 w∈E∩S m−1
wT ∆u ≤ 4δ,
(K.8)
where (K.7) holds given that the maximum occurs at an extreme point of the set
conv(E ∩ B2m ), because of the convexity of the function g(u).
Corollary 38. Suppose all conditions in Lemma 37 hold. Then ∀υ ∈ Cone(s),
υ T ∆υ
≤
2
4δ kυk2 .
(K.9)
Proof. It is sufficient to show that ∀υ ∈ Cone(s) ∩ S m−1 ,
υ T ∆υ
≤
4δ.
Denote by Cone := Cone(s). Clearly this set of vectors satisfy:
√ m
sB1 ∩ B2m .
Cone ∩ S m−1 ⊂
Thus (K.9) follows from (K.5).
Remark K.1. Suppose we relax the definition of Cone(s) to be:
√
Cone(s) := {υ : kυk1 ≤ 2 s kυk2 }.
Clearly, Cone(s, 1) ⊂ Cone(s). given that ∀u ∈ Cone(s, 1), we have
√
√
kuk1 ≤ 2 kuT0 k1 ≤ 2 s kuT0 k2 ≤ 2 s kuk2 .
Lemma 39. Suppose all conditions in Lemma 37 hold. Then for all υ ∈ Rm ,
2
υ T ∆υ ≤ 4δ(kυk2 +
1
2
kυk1 ).
s
(K.10)
Proof. The lemma follows given that ∀υ ∈ Rm , one of the following must hold:
if υ ∈ Cone(s)
υ T ∆υ
≤
otherwise
υ T ∆υ
≤
2
4δ kυk2 ;
4δ
2
kυk1 ,
s
(K.11)
(K.12)
leading to the same conclusion in (K.10).
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
90
We have shown (K.11) in Lemma 37. Let Cone(s)c be the complement set of Cone(s)
in Rm . That is, we focus now on the set of vectors such that
√
Cone(s)c := {υ : kυk1 ≥ s kυk2 }
√ v
,
and show that for u = s kvk
1
v T ∆v
2
kvk1
:=
1
1 T
u ∆u ≤ δ.
s
s
Now, the last inequality holds by Lemma 37 given that
[
√ m
u ∈ ( sB1 ∩ B2m ) ⊂ 2 conv
EJ ∩ B2m
|J|≤s
and thus
v T ∆v
≤
2
kvk1
1
1
uT ∆u ≤ 4δ.
sup
s u∈√sB1m ∩B2m
s
Appendix L: Proof of Corollary 25
First we show that for all υ ∈ Rm , (L.1) holds. It is sufficient to check that the
condition (K.4) in Lemma 37 holds. Then, (L.1) follows from Lemma 39: for υ ∈ Rm ,
2
υ T ∆υ ≤ 4δ(kυk2 +
1
3
1
2
2
2
kυk1 ) ≤ λmin (A)(kυk2 + kυk1 ).
k
8
k
(L.1)
The Lower and Upper RE conditions thus immediately follow. The Corollary is thus
proved.
Appendix M: Proof of Theorem 26
We first state the following preliminary results in Lemmas 40 and √
41; their proofs
appear in Section O. Throughout this section, the choice of C = C0 / c0 satisfies the
conditions on C in Lemmas 40 and 41, where recall min{C0 , C02 } ≥ 4/c for c as
defined in Theorem 31. For a set J ⊂ {1, . . . , m}, denote FJ = A1/2 EJ , where recall
EJ = span{ej : j ∈ J}. Let Z be an n × m random matrix with independent entries
2
Zij satisfying EZij = 0, 1 = EZij
≤ kZij kψ2 ≤ K. Let Z1 , Z2 be independent
copies of Z.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
91
Rudelson and Zhou/Errors-in-variables
Lemma 40. Suppose all conditions in Theorem 26 hold. Let
[
E=
EJ ∩ S m−1 .
|J|=k
Suppose that for some c0 > 0 and ε ≤
r(B) :=
tr(B)
kBk2
1
C,
√
where C = C0 / c0 ,
≥ c0 kK 4
log(3em/kε)
.
ε2
(M.1)
Then for all vectors u, v ∈ E∩S m−1 , on event B1 , where P (B1 ) ≥ 1−2 exp −c2 ε2 Ktr(B)
4 kBk
2
for c2 ≥ 2,
uT Z T BZv − EuT Z T BZv
≤
4Cεtr(B).
Lemma 41. Suppose that ε ≤ 1/C, where C is as defined in Lemma 40. Suppose that
(M.1) holds. Let
[
[
E=
EJ and F =
FJ .
(M.2)
|J|=k
|J|=k
Then on event B2 , where P (B2 ) ≥ 1 − 2 exp −c2 ε2 Ktr(B)
for c2 ≥ 2, we have for
4 kBk
2
all vectors u ∈ E ∩ S m−1 and w ∈ F ∩ S m−1 ,
wT Z1T B 1/2 Z2 u
Cεtr(B)
≤
1/2
1/2
(1 − ε)2 kBk2
≤ 4Cεtr(B)/kBk2 .
In fact, the same conclusion holds for all y, w ∈ F ∩ S m−1 ; and in particular, for
B = I, we have the following.
Corollary 42. Suppose all conditions in Lemma 40 hold. Suppose that F = A1/2 E
for E as defined in Lemma 40. Let
n ≥
c0 kK 4
log(3em/kε)
.
ε2
(M.3)
Then on event B3 , where P (B3 ) ≥ 1 − 2 exp −c2 ε2 n K14 , we have for all vectors
w, y ∈ F ∩ S m−1 and ε ≤ 1/C for C is as defined in Lemma 40,
y T ( n1 Z T Z − I)w
≤ 4Cε.
(M.4)
We prove Lemmas 40 and 41 and Corollary 42 in Section O. We are now ready to prove
Theorem 26.
Proof of Theorem 26.
b A − A :=
∆ := Γ
=
Let
1
T
nX X
( n1 X0T X0 − A) +
b
− n1 tr(B)I
m−A
1
n
W T X0 + X0T W +
1
n
b
W T W − tr(B)I
m ,
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
92
Rudelson and Zhou/Errors-in-variables
where recall X0 = Z1 A1/2 . Notice that
b A − A)υ = uT (X T X − tr(B)I
b
uT (Γ
m − A)υ
≤
uT ( n1 X0T X0 − A)υ + uT n1 (W T X0 + X0T W )υ + uT ( n1 W T W −
≤
uT A1/2 n1 Z1T Z1 A1/2 υ − uT Aυ + uT n1 (W T X0 + X0T W )υ
+ uT ( n1 Z2T BZ2 − τB Im )υ +
1
n
b
tr(B)
− tr(B) uT υ =: I + II + III + IV.
A1/2 u
For u ∈ E ∩ S m−1 , define h(u) :=
b
tr(B)
f Im )υ
kA1/2 uk2
. The conditions in (M.1) and (M.3) hold
for k.
We first bound the middle term as follows. Fix u, υ ∈ E ∩ S m−1 . Then on event B2 ,
for Υ = Z1T B 1/2 Z2 ,
uT (W T X0 + X0T W )υ
=
uT Z2T B 1/2 Z1 A1/2 υ + uT A1/2 Z1T B 1/2 Z2 υ
≤
uT ΥT h(v)
A1/2 v
+ h(u)T Υυ
A1/2 u
2
≤ 2
2
T
max
w Υυ
1/2
w∈F ∩S m−1 ,υ∈E∩S m−1
≤ 8Cεtr(B)
ρmax (k, A)
kBk2
ρ1/2
max (k, A)
.
We now use Lemma 40 to bound both I and III. We have for C as defined in Lemma 40,
on event B1 ∩ B3 ,
uT (Z2T BZ2 − tr(B)Im )υ
≤ 4Cεtr(B).
Moreover, by Corollary 42, we have on event B3 , for all u, v ∈ E ∩ S m−1 ,
uT ( n1 X0T X0 − A)υ
=
uT A1/2 Z T ZA1/2 υ − uT Aυ
=
h(u)T ( n1 Z T Z − I)h(υ)
≤
1
n
A1/2 u
A1/2 υ
2
2
maxw,y∈F ∩S m−1 wT (Z T Z − I)y ρmax (k, A)
≤ 4Cερmax (k, A).
Thus we have on event B1 ∩ B2 ∩ B3 and for τB := tr(B)/n,
!
1/2
ρmax (k, A)
I + II + III ≤ 4Cε ρmax (k, A) + 2τB
+ τB
kBk2
≤ 8Cε (τB + ρmax (k, A)) .
On event B6 , we have for D1 as defined in Lemma 5,
r
IV ≤ |b
τB − τB | ≤ 2C0 D1 K 2
The theorem thus holds by the union bound.
log m
.
mn
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
93
Rudelson and Zhou/Errors-in-variables
Appendix N: Proof of Lemma 32
Lemma 43 is a well-known fact.
Lemma 43. Let Auw := (u ⊗ w) ⊗ A, where u, w ∈ Sm−1 for m ≥ 2. Then
kAuw k2 ≤ kAk2 and kAuw kF ≤ kAkF .
Proof of Lemma 32. Let z1 , . . . , zn , z10 , . . . , zn0 ∈ Rm be the row vectors Z1 , Z2
respectively. Notice that we can write the quadratic form as follows:
X
uT Z1 A1/2 Z2T w =
ui wj zi A1/2 zj0
i,j=1,m
vec ZT
1
T
=: vec Z1
= vec ZT
=: vec ZT
=
uT ZAZ T w
T
(u ⊗ w) ⊗ A1/2 vec ZT
2
T
T 1/2
Auw vec Z2 ,
T
(u ⊗ w) ⊗ A vec ZT
T
Auw vec ZT
where clearly by independence of Z1 , Z2 ,
Evec ZT
1
T
(u ⊗ w) ⊗ A1/2 vec ZT
2
T
Evec { Z } (u ⊗ u) ⊗ A vec { Z }
=
0, and
=
tr (u ⊗ u) ⊗ A = tr(A).
Thus we invoke (C.1) and Lemma 43 to show the concentration bounds on event
{ uT Z1 A1/2 Z2T w > t}:
2
t
t
P uT Z1 A1/2 Z2T w > t
≤ 2 exp − min
2 ,
1/2
1/2
2
K Auw
K 4 Auw
2
F
!!
2
t
t
≤ 2 exp − min
,
.
K 4 tr(A) K 2 A1/2 2
Similarly, we have by Theorem 31 and Lemma 43,
P uT ZAZ T w − EuT ZAZ T w > t
≤ 2 exp −c min
≤ 2 exp −c min
The Lemma thus holds.
t2
t
2 , K 2 kA
4
uw k2
K kAuw kF
!!
t2
t
.
2 ,
K 4 kAkF K 2 kAk2
!!
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
94
Rudelson and Zhou/Errors-in-variables
Appendix O: Proof of Lemmas 40 and 41 and Corollary 42
tr(B)
kBk2 .
Let ε ≤ C1 where C is
√
large enough so that cc0 C 2 ≥ 4, and hence the choice of C = C0 / c0 satisfies our
need.
Throughout the following proof, we denote by r(B) =
Proof of Lemma 40. First we prove concentration bounds for all pairs of u, v ∈
Π0 , where Π0 ⊂ Sm−1 is an ε-net of E. Let t = CK 2 εtr(B). We have by Lemma 32,
and the union bound,
P ∃u, v ∈ Π0 , uT Z T BZv − EuT Z T BZv > t
"
!#
t
t2
0 2
≤ 2 |Π | exp −c min
2 ,
K 4 kBkF K 2 kBk2
CK 2 ε2 r(B)
2
≤ 2 |Π0 | exp −c min C 2 ,
ε
K4
≤ 2 exp −c2 ε2 r(B)/K 4 ,
2
where we use the fact that kBkF ≤ kBk2 tr(B) and
m
0
|Π | ≤
(3/ε)k ≤ exp(k log(3em/kε)),
k
while
CK 2
r(B)
c min C 2 ,
ε2 4
ε
K
tr(B)
kBk2 K 4
3em
3em
≥ cC02 k log
≥ 4k log
.
kε
kε
= cC 2 ε2
Denote by B2 the event such that for Λ :=
sup
v T Λu
u,v∈Π0
1
T
tr(B) (Z BZ
≤
− I),
0
Cε =: rk,n
holds. A standard approximation argument shows that under B2 and for ε ≤ 1/2,
y T Λx ≤
sup
x,y∈Sm−1 ∩E
The lemma is thus proved.
Proof of Lemma 41.
0
rk,n
≤ 4Cε.
(1 − ε)2
(O.1)
1/2
By Lemma 32, we have for t = Cεtr(B)/ kBk2
for
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
95
Rudelson and Zhou/Errors-in-variables
√
C = C0 / c0 ,
2
2
C 2 tr(B)
ε
Cεtr(B)
kBk2
≤ exp −c min
,
K 4 tr(B) K 2 kBk2
2 2
C ε rB CεrB
,
≤ 2 exp −c min
K4
K2
2
2 CK
2
4
≤ 2 exp −c min C ,
ε rB /K .
ε
P
wT Z1T B 1/2 Z2 u > t
Choose an ε-net Π0 ⊂ S m−1 such that
[
Π0 =
Π0J where Π0J ⊂ EJ ∩ S m−1
(O.2)
|J|=k
is an ε-net for EJ ∩ S m−1 and
m
|Π0 | ≤
(3/ε)k ≤ exp(k log(3em/kε)).
k
Similarly, choose ε-net Π of F ∩ S m−1 of size at most exp(k log(3em/kε)). By the
union bound and Lemma 32, and for K 2 ≥ 1,
1/2
P ∃w ∈ Π, u ∈ Π0 s.t. wT Z1T B 1/2 Z2 u ≥ Cεtr(B)/kBk2
≤ |Π0 | |Π| 2 exp −c min CK 2 /ε, C 2 ε2 rB /K 4
≤ exp (2k log(3em/kε)) 2 exp −cC 2 ε2 rB /K 4
≤ 2 exp −c2 ε2 rB /K 4 ,
where C is large enough such that cc0 C 2 := C 0 > 4 and for ε ≤
1
C,
tr(B)
c min CK 2 /ε, C 2 ε2
≥ C 0 k log(3em/kε) ≥ 4k log(3em/kε).
kBk2 K 4
Denote by Υ := Z1T B 1/2 Z2 . A standard approximation argument shows that if
wT Υu ≤ Cε
sup
w∈Π,u∈Π0
tr(B)
1/2
kBk2
=: rk,n ,
an event which we denote by B2 , then for all u ∈ E and w ∈ F ,
rk,n
wT Z1T B 1/2 Z2 u ≤
.
(1 − ε)2
The lemma thus holds for c2 ≥ C 0 /2 ≥ 2.
(O.3)
Proof of Corollary 42. Clearly (M.4) implies that (M.1) holds for B = I. Clearly (M.3)
holds following the analysis of Lemma 40 by setting B = I, while replacing event B1
with B3 , which denotes an event such that
sup
u,v∈Π
1
n
v T (Z T Z − I)u
≤ Cε.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
96
The rest of the proof follows by replacing E with F everywhere. The corollary thus
holds.
References
[1] AGARWAL , A., N EGAHBAN , S. and WAINWRIGHT, M. (2012). Fast global convergence of gradient methods for high-dimensional statistical recovery. Annals of
Statistics 40.
[2] A LLEN , G. and T IBSHIRANI , R. (2010). Transposable regularized covariance
models with an application to missing data imputation. Annals of Applied Statistics 4 764-790.
[3] B ELLONI , A., ROSENBAUM , M. and T SYBAKOV, A. (2014). Linear and
Conic Programming Estimators in High-Dimensional Errors-in-variables Models. arXiv:1408.0241.
[4] B ICKEL , P., R ITOV, Y. and T SYBAKOV, A. (2009). Simultaneous Analysis of
Lasso and Dantzig Selector. The Annals of Statistics 37 1705–1732.
[5] B ONILLA , E., C HAI , K. M. and W ILLIAMS , C. (2008). Multi-task Gaussian
process prediction. In In Advances in Neural Information Processing Systems 20
(NIPS 2010).
[6] C AND ÈS , E. and TAO , T. (2007). The Dantzig selector: statistical estimation
when p is much larger than n. Annals of Statistics 35 2313-2351.
[7] C ARROLL , R. J., G AIL , M. H. and L UBIN , J. H. (1993). Case-control studies
with errors in predictors. Journal of American Statistical Association 88 177 –
191.
[8] C ARROLL , R. J., G ALLO , P. P. and G LESER , L. J. (1985). Comparison of least
squares and errors-in-variables regression with special reference to randomized
analysis of covariance. Journal of American Statistical Association 80 929 – 932.
[9] C ARROLL , R. and WAND , M. (1991). Semiparametric estimation in logistic measurement error models. J. R. Statist. Soc. B 53 573-585.
[10] C ARROLL , R. J., RUPPERT, D., S TEFANSKI , L. A. and C RAINICEANU , C. M.
(2006). Measurement Error in Nonlinear Models (Second Edition). Chapman &
Hall.
[11] C HEN , Y. and C ARAMANIS , C. (2013). Noisy and Missing Data Regression:
Distribution-Oblivious Support Recovery. In Proceedings of The 30th International Conference on Machine Learning ICML-13.
[12] C HEN , S., D ONOHO , D. and S AUNDERS , M. (1998). Atomic decomposition by
basis pursuit. SIAM Journal on Scientific and Statistical Computing 20 33–61.
[13] C OHEN , M. R. and KOHN , A. K. (2011). Measuring and interpreting neuronal
correlations. Nature Neuroscience 14 809-811.
[14] C OOK , J. R. and S TEFANSKI , L. A. (1994). Simulation-extrapolation estimation in parametric measurement error models. Journal of the American Statistical
Association 89 1314–1328.
[15] DAWID , A. P. (1981). Some Matrix-Variate Distribution Theory: Notational Considerations and a Bayesian Application. Biometrika 68 265–274.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
97
[16] D EMPSTER , A., L AIRD , N. and RUBIN , D. (1977). Maximum likelihood from
incomplete data via the EM algorithm. Journal of the Royal Statistical Society,
Series B 39 1-38.
[17] D UTILLEUL , P. (1999). The MLE Algorithm for the matrix normal distribution.
Journal of Statistical Computation and Simulation 64 105–123.
[18] E FRON , B. (2009). Are a set of microarrays independent of each other? Ann. App.
Statist. 3 922–942.
[19] FAN , J. and L I , R. (2001). Variable selection via nonconcave penalized likelihood
and its oracle properties. Journal of American Statistical Association 96 1348–
1360.
[20] F ULLER , W. A. (1987). Measurement error models. John Wiley and Sons.
[21] G AUTIER , E. and T SYBAKOV, A. (2011). High-dimensional instrumental variables regression and confidence sets. arXiv:1105.2454.
[22] G UPTA , A. and VARGA , T. (1992). Characterization of Matrix Variate Normal
Distributions. Journal of Multivariate Analysis 41 80-88.
[23] H ALL , P. and M A , Y. (2007). Semiparametric estimators of functional measurement error models with unknown error. Journal of the Royal Statistical Society B
69 429-446.
[24] H WANG , J. T. (1986). Multiplicative Errors-in-Variables Models with Applications to Recent Data Released by the U.S. Department of Energy. Journal of
American Statistical Association 81 680–688.
[25] I TURRIA , S. J., C ARROLL , R. J. and F IRTH , D. (1999). Polynomial regression
and estimating functions in the presence of multiplicative measurement error.
Journal of the Royal Statistical Society, Series B, Methodological 61 547-561.
[26] K ALAITZIS , A., L AFFERTY, J., L AWRENCE , N. and Z HOU , S. (2013). The Bigraphical Lasso. In Proceedings of The 30th International Conference on Machine Learning ICML-13 1229-1237.
[27] K ASS , R., V ENTURA , V. and B ROWN , E. (2005). Statistical Issues in the Analysis of Neuronal Data. J Neurophysiol 94 8–25.
[28] L IANG , H., H ÄRDLE , W. and C ARROLL , R. J. (1999). Estimation in a semiparametric partially linear errors-in-variables model. Ann. Statist. 27 1519-1535.
[29] L IANG , H. and L I , R. (2009). Variable selection for partially linear models with
measurement Errors. Journal of the American Statistical Association 104 234248.
[30] L OH , P. and WAINWRIGHT, M. (2012). High-dimensional regression with noisy
and missing data: Provable guarantees with nonconvexity. The Annals of Statistics
40 1637–1664.
[31] L OH , P. and WAINWRIGHT, M. (2015). Regularized M-estimators with nonconvexity: Statistical and algorithmic theory for local optima. Journal of Machine
Learning Research 16 559–616.
[32] M A , Y. and L I , R. (2010). Variable selection in measurement error models.
Bernoulli 16 274-300.
[33] M ENDELSON , S., PAJOR , A. and T OMCZAK -JAEGERMANN , N. (2008). Uniform uncertainty principle for Bernoulli and subgaussian ensembles. Constructive
Approximation 28 277–289.
[34] N ESTEROV, Y. (2007). Gradient methods for minimizing composite objective
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
98
function. CORE DISCUSSION PAPER 2007/76, Center for Operations Research
and Econometrics (CORE), Catholic University of Louvain (UCL) 1-30.
ROSENBAUM , M. and T SYBAKOV, A. (2010). Sparse recovery under matrix uncertainty. The Annals of Statistics 38 2620-2651.
ROSENBAUM , M. and T SYBAKOV, A. (2013). Improved matrix uncertainty selector. IMS Collections 9 276-290.
RUDELSON , M. and V ERSHYNIN , R. (2013). Hanson-Wright inequality and subgaussian concentration. Electronic Communications in Probability 18 1–9.
RUDELSON , M. and Z HOU , S. (2013). Reconstruction from anisotropic random
measurements. IEEE Transactions on Information Theory 59 3434-3447.
RUDELSON , M. and Z HOU , S. (2015). High dimensional errors-in-variables
models with dependent measurements. arXiv:1502.02355.
RUFF , D. A. and C OHEN , M. R. (2014). Attention can either increase or decrease
spike count correlations in visual cortex. Nature Neuroscience 17 1591-7.
S ØRESEN , Ø., F RIGENSSI , A. and T HORESEN , M. (2014a). Measurement Error
in Lasso: Impact and Likelihood Bias Correction. Statistical Sinica Preprint.
S ØRESEN , Ø., F RIGENSSI , A. and T HORESEN , M. (2014b). Covariate Selection in High-Dimensional Generalized Linear Models with Measurement Error.
arXiv:1407.1070.
S T ÄDLER , N., S TEKHOVEN , D. J. and B ÜHLMANN , P. (2014). Pattern Alternating Maximization Algorithm for Missing Data in High-Dimensional Problems.
Journal of Machine Learning Research 15 1903-1928.
S TEFANSKI , L. A. (1985). The effects of measurement error on parameter estimation. Biometrika 72 583–592.
S TEFANSKI , L. A. (1990). Rates of convergence of some estimators in a class of
deconvolution problems. Statistics and Probability Letters 9 229–235.
S TEFANSKI , L. A. and C OOK , J. R. (1995). Simulation-extrapolation: The measurement error jackknife. Journal of the American Statistical Association 90
1247–1256.
S TRIMMER , K. (2003). Modeling gene expression measurement error: a quasilikelihood approach. BMC Bioinformatics 4.
T IBSHIRANI , R. (1996). Regression shrinkage and selection via the Lasso. J. Roy.
Statist. Soc. Ser. B 58 267-288.
T ROPP, J. A. (2004). Greed is good: Algorithmic results for sparse approximation. IEEE Trans. Inform. Theory 50 2231–2241.
T ROPP, J. and G ILBERT, A. (2007). Signal recovery from random measurements
via orthogonal matching pur- suit. IEEE Trans. Inform. Theory 53 4655-4666.
V IAL , J. P. (1982). Strong convexity of sets and functions. Journal of Mathematical Economics 9 187–205.
W ERNER , K., JANSSON , M. and S TOICA , P. (2008). On Estimation of Covariance Matrices With Kronecker Product Structure. IEEE Transactions on Signal
Processing 56 478 – 491.
X U , Q. and YOU , J. (2007). Covariate Selection for Linear Errors-in-Variables
Regression Models. Communications in Statistics – Theory and Methods 36.
Y U , K., L AFFERTY, J., Z HU , S. and G ONG , Y. (2009). Large-scale collaborative
prediction using a nonparametric random effects model. Proceedings of the 26th
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
Rudelson and Zhou/Errors-in-variables
99
International Conference on Machine Learning.
[55] Z HANG , C. H. (2010). Nearly unbiased variable selection under minimax concave penalty. Ann. Statist. 38 894-942.
[56] Z HOU , S. (2014). Gemini: Graph estimation with matrix variate normal instances. Annals of Statistics 42 532–562.
[57] Z HOU , S., L AFFERTY, J. and WASSERMAN , L. (2010). Time Varying Undirected
Graphs. Machine Learning 80 295-319.
imsart-ejs ver. 2014/10/16 file: eiv-rz.tex date: April 4, 2017
| 10math.ST
|
PSIque: Next Sequence Prediction of Satellite Images
using a Convolutional Sequence-to-Sequence Network
Seungkyun Hong ∗,1,2
Seongchan Kim 2
Minsu Joh 1,2
Sa-kwang Song †,1,2
Korea University of Science and Technology
2
Korea Institute of Science and Technology Information
{xo,sckim,msjoh,esmallj}@kisti.re.kr
arXiv:1711.10644v2 [cs.CV] 30 Nov 2017
1
Abstract
Predicting unseen weather phenomena is an important issue for disaster management. In this paper, we suggest a model for a convolutional sequence-to-sequence
autoencoder for predicting undiscovered weather situations from previous satellite
images. We also propose a symmetric skip connection between encoder and decoder modules to produce more comprehensive image predictions. To examine our
model performance, we conducted experiments for each suggested model to predict
future satellite images from historical satellite images. A specific combination
of skip connection and sequence-to-sequence autoencoder was able to generate
closest prediction from the ground truth image.
1
Introduction
Destruction caused by extreme weather events, such as cyclones and wildfires, has garnered increasing
attention in recent decades, and the need to predict such events has required more descriptive
nowcasting technologies. Currently, numerical weather prediction (NWP) models are performed to
forecast future weather situations; still, these simulations require intensive computation and long-term
cumulative datasets. Further, although NWP have advantages of accuracy and widespread simulation
performance, additional effort is required to fit these models to a responsive nowcasting system.
Recent improvements in earth observation systems have allowed us to observe the changes in global
weather on an hourly basis via remote sensing imageries. Since massive earth observation data with
long durations are now available, several works have surveyed the effectiveness of weather events
considering prediction tasks.
However, occasionally, there is a possibility of data loss from remote sensory systems. For example,
a geostationary satellite may fail to produce broad observation data owing to several issues such
as image merging, satellite repositioning on the designated orbit, special missions such as solar or
lunar eclipse, and so on. These operational failures can result in satellite stations facing critical
data loss. However, there could be possible sudden emergency situations wherein weather events
prediction without sufficient data sources or even on nonexistent moments may be required. Given
this background, it is unclear how event detection systems will forecast unseen or missed situations.
In this paper, we suggest an approach for predicting a visually unobserved situation by generating a
plausible image by referencing historical time-series visual clues. We found that specific information
propagation leads to better, reasonable prediction results. The remainder of this paper is organized
as followed. In the next section, we first discuss the recent research milestones in future image
prediction techniques. Next, we present our proposed model structure for future weather situation
prediction, and we conclude with comparison results and the conclusions.
†
Corresponding Author
Workshop on Deep Learning for Physical Sciences (DLPS 2017), NIPS 2017, Long Beach, CA, USA.
2
Related works
A prediction task of future weather conditions is a joint task comprising two separate meteorological
processes: (1) understanding visual clues for weather transition from spatiotemporal weather observations, and (2) image restoration from low-dimensional encoded information into high-dimensional
visual situations through time.
Recently, several studies focused on weather events prediction using either observation or simulation
data. Shi et al. [1] and Kim et al. [2] predicted future precipitation from historical multichannel radar
reflectivity images with a convolutional long short-term memory (ConvLSTM) network [3] based
on recurrent neural network (RNN) [4] and convolutional neural networks (CNNs). Racah et al. [5]
suggested an 3D convolutional autoencoder (AE) model for extreme weather events detection using
27-years CAM-5 simulation model results. Hong et al. [6] surveyed multiple CNNs for predicting
the coordinates of a typhoon eye from a single satellite image. Kim et al. [7] suggested a tropical
cyclone detection system based on GCM reanalysis data using 5-layer CNNs. These works depend
on historical visual datasets for predicting weather events.
However, it is not confirmed if unseen weather events can be predicted for specific occasions
without observation data such as an observation failure and future events prediction. For the issue of
missing/lost data, several studies on deep neural memory networks have been conducted for predicting
undiscovered data. [3, 4] Based on the LSTM-RNN, Srivastava et al. [8] surveyed an unsupervised
video representation on a fully-connected LSTM (fcLSTM) network with flatten data. Accordingly,
Patraucean et al. [9] used intermediate differentiable memory on a temporal video auto-encoder
network, and trained extensive optical flow transition differences through time between encoding and
decoding steps. Chen et al. [10] considered a 3D biomedical image as the spatial data continuum and
adopted a bi-directional LSTM structure for understanding 3D contexts beyond correlated 2D slices.
With the advantages of a memory network, several research studies also discovered the effectiveness
of structural separation into discrete encoder and decoder units for both sequenced input and output.
Cho et al. [11] suggested an RNN encoder–decoder architecture associated with two separated RNNs
as an encoder and divaricated decoder parts. Accordingly, Sutskever et al. [12] employed the structure
of [11] and an adopting LSTM cell [3] on the overall topology, entitled sequence-to-sequence
(Seq2Seq). Next, Chorowski et al. [13] introduced an external memory in Seq2Seq called attention
memory, as well as which is similar to a differentiable memory, which was used in a convolutional
attention-based Seq2Seq [14]. For image restoration issues, Mao et al. [15] introduced the residual
connectivity of the results of convolutional processes into the deconvolutional step for maximizing
plausible image reconstruction. Enriched with abundant visual clues from deep convolutional filters,
symmetric skip connections (SkipConx) helped achieve high performance as a denoising application,
and as a conventional image autoencoder.
In conclusion, few surveys have been conducted on image reconstruction for remote sensing imagery,
particularly on a satellite image dataset. Therefore, we studied the initial structure of learning weather
changes through time, with minimal available parameters. In other words, we studied how changes in
weather can be represented as fundamental structures of a memory network.
3
3.1
Methodology
Model Architecture
In this section, we discuss the architecture for predicting next one-hour satellite images using historical
observation datasets. To predict undiscovered situations, we used a basic Seq2Seq architecture that
consists of two RNNs using an encoder-decoder framework. In the encoder section, existing weather
situation data are plotted and encoded as reduced information through time. Next, the decoder
receives outputs and network states from the encoder, and uses this information to predict future
situations one-by-one.
Both the encoder and the decoder use a five-layer convolution operation to extract rich features or
restore encoded information during deconvolution operations. As convolutional filter weights of the
encoder and the decoder could be repeatedly utilized, the encoder and the decoder must be able to
save and restore their parameters within a scope of Seq2Seq memory networks because the model is
not restricted to a specific number of input and output sequences. The suggested model can therefore
2
Figure 1: Model topology for Seq2Seq Autoencoder with Selective Skip Connection (SkipConx)
be used for flexible input/output configurations. However, a fixed time gap between each data instance
(e.g., 1 hour gap per input) is suggested to achieve better prediction results.
As the basic LSTM network uses dense 1D information, we flattened the 2D image data into 1D
tensors to fit LSTM cells. To achieve better feature extraction from the spatial data space, we
introduce convolutional LSTM cells into both the encoder and the decoder. Further, we implement an
auxiliary link in the image regeneration step—a symmetric skip connection from the convolutional
encoder into the deconvolutional decoder—for better image restoration.
3.2
Dataset
We used four-channel infrared images observed from the COMS-1 satellite meteorological imager
(MI) for weather situation prediction. This dataset contains data from April 2011 to June 2017,
and it consists of only a partial subset owing to data collection issues. The pixel dimension of the
image dataset was resized to 256 × 256 × 4 for fast model training. Further, we designed a data
sequencer gateway for the Seq2Seq autoencoder, which receives four-step information to interpret
weather changes and produces two-step weather transitions. Meanwhile, a single instance of the
data sequencer for training Seq2Seq autoencoder requires six iterations of satellite observation, i.e.,
one hour per image, for a total time of 6 h. Based on those sequence settings and the status of data
collection, we constructed about 26,490 instances to train several Seq2Seq autoencoders and split the
set with a ratio of 0.81:0.09:0.1 (for the training, validation, and test configurations).
4
Evaluation
Figure 2: Model loss of each model configuration. Lower scores show better prediction results.
To evaluate the performance of the convolutional Seq2Seq model, we conducted multiple experiments
to identify the best image prediction setting. A basic LSTM cell was used for image restoration
(C2L2D), and a convolutional LSTM cell was also used (C2CL2D) for comparison. Next, a multiple
skip connection was designed between the encoder and the decoder in the C2CL2D model (C2CL2DSkipConx) to discover the optimal topology for plausible image restoration. Since our model uses
five convolutional layers, we created several combinations of layers for selective symmetric skip
connections: L5, L3/5, L2/3/5, and L1/3/5.
The accuracy of weather situation prediction models is defined as (1), where P is a set of pixels that
consists of width × height and the longitude of the ground truth, and P̂ is a set of pixels from the
prediction result.
N
2
1 X
M SEP ixels =
(P − P̂ )
(1)
N n=1
Each testing model was trained for 50 epochs and used an exponential linear unit (Elu) as an
activation function for increasing nonlinearity. Further, the Adam optimizer [16] is used for gradient
3
optimization with an initial learning rate 1e-3 to deal with large pixel values. All our models use the
toolkit TensorFlow [17] release 1.4 as the numerical computation framework with GPU acceleration
support from the CUDA/cuDNN framework. For model training and inference acceleration, we used
dual NVIDIA® Tesla® P100 GPU accelerators for boosting mini-batch training.
Table 1: MSE (Test) result of each model configuration after 50 epochs.
Model
Configuration
MSE
C2L2D (vanilla)
C2CL2D (vanilla)
C2CL2D (SkipConx)
C2CL2D (SkipConx)
C2CL2D (SkipConx)
C2CL2D (SkipConx)
fcLSTM
ConvLSTM
ConvLSTM, L5
ConvLSTM, L3/5
ConvLSTM, L2/3/5
ConvLSTM, L1/3/5
7537.908 [baseline]
215.039
206.413
208.698
214.000
197.222
Fig. 2 and Table. 1 shows the result of test errors for each situation prediction model. All test models
converged at some point. Compared to the C2L2D model, other C2CL2D models almost converged
around MSE 200. Unfortunately, the C2L2D model was trained too slowly and weakly, and the last
test error among 50 epoch trainings was almost MSE 7538, which is almost 35 times larger than the
other C2CL2D model configurations.
Figure 3: Ground truth (actual observation) vs. Prediction through time on Seq2Seq
Fig. 3 shows an example of qualitative prediction results from all model configurations, compared
with original observation data (201403131600 UTC, ch1-IR1). The prediction results of the C2L2D
model are almost unusable to figure out the weather situation, which show a lack of learning capability.
In contrast, the C2CL2D model shows more comprehensive weather predictions than the C2L2D
result. The C2CL2D model with the skip connection on Layer 1/3/5 achieved the lowest test error
among all models, and it presented the most reliable weather prediction result.
5
Conclusion
In this work, we studied how a Seq2Seq-based convolutional autoencoder can predict unseen weather
situations by referencing historical weather observation datasets. Further, we found that adding
symmetric skip connection from the convolutional encoder to the deconvolutional decoder in the
Seq2Seq autoencoder provides more reliable image prediction compared to bare LSTM or primitive
convolutional LSTM connectivity. Meanwhile, our proposed model still needs improvement in terms
of image resolution for more comprehensive weather situation prediction. We have planned several
optimizations, such as differentiable memory and usage of adversarial networks, to improve our work.
6
Acknowledgement
This work formed part of research projects carried out at the Korea Institute of Science and Technology
Information (KISTI). (Project No. K-17-L03-C03: Construction of HPC-based Service Infrastructure
Responding to National Scale Disaster; Project No. K-17-L05-C08: Research for Typhoon Track
Prediction using an End-to-End Deep Learning Technique) We also gratefully acknowledge the
support of NVIDIA Corporation with the donation of multiple GPUs used for this research.
4
References
[1] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-k. Wong, and W.-c. Woo, “Convolutional lstm
network: A machine learning approach for precipitation nowcasting,” in Advances in Neural
Information Processing Systems, pp. 802–810, 2015.
[2] S. Kim, S. Hong, M. Joh, and S.-k. Song, “Deeprain: A predictive convlstm network for
precipitation using multi-channel radar data,” in Proceedings of the 7th International Workshop
on Climate Informatics: CI 2017., 2017.
[3] S. Hochreiter and J. Urgen Schmidhuber, “Long short-term memory,” Neural Computation,
vol. 9, no. 8, pp. 1735–1780, 1997.
[4] J. L. Elman, “Finding structure in time,” Cognitive Science, vol. 14, no. 2, pp. 179–211, 1990.
[5] E. Racah, C. Beckham, T. Maharaj, Prabhat, and C. Pal, “Semi-supervised detection of extreme
weather events in large climate datasets,” arXiv preprint arXiv:1612.02095, 2016.
[6] S. Hong, S. Kim, M. Joh, and S.-k. Song, “Globenet: Convolutional neural networks for typhoon
eye tracking from remote sensing imagery,” Proceedings of the 7th International Workshop on
Climate Informatics: CI 2017., Aug. 2017.
[7] S. K. Kim, S. Ames, J. Lee, C. Zhang, A. Wilson C., and D. Williams, “Massive Scale Deep
Learning for Detecting Extreme Climate Events,” in Proceedings of the 7th International
Workshop on Climate Informatics: CI 2017., 2017.
[8] N. Srivastava, E. Mansimov, and R. Salakhutdinov, “Unsupervised learning of video representations using lstms,” in International Conference on Machine Learning, pp. 843–852, 2015.
[9] V. Patraucean, A. Handa, and R. Cipolla, “Spatio-temporal video autoencoder with differentiable
memory,” arXiv preprint arXiv:1511.06309, 2015.
[10] J. Chen, L. Yang, Y. Zhang, M. Alber, and D. Z. Chen, “Combining fully convolutional
and recurrent neural networks for 3d biomedical image segmentation,” Advances in Neural
Information Processing Systems, pp. 3036–3044, 2016.
[11] K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and
Y. Bengio, “Learning phrase representations using rnn encoder-decoder for statistical machine
translation,” in Proceedings of the 2014 Conference on Empirical Methods in Natural Language
Processing (EMNLP), pp. 1724–1734, 2014.
[12] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence learning with neural networks,”
in Advances in Neural Information Processing Systems, pp. 3104–3112, 2014.
[13] J. Chorowski, D. Bahdanau, D. Serdyuk, K. Cho, and Y. Bengio, “Attention-based models
for speech recognition,” in Advances in Neural Information Processing Systems, pp. 577–585,
2015.
[14] D. Lim, “Convolutional attention-based seq2seq neural network for end-to-end asr,” mathesis,
Korea University, 2017.
[15] X.-J. Mao, C. Shen, and Y.-B. Yang, “Image restoration using very deep convolutional encoderdecoder networks with symmetric skip connections,” in Advances in Neural Information Processing Systems, 2016.
[16] D. P. Kingma and J. L. Ba, “Adam: a Method for Stochastic Optimization,” International
Conference on Learning Representations 2015, pp. 1–15, 2015.
[17] M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, G. Irving,
M. Isard, M. Kudlur, J. Levenberg, R. Monga, S. Moore, D. G. Murray, B. Steiner, P. Tucker,
V. Vasudevan, P. Warden, M. Wicke, Y. Yu, X. Zheng, and G. Brain, “TensorFlow: A System
for Large-Scale Machine Learning TensorFlow: A system for large-scale machine learning,”
in 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI ’16),
pp. 265–284, 2016.
5
| 9cs.NE
|
SecSens: Secure State Estimation with Application
to Localization and Time Synchronization
Amr Alanwar
Bernhard Etzlinger
Henrique Ferraz
University of California, Los Angeles
[email protected]
Johannes Kepler University
[email protected]
University of California, Santa Barbara
[email protected]
João Hespanha
Mani Srivastava
arXiv:1801.07132v1 [cs.CR] 22 Jan 2018
University of California, Santa Barbara University of California, Los Angeles
[email protected]
[email protected]
Abstract—Research evidence in Cyber-Physical Systems (CPS)
shows that the introduced tight coupling of information technology with physical sensing and actuation leads to more vulnerability and security weaknesses. But, the traditional security
protection mechanisms of CPS focus on data encryption while
neglecting the sensors which are vulnerable to attacks in the
physical domain. Accordingly, researchers attach utmost importance to the problem of state estimation in the presence of sensor
attacks. In this work, we present SecSens, a novel approach for
secure nonlinear state estimation in the presence of modeling
and measurement noise. SecSens consists of two independent
algorithms, namely, SecEKF and SecOPT, which are based on
Extended Kalman Filter and Maximum Likelihood Estimation,
respectively. We adopt a holistic approach to introduce security
awareness among state estimation algorithms without requiring
specialized hardware, or cryptographic techniques. We apply
SecSens to securely localize and time synchronize networked
mobile devices. SecSens provides good performance at run-time
several order of magnitude faster than the state of art solutions
under the presence of powerful attacks. Our algorithms are
evaluated on a testbed with static nodes and a mobile quadrotor,
all equipped with commercial ultra-wide band wireless devices.
Index Terms—Security, state estimation, Kalman filter, optimization, localization, time synchronization
I. I NTRODUCTION
Last decades have witnessed a proliferation in CyberPhysical Systems (CPS) for observing a variety of physical
phenomena and provide appreciate control actions. Consider,
for instance, a wireless sensor network which offers many advantages and services in emergency rescue, homeland security,
military operations, habitat monitoring, and home automation
services. Such a network typically consists of spatially distributed sensor nodes to monitor the state of network [1].
Furthermore, modern vehicles are another example for CPS
which present a heterogeneous system for sensing dynamics
and providing the required control actions [2]. The new trend
in vehicle design is to move from typical isolated control
systems to more open automotive architectures that would
introduce new services such as remote diagnostics and code
updates, and vehicle-to-vehicle communication in order to
improved quality and reliability of such systems [3].
To use these systems securely and efficiently, we need a
paradigm shift in algorithmic design. For instance, focusing
Fig. 1: Time dispersion of measurements. At time ti of step i
a measurement is taken on one link in the network, where all
the links are under time varying attacks.
only on the use of data encryption is the wrong approach to
assure CPS security. An interdisciplinary approach is needed
that brings a diverse set of disciplines to bear on the secure
design process of CPS algorithms. Such an approach for CPS
should take care of security vulnerabilities that are easily
exploitable in life-critical applications and lead to catastrophic
consequences. These security vulnerabilities target sensing
software, hardware, and physical signals due to the new interaction between information technology and physical world
[4]. Such security vulnerabilities result from the dramatic
increase in the set of sensors functionalities and system design
complexity.
Attacks launched in the cyber domain led to calamitous
consensuses during the past decades [5]. For instance, one
popular attack is the StuxNet attack on Supervisory Control
and Data Acquisition (SCADA) controllers system which is
used in industrial processes control [6], [7]. Other security
issues on SCADA networks are shown in [8]. Examples
of wireless network attacks are shown in [9]. Furthermore,
the fragility of the underlying system structure of today’s
vehicles are exploited to a reasonable extent in [10], [11],
which target Electronic Control Units (ECU). They showed
that today’s vehicles are not resilient against attacks mounted
on its internal components. Thus, an attacker can easily disrupt
the operation of critical functions of a vehicle. This problem
is, even more, emerging with the rise of self-driving cars
[12]. Also, attacks on analog sensors which have increasingly
become an indispensable part of many modern systems are
shown in [13]. Similarly, attacks on CPS could affect the
underlying infrastructure as shown in the Maroochy Water
breach [14]. Moreover, the vulnerability of drones is a very
dangerous threat if attackers can take control of these deadly
weapons [15], [16]. Also, consider the case of generating a
fake GPS signal that appears identical to those sent out by the
real GPS [17] to take over some unmanned aircraft [18].
While considerable research has explored CPS security using cryptographical techniques, it is not sufficient to ensure for
securing CPS. Such cyber-security techniques cannot protect
against compromised physical environment around a sensor
node, which may inject a corrupted signal. Therefore, there is
an emerging need for secure systems [19] while considering
new techniques. Thus, researchers came up with techniques
that address the problem of secure state estimation and intrusion detection under the sensor, actuators, and communication
network attacks. Secure state estimation enables to estimate
the state of the CPS from corrupted/attacked measurements,
while intrusion detection aims to detect the presence of sensor
attacks.
CPS offer critical services which would require meeting
stringent timing and maintaining some guarantees on the accuracy of the state estimation during operating hours. However,
CPS have power consumption constraints [20], limitations in
terms of bandwidth [21], computation limitations [22], and
generally limited resources. Thus, we envision a new paradigm
in the designing and evaluating secure estimation algorithms.
Such paradigm requires meeting the real-time requirements
and the resource constraints of CPS. Extensive computation
and high computational overhead to provide secure estimate
estimation are not accepted in our paradigm. To realize this
vision, we propose algorithms for secure state estimation in the
presence of measurement noise and modeling errors. We show
that our algorithms secure the system against physical attacks
manifested by additive disruptions on the measurements while
keeping in mind the practicality of our proposed algorithms.
Up to our knowledge, SecSens is the first solution that meets
the real time requirements of CPS in our considered scenarios.
An entirely representative application for secure state estimation is the localization and time synchronization problem. With the growing prevalence of wireless devices, it is
important to coordinate timing among IoT and to provide
contextual information, such as location. Also, position estimation is necessary for different fields such as military [23],
indoor and outdoor localization [24], security surveillance,
and wildlife habitat monitoring. Also, maintaining a shared
notion of time is critical to the performance of many CPS,
Big Science [25], swarm robotics [26], high-frequency trading
[27], and global-scale databases [28]. Localization and time
synchronization involves a significant amount of collaboration between individual sensors to perform complex signal
processing algorithms. This introduces a considerable amount
of vulnerability to the network estimation process. Also, the
problem is challenging due to its nonlinearity, and due to the
high process and measurement noises. When we deal with
time; we are expecting a high effect of modeling errors and
measurement noises. Therefore, we picked localization and
time synchronization as the application driver to illustrate our
proposed secure state algorithm. In this work, we revisit the
secure estimation problem from a modern viewpoint of sensor
networks, to aid in the design of more efficient algorithms
with reduced overheads.
Inspired by the above discussion, we made the following
key contributions in this paper:
• Proposing SecSens which consists of two secure state
estimation algorithms suitable for applications in wireless
networks with nonlinear time evolution and measurement
model. Our solutions are applicable to networks with
pairwise measurements.
• Showing that our algorithm is resilient against attacks
on the sensor measurements while meeting the real-time
requirements of CPS.
• Applying the proposed strategy in the problem of simultaneous localization and synchronization of spatially
distributed nodes in an ad-hoc network.
• The proposed algorithms are evaluated on a real testbed
using ultra-wideband wireless devices and a quadrotor,
representing a network of both static and mobile nodes.
The rest of the paper is organized as follows: Section
II gives an overview of the relevant work in the domain.
Section III defines the secure state estimation problem and
the threat model under our study. We define the core concept
behind SecSens in IV. We then go through our proposed
algorithms in Sections VI-A and VI-B. Section VII illustrates
the experimental setup and evaluates the proposed algorithm
on static and mobile network of nodes. Finally, Section VIII
lists some concluding and discussion remarks.
II. R ELATED W ORK
In this section, we present the related work on secure state
estimation and its application to simultaneous localization and
synchronization.
A. Secure State Estimation
Several recent works have studied the problem of secure
state estimation against sensor attacks in dynamical systems.
We categorize the work to the following subcategories:
1) Secure Estimation for Noiseless Systems: Graphtheoretic conditions for the detectability of attacks for a
noiseless system are shown in [29]. Fawzi et al. show the
impossibility of accurately reconstructing the state of a system
if more than half the sensors are attacked [30]. Pasqualetti
et al. [31] propose a fully decentralized solution for attack
Identification.
2) Secure Estimation with Bounded non-stochastic Noise:
Satisfiability-Modulo-Theory (SMT) approach is proposed for
bounded noise in [32]. Another kind of work was based on
brute force search which is not applicable for large-scale
systems [33]. A practical solution is proposed in [34] that
considers real-time issues and synchronization errors.
3) Secure state estimation with Gaussian noise: Observing
and recording sensor readings for an amount of time, then repeat them afterward while carrying out an attack is commonly
defined as replay attack. Such kind of attack is considered in
[35] where all the existing sensors were attacked. The attacker
does not know model knowledge, but he can access and corrupt
the sensor data. Another work considered stochastic game
anachronism for detecting replay attacks [36]. Also, denial of
service attacks solution under Gaussian noise is proposed in
[37]. Furthermore, false data injection is solved by proposing
an ellipsoidal algorithm where attacker’s strategy is formulated
as a constrained control problem [38].
All the previous work assume linear state space model.
Thus, it can not be used with nonlinear state space. Recently,
Weber et al. [39] proposed Gordian SMT especially for secure
localization which is a non-linear state space. However, they
did not propose as a general solution for non-linear secure state
estimation of networks with pairwise measurements. Also,
during their evaluation, they only considered four links under
attack, which is very small number compared to the total
number of links. Furthermore, they required about 11 minutes
to be able to provide a secure state estimation which is a huge
overhead to recover a network state. Finally, they considered
a 2D localization problem only. On the other hand, SecSens
considers the non-linear state space model of networks with
pairwise measurements, while taking into consideration the
real-time requirements for CPS and assumes powerful attacks
comparing to the related work.
B. Secure Localization and Time Synchronization
Researchers attach great importance for secure localization
from a long time. For instance, SeRLoc [40] proposed a
localization scheme based on a two-tier network architecture.
Another type of work was based on partial homomorphic
encryption to achieve private and resilient localization [41].
For more details on secure localization, we refer the reader to
[42], [43]. On the other hand, secure time synchronization
alone got a lot of attention. TinySeRSync [44], for example, proposes secure single-hop pairwise time synchronization
using authenticated timestamping. Moreover, authors in [45]
analyze attacks on different time synchronization protocols and
propose a secure time synchronization toolbox to counter these
attacks. In [46] a holistic approach to maintain secure time
synchronization in a wireless network is proposed, and in [47]
pairwise secure synchronization concept is proposed tailored
to real-time applications in smart grids. For more details on
secure synchronization, we refer the reader to [48], [49].
Although these two problems, namely secure time synchronization and secure localization, share many aspects in
common, they are traditionally treated separately. That results
in a lower performance, inefficient use of resources, and nonscalable algorithms. Secnav [50] is the only work that covers
simultaneous secure localization and time synchronization up
to our knowledge at the time of writing. However, Secnav
assumes that sensor nodes cannot be compromised by an
adversary, and cares only about the attacked communication
channel. In contrast, SecSens considers attacks which compromise sensor nodes and links.
III. S ECURE S TATE E STIMATION P ROBLEM
In what follows, we discuss our threat model and the utility
measures. Then, provide a mathematical formulation of the
state estimation problem.
A. Threat Model
In the context of attack-resilient state estimation, we have
the following entities:
• Estimation Server: It attempts to securely, efficiently and
accurately estimate the state of a network of distributed
sensors given that we have measurements and modeling
noises. Also, sensors and links are compromised.
• Sensor Nodes: Entities that make some required sensing
in order to help in estimating the network state. The
adversary could compromise nodes.
• Adversary: He modifies the framework with the goal
of causing the server to misestimate the state of the
network. He can compromise sensor/links to corrupt
measurements. Denial of service attack is out of the scope
of our paper. Therefore, the receiver will still receive
the message from the sender, superimposed by the attacker’s messages. Also, we can look into the problem as
compromised messages where man-in-the-middle attack
(MITM) is presented to corrupt the delivered messages
to the server.
B. Resilience Metric
The accuracy of state estimation algorithm is the key for
its use in mission-critical systems. Besides, the CPS are often
resource-constrained and has to meet some time requirements.
To accommodate the above practical constraints while still
maintaining the desired resilience level we define the following
utility metrics:
• Estimation Error: The difference between the actual
network state (ground truth) and the estimated state.
• Execution Time: The time needed by the server to
estimate the network state based on the measurements.
C. System and Attack Model
Consider the following nonlinear time-varying system
x̃ki+1 = f˜(x̃ki ) + ñki
yik,j = h̃j (x̃ki , x̃ji ) + ṽik + ak,j
i ,
(1)
where xki is the state of node k at time i. yik,j is the measurement sent to node k from the neighborhood node j ∈ Nk . The
process and measurement noise are assumed to be uncorrelated
and zero mean white Gaussian noise. They are denoted by nki
and vik , respectively. Qi and Ri are the process covariance
and the measurement noise matrix at time i, respectively. The
attack vector ak,j
is a vector which models how an attacker
i
changes the sensor measurements between node k and node
j at time i. A non-zero elements in the vector a corresponds
to the attacked values on the corresponding sensor; otherwise
the measurement is not attacked. Thereby, the additive attack
vector ak,j
can account for both, a malicious node k and
i
a corrupted link (k, j). Moreover, the attack values can be
constant or time-varying.
SecSens considers secure state estimation problem over a
network of N nodes indexed by k ∈ {0, . . . , N − 1}. The
nodes are spatially distributed over some region in space. We
say that two nodes are connected if they can communicate
directly with each other. The neighborhood of a given node k
is denoted by the set Nk .
IV. S EC S ENS C ORE I DEA
Most of the related work on secure state estimation mitigates
the influence of attacks as described in Section II by identifying the attacked links and excluding them from the estimation
process. This is a highly complex combinatorial problem, and
even the convex relaxation techniques are not suitable for realtime applications. So, let’s go through SecSens core idea that
helps us to solve the problem while considering the practical
requirements.
To understand the idea behind SecSens, let’s consider the
following toy example which is representative of the core idea
behind our solution. Consider a network of N nodes, where
each node has a link with its neighbor. Now, let’s consider
Figure 1, node1 will communicate with node2 at t1 . Then
node1 will communicate with node2 and node3 at time t2
and t3 , respectively. Now, let’s go back to the general model
of the secure state estimation which is shown in Equation 1.
According to that model, we have three attacks which are
4,1
a12,1 , a3,1
on the three links. Related work tried to
2 , and a3
solve at each time instant t the secure state estimation problem
while considering the number of possible attacks (unknowns)
equals the number of links. It ends up in a very complicated
problem with the number of unknowns equals the number of
connections.
In contrast, we look into the problem from a different
perspective inspired by the real-life application’s constraints.
At each time instant, t each node can only communicate with
one neighbor and this is the practical scenario in any CPS.
For instance, measuring the distance using time of the flight
of two or three messages between two nodes is a type of
communication [51]. Thus, at time t1 in Figure 1, why should
we care about estimating a3,1
1 ?. What about renaming the three
attacks on the figure to a11 , a12 , and a13 as the couple time
instant ti and the node index k uniquely defines the neighbor
j because node k can only communicate with one node at
time ti . Thus, SecSens looks into the attack that is related to
node1 as a1t and deals with it as time-varying value. With this
paradigm shift, we are able to have the number of attack at
each time instant equals to the number of nodes, instead of
the number of links. Therefore, SecSens drastically reduces
the computational complexity. We will show the performance
of SecSens while attacking all the links with time-varying
attacks. State estimation while attacking all the links is a hard
problem and could not be solved in any of the related work
up to our knowledge.
We now utilize this idea to change the general model in (1).
We include the attack value of this time step to the state of the
node initiating the measurement. Following this procedure, the
state of node k would be extended to xki = [x̃ki T , aki T ]T . This
yields a modified system model with xki+1 = f (xki ) + nki and
yik = hj (xki , xji ) + vik . In this way, we decrease the number
of unknowns at each time step. Instead of being proportional
to the number of links, they are now only proportional to
the number of nodes. We should emphasize that this is just
a modeling trick and all the links can be under attack with
different values.
A valid question would be how to model the time-evolution
in the state-space model for unpredictable static or timevarying attacks. We choose to have aki = aki+1 , and the
variance of the attack entries of the modeling noise nki
would account for the time-varying aspect of the attack. We
would like to define two important concepts associated with
measurement values:
•
•
Outliers: These measurement values lie outside the main
expected range. For example, it would be the measured
distance between two nodes that is much larger than diagonal of the localization area. Also, if the measured speed
of moving node is higher than the hardware capabilities of
the nodes, then this speed value would be considered an
outlier. These Outliers can be easily detected and rejected
by setting some thresholds on the accepted values.
Attacks: the Smart attacker would corrupt the sensor
reading with new values that make sense. Thus, attacks
are values that can’t be removed by outliers’ detectors.
Attacked speed value, for example, would be within the
capabilities of the node motors.
SecSens makes use of this concept to simplify the complexity of secure state estimation algorithms. We care about
designing secure state estimation algorithm which would be
combined with outliers rejectors. Thus, attacks can only be in
a limited interval which can be represented by the modeling
variance. We used this concept in our previous work [51], [52]
where we use stationary model for process update function for
tracking a flying quadrotor. As long as the quadrotor moves
in the range of the modeling noise, D-SLATS would be able
perfectly to localize it. At the end of the day; the speed of
the quadrotor is limited by the hardware capabilities which
are public information. So why we do not make use of this
information to simplify the estimation algorithm?
In what follows, this approach will be applied to simultaneous localization and synchronization. It will be seen that
the system performance is greatly improved by this approach
while the computational complexity is significantly reduced
compared to the related work.
V. S EC S ENS IN R EAL L IFE A PPLICATION
VI. S EC S ENS ATTACK R ESILIENT S TATE E STIMATORS
Without loss of generality, we choose to apply SecSens
on simultaneous localization and time synchronization estimation problem while some nodes/links are under attack.
It a quite representative example of secure state estimation
problem where we have modeling noise of the clocks and the
dynamics of moving nodes in conjunction with measurements
noise. Our sensors are able to get the time difference and
the distance between two each other. More specifically, we
consider three types of measurements which are distinguished
by the number of messages exchanged between a pair of nodes.
The measurement vector at node k, from node j ∈ Nk has
the following form
k
dk,j
i + aoi
rk,j + adk ,
i
i
Rik,j + adki
yik,j
=
(2)
where dk,j
represents the counter difference at time i denoted
i
by dk,j
which is the measurement of the difference between
i
the clocks of each node. rik,j , on the other hand, represents
a noisy measurement due to frequency bias discrepancies
between k and j which is formally represented by singlesided two-way range. Finally, Rik,j is another distance
measurement between nodes k and j based on a trio of
messages between the nodes at time i. This is a more accurate
estimate than rik,j due to mitigation of frequency bias errors
from the additional message. It is formally called doublesided two-way range. For more details about the three types
of measurements, we encourage the reader to check [51].
However, these measurements are corrupted by some attacks
on counter difference and distance, namely, aoki and adki ,
respectively. We would like to note that subset of these
measurements may be used rather than the full set, i.e., we
can have experiments involving just rik,j , Rik,j , dk,j
or any
i
combinations. Also, we do not put any limitation on the way
of calculating these measurements. So, no special requirements
on the used hardware.
Thus, we are concerned with a state vector which consists
of three dimensional position vector pki , clock time offset
oki , and clock frequency bias bki for every node. We adopt
a convention where both oki and bki are described with respect
to the global time clock which is usually the clock of master
node, which can be any node. Also, there are attacks on the
distance measurements, which is denoted by adki , and the
attack on offset measurement values aoki which are added by
an attacked node. In summary our state vector is as following:
h
iT
xki = pki T oki bki aoki adki
(3)
We define the full network state xi at master time as the
concatenation of all states from the N participating devices at
time step i:
−1 T T
(4)
xi = x0i T x1i T .. xN
i
A. SecEKF Algorithm
The core idea behind SecEKF is to make use of Extended
Kalman filter (EKF) in order to estimate the network state
beside the introduced attacks. We periodically apply corrections when new measurements are received. We assume the
clock parameters evolve according to the first-order affine
approximation of the following dynamics oki+1 = oki + bki δt
and bki+1 = bki , where δt := tM (i + 1) − tM (i) given that
tM is the root node time which is the global time. The
process update is responsible for evolving the state estimates
for all N devices—clock offsets, frequency biases, and device
position. Therefore, we can write the process update function
as following:
pki
oki + bki δt
k
k
(5)
f (xi ) =
bi k
aoi
adki
Note that this simple process model can still be used to
capture node state dynamics and changing attack values simply
by setting the covariance of the corresponding state value
adequately high to account for high change rate. We will show
that our module can work with high rate changing attack and
moving nodes as well. On the other hand, The measurement
update step relates the state variables xki to a measurement
vector yik,j between two devices. Each of our measurements
depends on the of local clock values, distances between two
devices and the amount of the introduced attack on time and
distance. For any given node k in the network, the nodespecific measurement model with respect to a second node
j is given as:
oji − oki + aoki
j
hj (xki , xji ) = (1 + bki ) pi − pki + adki
(6)
2
j
k
(1 + bki ) pi − pki + adi
2
The time synchronization and localization problem aims
to estimate the clock parameters and the 3D position using
the available measurements and the model of the system. We
denote by x̂i the estimate of xi where we seek to minimize
the mean-square error Ekxi − x̂i k2 . Given this, we design
an EKF to get a secure state estimate over time, periodically
applying corrections when new timing and range measurements are received. Our EKF only updates upon arrival of a
new measurement (message) exchanged between two or more
nodes. The EKF steps are:
• Receive a new measurement: The server obtains a
new measurement at filter step i which is generated
by a message exchange between devices k and j. By
convention, communication is always defined to terminate
at device j. One of the nodes maybe under attack and
corrupt the sent measurement.
•
•
Calculate δt : The server maintains an estimate of master
time tM (i) as well as time offset and frequency bias
estimates for every node. This gives the capability to
calculate the time elapsed between measurements in the
master time reference.
Time update: Update the current secure estimates for the
location and time of each node while taking the attack
effect out of the measurement.
B. SecOPT Algorithm
j,k
i0 ∈{i−L,i}
where a time window comprising L measurements is used
for minimization. The L1-norm on the attack values is added
to quadratic minimization criteria for more robust attack
estimation. The resulting method is summarized in Algorithm
1. The algorithm is centralized in the sense that all the
exchanged messages are transmitted to a centralized server
where the computation and optimization are solved. While the
minimization is initialized for i = 0 with the state xinit = 0
for numerical optimization, run repeat the it multiple times by
setting xinit to the previously estimated state.
Algorithm 1 SecOPT
Set i = 0, and x̂0 = xinit . At the centralized server:
Step 1: Collect enough measurements between nodes
according to the window size.
Step 2: Solve the centralized optimization problem with
solver initial condition set to x̂i :
h
i
X k,j
ôji+1 p̂ji+1 ao
ˆ ji+1 = argmin
(di − (oki − oji )
oji ,pji ,aoji j∈Nk
h
p̂ji+1
We demonstrate SecSens state estimation algorithms in
simultaneous localization and time synchronization of wireless sensor nodes under attack while keeping on minds the
resilience metrics as defined in Subsection III-B. This attack
can be on the communication link or/and the whole sensor
node. We are going to evaluate our proposed algorithms under
different attack cases for static and mobile nodes.
C. Case Study: Static Nodes
The second algorithm for secure state estimation is based
on the Maximum Likelihood Estimator (MLE). By using
the model of the system and the knowledge of how the
unknown attack signal affects the measurements, we formulate
an optimization problem as
X
2
k
khj (xki0 , xji0 ) − yik,j
,
arg min
0 k − λkai0 k1
k
xk
i ,ai ∀k,j
B. Experiments
− pki − pji /c − aoji )2 + λkaoji k1
2
i
X k,j
j
ˆ
adi+1 = argmin
(Ri − pki − pji
bji ,pji ,adji j∈Nk
− adji )2 + λkadji k1
2
Step 3: Increment i by 1 and go back to Step 1
VII. E VALUATION
A. Experimental Setup
We evaluated the performance of the proposed secure state
estimation algorithms in a custom ultra-wideband RF testbed
based. The overall setup is described in details in the Appendix
and shown in Figure 6.
Nodes are placed in 8 distinct locations around the 10 × 9
m2 area, roughly in the positions indicated by Figure 6. The
goal of our algorithm is to accurately and securely estimate the
positions of all network devices relative to each other, as well
as, the relative clock offsets. This relative localization (graph
realization), is a well-researched field. Local minima, high
computational complexity, and restrictions on graph rigidity
are the main challenges in graph realization problems [53],
[54]. SecSens does not put restrictions on the connectivity
of the graph as we will show. An original extended Kalman
filter (OrigEKF) is implemented as a baseline for comparison
to show the criticality of estimating the attacks values in CPS.
1) Secure Position Estimation: We would like to start with
defining the attacker strategy to corrupt the network state
estimation. The common way in secure estimation field is
to consider a uniform random number or constant values as
attacks value without putting constraints on the rate of change
of the attacks values. Neglecting the convergence time of the
algorithm has catastrophic consensuses. Unlike the previous
work, we consider three types of attacks generator statistics.
Also, we consider the attacker switch between different parameters during an attack session. Function 1 in the Appendix
shows a Pseudocode for our attacker function. We consider
Uniform, Normal and Pareto Distributions as type 1, 2, and
3 attacks, respectively. Note that generating the attack values
according to Function 1 simulates time-varying attack values
that are different for all links. Also, we should highlight the
following points:
• Every new measurement gets different attack value, i.e,
the Function getDistanceAttackValue is called every
time instance for every new measurements.
• All the measurements under attack.
• The attacker changes the parameters of each type from
period to period as shown in the Pseudocode. Such
an attack is stronger than using fixed statistics random
generator.
• The attacks values must results in a location within the
localization area, as a smart attacker strategy. Otherwise,
the attack can be easily detected with a simple threshold
mechanism.
The proposed secure algorithms find relative positions.
Therefore, we first superimpose the estimated positions onto
the true positions of each node by use of a Procrustes transformation [55]. Specifically, the estimated network topology as a
whole is rotated and translated without scaling, until it most
closely matches the true node positions. Once transformed, the
error of a given node’s position is defined as the `2 norm of the
transformed position minus the true position. Figure 2 shows a
twenty seconds snapshot of the performance after running the
three types of attacks. We can notice the jump in the OrigEKF
behaviors due changing the attack statistics. However, SecSens
is resilient against such type of attacks. Table I summarizes
the localization error for every node using the three algorithm.
SecEKF outperforms SecOPT in the three types of attacks.
2) Secure Time Synchronization: We considered two attacker strategies in order to corrupt the network state estimation in term of time, for the space constraint. We define Type 4
attack where each node adds a constant value in order ∼ 100µ
seconds to measurement di j. While, Type 5 attack considers
adding uniform random values in order ∼ 100µ seconds to
measurement di j. The reason behind choosing these types of
attacks is discussed in the Appendix. We should highlight that
different attacked nodes add different attack values for both
Type 4 and Type 5. As mentioned before, we choose node 0
to be the reference node, i.e, b0i := 0 and o0i := 0 ∀i. Table II
shows the synchronization errors for all nodes with respect to
node 0. We also assume that the reference node is free of attack
which is not the same assumption in the secure localization
scenario. In order to test the secure time synchronization with
minimum uncertainty in our testing mechanism, we follow the
testing strategy in [56], and [57].
SecEKF comes with the best performance, then SecOPT
for both attack types. We should note that the synchronization
errors in Table II are reported for a fully connected network.
SecEKF algorithm achieved 0.03 and 0.25 µ seconds for Type
4 and Type 5 attack, respectively. On the other hand, SecOPT
algorithm achieved 0.34 and 9.10 µ seconds for Type 4 and
Type 5 attacks, respectively. Neglecting the secure estimation
techniques leads to catastrophic consequences as shown for
OrigEKF performance against Type 4 and Type 5 attacks. The
synchronization error is in order of attacks values, namely, 121
and 132 µ seconds for Type 4 and Type 5 attack, respectively.
D. Case Study: Mobile Nodes
We have shown that our proposed algorithms can be used to
localize and synchronize a network of static nodes under three
types of attacks. We now present the case of a heterogeneous
network containing both static and mobile nodes. We add one
mobile node in the form of a CrazyFlie quadrotor to the eight
anchor nodes topology used in Section VII-C, as shown in
Figure 6. We analyzed the results of running SecEKF, SecOPT
and OrigEKF to localize the mobile node flying in our lab.
Again, all the links are under attacks. For space limit, we will
show the results under attack 1 scenario only.
Some experiments were performed with quadrotors traveling with variable velocities. Figure 3 show the results of
localizing the traveling quadrotors using SecEKF, SecOPT
and OrigEKF algorithms. SecEKF achieved the best secure
localization estimation with a mean of 1.38 m and standard
deviation of 0.72 m. On the other hand, SecOPT reported
1.82m mean error and standard deviation of 0.78 m. If the
normal estimation does not consider attacks estimation, it will
suffer from 7.03m mean error and 3.54m standard deviation.
We repeated the experiments for the network topology
where every node is only connected to four neighbors instead
of eight nodes in the fully connected scenario, which means we
are dropping about 27 links out of 72 links. The performance
degraded as expected. As shown in Fig 4, SecEKF, SecOPT
and OrigEKF algorithms reported 3.48m, 3.36m, and 10.07m
mean localization error, respectively.
E. Case Study: Execution Time
Previous work [39] could achieve secure state estimation in
case of localization in about 11 minutes. They only consider
2D localization and only four links under attack. Eleven
minutes overhead is a very high one especially for the moving
nodes. We are going during this case study to analyze the
required execution time for the proposed secure estimation
algorithms to localize the CrazyFlie quadrotor, where all the
links are under Type 1 attack. The execution of SecOPT
depends on the number of considered measurements during
one run of the algorithms which we called the window
size. Figure 5 summarizes the 3D localization error for each
window size. For a window size of 300 measurements, we
could achieve a mean error of 1.59m at the cost of 4.6 seconds
of execution time. When we decreased the window size to 50
measurements, we get 4.7m localization error at the cost of
1.1 seconds of execution time. On the other hand, we can see
the SecEKF could achieve a mean error of 1.28m with a cost
of 0.004 seconds. SecEKF considers only a window size of
one measurement as it depends on EKF.
VIII. C ONCLUSION
This work has investigated the security aware-aspect of
the estimation algorithms for the multi-sensor system. More
specifically, we proposed SecEKF and SecOPT secure estimation algorithms for the wireless network where state evolution
and measurement models can follow nonlinear functions.
Then, we picked simultaneous localization and synchronization as a representative application for our secure estimation
problem. Several experiments in a network with static nodes
and a mobile quadrotor, all equipped with commercial ultrawideband wireless radios, were conducted. Our empirical
results indicate that SecSens offers reliable performance and
efficient computational procedures for the secure state estimation. Our proposed algorithm made it possible by stateof-the-art advances in commercial ultra-wideband radios to
securely estimate the state in asynchronous networks. Future
directions will deal with testing over a large-scale system and
more sophisticated attack scenarios.
R EFERENCES
[1] J. Yick, B. Mukherjee, and D. Ghosal, “Wireless sensor network survey,”
Computer networks, vol. 52, no. 12, pp. 2292–2330, 2008.
[2] U. Kiencke and L. Nielsen, “Automotive control systems: for engine,
driveline, and vehicle,” 2000.
(a) SecEKF
(b) SecOPT
(c) OrigEKF
(d) SecEKF
(e) SecOPT
(f) OrigEKF
(g) SecEKF
(h) SecOPT
(i) OrigEKF
Type 3 Type 2 Type 1
Fig. 2: Localization error for the eight nodes where all the measurements are under attack with different values. Sub-figures
(a), (b) and (c) correspond to Type 1 attack which is time varying uniform random number. Sub-figures (d), (e) and (f)
correspond to Type 2 attack is time varying normal distribution. Sub-figures (g), (h) and (i) correspond to Type 3 attack which
is following time varying Pareto Distribution. Function getDistanceAttackValue shows the attacker behavior for corrupting
the measurements Ri j and ri j.
Algorithm
SecEKF
SecOPT
OrigEKF
SecEKF
SecOPT
OrigEKF
SecEKF
SecOPT
OrigEKF
node 0
0.655
0.607
4.744
1.343
0.874
2.546
1.107
1.631
3.206
node 1
0.177
0.668
6.588
0.921
0.973
4.134
0.960
1.692
5.332
node 2
0.720
0.650
4.251
1.016
0.767
0.718
0.849
1.210
2.450
node 3
0.572
0.606
5.787
0.864
0.832
4.579
1.596
1.635
5.901
node 4
0.154
0.424
1.665
0.390
0.767
1.213
0.444
1.053
2.856
node 5
0.229
0.553
3.106
0.675
0.984
2.531
1.314
1.769
2.649
node 6
0.134
0.400
3.048
0.323
0.675
3.479
0.633
1.476
2.825
node 7
0.387
0.713
5.711
0.942
1.288
4.465
0.627
1.267
5.409
mean
0.378
0.578
4.362
0.809
0.895
2.958
0.941
1.467
3.829
std
0.240
0.113
1.671
0.336
0.190
1.462
0.386
0.260
1.448
TABLE I: Localization error for the eight nodes where all the measurements are under Type 1, 2, and 3 attack with different
values. Function getDistanceAttackValue shows the attacker behavior for corrupting the measurements. The three algorithms
are running in parallel on the same set of attacks values to have a fair comparison.
Type 5 Type 4
Algorithm
SecEKF
SecOPT
OrigEKF
SecEKF
SecOPT
OrigEKF
node 1
0.028
0.232
98.424
0.025
4.95
162.368
node 2
0.082
0.633
198.207
0.865
1.07
104.922
node 3
0.014
0.230
297.957
0.133
22.55
136.901
node 4
0.003
0.265
97.891
0.197
2.65
186.414
node 5
0.027
0.065
97.811
0.318
22.02
156.470
node 6
0.097
0.535
97.781
0.381
1.93
149.342
node 7
0.0507
0.791
82.600
0.1123
17.63
167.329
mean
0.038
0.344
121.334
0.254
9.101
132.968
std
0.035
0.279
89.050
0.279
9.842
58.803
TABLE II: Synchronization error (µ seconds) of different nodes with respect to node 0. Type 4 attack considers adding constant
values of order ∼ 100µ seconds to measurement di j. Type 5 attack considers adding uniform random values of order ∼ 100µ
seconds to measurement di j. The three algorithms are running in parallel on the same set of attacks values to have a fair
comparison.
Fig. 3: The localization error of the CrazyFlie quadrotor
where all the measurements are under Type 1 attack with
different values in a Fully connected network. Function
getDistanceAttackValue shows the attacker behavior for corrupting the measurements Ri j and ri j. The three algorithms
are running in parallel on the same set of attacks values to
have a fair comparison.
[3] H. Heinecke, K.-P. Schnelle, H. Fennel, J. Bortolazzi, L. Lundh,
J. Leflour, J.-L. Maté, K. Nishikawa, and T. Scharnhorst, “Automotive
open system architecture-an industry-wide initiative to manage the
complexity of emerging automotive e/e-architectures,” Convergence, pp.
325–332, 2004.
[4] A. A. Cárdenas, S. Amin, and S. Sastry, “Research challenges for the
security of control systems.” in HotSec, 2008.
[5] M. Uma and G. Padmavathi, “A survey on various cyber attacks and
their classification.” IJ Network Security, vol. 15, no. 5, pp. 390–396,
2013.
[6] D. Kushner, “The real story of stuxnet,” IEEE Spectrum, vol. 50, no. 3,
pp. 48–53, 2013.
[7] R. Langner, “Stuxnet: Dissecting a cyberwarfare weapon,” IEEE Security
& Privacy, vol. 9, no. 3, pp. 49–51, 2011.
[8] V. M. Igure, S. A. Laughter, and R. D. Williams, “Security issues in
scada networks,” Computers & Security, vol. 25, no. 7, pp. 498–506,
2006.
[9] D. Martins and H. Guyennet, “Wireless sensor network attacks and
security mechanisms: A short survey,” in Network-Based Information
Systems (NBiS), 2010 13th International Conference on. IEEE, 2010,
pp. 313–320.
Fig. 4: The localization error of the CrazyFlie quadrotor where
all the measurements are under Type 1 attack with different
values in a partial connected network where every node has
five neighbors only. Function getDistanceAttackValue shows
the attacker behavior for corrupting the measurements Ri j and
ri j. The three algorithms are running in parallel on the same
set of attacks values to have a fair comparison.
[10] S. Checkoway, D. McCoy, B. Kantor, D. Anderson, H. Shacham,
S. Savage, K. Koscher, A. Czeskis, F. Roesner, T. Kohno et al.,
“Comprehensive experimental analyses of automotive attack surfaces.”
in USENIX Security Symposium. San Francisco, 2011.
[11] K. Koscher, A. Czeskis, F. Roesner, S. Patel, T. Kohno, S. Checkoway,
D. McCoy, B. Kantor, D. Anderson, H. Shacham et al., “Experimental
security analysis of a modern automobile,” in Security and Privacy (SP),
2010 IEEE Symposium on. IEEE, 2010, pp. 447–462.
[12] J.-P. Hubaux, S. Capkun, and J. Luo, “The security and privacy of smart
vehicles,” IEEE Security & Privacy, vol. 2, no. 3, pp. 49–55, 2004.
[13] D. F. Kune, J. Backes, S. S. Clark, D. Kramer, M. Reynolds, K. Fu,
Y. Kim, and W. Xu, “Ghost talk: Mitigating emi signal injection attacks
against analog sensors,” in Security and Privacy (SP), 2013 IEEE
Symposium on. IEEE, 2013, pp. 145–159.
[14] J. Slay and M. Miller, “Lessons learned from the maroochy water
breach,” Critical infrastructure protection, pp. 73–82, 2007.
[15] M. Mohan, “Cybersecurity in drones,” Ph.D. dissertation, Utica College,
2016.
[16] N. Shachtman, “Computer virus hits us drone fleet,” CNN. com, 2011.
[17] A. H. Rutkin, “spoofers use fake gps signals to knock a yacht off course,”
MIT Technology Review, 2013.
[18] A. J. Kerns, D. P. Shepard, J. A. Bhatti, and T. E. Humphreys,
“Unmanned aircraft capture and control via gps spoofing,” Journal of
Fig. 5: The execution time and its corresponding localization
error of the CrazyFlie quadrotor with different window size
of measurements, where all the measurements are under Type
1 attack with different values in a Fully connected network.
Function getDistanceAttackValue shows the attacker behavior for corrupting the measurements Ri j and ri j.
Field Robotics, vol. 31, no. 4, pp. 617–636, 2014.
[19] C. Hartung, J. Balasalle, and R. Han, “Node compromise in sensor
networks: The need for secure systems,” Department of Computer
Science University of Colorado at Boulder, 2005.
[20] A. J. Goldsmith and S. B. Wicker, “Design challenges for energyconstrained ad hoc wireless networks,” IEEE wireless communications,
vol. 9, no. 4, pp. 8–27, 2002.
[21] G. J. Pottie and W. J. Kaiser, “Wireless integrated network sensors,”
Communications of the ACM, vol. 43, no. 5, pp. 51–58, 2000.
[22] K. Kumar, J. Liu, Y.-H. Lu, and B. Bhargava, “A survey of computation
offloading for mobile systems,” Mobile Networks and Applications,
vol. 18, no. 1, pp. 129–140, 2013.
[23] J. R. Lowell, “Military applications of localization, tracking, and targeting,” IEEE Wireless Communications, vol. 2, no. 18, pp. 60–65, 2011.
[24] K. Lingemann, H. Surmann, A. Nuchter, and J. Hertzberg, “Indoor and
outdoor localization for fast mobile robots,” in Intelligent Robots and
Systems, 2004.(IROS 2004). Proceedings. 2004 IEEE/RSJ International
Conference on, vol. 3. IEEE, 2004, pp. 2185–2190.
[25] M. Lipiński, T. Włostowski, J. Serrano, and P. Alvarez, “White rabbit: A
ptp application for robust sub-nanosecond synchronization,” in Precision
Clock Synchronization for Measurement Control and Communication
(ISPCS), 2011 International IEEE Symposium on. IEEE, 2011, pp.
25–30.
[26] G. Regula and B. Lantos, “Formation control of a large group of uavs
with safe path planning,” in Control & Automation (MED), 2013 21st
Mediterranean Conference on. IEEE, 2013, pp. 987–993.
[27] P. V. Estrela and L. Bonebakker, “Challenges deploying ptpv2 in a global
financial company,” in 2012 IEEE International Symposium on Precision
Clock Synchronization for Measurement, Control and Communication
Proceedings. IEEE, 2012, pp. 1–6.
[28] J. C. Corbett, J. Dean, M. Epstein, A. Fikes, C. Frost, J. J. Furman,
S. Ghemawat, A. Gubarev, C. Heiser, P. Hochschild et al., “Spanner:
Google’s globally distributed database,” ACM Transactions on Computer
Systems (TOCS), vol. 31, no. 3, p. 8, 2013.
[29] F. Pasqualetti, F. Dorfler, and F. Bullo, “Control-theoretic methods for
cyberphysical security: Geometric principles for optimal cross-layer
resilient control systems,” IEEE Control Systems, vol. 35, no. 1, pp.
110–127, 2015.
[30] H. Fawzi, P. Tabuada, and S. Diggavi, “Secure estimation and control for
cyber-physical systems under adversarial attacks,” IEEE Transactions on
Automatic Control, vol. 59, no. 6, pp. 1454–1467, 2014.
[31] F. Pasqualetti, F. Dörfler, and F. Bullo, “A divide-and-conquer approach
to distributed attack identification,” in Decision and Control (CDC),
2015 IEEE 54th Annual Conference on. IEEE, 2015, pp. 5801–5807.
[32] Y. Shoukry, P. Nuzzo, A. Puggelli, A. L. Sangiovanni-Vincentelli, S. A.
Seshia, and P. Tabuada, “Secure state estimation for cyber physical
systems under sensor attacks: a satisfiability modulo theory approach,”
IEEE Transactions on Automatic Control, 2017.
[33] M. S. Chong, M. Wakaiki, and J. P. Hespanha, “Observability of linear
systems under adversarial attacks,” in American Control Conference
(ACC), 2015. IEEE, 2015, pp. 2439–2444.
[34] M. Pajic, J. Weimer, N. Bezzo, P. Tabuada, O. Sokolsky, I. Lee, and G. J.
Pappas, “Robustness of attack-resilient state estimators,” in ICCPS’14:
ACM/IEEE 5th International Conference on Cyber-Physical Systems
(with CPS Week 2014). IEEE Computer Society, 2014, pp. 163–174.
[35] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in
Communication, Control, and Computing, 2009. Allerton 2009. 47th
Annual Allerton Conference on. IEEE, 2009, pp. 911–918.
[36] F. Miao, M. Pajic, and G. J. Pappas, “Stochastic game approach for
replay attack detection,” in Decision and control (CDC), 2013 IEEE
52nd annual conference on. IEEE, 2013, pp. 1854–1859.
[37] S. Amin, A. A. Cárdenas, and S. Sastry, “Safe and secure networked
control systems under denial-of-service attacks.” in HSCC, vol. 5469.
Springer, 2009, pp. 31–45.
[38] Y. Mo, E. Garone, A. Casavola, and B. Sinopoli, “False data injection
attacks against state estimation in wireless sensor networks,” in Decision
and Control (CDC), 2010 49th IEEE Conference on. IEEE, 2010, pp.
5967–5972.
[39] M. Weber, B. Jin, G. Lederman, Y. Shoukry, E. A. Lee, S. A. Seshia,
and A. L. Sangiovanni-Vincentelli, “Gordian smt: Untangling ranging
attacks in noisy sensor networks for secure localization,” 2017.
[40] L. Lazos and R. Poovendran, “Serloc: Robust localization for wireless
sensor networks,” ACM Transactions on Sensor Networks (TOSN),
vol. 1, no. 1, pp. 73–100, 2005.
[41] A. Alanwar, Y. Shoukry, S. Chakraborty, P. Martin, P. Tabuada, and
M. Srivastava, “Proloc: Resilient localization with private observers
using partial homomorphic encryption,” in Proceedings of the 16th
ACM/IEEE International Conference on Information Processing in Sensor Networks, ser. IPSN ’17. New York, NY, USA: ACM, 2017, pp.
41–52.
[42] A. Srinivasan and J. Wu, “A survey on secure localization in wireless
sensor networks,” Encyclopedia of Wireless and Mobile communications,
p. 126, 2007.
[43] A. Boukerche, H. A. Oliveira, E. F. Nakamura, and A. A. Loureiro,
“Secure localization algorithms for wireless sensor networks,” IEEE
Communications Magazine, vol. 46, no. 4, 2008.
[44] K. Sun, P. Ning, and C. Wang, “Tinysersync: Secure and resilient time
synchronization in wireless sensor networks,” in Proceedings of the 13th
ACM Conference on Computer and Communications Security, ser. CCS
’06. New York, NY, USA: ACM, 2006, pp. 264–277.
[45] S. Ganeriwal, C. Pöpper, S. Čapkun, and M. B. Srivastava, “Secure time
synchronization in sensor networks,” ACM Transactions on Information
and System Security (TISSEC), vol. 11, no. 4, p. 23, 2008.
[46] J. Ponniah, Y. Hu, and P. R. Kumar, “A clean slate design for secure
wireless ad-hoc networks - part 2: Open unsynchronized networks,” in
Int. Symp. Modeling and Optimization in Mobile, Ad Hoc, and Wireless
Networks (WiOpt), Mumbai, India, May 2015, pp. 183–190.
[47] A. Aichhorn, B. Etzlinger, S. Hutterer, and R. Mayrhofer, “Secure
communication interface for line current differential protection over
ethernet-based networks,” in IEEE Manchester PowerTech, Jun. 2017.
[48] H. Yih-Chun and A. Perrig, “A survey of secure wireless ad hoc routing,”
IEEE Security & Privacy, vol. 2, no. 3, pp. 28–39, 2004.
[49] A. Boukerche and D. Turgut, “Secure time synchronization protocols
for wireless sensor networks,” IEEE Wireless Communications, vol. 14,
no. 5, 2007.
[50] K. B. Rasmussen, S. Capkun, and M. Cagalj, “Secnav: secure broadcast
localization and time synchronization in wireless networks,” in Pro-
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
Function 1: Attack Generator Pseudocode
ceedings of the 13th annual ACM international conference on Mobile
computing and networking. ACM, 2007, pp. 310–313.
A. Alanwar, H. Ferraz, K. Hsieh, P. Martin, J. Hespanha, and M. Srivastava, “D-slats: Distributed simultaneous localization and time synchronization,” Proceedings of the 18th ACM International Symposium
on Mobile Ad Hoc Networking and Computing, 2017.
H. Ferraz, A. Alanwar, M. Srivastava, and J. P. Hespanha, “Node
localization based on distributed constrained optimization using jacobi’s
method,” 2017.
J. Aspnes, D. Goldenberg, and Y. R. Yang, “On the computational
complexity of sensor network localization,” in Algorithmic aspects of
wireless sensor networks. Springer, 2004, pp. 32–44.
B. Jackson and T. Jordán, “Connected rigidity matroids and unique
realizations of graphs,” Journal of Combinatorial Theory, Series B,
vol. 94, no. 1, pp. 1–29, 2005.
P. Zhang and Q. Wang, “On using the relative configuration to explore
cooperative localization,” IEEE Transactions on Signal Processing,
vol. 62, no. 4, pp. 968–980, Feb 2014.
A. Alanwar, F. Anwar, J. P. Hespanha, and M. Srivastava, “Realizing
uncertainty-aware timing stack in embedded operating system,” in Proc.
of the Embedded Operating Systems Workshop, Oct. 2016.
R. W. C Lenzen, P Sommer, “PulseSync: An Efficient and Scalable
Clock Synchronization Protocol,” in IEEE/ACM TRANSACTIONS ON
NETWORKING.
M. Quigley, K. Conley, B. P. Gerkey, J. Faust, T. Foote, J. Leibs,
R. Wheeler, and A. Y. Ng, “Ros: an open-source robot operating system,”
in ICRA Workshop on Open Source Software, 2009.
“Bitcraze crazyflie 2.0,” https://www.bitcraze.io/, accessed: 2017-10-01.
float getDistanceAttackValue ( type , currentT , totalT )
{
if ( currentT < totalT /3)
{
Shift = 2;
paretoScale = 3;
}
else if ( currentT > totalT /3 && currentT <2* totalT /3)
{
Shift = 6;
paretoScale = 6.5;
}
else
{
Shift = 1;
paretoScale = 2;
}
paretoShape = 3;
if ( type ==1)
// Uniform Distribution
attack = 2*( rand () + Shift );
A PPENDIX
else if ( type ==2)
A. Experimental Setup
The main components of our testbed can be summarized in
the following points:
• Motion capture system capable of 3D rigid body position
measurement with less than 0.5 mm accuracy. The system
consists of an eight-camera which are deployed in order
to provide accurate ground truth position measurements.
The results presented in this work treat the motion capture
estimate as a true position, though we qualify here that
all results are accurate to within the motion capture
accuracy. The ground truth position estimates from the
motion capture cameras are sent to a centralized server
which uses the Robot Operating System (ROS) [58] with
a custom package. We adopt a right-handed coordinate
system where y is the vertical axis, and x and z make up
the horizontal plane.
• The Fixed nodes used in the following experiments consist of custom-built circuit boards equipped with ARM
Cortex M4 processors t 196 MHz powered over Ethernet and communicating to Decawave DW1000 ultrawideband radios as shown in Figure 7 and Figure 8. Each
node performs a single and double-sided two-way range
with its neighbors. The used Decawave radio is equipped
with a temperature-compensated crystal oscillator with
frequency equals 38.4 MHz and a stated frequency stability of ±2 ppm. We installed eight UWB nodes in different
positions in a 10×9 m2 lab. More specifically, six nodes
are placed on the ceiling at about 2.5m high, and two
were placed at waist height at about 1 m in order to better
disambiguate positions on the vertical axis. Each node is
fully controllable over a TCP/IP command structure from
the central server. These nodes are placed so as to remain
// Normal Distribution
attack = 2* randn () + Shift ;
else if ( type ==3)
// Pareto Distribution
attack = randp ( paretoShape , paretoScale );
return attack ;
}
•
•
mostly free from obstructions, maximizing line-of-sight
barring pedestrian interference.
The Mobile Node used in the experiments are batterypowered mobile nodes also with ARM Cortex M4 processors based on the CrazyFlie 2.0 helicopter [59] and
equipped with the very same DW1000 radio as shown in
Figure 9. This allows for compatibility in the single and
double sided ranging technique used.
The server is a MacBook Pro laptop with Intel(R) Core
i7-4870HQ CPUs @2:5GHz, which is a normal PC.
B. Attacks on Distances
Function getDistanceAttackValue shows the attacker behavior for corrupting the measurements. We considered, as a
proof of concept, three types of attacks using Uniform, Normal
and Pareto Distributions. We gave them names as type 1, 2,
and 3 attacks, respectively. rand(), randn() and randp() are
considered APIs to get the required distribution. currentT is
the current system time and totalT is the total execution time.
We considered Pareto Distribution as an example of heavytailed distributions. The Pareto distribution is characterized
by a scale parameter paretoScale and a shape parameter
Fig. 6: Experimental setup overview, including, UWB nodes, motion capture cameras, and mobile quadrotor UWB nodes
Fig. 7: Custom anchor node with
ARM Cortex M4 processor and UWB
expansion.
Fig. 8: Ceiling-mounted anchor with
DW1000 UWB radio in 3D-printed
enclosure.
paretoShape. We picked the parameters of the distributions
to be logical attack values resulting in an acceptable position
in the localization area. Otherwise, the attack would be easily
detected.
Note that generating the attack values according to Function
1 simulates time-varying attack values that are different for all
links. The trend in the evaluation of the related work is to use
a uniform random number generator to define attack which
can look like some bias in the sensor reading from a practical
point of view. Thus, we used such time-varying parameters
that can evaluate the algorithm in a better way.
C. Attacks on Time
We would like to open a discussion in the types of the time
attacks, as they are tricky ones. Simple monitoring for evolving
time of every node can help in detecting many attacks. For
example, the time should not go backward; thus any reported
time value in the past can be flagged as an attack. Also,
the time can’t go faster than the specification of the crystal,
which is public information. However, we did not consider
this type of smart checkers in SecSens, we just let the secure
algorithms deal with that, but it can be a possible future work.
One popular attack in the time domain to provide a constant
delay to pretend, as if the time is shifted, that is why we
considered adding constant time attacks to the measurements
while evaluating SecSens.
Fig. 9: CrazyFlie 2.0 quadrotor helicopter with DW1000 UWB expansion.
D. Measurements Correlations
Many sensors in CPS reported correlated measurements. For
instance, some distance measurements could be based on the
time of flight calculations which is tightly coupled with time
measurements. Such tight coupling could be investigated to
detect attacks. We should note that SecSens does not at this
stage consider the measurements correlation and dealt with
the measurements as uncorrelated measurements. Many other
practical considerations could be investigated to enhance the
secure estimation algorithms.
E. CPS Physics
Understanding the underlying physics governing the CPS
itself would offer significant support to the secure estimation
algorithms using preprocessing phase which would help in
filtering the measurements. If a sensor observes a signal that
appears to violate the physics governing the CPS dynamics, an
attack flag can be easily raised to enhance the CPS security.
| 3cs.SY
|
Beyond the Pearson correlation: heavy-tailed risks,
weighted Gini correlations, and a Gini-type weighted
insurance pricing model
Edward Furman1
Department of Mathematics and Statistics, York University,
arXiv:1607.02623v1 [math.ST] 9 Jul 2016
Toronto, Ontario M3J 1P3, Canada
Ričardas Zitikis
Department of Statistical and Actuarial Sciences, University of Western Ontario,
London, Ontario N6A 5B7, Canada
Abstract. Gini-type correlation coefficients have become increasingly important in a variety of research areas, including economics, insurance and finance,
where modelling with heavy-tailed distributions is of pivotal importance. In
such situations, naturally, the classical Pearson correlation coefficient is of
little use. On the other hand, it has been observed that when light-tailed
situations are of interest, and hence when both the Gini-type and Pearson
correlation coefficients are well-defined and finite, then these coefficients are
related and sometimes even coincide. In general, understanding how the correlation coefficients above are related has been an illusive task. In this paper we
put forward arguments that establish such a connection via certain regressiontype equations. This, in turn, allows us to introduce a Gini-type Weighted
Insurance Pricing Model that works in heavy-tailed situation and thus provides a natural alternative to the classical Capital Asset Pricing Model. We
illustrate our theoretical considerations using several bivariate distributions,
such as elliptical and those with heavy-tailed Pareto margins.
Key words and phrases: Pearson correlation, weighted Gini correlation, capital
asset pricing model, weighted insurance pricing model, bivariate distributions.
JEL classification codes: IM10, IM30, IE43, G32.
1
Corresponding author. Email: [email protected]
1
1
Introduction
The topic of measuring association has been of profound interest in many theoretical
and applied research areas. The Pearson correlation coefficient, ρ := ρ[X, Y ] for two
random variables (r.v.’s) X and Y has arguably been one of the most popular members
of every researcher’s toolbox, notwithstanding its known shortcomings, and especially the
requirement of finite second moments. The latter requirement can be a real impediment
because modelling in economics, finance, and insurance frequently requires heavy-tailed
distributions. Furthermore, the coefficient ρ can turn out to be very small, and even zero,
when the r.v.’s X and Y are strongly dependent. This should not, of course, be surprising
because ρ aggregates, in the form of an integral, the values FX,Y (x, y) − FX (x)FY (y) over
all real x and y, where FX,Y denotes the joint cumulative distribution function (c.d.f.) of
X and Y , and FX and FY are the respective marginal c.d.f.’s. The desire to avoid some
of these, and other, limitations of the Pearson correlation coefficient has naturally grown
into a strong impetus for seeking alternatives (cf., e.g., Schechtman and Yitzhaki, 2013).
The monograph by Schechtman and Yitzhaki (2013) makes a number of convincing
arguments, supported by illustrative examples, in favour of the extended Gini correlation
coefficient
Γγ [X, Y ] =
Cov[X, (1 − FY (Y ))γ ]
,
Cov[X, (1 − FX (X))γ ]
(1.1)
indexed by parameter γ > 0. When γ = 1, then the extended Gini correlation coefficient
reduces to the classical Gini correlation coefficient, succinctly
Γ[X, Y ] =
Cov[X, FY (Y )]
.
Cov[X, FX (X)]
Various properties of Γγ [X, Y ] have been documented and discussed in the aforementioned
monograph (see also Samanthi et al., 2016). For example, we learn that when the pair
(X, Y ) follows the bivariate normal distribution, then Γγ [X, Y ] coincides with the Pearson
correlation coefficient ρ[X, Y ] irrespectively of the value of γ > 0. This property, obviously,
does not hold in general, which gives rise to a natural question as to when the two
coefficients are equal, but when they are not, then how much they differ. These are
among the problems that we tackle in this paper.
On a different note, the power weight function w(t) = tγ that transforms the decumulative distribution functions (d.d.f.’s) F X (x) := 1 − FX (x) and F Y (y) := 1 − FY (y)
in the definition of Γγ [X, Y ] might not be sufficiently flexible to adequately emphasize
(or de-emphasize) certain portions of the ranks of X and Y . For example, in prospect
theory (e.g., Tversky and Kahneman, 1992; Sriboonchita at al., 2009; Wakker, 2010; and
references therein) we find arguments in favour of sigmoidal and more complexly-shaped
2
weight functions. To accommodate various weighing designs, we therefore suggest using
the weighted Gini correlation coefficient
Cw [X, Y ] =
Cov[X, w(1 − FY (Y ))]
,
Cov[X, w(1 − FX (X))]
(1.2)
where the weight function w : [0, 1] → [0, 1] can be any non-decreasing function for which
the numerator and the denominator are well-defined and finite. The methodology that we
develop in this paper can successfully tackle this correlation coefficient for large classes of
weight functions and bivariate distributions.
We have organized the rest of this paper as follows. In Section 2 we present and discuss
general properties of the weighted Gini correlation coefficient, as well as its relationships
to the Capital Asset Pricing Model (CAPM) and Weighted Insurance Pricing Model
(WIPM). In Section 3, we present a technique that relates the weighted Gini correlation
coefficient to that of Pearson. Since the latter is closely linked to the bivariate normal
distribution, in the same section we discuss the aforementioned link in the case of bivariate
normal and, more generally, elliptical distributions. In Sections 4, 5 and 6 we deal, in
increasing complexity, with three bivariate Pareto-type distributions that model pairs
(X, Y ) with i) exchangeable margins and linear regression, ii) non-exchangeable margins
and linear regression, and iii) non-exchangeable margins and non-linear regression. Proofs
are relegated to Appendix A. To somewhat simplify the presentation, throughout the
paper we use W to denote the class of non-decreasing (weight) functions w : [0, 1] → [0, 1].
2
Properties of weighted Gini correlations
We have already noted several drawbacks of the Pearson correlation coefficient ρ[X, Y ]
when dealing with financial and insurance data, especially when the underlying r.v.’s do
not have finite second moments. Some other well-known drawbacks of relevance to our
present work are:
• The infinum and supremum of ρ[X, Y ] over all copulas governing the dependence
between X and Y generally depend on the marginal c.d.f.’s FX and FY (Shih and
Huang, 1992). Specifically, given an arbitrary pair of marginal c.d.f.’s, we cannot
claim that there is a pair (X, Y ) whose Pearson correlation coefficient is +1, but
there is such a pair (with its dependence described by the co-monotonic copula) in
the case of the Gini correlation coefficient.
• The Pearson correlation coefficient ρ[X, Y ] is not invariant under increasing and
non-linear transformations of the r.v.’s X and Y (Kendall and Stuart, 1979). In the
3
case of the Gini correlation coefficient we have that Γ[X, h(Y )] = Γ[X, Y ] for any
increasing function h.
Among the advantages of the Pearson correlation coefficient are its intuitive appeal
and computing easiness. The coefficient manifests naturally and plays a pivotal role in a
myriad of situations. For example, it is a parameter of the bivariate normal distribution,
and it measures the non-diversifiable risk in the Capital Asset Pricing Model (CAPM)
(Sharpe, 1964; Lintner, 1965; Black, 1972) and in its insurance counterpart (Furman and
Zitikis, 2009) called the Weighted Insurance Pricing Model (WIPM).
We next discuss properties of the weighted Gini correlation coefficient and in this way
highlight its superiority (at least in insurance and financial contexts) when compared to
the Pearson correlation coefficient.
Theorem 2.1 (Normalization). Let the c.d.f.’s FX and FY be continuous functions. Then
−λw [X] ≤ Cw [X, Y ] ≤ 1
(2.1)
for every w ∈ W, where
λw [X] =
Cov[X, w(FX (X))]
Cov[X, w ∗ (FX (X))]
with w ∗ (t) = 1 − w(1 − t). Furthermore,
(a) the upper bound is achieved with Y = X, in which case we have Cw [X, X] = 1;
(b) the lower bound is achieved with Y = −X, in which case Cw [X, −X] = −λw [X];
(c) when w(t) = t (the classical Gini correlation case), then λw [X] = 1.
Theorem 2.2 (Sign). For every w ∈ W, we have
(a) Cw [X, Y ] = 0 when X and Y are independent;
(b) Cw [X, Y ] ≥ 0 when X and Y are positively quadrant dependent (PQD);
(c) Cw [X, Y ] ≤ 0 when X and Y are negatively quadrant dependent (NQD).
Theorem 2.3 (Symmetry). Suppose there are real constants a, c ∈ R and positive constants b, d > 0, such that the random variables Xa,b = a + bX and Yc,d = c + dY are
interchangeable, that is, the distribution of (Xa,b , Yc,d) is equal to that of (Yc,d, Xa,b ). Then
for every w ∈ W we have
Cw [X, Y ] = Cw [Y, X].
4
(2.2)
Theorem 2.4 (Invariance under incrreasing transformations). Let a ∈ R and b > 0 be
some constants, and let h : [0, 1] → R be an increasing function. Then for every w ∈ W
we have
Cw [a + bX, h(Y )] = Cw [X, Y ].
(2.3)
Hence, in particular, Cw [a + bX, c + dY ] = Cw [X, Y ] for all real constants a, c ∈ R and
all positive constants b, d > 0.
Before stating our next theorem, we recall that for the set X of insurance risks X and
non-decreasing weight functions v : [0, ∞) → [0, ∞), the functional Πv : X × X → [0, ∞]
defined by the equation
E[Xv(Y )]
E[v(Y )]
Πv [X, Y ] =
(2.4)
is called the economic weighted premium calculation principle (p.c.p.), whereas the equation (Furman and Zitikis, 2009)
s
Πv [X, Y ] = E[X] + ρ[X, Y ]
Var[X]
(πv [Y ] − E[Y ]) ,
Var[Y ]
(2.5)
which holds with πv [Y ] := Πv [Y, Y ] under certain assumptions on the joint c.d.f.
of
(X, Y ), is referred to as the Weighted Insurance Pricing Model (WIPM). For example,
two natural choices of the r.v. Y would be i) leverage risk, which is of interest to the
pricing actuary, and ii) the aggregate risk S = X + Y . In the latter case, the WIPM
equation can be employed as an economic capital allocation rule, and we refer to Furman
and Zitikis (2009) for details on this topic.
An obvious disadvantage of the WIPM – akin to the CAPM – is its reliance on the
assumption that the variances of underlying r.v.’s are finite. Consequently, the WIPM
cannot be applied to a multitude of risks. To circumvent the problem, in definition (2.4)
we set v(y) = w ◦ (1 − FY (y)), which gives rise to the following Gini economic p.c.p.
ΠG,w [X, Y ] =
E[Xw(1 − FY (Y ))]
.
E[w(1 − FY (Y ))]
(2.6)
Theorem 2.5 (Connection to WIPM). Let (X, Y ) be a pair of r.v.’s in X . When there
exist constants α, β ∈ R such that E[X| Y ] = α + βY , then the Gini WIPM counterpart
to equation (2.5) is
ΠG,w [X, Y ] = E[X] + Cw [X, Y ]
Cov[X, w(1 − FX (X))]
(πG,w [Y ] − E[Y ]) ,
Cov[Y, w(1 − FY (Y ))]
where πG,w [Y ] := ΠG,w [Y, Y ].
5
(2.7)
Unlike the WIPM equation (2.5), Gini WIPM equation (2.7) does not need finiteness
of the second moments of X and Y , and as such it can be used for pricing/measuring
risks with infinite variances. Also, Theorem 2.5 shows that while the Pearson correlation
coefficient arises in the context of the CAPM, the weighted Gini correlation coefficient
Cw [X, Y ] manifests in the context of the Gini variant of WIPM. Finally, Theorem 2.5
provides yet another justification for the exploration of computational tractability of the
class of Gini correlations, with which we deal throughout the rest of this paper.
3
The Gini and Pearson correlations connected
At first sight, the weighted Gini correlation coefficient Cw [X, Y ] and the Pearson correlation coefficient ρ[X, Y ] seem to be quite different: the former aggregates the value of
X with the ranks of Y , and also with the ranks of X, whereas the Pearson correlation
coefficient couples the values of X and Y and does not rely on the ranks of any of them.
Nevertheless, the two coefficients are related as the following theorem shows.
Theorem 3.1. When there exist constants α, β ∈ R such that E[X| Y ] = α + βY , then
the equation
Cov[Y, w(1 − FY (Y ))]
(3.1)
Cov[X, w(1 − FX (X))]
holds for every w ∈ W. The converse is also true, assuming continuity of the c.d.f. FY .
Cw [X, Y ] = β
The most important part of Theorem 3.1 is of course the fact that having a linear
regression function allows us to decompose the weighted (and thus extended and classical) Gini correlation coefficient into i) covariance-type functionals based on the marginal
c.d.f.’s FX and FY , and ii) the slope β of the regression line whose immense role in the
classical Capital Asset Pricing Model (CAPM) has been well explored and documented
(e.g., Levy, 2011; and references therein). The main purpose of the converse of the theorem, which we establish under an additional but very mild condition, is to show that the
main part cannot be improved in general, that is, the linearity of the regression function
is pivotal for the form of Cw [X, Y ] spelled out on the right-hand side of equation (3.1).
We now illustrate Theorem 3.1 when the pair (X, Y ) follows the bivariate normal
distribution N2 (µ, Σ) with the vector µ = (µX , µY ) of the marginal expectations and
2
the symmetric matrix Σ whose main diagonal is made up of the variances σX
= Var[X]
and σY2 = Var[Y ], and the two off-diagonal entries are equal to the covariance σX,Y =
Cov[X, Y ]. It is well known that the equation E[X| Y ] = α + βY holds with the intercept
α = µX − βµY and the slope β = σX,Y /σY2 . Hence, an application of Theorem 3.1
establishes the equation
Cw [X, Y ] = ρ[X, Y ]
6
(3.2)
for every w ∈ W. This equation implies the well-known fact (e.g., Schechtman and
Yitzhaki, 2013) that the Pearson and the extended Gini correlation coefficients are equal
in the bivariate normal case. Hence, if the random pair (X, Y ) follows the bivariate normal
distribution, then the Gini-type WIPM equation reduces to
s
Var[X]
(πG,w [Y ] − E[Y ]) .
ΠG,w [X, Y ] = E[X] + ρ[X, Y ]
Var[Y ]
(3.3)
More generally, let the pair (X, Y ) follow the bivariate elliptical distribution E2 (µ, Σ)
with the vector µ = (µX , µY ) of marginal expectations and the positive-definite matrix Σ
2
whose diagonal entries are σX
and σY2 and the two identical off-diagonal entries are σX,Y .
We note that, unlike in the bivariate normal case, the just introduced σ’s may or may not
be the variances and covariances of X and Y . Nevertheless, we have (e.g., Fang et al.,
1990) the equation E[X| Y ] = α + βY with the intercept α = µX − βµY and the slope
β = σX,Y /σY2 , and so Theorem 3.1 applies and gives the equation
σX,Y
Cw [X, Y ] =
(3.4)
σX σY
for every w ∈ W. Consequently, just like in the bivariate normal case, the Pearson
and extended Gini correlation coefficients coincide whenever the former exists. When it
does not exist, which says that we are dealing with heavy-tailed random variables, then
Γγ [X, Y ] and likewise Cw [X, Y ] can be viewed as extensions of the Pearson correlation
coefficient to those pairs of X and Y that are outside the domain of definition of ρ[X, Y ].
Finally, the Gini-type WIPM equation is given in this case by
ΠG,w [X, Y ] = E[X] +
4
σX,Y
(πG,w [Y ] − E[Y ]) .
σY2
(3.5)
Exchangeable linearly-regressed margins
The elliptical distribution is a natural extension of the normal one and has served as an
adequate model in many financial problems. Its role in Economics has not been particularly pronounced, where the classical Pareto distribution with its numerous extensions
and variations have dominated the scene. Many insurance risk models have also relied on
extensions and generalizations of the Pareto distribution (e.g., Brazauskas and Serfling,
2003; Goovaerts et al., 2005). Because of this reason, and to emphasize that Theorem 3.1
does not hinge on the symmetry of the joint distribution of X and Y , we devote the rest
of this paper to several bivariate Pareto-type distributions in the context of Theorem 3.1.
We begin with the univariate Pareto distribution of the 2nd kind, whose d.d.f. is
−δ
x−µ
F (x) = 1 +
σ
7
for all x > µ, where µ ∈ R and σ > 0 are location and scale parameters, respectively, and
δ > 0 is tail index. The distribution has been a classical example of heavy-tailness: when
δ ∈ (1, 2], then the variance does not exist, and when δ ∈ (0, 1], then even the expectation
does not exist.
When it comes to bivariate extensions of this distribution, there are many of them to
consider. In this section we concentrate on the bivariate Pareto distribution of the 2nd
type (Arnold, 1983) whose joint d.d.f. is
−δ
x − µX y − µY
+
F X,Y (x, y) = 1 +
σX
σY
(4.1)
for all x > µX and y > µY , where µX , µY ∈ R are location parameters, σX , σY > 0 are
scale parameters, and δ > 0 is tail index. (Despite notational similarities, the two µ’s
are not the means, and the two σ’s are not the standard deviations of X and Y .) In
Figure 4.1 we depict this joint d.d.f. when µX = µY = 0, σX = σY = 1, and δ ≈ 5.87.
1
0.8
F (x, y)
0.6
0.4
0.2
0
-0.2
1.5
1.5
1
1
0.5
y
0.5
0
0
x
Figure 4.1: De-cumulative distribution function (4.1).
Consequently, ρ[X, Y ] ≈ 0.17. The distribution is symmetric, because X and Y are
identically distributed, and quite light-tailed, because δ ≈ 5.87 is a fairly large tail-index
value.
The pair (X, Y ) with joint d.d.f. (4.1) enjoys a number of attractive properties, which
have arguably been a reason for the popularity of this distribution. Among the properties
is the stochastic representation
EY
EX
,
, µY + σY
(X, Y ) =d µX + σX
G
G
8
(4.2)
where “=d ” denotes equality in distribution, EX and EY are independent exponential r.v.’s
with unit rates, and G is an independent gamma-distributed r.v. with shape parameter
δ > 0 and unit scale. Succinctly, we write EX ∼ Exp(1), EY ∼ Exp(1) and G ∼ Ga(δ, 1).
Note 4.1. The two exponential random variables EX and EY can be viewed as models
of idiosyncratic risks, and the third random variable G as the model for a background,
or underlying, risk. Naturally, one may need to deal with more than two dimensions and
with other distributions than the exponential and gamma, and all this is indeed possible
(cf., e.g., Asimit et al., 2016; and references therein).
A very important for us property of this bivariate distribution is that its regression
function is linear, that is, the equation E[X| Y ] = α + βY holds, and the parameters are
µY
σX
σX
1−
and β =
.
α = µX +
δ
σY
δσY
Consequently, Theorem 3.1 allows us to calculate the weighted Gini correlation coefficient
at a stroke:
Cw [X, Y ] =
σX Cov[Y, w(1 − FY (Y ))]
1
= ,
δσY Cov[X, w(1 − FX (X))]
δ
(4.3)
where the right-most equation holds because the ratio of the covariances is equal to
σY /σX , which is a simple consequence of the fact that (X − µX )/σX and (Y − µY )/σY
are identically distributed.
Note 4.2. For the covariances in equation (4.3) to be finite, the means E[X] and E[Y ]
have to be finite, and this is true whenever δ > 1, which we assume. On the other
hand, the Pearson correlation coefficient ρ[X, Y ] exists and is equal to 1/δ only when
δ > 2. Consequently, we can say that the weighted Gini correlation coefficient Cw [X, Y ]
is an extension of the Pearson correlation coefficient to a wider class of random pairs
(X, Y ), and in particular to those following d.d.f. (4.1) with the tail-index δ ∈ (1, ∞).
This extension is useful because tail-index values δ ∈ (1, 2] have manifested in numerous
real-life data sets: insurance, financial, and those related to income inequality (cf., e.g.,
Greselin et al., 2014; and references therein).
5
Non-exchangeable linearly-regressed margins
In some situations, exchangeability might be a drawback because identical tail indices
of the margins X and Y might contradict empirical evidence. This suggests that the
two idiosyncratic r.v.’s EX and EY in stochastic representation (4.2) might be affected
differently by the background r.v.
G. To rectify this situation, we can proceed by
9
introducing an auxiliary ‘background’ random variable GY ∼ Ga(δY , 1) with parameter
δY > 0, and require it to be independent of all the other random variables. That is, we
let the pair (X, Y ) admit the stochastic representation
EY
EX
.
, µY + σY
(X, Y ) =d µX + σX
G
GY + G
(5.1)
with location parameters µX , µY ∈ R and scale parameters σX , σY > 0. The corresponding d.d.f. is (Su and Furman, 2016)
−δ
−δY
y − µY
x − µX y − µY
+
1+
F X,Y (x, y) = 1 +
σX
σY
σY
(5.2)
for all x > µX and y > µY , where δ > 0 and δY > 0 are tail indices. Obviously,
(X − µX )/σX is Pareto of the 2nd kind with the tail index δ, and (Y − µY )/σY is Pareto
of the 2nd kind with the tail index δY∗ = δ + δY . In Figure 5.1 we depict this joint d.d.f.
1
0.8
F (x, y)
0.6
0.4
0.2
0
-0.2
1.5
1.5
1
1
0.5
y
0.5
0
0
x
Figure 5.1: De-cumulative distribution function (5.2).
when µX = µY = 0, σX = σY = 1, δ = 2.1 and δY∗ = 2.6254. As a result of these choices,
we have ρ[X, Y ] ≈ 0.17, which is an approximately the same Pearson correlation value
as in the previous section. Since the tail indices are different, the joint distribution is
not symmetric. Note also that the current distribution is more heavy-tailed than that
depicted in Figure 4.1.
Remarkably, despite being an asymmetric distribution, it nevertheless has a linear
regression function (Su and Furman, 2016, Theorem 2.3). Indeed, the equation E[X| Y ] =
10
α + βY holds with the parameters
α=
σX (δY∗ − 1)
δY∗ (δ − 1)
and β =
σX (δY∗ − 1)
.
σY δY∗ (δ − 1)
(5.3)
Hence, only with a slightly more complex stochastic representation than that in equation (4.2), we have succeeded in departing from the symmetry of margins X and Y but
preserved the linearity of their regression function. Consequently, we can still enjoy the
computational tractability of the weighted Gini correlation coefficient. Namely, Theorem
3.1 implies the equation
Cw [X, Y ] =
σX (δY∗ − 1) Cov[Y, w(1 − FY (Y ))]
,
σY δY∗ (δ − 1) Cov[X, w(1 − FX (X))]
(5.4)
and so we only need to calculate the two covariances on the right-hand side of equation
(5.4). We note at the outset that, unlike in the previous section, we are now dealing with
the case when the standardized covariances do not cancel out, because (X − µX )/σX and
(Y − µY )/σY have different tail indices. Hence, some additional effort is required.
To illustrate, let w(t) = tγ , in which case Cw [X, Y ] reduces to the extended Gini
correlation coefficient Γγ [X, Y ]. With technicalities relegated to Appendix A, we have the
formula
Γγ [X, Y ] =
δ(γ + 1) − 1
1
× ∗
δ δY (γ + 1) − 1
(5.5)
whenever δ > 1 and δY > 0. For comparison, the Pearson correlation coefficient is (Su
and Furman, 2016, Proposition 2.2 and Corollary 3.2)
s
δ−2
ρ[X, Y ] =
∗
δδY (δY∗ − 2)
(5.6)
whenever δ > 2 and δY > 0. Thus, in general, Γγ [X, Y ] and ρ[X, Y ] do not coincide.
We see from equations (5.5) and (5.6) that the extended Gini correlation coefficient is a
decreasing function of δ, whereas the Pearson correlation coefficient is not. We also see
this phenomenon in Figure 5.2, where µX = µY = 0, σX = σY = 1, and δY = 0.5254.
Our next illustration of Theorem 3.1 concerns the possibly S-shaped weight function
w(t) = wa,b (t), where
1
wa,b (t) =
B(a, b)
Z
t
sa−1 (1 − s)b−1 ds
0
for all t ∈ (0, 1), with parameters a > 0 and b > 0, whose values determine whether the
weight function is convex, concave, or S-shaped. Obviously, when b = 1, then wa,b (t)
reduces to the power function ta that leads to the extended Gini correlation coefficient
Γγ [X, Y ] with γ = a. (Certainly, the function wa,b is known in Calculus as the regularized
incomplete beta function, whereas in Statistics it is known as the beta c.d.f.)
11
0.5
0.45
0.4
0.35
Correlation
0.3
0.25
0.2
0.15
0.1
0.05
0
1
2
3
4
5
6
7
8
9
10
δ
Figure 5.2: Extended Gini (blue curve, decreasing) and Pearson (red curve, nonmonotonic) correlation coefficients as functions of δ.
Just like in the (convex or concave) case w(t) = tγ considered earlier, we can now
equally successfully employ Theorem 3.1 and concentrate on calculating the two covariances on the right-hand side of equation (5.4) with w(t) = wa,b (t). With the technical
details relegated to Appendix A, we obtain the equation
b
B (a + 1 − 1/δY∗ , b)
−
1
B(a, b)
a+b
Cwa,b [X, Y ] = ×
.
B (a + 1 − 1/δ, b)
b
δ
1−
−
B(a, b)
a+b
1−
(5.7)
When b = 1, then the right-hand side of equation (5.7) reduces to that of equation (5.5),
which is natural because wa,b (t) is the power function ta when b = 1.
Note 5.1. One may naturally wonder whether formulas like (5.5)–(5.7) are of practical value, given theoretical challenges when deriving them. The answer is definitely in
affirmative because parametric statistical inference crucially hinges on such formulas.
6
Non-exchangeable nonlinearly-regressed margins
Here we discuss yet another useful Pareto-type bivariate distribution. We note at the
outset that it does not have a linear regression function and thus Theorem 3.1 cannot be
12
applied. Namely, let the pair (X, Y ) admit the stochastic representation
EY
EX
,
, µY + σY
(X, Y ) =d µX + σX
GX + G
GY + G
(6.1)
where EX and EY are two independent exponential r.v.’s with unit scales, and G ∼
Ga(δ, 1), GX ∼ Ga(δX , 1) and GY ∼ Ga(δY , 1) are three independent gamma r.v.’s, which
are also independent of EX and EY . Hence, µX and µY ∈ R are location parameters, σX
and σY > 0 are scale parameters, and δ, δX and δY > 0 are tail indices. The joint d.d.f.
of the random pair (X, Y ) is (Su and Furman, 2016)
−δ
−δX
−δY
x − µX y − µY
y − µY
x − µX
F X,Y (x, y) = 1 +
1+
+
1+
σX
σY
σX
σY
(6.2)
for all x > µX and y > µY . We see from the d.d.f. that (X − µX )/σX is Pareto of the
∗
2nd kind with the tail index δX
= δ + δX , and (Y − µY )/σY is Pareto of the 2nd kind with
the tail index δY∗ = δ + δY . In Figure 6.1 we depict this joint d.d.f. when µX = µY = 0,
1
0.8
F (x, y)
0.6
0.4
0.2
0
-0.2
1.5
1.5
1
1
0.5
y
0.5
0
0
x
Figure 6.1: De-cumulative distribution function (6.2).
∗
σX = σY = 1, δX
= 3 and δY∗ = 2.5. Again, we have chosen the parameters so that
ρ[X, Y ] ≈ 0.17, which is an approximately the same Pearson correlation value as in the
previous sections. The joint distribution is not symmetric due to the different values of
∗
δX
and δY∗ , and it is more heavily-tailed than that depicted in Figure 4.1.
Unlike in the previous sections, the current bivariate Pareto distribution does not have
a linear regression function (Su and Furman, 2016, Theorem 3.3), and so we cannot rely
13
on Theorem 3.1 to easily calculate Gini-type correlation coefficients. Nevertheless, we can
still compute them directly and get closed-form formulas, but the task is considerably
more involved.
As an example, consider the extended Gini correlation coefficient Γγ [X, Y ]. We need
the (q + 1) × q hypergeometric function (Gradshteyn and Ryzhik, 2007)
q+1 Fq (a1 , . . . , aq+1 ; b1 , . . . , bq ; z)
:=
∞
X
(a1 )k , . . . , (aq+1 )k z k
k=0
(b1 )k , . . . , (bq )k k!
,
(6.3)
where q ≥ 0 is a non-negative integer and (a)k is the Pochhammer symbol. When
a1 , . . . , aq+1 are positive, and this is the case upon which we rely, then the radius of convergence of the series is the open disk |z| < 1 of the complex plane. On the boundary |z| = 1
of the disc, the series converges absolutely when h := b1 + · · · + bq − a1 − · · · − aq+1 > 0,
and converges except at z = 1 when −1 < h ≤ 0.
Let i = (i1 , i2 , i3 ) be a non-negative integer-valued triplet such that i1 + i2 + i3 = 2,
and let I denote the set of all such triplets. We prove (details in Appendix A) that for
∗
∗
γ > 0, δX,1,3
= δX
+ i1 + i3 , and appropriately defined constants di , the extended Gini
correlation coefficient is
∗
δX
(γ + 1) + 1
∗
δX
γ
∗
∗
∗
(γ + 1) − 1)3 F2 (δ + i3 , 2, 1; δX,1,3
, (γ + 1)δY∗ + i2 + i3 ; 1)
1 X (δX − 1)(γ + 1)(δX
di
− ∗
.
∗
∗
δX γ i∈I
(δX,1,3
− 2)(δX,1,3
− 1)((γ + 1)δY∗ + i2 + i3 − 1)
Γγ [X, Y ] =
(6.4)
We admit that it is a cumbersome formula, but it is a natural one because it contains,
as special cases, earlier derived formulas (4.3) and (5.5). To demonstrate, consider first
d.d.f. (4.1), in which case we let δX ↓ 0 and δY ↓ 0, then set
(
1 when i = (0, 0, 2),
di =
0 otherwise,
and finally use equation (6.3) to obtain
3 F2 (δ
∗
+ i3 , 2, 1; δX,1,3
, (γ + 1)δY∗ + i2 + i3 ; 1) = 2 F1 (2, 1; (γ + 1)δ + 2; 1)
=
δ(γ + 1) + 1
.
δ(γ + 1) − 1
With these facts, formula (4.3) follows easily. In a similar fashion, consider d.d.f. (5.2),
in which case we let δX ↓ 0, then set
(
1 when i = (0, 0, 2) or (0, 1, 1),
di =
0 otherwise,
and arrive at formula (5.5).
14
Note 6.1. Reflecting upon the bivariate distribution considered in this section, we see
that the extended and, more generally, weighted Gini correlations can be calculated in
cases that do not have linear regression functions. Indeed, a careful analysis of the proof
of Theorem 3.1 shows that it is possible to accommodate non-linear regression functions
as well, but the right-hand side of equation (3.1) changes, which is of course natural.
Acknowledgements
This work is a part of our research under the grant “The Capital Asset Pricing Model:
an Insurance Variant” awarded by the Casualty Actuarial Society (CAS).
A
Appendix: proofs
Recall the notation w ∗ (t) = 1 − w(1 − t), which we use throughout the appendix.
Proof of Theorem 2.1. First, we rewrite the weighted correlation coefficient as follows:
Cw [X, Y ] =
Cov[X, w ∗ (V )]
,
Cov[X, w ∗ (U)]
(A.1)
where V = FY (Y ), and U = FX (X). Using Cuadras’s (2002) generalization of Hoeffding’s
covariance representation, we have
ZZ
∗
Cov[X, w (V )] =
P[X ≤ x, V ≤ t] − FX (x)t dxdw ∗ (t)
(A.2)
because FV (t) = t due to the fact that V is distributed uniformly on the interval [0, 1].
The upper Fréchet bound for the joint probability P[X ≤ x, V ≤ t] is min{FX (x), t},
which is equal to P[X ≤ x, U ≤ t] because U = FX (X). Thus
P[X ≤ x, V ≤ t] − FX (x)t ≤ P[X ≤ x, U ≤ t] − FX (x)FU (t)
because FU (t) = t due to the fact that U is distributed uniformly on the interval [0, 1].
Consequently, Cov[X, w ∗ (V )] ≤ Cov[X, w ∗ (U)], which proves Cw [X, Y ] ≤ 1.
To prove Cw [X, Y ] ≥ −λw [X], we again start with equation (A.2) but this time use
the lower Fréchet bound for P[X ≤ x, V ≤ t], which is max{FX (x) + t − 1, 0}. Note that
max{FX (x)+t−1, 0}−FX (x)t = −(min{FX (x), 1−t}−FX (x)(1−t)). Since U = FX (X),
we thus have
P[X ≤ x, V ≤ t] − FX (x)t ≥ − P[X ≤ x, U ≤ 1 − t] − FX (x)(1 − t)
= P[X ≤ x, 1 − U ≤ t] − FX (x)t
= P[X ≤ x, 1 − U ≤ t] − FX (x)F1−U (t).
15
Consequently, Cov[X, w ∗ (V )] ≥ Cov[X, w ∗ (1 − U)], which proves the bound
Cw [X, Y ] ≥
Cov[X, w ∗ (1 − U)]
Cov[X, w ∗ (U)]
whose right-hand side is equal to −λw [X]. This completes the proof of Theorem 2.1.
Proof of Theorem 2.2. Part (a) is obvious. To prove part (b), we rewrite the weighted
correlation coefficient as follows:
Cw [X, Y ] =
Cov[X, w ∗ ◦ FY (Y )]
.
Cov[X, w ∗ ◦ FX (X)]
(A.3)
Since the function w ∗ ◦FX is non-decreasing, we know from Lehmann (1966) that Cov[X, w ∗ ◦
FX (X)] ≥ 0, and so we only need to show that Cov[X, w ∗ ◦FY (Y )] ≥ 0 whenever X and Y
are PQD. Using Cuadras’s (2002) generalization of Hoeffding’s covariance representation,
we have
∗
Cov[X, w ◦ FY (Y )] =
ZZ
P[X ≤ x, Y ≤ y] − FX (x)FY (y) dxdw ∗ ◦ FY (y).
(A.4)
Since the function w ∗ ◦ FY is non-decreasing, the integral is non-negative whenever the
integrand is non-negative, which is so whenever X and Y are PQD, that is, P[X ≤ x, Y ≤
y] ≥ FX (x)FY (y) for all x and y. This proves part (b). The proof of part (c) is analogous.
This concludes the proof of Theorem 2.2.
Proof of Theorem 2.3. We start with the elementary observation that
Cw [X, Y ] = Cw [Xa,b , Yc,d].
(A.5)
Next we rewrite the right-hand side of equation (A.5) in the form of a ratio analogous
to that on the right-hand side of equation (A.3). Then, for both the numerator and the
denominator, we apply Cuadras’s (2002) generalization of Hoeffding’s covariance representation, analogously to equation (A.4). In this way, we express Cw [Xa,b , Yc,d] in terms of
the bivariate cdf of (Xa,b , Yc,d) and the marginal c.d.f.’s of Xa,b and Yc,d. The interchangeability assumption implies that the latter three c.d.f.’s are equal to those of (Yc,d, Xa,b ),
Yc,d, and Xa,b , respectively, thus proving that Cw [Xa,b , Yc,d] is equal to Cw [Yc,d , Xa,b ], which
is obviously equal to Cw [Y, X]. This concludes the proof of Theorem 2.3.
Proof of Theorem 2.4. Since Cw [a + bX, h(Y )] = Cw [X, h(Y )], we only need to show that
Cw [X, h(Y )] = Cw [X, Y ], which follows from the fact that Fh(Y ) (h(y)) = FY (y). This
concludes the proof of Theorem 2.4.
Proofs of Theorems 2.5 and 3.1 rely on the following lemma.
16
Lemma A.1. When there are constants α, β ∈ R such that E[X | Y ] = α + βY , then the
equation
Cov[X, w(1 − FY (Y ))] = βCov[Y, w(1 − FY (Y ))].
(A.6)
holds for every w ∈ W.
Proof. We have
Cov[X, w(1 − FY (Y ))] = E[E[X | Y ]w(1 − FY (Y ))] − E[X]E[w(1 − FY (Y ))]
= βE[Y w(1 − FY (Y ))] + (α − E[X])E[w(1 − FY (Y ))]
= βE[Y w(1 − FY (Y ))] − βE[Y ]E[w(1 − FY (Y ))]
= βCov[Y, w(1 − FY (Y ))],
which establishes equation (A.6).
Proof of Theorem 2.5. We start with the equations
ΠG,w [X, Y ] − E[X] = α − E[X] + βE[Y ] + β (πG,w [Y ] − E[Y ])
= β (πG,w [Y ] − E[Y ]) ,
(A.7)
that follow from the definitions of ΠG,w [X, Y ] and πG,w [Y ] combined with simple algebra,
and the elementary equation E[X] = α + βE[Y ]. It now suffices to note that, in view of
equation (A.6), we have
Cov[X, w(1 − FY (Y ))]
Cov[Y, w(1 − FY (Y ))]
Cov[X, w(1 − FX (X))]
= Cw [X, Y ]
Cov[Y, w(1 − FY (Y ))]
β=
which establishes equation (2.7) and concludes the proof of Theorem 2.5.
Proof of Theorem 3.1. Equations (3.1) and (A.6) are equivalent, and thus the equation
E[X| Y ] = α + βY holds. To prove the second (i.e., converse) part of Theorem 3.1, we
start with the assumption that equation (A.6) holds for every w ∈ W. With the notation
Z = X − βY , equation (A.6) becomes equivalent to
E (Z − E[Z])w(1 − FY (Y )) = 0,
(A.8)
which by assumption must hold for all the weight functions w ∈ W, and in particular for
those that safisfy w(t) > 0 for all t ∈ (0, 1). What remains to be shown, therefore, is that
the validity of equation (A.8) for all w ∈ W implies the validity of
Z − E[Z] = 0 almost surely.
17
(A.9)
Indeed, statement (A.9) is equivalent to X − βY − E[X] + βE[Y ] = 0, and the latter
equation obviously implies E[X| Y ] = α + βY with α = E[X] − βE[Y ], thus concluding
the proof of the converse of Theorem 3.1.
To verify that statement (A.9) holds, we proceed by contradiction, that is, we assume
that statement (A.9) does not hold. Then there is a subset A of the underlying sample
space Ω such that P(A) > 0 and Z(ω) − E[Z] 6= 0 for all ω ∈ A. We can refine this
statement by decomposing A = A− ∪ A+ with either A− or A+ or both having (strictly)
positive probabilities and such that Z(ω)−E[Z] < 0 for all ω ∈ A− and Z(ω)−E[Z] > 0 for
all ω ∈ A+ . Assume for concreteness that P(A+ ) > 0, with the case P(A− ) > 0 analyzed
analogously. We arrive at the required contradiction if w(1 − FY (Y (ω))) > 0 for P-almost
all ω ∈ A+ . The latter holds if FY (Y (ω)) ∈ (0, 1) because we deal with the subclass of
weight functions such that w(t) > 0 for all t ∈ (0, 1). Statement FY (Y (ω)) ∈ (0, 1) holds
for P-almost all ω ∈ Ω because the c.d.f. FY is continuous by assumption. This concludes
the proof of Theorem 3.1.
Proof of equation (5.5). We need to calculate the two covariances on the right-hand side
of equation (3.1). We start with the denominator and have
Z ∞
γ
E[(X − µX )(1 − FX (X)) ] = σX
x(1 + x)δγ δ(1 + x)−(δ+1) dx
0
=
σX
.
(γ + 1)(δ(γ + 1) − 1)
Setting γ = 0, we have E[X−µX ] = σX /(δ−1). Furthermore, E[(1−FX (X))γ ] = 1/(γ+1).
From these formulas, we obtain
Cov[X, (1 − FX (X))γ ] = −
γ
δ
σX
.
γ + 1 (δ − 1)(δ(γ + 1) − 1)
(A.10)
To calculate the covariance Cov[Y, (1 − FY (Y ))γ ], we recall that (Y − µY )/σY is Pareto
of the 2nd kind with the tail index δY∗ = δ + δY . Hence, by changing α into α + αY and
also σX into σY on the right-hand side of equation (A.10), we arrive at
Cov[Y, (1 − FY (Y ))γ ] = −
δY∗
γ
σY ∗
.
γ + 1 (δY − 1)((δY∗ )(γ + 1) − 1)
(A.11)
Using equations (A.10) and (A.11) on the right-hand side of equation (3.1), we obtain
equation (5.5).
Proof of equation (5.7). With the notation Z = (X −µX )/σX and the fact that FZ−1 (u) =
18
(1 − u)−1/δ − 1, we have
E[(X − µX )wa,b (1 − FX (X))] = σX E[FZ−1(FZ (Z))wa,b (1 − FZ (Z))]
Z 1
= σX
(u−1/δ − 1)wa,b (u)du
0Z 1
b
−1/δ
.
= σX
u
wa,b (u)du −
a+b
0
Furthermore,
Z
1
u
−1/δ
0
1
u a−1
(1 − t)b−1
wa,b (u)du =
dtdu
u
B(a, b)
0
0
Z 1 a−1
t (1 − t)b−1
δ
1 − t(δ−1)/δ dt
=
δ−1 0
B(a, b)
δ
B (a + (δ − 1)/δ, b)
=
1−
.
δ−1
B(a, b)
Z
−1/δ
Z
t
Hence, with the formulas E[X − µX ] = σX /(δ − 1) and E[wa,b (1 − FX (X))] = b/(a + b),
we obtain
σX δ
Cov[X, wa,b (1 − FX (X))] =
δ−1
B (a + 1 − 1/δ, b)
b
1−
.
−
B(a, b)
a+b
(A.12)
(Note that when b = 1, then equation (A.12) reduces to equation (A.10).) Changing δ
into δ + δY and σX into σY on the right-hand side of equation (A.12), we arrive at
B (a + 1 − 1/(δ + δY ), b)
b
σY (δ + δY )
1−
. (A.13)
−
Cov[Y, wa,b (1−FY (Y ))] =
δ + δY − 1
B(a, b)
a+b
Using covariance formulas (A.12) and (A.13) on the right-hand side of equation (5.4), we
obtain equation (5.7).
Proof of equation (6.4). The bivariate p.d.f. of (X − µX )/σX and (Y − µY )/σY is given
by
p(x, y; δX , δY , δ) =
X
di (1 + x)−(δX +i1 ) (1 + y)−(δY +i2 ) (1 + x + y)−(δ+i3 ) ,
(A.14)
i∈I
where I is the set of all non-negative and integer-valued triplets i = (i1 , i2 , i3 ) such that
i1 + i2 + i3 = 2. We have
1
E[(X − µX )(1 − FY (Y ))γ ]
σX
X Z ∞Z ∞
∗
=
di
x(1 + y)δY γ (1 + x)−(δX +i1 ) (1 + y)−(δY +i2 ) (1 + x + y)−(δ+i3 ) dxdy
0
i∈I
=
X
i∈I
di
Z
0
∞
(1 + y)
0
−((γ+1)(δY +δ)+i2 +i3 )
Z
∞
x(1 + x)
0
19
−(δX +i1 )
x
1+
1+y
−(δ+i3 )
dxdy.
We continue with the help of the hypergeometric function noted in Section 6 and have
1
E[(X − µX )(1 − FY (Y ))γ ]
σX
X Z ∞ (1 + y)−((γ+1)(δY +δ)+i2 +i3 )
∗
=
di
2 F1 (δ + i3 , 2; δX,1,3 ; y/(1 + y))dy
∗
∗
(δ
−
2)(δ
−
1)
0
X,1,3
X,1,3
i∈I
Z
1
((γ+1)(δ
+δ)+i
X
2 +i3 −2)
Y
(1 − x)
∗
=
di
2 F1 (δ + i3 , 2; δX,1,3 ; x)dx
∗
∗
(δX,1,3 − 2)(δX,1,3 − 1)
0
i∈I
=
X
di
i∈I
∗
∗
3 F2 (δ + i3 , 2, 1; δX,1,3 , (γ + 1)δY
∗
∗
(δX,1,3
− 2)(δX,1,3
− 1)((γ + 1)δY∗
+ i2 + i3 ; 1)
.
+ i2 + i3 − 1)
∗
Since E[X − µX ] = σX /(δX
− 1) and E[(1 − FX (X))γ ] = 1/(γ + 1), equation (6.4)
follows.
References
Arnold, B.C. (1983) Pareto Distributions. International Cooperative Publishing House,
Fairland, MD.
Asimit, A.V., Vernic, R. and Zitikis, R. (2016). Background risk models and stepwise
portfolio construction. Methodology and Computing in Applied Probability (in press).
Black, F. (1972). Capital market equilibrium with restricted borrowing, Journal of Business, 45, 444–455.
Brazauskas, V. and Serfling, R. (2003). Favourable estimators for fitting Pareto models: a
study using goodness-of-fit measures with actual data. ASTIN Bulletin: The Journal
of the International Actuarial Association, 33, 365–381.
Cuadras, C.M. (2002). On the covariance between functions. Journal of Multivariate
Analysis, 81, 19–27.
Fang, K.T., Kotz, S. and Ng, K.W. (1990). Symmetric Multivariate and Related Distributions. Chapman & Hall, London.
Furman, E. and Zitikis, R. (2009). Weighted pricing functionals with application to
insurance: An overview. North American Actuarial Journal, 13, 483–496.
Goovaerts, M. J., Kaas, R., Laeven, R. J. A., Tang, Q. and Vernic, R. (2005). The
tail probability of discounted sums of Pareto-like losses in insurance. Scandinavian
Actuarial Journal, 6, 446–461.
Gradshteyn, I.S. and Ryzhik, I.M. (2007). Table of Integrals, Series and Products (7th
edition). Academic Press, New York.
Greselin, F., Pasquazzi, L. and Zitikis, R. (2014). Heavy tailed capital incomes: Zenga
index, statistical inference, and ECHP data analysis. Extremes, 17, 127–155.
20
Kendall, M. and Stuart, A. (1979). The Advanced Theory of Statistics. Charles Griffin,
London.
Levy, H. (2011). The Capital Asset Pricing Model in the 21st Century: Analytical,
Empirical, and Behavioral Perspectives. Cambridge University Press, Cambridge.
Lintner, J. (1965). Security prices, risk, and maximal gains from diversification, Journal
of Finance, 20, 587–615.
Samanthi, R.G.M., Wei, W. and Brazauskas, V. (2016). Ordering Gini indexes of multivariate elliptical risks. Insurance: Mathematics and Economics, 68, 84–91.
Schechtman, E. and Yitzhaki, S. (2013). The Gini Methodology: A Primer on a Statistical
Methodology. Springer, New York.
Sharpe, W.F. (1964). Capital asset prices: A theory of market equilibrium under conditions of risk. Journal of Finance, 19, 425–442.
Shih, W.J. and Huang, W.M. (1992). Evaluating correlations with proper bounds. Biometrics, 48, 1207–1213.
Sriboonchita, S., Wong, W.K., Dhompongsa, D., Nguyen, H.T. (2009). Stochastic Dominance and Applications to Finance, Risk and Economics. Chapman and Hall, Boca
Raton, FL.
Su, J. and Furman, E. (2016). A form of multivariate Pareto distribution with applications
to financial risk measurement. ASTIN Bulletin: The Journal of the International
Actuarial Association (in press).
Tversky, A. and Kahneman, D. (1992). Advances in prospect theory: cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5, 297–323.
Wakker, P.P. (2010). Prospect Theory: For Risk and Ambiguity. Cambridge University
Press, Cambridge.
21
| 10math.ST
|
SUBMITTED TO IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
1
Improving Dermoscopic Image Segmentation with
Enhanced Convolutional-Deconvolutional Networks
arXiv:1709.09780v1 [cs.CV] 28 Sep 2017
Yading Yuan* and Yeh-Chi Lo
Abstract—Automatic skin lesion segmentation on dermoscopic
images is an essential step in computer-aided diagnosis of
melanoma. However, this task is challenging due to significant
variations of lesion appearances across different patients. This
challenge is further exacerbated when dealing with a large
amount of image data. In this paper, we extended our previous
work by developing a deeper network architecture with smaller
kernels to enhance its discriminant capacity. In addition, we
explicitly included color information from multiple color spaces
to facilitate network training and thus to further improve the
segmentation performance. We extensively evaluated our method
on the ISBI 2017 skin lesion segmentation challenge. By training
with the 2000 challenge training images, our method achieved an
average Jaccard Index (JA) of 0.765 on the 600 challenge testing
images, which ranked itself in the first place in the challenge.
Index Terms—Dermoscopic images, deep learning, fully convolutional neural networks, image segmentation, jaccard distance,
melanoma
I. I NTRODUCTION
ALIGNANT melanoma is among the most rapidly
growing cancers in the world [1] and dermoscopy is the
most commonly used in vivo imaging modality that provides
a better visualization of subsurface structures of pigmented
skin lesions [2]. While this technique allows dermatologists
to detect early stage melanoma that are not visible by human
eyes, visual interpretation alone is a time-consuming procedure
and prone to inter- and intra-observer variabilities. Therefore,
automated and accurate analysis of melanoma has become
highly desirable in assisting dermatologists for improving
their efficiency and objectivity when interpreting dermoscopic
images in clinical practice [3].
Automatic skin lesion segmentation is an essential component in computer-aided diagnosis (CAD) of melanoma [4]
[5]. However, this is a very challenging task due to significant
variations in location, shape, size, color and texture across
different patients. In addition, some dermoscopic images have
low contrast between lesion and surrounding skin, and suffer
from artifacts and intrinsic features such as hairs, frames,
blood vessels and air bubbles. Existing lesion segmentation
methods based on clustering, thresholding, region growing, or
deformable models have shown limited success in solving this
difficult problem when applying to a large amount of image
data [3], [6], [7].
Recent development of deep learning has revolutionized the
field of machine learning and computer vision. Deep learning
M
Y. Yuan and Y-C Lo are with the Department of Radiation Oncology,
Icahn School of Medicine at Mount Sinai, New York, NY, USA, e-mail:
[email protected], [email protected].
techniques, especially deep convolutional neural networks [8],
have been rapidly adopted in various medical image analysis
problems, including body recognition [9], lesion detection
[10], image registration [11], segmentation [12] and classification [13]. In particular, Yu et al. [14] introduced a deep
residual network with more than 50 layers for automatic skin
lesion segmentation, in which several residual blocks [15]
were stacked together to increase the representative capability
of their model. In [16], Bi et al. proposed a multi-stage
approach to combine the outputs from multiple cascaded fully
convolutional networks (FCNs) to achieve a final skin lesion
segmentation. In our recent study [17], we developed a fully
automatic method for skin lesion segmentation by leveraging
a 19-layer deep FCN that is trained end-to-end and does not
rely on prior knowledge of the data. Furthermore, we designed
a novel loss function based on Jaccard distance that is directly
related to image segmentation task and eliminates the need
of sample re-weighting. Experimental results showed that our
method outperformed other state-of-the-art algorithms on two
benchmark datasets - one is from ISBI 2016 challenge titled as
skin lesion analysis towards melanoma detection [18],
and the other is the PH2 dataset [19].
In this paper, we present a major extension of our previous
work to further enhance our model in automatic skin lesion
segmentation. Specifically, 1) we investigate the potential of
using a deeper network architecture with smaller convolutional
kernels such that the new model has increased discriminative
capacity to handle a larger variety of image acquisition conditions; 2) besides Red-Green-Blue (RGB) channels, we also
investigate the use of channels in other color spaces, such as
Hue-Saturation-Value (HSV) and CIELAB [20], as additional
inputs to our network that aim for a more efficient training
while controlling over-fitting; 3) we evaluate the proposed
framework on ISBI 2017 Skin Lesion Analysis T owards
M elanoma Detection challenge1 datasets. Experimental results demonstrated a significant performance gain as compared
to our previous model, ranking itself as the first place among
21 final submissions.
II. DATASET AND P RE - PROCESSING
We solely used ISBI 2017 challenge datasets for training
and validating the proposed deep learning model named as
convolutional-deconvolutional neural network (CDNN). As
compared to ISBI 2016 challenge, the image database in 2017
is doubled in size and includes a larger variety of tumor appearance on dermoscopic images, which makes the automatic
1 https://challenge.kitware.com/#challenge/583f126bcad3a51cc66c8d9a
SUBMITTED TO IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
Fig. 1. The appearance of a dermoscopic image with color manipulation.
The left image in the first row shows the original image, and the rest are
the adjusted images by normalizing the contrast of each channel to [5, 95]
percentile window, respectively.
lesion segmentation much more challenging. Specifically, the
2017 database includes a training dataset with 2000 annotated
dermoscopic images (374 melanomas), and a blind held-out
testing dataset with 600 images (117 melanomas). In order
to facilitate model validation, a small independent dataset
with 150 images (30 melanomas) is publicly available for
challenge participants to fine tune the hyper-parameters before
submitting the final results on the testing dataset. The image
size ranges from 540 × 722 to 4499 × 6748. By observing
most of the images in the training set have a height to width
ratio of 3 : 4, we resized all the images to 192 × 256 using
bi-linear interpolation to keep balance between segmentation
performance and computation cost.
While RGB is the most popular space to represent color
information of natural images, its usage in image segmentation
is however restricted by the fact that the RGB channels are
not independent to each other. Thus, when applying image
augmentation onto each channel during network training, the
resultant image may have unrealistic appearance that makes
the training less efficient. In order to address this issue and
to use color information more effectively, we added the three
channels from HSV color space and the lightness channel (L)
from CIELAB color space, which separate luma (image intensity) from chroma (color information) to allow independent
process on these two types of information. Figure 1 shows
how a dermoscopic image can be adjusted by normalizing the
contrast of each channel to [5, 95] percentile window. No other
pre-processing was performed, so the input dimension to our
CDNN model is 192 × 256 × 7.
III. M ETHODS
The CDNN model we proposed belongs to FCNs category,
which allows an entire image segmentation in a single pass
instead of classifying the central pixel of a small image patch
each time. Using an entire image as input not only makes
model training and inference more efficient , but also includes
much richer contextual information as compared to small
patches, which usually leads to more robust and accurate
segmentation.
2
model consists of two pathways, in which the convolutional
path resembles a traditional CNN that extracts a hierarchy
of image features from low to high complexity, and the
deconvolutional path transforms the aggregated features and
reconstructs the segmentation map from coarse to fine resolution. In this way, CDNN can take both global information and
fine details into account for tumor segmentation.
In order to improve the discriminative power of CDNN to
handle a much larger variety of image acquisition conditions,
we extend the previous model in the following two aspects.
Firstly, we employ small 3 × 3 kernels in nearly all the
convolutional and de-convolutional layers and maintain the
same effective receptive field size of larger kernels by adding
additional layers. This deeper architecture reduces the number
of weights while allowing more non-linear transformations on
the data. Then we increase the number of features in each
layer in order to boost the capacity of the CDNN model.
All convolutional and de-convolutional layers use Rectified
Linear Units (ReLUs) as the nonlinear activation function [21],
and batch normalization [22] is applied to reduce the internal
covariate shift for each training mini-batch . The batch size
was set as 18 in this study. Eventually the network contains
29 layers with 5, 042, 589 trainable parameters. The filter size
and number of channels of the output feature maps are shown
in Table I.
TABLE I
A RCHITECTURAL DETAILS OF THE PROPOSED CDNN MODEL
(A BBREVATIONS : CONV: CONVOLUTIONAL LAYER ; POOL : MAX - POOLING
LAYER ; DECV: DECONVOLUTIONAL LAYER , UPS : UPSAMPLING LAYER ).
Conv
Filter
Features
Deconv
Filter
Features
conv-1-1
conv-1-2
pool-1
conv-2-1
conv-2-2
pool-2
conv-3-1
conv-3-2
pool-3
conv-4-1
conv-4-2
pool-4
conv-5
3×3
3×3
2×2
3×3
3×3
2×2
3×3
4×4
2×2
3×3
3×3
2×2
3×3
16
32
32
64
64
64
128
128
128
256
256
256
512
decv-1
ups-1
decv-2-1
decv-2-2
ups-2
decv-3-1
decv-3-2
ups-3
decv-4-1
decv-4-2
ups-4
decv-5-1
output
3×3
2×2
3×3
3×3
2×2
4×4
3×3
2×2
3×3
3×3
2×2
3×3
3×3
256
256
256
128
128
128
128
128
64
32
32
16
1
The CDNN model represents an end-to-end mapping from
the input image to a segmentation map, where each element is
the probability that the corresponding input pixel belongs to
the tumor. Those trainable parameters are learned from training
data by minimizing a loss function. We use a loss function
based on Jaccard distance [17]:
P
(tij pij )
i,j
P
P
P
,
(1)
LdJ = 1 −
t2ij + p2ij − (tij pij )
i,j
A. CDNN model
In our previous work [17], we proposed a CDNN model
containing 19 layer with 290, 129 trainable parameters. This
i,j
i,j
where tij and pij are target and the output of pixel (i, j),
respectively. As compared to the conventionally used crossentropy, the proposed loss function is directly related to image
SUBMITTED TO IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
segmentation task because Jaccard index is a common metric
to assess medical image segmentation accuracy, especially in
this ISBI 2017 challenge. Meanwhile, this loss function is
well adapted to the problems with high imbalance between
foreground and background classes as it doesn’t require any
class re-balancing.
Given a new test dermoscopic image, it is firstly rescaled
to 192 × 256 with 7 color channels and the CDNN model is
applied to yield a segmentation map. A dual-thresholds method
is then developed to generate a binary tumor mask from the
CDNN output. In this method, a relatively high threshold
(thH = 0.8) is firstly applied to determine the tumor center,
which is calculated as the centroid of the region that has the
largest mass among the candidates from thresholding. Then
a lower threshold, thL = 0.5, is applied to the segmentation
map. After filling small holes with morphological dilation, the
final tumor mask is determined as the region that embraces the
tumor center.
B. Implementation details
Our CDNN model is implemented with Python using publicly available Theano [23] and Lasagne 2 packages. The
model was trained from scratch using Adam stochastic optimization method [24] that adaptively adjusts the learning
rate based on the first and the second-order moments of the
gradient at each iteration. The initial learning rate α was set
as 0.003.
In order to reduce overfitting, we add two dropout layers
with p = 0.5 before conv-4-1 and decv-5-1 in Table I. In
addition, we implement two types of image augmentation to
further improve the robustness of the proposed model under
a wide variety of image acquisition conditions. One consists
of a series of geometric transformations, including randomly
flipping, shifting, rotating as well as scaling. The other type
focuses on randomly normalizing the contrast of each channels
in the training images. Note that these augmentations only
require little extra computation, so the transformed images are
generated from the original images for every mini-batch with
each iteration.
We used five-fold cross validation to evaluate the performance of our model on the challenge training dataset, in which
a few hype-parameters were also experimentally determined
via grid search. The total number of iterations was set as
600 for each fold. When applying the trained models onto
the testing dataset, a bagging-type ensemble strategy was
implemented to average the outputs of the six models to further
improve the segmentation performance. One iteration in model
training took about 60 seconds using a single NVIDIA Geforce
GTX 1060 GPU with 1280 cores and 6GB memory. Applying
the entire segmentation framework on a new test image was,
however, very efficient, taking about 0.2 second for a typical
768 × 1024 image.
2 http://github.com/Lasagne/Lasagne
3
TABLE II
C OMPARISON OF DIFFERENT ARCHITECTURES ON VALIDATION DATASET
Image size
CDNN-19
CDNN-29
AC
0.945
0.953
DI
0.839
0.865
JA
0.749
0.783
SE
0.854
0.879
SP
0.982
0.979
TABLE III
C OMPARISON OF
Image size
CDNN-19
CDNN-29
DIFFERENT ARCHITECTURES ON TESTING DATASET
AC
0.921
0.934
DI
0.824
0.849
IV. E XPERIMENTS
JA
0.736
0.765
SE
0.796
0.825
SP
0.979
0.975
AND RESULTS
A. Evaluation metrics
We applied the challenge evaluation metrics to evaluate
the performance of our method by comparing the computergenerated lesion masks with the ground truths created by
human experts. The evaluation metrics include pixel-wise accuracy (AC), sensitivity (SE), specificity (SP), dice coefficient
(DI), and Jaccard index (JA):
• AC = (T P + T N )/(T P + F P + T N + F N )
• DI = 2 · T P/(2 · T P + F N + F P )
• JA = T P/(T P + F N + F P )
• SE = T P/(T P + F N )
• SP = T N/(T N + F P ),
where T P , T N , F P , F N refer to the number of true
positives, true negatives, false positives, and false negatives
respectively. The final rank was based on JA for this challenge.
B. Experiments on network architectures
In order to investigate if the increase of network depth with
smaller kernel size can improve the discriminative capability
of CDNN and thus yield a better segmentation performance,
we compared the performance of the proposed deep CDNN,
which we denoted as CDNN-29, with the one proposed in
our previous work, which we denoted as CDNN-19. Since
only the results on the validation dataset could be obtained
during the challenge, this comparison was initially conducted
on the validation dataset, as shown in Table II. The ground
truth of testing dataset was held out by the organizer for
final performance evaluation during the challenge, and later
released to the public to encourage further investigations. So
we extended this comparison to the testing dataset in this
experiment, as shown in Table III.
It is clear to see that the new CDNN model achieved
better segmentation performance in most of the metrics on
both validation and testing datasets. These results demonstrate
that increasing network depth while maintaining equivalent
receptive field by reducing filter size can effectively improve
the discriminative capability of CDNN.
C. Experiments on input channels
Using CDNN-29 model, we evaluated how the additional
HSV and L channels affect skin lesion segmentation performance. Figure 2 shows how the loss function is minimized as
SUBMITTED TO IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
4
TABLE V
C OMPARISON OF DIFFERENT INPUT CHANNELS ON TESTING DATASET
Image size
RGB
RGB+HSV+L
AC
0.931
0.934
DI
0.840
0.849
JA
0.754
0.765
SE
0.813
0.825
SP
0.976
0.975
TABLE VI
R ESULTS OF 2017 ISBI CHALLENGE ON SKIN LESION SEGMENTATION
Fig. 2. Comparison of network inputs with and without the additional HSV+L
color channels for training and validating of the proposed model.
TABLE IV
C OMPARISON OF DIFFERENT INPUT CHANNELS ON VALIDATION DATASET
Image size
RGB
RGB+HSV+L
AC
0.951
0.953
DI
0.848
0.865
JA
0.765
0.783
SE
0.868
0.879
SP
0.976
0.979
network training evolves. The black lines represent the model
training and validation using RGB channels, and the red lines
add additional four channels as model input. It is clear to
see that the gap between training and validation errors was
effectively narrowed down when including the additional four
color channels as network input. Meanwhile, the validation
error with additional color channels was consistently lower
than that of RGB alone, yielding improved performance on
the new testing images. As shown in Table IV and V, the
additional four input channels improved the Jaccard index by
2.4% and 1.5% on validation and testing datasets, respectively.
D. Comparison with other methods in the challenge
During the 2017 ISBI challenge on skin lesion segmentation, 39 teams evaluated their algorithms during validation
phase, but only 21 teams were able to participate the final
official challenge by submitting their results on the 600 testing
images. Table VI lists the results from the top ten teams.
Because the previous studies have shown the advantage of
using CNNs in skin lesion segmentation over other traditional
methods [14], [17], most of the top teams in this challenge
included various deep neural network models in their segmentation methods, with the exception that the NedMos team
employed active contour on the saliency map [25]. While each
team was allowed for multiple submissions, only the most
recent one would be considered to determine the team’s final
score. Our method achieved an average Jaccard Index (JA) of
0.765, ranking as the first place in this challenge.
Figure 3 shows a few challenging examples of automatic
skin lesion segmentation with our CDNN model, demonstrating the robustness of the proposed model under various
imaging acquisition conditions.
V. D ISCUSSIONS
While deep convolutional neural networks have shown
remarkable success in various medical image segmentation
Team
MtSinai (ours)
NLP LOGIX
USYD (Bi)
USYD (Ann)
RECOD
Jer
NedMos
INESC
Shenzhen U (Lee)
GAMMA
AC
0.934
0.932
0.934
0.934
0.931
0.930
0.930
0.922
0.922
0.915
DI
0.849
0.847
0.844
0.842
0.839
0.837
0.839
0.824
0.810
0.797
JA
0.765
0.762
0.760
0.758
0.754
0.752
0.749
0.735
0.718
0.715
SE
0.825
0.820
0.802
0.801
0.817
0.813
0.810
0.813
0.789
0.774
SP
0.975
0.978
0.985
0.984
0.970
0.976
0.981
0.968
0.975
0.970
tasks, it is still a challenging task to extend those models
into a scaled-up study where a large amount of image data
are involved. In this study, we investigated if a deeper CDNN
model coupled with additional input information is capable of
handling the increased complexity introduced from the much
dynamic appearance of the targeting objects. It should be noted
that deep neural networks have a very flexible architecture and
different network design certainly matters. However, there is
no clear guideline about what the optimal network architecture
would be for a given application and available data. While it is
infeasible to explore all the possible options in our study due to
the long training time, we participated the ISBI 2017 challenge
on skin lesion segmentation, which not only provides a large
database to evaluate the performance of our model, but also
allows researchers to compare different designs using the same
benchmark data.
All of the top five teams employed deep learning in their
segmentation methods, demonstrating the popularity and effectiveness of FCN-based methods in medical image segmentation. These methods cover a large spectrum of FCN
models and training strategies. For example, Berseth from
NLP LOGIX [26] employed a U-Net architecture [27] and
tried to include a conditional random field (CRF) [28] as
post-processing in their method. Besides the 2000 training
images provided by the challenge organizers, Bi and Ann
from USYD [29] trained their deep residual networks with
additional ∼ 8000 images from ISIC archive. We attribute the
superior performance of our CDNN model to the following
three aspects: 1) A deeper architecture with small kernel
allowed more non-linear transformations on the data while
reducing the number of trainable parameters as compared
to its shallower counterpart, yielding significantly improved
discriminant capability; 2) Additional input features made network training more efficient and robust by explicitly including
complementary but useful information from other color spaces;
3) The loss function based on Jaccard distance enabled the
network training to naturally focus more on lesion pixels
over background, which further improved the segmentation
performance by lifting sensitivity (SE), as demonstrated in
SUBMITTED TO IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
5
Fig. 3. Examples of automatic segmentation on the ISBI 2017 challenge testing dataset, including four nevus lesions in the top row, four seborrheic keratosis
lesions in the middle rows and four melanoma lesions in the bottom row. In each figure, the red dash line indicates the outline contoured by dermatologist
and the yellow solid line is the result of automatic segmentation.
Table VI.
The segmentation performance on some cases is still rather
low, as shown in two examples in Figure 4. Thus, further
improvement is certainly needed. In additional to better network architecture and more effective training strategies, one
possible way is to combine the deep learning models with
the conventional image segmentation methods, such as active
contour. For example, we found the correlation between the
outputs of our model and MedMos (active contour model) was
0.759, while it was 0.873 with NLP LOGIX (another deep
FCN model). The scatter plots are shown as Figure 5. Since the
active contour is less correlated with our model, it may provide
more complementary information when ensembling multiple
models for further segmentation. Another option is to integrate
other post-processing techniques, such as CRF, into our CDNN
model. We actually attempted to employ this technique in
this challenge, but ultimately discard it due to the inferior
performance. One possible reason is that the parameters to
determine the unary and pariwise potentials were pre-set and
thus lack of flexibility to handle the large variations of the
data. A more dynamic mechanism, such as plugging CRF in
as a part of FCN by modeling it as recurrent neural networks
(RNNs) [30], may resolve this issue and thus improve the
segmentation performance. These are all interesting research
topics and worthy of further investigations.
Fig. 4. Cases where our model failed due to mis-segmentation to the wrong
sites. The red dash and yellow solid contours indicate the ground truth and
the segmentation results, respectively.
Fig. 5. Scatter plots of our model vesus NedMos (left), and NLP LOGIX
(right) .
VI. C ONCLUSION
In this paper, we present a major extension of our previously
proposed CDNN model in automatic skin lesion segmentation on dermoscopic images. Our new model leverages the
increased discriminant capability of deeper network structures
with smaller convolutional kernels to segment skin lesions in
a much larger variety of image acquisition conditions. The
segmentation performance was further boosted by combining
information from multiple color spaces. Our approach excelled
other state-of-the-art methods when evaluating on ISBI 2017
challenge of skin lesion segmentation. Our network architectures and training strategies are inherently general and can be
easily extended to other applications.
SUBMITTED TO IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS
VII. ACKNOWLEDGMENT
The authors are grateful to the organizers of 2017 International Symposium on Biomedical Imaging (ISBI 2017)
challenge of Skin Lesion Analysis T owards M elanoma
Detection for organizing such an inspiring challenge.
R EFERENCES
[1] R. L. Siegel, K. D. Miller, and A. Jemal, “Cancer statistics, 2016,” CA
Cancer J Clin., vol. 66, pp. 7–30, 2016.
[2] M. Vestergaard, P. Macaskill, P. Holt, and S. Menzies, “Dermoscopy
compared with naked eye examination for the diagnosis of primary
melanoma: a meta-analysis of studies performed in a clinical setting,”
Br. J. Dermatol., vol. 159, no. 3, pp. 669–676, 2008.
[3] K. Korotkov and R. Garcia, “Computerized analysis of pigmented skin
lesions: a review,” Artif. intell. Med., vol. 56, no. 2, pp. 69–90, 2012.
[4] H. Ganster, P. Pinz, R. Rohrer, E. Wildling, M. Binder, and H. Kittler,
“Automated melanoma recognition,” IEEE Trans. Med. Imag., vol. 20,
no. 3, pp. 233–239, 2001.
[5] M. E. Celebi, H. A. Kingravi, B. Uddin, H. Iyatomi, Y. A. Aslandogan,
W. V. Stoecker, and R. H. Moss, “A methodological approach to the
classification of dermoscopy images,” Comput. Med. Imag. Graph.,
vol. 31, no. 6, pp. 362–373, 2007.
[6] M. Silveira, J. C. Nascimento, J. S. Marques, A. R. Marçal, T. Mendonça, S. Yamauchi, J. Maeda, and J. Rozeira, “Comparison of segmentation methods for melanoma diagnosis in dermoscopy images,” IEEE
J. Sel. Top. Signal Process., vol. 3, no. 1, pp. 35–45, 2009.
[7] M. E. Celebi, H. Iyatomi, G. Schaefer, and W. V. Stoecker, “Lesion
border detection in dermoscopy images,” Comput. Med. Imag. Graph.,
vol. 33, no. 2, pp. 148–153, 2009.
[8] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[9] Z. Yan, Y. Zhan, Z. Peng, S. Liao, Y. Shinagawa, S. Zhang, D. N.
Metaxas, and X. S. Zhou, “Multi-instance deep learning: Discover
discriminative local anatomies for bodypart recognition,” IEEE Trans.
Med. Imag., vol. 35, no. 5, pp. 1332–1343, 2016.
[10] D. C. Cireşan, A. Giusti, L. M. Gambardella, and J. Schmidhuber,
“Mitosis detection in breast cancer histology images with deep neural
networks,” in Proc. MICCAI 2013. Springer, 2013, pp. 411–418.
[11] S. Miao, Z. J. Wang, and R. Liao, “A CNN Regression Approach for
Real-Time 2D/3D Registration,” IEEE Trans. Med. imag., vol. 35, no. 5,
pp. 1352–1363, 2016.
[12] K. H. Cha, L. Hadjiiski, R. K. Samala, H.-P. Chan, E. M. Caoili, and
R. H. Cohan, “Urinary bladder segmentation in CT urography using
deep-learning convolutional neural network and level sets,” Med. Phys.,
vol. 43, no. 4, pp. 1882–1896, 2016.
[13] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau,
and S. Thrun, “Dermatologist-level classification of skin cancer with
deep neural networks,” Nature, vol. 542, no. 7639, pp. 115–118, 2017.
[14] L. Yu, H. Chen, Q. Dou, J. Qin, and P. A. Heng, “Automated melanoma
recognition in dermoscopy images via very deep residual networks,”
IEEE Trans. Med. Imag., vol. 36, no. 4, pp. 994–1004, 2017.
[15] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in CVPR 2016. IEEE, 2016, pp. 770–778.
[16] L. Bi, J. Kim, E. Ahn, A. Kumar, and D. Feng, “Dermoscopic image
segmentation via multistage fully convolutional networks,” IEEE Trans.
Biomed. Eng., vol. 64, no. 9, pp. 2065–2074, 2017.
[17] Y. Yuan, M. Chao, and Y.-C. Lo, “Automatic skin lesion segmentation
using deep fully convolutional networks with jaccard distance,” IEEE
Trans. Med. Imaging, vol. 36, no. 9, pp. 1876–1886, 2017.
[18] D. Gutman, N. C. Codella, E. Celebi, B. Helba, M. Marchetti, N. Mishra,
and A. Halpern, “Skin Lesion Analysis toward Melanoma Detection:
A Challenge at the International Symposium on Biomedical Imaging
(ISBI) 2016, hosted by the International Skin Imaging Collaboration
(ISIC),” arXiv preprint arXiv:1605.01397, 2016.
[19] T. Mendonça, P. M. Ferreira, J. S. Marques, A. R. Marcal, and J. Rozeira,
“PH 2-A dermoscopic image database for research and benchmarking,”
in Proc. EMBC 2013. IEEE, 2013, pp. 5437–5440.
[20] K. Plataniotis and A. N. Venetsanopoulos, Color image processing and
applications. Springer Science & Business Media, 2013.
[21] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Adv. Neural Inf. Process.
Sys., 2012, pp. 1097–1105.
6
[22] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” arXiv preprint
arXiv:1502.03167, 2015.
[23] T. T. D. Team, R. Al-Rfou, G. Alain, A. Almahairi, C. Angermueller,
D. Bahdanau, N. Ballas, F. Bastien, J. Bayer, A. Belikov et al., “Theano:
A python framework for fast computation of mathematical expressions,”
arXiv preprint arXiv:1605.02688, 2016.
[24] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
arXiv preprint arXiv:1412.6980, 2014.
[25] M. Jahanifar, N. Z. Tajeddin, A. Goodya, and B. M. Asl, “Segmentation
of lesions in dermoscopy images using saliency map and contour
propagation,” arXiv preprint arXiv:1703.00087, 2017.
[26] M. Berseth, “ISIC 2017 - Skin lesion analysis towards melanoma
detection,” arXiv preprint arXiv:1703.00523, 2017.
[27] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks
for biomedical image segmentation,” in Proc. MICCAI 2015. Springer,
2015, pp. 234–241.
[28] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille,
“Semantic image segmentation with deep convolutional nets and fully
connected CRFs,” arXiv preprint arXiv:1412.7062, 2014.
[29] L. Bi, K. Kim, E. Ann, and D. Feng, “Automatic skin lesion analysis
using large-scale dermoscopy images and deep residual networks,” arXiv
preprint arXiv:1703.04197, 2017.
[30] S. Zheng, S. Jayasumana, B. Romera-Paredes, V. Vineet, Z. Su, D. Du,
C. Huang, and P. Torr, “Conditional random fields as recurrent neural
networks,” in ICCV, 2015, pp. 1529–1537.
| 1cs.CV
|
The interplay between system identification and machine learning
Gianluigi Pillonetto a
arXiv:1612.09158v1 [cs.SY] 29 Dec 2016
a Department
of Information Engineering, University of Padova, Padova, Italy (e-mail: [email protected])
Abstract
Learning from examples is one of the key problems in science and engineering. It deals with function reconstruction from a finite set of
direct and noisy samples. Regularization in reproducing kernel Hilbert spaces (RKHSs) is widely used to solve this task and includes
powerful estimators such as regularization networks. Recent achievements include the proof of the statistical consistency of these kernelbased approaches. Parallel to this, many different system identification techniques have been developed but the interaction with machine
learning does not appear so strong yet. One reason is that the RKHSs usually employed in machine learning do not embed the information
available on dynamic systems, e.g. BIBO stability. In addition, in system identification the independent data assumptions routinely adopted
in machine learning are never satisfied in practice. This paper provides new results which strengthen the connection between system
identification and machine learning. Our starting point is the introduction of RKHSs of dynamic systems. They contain functionals over
spaces defined by system inputs and allow to interpret system identification as learning from examples. In both linear and nonlinear
settings, it is shown that this perspective permits to derive in a relatively simple way conditions on RKHS stability (i.e. the property of
containing only BIBO stable systems or predictors), also facilitating the design of new kernels for system identification. Furthermore, we
prove the convergence of the regularized estimator to the optimal predictor under conditions typical of dynamic systems.
Key words: learning from examples; system identification; reproducing kernel Hilbert spaces of dynamic systems; kernel-based
regularization; BIBO stability; regularization networks; generalization and consistency;
1 Introduction
Learning from examples is key in science and engineering,
considered at the core of intelligence’s understanding [56].
In mathematical terms, it can be described as follows. We
are given a finite set of training data (xi , yi ), where xi is the
so called input location while yi is the corresponding output measurement. The goal is then the reconstruction of a
function with good prediction capability on future data. This
means that, for a new pair (x, y), the prediction g(x) should
be close to y.
To solve this task, nonparametric techniques have been extensively used in the last years. Within this paradigm, instead
of assigning to the unknown function a specific parametric
structure, g is searched over a possibly infinite-dimensional
functional space. The modern approach uses Tikhonov regularization theory [74,13] in conjunction with Reproducing Kernel Hilbert Spaces (RKHSs) [8,12]. RKHSs possess many important properties, being in one to one corre1
This research has been partially supported by the MIUR FIRB
project RBFR12M3AC-Learning meets time: a new computational
approach to learning in dynamic systems and by the Progetto
di Ateneo CPDA147754/14-New statistical learning approach for
multi-agents adaptive estimation and coverage control. This paper
was not presented at any IFAC meeting. Corresponding author
Gianluigi Pillonetto Ph. +390498277607.
Preprint submitted to Automatica
spondence with the class of positive definite kernels. Their
connection with Gaussian processes is also described in
[35,42,11,5].
While applications of RKHSs in statistics, approximation
theory and computer vision trace back to [14,76,54], these
spaces were introduced to the machine learning community
in [29]. RKHSs permit to treat in an unified way many different regularization methods. The so called kernel-based
methods [25,62] include smoothing splines [76], regularization networks [54], Gaussian regression [57], and support
vector machines [23,75]. In particular, a regularization network (RN) has the structure
(yi − f (xi ))2
+ γ k f k2H
∑
N
i=1
N
ĝ = arg min
f ∈H
RN
(1)
where H denotes a RKHS with norm k · kH . Thus, the
function estimate minimizes an objective sum of two contrasting terms. The first one is a quadratic loss which
measures the adherence to experimental data. The second
term is the regularizer (the RKHS squared norm) which
restores the well-posedness and makes the solution depend
continuously on the data. Finally, the positive scalar γ is
the regularization parameter which has to suitably trade off
these two components.
The use of (1) has significant advantages. The choice of
30 December 2016
identification and machine learning stems also from the (apparently) different contexts these disciplines are applied to.
In machine learning one typically assumes that data (xi , yi )
are i.i.d. random vectors assuming values on a bounded subset of the Euclidean space. But in system identification, even
when the system input is white noise, the input locations are
not mutually independent. Already in the classical Gaussian
noise setting, the outputs are not even bounded, i.e. there is
no compact set containing them with probability one. Remarkably, this implies that none of the aforementioned consistency results developed for kernel-based methods can be
applied. Some extensions to the case of correlated samples
can be found in [78,32,68] but still under conditions far from
the system identification setting.
an appropriate RKHS, often obtained just including function smoothness information [62], and a careful tuning of
γ , e.g. by the empirical Bayes approach [43,3,4], can well
balance bias and variance. One can thus obtain favorable
mean squared error properties. Furthermore, even if H is
infinite-dimensional, the solution ĝ is always unique, belongs to a finite-dimensional subspace and is available in
closed-form. This result comes from the representer theorem [34,61,7,6]. Building upon the work [77], many new
results have been also recently obtained on the statistical
consistency of (1). In particular, the property of ĝ to converge to the optimal predictor as the data set size grows
to infinity is discussed e.g. in [66,81,80,46,55]. This point
is also related to Vapnik’s concepts of generalization and
consistency [75], see [25] for connections among regularization in RKHS, statistical learning theory and the concept
of Vγ dimension as a measure of function class complexity
[2,24]. The link between consistency and well-posedness is
instead discussed in [15,46,55].
In this paper we provide some new insights on the interplay
between system identification and machine learning in a
RKHS setting. Our starting point is the introduction of what
we call RKHSs of dynamic systems which contain functionals over input spaces X induced by system inputs u. More
specifically, each input location x ∈ X contains a piece of
the trajectory of u so that any g ∈ H can be associated to a
dynamic system. When u is a stationary stochastic process,
its distribution then defines the probability measure on X
from which the input locations are drawn. Again, we stress
that this framework has been (at least implicitly) used in
previous works on nonlinear system identification, see e.g.
[64,51,73,38,63]. However, it has never been cast and studied in its full generality under a RKHS perspective.
Parallel to this, many system identification techniques have
been developed in the last decades. In linear contexts, the
first regularized approaches trace back to [60,1,36], see
also [30,41] where model error is described via a nonparametric structure. More recent approaches, also inspired by
nuclear and atomic norms [17], can instead be found in
[39,31,45,58,48]. In the last years, many nonparametric
techniques have been proposed also for nonlinear system
identification. They exploit e.g. neural networks [38,63],
Volterra theory [26], kernel-type estimators [37,51,82]
which include also weights optimization to control the mean
squared error [59,9,10]. Important connections between
kernel-based regularization and nonlinear system identification have been also obtained by the least squares support
vector machines [72,71] and using Gaussian regression for
state space models [27,28]. Most of these approaches are
inspired by machine learning, a fact not surprising since
predictor estimation is at the core of the machine learning
philosophy. Indeed, a black-box relationship can be obtained through (1) using past inputs and outputs to define
the input locations (regressors). However, the kernels currently used for system identification are those conceived by
the machine learning community for the reconstruction of
static maps. RKHSs suited to linear system identification,
e.g. induced by stable spline kernels which embed information on impulse response regularity and stability, have been
proposed only recently [52,50,18]. Furthermore, while stability of a RKHS (i.e. its property of containing only stable
systems or predictors) is treated in [16,53,22], the nonlinear
scenario still appears unexplored. Beyond stability, we also
notice that the most used kernels for nonlinear regression,
like the Gaussian and the Laplacian [62], do not include
other important information on dynamic systems like the
fact that output energy is expected to increase if input energy augments.
At first sight, our approach could appear cumbersome. In
fact, the space X can turn out complex and unbounded just
when the system input is Gaussian. Also, X could be a
function space itself (as e.g. happens in continuous-time). It
will be instead shown that this perspective is key to obtain
the following achievements:
• linear and nonlinear system identification can be treated
in an unified way in both discrete- and continuous-time.
Thus, the estimator (1) can be used in many different contexts, relevant for the control community, just changing
the RKHS. This is important for the development of a
general theory which links regularization in RKHS and
system identification;
• system input’s role in determining the nature of the RKHS
is made explicit. This will be also described in more detail
in the linear system context, illustrating the distinction
between the concept of RKHSs H of dynamic systems
and that of RKHSs I of impulse responses;
• for linear systems we provide a new and simple derivation of the necessary and sufficient condition for RKHS
stability [16,53,22] that relies just on basic RKHS theory;
• in the nonlinear scenario, we obtain a sufficient condition
for RKHS stability which has wide applicability. We also
derive a new stable kernel for nonlinear system identification;
• consistency of the RN (1) is proved under assumptions
suited to system identification, revealing the link between
Another aspect that weakens the interaction between system
2
Definition 2 (Positive definite kernel and kernel section)
A symmetric function K : X × X → R is called positive
definite kernel if, for any integer p, it holds
consistency and RKHS stability.
The paper is organized as follows. In Section 2 we provide a brief overview on RKHSs. In Section 3, the concept
of RKHSs of dynamic systems is defined by introducing
input spaces X induced by system inputs. The case of
linear dynamic systems is then detailed via its relationship
with linear kernels. The difference between the concepts
of RKHSs of dynamic systems and RKHSs of impulse responses is also elucidated. Section 4 discusses the concept
of stable RKHS. We provide a new simple characterization
of RKHS stability in the linear setting. Then, a sufficient
condition for RKHS stability is worked out in the nonlinear
scenario. We also introduce a new kernel for nonlinear system identification, testing its effectiveness on a benchmark
problem. In Section 5, we first review the connection between the machine learning concept of regression function
and that of optimal predictor encountered in system identification. Then, the consistency of the RN (1) is proved
in the general framework of RKHSs of dynamic systems.
Conclusions end the paper while proofs of the consistency
results are gathered in Appendix.
p
p
∑ ∑ ci c j K (xi , x j ) ≥ 0,
∀(xk , ck ) ∈ (X , R) ,
i=1 j=1
k = 1, . . . , p.
The kernel section Kx centered at x is the function from X
to R defined by
Kx (a) = K (a, x) ∀a ∈ X .
The following theorem provides the one-to-one correspondence between RKHSs and positive definite kernels.
Theorem 3 (Moore-Aronszajn and reproducing property)
To every RKHS H there corresponds a unique positive definite kernel K such that the so called reproducing property
holds, i.e.
hKx , gi = g(x) ∀(x, g) ∈ (X , H )
In what follows, the analysis is always restricted to causal
systems and, to simplify the exposition, the input locations
contain only past inputs so that output error models are considered. If an autoregressive part is included, the consistency
analysis in Section 5 remains unchanged while the conditions developed in Section 4 guarantee predictor (in place
of system) stability.
Conversely, given a positive definite kernel K , there exists
a unique RKHS of real-valued functions defined over X
where (3) holds.
Theorem 3 shows that a RKHS H is completely defined by
a kernel K , also called the reproducing kernel of H . More
specifically, it can be proved that any RKHS is generated
by the kernel sections in the following manner. Let S denote
the subspace spanned by {Kx }x∈X and for any g ∈ S, say
p
g = ∑i=1
ci Kxi , define the norm
2 Brief overview on RKHSs
p
We use X to indicate a function domain. This is a nonempty set often referred to as the input space in machine
learning. Its generic element is the input location, denoted
by x or a in the sequel. All the functions are assumed real
valued, so that g : X → R.
In function estimation problems, the goal is to estimate maps
to make predictions over the whole X . Thus, a basic requirement is to use an hypothesis space H with functions
well defined pointwise for any x ∈ X . In particular, assume
that all the pointwise evaluators g → g(x) are linear and
bounded over H , i.e. ∀x ∈ X there exists Cx < ∞ such that
|g(x)| ≤ Cx kgkH ,
∀g ∈ H .
(3)
p
kgk2H = ∑ ∑ ci c j K (xi , x j ).
(4)
i=1 j=1
Then, one has that H is the union of S and all the limits w.r.t. k · kH of the Cauchy sequences contained in S. A
consequence of this construction is that any g ∈ H inherits
kernel properties, e.g. continuity of K implies that all the
g ∈ H are continuous [19][p. 35].
The kernel sections play a key role also in providing the
closed-form solution of the RN (1), as illustrated in the famous representer theorem.
(2)
Theorem 4 (Representer theorem) The solution of (1) is
unique and given by
This property already leads to the spaces of interest.
N
ĝ = ∑ ĉi Kxi ,
Definition 1 (RKHS) A reproducing kernel Hilbert space
(RKHS) H over X is a Hilbert space containing functions
g : X → R where (2) holds.
(5)
i=1
where the scalars ĉi are the components of the vector
RKHSs are connected to the concept of positive definite
kernel, a particular function defined over X × X .
ĉ = (K + γ NIN )−1 Y,
3
(6)
Y is the column vector with i-th element yi , IN is the N × N
identity matrix and the (i, j) entry of K is K (xi , x j ).
t ∈ Z . Different cases arise depending on the postulated
system model. For example, one can have
Another RKHS characterization useful in what follows is
obtained when the kernel can be diagonalized as follows
xt = [ut ut−1 . . . ut−m+1 ]T ,
where m is the system memory. This construction is connected to FIR or NFIR models and makes X a subset of
the classical Euclidean space Rm .
Another scenario is
∞
K (a, x) = ∑ ζi ρi (a)ρi (x), ζi > 0 ∀i.
(11)
(7)
i=1
The RKHS is then separable and the following result holds,
e.g. see [25][p. 15] and [19][p. 36].
xt = [ut ut−1 ut−2 . . .]T ,
Theorem 5 (Spectral representation of a RKHS) Let (7)
hold and assume that the ρi form a set of linearly independent functions on X . Then, one has
where any input location is a sequence (an infinitedimensional column vector) and the input space X becomes a subset of R∞ . The definition (12) is related to
infinite memory systems, e.g. IIR models in linear settings.
H =
and
(
)
c2i
g | g(x) = ∑ ci ρi (x) s.t. ∑ < ∞ ,
i=1
i=1 ζi
∞
∞
b i ci
,
i=1 ζi
h f , giH = ∑
∞
∞
k f k2H = ∑
∞
where f = ∑∞
i=1 bi ρi and g = ∑i=1 ci ρi .
i=1
b2i
ζi
,
(8)
Continuous-time The continuous-time input is the map
u : R → R. In this case, the input location xt becomes the
function xt : R+ → R defined by
(9)
xt (τ ) = u(t − τ ),
X
K (·, x)ρi (x)d µx (x) = ζi ρi (·),
τ ≥ 0,
(13)
i.e. xt contains the input’s past up to the instant t. In many
circumstances, one can assume X ⊂ P c , where P c contains piecewise continuous functions on R+ . When the input
is causal, and ut is smooth for t ≥ 0, the xt is indeed piecewise continuous.
Note that (13) is the continuous-time counterpart of (12)
while that of (11) can be obtained just zeroing part of the
input location, i.e.
The expansion (7) can e.g. be obtained by the Mercer theorem [44,33]. In particular, let µx be a nondegenerate σ -finite
measure on X . Then, under somewhat general conditions
[70], the ρi and ζi in (7) can be set to the eigenfunctions and
eigenvalues of the integral operator induced by K , i.e.
Z
(12)
0 < ζ1 ≤ ζ2 ≤ . . .
(10)
In addition, the ρi form a complete orthonormal basis in the
µ
classical Lebesgue space L2 x of functions square integrable
under µx . 1
xt (τ ) = u(t − τ )ξT (τ ),
τ ≥ 0,
(14)
where ξT is the indicator function of the interval [0, T ].
In linear systems, (14) arises when the impulse response
support is compact.
3 RKHSs of dynamic systems and the linear system
scenario
RKHSs H of functions over domains X , induced by system inputs u as illustrated above, are hereby called RKHSs
of dynamic systems. Thus, if g ∈ H , the scalar g(xt ) is the
noiseless output at t of the system fed with the input trajectory contained in xt . Note that g in general is a functional: in
the cases (12-14) the arguments xt entering g(·) are infinitedimensional objects.
3.1 RKHSs of dynamic systems
The definition of RKHSs of dynamic systems given below
relies on simple constructions of input spaces X induced
by system inputs u.
3.2 The linear system scenario
Discrete-time First, the discrete-time setting is considered. Assume we are given a system input u : Z → R. Then,
we think of any input location in X indexed by the time
RKHSs of linear dynamic systems are now introduced also
discussing the structure of the resulting RN.
Linear system identification was faced In [52] and [53][Part
III] by introducing RKHSs of impulse responses. These are
spaces I induced by kernels K defined over subsets of
R+ × R+ . They thus contain causal functions, each of them
1
Thus, the representation (8) is not unique since spectral maps
are not unique. Eigendecompositions depend on the measure µx
but lead to the same RKHS.
4
representing an impulse response θ . The RN which returns
the impulse response estimate was
(yi − (θ ⊗ u)ti )2
+ γ kθ k2I ,
θ̂ = arg min ∑
θ ∈I i=1
N
row equal to xTi , where
yi := yti ,
xi := xti
N
(15)
and ti is the time instant where the i-th output is measured.
Then, after plugging the representation g(x) = θ T x in (1),
one obtains ĝ(x) = θ̂ T x with
where (θ ⊗ u)ti is the convolution between the impulse response and the input evaluated at ti .
The RKHSs of linear dynamic systems here introduced are
instead associated to (output) linear kernels K defined on
X × X through convolutions of K with system inputs. In
particular, if xt and xτ are as in (11-14), one has 2
θ̂ = arg min kY − Φθ k2 + γθ T K −1 θ
θ ∈Rm
T
= (Φ Φ + γ P−1)−1 ΦT Y.
which e.g. in continuous-time becomes
K (xt , xτ ) =
0
u(t − α )
Z
0
+∞
IIR models Consider now the input locations defined by (12). The input space contains sequences and
X ⊆ R∞ . Interpreting any input location as an infinitedimensional column, we can use ordinary algebra’s notation to handle infinite-dimensional objects. For example,
if (a, x) ∈ (X , X ) then aT x = ha, xi2 , where h·, ·i2 is the
inner-product in the classical space ℓ2 of squared summable
sequences.
Let K be symmetric and positive semidefinite infinitedimensional matrix K (the nature of K is discussed also in
Section 3.3). Then, the function
u(τ − β )K(α , β )d β d α .
(16)
These kernels lead to the RN (1) which corresponds to (15)
after the “reparametrization” g(xt ) = (θ ⊗ u)t so that, in
place of the impulse response θ , the optimization variable
becomes the functional g(·).
The kernels K arising in discrete- and continuous-time are
described below in (17,19) and (22). The distinction between
the RKHSs induced by K and K will be further discussed
in Section 3.3.
K (x, a) = xT Ka,
FIR models We start assuming that the input location is
defined by (11) so that any xt is an m-dimensional (column)
vector and X ⊆ Rm . If K ∈ Rm×m is a symmetric and positive semidefinite matrix, a linear kernel is defined as follows
K (a, x) = aT Kx,
(a, x) ∈ Rm × Rm .
(18b)
The nature of the input locations (11) shows that θ̂ is the
impulse response estimate. Thus, (18) corresponds to regularized FIR estimation as e.g. discussed in [18].
K (xt , xτ ) = (u ⊗ (K ⊗ u)τ )t
Z +∞
(18a)
(x, a) ∈ R∞ × R∞
(19)
defines a linear kernel on X × X . Following arguments
similar to those developed in the FIR case, one can see that
the RKHS associated to such K contains linear functions of
the form g(x) = aT Kx with a ∈ R∞ . Note that each g ∈ H
is a functional defined by the sequence aT K which represents an impulse response. In fact, one can deduce from (12)
that g(xt ) is the discrete-time convolution, evaluated at t, between u and aT K.
The RN with H induced by (19) now implements regularized IIR estimation. Roughly speaking, (1) becomes the
limit of (18) for m → ∞. The exact solution can be obtained
by the representer theorem (5) and turns out
(17)
All the kernel sections are linear functions. Their span defines a finite-dimensional (closed) subspace that, in view
of the discussion following Theorem 3, coincides with the
whole H . Hence, H is a space of linear functions: for any
g ∈ H , there exists a ∈ Rm such that
g(x) = aT Kx = Ka (x).
N
ĝ(x) = ∑ ĉi Kxi (x) = θ̂ T x,
(20)
i=1
If K is full rank, it holds that
where the ĉi are the components of (6) while the infinitedimensional column vector
||g||2H = ||Ka ||2H = hKa , Ka iH
= K (a, a) = aT Ka
= θ T K −1 θ with θ := Ka.
N
θ̂ := ∑ ĉi Kxi
(21)
i=1
Let us use the H associated to (17) as hypothesis space for
the RN in (1). Let Y = [y1 . . . yN ]T and Φ ∈ RN×m with i-th
contains the impulse response coefficients estimates.
Continuous-time The continuous-time scenario arises
considering the input locations defined by (13) or (14). The
input space X now contains causal functions. Considering
2
Translated in a stochastic setting, the (output) kernel K can be
seen as the covariance of a causal random process of covariance
K filtered by u.
5
(13), given a positive-definite kernel K : R+ × R+ → R, the
linear kernel K is
K (x, a) =
Z
spaces are isometrically isomorphic since
∞
R+ ×R+
K(t, τ )x(t)a(τ )dtd τ
c2i
= kθ k2I .
ζ
i
i=1
kgk2H = ∑
(22)
which coincides with (16) when x = xt and a = xτ . Each
kernel section Kx (·) is a continuous-time
linear system with
R
impulse response θ (·) = R+ K(·,t)x(t)dt. Thus, the corresponding RKHS contains linear functionals and (1) now
implements regularized system identification in continuoustime. Using the representer theorem, the solution of (1) is
N
ĝ(x) = ∑ ĉi Kxi (x) =
i=1
This result holds also in continuous-time where I is now the
RKHS associated to the kernel K : R+ × R+ → R. Letting ψi
be real-valued functions on R+ , this comes from the same
arguments adopted in discrete-time but now applied to the
expansions
∞
K(t, τ ) = ∑ ζi ψi (t)ψi (τ ),
i=1
Z
R+
θ̂ (τ )x(τ )d τ
(23)
and
N
θ̂ (τ ) := ∑ ĉi
i=1
Z
R+
K(τ ,t)xi (t)dt.
Z
K(t, τ )x(t)a(τ )dtd τ
Z
= ∑ ζi
ψi (t)x(t)dt
ψi (τ )a(τ )d τ .
R
i=1 | R+
{z
} | + {z
}
where ĉ is still defined by (6) while θ̂ is the impulse response
estimate given by
K (x, a) =
(24)
R+ ×R+
Z
∞
ρi (x)
ρi (a)
Remark 6 The functionals ρi could turn out linearly dependent even if the ψi composing K are linearly independent. This depends on the nature of the input space. For instance, let X contain only the input locations induced by
ut = sin(ω t). Then, if ψi is a rational transfer function with
zeros ± jω , the functional ρi associated to ψi vanishes over
X . In these cases, there is no isometry between H and
I : the same dynamic system g could be defined by different
impulse responses θ i ∈ I . In particular, results on RKHSs
induced by sums of kernels reported in [8][Section 6 on p.
352] allow us to conclude that kgkH = mini kθ i kI . Thus,
among all the possible equivalent representations in I of
the dynamic system g ∈ H , the complexity of g is quantified by that of minimum norm. This is illustrated through the
following simple continuous-time example. Assume that
3.3 Relationship between RKHSs of impulse responses and
RKHSs of dynamic systems
In (19), the infinite-dimensional matrix K represents a kernel over N × N. Then, let I be the corresponding RKHS
which contains infinite-dimensional column vectors θ =
[θ1 θ2 . . .]T . We will now see that I is the RKHS of impulse
responses associated to H , i.e. each θ ∈ I is the impulse
response of a linear system g ∈ H . In particular, let K admit the following expansion in terms of linearly independent
infinite-dimensional (column) vectors ψi :
∞
K = ∑ ζi ψi ψiT .
i=1
K(t, τ ) = ζ1 ψ1 (t)ψ1 (τ ) + ζ2 ψ2 (t)ψ2 (τ ),
According to Theorem 5, the span of the ψi provides all the
∞
2
θ ∈ I . Moreover, if θ = ∑∞
i=1 ci ψi , then kθ kI = ∑i=1
The equality
T
K (a, x) = a Kx = a
T
∞
∑
ζi ψi ψiT
i=1
∞
= ∑ ζi aT ψi ψiT x ,
i=1 | {z } | {z }
ρi (a)
!
c2i
ζi .
where the Laplace transforms of ψ1 and ψ2 are given, respectively, by the rational transfer functions
2s
√ ,
W1 (s) =
s+1+ 2
x
√
s+1− 2
W2 (s) =
,
s+1
√
√
which satisfy W1 ( −1) = W2 ( −1). Let the input space
contain only the input locations induced by ut = sin(t). Then,
the functionals (ρ1 , ρ2 ) associated, respectively, to (ψ1 , ψ2 )
coincide over the entire X . Thus, the two impulse responses
ψ1 and ψ2 induce the same system g ∈ H . Using Theorem
5, one has
ρi (x)
also provides the expansion of K in terms of functionals
ρi (·) defined by ρi (x) := ψiT x. Assuming that such functionals are linearly independent, it comes from Theorem 5
that each dynamic system g ∈ H has the representation
g(·) = ∑∞
i=1 ci ρi (·). It is now obvious that such system is associated to the impulse response θ = ∑∞
i=1 ci ψi and the two
kψ1 k2I =
6
1
,
ζ1
kψ2 k2I =
1
,
ζ2
which implies
4.1 The linear system scenario
kgk2H = min
1 1
,
ζ1 ζ2
We start studying the stability of RKHSs of linear dynamic
systems. Obviously, all the FIR kernels (17) induce stable
RKHSs. As for the IIR and continuous-time kernels in (19)
and (22), first it is useful to recall the classical result linking
BIBO stability and impulse response summability.
.
4 Stable RKHSs
Proposition 9 (BIBO stability and impulse response
summability) Let θ be the impulse response of a linear system. Then, the system is BIBO stable iff θ ∈ ℓ1 in
discrete-time or θ ∈ L1 in continuous-time.
BIBO stability of a dynamic system is a familiar notion
in control. In the RKHS context, we have the following
definition.
Definition 7 (stable dynamic system) Let u be any
bounded input, i.e. satisfying |ut | < Mu < ∞ ∀t. Then, the
dynamic system g ∈ H is said to be (BIBO) stable if there
exits a constant My < ∞ such that |g(xt )| < My for any t
and any input location xt induced by u.
The next proposition provides the necessary and sufficient
condition for RKHS stability in the linear scenario.
Proposition 10 (RKHS stability in the linear case) Let
H be the RKHS of dynamic systems g : X → R induced
by the IIR kernel (19). Then, the following statements are
equivalent
Note that, for g ∈ H to be stable, the above definition implicitly requires the input space of H to contain any xt induced by any bounded input.
Definition 8 (stable RKHS) Let H be a RKHS of dynamic
systems induced by the kernel K . Then, H and K are said
to be stable if each g ∈ H is stable.
(1) H is stable;
(2) The input space X contains ℓ∞ so that
g(a) < ∞ for any (g, a) ∈ (H , ℓ∞ );
To derive stability conditions on the kernel, let us first introduce some useful Banach spaces. The first two regard the
discrete-time setting:
∞
(3) ∑∞
i=1 ∑ j=1 K(i, j)a j < ∞ for any a ∈ ℓ∞ .
• the space ℓ1 of absolutely summable real sequences a =
[a1 a2 . . .], i.e. such that ∑∞
i=1 |ai | < ∞, equipped with the
norm
Let instead H be the RKHS induced by the continuous-time
kernel (22). The following statements are then equivalent
∞
kak1 = ∑ |ai |;
(1) H is stable;
(2) The input space X contains L∞ so that
i=1
• the space ℓ∞ of bounded real sequences a = [a1 a2 . . .],
i.e. such that supi |ai | < ∞, equipped with the norm
g(a) < ∞ for any (g, a) ∈ (H , L∞ );
kak∞ = sup |ai |.
(3)
i
The other two are concerned with continuous-time:
Z
R+
R+
R
R+ K(t, τ )a(τ )d τ
dt < +∞ for any a ∈ L∞ .
Proof: The proof is developed in discrete-time. The
continuous-time case follows exactly by the same arguments with minor modifications.
(1) → (2) Recalling Definition 7 and subsequent discussion, this is a direct consequence of the BIBO stability
assumption of any g ∈ H .
(2) → (1) Given any g ∈ H , let θ = [θ1 θ2 . . .]T its associated impulse response and define
• the Lebesgue space L1 of functions
a : R+ → R absoR
lutely integrable, i.e. such that R+ |a(t)|dt < ∞, equipped
with the norm
kak1 =
R
|a(t)|dt;
xt = [sign(θ1 ) sign(θ2 ) . . .]T .
• the Lebesgue space L∞ of functions a : R+ → R essentially bounded, i.e. for any a there exists Ma such that
(27)
The assumption g(xt ) = θ T xt = kθ k1 < ∞ implies that
θ ∈ ℓ1 and the implication follows by Proposition 9.
(2) → (3) By assumption, the kernel is well defined over
the entire ℓ∞ × ℓ∞ . Hence, any kernel section Ka centred on
a ∈ ℓ∞ is a well defined element in H and corresponds to a
dynamic system with associated impulse response θ = Ka.
|a(t)| ≤ Ma almost everywhere in R+ ,
equipped with the norm
kak∞ = inf {M s.t. |a(t)| ≤ M a.e.} .
7
With xt still defined by (27), one has Ka (xt ) = kθ k1 < ∞
which implies Ka ∈ ℓ1 and proves (3).
(3) → (2) By assumption, any impulse response associated
to any kernel section centred on a ∈ ℓ∞ belongs to ℓ1 . This
implies that the kernel K associated to H is well defined
over the entire ℓ∞ × ℓ∞ . Recalling Definition 1 and eq. (2),
RKHS theory then ensures that any g ∈ H is well defined
pointwise on ℓ∞ and g(a) < ∞ ∀a ∈ ℓ∞ .
Proof: Let the system input u ∈ ℓ∞ in discrete-time or u ∈ L∞
in continuous-time. Then, we can find a closed ball Br∞ containing, for any t, all the input locations xt induced by u as
defined in (11-14). For any g ∈ H and xt ∈ Br∞ , exploiting
the reproducing property and the Cauchy-Schwartz inequality, one obtains
|g(xt )| = |hg, Kxt iH | ≤ kgkH kKxt kH
p
p
= kgkH K (xt , xt ) ≤ kgkH Cr ,
hence proving the stability of H .
Point (3) contained in Proposition 10 was also cited in
[53,22] as a particularization of a quite involved and abstract result reported in [16]. The stability proof reported
below turns instead out surprisingly simple. The reason is
that, with the notation adopted in (19) and (22), the outcomes in [16] were obtained starting from spaces I of impulse responses induced by K. Our starting point is instead
the RKHSs H of dynamic systems induced by K (in turn
defined by K). This different perspective permits to greatly
simplify the analysis: kernel stability can be characterized
just combining basic RKHS theory and Proposition 9.
The following result will be also useful later on. It derives
from the fact that kernels products (sums) induce RKHSs of
functions which are products (sums) of the functions induced
by the single kernels [8][p. 353 and 361].
Proposition 12 (RKHS stability with kernels sum and
product) Let K1 and K2 be stable kernels. Then, the RKHSs
induced by K1 × K2 and K1 + K2 are both stable.
The two propositions above allow to easily prove the stability
of a very large class of kernels, as discussed in discrete-time
in the remaining part of this section.
Proposition 10 shows that RKHS stability is implied by the
absolute integrability of K, i.e. by
∞
∞
∑ ∑ |K(i, j)| < ∞ or
i=1 j=1
Z
R+ ×R+
|K(t, τ )|dtd τ < ∞ (28)
Radial basis kernels First, consider the input locations
(11) contained in X ⊆ Rm . As already mentioned in Section
4.1, radial basis kernels K (x, a) = h(|x− a|), with |·| now to
indicate the Euclidean norm, are widely adopted in machine
learning. Important examples are the Gaussian kernel
in discrete- and continuous-time, respectively. The condition (28) is also necessary for nonnegative-valued kernels
[53][Section 13]. Then, considering e.g. the continuous-time
setting, the popular Gaussian and Laplacian kernels, which
belong to the class of radial basis kernels K(t, s) = h(|s − t|)
for t, s ≥ 0, are all unstable. Stability instead holds for the
stable spline kernel [49] given by:
K(t, s) = e−β max(t,s)
t, s ≥ 0
|x − a|2
K (x, a) = exp −
,
η
η >0
(31)
η > 0.
(32)
and the Laplacian kernel
(29)
|x − a|
,
K (x, a) = exp −
η
where β > 0 is related to the impulse response’s dominant
pole.
From Proposition 11 one immediately sees that both these
kernels are stable. More in general, all the radial basis kernels are stable 3 since they are constant along their diagonal
(K (x, x) = h(0)).
However, despite their stability, some drawbacks affect the
use of (31,32) in system identification. First, the fact that
K (x, x) is constant implies that these models do not include
the information that output energy is likely to increase if
4.2 The nonlinear system scenario
Let us now consider RKHSs of nonlinear dynamic systems
with input locations (11-14). A very simple sufficient condition for RKHS stability is reported below.
Proposition 11 (RKHS stability in the nonlinear case)
Let H be a RKHS of dynamic systems induced by the kernel K . Let Br∞ denote the closed ball of radius r induced
by k · k∞ , contained in ℓ∞ or Rm in discrete-time, or in L∞
in continuous-time. Assume that, for any r, there exists Cr
such that
K (x, x) < Cr < ∞, ∀x ∈ Br∞ .
Then, the RKHS H is stable.
3 This statement should not be confused with the result discussed
in Section 4.1 in the linear system scenario. There, we have seen
that radial basis kernels lead to unstable linear kernels K when
used to define K in the IIR (19) and continuous-time (22) case.
Here, the Gaussian and Laplace kernels are instead used to define
directly K in the nonlinear system scenario.
8
where all the ut and et are independent Gaussian noises of
variance 4. Note that (S2) contains the sum of a linear time
invariant system (details on the impulse response θ are given
below) and the nonlinear FIR in (S1). Our aim is to identify
the two systems from 1000 input-output pairs (xt , yt ) via (1).
The performance will be measured by the percentage fit on
a test set of 1000 noiseless system outputs contained in the
vector ytest , i.e.
input energy augments. Second, they measure the similarity among input locations without using the information that
ut−τ is expected to have less influence on the prediction of
yt as the positive lag τ augments. Such limitation is also
in some sense hidden by the finite-dimensional context. In
fact, if the input locations are now defined by (12), i.e. the
system memory is infinite, the Gaussian kernel becomes
kx − ak2
.
K (x, a) = exp −
η
(33)
|ytest − ŷtest |
,
100% 1 − test
|y − ȳtest |
This model is not reasonable: it is not continuous around
the origin of R∞ and, out of the diagonal, is null for many
input locations. This reveals the importance of finding an
appropriate metric to measure the distance between different
input trajectories.
where ȳtest is the mean of the components of ytest while ŷtest
is the prediction from an estimated model. We will display
MATLAB boxplots of the 100 fits achieved by (1) after a
Monte Carlo of 100 runs, using different kernels. At any run,
independent noises are drawn to form new identification and
test data. The impulse response h in (S2) also varies. It is a
10-th order rational transfer function with ℓ2 norm equal to
10 (this makes similar the contribution to the output variance
of the linear and nonlinear system components) and poles
inside the complex circle of radius 0.95, randomly generated as detailed in [48][section 7.4].
First, we use the Gaussian kernel (31) over an m-dimensional
input space. Plugged in (1), it defines the hyperparameter
vector [m η γ ]. For tuning the Gaussian kernel hyperparameters, the regressor vector dimension m is chosen by an oracle not implementable in practice. Specifically, at any run,
for each m ∈ {1, . . . , 50} the pair (η , γ ) is determined via
marginal likelihood optimization [20] using only the identification data. Multiple starting points have been adopted
to mitigate the effect of local minima. The oracle has then
access to the test set to select, among the 50 couples, that
maximizing the prediction fit (35). The two left boxplots in
Fig. 1 report the prediction fits achieved by this procedure
applied to identify the first (top) and the second (bottom)
system. Even if the Gaussian kernel is equipped with the oracle, its performance is satisfactory only in the (S1) scenario
while the capability of predicting outputs in the (S2) case
is poor during many runs. In place of the marginal likelihood, a cross-validation strategy has been also used for tuning (η , γ ), obtaining results similar to those here displayed.
During the Monte Carlo, even when the number of impulse
response coefficients θk different from zero is quite large,
we have noticed that a relatively small value for m is frequently chosen, i.e. the oracle tends to use few past input
values (ut−1 , ut−2 , . . .) to predict yt . This indicates that the
Gaussian kernel structure induces the oracle to introduce a
significant bias to guard the estimator’s variance.
Now, we show that in this example model complexity can
be better controlled avoiding the difficult and computationally expensive choice of discrete orders. In particular, we
set m = ∞ and use the new NSS kernel (34). For tuning the
NSS hyperparameter vector (α , η , γ ), no oracle having access to the test set is employed but just a single continuous
optimization of the marginal likelihood that uses only the
identification data. The fits achieved by NSS after the two
Monte Carlo studies are in the right boxplots of the two fig-
New kernel for nonlinear system identification We now
show how stable spline kernels, which embed exponential
stability of linear systems [52], can be useful also to define nonlinear models. Specifically, let Kα be a stable spline
kernel, e.g. diagonal with Kα (i, i) = α max(i, j) or given by
Kα (i, j) = α max(i, j) for any integer i and j. Then, define the
nonlinear stable spline (NSS) kernel as
(a − x)T Kα (a − x)
K (a, x) = a Kα x × exp −
η
T
NSS,
(34)
which corresponds to the product between a linear kernel
and a modified version of (31). Such kernel defines a new
infinite-dimensional RKHS suited for identification of nonlinear output error models. Being no more constant along
the diagonal, it embeds the information that output energy
augments if input energy increases, preserving BIBO stability (as one can easily deduce from Propositions 11 and 12).
Note also that, letting the dimensionality m of the regression
space go to infinity, the difficult selection of the discrete dimension of the regressors x has been eliminated. In fact, input locations similarity is regulated by the hyperparameter
α that includes the information that the influence of ut−τ on
yt goes to zero as the time lag τ increases.
4.3 Numerical experiment
The following nonlinear system is taken from [67]:
2
f (xt ) = ut + 0.6ut−1 + 0.35 (ut.2 + ut−4 ) − 0.25ut−3
3
+ 0.2(ut−5 + ut−6 ) + 0.9ut−3 + 0.25ut ut−1 + 0.75ut−2
− ut−1 ut−2 + 0.5 ut2 + ut ut−2 + ut−1ut−3
Then, consider the identification of the following two systems, called (S1) and (S2):
∞
yt = f (xt )+ et (S1),
yt =
∑ θk ut−k + f (xt )+ et
(35)
(S2),
k=1
9
Fits - S1
The following result is well known and characterizes the
minimizer of (36) which goes under the name of regression
function in machine learning.
95
90
Theorem 13 (The regression function) We have
fρ = arg min E (y − f (x))2 ,
85
f
80
where fρ is the regression function defined for any x ∈ X by
Gaussian+Oracle
NSS+MargLik
100
80
R
y d µy|x (y|x).
(37)
In our system identification context, fρ is the dynamic system associated to the optimal predictor that minimizes the
expected quadratic loss on a new output drawn from µyx .
60
5.2 Consistency of regularization networks for system
identification
40
20
Z
fρ (x) =
Fits - S2
Gaussian+Oracle
Consider a scenario where µy|x (and possibly also µx ) is unknown and only N samples {xi , yi }Ni=1 from µyx are available.
We study the convergence of the RN in (1) to the optimal
predictor fρ as N → ∞ under the input-induced norm
NSS+MargLik
Fig. 1. Boxplots of the 100 test set fits achieved in the two scenaria
by (1) equipped with the Gaussian kernel (31) relying on the oracle
to select the regressors space dimension (left boxplots) and by
the new kernel NSS (34) with all the hyperparameters tuned via
marginal likelihood optimization (right boxplots).
k f k2x =
ures above: in both the cases the new estimator behaves very
nicely.
Z
X
f 2 (x)d µx (x).
µ
This is the norm in the classical Lebesgue space L2 x already
introduced at the end of Section 2.
We assume that the reproducing kernel of H admits the
expansion
5 Consistency of regularization networks for system
identification
∞
K (x, a) = ∑ ζi ρi (x)ρi (a), ζi > 0 ∀i,
5.1 The regression function
i=1
In what follows, the system input u is a stationary stochastic process over Z in discrete-time or R in continuous-time.
The distribution of u induces on X the (Borel non degenerate) probability measure µx , from which the input locations xi are drawn. In view of their dependence on u, the
xi are in general correlated each other and unbounded, e.g.
for Gaussian u no bounded set contains xi with probability
one. Such peculiarities, inherited by the system identification setting, already violate the data generation assumptions
routinely adopted in machine learning.
The identification data {xi , yi }∞
i=1 are assumed to be a stationary stochastic process. In particular, each couple (x, y) has
joint probability measure µyx (y, x) = µy|x (y|x)µx (x) where
µy|x is the probability measure of the output y conditional
on a particular input location x.
Given a function (dynamic system) f , the least squares error
associated to f is
E (y − f (x))2 =
Z
X ×R
(y − f (x))2 d µyx (y, x).
with ζi and ρi defined via (10) and the probability measure
µx .
Exploiting the regression function, the measurements process can be written as
yi = fρ (xi ) + ei,
(38)
where the errors ei are zero-mean and identically distributed.
They can be correlated each other and also with the input
locations xi . Given any f ∈ H and the ℓ-th kernel eigenfunction ρℓ , we define the random variables {vℓi }i∈Z by
combining f , ρℓ and the errors ei defined by (38) as follows
vℓi = ( f (xi ) + ei ) ρℓ (xi ).
(39)
Let H be stable so that the {vℓi } form a stationary process.
In particular, note that each vℓi is the product of the outputs
from two stable systems: the first one, f (xi )+ ei , is corrupted
by noise while the second one, ρℓ (xi ), is noiseless.
(36)
10
Now, by summing up over ℓ the cross covariances of lag k
using kernel eigenvalues ζℓ as weights one obtains 4
∞
ck :=
∑
ζℓCov vℓi , vℓ,i+k
ℓ=1
The relevance and usefulness of (42) further emerges if more
specific experimental conditions are considered. An example is given below by specializing Proposition 14 to the
continuous-time linear setting. Here, the aim is to reconstruct the continuous-time impulse response of a linear system fed with a stationary input process from a sampled and
noisy version of the output. Below, one can e.g. think of
the sampling instants as ti = i∆ + δi where δi are identically
distributed random variables with support on [0, ∆]. Then, it
is shown that, for Gaussian u, a weak condition on the input covariance’s decay rate already guarantees consistency.
This outcome can also be seen as a non trivial extension to
the dynamic context of studies on functional linear regression like e.g. that illustrated in [81] under independent data
assumptions.
(40)
where Cov(·, ·) is the covariance operator. The next proposition shows that summability of the ck is key for consistency.
Proposition 14 (RN consistency in system identification)
Let H be the RKHS with kernel
∞
K (x, a) = ∑ ζi ρi (x)ρi (a), ζi > 0 ∀i,
(41)
i=1
where (ζi , ρi ) are the eigenvalues/eigenfunctions pairs defined by (10) under the probability measure µx . Assume that,
for any r > 0 and f s.t. k f kH ≤ r, there exists a constant
Cr such that
∞
∑
k=0
|ck | < Cr < ∞,
Proposition 15 (RN consistency in continuous-time linear system identification) Let H be the RKHS with kernel K defined by the continuous-time stable spline kernel
K(t, s) = e−β max(t,s) with support restricted to any compact
set of R+ × R+. Assume that
(42)
• the regression function fρ is a continuous-time and timeinvariant
linear system with impulse response θ satisfying
R 2
θ̇ (t)dt < ∞;
• the system input u is a stationary Gaussian process with
Cov(u(t + τ ), u(t)) decaying to zero as 1/τ 1+δ for δ > 0;
• the errors ei in (38) are white, independent of the system
input.
with ck defined in (40). Let also
γ∝
1
,
Nα
(43)
where α is any scalar in (0, 21 ). Then, if fρ ∈ H and ĝN is
the RN in (1), as N → ∞ one has
kĝN − fρ kx −→ p 0,
Then, letting γ satisfy (43), the RN in (1) is consistent, i.e.
(44)
kĝN − fρ kx −→ p 0.
where −→ p denotes convergence in probability.
The fact that ∑∞
ℓ=1 |ζℓ | < ∞ is already an indication that
(42) is not so hard to be satisfied. Indeed, to the best of our
knowledge, (42) is the weakest RKHS condition currently
available which guarantees RN consistency. In fact, previous works, like [83,69], beyond considering only noises
with densities of compact support use mixing conditions
(which rule out infinite memory systems) with fast mixing
coefficients decay. These assumptions largely imply (42) as
it can e.g. be deduced by Lemma 2.2 in [21].
(45)
Remark 16 (Convergence to the true impulse response)
The optimal predictor fρ is a dynamic system, unique as
map X → R. However, considering e.g. the linear scenario
in Proposition 15, such functional can be associated to
different impulse responses (as illustrated in Section 3.3 by
discussing the relationship between the space H of linear
dynamic systems and the space I of impulse responses).
Convergence to θ can be guaranteed under persistently
exciting conditions related to u and the sampling instants ti
[40]. One can then wonder which kind of impulse response
estimate is obtained if such conditions are not satisfied. The
answer is obtained by [53][Theorem 3 on p. 371] which
reveals that the impulse response estimate (24) associated
to ĝN coincides with (15). Thus, among all the possible
impulse responses defining the optimal predictor fρ , the estimator (24) will asymptotically privilege that of minimum
norm in I .
From (40) one can see that (42) essentially reduces
to studying summability of the covariances of the
vℓi = ( f (xi ) + ei ) ρℓ (xi ). In particular, the covariance of such
product sequences depends on the first four moments of the
component sequences, see eq. 3.1 in [79]. The stability of
H is thus crucial to ensure the existence of the moments
of the vℓi . This is e.g. connected with the use of kernels like
(19,22,34) where the influence of past input locations on
the output decays exponentially to zero as time progresses.
6 Conclusions
We have introduced a new look at system identification in
a RKHS framework. Our approach uses RKHSs whose elements are functionals associated to dynamic systems. This
4
As shown in Appendix, the ck in (40) are invariant w.r.t. the
particular spectral decomposition used to obtain the pairs (ζi , ρi ).
11
Lemma 18 Let r = 2k fρ kH . Then, for any γ > 0 one has
perspective establishes a solid link between system identification and machine learning, with focus on the problem of
learning from examples.
Such framework has led to simple derivations of RKHSs
stability conditions in both linear and nonlinear scenarios.
It has been also shown that stable spline kernels can be
used as basic building blocks to define other models for
nonlinear system identification.
In the last part of the paper, RN convergence to the optimal
predictor has been proved under assumptions tailored to
system identification, also pointing out the link between
consistency and RKHS stability. This general treatment will
hopefully pave the way for an even more fruitful interplay
between RKHS theory and regularized system identification.
1
E kĝN − fˆkx ≤
γ
Then, we have
Sx∗ Sx
+ γ I ĝN
N
Sx∗
=
K (K + N γ IN )−1 + N γ (K + N γ IN )−1 Y
N
Sx∗
= Y.
N
Hence, one has
We start reporting three useful lemmas instrumental to the
main proof. First, define
and
ĝN − fˆ =
(46)
=
ηi (·) = yi − fˆ(xi ) K (xi , ·).
(47)
→ RN
N
∀ c ∈ RN .
i=1
Sx∗ Sx
+ γI
N
Sx∗ Sx
+ γI
N
−1
−1
Sx∗Y Sx∗ Sx fˆ
ˆ
−
−γ f
N
N
1 N
∑ (ηi − E [ηi ]) ,
N i=1
where we used the equality Sx∗Y − Sx∗Sx fˆ = ∑Ni=1 ηi and (49).
Using (50) in Lemma 17, we then obtain
In addition, the notation Sx : H
is the sampling operator defined by Sx f = [ f (x1 ) . . . f (xn )] while Sx∗ is its adjoint
given by
Sx∗ c = ∑ ci Kxi
(51)
ĝN = Sx∗ (K + N γ IN )−1 Y.
Proof of Proposition 14
f ∈H
r
2Cr
.
max 1, max ζi
i
N
Proof: It comes from the representer theorem and (48) that
Appendix
fˆ = arg min k f − fρ k2x + γ k f k2H
s
1 1 N
kĝN − fˆkH ≤
∑ (ηi − E [ηi ])
γ N i=1
(48)
.
(52)
H
The first lemma below involves the definitions of ηi , fˆ, Sx
and Sx∗ given above. It is derived from [66] and Proposition
1 in [65].
Now, let f := fρ − fˆ. With the ei defined in (38), we can
write
Lemma 17 It holds that
ηi (·) − E ηi (·)
= [ f (xi ) + ei ] K (xi , ·) − E [ f (xi ) + ei ] K (xi , ·)
E ηi = γ fˆ.
∞
(49)
=
ℓ=1
∞
Furthermore, if v ∈ H and f satisfies
Sx∗ Sx
+ γI
N
=
∑ ζℓ [vℓi − mℓ] ρℓ(·)
ℓ=1
f = v,
where vℓi = ( f (xi ) + ei ) ρℓ (xi ) and mℓ = E vℓi . Now, the structure of the RKHS norm outlined in (9) allows us to write
where I denotes the identity operator, one has
1
k f kH ≤ kvkH .
γ
∑ ζℓ [( f (xi ) + ei)ρℓ (xi ) − E ( f (xi ) + ei)ρℓ (xi )] ρℓ (·)
hηi (·) − E ηi , η j (·) − E η j iH
(50)
∞
∞
= h ∑ ζℓ (vℓi − mℓ ) ρℓ (·), ∑ ζℓ vℓ j − mℓ ρℓ (·)iH
ℓ=1
∞
The second lemma states a bound between the expected
RKHS distance between ĝN and fˆ.
=
∑ ζℓ (vℓi − mℓ)
ℓ=1
12
ℓ=1
v ℓ j − mℓ .
γ
r
fˆ − fρ = − ∑∞
i=1 ζi +γ ζi di ρi and the same manipulations contained in the proof of Theorem 4 in [65][p. 295] lead to the
following result.
So, using definition (40)
E hηi − E ηi , η j − E η j iH
∞
= ∑ ζℓCov vℓi , vℓ, j
Lemma 19 For any 0 < r ≤ 1/2, one has
ℓ=1
= c|i− j| .
k fˆ − fρ kx ≤ γ r kL−r
K f ρ kx
Comparing the values of the objective in (46) at the optimum
fˆ and at fρ , one finds k fˆkH ≤ k fρ kH so that
Combining (51) and (53), for any 0 < r ≤ 1/2 it holds that
k fρ − fˆkH = k f kH ≤ 2k fρ kH .
∞
k=0
|ck | < Cr < ∞,
Hence, we obtain
1 N
E
∑ (ηi − E [ηi ])
N i=1
N
N
2
H
r
r = 2k fρ kH .
1
kL−r
K f ρ kx + γ
Proof of Proposition 15
Let I be the space of impulse responses with compact
support e.g. on [0, T ] induced by the stable spline kernel
R
K. Then, it comes from [49] that kθ k2I ∝ 0T θ̇ 2 (t)eβ t dt.
So, finite energy of the first derivative of θ ensures that the
optimal predictor fρ belongs the RKHS H defined by the
linear kernel K induced by K.
Now, we have just to prove that condition (42) holds. Recall
that the ck are invariant w.r.t. the particular kernel expansion
of K adopted. For the stable spline kernel K we choose
the expansion K(t, τ ) = ∑∞
ℓ=1 ζℓ ψℓ (t)ψℓ (τ ) derived in [49]
where
2Cr
1
.
≤ 2 ∑ ∑ |c|i− j| | ≤
N i=1 j=1
N
∞
2
Now, recall that, if f = ∑∞
i=1 ai ρi , then k f kH = ∑i=1
2
while k f k2x = ∑∞
i=1 ai . Thus, for any f ∈ H , one has
a2i
ζi
k f k2x ≤ max 1, max ζi k f k2H .
i
The use of Jensen’s inequality and (52) then completes the
proof.
√
ψℓ (t) = 2 sin
LK [ρi ] :=
X
e− β s
p
ζℓ
!
,
ζℓ =
1
.
(ℓπ − π /2)2
The eigenfunctions thus satify
Now, we need to set up some additional notation. Following
[65,66], given the integral operator
Z
s
r
2Cr
E kĝN − fρ kx ≤ γ
max 1, max ζi
.
i
N
(54)
Hence, when γ is chosen according to (43), E kĝN − fρ kx
converges to zero as N grows to ∞. Using the Markov inequality, (44) is finally obtained.
This, combined with (42), implies that for any γ > 0
∑
(53)
|ψℓ (t)| <
K (·, u)ρi (u)d µx (u) = ζi ρi ,
√
2 ∀(ℓ,t).
(55)
Now, let f be any dynamic system satisfying k f kH ≤ r and
let θ be the associated impulse response of minimum norm
living in I . 5 From the arguments discussed in section 3.3
one then has kθ kI ≤ r. It then holds that
for r > 0 we define
∞
ci
r ρi .
ζ
i=1 i
L−r
K [f] = ∑
|θ (t)| = |hθ , Kt iI | ≤ r
p
K(t,t) =⇒ max |θ (t)| ≤ Ar < ∞
t∈[0,T ]
(56)
with Ar independent of the particular f chosen inside the
ball of radius r of H .
The third lemma reported below contains the inequality (53)
which was also derived in [65] assuming a compact input space. However, the bound holds just assuming the validity of the kernel expansion (41). To see this, recalling
µx
Theorem 5, first note that L−r
K f ρ ∈ L2 for any 0 ≤ r ≤
µx
1/2. So, there exists g ∈ L2 , say g = ∑∞
i=1 di ρi , such that
r d ρ . After simple computations, one obtains
fρ = ∑∞
ζ
i
i
i=1 i
5
The minimum norm impulse response is chosen without loss
of generality since any other θ associated with f would induce
the same input-output relationship.
13
Without loss of generality, the input u is now assumed zeromean so that the f (xi ) and ρℓ (xi ) become zero-mean Gaussian processes. Using eq. 3.2 in [79], for k > 1 one obtains
[15] O. Bousquet and A. Elisseeff. Stability and generalization. J. Mach.
Learn. Res., 2:499–526, 2002.
[16] C. Carmeli, E. De Vito, and A. Toigo. Vector valued reproducing
kernel Hilbert spaces of integrable functions and Mercer theorem.
Analysis and Applications, 4:377–408, 2006.
Cov vℓi , vℓ,i+k = Cov ( f (xi )ρℓ (xi ), f (xi+k )ρℓ (xi+k ))
= Cov ( f (xi ), f (xi+k ))Cov (ρℓ (xi ), ρℓ (xi+k ))
+ Cov ( f (xi ), ρℓ (xi+k ))Cov ( f (xi ), ρℓ (xi−k )) .
(57)
[17] V. Chandrasekaran, B. Recht, P.A. Parrilo, and A.S. Willsky.
The convex geometry of linear inverse problems. Foundations of
Computational Mathematics, 12(6):805–849, 2012.
[18] T. Chen, H. Ohlsson, and L. Ljung. On the estimation of
transfer functions, regularizations and Gaussian processes - revisited.
Automatica, 48(8):1525–1535, 2012.
Now, let h be any of the four covariances in the r.h.s. of
(57). Combining (55,56) and classical integral formulas for
covariances computations, as e.g. reported in [47][p. 308313], it is easy to obtain a constant Br independent of ℓ such
that |h(k)| ≤ Br /k1+ε . Condition (42) thus holds true and
this completes the proof.
[19] F. Cucker and S. Smale. On the mathematical foundations of learning.
Bulletin of the American mathematical society, 39:1–49, 2001.
[20] G. De Nicolao, G. Sparacino, and C. Cobelli. Nonparametric input
estimation in physiological systems: problems, methods and case
studies. Automatica, 33:851–870, 1997.
[21] H. Dehling and W. Philipp. Almost sure invariance principles for
weakly dependent vector-valued random variables. The Annals of
Probability, 10(3):689–701, 1982.
References
[22] F. Dinuzzo. Kernels for linear time invariant system identification.
SIAM Journal on Control and Optimization, 53(5):3299–3317, 2015.
[1] H. Akaike. Smoothness priors and the distributed lag estimator.
Technical report, Department of Statistics, Stanford University, 1979.
[23] H. Drucker, C.J.C. Burges, L. Kaufman, A. Smola, and V. Vapnik.
Support vector regression machines. In Advances in Neural
Information Processing Systems, 1997.
[2] N. Alon, S. Ben-David, N. Cesa-Bianchi, and D. Haussler. Scalesensitive dimensions, uniform convergence, and learnability. J. ACM,
44(4):615–631, 1997.
[24] T. Evgeniou and M. Pontil. On the Vγ dimension for regression in
reproducing kernel Hilbert spaces. In Algorithmic Learning Theory,
10th International Conference, ALT ’99, Tokyo, Japan, December
1999, Proceedings, volume 1720 of Lecture Notes in Artificial
Intelligence, pages 106–117. Springer, 1999.
[3] A. Aravkin, J.V. Burke, A. Chiuso, and G. Pillonetto. On
the estimation of hyperparameters for empirical bayes estimators:
Maximum marginal likelihood vs minimum mse. IFAC Proceedings
Volumes, 45(16):125 – 130, 2012.
[25] T. Evgeniou, M. Pontil, and T. Poggio. Regularization networks and
support vector machines. Advances in Computational Mathematics,
13:1–50, 2000.
[4] A. Aravkin, J.V. Burke, A. Chiuso, and G. Pillonetto. Convex vs
non-convex estimators for regression and sparse estimation: the mean
squared error properties of ard and glasso. Journal of Machine
Learning Research, 15:217–252, 2014.
[26] M.O. Franz and B. Schölkopf. A unifying view of Wiener and volterra
theory and polynomial kernel regression. Neural Computation,
18:3097–3118, 2006.
[5] A.Y. Aravkin, B.M. Bell, J.V. Burke, and G. Pillonetto. The
connection between Bayesian estimation of a Gaussian random
field and RKHS. Neural Networks and Learning Systems, IEEE
Transactions on, 26(7):1518–1524, 2015.
[27] R. Frigola, F. Lindsten, T.B. Schon, and C.E. Rasmussen. Bayesian
inference and learning in Gaussian process state-space models with
particle mcmc. In Advances in Neural Information Processing
Systems (NIPS), 2013.
[6] A. Argyriou and F. Dinuzzo. A unifying view of representer theorems.
In Proceedings of the 31th International Conference on Machine
Learning, volume 32, pages 748–756, 2014.
[28] R. Frigola and C.E. Rasmussen. Integrated pre-processing for
Bayesian nonlinear system identification with Gaussian processes.
In Proceedings of the 52nd Annual Conference on Decision and
Control (CDC), 2013.
[7] A. Argyriou, C.A. Micchelli, and M. Pontil. When is there a
representer theorem? vector versus matrix regularizers. J. Mach.
Learn. Res., 10:2507–2529, 2009.
[29] F. Girosi. An equivalence between sparse approximation and support
vector machines. Technical report, Cambridge, MA, USA, 1997.
[8] N. Aronszajn. Theory of reproducing kernels. Trans. of the American
Mathematical Society, 68:337–404, 1950.
[30] G.C. Goodwin, M. Gevers, and B. Ninness. Quantifying the error in
estimated transfer functions with application to model order selection.
IEEE Trans. on Automatic Control, 37(7):913–928, 1992.
[9] E. Bai and Y. Liu. Recursive direct weight optimization in nonlinear
system identification: a minimal probability approach. IEEE Trans.
Automat. Contr., 52(7):1218–1231, 2007.
[10] E. W. Bai.
Non-parametric nonlinear system identification:
An asymptotic minimum mean squared error estimator. IEEE
Transactions on Automatic Control, 55(7):1615–1626, 2010.
[31] C. Grossmann, C.N. Jones, and M. Morari. System identification via
nuclear norm regularization for simulated moving bed processes from
incomplete data sets. In Proceedings of the 48th IEEE Conference
on Decision and Control (CDC), pages 4692–4697, 2009.
[11] B.M. Bell and G. Pillonetto. Estimating parameters and stochastic
functions of one variable using nonlinear measurement models.
Inverse Problems, 20(3):627, 2004.
[32] Z.C. Guo and D.X. Zhou. Concentration estimates for learning with
unbounded sampling. Adv. Comput. Math., 38(1):207–223, 2013.
[12] S. Bergman. The Kernel Function and Conformal Mapping.
Mathematical Surveys and Monographs, AMS, 1950.
[34] G. Kimeldorf and G. Wahba. A correspondence between bayesian
estimation on stochastic processes and smoothing by splines. The
Annals of Mathematical Statistics, 41(2):495–502, 1970.
[33] H. Hochstadt. Integral equations. John Wiley and Sons, 1973.
[13] M. Bertero. Linear inverse and ill-posed problems. Advances in
Electronics and Electron Physics, 75:1–120, 1989.
[35] G. Kimeldorf and G. Wahba. A correspondence between Bayesan
estimation of stochastic processes and smoothing by splines. Ann.
Math. Statist., 41(2):495–502, 1971.
[14] M. Bertero, T. Poggio, and V. Torre. Ill-posed problems in early
vision. In Proceedings of the IEEE, pages 869–889, 1988.
14
[58] C.R. Rojas, R. Toth, and H. Hjalmarsson. Sparse estimation of
polynomial and rational dynamical models. IEEE Transactions on
Automatic Control, 59(11):2962–2977, 2014.
[36] G. Kitagawa and W. Gersch. Smoothness priors analysis of time
series. Springer, 1996.
[37] W. E. Leithead, E. Solak, and D. J. Leith. Direct identification
of nonlinear structure using Gaussian process prior models. In
Proceedings of European Control Conference (ECC 2003), 2003.
[59] J. Roll, A. Nazin, and L. Ljung. Nonlinear system identification via
direct weight optimization. Automatica, 41(3):475–490, 2005.
[38] T. Lin, B.G. Horne, P. Tino, and C.L. Giles. Learning long-term
dependencies in NARX recurrent neural networks. IEEE Trans. on
Neural Networks, 7:1329 – 1338, 1996.
[60] R.J. Schiller. A distributed lag estimator derived from smoothness
priors. Economics Letters, 2(3):219 – 223, 1979.
[61] B. Schölkopf, R. Herbrich, and A. J. Smola. A generalized representer
theorem. Neural Networks and Computational Learning Theory,
81:416–426, 2001.
[39] Z. Liu and L. Vandenberghe. Interior-point method for nuclear
norm approximation with application to system identification. SIAM
Journal on Matrix Analysis and Applications, 31(3):1235–1256,
2009.
[62] B. Schölkopf and A. J. Smola. Learning with Kernels: Support Vector
Machines, Regularization, Optimization, and Beyond. (Adaptive
Computation and Machine Learning). MIT Press, 2001.
[40] L. Ljung. System Identification - Theory for the User. Prentice-Hall,
Upper Saddle River, N.J., 2nd edition, 1999.
[63] S. Shun-Feng and F.Y.P. Yang. On the dynamical modeling with
neural fuzzy networks. IEEE Transactions on Neural Networks,
13:1548 – 1553, 2002.
[41] L. Ljung, G.C. Goodwin, and J. C. Agero. Stochastic embedding
revisited: A modern interpretation. In 53rd IEEE Conference on
Decision and Control, pages 3340–3345, 2014.
[42] M.N. Lukic and J.H. Beder. Stochastic processes with sample paths
in reproducing kernel Hilbert spaces. Trans. Amer. Math. Soc.,
353:3945–3969, 2001.
[64] J. Sjöberg, Q. Zhang, L. Ljung, B. Delyon A. Benveniste,
P. Glorennec, H. Hjalmarsson, and A. Juditsky. Nonlinear black-box
modeling in system identification: A unified overview. Automatica,
31(12):1691–1724, December 1995.
[43] J. S. Maritz and T. Lwin. Empirical Bayes Method. Chapman and
Hall, 1989.
[65] S. Smale and D.X. Zhou. Shannon sampling II: connections to
learning theory. Appl. Comput. Harmon. Anal., 19:285–302, 2005.
[44] J. Mercer. Functions of positive and negative type and their
connection with the theory of integral equations. Philos. Trans. Roy.
Soc. London, 209(3):415–446, 1909.
[66] S. Smale and D.X. Zhou. Learning theory estimates via integral
operators and their approximations. Constructive Approximation,
26:153–172, 2007.
[45] K. Mohan and M. Fazel. Reweighted nuclear norm minimization with
application to system identification. In American Control Conference
(ACC), pages 2953–2959, 2010.
[67] W. Spinelli, L. Piroddi, and M. Lovera. On the role of prefiltering
in nonlinear system identification. IEEE Transactions on Automatic
Control, 50(10):1597–1602, 2005.
[46] S. Mukherjee, P. Niyogi, T. Poggio, and R. Rifkin. Learning
theory: stability is sufficient for generalization and necessary and
sufficient for consistency of empirical risk minimization. Advances
in Computational Mathematics, 25(1):161–193, 2006.
[68] I. Steinwart, D. Hush, and C. Scovel. Learning from dependent
observations. Journal of Multivariate Analysis, 100(1):175 – 194,
2009.
[47] A. Papoulis. Probability, Random Variables and Stochastic Processes.
Mc Graw-Hill, 1991.
[69] H. Sun and Q. Wu. Regularized least square regression with
dependent samples.
Advances in computational mathematics,
32(2):175 – 189, 2010.
[48] G. Pillonetto, T. Chen, A. Chiuso, G. De Nicolao, and L. Ljung.
Regularized linear system identification using atomic, nuclear and
kernel-based norms: The role of the stability constraint. Automatica,
69:137 – 149, 2016.
[70] Hongwei Sun. Mercer theorem for RKHS on noncompact sets. J.
Complexity, 21(3):337–349, 2005.
[71] J.A.K. Suykens, C. Alzate, and K. Pelckmans. Primal and dual model
representations in kernel-based learning. Statist. Surv., 4:148–183,
2010.
[49] G. Pillonetto, A. Chiuso, and G. De Nicolao. Regularized estimation
of sums of exponentials in spaces generated by stable spline kernels.
In Proceedings of the IEEE American Cont. Conf., Baltimora, USA,
2010.
[72] J.A.K. Suykens, T. Van Gestel, J. De Brabanter, B. De Moor, and
J. Vandewalle. Least Squares Support Vector Machines. World
Scientific, Singapore, 2002.
[50] G. Pillonetto, A. Chiuso, and G. De Nicolao. Prediction error
identification of linear systems: a nonparametric Gaussian regression
approach. Automatica, 47(2):291–305, 2011.
[73] J.A.K. Suykens, T. Van Gestel, J. De Brabanter, B. De Moor, and
J. Vandewalle. Least Squares Support Vector Machines. World
Scientific, Singapore, 2002.
[51] G. Pillonetto, A. Chiuso, and Minh Ha Quang. A new kernelbased approach for nonlinear system identification. IEEE Trans. on
Automatic Control, 56(12):2825–2840, 2011.
[74] A.N. Tikhonov and V.Y. Arsenin. Solutions of Ill-Posed Problems.
Washington, D.C.: Winston/Wiley, 1977.
[52] G. Pillonetto and G. De Nicolao. A new kernel-based approach for
linear system identification. Automatica, 46(1):81–93, 2010.
[75] V. Vapnik. Statistical Learning Theory. Wiley, New York, NY, USA,
1998.
[53] G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and L. Ljung.
Kernel methods in system identification, machine learning and
function estimation: a survey. Automatica, 50(3):657–682, 2014.
[76] G. Wahba. Spline models for observational data. SIAM, Philadelphia,
1990.
[77] Grace Wahba. Practical approximate solutions to linear operator
equations when the data are noisy. SIAM Journal on Numerical
Analysis, 14(4):651–667, 1977.
[54] T. Poggio and F. Girosi. Networks for approximation and learning.
In Proceedings of the IEEE, volume 78, pages 1481–1497, 1990.
[55] T. Poggio, R. Rifkin, S. Mukherjee, and P. Niyogi. General conditions
for predictivity in learning theory. Nature, 428(6981):419–422, 2004.
[78] C. Wang and D.X. Zhou. Optimal learning rates for least
squares regularized regression with unbounded sampling. Journal of
Complexity, 27(1):55 – 67, 2011.
[56] Tomaso Poggio and Christian R. Shelton. Machine learning, machine
vision, and the brain. AI Magazine, 20(3):37–55, 1999.
[79] W.E. Wecker. A note on the time series which is the product of two
stationary time series. Stochastic Processes and their Applications,
8(2):153 – 157, 1978.
[57] C.E. Rasmussen and C.K.I. Williams. Gaussian Processes for
Machine Learning. The MIT Press, 2006.
15
[80] Q. Wu, Y Ying, and D.X. Zhou. Learning rates of least-square
regularized regression. Foundations of Computational Mathematics,
6:171–192, 2006.
[81] M. Yuan and T. Tony Cai. A reproducing kernel Hilbert space
approach to functional linear regression. Annals of Statistics,
38:3412–3444, 2010.
[82] W. Zhao, H.F. Chen, E. Bai, and K. Li. Kernel-based local order
estimation of nonlinear nonparametric systems. Automatica, 51:243–
254, 2015.
[83] B. Zou, L. Li, and Z. Xu. The generalization performance of erm
algorithm with strongly mixing observations. Machine Learning,
75:275–295, 2009.
16
| 3cs.SY
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:cs/0406012v1 [cs.PL] 7 Jun 2004
Secure Prolog Based Mobile Code
SENG WAI LOKE∗
DSTC, Monash University,
Caulfield, Victoria 3145, Australia
(e-mail: [email protected])
ANDREW DAVISON
Dept. of Computer Engineering
Prince of Songkla University,
Hat Yai, Songkhla 90112, Thailand
(e-mail: [email protected])
Abstract
LogicWeb mobile code consists of Prolog-like rules embedded in Web pages, thereby adding
logic programming behaviour to those pages. Since LogicWeb programs are downloaded
from foreign hosts and executed locally, there is a need to protect the client from buggy
or malicious code. A security model is crucial for making LogicWeb mobile code safe to
execute. This paper presents such a model, which supports programs of varying trust levels
by using different resource access policies. The implementation of the model derives from
an extended operational semantics for the LogicWeb language, which provides a precise
meaning of safety.
1 Introduction
A significant development of the World Wide Web, which has attracted considerable
interest in recent years, is the addition of mobile code to Web pages; Java applets and
JavaScript are popular examples. Code of this kind adds sophisticated interactive
behaviour to Web documents, allowing more useful tasks to be performed. However,
a key issue is security, since mobile code is downloaded from foreign hosts and
executed locally.
Malicious or buggy code can attack a local host in three main ways (Brown 1996;
Ousterhout et al. 1997):
• integrity attacks: attempts to delete or modify information in the local environment in unauthorised ways. For example, a program issues an operating
system command to delete local files.
• privacy attacks: attempts to steal or leak information to unauthorised parties.
For example, a program reads a file and transmits the contents to another
host.
∗ This author is grateful to Leon Sterling who (with Andrew Davison) co-supervised the work on
which this paper is based.
2
S.W. Loke and A. Davison
• denial of service attacks: attempts to occupy resources to an extent which
interferes with the normal operation of the host. An example is a program
executing an infinite loop, resulting in CPU cycles wasted and a depleted heap
space.
These attacks are only possible if a foreign program has unregulated access to
local system resources such as the file system, network services (e.g., socket connections), CPU cycles, internal memory, input/output devices, the program environment (e.g., environment variables), and operating system commands. At the other
extreme, a foreign program which is denied access to all resources cannot do any
damage, but is unlikely to be able to do much useful work.
There is no one applicable policy for all foreign code: some code should be more
trusted than others, and hence should be given more privileges. Providing a security
model that provides the right level of functionality with the appropriate degree
of security at the right time is a challenge. Indeed, security for mobile code is
an active research area (Vitek and Tschudin 1997; Brown 1996; Cugola et al. 1997;
Thorn 1997).
An emerging trend is the use of formal language semantics to reason with mobile code and prove security properties. For example, Leroy and Rouaix (1998) have
developed formal techniques for validating a typed functional language with assignments, ensuring that memory locations always contain appropriate values. The use
of a rewriting logic is advocated in (Meseguer and Talcott 1997) as the basis for a
mobile code programming language where security properties can be formally verified. Volpano (1997) argues for the development of provably-secure programming
languages which can be remotely evaluated, and whose semantics permit the formal
proof of security properties. As reported in (Dean et al. 1996), a number of security
loopholes have been discovered with Java applets, and this seems related to its lack
of a formal semantics. Recently, this weakness has been addressed by giving semantics to subsets of Java (Drossopoulou and Eisenbach 1998; Jensen et al. 1998).
Declarative languages come with simple and elegant semantics, providing a convenient basis for reasoning about program execution and for introducing security
mechanisms. This paper examines a security extension for a Prolog-based mobile
code language called the LogicWeb language (Loke and Davison 1998a). Underpinning the language is a view of the Web as a collection of logic programs called LogicWeb programs. LogicWeb clauses embedded within Web pages add logic programming behaviour to those pages. A LogicWeb program can retrieve and manipulate
other Web pages (which may optionally contain embedded LogicWeb code themselves) using compositional logic programming techniques. Applications of LogicWeb include Web-based distributed deductive databases, and extending the semantics of Web links (Loke et al. 1996b; Loke et al. 1996a; Loke and Davison 1997).
In this paper, we present a security model for LogicWeb which provides for the
safe (i.e., host-protecting) execution of programs without overly restricting their
capabilities. The key features of the model are:
• flexible: it supports varying degrees of trust and access to resources. More
trusted programs can be given more privileges to achieve greater function-
Secure Prolog Based Mobile Code
3
ality. Policies determine what resources are available to whom, and how the
resources must be used.
• formally specified: the model provides a precise meaning of safety.
• implemented based on the model’s specification: the implementation takes the
form of a meta-interpreter which is closely allied to the underlying operational
semantics.
The security model is specified as an extension of the operational semantics of
the LogicWeb language, and so permits a straightforward proof that the model protects the host according to specified policies. We describe how meta-programming
provides the necessary control over execution behaviour for a flexible yet safe execution environment, including control over resource usage and looping. We also
show how security policies can support different trust levels, and how policies can
be combined to control the execution of an application composed from programs
with different trust levels. The security model is implemented as an extension of
the system described in (Loke and Davison 1998b).
The rest of this paper is organised as follows. §2 provides background on LogicWeb, §3 gives an overview of the security model, §4 describes how we utilise digital
signatures, and §5 examines how security policies are specified. The need to combine security policies during goal evaluation is explained in §6. §7 describes how
policies are enforced by policy programs. An implementation of the security model
is given in §8. Control of resource usage using meta-interpreters is illustrated in §9.
§10 reviews related work, and §11 concludes.
2 Preliminaries
This section gives an overview of LogicWeb programs, the language, and the architecture of the client-side LogicWeb system.
2.1 LogicWeb Programs
LogicWeb programs are constructed from the data returned by HEAD, GET or
POST HTTP requests. When successful, a HEAD request returns meta-information
about a page, whereas GET and POST requests return meta-information and Web
pages.
The LogicWeb program derived from a HEAD response consists of two types of
facts for holding the page’s meta-information and URL:
• about(FieldName, Value). about/2 facts store the page’s meta-information
as supplied by the server (e.g., the content length).
• actual url(URL). This holds the URL of the page whose meta-information
was retrieved from the server. The actual URL would be different from the
URL used in the original HTTP request if the request was redirected by the
server.
The program corresponding to a HEAD response uses the identifier lw(head,
4
S.W. Loke and A. Davison
URL). Programs corresponding to GET and POST responses use lw(get, URL)
and lw(post(Data), URL) respectively. A GET or POST response is translated
into five types of facts based on the meta-information returned by the page’s server
and the HTML text in the page. Additional LogicWeb clauses may also be included
within the page.
The facts include about/2 and actual url/1 described above, and the following:
• my id(Type, URL). Type stores the type of the program, which is either the
term get or post(Data). Data is the information posted to the CGI script
at URL.
• h text(HTMLText). This contains the HTML text of the page (apart from
any additional LogicWeb clauses).
• link(Label, URL). This stores a page’s link information. For instance, the
anchor:
...<A HREF="http://www.cs.mu.oz.au/">Melb. U</A>...
becomes:
link("Melb. U", "http://www.cs.mu.oz.au/").
Additional facts can be readily extracted from the text of the page, as described
in (Loke and Davison 1998b).
Clauses are included within a page between the tags “<LW CODE>” and
“</LW CODE>”. Typically, the code appears inside a verbatim container
“<PRE>...</PRE>” or a comment container “<!--...-->” so that it is uninterpreted by the browser. Figure 1 shows a page holding LogicWeb clauses about
research interests.
interested in/1 defines the author’s interests using Prolog-style facts and rules.
The LogicWeb operator “#>” evaluates the interested in/1 goal against the program identified by lw(get, URL).
Assuming that the page shown in Figure 1 was retrieved from
http://www.cs.mu.oz.au/~swl/ using the GET method, then the resulting
LogicWeb program contains:
1.
2.
3.
4.
Several about/2 facts concerning the page’s meta-information.
actual url("http://www.cs.mu.oz.au/~swl/").
my id(get,"http://www.cs.mu.oz.au/~swl/").
h text("<HTML>...</HTML>"). This fact stores all the HTML text from the
page except the clauses between the “<LW CODE>” and “</LW CODE>” tags.
5. Two link/2 facts:
link("Department of Computer Science", "http://www.cs.mu.oz.au/").
link("University of Melbourne", "http://www.unimelb.edu.au/").
6. The clauses between the “<LW CODE>” and “</LW CODE>” tags.
A LogicWeb application comprises a set of LogicWeb programs, with one singled
out (by the programmer) as the main program. Each application has an identity
which is the program identifier of the main program, and a visible interface with
Secure Prolog Based Mobile Code
5
<HTML>
<HEAD>
<TITLE>Seng Wai Loke’s Home Page</TITLE>
</HEAD>
<BODY><H1>Seng Wai Loke’s Home Page</H1>
I’m from the<A HREF="http://www.cs.mu.oz.au/">
Department of Computer Science</A> at the
<A HREF="http://www.unimelb.edu.au/">
University of Melbourne</A>.
<!-<LW_CODE>
interests(["Logic Programming", "AI", "Web", "Agents"]).
friend_page("http://www.cs.mu.oz.au/~f1/").
friend_page("http://www.cs.mu.oz.au/~f2").
interested_in(X) :- interests(Is), member(X, Is).
interested_in(X) :friend_page(URL), lw(get, URL)#>interested_in(X).
</LW_CODE>
-->
</BODY>
</HTML>
Fig. 1. A page with clauses describing research interests.
which the user interacts. A typical interface consists of a HTML form and/or Web
links. Clauses in the main program define the forms interface and the mapping of
link selections to goals.
2.2 The LogicWeb Language
The LogicWeb language utilises Prolog and special LogicWeb operators. We will
assume that the reader is familiar with Prolog, and concentrate instead on the
operators.
2.2.1 LogicWeb Operators
A LogicWeb goal is formed using the operator “#>”, which is known as context
switching. For example, the LogicWeb goal lw(get, URL)#>Goal applies Goal to
the program specified by the lw(get, URL) identifier.
If the program is not already present on the client-side, then its page will be
downloaded and translated into a LogicWeb program before the query is evaluated. However, if the program is present, then the goal is executed immediately.
Thus, the “#>” operator permits the programmer to think of Web computation as
goals applied to programs, with no need for explicit Web page retrieval, parsing, or
storage.
The current context of a LogicWeb goal, i.e. the program (or composition of
programs) where the goal is located, is ignored when “#>” is evaluated.
6
S.W. Loke and A. Davison
LogicWeb programs can be manipulated using five LW-composition operators.
Four are based on Brogi’s algebraic program composition framework (Brogi 1993),
and the fifth on reduce from functional programming. The operators are: LW-union
(“+”), LW-intersection (“*”), LW-restriction (“/”), LW-encapsulation (“@”), and
LW-reduce (“<>”). An expression formed with LW-composition operators is called
a program expression.
LogicWeb programs are composed together after they have been retrieved from
the Web, which means that the semantics of the operators can be viewed as a variant
of Brogi’s definitions, extended to address issues related to page downloading and
the translation of pages to programs. The operational semantics are specified in
§2.2.3.
The sixth operator, the context operator, denoted by “(#)”, represents the current
context in a program expression. For instance:
?- lw(get, "URL0")#>(((#) + lw(get, "URL1"))#>interested_in(X)).
“(#)” is instantiated to lw(get, "URL0") when the goal is evaluated. “(#)” can
be used in place of a program identifier in any expression, which provides very
useful expressive power. For instance, it can readily model other composition formalisms (Loke and Davison 1998a).
2.2.2 EBNF Syntax
A pure LogicWeb program is a finite set of clauses of the form [∀x](A :- G) where
G is defined recursively as:
G ::= A | E #> G | (G, G)
E defines LogicWeb program expressions:
E ::= P | E+E | E*E | E/P | @E | (/)<>(E,L(P)) | (⊕)<>L(E)
P ::= lw(head, U RL) | lw(get, U RL) | lw(post(L(F ) ), U RL) | (#)
F ::= field(N ame,V alue)
L(I) ::= [] | [I|L(I) ]
⊕ ::= + | *
L(I) defines a Prolog list of items, each of which is described by a nonterminal
I. U RL is a URL, and F a query attribute submitted to a CGI script. N ame is
the name of a query attribute, and V alue is the value submitted to the server for
the corresponding attribute. The context operator “(#)” can appear anywhere a
program identifier can, but must be instantiated to a program identifier when used
as the right-hand argument of LW-restriction.
2.2.3 Operational Semantics
The operational semantics of the LogicWeb language is detailed in (Loke and Davison 1998a),
but so that this paper is self-contained, the semantics are summarised here. In particular, we do not consider LW-restriction and LW-reduce.
Secure Prolog Based Mobile Code
7
An oracle function is defined to model the accessing of LogicWeb programs.
Definition 2.1 (oracle function)
The oracle function
download : LW P rogramIDs → LW P rograms ∪ {⊥}
takes a LogicWeb program identifier P (of the form defined by P), and returns the
program DP if it is successfully created. Failure to obtain a program is represented
by returning the symbol ⊥.
DP if the program denoted by identifier P is
download(P ) =
successfully created,
⊥
otherwise.
download attempts to download a HTTP response object and translate it into a
LogicWeb program in the way specified in §2.1.
The set of downloaded programs is extended by calling the function
add programs, which takes the set of existing LogicWeb programs and maps it
to a new set using download.
Definition 2.2 (addition of LogicWeb programs)
Let ℘ denote powerset. The function
add programs : ℘(LW P rograms) × ℘(LW P rogramIDs) → ℘(LW P rograms)
takes a set S of programs and a set I of program identifiers and returns a new
set add programs(S, I) consisting of S augmented with newly created programs
mentioned in I but previously not in S:
add programs(S, I) = S ∪ {DP | P ∈ (I \ ids(S)), DP = download(P ), DP 6= ⊥}
where ids is a function that takes a set of programs and returns the identifiers of
the programs in the set, i.e. ids(S) is the set of identifiers of the programs in S.
A derivation relation involving the set of downloaded LogicWeb programs specifies the computation model involving interactions with the Web.
Definition 2.3 (derivation relation)
′
For any goal formula G and program expression E, we denote by S, E ⊢Sθ G the
fact that there exists a top-down derivation of G in E starting with the set S of
existing LogicWeb programs and ending with computed answer substitution θ and
created program set S ′ . A top-down derivation or proof of G in E starting with S
and ending with S ′ and θ is a finite tree such that:
′
1. the root node (bottom node) is labelled by the string “S, E ⊢Sθ G”;
2. the internal nodes are derived according to the set of inference rules given
below; and
3. all the leaves of the tree are either empty or labelled by a string not containing
the symbol “⊢” (e.g., the label “(A :- G) ∈ P ”).
8
S.W. Loke and A. Davison
⊢Sθ is defined to be the smallest relation satisfying the inference rules below. If
′
S, E ⊢Sθ G, then the goal G succeeded when evaluated in E using S. Otherwise,
the goal G f ailed when evaluated in E using S.
′
Given a top-down derivation S, E ⊢Sθ G, if the derivation does not involve interactions with the Web or if no programs are downloaded, then S ′ is the same as S
(i.e., S ′ = S). If programs are downloaded, then S is extended to S ′ (⊇ S). The
difference S ′ \ S represents the effect of Web interactions during the derivation.
A context is defined for each node in a top-down derivation whose label is of the
′
form “S, E ⊢Sθ G”.
Definition 2.4 (context of a goal)
′
Given the node label “S, E ⊢θS G”, E is the context (of G).
In the rules, P denotes single program identifiers of the form P, and E and F
denote program expressions of the form E as defined in §2.2.2. L(I) denotes a list
of the form LI . ǫ denotes the empty (identity) substitution.
The following rules define derivation in pure Prolog taking into account the creation of LogicWeb programs:
True.
(1)
S, E ⊢Sǫ true
true is always derivable in any program expression E without any change to the
set of created programs.
Conjunction.
′
′′
S, E ⊢Sθ G1 ∧ S ′ , E ⊢Sγ G2 θ
′′
S, E ⊢Sθγ G1 , G2
(2)
S ′ , which may or may not be the same as S, is the result of Web interactions from
G1 ’s derivation. The result of these interactions are propagated to G2 by starting
the derivation of G2 θ with S ′ . Since G2 θ started with S ′ , S ′′ is the result of Web
interactions during the derivation of the conjunction.
Atomic formula.
′
S, E ⊢Sθ (H :- G)
∧
γ = mgu(A, Hθ)
′′
S, E ⊢Sθγδ A
∧
′′
S ′ , E ⊢Sδ Gθγ
(3)
Obtaining clauses from E can involve the creation of new programs due to LW-encapsulation (see rule (8)), and so, S is changed to S ′ . The proof of the body starts
with the computed program set S ′ and returns the new set S ′′ and the answer
substitution δ.
Obtaining clauses from a single program.
(A :- G) ∈ P
S, P ⊢Sǫ (A :- G)
(4)
Secure Prolog Based Mobile Code
9
The answer substitution is ǫ, and there is no change to S.
Below, we present the rules for choosing clauses from LW-union, LW-intersection,
LW-encapsulation, and context switching.
LW-union.
′
S, E ⊢Sθ (A :- G)
′
S, E+F ⊢Sθ (A :- G)
(5)
′
S, F ⊢Sθ (A :- G)
′
S, E+F ⊢Sθ (A :- G)
(6)
A clause is chosen from a LW-union E+F by choosing a clause from either E or F .
LW-intersection.
′
S, E ⊢Sθ1 (H1 :- G1 )
′′
∧ S ′ , F ⊢Sθ2 (H2 :- G2 ) ∧ γ = mgu(H1 θ1 , H2 θ2 )
(7)
′′
S, E*F ⊢Sθ1 θ2 γ (H1 :- G1 , G2 )
A clause H:-G is obtained from the LW-intersection E*F if there exists a clause
H1 :-G1 in E and a clause H2 :-G2 in F such that H unifies with H1 and H2 ,
and G = (G1 , G2 ). This rule utilises a left-to-right ordering in choosing clauses
from E*F . Rules are first chosen from E returning S ′ , and then, S ′ is used when
selecting clauses from F ending up with S ′′ .
LW-encapsulation.
′
S, E ⊢Sθ A
′
S, @E ⊢Sθ A :- true
(8)
A clause A:- true is obtained from the encapsulation of E, @E, if A is provable
in E. LogicWeb programs can be created in the proof of A, i.e. a new program set
S ′ is computed starting with S.
Context Switching. We first define the function
expids : P rogramExpressions → ℘(LW P rogramIDs)
10
S.W. Loke and A. Davison
to refer to the program identifiers within a program expression. expids is defined
recursively based on the syntax of program expressions:
expids(P ) = {P }
expids(E1 + E2 ) = expids(E1 ) ∪ expids(E2 )
expids(E1 * E2 ) = expids(E1 ) ∪ expids(E2 )
expids(E / P ) = expids(E) ∪ expids(P )
expids(@E) = expids(E)
expids((/)<>(E,L(P))) = expids(E) ∪ expids(L(P) )
expids((⊕)<>L(E)) = expids(L(E) )
[
expids(L(E) ) =
expids(E)
E∈L(E)
In the above, we have used ∈ to represent list membership, and a L(P) is a L(E)
from the EBNF definition in §2.2.2.
We also define the function:
insertCC : P rogramExpressions×P rogramExpressions → P rogramExpressions
which substitutes every occurrence of the operator (#) in a program expression
(the first argument) with the current context (the second argument):
insertCC((#), C) = C
insertCC(P, C) = P
insertCC(E1 + E2 , C) = insertCC(E1 , C) + insertCC(E2 , C)
insertCC(E1 * E2 , C) = insertCC(E1 , C) * insertCC(E2 , C)
insertCC(E / P, C) = insertCC(E, C) / insertCC(P, C)
insertCC(@E, C) = @insertCC(E, C)
insertCC((/)<>(E,L(P)), C) = (/)<>(insertCC(E, C),insertCC(L(P), C))
insertCC((⊕)<>L(E), C) = (⊕)<>insertCC(L(E), C)
insertCC(L(E) , C) = [insertCC(E, C) | E ∈ L(E) ]
The rule defining context switching is the following:
′′
I ⊆ ids(S ′ ) ∧ S ′ , F ′ ⊢Sθ G
′′
S, E ⊢Sθ F #> G
(9)
where F ′ = insertCC(F, E), I = expids(F ), and S ′ = add programs(S, I).
The rule states that the goal F #> G is provable in E starting with the program
set S if the goal G is provable in F ′ starting with the updated program set S ′ which
contains all the programs mentioned in I (and hence, in F ).
2.3 The LogicWeb System Architecture
The architecture of the LogicWeb system (its components and data-flows) is shown
in Figure 2.
Secure Prolog Based Mobile Code
useractions
User
displayed
result
LogicWeb System
Graphical
User Interface
(Web browser)
query
11
result
Prolog Engine
Query
Processing
Engine
program
Program
Store
program
request
program
Page Loader
and Translator
Client-side
page request
page
Web Servers
Fig. 2. Architecture of the LogicWeb system.
The main components are:
•
•
•
•
a graphical user interface (GUI);
an engine for processing user queries;
a page loader and (page to program) translator; and
a store of LogicWeb programs.
The Prolog engine consists of three components: the query processing engine, the
program store, and the page loader and translator. The query processing engine
consists of the LogicWeb program interpreter (see §8.1) and predicates to translate
user actions into the evaluation of specific goals in LogicWeb programs. Downloaded
LogicWeb programs are stored inside the program store (actually as facts in a
SWI-Prolog database). The page loader and translator consist of predicates which
download and parse HTML documents to extract the clauses making up LogicWeb
programs.
The system converts a user action (i.e., link selection or form submission) into a
query. It computes the result of the query with respect to a program by invoking
the LogicWeb interpreter. Processing of the query may result in other pages being
downloaded and translated into LogicWeb programs. When query processing ends,
the system formats the result and shows it to the user via the browser.
The above architecture has been implemented using the NCSA Mosaic browser
and SWI-Prolog (for the Prolog engine). The two are integrated with Mosaic’s Com-
12
S.W. Loke and A. Davison
mon Client Interface (CCI) API (Loke and Davison 1998b). This implementation
is extended in §8.
3 Overview of the Security Model
The LogicWeb security model assumes the following about the downloading and
execution of programs:
• Program source is downloaded. Since the local host receives the source rather
than a compiled form (e.g., binaries or bytecode), each goal is visible at the
interpreter level. Also, if the source is unencrypted then it is open to tampering
– it may be intercepted during transmission and modified.
• Downloaded programs are executed by an interpreter based on the operational
semantics presented earlier. Such an interpreter is described in (Loke and Davison 1998b),
and is extended in §8.
• A LogicWeb program can invoke goals in other programs. Therefore, the
model must handle communication between programs possessing different levels of trust. A less trusted program should not be able to use the privileges
of a more trusted program.
We have adopted a sandbox model, where each program is executed in a controlled
space, limiting its access to resources and controlling the resources used. A sandbox
for a LogicWeb program consists of its interpreter and a security policy, or policy for
short, which controls the program’s utilisation of system resources. A less trusted
LogicWeb program is assigned a policy which restricts resource usage, while a more
trusted program is granted a more liberal policy. Security policies are defined by
each user of the LogicWeb system.
A LogicWeb program accesses the local environment and system resources via
system calls, which are SWI-Prolog and LogicWeb built-in predicates. LogicWeb
built-ins include predicates for displaying pages, constructing and parsing HTML
documents, fast string matching, checking if a particular program exists in the
program store, and deleting programs from the program store.
A policy specifies what system calls are permitted, and the ways they can be
used. This information is encoded as a set of predicates in a policy program created
by the LogicWeb system user. A policy program also states what programs can be
utilised via context switching. Policy programs may be stored at remote sites or
locally as long as they are protected from tampering. They can be integrated into
the system using the same downloading mechanism as the rest of the LogicWeb
system, including the use of composition to combine policies.
Assignment of policies to LogicWeb programs is owner-based. The choice of policy program assigned to a LogicWeb program is determined by how much the LogicWeb program is trusted which, in turn, depends on the owner of the program.
The owner is identified by a mechanism external to the LogicWeb language, by
authenticating the LogicWeb program using a PGP (Pretty Good Privacy) digital
signature (Garfinkel 1995).
Secure Prolog Based Mobile Code
13
Policy programs restrict the use of system calls, thereby preventing integrity and
privacy attacks. Denial-of-service attacks are addressed by controlling the use of
resources using policy programs (as shown in §9.1) and meta-interpreters (discussed
in §9.2).
4 Digital Signatures for LogicWeb Programs
Information can be encrypted in PGP using a secret private key, and decrypted
using a corresponding public key which is distributed publically. PGP keys are used
for authentication in the following way. Suppose A encrypts a message and sends
it to B. B wants to ensure that the encrypted message is really from A, and that
the message has not been tampered with. To do this, B attempts to decrypt the
message using A’s public key. If the decryption succeeds, it means that the message
was encrypted using A’s private key and that it has not been modified. Also, since
the key is only known to A, the message must have come from A. If the decryption
fails, B cannot conclude that A encrypted the message.
PGP keys can also be used to sign documents or programs. Figure 3 shows
the use of a digital signature when downloading a program from A to B, and
the process of assigning a policy program. A PGP digital signature is created by
first mapping the program to a string (actually, a single large number) using the
MD5 algorithm (Garfinkel 1995), which almost uniquely identifies the message.1
The MD5 string is then encrypted using the sender’s private key, and this encrypted
MD5 string becomes the digital signature. The receiver of a signed program (the
program and its signature) uses the sender’s public key to decrypt the signature,
producing the original MD5 string and the PGP ID. The receiver then checks that
the program corresponds to the MD5 string, i.e. the program has not been modified
in any way, and assigns a policy to the program by consulting a database maintained
in the system which maps PGP IDs to policy program IDs. Note that the system
uses three kinds of IDs: PGP IDs, policy program IDs and IDs of an application’s
LogicWeb programs, and that since policy programs are LogicWeb programs, policy
program IDs are LogicWeb program IDs. Unsigned programs, or programs where
authentication failed, should not be trusted, and are given restricted access to
system resources. The system stores the downloaded program, and records the
policy assignment for later use in query evaluation. Policy programs not already
present in the system are downloaded during query evaluation. §8 provides details
on policy assignment and use.
A digital signature proves who sent the program, and that the program was not
altered either by error or design. The signature also provides non-repudiation, which
means that the sender cannot easily disavow his signature on the program.
One reason for using an MD5 string is that it is faster to encrypt and decrypt than
the entire program. Also, since the program is not encrypted, the program source is
1
In theory, two different messages could map to the same MD5 string, although this has never
been known to occur (Garfinkel 1995).
A
B
program
to send
received
program
program
to send
received
program
S.W. Loke and A. Davison
transmission
MD5
algorithm
MD5
string
received
digital
signature
digital
signature
to send
decrypt with
A’s public key
MD5
string
encrypt with
A’s private key
used in
query
PGP ID
to
policy program query
ID database
result
PGP ID
for A
A’s policy
program
ID
used by
used by
Fig. 3. Downloading a program from A to B, and the subsequent assignment of a policy program.
14
check
correspondence
LogicWeb
program
interpreter
Secure Prolog Based Mobile Code
15
available even if authentication fails, and so can still be executed (though with limited privileges). However, if program confidentiality is required during transmission,
the code can be encrypted.
We utilise PGP since it is difficult to break, equipped with public and private
keys generation, widely available, supported on a range of operating system platforms, and one of the most popular encryption techniques. PGP public keys can
be obtained either through personal communication between the LogicWeb system
user and program owners, or via the Web: PGP public keys are increasingly being
placed on homepages, where they can be readily retrieved by the LogicWeb system.
The distribution of PGP keys is discussed further in §11.
5 Specifying Security Policies
A LogicWeb program accesses system resources via system calls and context switching. The resources used via context switching are socket connections for downloading pages and local storage for programs.
A policy program defines three predicates for controlling the system calls and
context switching of its designated LogicWeb programs. The predicates are:
• valid program(Type, URL): this predicate specifies the LogicWeb programs
which can be used via context switching. For example, restricting programs
to those from the domain www.cs.mu.oz.au is given by:
valid_program(get, URL) :- contains(URL, "http://www.cs.mu.oz.au/").
The following query will fail when using that policy since the goal refers to
an invalid domain:
?- lw(get, "http://www.cs.rmit.edu.au/")#>h_text(Src).
• valid systemCall(Call): this predicate defines the set of system calls a
program is allowed to invoke. Call is a term representing the form the system
call may take.
• call system(Call): this predicate is a wrapper for the allowed predicates
defined by valid systemCall/1. Call is a term representing the goal which is
to be invoked in a special way. Typically, call system/1 is used to implement
more restrictive versions of system calls.
The difference between valid systemCall/1 and call system/1 is execution.
Call patterns specified in valid systemCall/1 are used to test system calls, rejecting ones which do not match the necessary requirements. call system/1 is
used to execute system calls in novel (usually restricted) ways. A policy program
could be written with the body code of valid systemCall/1 implemented in
call system/1, doing away with the need for the valid systemCall/1 predicate.
However, the distinction between testing and execution would then be much less
clear.
The following policy program permits the retrieval (using the GET method) of
all URLs except http://www.cs.mu.oz.au/~swloke/private.html. It disallows
all calls to system/1 and open/3, except when open/3 reads dump.txt:
16
S.W. Loke and A. Davison
valid_program(get, URL) :URL \== "http://www.cs.mu.oz.au/~swloke/private.html".
valid_systemCall(open(’/home/pgrad/swloke/lws/dump.txt’, read, S)).
valid_systemCall(Call) :Call \= open(_, _, _),
Call \= system(_).
call_system(Call) :built_ins:call_builtin(Call).
call builtin/1 is a system predicate which carries out type checking on a call
before executing it. For example, the call builtin/1 clause for open/3 is:
call_builtin(open(FileName, Mode, Stream)) :atom(FileName),
( Mode = read
; Mode = write
),
var(Stream),
open(FileName, Mode, Stream).
FileName must be an atom, Mode either read or write, and Stream a variable.
Type checking increases the robustness of the system by preventing instantiation
faults when the arguments are the wrong type.
We briefly consider two examples that show the flexibility of security policies.
5.1 A Historical Policy
State information can be used to implement a history-based policy. For example,
the following policy lets a program carry out context switching (i.e. execute goals
such as lw(get, URL)#>Goal) until a file is accessed. A file access is detected by
having call system/1 monitor open/3 calls.
file_accessed(no).
% state information: no file has been accessed
valid_program(_, _) :file_accessed(no).
% only context switch when no
% files have been accessed
call_system(Call) :Call \= open(_, _, _),
built_ins:call_builtin(Call).
call_system(open(F, R, S)) :built_ins:call_builtin(open(F,
(file_accessed(no) ->
retract(file_accessed(no)),
assert(file_accessed(yes))
;
true
).
% process non-open/3 calls
% process open/3 call
R, S)),
% record file access
Once open/3 has been called, file accessed/1 will be changed to hold the value
yes. This will cause subsequent calls to valid program/2 to fail, disabling context
switching for the program.
Secure Prolog Based Mobile Code
17
5.2 Levels of Trust
It is straightforward to build policies with varying levels of trust for different programs. In the following example we use three levels: dangerous, ok, and safe. A
‘dangerous’ program cannot download any programs, or execute any system calls.
An ‘ok’ program can download programs but it can only write to the directory /tmp
and cannot use system/1 to delete files. A ‘safe’ program has no restrictions placed
upon it, except that a message is printed out when a file deletion succeeds. Each
level of trust is represented by a separate policy program.
The policy program for ‘dangerous’:
valid_program(_, _) :- fail.
valid_systemCall(_) :- fail.
call_system(_) :- fail.
Any call that uses system resources will fail.
The policy program for ‘ok’ is given as:
valid_program(_, _).
% no restrictions on downloads
valid_systemCall(open(Path, write, _)) :!, append("/tmp/", _, Path).
% only write to /tmp
valid_systemCall(system(Cmd)) :append("rm ", _, Cmd), !, fail.
% disallow file deletions
valid_systemCall(_).
% everything else allowed
call_system(Call) :built_ins:call_builtin(Call).
% no restrictions on execution
The policy program for ‘safe’:
valid_program(_, _).
% no restrictions on downloads
valid_systemCall(_).
% all system calls permitted
call_system(Cmd) :built_ins:call_builtin(Cmd),
(append("rm ", _, Cmd) ->
report_deletion(Cmd)
;
true
).
% report deletions when executed
We could call report deletion/1 in valid systemCall/1 but reporting a file deletion has nothing to do with testing the validity of a system call; it is a diagnostic
associated with execution.
6 Combining Security Policies
A LogicWeb program must not be allowed to perform an illegal system operation
by invoking a goal in a more privileged program. Hence, the security model must
ensure the following:
18
S.W. Loke and A. Davison
A
B#>read_file(Contents)
For B#>read_file(Contents),
consult the policy of A.
B
... read_file(Contents)...
For read_file(Contents),
consult the policy of A and that of B.
Fig. 4. The invocation of B#>read file(Contents) in A.
1. Context switching must not transfer privileges between programs. The system
has to ensure that an untrusted program does not illegally access resources
by invoking a goal in a trusted program. For example, assume that a program
P is not allowed to read a file (as determined by its policy) while another
program Q can (as determined by Q’s policy). This means that P should not
be able to call a goal like Q#>read file(Contents) to read a file. Clearly, it
is too simplistic to validate a call to read file/1 in Q using only the policy
of Q; the policy of P must also be considered.
In a similar manner, a LogicWeb goal cannot be allowed to perform an illegal
system operation because it is invoked in a trusted program. For example, if
A can read a file which B can not, then the goal B#>read file(Contents)
invoked in A should fail. This means that although A uses B, A should not pass
its privileges to B, since the system trusts A but not B. Hence, the policies of
both A and B must be consulted to determine if a read file/1 call should be
allowed, as Figure 4 illustrates.
When more than two programs are involved, a similar but slightly more complex situation arises. For example, consider a goal B#>(C#>G) invoked in the
program A, where G is a system call. If G is allowed by the policies of A and
C, but B does not permit G, then B is performing an illegal system operation
through C.
Essentially, disregarding any one of the policies potentially allows an illegal
system operation. This means that to determine if G should be allowed, all
the policy programs used by the code must be considered.
Figure 5 shows the changes in goal evaluation contexts starting from the goal
B#>(C#>G) in program A and ending with goal G in C. The policies of all
the programs involved in the evaluation of the goal starting from the first
program (i.e., A) are required. Hence, the proof rules for the security model
must incorporate a notion of the current set of policy programs which grows
as goal evaluation progresses.
Secure Prolog Based Mobile Code
19
A
... B#>(C#>G) ...
For B#>(C#>G),
consult the policy of A.
B
C
... C#>G ...
... G ...
For C#>G,
consult the policy of
A and that of B.
For G, consult the
policy of A, and that of B and of C.
Fig. 5. The evaluation B#>(C#>G) in program A.
2. LW-compositions must not transfer privileges between programs. A LogicWeb
program can invoke the clauses of another program in a LW-composition. For
example, consider a goal G evaluated in the LW-composition of programs S
and T: (S + T)#>G. G’s evaluation may utilise some clauses from S and some
from T, so in general G must be allowed by both the policies for S and T.
The observations made above can be made more formal by restating them using
the LW-encapsulation and LW-intersection operators (@ and *). Namely, given a set
of policies (e.g., {P,Q}), the allowable set of system calls and downloads are those
permitted by the composition @P * @Q.
LW-encapsulation is used to express that each policy must separately validate
the system calls and downloads. LW-intersection is used to represent that a given
system call or download must be permitted by every policy program. For example,
an open/3 call is permitted by both P and Q when the following goal succeeds:
?- (@P * @Q)#>valid_systemCall(open(_,_,_)).
7 Enforcing Security Policies
This section defines a new derivation relation, extending the operational semantics
in §2.2.3 to use policy programs. Previously, a goal was evaluated in the current
context, with the current set of created programs. In the new derivation relation, we
also consider the current set of policy programs. The derivation relation is extended
as follows:
For any goal formula G and program expression E,
′
Σ, S, E ⊢S
θ G
denotes the fact that there exists a top-down derivation of G in E starting with the set
S of existing LogicWeb programs and the ordered set of policy programs Σ, and ending
with the computed answer substitution θ and created program set S ′ .
20
S.W. Loke and A. Davison
The proof rules which define this new relation are an extension of the rules
in §2.2.3. Rules (1) to (8) are modified to form corresponding rules for the new
derivation relation by using the following syntactic mapping:
Replace every occurrence of an expression involving the derivation relation:
′
S, E ⊢S
θ G
with the corresponding expression:
′
Σ, S, E ⊢S
θ G
For instance, the following rule is obtained by applying this mapping to rule (2)
(conjunction):
′
′′
Σ, S, E ⊢Sθ G1 ∧ Σ, S ′ , E ⊢Sγ G2 θ
′′
Σ, S, E ⊢Sθγ G1 , G2
The same set of policy programs Σ is used for each of the conjuncts since they
occur in the same rule (and hence, in the same program).
7.1 Mapping from Programs to Policies
To aid the discussion which follows, two functions are defined: one maps a LogicWeb
program to its policy program, and the other relates a set of LogicWeb programs to
their policy programs. Φ denotes the set of all policy programs used by the system.
A policy program is permitted free access to system resources, and is not itself
assigned a policy program.
Definition 7.1 (policy program assignment)
The function
pol : (LW P rogramIDs \ ids(Φ)) → Φ
takes a program identifier (which is not that of a policy program) and returns a
policy program for that program from Φ.
pol is not defined on policy programs, and the empty program ξ is not assigned a
policy program.
Definition 7.2 (policies for a set of programs)
The function
pols : ℘(LW P rogramIDs \ ids(Φ)) → ℘(Φ)
takes a set of (non-policy) program identifiers I and returns the set of policy programs assigned to the programs in I. pols is defined by:
pols(I) = {pol(i) | i ∈ I}
7.2 Context Switching Revisited
*Σ@ denotes the LW-intersection of the LW-encapsulation of the policy programs
in Σ (i.e., if Σ = {P1 , ..., Pn } then *Σ@ = @P1 *...* @Pn ), and G denotes a goal.
Secure Prolog Based Mobile Code
21
Evaluating a goal against *Σ@ has the effect of evaluating the goal against each
program in Σ separately.
The new definition of the context switching rule (9) is:
Σ 6= ∅ ∧
W here each lw(Typei, URLi) ∈ expids(F ′ ) \ ids(Φ),
′
∅, S, ξ ⊢Sγ *Σ@ #>(valid program(Type1, URL1),
. . . ,valid program(TypeN, URLN))
I ⊆ ids(S ′ ) ∧
∧
′′
(pols(I \ ids(Φ)) ∪ Σ), S ′ , F ′ ⊢Sθ G
′′
Σ, S, E ⊢Sθ F #> G
(9’a)
The new rule has been augmented with:
i. A test to determine if each non-policy program to be used in the new context is
allowed by the current set of policy programs Σ. The test is done by invoking the
predicate valid program/2 in *Σ@ for each such non-policy program identified
by lw(Typei, URLi) using a goal of the form:
′
∅, S, ξ ⊢Sγ *Σ@ #>(valid program(Type1, URL1), . . . ,
valid program(TypeN, URLN))
N is the number of program identifiers in expids(F ′ ) \ ids(Φ). A LogicWeb goal
is used when evaluating the valid program/2 goals in order to download the
policy programs. Note that goal evaluation fails if not all the policy programs
are downloaded. Also, a policy program can invoke non-policy programs in its
rules, and the evaluation of the valid program/2 goals begins with an empty
policy program set. The rule for the case where Σ is empty is given below.
Note that we test all the programs in F ′ = insertCC(F, E) instead of F , which
means that we test the programs in the current context as well.
ii. An extension of Σ with new policy programs. New policy programs are added
to the front (left) of the ordered set Σ:
pols(I \ ids(Φ)) ∪ Σ
The last (rightmost) program in the ordered set is, chronologically, the first
policy program, and its significance is explained below.
The following variant of the context switching rule caters for when Σ is empty.
In that case, no policy programs are employed, no checks are made:
I ⊆ ids(S ′ ) ∧
′′
pols(I \ ids(Φ)), S ′ , F ′ ⊢Sθ G
′′
∅, S, E ⊢Sθ F #> G
(9’b)
In (i), we applied valid program/2 to all the programs used in the goal’s program expression and all the programs in the current context. We illustrate why by
considering an example:
22
S.W. Loke and A. Davison
valid_program(get, ’A’).
valid_program(get, ’B’).
valid_program(get, ’C’) :- fail.
% may be omitted but added for clarity
If the goal ((A+B)*C)#>G is used in A then we would expect it to fail since A is
not permitted to use B in context switching. Suppose instead that the goal was
((#)*C)#>G and the current context (as represnted by #) was A+B. The goal is
equivalent to the earlier one; should it be permitted or not?
If the goal is analysed by testing only C and ignoring the current context, then
A would be allowed to use B in context switching, so violating A’s policy. However,
this behaviour appears harmless since B must already have been retrieved if it is in
the current context, and so its utilisation will require no additional resources.
This would not be the case if the semantics of add programs in Definition 2.2
were different. Suppose that A’s use of B did result in the retrieval of a new version
of B, then resources would be used. Or suppose that programs are only retrieved
when their clauses are actually needed, and so B may not have been retrieved at
the time of G’s evaluation.
By conservatively testing the current context, we stay faithful to A’s policy: A
cannot use B via context switching, regardless of whether B is downloaded. This
design is more general since it makes the security model more tolerant of changes
in the system’s download and storage mechanisms. Also, the intention of A’s policy
is arguably to unconditionally bar A from using B via context switching.
We could easily implement the less conservative view: namely that A cannot use
B via context switching, unless B is already downloaded (or in the current context).
A’s policy program would become:
valid_program(get, ’A’).
valid_program(get, ’B’).
valid_program(get, ’C’) :- program_exists(’B’).
program exists/1 is a LogicWeb built-in which succeeds if the specified program
is in the local cache.
7.3 System Call Rules
Two new rules called system call rules are added to specify how policy programs
are utilised with system calls. The first rule specifies a test to determine if a goal
is a valid system call. It invokes call system/1 in the last policy program in Σ,
which is the policy for the main program of the application. This means that any
state changes carried out by call system/1 (for instance, see the example in §5.1)
will be done in the main program’s policy program.
Σ 6= ∅ ∧
′
∅, S, ξ ⊢Sγ *Σ@ #>valid systemCall(G) ∧
Σ = Σ′ ∪ Pn ∧
′′
∅, S ′ , ξ ⊢Sθ Pn #>call system(G)
′′
Σ, S, E ⊢Sθ G
(10’a)
Secure Prolog Based Mobile Code
23
Substitutions resulting from calling valid systemCall/1 are discarded. Instead,
the substitutions computed from call system/1 are kept. As before, proofs occurring in policy programs proceed with an empty policy program set.
For the case where Σ = ∅, the rule is
′
∅, S, E ⊢S
γ built ins:builtin(G)
∧
∅, S ′ , E ⊢S
θ
∅, S, E ⊢S
θ
′′
′′
built ins:call builtin(G)
G
(10’b)
A call to a built-in predicate in the module built ins is represented by:
G succeeds with θ
∅, S, E ⊢Sθ built ins:(G)
(11)
The evaluation of the goal is outside the scope of the inference rules, but θ is
assumed to be the computed answer substitution.
7.4 Properties of the Inference Rules
We show how the inference rules disallow illegal system operations. First, context
sequence and illegal system operation are defined.
Goal derivation involving LogicWeb goals may result in several sequences of context switches, each represented by a sequence of contexts. For example, the sequence
of contexts in Figure 5 is:
A, B, C
Note that the only operational semantic rules which change the context are the
context switching rules (i.e. 9’a and 9’b).
Definition 7.3 (context sequence)
Given a sequence of applications of inference rules in a top-down derivation, suppose
that there are n − 1 applications of the context switching rules. Further suppose
that the ith application of these rules causes the context to switch from Ei to Ei+1 ,
where 1 ≤ i ≤ n − 1. The context sequence is the sequence of contexts E1 , . . . , En .
Informally, given a context sequence E1 , . . . , En , an illegal system operation is
performed in the context Ei if a system call, or the use of a program in context
switching, is disallowed by the policy of some program in Ei , but is attempted in
Ei or some later context.
Definition 7.4 (illegal system operation)
Given a context sequence E1 , . . . , En , for some Ei , P ∈ pols(expids(Ei ) \ ids(Φ)),
i ≤ j ≤ n, and system call G, an illegal system operation is performed in
Ei if the goal call system(G) is invoked when the context of G is Ej but
′
∅, S, P ⊢Sθ valid systemCall(G) does not hold for any S, S ′ , and θ, or the oracle
function is called on the identifier lw(Type, URL) when the current context is Ej
′
but ∅, S, P ⊢Sθ valid program(Type, URL) does not hold for any S, S ′ , and θ.
The context switching rules guarantee that all policy programs which must be
consulted during a derivation are present.
24
S.W. Loke and A. Davison
Lemma 7.1
Given the context sequence E1 , . . . , En for a top-down derivation, for any i ∈
{1, . . . , n}, at the node with context Ei , the current set of policy programs Σ contains the policy programs of all the non-policy programs occurring in E1 , . . . , Ei .
Proof
The proof is by induction on i. For i = 1, the goal evaluation starts in the empty
program ξ with Σ = ∅. For i ≥ 2, (by the inductive hypothesis) when the goal is
evaluated in Ei−1 , Σ contains the policy programs of all the non-policy programs
occurring in E1 , . . . , Ei−1 . Ei−1 changes to Ei using one of the context switching
rules. This rule updates Σ (say, to Σ′ ) by adding to Σ the policy programs of the nonpolicy programs occurring in Ei . Hence, the goal evaluation continues in Ei with Σ′
containing all the policy programs for the non-policy programs in E1 , . . . , Ei .
The theorem below implies soundness of the security model: all goals which have
a successful derivation have the desired security property of not performing an
illegal system operation.
Theorem 7.1 (safety property)
No illegal system operation is performed in any context during a top-down derivation using the above rules.
Proof
Given the context sequence E1 , . . . , En , for any i ∈ {1, . . . , n}, suppose that an
illegal system operation was performed in Ei , say when the current context was Ej ,
where i ≤ j ≤ n, the set of policy programs was Σ, and the set of created programs
was S. To perform this operation either (1) the system call rule for Σ 6= ∅ or (2) the
new context switching rule for Σ 6= ∅ must have been used. In case (1), according
to the definition of the system call rule used, since call system(G) was invoked,
the goal:
′
∅, S ∪ Σ, *Σ@ ⊢Sγ valid systemCall(G)
must have succeeded. In case (2), suppose that the LogicWeb goal was F #>G
and F ′ = insertCC(F, Ej ), and the oracle function was called on the identifier
lw(Type, URL) ∈ expids(F ′ ) \ ids(Φ), then according to the definition of the context switching rule used, the goal:
′
∅, S ∪ Σ, *Σ@ ⊢Sγ valid program(Type, URL)
must have succeeded. But by the definition of an illegal system opera′
tion, for some P ∈ pols(expids(Ei ) \ ids(Φ)), in case (1), ∅, T, P ⊢Tθ
valid systemCall(G) does not hold for any T , T ′ , and θ, and in case (2),
′
∅, T, P ⊢Tθ valid program(Type, URL) does not hold for any T , T ′ , and θ. In
either case, it must have been that P 6∈ Σ. By the above lemma, Σ contains the
policy programs of all the non-policy programs occurring in E1 , . . . , Ej , namely,
Σ ⊇ pols(expids(Ei ) \ ids(Φ)). This means P ∈ Σ, and hence there is a contradiction.
Secure Prolog Based Mobile Code
25
8 Implementation
8.1 Interpreter with Security Mechanisms
The inference rules presented above provide the basis for an interpreter for evaluating goals in the presence of policy programs. Program 1 shows this interpreter,
which is an extension of the one described in (Loke and Davison 1998b).
% demo/3 with LogicWeb goal
demo(PL, E, F#>G) :establish_context(PL, F, E, PL, NPL, F1),
demo(NPL, F1, G).
demo(PL, E, G) :allowed_systemCall(PL, G), invoke_systemCall(PL, G).
demo(_PL, P, built_ins:G) :pgpID_to_policyID(_, P), call(built_ins:G).
demo(_PL, _E, true).
demo(PL, E, (A, B)) :- demo(PL, E, A), demo(PL, E, B).
demo(PL, E, A) :- select_clause(PL, E, (A :- B)), demo(PL, E, B).
% establish a context
establish_context(OPL, E + F, C, PL, PL2, E1 + F1) :establish_context(OPL, E, C, PL, PL1, E1),
establish_context(OPL, F, C, PL1, PL2, F1).
establish_context(OPL, E * F, C, PL, PL2, E1 * F1) :establish_context(OPL, E, C, PL, PL1, E1),
establish_context(OPL, F, C, PL1, PL2, F1).
establish_context(OPL, E / P, C, PL, PL2, E1 / P1) :establish_context(OPL, E, C, PL, PL1, E1),
establish_context(OPL, P, C, PL1, PL2, P1).
establish_context(OPL, @E, C, PL, PL1, @E1) :establish_context(OPL, E, C, PL, PL1, E1).
establish_context(OPL, (/)<>(E, L), C, PL, PL2, (/)<>(E1, L1)) :establish_context(OPL, E, C, PL, PL1, E1),
establish_contextL(OPL, L, C, PL1, PL2, L1).
establish_context(OPL, Op<>L, C, PL, PL1, Op<>L1) :establish_contextL(OPL, L, C, PL, PL1, L1).
establish_context(OPL, lw(T, U), _C, PL, NPL, lw(T, U)):allowed_program(OPL, lw(T, U)), download(T, U),
add_policyID(lw(T, U), PL, NPL).
establish_context(OPL, (#), C, PL, PL, C) :allowed_programs(OPL, C).
establish_contextL(_OPL, [], _C, PL, PL, []).
establish_contextL(OPL, [E|Es], C, PL, PL2, [E1|Es1]) :establish_context(OPL, E, C, PL, PL1, E1),
establish_contextL(OPL, Es, C, PL1, PL2, Es1).
% predicate to ensure that only allowed system calls are invoked
allowed_systemCall([], G) :built_ins:builtin(G).
allowed_systemCall([Pol|Pols], G) :demo([], empty, Pol#>valid_systemCall(G)),
allowed_systemCall(Pols, G).
Program 1. The interpreter for pure LogicWeb programs which use policy programs.
26
S.W. Loke and A. Davison
% predicate to invoke system calls
invoke_systemCall([], G) :built_ins:call_builtin(G).
invoke_systemCall(PL, G) :last(P, PL), demo([], empty, P#>call_system(G)).
% add a policy program ID to list of IDs
add_policyID(Id, PL, PL) :pgpID_to_policyID(_, Id), !.% no policy program for policy programs
add_policyID(Id, PL, NPL) :policyID(Id, NewPolicyId),
(member(NewPolicyId, PL) ->
NPL = PL
;
NPL = [NewPolicyId|PL]
).
% predicate to ensure that only allowed programs are downloaded
allowed_program([], _Id).
allowed_program([Pol|Pols], lw(Type, URL)) :demo([], empty, Pol#>valid_program(Type, URL)),
allowed_program(Pols, lw(Type, URL)).
% select a clause from the program store
select_clause(_PL, lw(Type, URL), A :- B) :lw(Type, URL)::(A :- B).
select_clause(PL, E + _F, A :- B) :select_clause(PL, E, A :- B).
select_clause(PL, _E + F, A :- B) :select_clause(PL, F, A :- B).
select_clause(PL, E * F, A :- (B,C)) :select_clause(PL, E, A :- B), select_clause(PL, F, A :- C).
select_clause(PL, E / P, A :- B) :select_clause(PL, E, A :- B), not defined(A, P).
select_clause(PL, @E, A :- true) :- demo(PL, E, A).
select_clause(PL, (/)<>(E, []), A :- B) :select_clause(PL, E, A :- B).
select_clause(PL, (/)<>(E, [P|Ps]), A :- B) :select_clause(PL, (/)<>[(E / P)|Ps], A :- B).
select_clause(PL, _Op<>[E], A :- B) :select_clause(PL, E, A :- B).
select_clause(PL, Op<>[E1,E2|Es], A :- B) :C =.. [Op, E1, E2], select_clause(PL, Op<>[C|Es], A :- B).
defined(A, P) :functor(A, Functor, Arity), functor(H, Functor, Arity),
P::(H :- _B).
Program 1. (Continued)
demo/3 is derived from a standard vanilla Prolog meta-interpreter (Sterling and Beer 1989;
Sterling and Shapiro 1994), and executes a goal (the third argument of demo/3)
against its program context (the second argument). The first argument of demo/3
stores the IDs of downloaded policy programs.
In the first clause of demo/3, establish context/6 returns the new policy program identifiers for the programs in the new context.
Secure Prolog Based Mobile Code
27
The second clause checks for and invokes an allowed system call. In
allowed systemCall/2, valid systemCall/1 is invoked in each policy program.
Policy programs known to the system are recorded in pgpID to policyID/2, which
is described below. invoke systemCall/2 invokes the system call in the last program mentioned in the list of policy program identifiers (PL). This is the policy
program for the main program, as mentioned earlier.
The third clause of demo/3 allows policy programs to invoke goals in the Prolog
module built ins.
The last three clauses of demo/3 implement the basic meta-interpreter.
select clause/3 in the final clause evaluates the goal A by looking for a suitable clause (A :- B) from the programs defined by the program expression E. The
clauses of downloaded programs are held in the program store (see Figure 2) in the
format lw(Type,URL)::(Clause).
The interpreter is called using a goal such as:
?- demo([], empty, lw(get, "http://main.program")#>query).
Goal derivation begins with the empty context without any policy programs and a
goal where lw(get, "http://main.program") is the main program of an application.
establish context/6 carries out several tasks. One of these is to expand the
program expression F (which was part of a context switching goal F #>G), so that
any context operators (#) in F are replaced by the current context. This results in
a new program expression F1. Note that two copies of the original list of policy
programs is passed into establish context/6 initially. The fourth and fifth arguments of establish context/6 represent the current extended list and the newly
extended list respectively. Because the current extended list may not be the original
and the security tests are done using the original, the extra copy is supplied.
The last two clauses of establish context/6 relate to the security model. The
penultimate clause checks a new program ID against the existing policies, and may
download a new policy for that program, which causes the list of policies IDs (PL)
to be extended.
allowed program/2 checks a program ID (lw(T,U)) against the existing policies
by calling valid program/2 against each policy in turn. This mimics the LWintersection of LW-encapsulated policies described earlier.
add policyID/3 attempts to retrieve a policy ID for the specified program ID
(lw(T,U)), and add it to the policy IDs list. This process is complicated by the
possibility that the program ID may actually be a policy ID, which is ascertained
by checking it against pgpID to policyID/2. However, if lw(T,U) is an ordinary
LogicWeb program ID, then policyID/2 will contain a mapping from the program
ID to its matching policy ID. How this is achieved will be explained shortly.
The final clause of establish context/6 checks all the programs in the current context against the existing policies. This check is required only once, but for
simplicity, we have permitted the code to perform this check every time a context
operator (#) is encountered in a program expression.
28
S.W. Loke and A. Davison
8.2 Installing Programs
The predicate for downloading a page and installing its corresponding LogicWeb
program is given below. Its operation parallels the stages shown in Figure 3.
download(Type, URL) :created(Type, URL), !.
download(Type, URL) :retrieve(Type, URL, Contents),
create_program(Type, URL, Contents),
assign_policyID(Type, URL, Contents).
%
%
%
%
%
program already exists
program does not exist
retrieve from the Web
create the program
assign a policy program
If the program stored at URL is not already present in the program store then its
page is retrieved from the Web, converted to a program, and assigned a policy.
The policy mechanism is located in assign policyID/3, which is defined as:
assign_policyID(Type, URL, Contents) :(pgpID_to_policyID(_, lw(Type, URL)) ->
% if policy program
true
% lw(Type, URL) is not assigned a policy program
;
determine_policyID(URL, Contents, PolicyID),
record_policyID(Type, URL, PolicyID)
).
determine_policyID(URL, Contents, PolicyID) :(pgp_signed(URL) ->
% if digitally signed
authenticate(Contents, PGPID), % try authentication
pgpID_to_policyID(PGPID, PolicyID)
;
pgpID_to_policyID(unknown, PolicyID)
).
record_policyID(Type, URL, PolicyID) :assert(policyID(lw(Type, URL), PolicyID)).
assign policyID/3 tries to obtain a security policy for a downloaded (nonpolicy) LogicWeb program by calling determine policyID/3. It bases its search on
the result of calling pgp signed/1 which checks if the contents of the downloaded
page are digitally signed. It does this by inspecting the URL for the extension
“.lwpgp.html”, which denotes a digitally signed LogicWeb program.
If the program is signed then the signature is authenticated with
authenticate/2. It invokes PGP’s authentication procedure on Contents, which
contains the HTML text and its digital signature. This stage is depicted in Figure 3
in box B. PGP extracts the digital signature from Contents, and attempts to decrypt it to obtain an MD5 string by employing its collection of public keys. These
keys must have been previously added to PGP by the LogicWeb administrator. On
successful decryption, PGP checks the HTML text from Contents against the MD5
string, and returns the PGP ID labelling the key that decrypted the signature as
PGPID. On authentication failure (i.e., if the signature is not decrypted or if the
MD5 string does not match the HTML text), PGPID is instantiated to unknown.
The signature ID returned by authenticate/2 is utilised by
pgpID to policyID/2 to lookup a policy ID. A typical pgpID to policyID/2 fact:
Secure Prolog Based Mobile Code
29
pgpID_to_policyID(’Seng W. Loke <[email protected]>’,
lw(get, "http://www.cs.mu.oz.au/~swloke/my_policy.html").
The first argument is Seng Wai Loke’s PGP ID, and the second a policy program
identifier.
If the downloaded LogicWeb program is not signed then determine policyID/3
makes use of the default policy ID associated with the ‘unknown’ signature ID:
pgpID_to_policyID(unknown,
lw(get, "http://www.cs.mu.oz.au/~swloke/default_policy.html").
Back in assign policyID/3, the policy ID returned by determine policyID/3
is passed to record policyID/3. It asserts a policyID/2 fact which has as its first
argument the LogicWeb program ID and its second is the matching policy ID. This
fact can be used subsequently to map from a program ID to its policy. The fact
is not asserted into the LogicWeb program itself since it must be impossible for a
rogue program to redefine policyID/2.
9 Control of Resource Usage
This section considers denial-of-service attacks by showing how it is possible to
control an operation or resource with policy programs and meta-interpreters.
9.1 Resource Control Using Policy Programs
Resource usage can be monitored by keeping contextual or state information within
a policy program, allowing decisions about the utilisation of resources to be based
on the execution history. For example, a limit can be imposed on the frequency of
system calls.
The following definition of call system/1 permits up to ten files to be opened
at a time, thereby limiting the number of allocated file descriptors.
30
S.W. Loke and A. Davison
call_system(P) :P \= open(_, _, _),
P \= close(_),
built_ins:call_builtin(P).
call_system(open(F, R, S)) :open_count(N), N < 10,
increment_open_count,
built_ins:call_builtin(open(F, R,
call_system(close(S)) :open_count(N), N > 0,
decrement_open_count,
built_ins:call_builtin(close(S)).
% for non- open/3, close/1 calls
% open a file
% increment count of open files
S)).
% close a file
% decrement count of open files
increment open count/0 and decrement open count/0 update the value stored
in open count/1.
9.2 Resource Control Using Meta-interpreters
As pointed out in (Ousterhout et al. 1997), denial-of-service attacks are not as
severe as other kinds of attack since the user can always hit the “kill key”
and exit the system. However, the graceful termination of goal evaluation is
preferable so that the system and application can at least continue running.
Execution control can be readily incorporated into the system by using metainterpreters (Sterling and Beer 1989; Sterling and Shapiro 1994). Their use for loop
checking and detecting resource limits is described below.
9.2.1 Loop Checking
Our approach to loop checking uses two meta-interpreters. solve ad/2 evaluates
a goal and also stores information about ancestor goals and the recursion depth.
Program 2 shows a simplified version of solve ad/2 for pure Prolog.
solve_ad(_Depth_Ancs, true).
% evaluate true
solve_ad(Depth-Ancs, (G1, G2)) :% evaluate conjunction
solve_ad(Depth-Ancs, G1),
solve_ad(Depth-Ancs, G2).
solve_ad(Depth-Ancs, Goal) :% evaluate goal
copy_term(Goal, CopyGoal),
% make a copy of the goal
clause(Goal, Body),
% select a matching clause
spy_point(Depth-Ancs, Goal),
% insert a spy_point
Depth1 is Depth + 1,
% increment recursion depth
solve_ad(Depth1-[CopyGoal|Ancs], Body).
Program 2. A meta-interpreter for pure Prolog with an argument carrying the
recursion depth and a list of ancestor goals.
The purpose of the spy point/2 call will be explained shortly.
The solve t/2 meta-interpreter terminates goal evaluation whenever some predefined termination condition becomes true. Such conditions are specified as
Secure Prolog Based Mobile Code
31
terminate/1 clauses. solve t/2 also sets a termination flag to ‘terminated’ when
a termination condition becomes true. Program 3 shows a simplified version of
solve t/2 for pure Prolog.
solve_t(true, _).
% evaluate true
solve_t(Goal, T) :terminate(Goal), !,
% check termination conditions
T = terminated.
solve_t((G1, G2), T) :% evaluate conjunction
solve_t(G1, T),
(T == terminated ->
true
;
solve_t(G2, T)
).
solve_t(Goal, T) :% evaluate goal
clause(Goal, Body),
solve_t(Body, T),
(T == terminated, ! ; true).
Program 3. A meta-interpreter for pure Prolog which checks for termination conditions.
The next step is to combine these meta-interpreters with the LogicWeb demo/3
interpreter described in the previous section. Instead of invoking demo/3 directly,
as in:
?- demo([], empty, LogicW ebGoal).
it will be called nested inside the two meta-interpreters:
?- solve_t( solve_ad(0-[], demo([], empty, LogicW ebGoal)), _T ).
solve ad/2 will record ancestor goals and recursion depth information about
demo/3. solve t/2 can check for termination conditions by looking for suitable
patterns in the spy point/2 calls in solve ad/2. For example:
terminate(spy_point(Depth-Ancs, demo(_, E, G))) :(
member(demo(_, E1, G1), Ancs),
E = E1,
% the goals are in the same context
variant(G, G1),
% and they are variants of each other
write(’loop found’)
;
Depth > 40,
% recursion depth too large
write(’maximum recursion depth exceeded’)
).
demo( , E, G) is the current LogicWeb interpreter goal, Depth is the current recursion depth, and Ancs is the list of ancestor goals. Note that the condition E = E1
will not detect loops where the context grows indefinitely. To detect such loops, the
size of the context can be compared against a preset limit.
The essential utility of this approach is the relative ease of combining distinct
32
S.W. Loke and A. Davison
meta-interpreters which monitor/control different aspects of the computation. Other
loop checking techniques are studied in (Bol 1991).
9.2.2 Resource Limits
The techniques from the last section can be easily utilised to monitor resource limits.
We will measure two resources: the number of LogicWeb programs downloaded,
and the number of clause applications. The first count will persist across different
solve t/2 invocations, the second will be reset at each invocation of solve t/2.
The necessary terminate/1 clauses for these measures are:
terminate(invoked_builtin(download(_,_))) :program_count(N),
N > 100,
% permit up to 100 program downloads
write(’maximum LogicWeb program count exceeded’).
terminate(invoked_builtin(clause(_,_))) :retract(clause_count(N)),
N1 is N + 1,
assert(clause_count(N1)),
(N > 500 ->
% permit up to 500 clause applications
writef(’maximum clause count exceeded’)
;
!,
fail
).
This code assumes that solve ad/2 can deal with built-in predicates, which would
require a minor extension to Program 2.
The first terminate/1 clause monitors program downloads by looking for calls
to the download/2 built-in. When there is such a call, program count/1 is checked
to see if the number of programs downloaded exceeds 100. If it does then execution is terminated. The use of program count/1 would require a change to
create program/3 (called by download/2) which translates a page into a program;
it would also have to increment program count/1 each time a new program was
created.
The second terminate/1 clause restricts the number of clause applications by
monitoring calls to the clause/2 built-in. The terminate/1 clause increments the
number stored in clause count/1 until it exceeds 500.
An oft-stated problem with nested meta-interpreters is the penalty incurred upon
execution speed. In fact, the termination checks described here could be implemented more efficiently by directly modifying demo/3. However, multiple metainterpreters are less complex to understand than a single, monolithic piece of
code. Also, they permit a compositional approach to implementing execution control: functionality is separately implemented and introduced. We also believe that
the speed costs are relatively unimportant if the application carries out frequent
Web requests, and hence spends most of its wall-clock execution time interacting with the Web. Moreover, partial reduction techniques for translating meta-
Secure Prolog Based Mobile Code
33
interpreters into specialised forms can be explored to remove the levels of interpretation (Sterling and Shapiro 1994).
10 Comparisons with Security Models in Other Mobile Code Systems
We review security models which share certain common features with LogicWeb.
First, we consider security approaches in some interpreted languages, and then
examine systems which utilise policy modules and authentication.
10.1 Security Models in Two Interpreted Languages
10.1.1 Safe-Tcl
Tcl is an interpreted imperative language, and access to system resources is via permitted commands of the interpreter. In the Safe-Tcl security model (Ousterhout et al. 1997),
security is enforced by making dangerous commands unavailable to scripts running
in the safe interpreter. Potentially dangerous operations, such as opening sockets,
can be carried out via wrappers or aliases. The wrappers ensure that the commands
are used in a controlled manner (e.g., only socket connections to some hosts are
permitted).
The security model in LogicWeb is partly motivated by the Safe-Tcl approach: LogicWeb’s valid system calls correspond to Safe-Tcl commands, and
call system/1 corresponds to wrappers for those commands. A safe interpreter in
Safe-Tcl corresponds to the LogicWeb program interpreter with appropriate policy
programs. However, the Safe-Tcl security model has no notion of authentication.
Safe-Tcl allows programs running in different interpreters (each with its own security policy) to communicate. Such communication effectively composes the security
policies of the programs, i.e. a program can use the privileges of other programs.
The dangers of this are highlighted in (Ousterhout et al. 1997), but not solved in
a structured way. In LogicWeb, the composition of policies do not violate safety
criteria.
10.1.2 Java Applets
A Java applet is transmitted in bytecode format, and then executed by an interpreter on the local host. This contrasts with Safe-Tcl and LogicWeb where source
code is transferred.
The LogicWeb system does not support concurrency or sophisticated GUI programming, and so its security model does not need to deal with multi-threading or
screen resources (e.g., windows). The Java language enforces type safety by having
the compiler ensure that class methods do not access memory in ways that are
inappropriate for their type (Sun Microsystems, Inc 1997). LogicWeb programs are
not explicitly typed and there is no class abstraction. However, type conversions
(e.g., strings to atoms) are done via system calls, and so the dangers of type conversion are avoided at call time. There are two trust levels for classes in the current
34
S.W. Loke and A. Davison
Java security model: trusted classes are local and part of the Java system, while
untrusted classes are ones that are downloaded. Multiple levels of trust are possible
with signed LogicWeb programs.
Security for Java applets, as implemented in Microsoft’s Internet Explorer 3.0
and Netscape Navigator 3.0, disallows applets from reading from or writing to
local files, or establishing network connections except to the originating host.
However, there are ways around these problems, including the use of server-side
databases and proxy servers. The digital signing of JAR (Java ARchive) files, as
supported in JDK 1.1.52 , bundles Java code and related files together into one
file (Sun Microsystems, Inc 1997). In a similar way, a collection of programs making up a LogicWeb application can be bundled, signed, and transported as one
archive file.
Microsoft Internet Explorer 4.0 now allows finer grained control over the capabilities granted to Java code, such as access rights to local files and network
connections (Microsoft Corporation 1999a; Microsoft Corporation 1999b). This is
done by including a list of capabilities requested by the applet in the applet’s signature. Approval for the capabilities are either pre-set or given via user dialog. The
implementation prevents transfer of privileges between classes in a similar way to
our model: if a class has not been directly granted a capability, it cannot obtain
that capability regardless of its caller’s capabilities. A class without a capability
cannot gain that capability by invoking a more privileged class.
10.2 Security Policy Modules in Two Mobile Code Systems
10.2.1 SERC’s Safer Erlang (SSErl)
The security mechanism in LogicWeb was motivated by the policy modules used
in SSErl, a functional declarative language for programming concurrent and distributed systems (Brown 1997a; Brown 1997b). In the SSErl framework, a policy
module is not associated with a particular program as in LogicWeb, but with a
SSErl node, a platform where multiple processes may run concurrently. System
operations are performed via built-in functions. A policy module specifies the allowable built-in functions for all the processes running at a node.
In contrast to SSErl, execution of LogicWeb programs is single-threaded, although the thread of execution can proceed from one program to another. At any
one time, at most one LogicWeb application is running. The overall policy for the
running application is represented by the current set of policy programs, and is determined dynamically since it is generally not possible to determine a priori which
programs an application will use.
10.2.2 Java Aglets
The Aglet workbench (Karjoth et al. 1997) allows Java programs (called aglets) to
be transferred between hosts as mobile agents. A specialised language for writ2
JDK 1.1.5 (final version) is available at http://java.sun.com/products/jdk/1.1/.
Secure Prolog Based Mobile Code
35
ing aglet security policy modules has been proposed which makes use of specific
roles such as the aglet’s manufacturer, owner, execution platform (e.g., the URL of
the host), and domain. In the LogicWeb implementation, the assignment of policy
programs is currently based solely on a program’s signature (as implemented by
determine policyID/3 in §8.2).
10.3 Authentication in Two Mobile Code Systems
10.3.1 Agent Tcl
Programs are authenticated with PGP in Agent Tcl (Gray 1996). PGP is invoked
externally in the same way as in LogicWeb, thereby keeping the Agent Tcl system
simpler and more flexible (the encryption mechanism can be easily replaced). Depending on the identity of the agent’s owner, an Agent Tcl program is assigned
resources such as CPU time, windows, the file system, external programs, and network connections.
10.3.2 ActiveX
ActiveX controls (Ernst 1996) are programs (in native x86 code) that can be executed on the client-side in Web pages. A digital signature is attached to each
ActiveX control which identifies the control’s author. The advantage of a signature
is that since the author is known, an ActiveX control is permitted more freedom to
access resources.
Unlike a LogicWeb program, which is sandboxed by policy programs and metainterpreters, a trusted ActiveX control is not restricted in any way and has access
to all operating system services. Since an ActiveX control is native code, a suitable
sandbox mechanism may be more difficult to implement (Object Management Group 1999).
11 Summary and Future Work
Security is a major concern for any system that executes downloaded code. This
paper has described a security model for LogicWeb which protects the client host
from integrity, privacy, and denial-of-service attacks. The security model is specified
using proof rules in the operational semantics of the language. This specification
permits a straightforward proof of the soundness of the model with respect to illegal
system operations. Resource control of predicates invoked at the application level
is carried out using policy programs. Resource control of programs using resource
counts and loop checking is performed by meta-interpreters. Authentication provides varying trust levels, and different policy programs define resource access for
the different trust levels.
The security model dynamically extends the LogicWeb system with selected policy programs. Since policy programs are separate from the system (and integrated
only on-demand) they are a convenient means for imposing consistent security restrictions on multiple hosts. For example, similar Intranet-wide security restrictions
36
S.W. Loke and A. Davison
can be imposed on a group of hosts by having them download the same policy programs (which, if needed, can be specialised by being composed with host-specific
policies).
As noted in (Ousterhout et al. 1997), it may be difficult to distinguish between
legitimate behavior and denial-of-service attacks. For instance, it is easy to introduce resource bounds but difficult to state appropriate values for those bounds. If a
value is too low, a useful task may be terminated prematurely; if too high, resource
wastage occurs. A possible solution to this dilemma is to involve the user’s judgment by providing query-the-user facilities via a meta-interpreter. The user could
also utilise trace information supplied by the meta-interpreter to detect unwanted
behavior, such as the movement of data between sites using the host’s resources.
Some Prolog systems, such as SICStus Prolog, have a time-out facility for goal
evaluation – the goal is terminated if its evaluation exceeds a given duration (Intelligent Systems Laboratory, Swedis
This can be incorporated as a hard limit after other checks in our meta-interpreters:
?- time_out(solve_t(solve_ad(0-[], demo([], empty, LogicW ebGoal)), _T),
Time, Result).
time out/3 terminates the execution of solve t/2 after Time milliseconds, with
Result being set to timeout or success.
Policy programs could be allocated based on the more general idea of credentials
instead of digital signatures alone. For instance, a program might be allowed certain
privileges not only because of its signature, but because it had sufficient advocates.
Recent work by Seamons et al (1997) on using Prolog to encode digital credential
acceptance policies and their verification logic could be exploited.
Policy programs simplify the administration of privileges to applications. Our
policy assignment depends solely on the digital signature of the LogicWeb program. A more flexible method would involve multiple roles such as the program’s
manufacturer, owner, execution platform, or domain. For instance, suppose we trust
particular domains (or URLs) and assume that data transfer is not intercepted and
modified. We can assign policies based on the domains or URLs of the LogicWeb
programs, regardless of whether they are signed. We modify determine policyID/3
of §8.2 as follows:
safe("http://www.cs.mu.oz.au/~swl").
safe("http://www.cs.ait.ac.th/~ad").
% trusted URLs
determine_policyID(URL, _Contents, PolicyID) :- % safe URLs get a standard policy
safe(URL), !,
PolicyID = lw(get, "http://www.policyforthetrusted.com").
determine_policyID(URL, Contents, PolicyID) :- % check signature for everything else
(pgp_signed(URL) ->
authenticate(Contents, PGPID),
pgpID_to_policyID(PGPID, PolicyID)
;
pgpID_to_policyID(unknown, PolicyID)
).
Domain or zone-based policy assignment can be implemented this way.
Secure Prolog Based Mobile Code
37
LW-composition operators could be used to create policies from dynamically composed programs. For instance, assume that the program at
http://www.usedownloadedpgms.com stores the clause:
valid program(X) :- program_exists(X).
determine policyID/3 can be modified to always combine this policy with the one
being downloaded:
determine_policyID(URL, Contents, Policy) :(pgp_signed(URL) ->
authenticate(Contents, PGPID),
pgpID_to_policyID(PGPID, PolicyID)
;
pgpID_to_policyID(unknown, PolicyID)
),
Policy = lw(get,"http://www.usedownloadedpgms.com")+PolicyID. % combine policies
We have only considered policy programs created by the user of the LogicWeb
system. If policies could be created by other parties, authentication would then
become necessary for policy programs in addition to ordinary LogicWeb programs.
A limitation of our model is that the system must contain all the public keys
required for the authentication of incoming programs. An automatic distribution
mechanism for PGP public keys is needed. For example, the LogicWeb system
could extract required public keys from homepages given their URLs, or could
query Internet PGP public key servers3 (Garfinkel 1995). But this requires that
the information be trusted, has not been tampered with, and is reliable. The use
of credentials, i.e. certification by other authorities, would be required.
The LogicWeb security model is conservative in that a trusted program is not
allowed to transfer its privileges to an untrusted program it is using. Besides resource usage privileges, another kind of privilege could be introduced: the right to
transfer privileges. This would be useful when the system did not know the signatories of all the programs that a trusted program utilised. The policy program
would have to specify which privileges were transferable, and whether the right to
transfer privileges was itself transferable.
References
Bol, R. 1991. Loop Checking in Logic Programming. Ph.D. thesis, CWI, Amsterdam,
The Netherlands.
Brogi, A. 1993. Program Construction in Computational Logic. Ph.D. thesis, Universita
di Pisa-Genova-Udine.
Brown, L. 1996. Mobile Code Security. Tech. rep., Australian Defence Force Academy,
Canberra, Australia. TR - CS07/96. September. Presented to AUUG96, Melbourne,
Australia and available at <http://www.adfa.oz.au/~lpb/TR/mcode96.html>.
Brown, L. 1997a.
Custom Security Policies in SSErl (DRAFT). Available at
<http://www.adfa.oz.au/~lpb/TR/ssp97/sserl pol97.html>.
3
A PGP public key server is located at
http://www-swiss.ai.mit.edu/~bal/keyserver.html.
38
S.W. Loke and A. Davison
Brown, L. 1997b.
Introducing SERCs Safer Erlang (DRAFT).
<http://www.adfa.oz.au/~lpb/TR/ssp97/sserl97a.html>.
Available at
Cugola, G., Ghezzi, C., Picco, G., and Vigna, G. 1997. Analyzing Mobile Code
Languages. In Mobile Object Systems: Towards the Programmable Internet (Lecture
Notes in Computer Science No. 1222). Springer-Verlag, 93–110.
Dean, D., Felten, E., and Wallach, D. 1996.
Java Security: From HotJava to Netscape and Beyond.
In Proceedings of the 1996 IEEE Symposium
on Security and Privacy. Oakland, California, U.S.A., 190 – 200. Available at
<http://www.cs.princeton.edu/sip/pub/secure96.html>.
Drossopoulou, S. and Eisenbach, S. 1998.
Towards an Operations Semantics and Proof of Type Soundness for Java.
Available at
<http://outoften.doc.ic.ac.uk/projects/slurp/pubs/LNCSJavaSem.ps.Z>.
Ernst, W. 1996. Presenting ActiveX. Sams.net.
Garfinkel, S. 1995. PGP: Pretty Good Privacy. O’ Reilly and Associates, Inc.
Gray, R. 1996.
Agent Tcl: A Flexible and Secure Mobile-agent System.
In Proceedings of the 4th Annual Tcl/Tk Workshop (TCL 96),
M. Diekhans and M. Roseman, Eds. Monterey, California, U.S.A. Available at
<http://www.cs.dartmouth.edu/~agent/papers/tcl96.ps.Z>.
Intelligent
Systems
Laboratory,
Swedish
Institute
puter Science. 1996.
SICStus Prolog User’s Manual.
<ftp://ftp.sics.se/archive/sicstus3/sicstus.ps.gz>.
of
ComAvailable at
Jensen, T., Mtayer, D. L., and Thorn, T. 1998.
Security and Dynamic
Class Loading in Java: A Formalisation.
In Proceedings of the IEEE International Conference on Computer Languages. Illinois, U.S.A.
Available at
<ftp://ftp.irisa.fr/local/lande/tjdlmtt-iccl98.ps.gz>.
Karjoth, G., Lange, D., and Oshima, M. 1997.
A Security Model
for Aglets.
IEEE Internet Computing 1, 4 (July/August).
Available at
<http://www.computer.org/internet/ic1997/w4toc.htm>.
Leroy, X. and Rouaix, F. 1998. Security Properties of Typed Applets. In Proceedings of the 25th ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages. San Diego, California, U.S.A.
Loke, S. and Davison, A. 1997. A Two-level World Wide Web Model with Logic
Programming Links. In Proceedings of the 2nd Workshop on Logic Programming Tools
for Internet Applications (in conjunction with the 14th International Conference on
Logic Programming), K. De Bosschere, M. Hermenegildo, and P. Tarau, Eds. 41 – 54.
Available at <http://clement.info.umoncton.ca/~lpnet/proceedings97/loke.ps>.
Loke, S. and Davison, A. 1998a. LogicWeb: Enhancing the Web with Logic Programming. Journal of Logic Programming 36, 3 (September), 195 – 240.
Loke, S. and Davison, A. 1998b. The Implementation of a System for Declarative
Mobile Code and Web Querying. In Proceedings of the 11th International Conference
on Applications of Prolog. 197 – 204.
Loke, S., Davison, A., and Sterling, L. 1996a. Lightweight Deductive Databases
on the World Wide Web. In Proceedings of the 1st Workshop on Logic Programming Tools for Internet Applications (in conjunction with the Joint International
Conference and Symposium on Logic Programming), P. Tarau, A. Davison, K. De
Bosschere, and M. Hermenegildo, Eds. Bonn, Germany, 91 – 106. Available at
<http://clement.info.umoncton.ca/~lpnet/lp-internet/lwddbs/lwddbs.html>.
Loke, S., Davison, A., and Sterling, L. 1996b. CiFi: An Intelligent Agent for Citation
Finding on the World Wide Web. In Topics in Artificial Intelligence, Proceedings of
Secure Prolog Based Mobile Code
39
the 4th Pacific Rim International Conference on Artificial Intelligence (Lecture Notes
in Artificial Intelligence No. 1114). 580 – 592.
Meseguer, J. and Talcott, C. 1997.
Rewriting Logic and Secure Mobility.
In Proceedings of Foundations for Secure Mobile Code Workshop. Available at
<http://www.cs.nps.navy.mil/research/languages/statements/meseguer.ps>.
Microsoft
Corporation.
1999a.
Microsoft
Security
Management
Architecture
White
Paper.
Available
at
<http://microsoft.com/ie/security/?/ie/security/ie4security.htm>.
Microsoft Corporation. 1999b.
Trust-Based Security for Java.
Available at
<http://www.microsoft.com/java/security/jsecwp.htm>.
Object Management Group. 1999. Comparing ActiveX and CORBA/IIOP. Available
at <http://www.omg.org/library/activex.html>.
Ousterhout, J., Levy, J., and Welch, B. 1997.
The Safe-Tcl Security
Model.
Tech. rep., Sun Microsystems Research Lab. March.
Available at
<http://www.sun.com/research/techrep/1997/smli tr-97-60.pdf>.
Seamons, K., Winsborough, W., and Winslett, M. 1997. Internet Credential Acceptance Policies. In Proceedings of the 2nd Workshop on Logic Programming Tools for
Internet Applications (in conjunction with the 14th International Conference on Logic
Programming), K. De Bosschere, M. Hermenegildo, and P. Tarau, Eds. 13 – 27. Available
at <http://clement.info.umoncton.ca/~lpnet/proceedings97/winsborough.ps>.
Sterling, L. and Beer, R. 1989. Metainterpreters for Expert System Construction.
Journal of Logic Programming 6, 1 & 2 (January/March), 163–178.
Sterling, L. and Shapiro, E. 1994. The Art of Prolog. MIT Press.
Sun
Microsystems,
Inc.
1997.
Secure
Computing
with
Java:
Now
and
the
Future
(Executive
Summary).
Available
at
<http://www.javasoft.com/marketing/collateral/security.html>.
Thorn, T. 1997. Programming Languages for Mobile Code. ACM Computing Surveys,
213 – 239.
Vitek, J. and Tschudin, C., Eds. 1997. Mobile Object Systems: Towards the Programmable Internet (Lecture Notes in Computer Science No. 1222). Springer-Verlag.
Volpano, D. 1997. Provably-Secure Programming Languages for Remote Evaluation. In Proceedings of Foundations for Secure Mobile Code Workshop. Available at
<http://www.cs.nps.navy.mil/research/languages/statements/necula.ps>.
| 6cs.PL
|
Controllability Gramian Spectra of Random Networks
arXiv:1609.04727v1 [cs.SY] 15 Sep 2016
Victor M. Preciado∗ & M. Amin Rahimian
Abstract—We propose a theoretical framework to study the
eigenvalue spectra of the controllability Gramian of systems with
random state matrices, such as networked systems with a random graph structure. Using random matrix theory, we provide
expressions for the moments of the eigenvalue distribution of
the controllability Gramian. These moments can then be used
to derive useful properties of the eigenvalue distribution of the
Gramian (in some cases, even closed-form expressions for the
distribution). We illustrate this framework by considering system
matrices derived from common random graph and matrix ensembles, such as the Wigner ensemble, the Gaussian Orthogonal
Ensemble (GOE), and random regular graphs. Subsequently, we
illustrate how the eigenvalue distribution of the Gramian can be
used to draw conclusions about the energy required to control
random system.
Index Terms—Controllability; Gramian spectrum; Complex
Networks; Random graphs; Random matrix theory.
I. I NTRODUCTION
Controllability is a classical tool in systems theory, as
developed by Kalman et al. [1]. More recently, controllability
of complex networked systems has attracted much attention,
beginning with the study of leader-follower multi-agent systems in [2], [3], [4], [5], [6], followed by several works
considering the network controllability problem in special
topologies such as paths [7], circulant networks [8], grids
[9], distance regular graphs [10], and networks-of-networks
[11], as well as works which employ the concept of structural
controllability in networked settings such as [12], [13], [14].
The concept of controllability is closely related to the
minimum energy needed to steer a linear system through
its state space [1], [15], [16]. In particular, the minimum
input energy needed to drive the system from an initial to
a desired state can be expressed in terms of the inverse of
the controllability Gramian matrix, which we denote by W
[17], [16]. This relationship has motivated several Gramianbased metrics to quantify the energy required to steer a system,
such as the trace of the Gramian inverse tr(W −1 ), the inverse
of the minimum eigenvalue 1/λmin (W ), or − log(det(W )),
[16]. These measures have found recent applications in the
analysis and design of networked control systems [18], [19],
[20]. These energy control metrics are closely related to the
Gramian eigenvalue spectrum, which is the central topic of
our work.
In this paper, we focus our attention on the eigenvalue
spectrum of the controllability Gramian of systems with random state matrices. In this case, the controllability Gramian
is also a random (symmetric) matrix with random (real)
eigenvalues following a probability distribution, called the
Gramian spectral distribution. The study of the eigenvalues
∗ The authors are with the Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104-6228 USA. (email:
[email protected]). This work was supported by the National
Science Foundation under grants CNS-1302222 and IIS-1447470.
of random matrices started in the pioneering work of Wigner
[21], [22], Mehta and Gaudin [23], and Dyson [24], and has
attracted a great deal of attention since then. Over the last
decades, random matrix theory has become a mature field and
found applications in a wide variety of disciplines [25], [26].
A particularly relevant application of random matrix theory
can be found in the study of the eigenvalues of random graph
models [27], [28], [29], [30]. In this direction, many graphrelated matrices—such as the adjacency, the Laplacian, and
the normalized Laplacian—have been widely studied in the
literature. In contrast, theoretical results about the eigenvalues
of the controllability/observability Gramians of random graphs
and matrices are very scarce.
The are only a few works that attempt to address some
aspects of this problem [31], [32], [33]. The author in [31]
introduces a stochastic notion of controllability for jump
systems, where the state and input matrices vary between N
designated matrix pairs and the variations follow a Markov
process. Accordingly, the linear jump system is controllable
if transitions between any two states occur with positive
probability and in an almost surely finite length of time;
variations of this definition when the transitions occur with
probability one and the required time-span has finite expectation are also considered, and an algebraic criterion for
stochastic controllability is expressed as a rank condition on
the N state and input matrix pairs. Parallel results are derived
in [32], but for the case where the state matrix is given
by the Laplacian of a random graph process. In [33] the
authors address related observability problems by investigating
the recovery of sparse initial states using independent and
randomly populated measurement matrices and under certain
conditions on the state transition matrix.
In this paper, we study the Gramian spectral distribution
of random systems, paying special attention to the Gramian
spectral moments (defined as the moments of the eigenvalue
distribution). In particular, we establish a connection between
the spectral moments of the (random) state matrix of a system and the spectral moments of its (random) controllability
Gramian. We then exploit this connection to derive closedform expressions for the Gramian spectral distribution of
important random matrices, such as Wigner matrices, the
Gaussian Orthogonal Ensemble (GOE), or random regular
graphs. From the Gramian spectral distribution, we then draw
conclusions about the energy required to steer a random
system and study how this energy is affected by the system
parameters and the dimension of the state space.
The paper is organized as follows. Notation and preliminary
results are provided in Section II. Our main results are stated in
Section III and proofs are included in Appendix A. Illustrative
examples and additional discussions are presented in Section
IV and the paper is concluded in Section V.
II. N OTATION & P RELIMINARY R ESULTS
Notation.: Throughout the paper, R is the set of real
numbers, boldface letters denote random variables, E{·} is the
expectation operator, vectors are denoted by a bar over their
respective lower-case letters, matrices are denoted by upper
case letters, T denotes matrix transpose and In is the n × n
identity matrix. We use ∼ to indicate asymptotic equivalence
(n)
= 1.
as n → ∞, i.e., f (n) ∼ g(n) indicates that limn→∞ fg(n)
A. The Method of Moments
Let us consider a random system characterized by the matrix
pair (An , Bn ), where An is a random n × n state matrix and
Bn is an input matrix. We investigate the limiting distribution
of the eigenvalues of the controllability Gramian associated
with the matrix pair (An , Bn ) as n → ∞. It is noted in [34],
[35] that the required control energy can increase unbounded
with the growing network size, unless Bn = In ; hence, we
restrict attention to the case Bn = In . We also assume that
An is symmetric1 . Given an n × n symmetric random matrix
C, let λ1 (C) ≤ λ2 (C) ≤ . . . ≤ λn (C) be n random variables
representing the n real eigenvalues of An . We
Pn consider the
random probability measure LC {·} = n1 i=1 δλi (C) {·},
δλi (C) being Dirac’s delta function centered at λi (C), as the
random probability measure on the real line that assigns mass
uniformly to each of the n eigenvalues of the random matrix
C. The corresponding distribution Fn (x) = LC {(−∞, x]}
= n1 |{i ∈ [n] : λi (C) ≤ x}|, is a random variable for
each x ∈ R and is referred to as the empirical spectral
distribution (ESD) for the random matrix C. The k-th spectral
moments of a random matrix C is the k-th moment of its
spectral
These moments can be written as mk (C) :=
Pn density.
1
k
λ
(C)
[25]. In our derivations, we also make use
i=1 i
n
of the centralized
Pn spectral moments, which are defined as
m
e k (C) := n1 i=1 (λi (C) − m1 (C))k .
In this paper, we will use the method of moments to
derive the limiting spectral distributions of the controllability
Gramians herein considered. According to this method, when
dealing with compactly supported distributions (as will be the
case for us), one can show that a sequence of empirical spectral
distributions Fn (·) converges in probability to some limiting
distribution F (·) by showing that
R +∞thek sequence of expected
moments for Fn (·), given by { 0 x dFn (x), k ∈ N}, converges point-wise (for
every k) to the corresponding moments
R +∞
of F (·), given by { 0 xk dF (x), k ∈ N} [26, Theorem 2.2.9
and Section 2.4.2].
In the following subsections, we describe two particular
random matrices that will be useful to illustrate our results.
B. Random Wigner Matrices: Semicircle Law
A Wigner matrix An is a random matrix where the entries
above the main diagonal are i.i.d. zero-mean with identical
variance σ 2 . In his seminal papers [21], [22], Eugene Wigner
1 More
general input and state matrices, such as random input matrices
and non-symmetric state matrices, are considered in an upcoming extended
version of this paper.
√
proved that the empirical spectral distribution of (1/ n)An
converges (in probability) to the semicircular density given by
1 p 2
4σ − x2 ,
(1)
fSC (x) =
2πσ 2
for x ∈ [−2σ, 2σ] and fSC (x) = 0, otherwise. The even-order
moment of the semi-circular distribution are given by mSC
2j =
σ 2j 2j
;
odd-order
moments
are
zero.
In
fact,
for
this
result
to
j+1 j
hold we do not need the variables to be identically distributed
and the result continues to hold for independent zero-mean
entires with common variance σ 2 [36], under certain mild
conditions on higher-order moments. An important special
case of Wigner’s random matrix is the Gaussian Orthogonal
Ensemble (GOE), in which the off-diagonal entries are real
Gaussian variables with mean zero and variance σ 2 . The
diagonal entries are real Gaussian variables with mean zero
and variance σ 2 /2. We will use this particular ensemble to
illustrate our results in Section IV.
C. Random Regular Graphs: McKay Law
Consider a random graph drawn uniformly over the space
of all undirected d-regular graphs on n vertices, where we
assume d > 2. As n → ∞, the empirical spectral distribution
of the adjacency matrix of a random d-regular graph converges
in probability to the McKay law [37], whose density is given
by
p
d 4 (d − 1) − x2
,
fM K (x) =
2π (d2 − x2 )
√
for |x| ≤ 2 d − 1, and fM K (x) = 0, otherwise. The moments
of the McKay distribution are given by
k/2
X
k k − 2r + 1
r
MK
(d − 1) ,
(2)
mk =
k
−
r
+
1
r
r=1
for k even, and 0 for k odd. In our derivations, we shall
also make use of the centralized spectral moments of the
Laplacian matrix. Let Cn be the (random) adjacency matrix
of the graph (i.e., [Cn ]ij = 1 if nodes i and j are connected,
and 0 otherwise). For a random d-regular graph, the Laplacian
matrix can be written as Ln = dIn −Cn . One can easily prove
that the centralized spectral moments of the Laplacian satisfy
K
E{m
e k (Ln )} ∼ mM
as n → ∞, in the case of random dk
regular graphs. This fact will be useful in future derivations.
III. M OMENT-BASED C HARACTERIZATION OF THE
G RAMIAN S PECTRUM
Consider a discrete-time (DT) linear time-invariant system
x [t + 1] = An x[t] + Bn u[t], where Bn = In , An is an n × n
random symmetric matrix, and u[t] ∈ Rn . Assuming An is
Schur stable and symmetric, the discrete-time controllability
Gramian is an n × n symmetric random matrix given by
Wd =
∞
X
τ =0
Aτn (ATn )τ = (In − A2n )−1 .
Similarly, for a continuous-time (CT) linear time-invariant
d
system described by dt
x(t) = An x(t) + u(t), when An
is an n × n Hurwitz stable, symmetric random matrix the
continuous-time controllability Gramian Wc is given by
Z ∞
Z ∞
T
e2An τ dτ
eAn τ eAn τ dτ =
Wc =
0
0
n
Z ∞
exp (2λi (An ) τ )dτ
VT , (3)
= Vdiag
0
i=1
where in last equality we use the fact that An is symmetric
and, therefore, orthogonally diagonalizable.
In the following subsections, for both random discretetime and continuous-time systems, we first provide theorems
to relate the spectral moments of the state matrix An to
the spectral moments of the controllability matrix. Since the
spectral moments of An are well-understood for a number of
random matrix ensembles, our result allows us to calculate the
spectral moments of the controllability Gramian. Finally, using
the method of moments, we will characterize and completely
determine the limiting spectral distribution of the Gramian
matrix for a number of well-known random matrix ensembles.
A. Gramian of Random Discrete-Time Systems
In the following lemma, we provide a useful relationship
between the moments of the state matrix An and the moments
of the associated controllability gramian Wd for random
discrete-time systems.
Lemma 1 (Spectral Moments of DT Gramian). Let An be
an n × n symmetric, Schur-stable state matrix with spectral
moments given by mk := mk (An ) for k ∈ N. Then,
the spectral moments of the controllability Gramian of the
discrete-time system (An , In ) are given by
∞
X
j+k−1
m2j .
(4)
mk (Wd ) =
k−1
j=0
We will prove the above lemma in Appendix A. Lemma 1
can be used to, for example, characterize the Gramian spectrum when the state matrix √
An is a Wigner random matrix.
In particular, let An = (1/ n)Hn where Hn is a Wigner
random matrix with the second moment satisfying σ < 1/2.
Then, explicit expressions for the spectral moments of the
Gramian and the Gramian spectrum are given in the following
theorem:
Theorem 1 (The Wigner Ensemble). Consider a DT system
with state matrix An = √1n Hn , where Hn is a Wigner random
matrix with independent zero-mean entries and common variance σ 2 < 1/4, and the identity input matrix, Bn = In . Then,
the expected spectral moments of the controllability Gramian
Wd are asymptotically given by
1
mk (Wd ) ∼ 2 F1 ( , k; 2, 4σ 2 ),
2
(5)
where 2 F1 (a, b; c; z) is the hyper-geometric function2 . Furthermore, the limiting spectral distribution of Wd is characterized by the following density function
(
q
1
1
4σ 2 x−x+1
, if x ∈ [1, 1−4σ
2 x2
2 ].
2πσ
x−1
(6)
fW (x) =
0,
otherwise.
While proving Theorem 1 in Appendix B, we provide a
series of tools that can be directly applied to many other
random matrix ensembles, as soon as they are Hurwitz stable.
In the particular case treated in Theorem 1, the assumption
σ 2 < 1/4 guarantees the matrix ensemble An to be Hurwitz
stable. We now develop parallel results for random systems in
the continuous-time case.
B. Gramian of Random Continuous-Time Systems
Similarly to Subsection III-A, we first provide a lemma
(proved in Appendix C) that allows us to calculate the asymptotic spectral moments of the Gramian from those of the state
matrix An . We then apply this result to study the Gramian of
a system whose state matrix is related to the Laplacian of a
random regular graph.
Lemma 2 (Spectral Moments of the CT Gramian). Let
An be an n × n symmetric, Schur-stable state matrix with
spectral moments given by mk := mk (An ) and centralized
e k := m
spectral moments given by m
e k (An ) for k ∈ N. Then,
the spectral moments of the continuous-time controllability
Gramian Wc are given by:
∞
ej
1 X −k
m
mk (Wc ) =
.
(7)
(−2)k j=0 j (m1 )k+j
In what follows, we illustrate the usage of Lemma 2
by considering the Laplacian matrix of a random d-regular
graph, described in Subsection II-C. Since the Laplacian
matrix is marginally stable (has an eigenvalue at zero), the
Gramian of the continuous-time system (−Ln , In ) is illdefined. To overcome this issue, we use a stabilized version
of the Laplacian dynamics, where the state matrix is given
e n := −Ln − d 11T . Notice that this matrix
by An = −L
n
is not singular for connected graphs; thus, the Gramian of
e n , In ) is well-defined and satisfies (3). The
the system (−L
following theorem, proved in Appendix D, provides explicit
expressions for the spectral moments of the Gramian and the
Gramian spectrum.
Theorem 2 (Random Regular Graphs). Consider a CT system
e n , where
with input matrix In and state matrix An = −L
e
Ln is the stabilized Laplacian matrix of a random d-regular
graph with n nodes and d ≥ 3. Then, the asymptotic expected
spectral moments of the controllability Gramian Wc are given
2 The
2 F1
hyper-geometric function is defined for |z| < 1 by the power series
P
(q)j (b)j z j
(q, b; c; z) = ∞
, where (q)j = q(q + 1) . . . (q + j − 1).
j=0
(c)
j!
j
by
∞
1 X 2j + k − 1 −2j M K
mk (Wc ) ∼
d m2j ,
(2d)k j=0
2j
(8)
K
where mM
are the spectral moments of the McKay law,
2j
defined in (2). Furthermore, the limiting spectral density of
Wc is given by
q
1 2
d
4 (d − 1) − d − 2x
1
,
(9)
fL (x) =
1 2
4πx2
d2 − d − 2x
h
i
for x ∈ 2d+41√d−1 , 2d−41√d−1 , and fL (x) = 0, otherwise.
IV. E XAMPLES
To illustrate the applicability of our results we consider two
special cases where the state matrices are given by well-known
random matrix and graph models.
Example 1. Gaussian-weighted Erdős-Rényi Ensemble
To illustrate our results, we consider a state matrix with the
sparsity structure of an Erdős-Rényi random graph with edge
probability pn . Additionally, we assume that the weights of
the edges in the random graph are also random. In particular,
we assume the edge weights are Gaussian variables with
2
zero mean and variance σn2 = 4pαn n for any 0 < α < 1.3
We call this random matrix ensemble the Gaussian-weighted
Erdős-Rényi Ensemble (GER) with parameters (n, pn , σn2 ).
Let An be an instance from this ensemble. The variance
of the entries of An are pn σn2 = α2 /4n, and they have
zero mean. According to the semicircle law, as n → ∞
the eigenvalue distribution is supported in the range [−α, α];
thus, An is Schur stable. The limiting spectral density of
GER (n,√pn , σn2 ) can be derived from (1) as fGER (x) =
(2/πα2 ) α2 − x2 , for x ∈ [−α, α] and 0 otherwise. Fig. 1
shows the histogram of eigenvalues of An from a Gaussianweighted Erdős-Rényi Ensemble with parameters n = 1000,
pn = 2 log n/n = 0.006, and α = 0.5.
Similarly, the limiting
moments of GER are
spectral
2k
1
GER
= ( α2 )2k k+1
mGER
= 0 for j odd. The
2k
k , and mj
−1
trace of the Gramian inverse, tr(Wd ), is a Gramian-based
control energy metric that quantifies the average steering
energy along all directions in the state space. Accordingly,
for the controllability Gramian of a random DT system with
a state matrix drawn from the GER (n, pn , σn2 ), we have
that asymptotically E{tr(Wd−1 )} ∼ limn→∞ E{tr(In − A2n )}
= limn→∞ (n − nE {m2 (An )}). For the particular case of
GER, we have that E{m2 (An )} ∼ mGER
= α2 /4 so that
2
−1
2
E{tr(Wd )} ∼ n(1 − α /4), and the average control energy
increases linearly with the increasing size n. Notice that the
control energy decreases as α, which is related to the variance
of the edge weights, increases. Fig. 2 shows the histogram of
eigenvalues of controllability Gramian for An using the same
parameters as in Fig. 1.
3 In fact, we can use any distribution (not necessarily the normal density)
2.
that has zero mean, with variance σn
Fig. 1. Eigenvalue Histogram of a Gaussian-weighted
Erdős-Rényi ensemble
√
with parameters n = 1000, pn = 2 log n/ n, and α = 0.5 is given by a
semi-circular density that is supported over [−0.5, 0.5].
Fig. 2. Eigenvalue Histogram of the Gramian of a discrete-time system with
the state-matrix An given by a Gaussian-weighted Erdős-Rényi ensemble
with the same parameters as in Fig. 1. The Gramian spectrum in this case
follows the limiting spectral density fW given in Theorem 1.
Example 2. Random Regular Graph Laplacian Dynamics
As a second example, consider the stabilized Laplacian dynamics on a random d-regular graph considered in Theorem 2.
In Fig. 3 we have plotted the histogram of the eigenvalues of
the stabilized Laplacian L̃n with d = 3 and n = 1000. The
eigenvalue spectrum of the associated controllability Gramian
is plotted in Fig. 4. Here, we use 1/λmin (Wc ) as a measure of
the worst-case minimum required control energy for making
a unit transfer in the state space.
√ Applying Theorem 2 we
get that 1/λmin (Wc ) ∼ 2d + 4 d − 1; hence, the minimum
required energy increases with the increasing degree.
about the energy required to control a random system.
A PPENDIX
P ROOFS OF THE M AIN R ESULTS
A. Proof of Lemma 1
e n := Ln + d 11T
Fig. 3. Eigenvalue Histogram of the stabilized Laplacian L
n
with d = 3 and n = 1000. The limiting spectral distribution in this case is
given by the McKay law (Subsection II-C), shifted so that the eigenvalue
spectrum is centered around d = 3.
Given the input matrix Bn = In and a symmetric state
T
matrix
n with spectral moments given by mk :=
Pn An = A
1
k
λ
(A
)
= n1 tr(Akn ), we P
can write the spectral
i
n
i=1
n
∞
moments of the DT Gramian Wd := τ =0 A2τ
n , as follows:
!k
∞
X
1
1
mk (Wd ) := tr(Wdk ) = tr
A2τ
n
n
n
τ =0
=
1
tr (I − A2n )−k .
n
2j
P∞
From the Taylor expansion (I −A2n )−k = j=0 j+k−1
k−1 An ,
we have that:
∞
1 X j + k − 1
mk (Wd ) = tr
A2j
n
n
k
−
1
j=0
∞
X
j+k−1 1
tr(A2j
=
n )
n
k
−
1
j=0
∞
X
j+k−1
=
m2j ,
k−1
j=0
as stated in Lemma 1.
B. Proof of Theorem 1
Fig. 4. Eigenvalue Histogram of the Gramian of the stabilized Laplacian
dynamics on a random d-regular graph with d = 3 and n = 1000. The
limiting spectral distribution of the Gramian in this case has the density
function fL given in Theorem 2.
V. C ONCLUSIONS
In this paper, we have introduced a theoretical framework
to study the eigenvalue spectra of the controllability Gramian
of systems with random state matrices. Using tools from
random matrix theory, such as the method of moments, we
have derived expressions for the spectral moments of the
controllability Gramian for both continuous- and discrete-time
systems. We have used these moments to derive closed-form
expressions for the spectral distribution of the Gramian of random systems derived from popular random matrix and graph
ensembles. Finally, we have illustrated how the eigenvalue
distribution of the Gramian can be used to draw conclusions
We begin by showing that the spectral moments of the
Gramian are asymptotically given by (5). Starting form (4)
and replacing the moments mk of the Wigner ensemble, we
obtain
∞
X
1
j + k − 1 2j 1
2j
k
E tr[Wd ] =
σ
n
k−1
j+1 j
j=0
∞
X j+k−1
2−2j (2j)!
=
(4σ 2 )j
. (10)
k−1
j!(j + 1)!
j=0
Next, we can write (2j)! = (2j )(j!)(3 × 5 × 7 × . . . (2j − 1))
so that
2−2j (2j)!
2−j (3 × 5 × 7 × . . . (2j − 1))
=
j!(j + 1)!
(j + 1)!
(1/2)(1/2 + 1)(1/2 + 2)(1/2 + 3) . . . (1/2 + j − 1)
=
2(2 + 1)(3 + 1) . . . (2 + j − 1)
(1/2)j
=
,
(11)
(2)j
where (q)j = q(q + 1) . . . (q + j − 1). Replacing (11) in (10)
yields
∞
X
j + k − 1 (1/2)j
1
k
E tr[W ] =
(4σ 2 )j
n
(2)
k
−
1
j
j=0
=
∞
X
(k)j (1/2)j
j=0
(j!)(2)j
(4σ 2 )j
1
= 2 F1 ( , k; 2, 4σ 2 ),
2
when 4σ 2 < 1, as we wanted to prove.
We now proceed to determine the limiting spectral density
of the Gramian when An is a random matrix from the Wigner
ensemble. We begin by noting two identities, which hold true
for hyper-geometric functions
3
4σ 2
1
2
2 −k
), (12)
2 F1 (k, ; 2,
2 F1 ( , k; 2, 4σ ) = (1−4σ )
2
2
4σ 2 − 1
Z 1 b−1
x (1 − x)c−b−1
−1
dx,
2 F1 (a, b; c; z) = B(b, c − b)
(1 − zx)a
0
(13)
for c > b > 0 and |z| < 1, where B(·, ·) is the Beta function.4
The first formula follows from Kummer solutions to Gaussian
hyper-geometric differential equation, after a proper change
of variables [38, Section 3.7]. The second one is an integral
representation of hyper-geometric functions known as Euler
formula [38, Section 3.6]. We can use (13) to rewrite the lefthand side of (12) as follows
1
2
2 F1 ( , k; 2, 4σ )
2
Z
−k 2 1 1
1
4σ 2
= 1 − 4σ 2
x 2 (1 − x)− 2 (1 − 2
x)−k dx
π 0
4σ − 1
Z
1
2 1 1
=
x 2 (1 − x)− 2 1 − 4σ 2 + 4σ 2 x −k dx
π 0
s
s
Z1
1
4yσ 2 − y + 1 4yσ 2 k dy
=−
y
2πσ 2
4yσ 2
y − 1 y2
1/(1−4σ 2 )
=
=
2
1/(1−4σ
)
Z
Z
R
1
2πσ 2 y 2
1
s
4yσ 2 − y + 1 k
y dy
y−1
y k fW (y) dy,
where we have used the fact that B( 32 , 12 ) = π/2, together with
2
−y+1
,
the change of variables y −1 = 1−4σ 2 +4σ 2 x, x = 4yσ4yσ
2
d
y
and dx = −
for the integration. Hence, the moments of
4σ 2 y 2
the spectral distribution in (6) are indeed given by the spectral
moments in (5). Therefore, the claim in our theorem follows
from the method of moments.
4 The
Beta function is defined in terms of the Gamma function as B(x, y) =
Γ(c)
.
Γ(b)Γ(c−b)
C. Proof of Lemma 2
Since the CT Gramian Wc satisfy (3), its spectral moments
are given by
"Z
k #
∞
1
2An τ
e
dτ
mk (Wc ) = tr
n
0
k
n Z ∞
1X
e2λi (An )τ dτ
,
=
n i=1
0
R∞
Since 0 e2λi (An )τ dτ = −1/(2λi (An )), we have that,
mk (Wc ) =
n
X
1
(λi (An ))−k .
n(−2)k i=1
(14)
The Taylor expansion of x−k around m1 is given by
∞
X
−k
x−k =
(m1 )−k−j (x − m1 )j
j
j=0
We can now replace for (λi (An ))−k in (14) to get
n X
∞
j
X
1
−k (λi (An ) − m1 )
mk (Wc ) =
n(−2)k i=1 j=0 j
(m1 )k+j
Pn
∞
j
1 X −k
i=1 (λi (An ) − m1 )
=
k
k+j
(−2) j=0 j
n(m1 )
∞
ej
m
1 X −k
=
,
(−2)k j=0 j (m1 )k+j
as claimed in Lemma 2.
D. Proof of Theorem 2
e n , we can replace for the moGiven that An = −L
ments of An in (7). Hence, using the asymptotic moments
K
e n ) ∼ −d and m
e j = mM
m1 = n1 Tr(−L
j , as discussed in
Subsection II-C, (7) becomes
∞
1 X −k
−k−j
K
E{mk (Wc )} ∼
(−d)
mM
j
k
j
(−2) j=0
∞
1 X 2j + k − 1 1 M K
=
m
k
2j
d2j 2j
(2d) j=0
where in the last equality, we have used j+k−1
(−1)j = −k
j
j
K
and substituted j by 2j to account for the fact that mM
=0
j
for j odd. This proves the expression claimed in (8) for the
asymptotic spectral moments of the CT Gramian.
Once we have a closed-form expression for the expected
spectral moments of the Gramian, we need to prove that these
moments correspond to those of the density in (9). We begin
by calculating the moments of (9) as follows
Z
xk fL (x)dx
R
=
=
√
1/(2d−4
Z d−1)
xk d
1 2
2x
dx
1 2
2x )
4 (d − 1) − d −
4πx2 d2 − (d −
√
1/(2d+4 d−1)
√
2 Zd−1
√
−2 d−1
q
p
d 4 (d − 1) − u2
du,
(d − u)k (2π)2k (d2 − u2 )
(15)
where in the last equality we have used the change of variables
u = d − (1/2x). We now use the Taylor expansion of
−k
(d − u) around d, which is given by
∞
X
j + k − 1 uj
−k
.
(d − u) =
dk+j
j
j=0
Using this expansion in (15), we get that
Z
∞
1 X j + k − 1 −k−j
k
x fL (x)dx = k
d
× ...
2 j=0
j
R
√
2 Zd−1
p
uj d 4 (d − 1) − u2
du
2π(d2 − u2 )
√
−2 d−1
∞
X
2j + k − 1 −k−2j M K
d
m2j
2j
j=0
∞
1 X 2j + k − 1 −2j
=
d
× ...
k
2j
(2d) j=0
j
X
2j 2j − 2r + 1
r
(d − 1)
r
2j
−
r
+
1
r=1
a 1
= k
2
= E{mk (Wc )},
a
K
where in (=) we have used the fact that mM
= 0 for j.
j
Having thus shown that the moments of fL (·) coincides with
those of the Gramian spectrum the proof follows from the
method of moments, described in Section II.
R EFERENCES
[1] R. Kalman, Y. C. Ho, and K. S. Narendra, “Controllability of linear
dynamical systems,” Contrib. Differential Equations, vol. 1, pp. 189–
213, 1962.
[2] H. Tanner, “On the controllability of nearest neighbor interconnections,”
in Proceedings of the 43rd IEEE Conference on Decision and Control,
Dec. 2004, pp. 2467–2472.
[3] A. Rahmani, M. Ji, M. Mesbahi, and M. Egerstedt, “Controllability of
multi-agent systems from a graph-theoretic perspective,” SIAM Journal
on Control and Optimization, vol. 48, no. 1, pp. 162–186, 2009.
[4] S. Martini, M. Egerstedt, and A. Bicchi, “Controllability analysis of
multi-agent systems using relaxed equitable partitions,” Int. J. Systems,
Control and Communications, vol. 2, no. 1,2,3, pp. 100–121, 2010.
[5] Z. Ji, Z. Wang, H. Lin, and Z. Wang, “Interconnection topologies for
multi-agent coordination under leaderfollower framework,” Automatica,
vol. 45, no. 12, pp. 2857–2863, 2009.
[6] B. Liu, T. Chu, L. Wang, and G. Xie, “Controllability of a leaderfollower dynamic network with switching topology,” IEEE Transactions
on Automatic Control, vol. 53, no. 4, pp. 1009–1013, 2008.
[7] G. Parlangeli and G. Notarstefano, “On the reachability and observability
of path and cycle graphs,” IEEE Transactions on Automatic Control,
vol. 57, no. 3, pp. 743–748, 2012.
[8] M. Nabi-Abdolyousefi and M. Mesbahi, “On the controllability properties of circulant networks,” IEEE Transactions on Automatic Control,
vol. 58, no. 12, pp. 3179–3184, 2013.
[9] G. Notarstefano and G. Parlangeli, “Controllability and observability
of grid graphs via reduction and symmetries,” IEEE Transactions on
Automatic Control, vol. 58, no. 7, pp. 1719–1731, 2013.
[10] S. Zhang, M. K. Camlibel, and M. Cao, “Controllability of diffusivelycoupled multi-agent systems with general and distance regular coupling
topologies,” in 50th IEEE Conference on Decision and Control and
European Control Conference (CDC-ECC), 2011, pp. 759–764.
[11] A. Chapman, M. Nabi-Abdolyousefi, and M. Mesbahi, “Controllability
and observability of network-of-networks via cartesian products,” Automatic Control, IEEE Transactions on, vol. 59, no. 10, pp. 2668–2679,
Oct 2014.
[12] Y.-Y. Liu, J.-J. Slotine, and A.-L. Barabási, “Controllability of complex
networks,” Nature, vol. 473, no. 7346, pp. 167–173, 2011.
[13] M. A. Rahimian and A. G. Aghdam, “Structural controllability of multiagent networks: Robustness against simultaneous failures,” Automatica,
vol. 49, no. 11, pp. 3149–3157, 2013.
[14] A. Chapman and M. Mesbahi, “On strong structural controllability of
networked systems: a constrained matching approach,” in American
Control Conference (ACC), 2013, pp. 6126–6131.
[15] C. Johnson, “Optimization of a certain quality of complete controllability
and observability for linear dynamical systems,” Journal of Fluids
Engineering, vol. 91, no. 2, pp. 228–237, 1969.
[16] P. Muller and H. Weber, “Analysis and optimization of certain qualities of controllability and observability for linear dynamical systems,”
Automatica, vol. 8, no. 3, pp. 237 – 246, 1972.
[17] W. Rugh, Linear System Theory. Prentice Hall, 1993.
[18] T. H. Summers, F. L. Cortesi, and J. Lygeros, “On submodularity and
controllability in complex dynamical networks,” ArXiv e-prints, Apr.
2014.
[19] F. Pasqualetti, S. Zampieri, and F. Bullo, “Controllability metrics,
limitations and algorithms for complex networks,” Control of Network
Systems, IEEE Transactions on, vol. 1, no. 1, pp. 40–52, March 2014.
[20] V. Tzoumas, M. A. Rahimian, G. J. Pappas, and A. Jadbabaie, “Minimal
actuator placement with bounds on control effort,” IEEE Transactions
on Control of Network Systems, 2015, in press.
[21] E. Wigner, “Characteristic vectors of bordered matrices with infinite
dimensions,” Ann. of Math., vol. 62, pp. 548–564, 1955.
[22] ——, “On the distribution of the roots of certain symmetric matrices,”
Annals of Mathematics, vol. 67, no. 2, pp. pp. 325–327, 1958.
[23] M. Mehta and M. Gaudin, “On the density of eigenvalues of a random
matrix,” Nuclear Physics, vol. 18, pp. 420–427, 1960.
[24] F. J. Dyson, “A brownian-motion model for the eigenvalues of a random
matrix,” Journal of Mathematical Physics, vol. 3, no. 6, pp. 1191–1198,
1962.
[25] G. Anderson, A. Guionnet, and O. Zeitouni, An Introduction to Random
Matrices, 1st ed. Cambridge: Cambridge University Press, 2009.
[26] T. Tao, Topics in random matrix theory. Graduate Studies in Mathematics, American Mathematical Society, 2012.
[27] F. Chung, L. Lu, and V. Vu, “Spectra of random graphs with given
expected degrees,” Proceedings of the National Academy of Sciences,
vol. 100, no. 11, pp. 6313–6318, 2003.
[28] F. Chung and M. Radcliffe, “On the spectra of general random graphs,”
the Electronic Journal of Combinatorics, vol. 18, no. 1, p. P215, 2011.
[29] R. R. Nadakuditi and M. E. Newman, “Spectra of random graphs with
arbitrary expected degrees,” Physical Review E, vol. 87, no. 1, p. 012803,
2013.
[30] L. Erdős, A. Knowles, H.-T. Yau, and J. Yin, “Spectral statistics of
Erdős–Rényi graphs I: Local semicircle law,” The Annals of Probability,
vol. 41, no. 3B, pp. 2279–2375, 2013.
[31] M. Mariton, “Stochastic controllability of linear systems with markovian
jumps,” Automatica, vol. 23, no. 6, pp. 783–785, 1987.
[32] M. Nabi-Abdolyousefi and M. Mesbahi, “System theory over random
consensus networks: controllability and optimality properties,” in IEEE
Conference on Decision and Control and European Control Conference
(CDC-ECC), 2011, pp. 2323–2328.
[33] B. M. Sanandaji, M. B. Wakin, and T. L. Vincent, “Observability with
random observations,” IEEE Transactions on Automatic Control, vol. 59,
no. 11, pp. 3002–3007, 2014.
[34] C. Enyioha, M. A. Rahimian, G. J. Pappas, and A. Jadbabaie, “Controllability and fraction of leaders in infinite networks,” in IEEE 53rd Annual
Conference on Decision and Control (CDC), Dec 2014, pp. 1359–1364.
[35] G. Yan, G. Tsekenis, B. Barzel, J.-J. Slotine, Y.-Y. Liu, and A.-L.
Barabasi, “Spectrum of controlling and observing complex networks,”
Nature Physics, vol. 11, no. 9, pp. 779–786, Oct 2014.
[36] T. Tao and V. Vu, “Random matrices: The universality phenomenon for
wigner ensembles,” arXiv preprint, 2012.
[37] B. D. McKay, “The expected eigenvalue distribution of a large regular
graph,” Linear Algebra and its Applications, vol. 40, pp. 203–216, 1981.
[38] Y. L. Luke, The special functions and their approximations. Academic
press, 1969.
| 3cs.SY
|
1
arXiv:1711.03910v1 [cs.IT] 10 Nov 2017
LDPC-Based Code Hopping for Gaussian Wiretap
Channel With Limited Feedback
Zhao Chen1,3, Liuguo Yin2,3 , and Jianhua Lu1
State Key Laboratory on Microwave and Digital Communications
Tsinghua National Laboratory for Information Science and Technology
1 Department of Electronic Engineering, Tsinghua University
2 School of Aerospace, Tsinghua University
3 EDA Lab, Research Institute of Tsinghua University in Shenzhen
Email: [email protected]; {yinlg, lhh-dee}@tsinghua.edu.cn
Abstract—This paper presents a scheme named code hopping
(CodeHop) for gaussian wiretap channels based on nonsystematic
low-density parity-check (LDPC) codes. Different from traditional communications, in the CodeHop scheme, the legitimate
receiver (Bob) rapidly switches the parity-check matrix upon
each correctly received source message block. Given an authenticated public feedback channel, the transmitter’s (Alice) paritycheck matrix can also be synchronized with the receiver’s. As
a result, once an eavesdropper (Eve) erroneously decodes a
message block, she may not be able to follow the update of
subsequent parity-check matrices. Thus, the average BER of Eve
will be very close to 0.5 if the transmitted number of message
blocks is large enough. Focused on the measure of security gap
defined as the difference of channel quality between Bob and Eve,
numerical results show that the CodeHop scheme outperforms
other solutions by sufficiently reducing the security gap without
sacrificing the error-correcting performance of Bob.
I. I NTRODUCTION
Traditionally, security of data transmissions has been handled at upper protocol layers, which presents wiretappers (Eve)
by cryptographic algorithms with a secret key privately shared
between a transmitter (Alice) and legitimate receivers (Bob).
This notion of security relies on the assumption that Eve has
limited resources. Thus, perfect secrecy was introduced by
Shannon [1], which proved that if the secret key rate is lager or
equal to the transmission rate, information-theoretical security
can be achieved such that the encrypted message reveals no
information of the source message.
Owing to the stochastic nature of communication channels,
the work of Shannon was extended to the model of wiretap
channel by Wyner [2], where information-theoretically secure
communication is possible when Eve’s channel condition is
worse than Bob’s, and secrecy capacity is measured as the
highest transmission rate that can be achieved by Bob without
any information leaked to Eve. Hereafter, Wyner’s work was
refined by Csiszar and Korner in [3], and then generalized to
gaussian wiretap channel in [4]. Note that the equivocation rate
at Eve can also be a metric for security, which is defined as
the conditional entropy of the source message with given Eve’s
noisy observation. More recently, scenarios including fading
channels [5], MIMO channels [6] and multi-user channels [7]
have also been investigated in the literature.
Recently, coding techniques attract much research interests
to achieve better security performance. In [8], for binary
erasure channel (BEC) and binary symmetric channel (BSC),
low-density parity-check (LDPC)-based wiretap codes are
constructed. For gaussian wiretap channels, lattice codes are
also developed in [9]. Moreover, with the emerging polar
codes techniques, a nested code structure is established for
binary-input symmetric-output (BISO) channels in [10]. All
the designs above are for asymptotically large block lengths,
which may be difficult to implement in practical systems. To
this end, an LDPC code design for BPSK-constrained gaussian
wiretap channel was shown in [11]. A joint error correction
and encryption using LDPC codes is also studied in [12].
Recently, in order to construct practical codes, the bit error
rate (BER) of message bits is measured for physical layer
security [13]. Specifically, if Eve’s BER is ensured to be close
to 0.5 and the error bits are i.i.d. distributed, she would be
unable to recover any information from the decoded messages.
Then, security gap was introduced by Klinc [14], which is
defined as the difference of the channel quality between Bob
and Eve, requiring Eve’s BER to be sufficiently secure, while
keeping Bob’s BER low enough to reliably receive messages.
In [15], Baldi reduced the security gap by implementing
nonsystematic LDPC codes using scrambling matrices, where
security is enhanced by considering automatic repeat-request
(ARQ). Also, a feedback-aided secure transmission scheme
using stopping sets of LDPC codes can be found in [16].
In this paper, the gaussian wiretap channel enhanced with
an authenticated feedback channel is considered, where an
LDPC-based code hopping (CodeHop) scheme is proposed
to improve physical layer security by reducing the security
gap. Using a structured-random LDPC code generator, the
parity-check matrix of CodeHop will be rapidly updated upon
a correctly received message block by Bob, which is then
notified to Alice with one bit feedback. Once Eve encounters
a decoding error, she may not be able to decode any more
message block without knowing the parity-check matrix in the
following transmission and then her average BER will be close
to 0.5. Thus, the security gap can be significantly reduced.
The rest of the paper is organized as follows. In Section II,
system model of the CodeHop scheme is introduced. Then, the
construction of the structured-random LDPC codes is shown
2
fi
mi−1
Feedback Channel
mi
Encoder
xi
Main
Channel
yi
Decoder
Transmitter (Alice)
ˆi
m
z −1
Receiver (Bob)
Wiretap
Channel
zi
Decoder
Parity-Check Matrix
Generator
mi
f i −1
Hi
xi
LDPC Encoder
mi
(a) Encoder.
Eavesdropper (Eve)
Fig. 1. System model of the gaussian wiretap channel with limited feedback.
fi
Integrity Check
z −1
in Section III. In Section IV, the security gap performance is
analyzed. Finally, we conclude this paper in Section V.
f i −1
II. T HE LDPC-BASED C ODE H OPPING S CHEME
In this section, system model of the gaussian wiretap
channel with limited feedback will be introduced, along with
the design of the encoder and the decoder.
ˆ i−1
m
Parity-Check Matrix
Generator
z −1
Hi
yi
LDPC Decoder
ˆi
m
(b) Decoder.
A. System Model
As depicted in Fig. 1, for i = 1, 2, . . ., the secret source
to be transmitted is divided into k-bit message blocks mi ,
which are encoded into n-bit codewords xi by the encoder at
Alice and then transmitted over an AWGN channel to Bob.
In the meantime, codewords are also perceived by Eve over a
noisier and independent AWGN channel. For Bob and Eve,
the received codewords are denoted by yi and zi , which
are recovered by the decoders as m̂i and m̄i , respectively.
Additionally, in our model, legitimate receivers are always
given a public feedback channel, which can be used to inform
Alice whether the current codeword is correctly decoded with
just one bit feedback fi .
Let PeB and PeE define the receiving BERs of Bob and
Eve, respectively. To guarantee both information reliability and
security, it is desired to constrain PeB to be lower than a fixed
B
threshold Pe,max
(≈ 0) and PeE to be greater than a threshold
E
Pe,min (≈ 0.5). If the receiving SNRs at Bob and Eve are
denoted by SNRB and SNRE , it must hold that
(
B
SNRB ≥ SNRB,min = p−1 (Pe,max
),
(1)
E
SNRE ≤ SNRE,max = p−1 (Pe,min
),
(2)
where p(·) denotes the BER as a function of the SNR, and the
inverse p−1 (·) gives the target SNR for a given BER threshold.
Then, the security gap is defined as
Sg = SNRB,min − SNRE,max ,
(3)
where the SNRs are expressed in decibels (dB).
If both (1) and (2) are satisfied, the secrecy rate Rs can be
defined as the efficient transmission rate
k
Rs ≈ R = ,
(4)
n
where in fact Rs is slightly less than R due to some overhead
bits in mi for integrity check. Note that the security gap Sg
measures the minimum required difference between Bob and
Eve’s SNR in dB for secure communication, and our target is
to reduce the security gap as small as possible.
Fig. 2. Block diagrams of the encoder and the decoder. Note that the z −1
block denotes a delay unit.
B. Encoder and Decoder Design
In order to minimize the security gap of gaussian wiretap
channel, the CodeHop scheme is implemented such that each
source message block mi will be encoded by a rapidly hopping
parity-check matrix of LDPC codes Hi . That is, Hi is not
fixed for all codewords. On the contrary, Hi will be generated
in real time according to the past correctly received source
message block mi−1 by a parity-check matrix generator. Block
diagrams of the encoder and the decoder are both illustrated
in Fig. 2.
In the encoder of Alice, starting from a public agreed paritycheck matrix H0 and feedback f0 = 0, the LDPC parity-check
matrix Hi is generated by
(
H(hash(mi−1 )), if fi−1 = 1;
Hi =
(5)
Hi−1 ,
if fi−1 = 0;
where hash(·) is a hash function, and H(·) is the mapping
of structured-random LDPC codes described in Section III.
Thus, the parity-check matrix Hi will be updated when the
past source message mi−1 is successfully recovered by Bob,
i.e. fi−1 = 1. Otherwise, Hi will be kept the same as Hi−1 .
In the decoder of Bob, the public feedback fi is the result of
integrity check for m̂i . Although the correctness of mi can be
determined by the syndrome of the decoded codeword in most
cases, it is still possible that there exist some undetected errors
in m̂i when the decoded codeword converges to another valid
codeword of Hi . Thus, integrity check, e.g. cyclic redundancy
check (CRC), is still needed to avoid such undetected errors,
and some overhead bits are required in mi to store the integrity
check value. As for the parity-check matrix Hi , it will be
generated as in the encoder of Alice.
3
c2
c4
v2
v6
v4
v1
v5
c1
v3
c3
Fig. 3. Protograph P optimized for the rate-1/2 nonsystematic LDPC code.
III. S TRUCTURED -R ANDOM LDPC C ODES
In this section, we will show how to generate nonsystematic
LDPC codes in real time via structured-random protograph
expanding. As presented by Thorpe in [17], protograph can
be regarded as the minimal base Tanner graph to describe an
LDPC code. Using protograph, our target is to puncture all of
the k information bits and transmit only the n parity bits in
the original (n + k)-bit codeword. In order to avoid stopping
sets which make the iterative decoding of nonsystematic codes
difficult to converge, the code doping method [18] is adopted.
As for a rate-1/2 nonsystematic LDPC code, the result of
an optimized protograph P = (V, C, E) is given in Fig. 3,
which is equivalent to a 4 × 6 base parity-check matrix HB,0
as shown below,
1 1 0 0 1 0
1 1 0 0 0 1
HB,0 =
(6)
3 1 1 1 0 0 .
0 1 1 2 0 0
In the figure, there are a variable node set V = {v1 , v2 , ..., v6 },
a check node set C = {c1 , c2 , ..., c4 } and an edge set
E = {e1 , e2 , ..., e16 } in the protograph P . Among all variable
nodes, the information nodes denoted by v1 and v2 will be
punctured. Thus, according to the requirement of code doping,
the check node c4 is designated to be connected to only one
punctured variable node v2 to trigger the convergence of the
iterative decoding.
Starting with the protograph P , a larger tanner graph can
be obtained by copy-and-permute operation. Let each edge
e ∈ E stand for a unique edge type. If expanded with a
factor of T , the copy-and-permute operation firstly replicates
the protograph P for T times and forms a set of T edge
copies for each edge type, whose endpoints are then permuted
among the variable and check nodes in the set. Thus the T
copies of the protograph are all interconnected in the derived
graph, which defines the Tanner graph for the derived code.
Note that the operation of edge expanding is equivalent to
matrix expanding for each element in the base matrix HB,0 .
Each element of value w will be expanded to a T × T matrix
with w ones in each row or column, and thus the equivalent
parity-check matrix can be of size 4T × 6T .
Instead of using random permutations on the set of each
edge type, permutations defined by algebraic structures such
as cyclic shift are preferred for the ease of description and
efficient implementation. That is to say, the protograph can be
expanded by choosing appropriate T ×T circulant permutation
matrices IT (t) for each edge type, and the derived parity-check
matrix will become a T -circulant matrix. Note that each row
in IT (t) can be obtained by one bit cyclic right shifting its
previous row, and thus IT (t) is defined by its first row uT (t),
where uT (t) = (0, . . . , 0, 1, 0, . . . , 0) and the shift value t ∈
J0, T −1K 1 denotes the position of the unique one in the vector.
Therefore, given the protograph P , the derived LDPC code can
be described by shift values of all the circulant permutations.
Usually searching for only one good code is satisfactory
for channel coding. However, now it is far more challenging
to construct a large amount of highly efficient LDPC codes
for the CodeHop scheme. Thus, expanding with just one
single stage is not enough. A multi-stage expanding scheme
is developed, which we referred to as a structured-random
protograph expanding technique. Specifically, the protograph
P will be expanded via L > 1 stages such that the total
expanding factor T = T1 T2 · · · TL , where Tl is the expanding
factor of the lth stage for l ∈ J1, LK and the equivalent
expanded parity-check matrix is denoted by HB,l .
• Structured expanding: In the first L − 1 stages, the protograph is carefully expanded to avoid short loops, lowweight codewords or parallel edges, which can restrict
the general structure of the code. After L − 1 stages, all
the non-zero elements in HB,L−1 will equal 1.
• Random expanding: In the Lth stage, all of the edges in
the tanner graph will be randomly expanded, or equivalently, all the non-zero elements in HB,L−1 will be
expanded to circulant permutations according to the hash
value of the past correctly decoded message mi−1 .
After expanding the protograph P in Fig. 3, each paritycheck matrix Hi is a TL -circulant matrix with a size of
n × (n + k), which can be written as Hi = [A, B] such that
1
A11 A112
A121 A122
,
A = Aw
(7)
αβ 2×4 = A3
A132
31
0
A142
0
0
B113
0
0
0
0
B124
.
B = Bw
(8)
1
αβ 4×4 = B1
0
0
31 B32
1
2
B41 B42
0
0
Note that n = 4T , k = 2T , and the first k = 2T nodes are
punctured as information nodes among all the n + k = 6T
variable nodes. There are totally J = |E|T /TL non-zero
elements in HB,L−1 , each of which will be then replaced by
a TL × TL circulant permutation matrix ITL (t). Meanwhile,
all the zero elements in HB,L−1 will be replaced by a
TL × TL zero matrix 0TL ×TL . Now, we rewrite the hash value
hash(mi+1 ) as a binary vector
ri = hash(mi−1 ) = (ri,0 , ri,1 , . . . , ri,j , . . . , ri,J−1 ),
(9)
where each number ri,j ∈ J0, TL −1K is represented by log2 TL
bits in the vector ri . Therefore, all the random shift values
in the Lth stage will be controlled by the vector ri , whose
length is required to be h = J log2 TL bits. If the bit length
of hash(mi−1 ) is smaller than h, ri can be filled up by using
hash(mi−1 ) repeatedly. As for the parameters selected in the
first L−1 stages, i.e. all the expanding factors and shift values,
they are constant and will be publicly sent to Bob in advance.
1 Note that Ja, bK := {a, a + 1, . . . , b}, which denotes the set of integers
between a and b.
4
the feedback fi . As a result, an event which we referred to
as synchronization error may happen. That is, there exists an
index iTH ∈ N, such that ziTH is not correctly decoded by Eve,
but yiTH is successfully recovered by Bob. Thus, Eve will be
not able to follow the update of the parity-check matrix Hi
for i ≥ iTH , and the subsequent source message blocks mi
cannot be decoded by Eve without any knowledge of Hi .
Let SNRB and SNRE denote the receiving SNRs of Bob
and Eve, respectively. The BER of Bob is given by
0
10
−1
10
−2
10
−3
BER
10
−4
10
−5
10
CCSDS (Systematic)
BER: CodeHop
FER: CodeHop
Uncoded
−6
10
PeB = Pe (SNRB ),
−7
10
0.5
1
1.5
E /N (dB)
b
2
2.5
o
Fig. 4. For (2048, 1024) LDPC codes, the average BER and FER of the
nonsystematic LDPC codes in CodeHop is compared with the BER of the
systematic code specified in CCSDS.
A. An example
With the structured-random expanding technique, a hopping
sequence of parity-check matrices for (2048, 1024) nonsystematic LDPC codes is generated. From the protograph P in
Fig. 3, each parity-check matrix Hi will be constructed by
a total expanding factor T = T1 T2 T3 = 4 × 4 × 32 = 512
via L = 3 stages. In the first two stages, the protograph P is
expanded twice by factors of T1 = T2 = 4, which is intended
to separate all the parallel edges and guarantees the girth is
larger than 4 in the derived Tanner graph. In the third stage,
totally there are |E|T /T3 = 256 edges to be randomly expanded by a factor of T3 = 32, which is described by a vector
ri with length of h = J log2 TL = 1280 bits. If SHA-256 is
adopted as the hash(·), the 256-bit hash value hash(mi−1 )
will be repeated by four times to represent ri . Thus, a large
set of parity-check matrices H = {H(r) : r ∈ J0, 2256 − 1K}
is available to be randomly generated in the hopping sequence
of CodeHop.
To evaluate the error correcting performance of CodeHop,
it is different from the evaluation of a fixed LDPC paritycheck matrix. The BER of one hopping sequence will be
evaluated as how the CodeHop scheme works, that is, by performing numerical simulations with rapidly switching paritycheck matrices. For the average performance of CodeHop,
random initial H0 will be tested to generate different hopping
sequences, whose BERs and FERs are averaged as the final
result. As shown in Fig. 4, the average BER and FER of the
structured-random nonsystematic (2048, 1024) LDPC codes
are plotted. The number of iterations is restricted by 63. It
can be seen that the average BER of nonsystematic LDPC
codes is slightly degraded by no more than 0.2 dB in Eb /N0 ,
compared with the systematic codes specified in CCSDS.
IV. S ECURITY G AP A NALYSIS
In this section, the analysis of security gap using CodeHop
for gaussian wiretap channel with limited feedback will be
presented. Owing to the public feedback fi , the parity-check
matrix Hi will be always perfectly synchronized between
Alice and Bob. However, Eve is restricted to passively receive
(10)
where Pe (SNR) is the BER of the CodeHop scheme with
perfect synchronization that can be obtained from Fig. 4.
As for Eve’s BER, firstly the event of synchronization
error needs to be investigated. Assume that the total number
of source message blocks is N . Given Pf (SNR) as the
frame error probability (FER) of the CodeHop scheme with
perfect synchronization, the index iTH of the first erroneously
synchronized message block follows a geometric distribution.
Thus, for 1 ≤ i0 ≤ N , the probability of the event iTH = i0
can be derived by
i0 −1
Pr(iTH = i0 ) = 1 − (1 − PfB )PfE
(1 − PfB )PfE , (11)
where we define PfB = Pf (SNRB ) and PfE = Pf (SNRE ). If
synchronization error never happens, the probability is denoted
by
Pr(iTH ≥ N + 1) = [1 − (1 − PfB )PfE ]N .
(12)
Then, for a fixed index iTH , Eve’s erroneously received source
message blocks can be divided into two categories. The first
category involves the blocks mi for 1 ≤ i ≤ iTH − 1, which
are erroneously decoded with knowing the parity-check matrix
Hi . Actually, the number of block errors in the first category
follows a binomial distribution and can be estimated by the
mean value
NER (iTH ) =
(iTH − 1)PfB PfE
1 − (1 − PfB )PfE
,
(13)
where the averaged number of bit errors in such an erroneous
message block is given by
kER =
Pe (SNRE ) · k
.
Pf (SNRE )
(14)
The second category includes all the unknown message blocks
for iTH ≤ i ≤ N , which is caused by synchronization error
and cannot be decoded without knowing Hi . That is, all of the
message bits in mi are randomly decoded and half of them
will be error bits. Finally, the BER of Eve can be obtained by
PeE =
N
X
kER · NER (i) + 0.5k · (N − i + 1)
· Pr(iTH = i)
k·N
i=1
kER · NER (N + 1)
· Pr(iTH ≥ N + 1). (15)
k·N
From (10) and (15), we can characterize the security gap of
the CodeHop scheme for gaussian wiretap channel. As shown
in Fig. 5 and Fig. 6, the BER curves for different security
gaps Sg and different number of transmitted message block N
using (2048, 1024) nonsystematic LDPC codes are illustrated.
+
5
N = 1000
0
10
−1
10
−2
10
−3
BER
10
−4
10
Bob
Eve (Sg = 0 dB)
−5
10
Eve (Sg = 0.2 dB)
−6
10
Eve (Sg = 0.5 dB)
−7
10
0.8
1
1.2
1.4
1.6
Eb/N0 (Bob)
1.8
2
2.2
Fig. 5. BER of the CodeHop scheme for gaussian wiretap channel using
(2048, 1024) nonsystematic LDPC codes for different security gaps when
N = 1000 message blocks are transmitted.
N = 10000
0
10
−2
10
−3
BER
10
−4
10
Bob
Eve (Sg = 0 dB)
−5
10
Eve (Sg = 0.2 dB)
−6
10
Eve (S = 0.5 dB)
g
−7
0.8
ACKNOWLEDGMENT
This work was supported in part by the Natural Science
Foundation of China (61101072, 61132002) and the new
strategic industries development projects of Shenzhen city
(ZDSY20120616141333842).
R EFERENCES
−1
10
10
on the most recent successfully recovered message block at
Bob. Enhanced by an authenticated public feedback channel,
Alice can be notified of the decoding state of Bob and
synchronize to the updated parity-check matrix with Bob.
However, Eve is restricted to passively follow the update of the
parity-check matrix. Once the event of synchronization error
occurs, she would be unable to decode the rest of the message
blocks. In the CodeHop scheme, security gap is considered as a
metric for physical layer security. It is verified with numerical
simulations that the security gap can be highly reduced with
finite length LDPC codes. In the future, more scenarios such as
fading channels will also be studied by applying the CodeHop
scheme.
1
1.2
1.4
1.6
Eb/N0 (Bob)
1.8
2
2.2
Fig. 6. BER of the CodeHop scheme for gaussian wiretap channel using
(2048, 1024) nonsystematic LDPC codes for different security gaps when
N = 10000 message blocks are transmitted.
E
If the BER threshold of Eve is set to be Pe,min
= 0.49, for
N = 1000, we can see in Fig. 5 that Sg = 0.5dB is achievable
B
B
< 10−5 . If Pe,max
is
for the BER threshold of Bob Pe,max
−3
relaxed to be less than 10 , Sg = 0dB can be also achieved.
It can be inferred that with the help of public feedback fi and
rapidly hopping of the parity-check matrix Hi , the security
gap can be significantly reduced. Moreover, if N = 10000,
i.e. more source message blocks are transmitted, the security
B
gap performance can be further improved. E.g., for Pe,max
<
−5
10 , the security gap Sg can be reduced to 0.2dB. This is
owed to the fact that all message blocks mi after iTH cannot
be decoded by Eve. Thus, more source message blocks to
be transmitted makes Eve suffer from more random decoded
blocks. It is worth noting that to make the error distributed
uniformly in the decoded message bits of Eve, interleaving is
needed for all the N transmitted source message blocks.
V. C ONCLUSIONS
Design of practical coding techniques for gaussian wiretap
channel is investigated in this paper. Based on nonsystematic
LDPC codes, we propose the CodeHop scheme in which the
LDPC parity-check matrix will be rapidly switched depending
[1] C. E. Shannon, “Communication theory of secrecy systems,” Bell system
technical journal, vol. 28, no. 4, pp. 656–715, 1949.
[2] A. D. Wyner, “The wire-tap channel,” Bell system technical journal,
vol. 54, pp. 1355–1387, 1975.
[3] I. Csiszár and J. Korner, “Broadcast channels with confidential messages,” IEEE Trans. Inf. Theory, vol. 24, no. 3, pp. 339–348, 1978.
[4] S. Leung-Yan-Cheong and M. Hellman, “The gaussian wire-tap channel,” IEEE Trans. Inf. Theory, vol. 24, no. 4, pp. 451–456, 1978.
[5] P. K. Gopala, L. Lai, and H. El Gamal, “On the secrecy capacity of
fading channels,” IEEE Trans. Inf. Theory, vol. 54, no. 10, pp. 4687–
4698, 2008.
[6] F. Oggier and B. Hassibi, “The secrecy capacity of the mimo wiretap
channel,” IEEE Trans. Inf. Theory, vol. 57, no. 8, pp. 4961–4972, 2011.
[7] E. Ekrem and S. Ulukus, “The secrecy capacity region of the gaussian
mimo multi-receiver wiretap channel,” IEEE Trans. Inf. Theory, vol. 57,
no. 4, pp. 2083–2114, 2011.
[8] A. Thangaraj, S. Dihidar, A. R. Calderbank, S. W. McLaughlin, and J.M. Merolla, “Applications of ldpc codes to the wiretap channel,” IEEE
Trans. Inf. Theory, vol. 53, no. 8, pp. 2933–2945, 2007.
[9] F. Oggier, P. Solé, and J.-C. Belfiore, “Lattice codes for the wiretap gaussian channel: Construction and analysis,” arXiv preprint
arXiv:1103.4086, 2011.
[10] O. O. Koyluoglu and H. El Gamal, “Polar coding for secure transmission
and key agreement,” IEEE Trans. Inf. Forensics Security, vol. 7, no. 5,
pp. 1472–1483, 2012.
[11] C. W. Wong, T. F. Wong, and J. M. Shea, “LDPC code design for
the bpsk-constrained gaussian wiretap channel,” in IEEE GLOBECOM
Workshops, 2011, pp. 898–902.
[12] Z. Chen, L. Yin, Y. Pei, and J. Lu, “Codehop: physical layer error
correction and encryption with ldpc-based code hopping,” Science China
Information Sciences, vol. 59, no. 10, p. 102309, 2016.
[13] Z. Chen, L. Yin, and J. Lu, “Hamming distortion based secrecy systems:
To foil the eavesdropper with finite shared key,” IEEE Communications
Letters, vol. 19, no. 5, pp. 711–714, 2015.
[14] D. Klinc, J. Ha, S. W. McLaughlin, J. Barros, and B.-J. Kwak, “LDPC
codes for physical layer security,” in IEEE Global Telecommunications
Conference (GLOBECOM), 2009, pp. 5765–5770.
[15] M. Baldi, M. Bianchi, and F. Chiaraluce, “Coding with scrambling,
concatenation, and harq for the awgn wire-tap channel: A security gap
analysis,” IEEE Trans. Inf. Forensics Security, vol. 7, no. 3, pp. 883–894,
2012.
[16] W. K. Harrison, J. Almeida, S. W. McLaughlin, and J. Barros, “Coding
for cryptographic security enhancement using stopping sets,” IEEE
Trans. Inf. Forensics Security, vol. 6, no. 3, pp. 575–584, 2011.
[17] J. Thorpe, “Low-density parity-check (LDPC) codes constructed from
protographs,” IPN progress report, vol. 42, no. 154, pp. 42–154, 2003.
[18] S. ten Brink and G. Kramer, “Design of repeat-accumulate codes
for iterative detection and decoding,” IEEE Transactions on Signal
Processing, vol. 51, no. 11, pp. 2764–2772, 2003.
| 7cs.IT
|
МАТЕМАТИЧЕСКИЙ И ПРОГРАММНЫЙ АППАРАТ
АВТОМАТИЗИРОВАННОГО ПРОЕКТИРОВАНИЯ
АКСОНОМЕТРИЧЕСКИХ СХЕМ ТРУБОПРОВОДНЫХ СИСТЕМ
В.В. Мигунов
Рассматривается применение модульной технологии разработки
проблемно-ориентированных расширений систем автоматизированного
проектирования [1] к задаче автоматизации подготовки аксонометрических
схем трубопроводных систем (АСТС). В [2] выявлено единство состава
схем для специальных технологических трубопроводов, систем
водопровода и канализации, отопления, теплоснабжения, вентиляции,
кондиционирования
воздуха
и
приведено
структурированное
параметрическое представление схем, включая свойства объектов и их
связи, общие установки и установки по умолчанию. В [3] описаны
особенности реализации типовых операций, состав и реализация
специальных операций проектирования.
Реализация плоских схем трехмерных каркасных объектов (принято
присваивать таким моделям размерность 2,5) ввиду относительно
большого объема вычислений требует специализированного модельного,
методического и математического подходов, в общем случае
отсутствующих в двумерных графических редакторах САПР. Эти
подходы, реализованные в САПР TechnoCAD GlassX, излагаются ниже.
Системы координат
Во всех системах координат единицей измерения являются
миллиметры. В самом чертеже имеются системы координат Натура (Re) и
Бумага (Pr), отличающиеся масштабом всего чертежа, указываемом в
основной надписи. У экранной системы координат Sc ось X идет вправо, Y
- вниз, единица измерения - пиксел. Специально для АСТС вводятся
дополнительные декартовы системы координат.
ReIn - правая трехмерная. Соответствует представлению концов труб
в реальном трехмерном пространстве. Начало координат задается точкой
привязки АСТС в чертеже. Единица измерения - мм Натуры.
ReVi - видимая трехмерная. Соответствует представлению концов
труб и всех других элементов АСТС в трехмерном пространстве после
всех смещений. Во всем совпадает с ReIn, но расстояния в ней не
соответствуют реальным длинам, если действуют смещения. Смещения в
АСТС
с
условным
изображением
длины
труб
разрешены
соответствующими ГОСТами. Общие смещения действуют на
полупространства, местные - на ветки труб, идущие от одной или
нескольких параллельных труб.
Li - библиотечная, двумерная. Соответствует представлению блоков
(условных графических обозначений трубопроводной арматуры и
элементов трубопроводов) в графической библиотеке АСТС. Начало
координат является точкой привязки блока, ее координата вдоль трубы
задается при привязке блока к трубе. Ось X является осью привязки блока
к трубопроводу, совпадает с осью трубы после привязки. Вдоль этой оси
задается отрезок, вырезаемый блоком на трубе при установке на нее.
Единица измерения - мм Бумаги.
Типы данных
Используются четырехбайтные вещественные числа (Float), что
продиктовано требованием компактности хранения АСТС. Вводятся типы
данных: V2, V3 - векторы из 2 и 3 компонентов Float; M22, M32, M23, M33
- матрицы с соответствующим числом строк и столбцов из элементов Float.
В терминах таких типов задаются направленные отрезки (V2, V2) в Li,
Re, Pr, Sc и (V3, V3) в ReIn, ReVi; ориентированная плоскость в ReIn (V3,
Float) - орт нормали в положительном направлении и координата
плоскости на этом направлении; оператор проецирования (V2, M23) из
ReIn и ReVi в Re, Pr, Li, Sc - линейное преобразование и сдвиг;
аналогичные операторы размещения блока (V3, M32) из Li в ReIn и
оператор перехода (V2, M22) из Li в Re, Pr, Sc.
Общие принципы реализации операций над данными
Состав и реализация операций над такими данными подчинены
требованиями компактности и наглядности использующего их
программного кода, а также быстродействия. Во всех случаях передача
аргументов длиной более 4 байтов осуществляется по ссылке, а не по
значению. Для использования одних и тех же процедур для векторов и
строк/столбцов матриц в описании интерфейса аргументы указываются как
нетипированные переменные.
Состав основных операций
Чтение и запись значений Float, расположенных в элементе массива с
заданным номером.
Проверки с заданной погрешностью равенства нулю и единице чисел
Float, покомпонентного равенства векторов V2, V3.
Обнуление, смена знака, сложение и вычитание векторов V2 и V3.
Умножение векторов V2 и V3 на скаляр, умножение на скаляр
заданного числа идущих подряд элементов Float.
Скалярные произведения векторов V2 и V3. Скалярное произведение
X: V2 на нормаль к Y: V2, получаемую путем поворота Y на 90° против
часовой стрелки. Модуль этого произведения равен расстоянию от конца X
до линии, продолжающей Y, если Y - орт.
Вычисление длин и ортов векторов V2 и V3, расстояний между
точками V2 и V3 и ортов направлений от одной к другой.
Векторное умножение векторов V3.
Определение скаляра k в формуле Y = k * X для коллинеарных X, Y:
V3 с заданной погрешностью.
Определение с заданной погрешностью индекса оси (1 - +X, 2 - +Y, 3 -
+Z, -1 - -X, -2 - -Y, -3 - -Z), в направлении которой идет вектор V3.
Обратная операция - вычисление орта оси по ее индексу и орта трубы по ее
номеру в списке. Эта довольно необычная операция нужна для выяснения
ориентации труб при нанесении цепных размеров в ReVi, и
соответственно, при установке блоков на трубы. Дело в том, что выносные
линии должны находиться в плоскости блока и одновременно идти вдоль
одной из осей координат. Допустимый набор ориентации плоскостей
блоков и размеров зависит от взаимного расположения трубы и осей
координат ReIn.
Умножение векторов на матрицы: M22*V2, M23*V3, M32*V2,
M33*V3, M22*V2.
Умножение матриц M23*M32 => M22.
Вычисление определителя матрицы M22.
Пересчеты координат точек V3 при переходах между системами
координат ReIn и ReVi с учетом общих, местных и всех смещений.
Пересчет координат точек V3 из ReVi в систему координат Бумага
чертежа.
Проверка того, что матрица M23 проецирования из ReVi в Pr
превращает плоскость, образуемую двумя непараллельными ненулевыми
векторами M3Rec, в прямую линию.
Приведенные сведения могут быть полезны при разработке
программных средств, оперирующих в пространстве размерности 2,5.
Литература
1. Мигунов В.В. Модульная технология разработки проблемноориентированных расширений САПР реконструкции предприятия /
Материалы Второй международной электронной научно-технической
конференции "Технологическая системотехника" (ТСТ'2003), г.Тула,
01.09.2003-30.10.2003 [Электронный ресурс] / Тульский государственный
университет. – Режим доступа: http://www.tsu.tula.ru/aim/, свободный. –
Загл. с экрана. – Яз. рус., англ.
2. Мигунов В.В., Кафиятуллов Р.Р., Сафин И.Т. Модульная
технология разработки расширений САПР. Аксонометрические схемы
трубопроводов. Параметрическое представление / Там же.
3. Сафин И.Т., Мигунов В.В., Кафиятуллов Р.Р. Модульная
технология разработки расширений САПР. Аксонометрические схемы
трубопроводов. Типовые и специальные операции / Там же.
Мигунов Владимир Владимирович; к.т.н.; ст. науч. сотр.; член-корр.
Международной академии информатизации; нач. отдела инф-х технологий
Центра экономических и социальных исследований Республики Татарстан;
420088 Казань, Губкина, д.50; (8432) 73-37-14; [email protected].
| 5cs.CE
|
Karlsruhe
October 03 2012
Revised: March 03 2017
The field Q(2 cos
π
n
), its Galois group, and length ratios in the
regular n-gon
arXiv:1210.1018v2 [math.GR] 7 Mar 2017
Wolfdieter L a n g
1
Abstract
π
The normal field extension Q(ρ(n)), with the algebraic number ρ(n) := 2 cos
, for n ∈ N,
n
is related to ratios of the lengths between diagonals and the side of a regular n−gon. This has been
considered in a paper by P. Steinbach. These ratios, numbered k = 1, ..., n−1, are given by Chebyshev
polynomials S(k − 1, x = ρ(n)). The product formula for these ratios was found by Steinbach and is
re-derived here from a known formula for the product of Chebyshev S−polynomials. It is shown that
it follows also from the S−polynomial recurrence and certain rules following from the trigonometric
nature of the argument x = ρ(n). The minimal integer polynomial C(n, x) for ρ(n) is presented, and
its simple zeros are expressed in the power-basis of Q(ρ(n)). Also the positive zeros of the Chebyshev
polynomial S(k − 1, ρ(n)) are rewritten in this basis. The number of positive and negative zeros of
C(n, x) is determined. The coefficient C(n, 0) is computed for special classes of n values. Theorems
on C(n, x) in terms of monic integer Chebyshev polynomials of the first kind (called here t̂) are given.
These polynomials can be factorized in terms of the minimal C-polynomials. A conjecture on the
discriminant of these polynomials is made. The Galois group is either Zδ(n) , the cyclic group of the
order given by the degree δ(n) of C(n, x), or it is a direct product of certain cyclic groups. In order
to determine the cycle structure a novel modular multiplication, called Modd n is introduced. On the
reduced odd residue system Modd n this furnishes a group which is isomorphic to this Galois group.
1
Introduction and Summary
Length ratios between diagonals and the side of a regular n−gon (called diagonal/side ratios, abbreviated
DSRs) have been considered by Steinbach [29]. He gave a product formula for these ratios (called by
him diagonal product formula (DPF)), and for
√ the pentagon and heptagon details were given. In the
pentagon case the quadratic number field Q( 5) with the basis < 1, ϕ > for integers of this field turns
up. Here ϕ, the golden section, is identified with ρ(5) which is the length ratio between any of the two
diagonals and the pentagon side. For the heptagon there are two different diagonal length and the two
ratios between them and the side length have been called ρ := ρ(7) and σ. The DPF allowed to reduce
all products and quotients of ρ and σ to Q-vectors with the basis < 1, ρ, σ > . For example, ρ2 = 1 + σ,
and one can use instead the power basis < 1, ρ, ρ2 > of the algebraic number field Q(ρ(7)) (the use of the
same acute bracket notation for different bases should not lead to a confusion). The minimal polynomial
of ρ(7), in the present work called C(7,x), is x3 − x2 − 2 x + 1, and this was also given in [29]. Therefore,
[Q(ρ(7)) : Q] = 3, which is the degree of this field extension. For odd n, n = 2 k + 1, Steinbach gave
a polynomial with one of its roots ρ(2 k + 1). In the present work the minimal
polynomial for ρ(n),
2π
called C(n,x), for every n ∈ N, is given in terms of the known one for cos
, called Ψ(N, x), with
N
N = 2 n. The connection between Ψ(N, x) and divisor product representations of N , and to Chebyshev
T −polynomials has been worked out earlier by the author. See the links under OEIS [27] A181875 and
A007955 (in the sequel OEIS Anumbers will appear without repeating this reference). C(n, x) turns out
1
[email protected],
http://https://www.itp.kit.edu/~ wl
1
to be an integer polynomials of degree δ(n) = A055034(n). All its zeros are known, and they are simple.
Q(ρ(n)) is the splitting field for the minimal polynomial C(n, x) of ρ(n). It is a normal extension of
the rational field. These C-polynomial zeros are written in the power basis of Q(ρ(n)) with the help of
certain scaled Cheyshev T −polynomials (called here t̂ with their integer coefficient array A128672). This
array can also be used to rewrite the positive zeros of the Chebyshev S(n − 1, x) polynomial, related to
the DSRs, in this power basis. The use of S−polynomial technology allows also for a re-derivation of the
product formulae (DPF) for the length ratios between the regular n−gon diagonals and side.
In section 2 S(k − 1, ρ(n)) is shown to yield the DSRs of the regular n-gon. The DPF is also re-derived
there, and the independent products are extracted. This is called reduced algebra over Q. In section 3
the minimal polynomial C(n, x) is presented, and its zeros are related to the power basis of the algebraic
number field Q(ρ(n)). In section 4 the Q-automorphisms of this field, the Galois group G(Q(ρ(n))/Q),
is treated. It is the cyclic group Zδ(n) , except for infinitely many n-values where it is the direct product
of cyclic groups, hence non-cyclic. For n from 1..100 there are 30 non-cyclic groups. A novel modular
multiplication on the odd numbers, called Modd n, is introduced which serves to determine the cycle
graph structure of this Galois group. The multiplicative group Modd n is isomorphic to this Galois group.
On the reduced odd residue system Modd n this multiplicative group is isomorphic to the Galois group.
2
Regular polygon diagonals/side ratios (DSRs)
Motivated by a paper of Steinbach [29] we consider the diagonals and the side in the regular n−gon
(n)
(inscribed in the unit circle). The vertices on the unit circle are called Vk , k = 0, ...., n − 1, n= 2, 3, ...;
π
(n)
. The
V0 has coordinates (1, 0); the side length (with the radius’ length 1 unit) is s(n) = 2 sin
n
πk
(n)
(n) (n)
length of V0 Vk is dk = 2 sin
, for k = 1, ..., n − 1, hence d1 = s(n). We only need one side
n
and the diagonals in the upper half plane, i.e. it suffices to consider k ∈ {1, ..., ⌊n/2⌋}. For even n, the
n
largest diagonal (k = ), of length 2, lies on the negative real axis. See Fig. 1, and Fig. 2, for the case
2
n = 10 (decagon), and for n = 11 (hendecagon), respectively. The length ratios of interest, the DSRs,
(n)
d
(n)
are Rk = k(n) for the given k−values. Here Chebyshev S−polynomials in their trigonometric form
d1
enter the stage (see A049310 for their coefficients):
(n)
Rk
= S(k − 1, ρ(n)),
(1)
(n)
where we use the second ratio (for the smallest diagonal) ρ(n) := R2
π
. Remember that
n
S(n, x) := U (n, x/2) with Cheyshev U −polynomial
(second kind). It is well known that the zeros of the
kπ
(n−1)
, for k = 1, ..., n−1
polynomials S(n − 1, x) are xk,± = ±2 cos
. Note that in the even n case
2
n
the zero x = 0 appears twice from ±0. The S−polynomials are orthogonal polynomials defined on the
real interval [−2, +2]. Hence their zeros are guaranteed to be simple. They belong to the Jacobi class.
Due to (anti)symmetry it is sufficient to consider only the positive zeros (we disregard x = 0 in the even
n case). Now also Chebyshev T −polynomials (first kind; see the coefficient table A053120) enter, which
are written in terms of S−polynomials. (The T −polynomials appear as trace polynomials in a 2 × 2
transfer matrix when one solves the three term recurrences of these classical orthogonal polynomials.)
(n−1)
xk
(n−1)
≡ xk,+
=
=
= 2 cos
2 T (k, ρ(n)/2) =: t̂k (ρ(n)) ≡ t̂(k, ρ(n))
n−1
.
S(k, ρ(n)) − S(k − 2, ρ(n)) , k = 1, ...,
2
2
(2)
(3)
Note that these positive zeros decrease with k, and that t̂k is an integer polynomial of degree k because the
S−polynomials are integer. The coefficient array for these monic t̂-polynomials is A127672. In [1] these
polynomials are called Chebyshev C-polynomials, but we use the letter C for the minimal polynomials
of ρ(n). The zeros can be replaced by DSRs from the above given eq. (1).
Positive zeros of S(n − 1, x) from DSRs:
(n−1)
xk
Figure 1: Decagon n = 10
(n)
(n)
= Rk+1 − Rk−1 .
(4)
Figure 2: Hendecagon (n=11)
As an aside we give the factorization of S(n − 1, x) in terms of the polynomial
k
Y
lπ
P (k, x) :=
with positive zeros.
x − 2 cos
2k + 1
l=1
S(n − 1, x) = xΘ(n−1) P
n−1
n−1
n−1
, x (−1)⌊ 2 ⌋ P
, −x ,
2
2
(5)
where θ(n − 1) := 0 if n is odd, and 1 if n is even. This factorization is also considered in [4] for the
U (n, x) polynomials for even and odd n separately. For example, for the heptagon n = 7 one finds the
factorization of S(6, x) with the following real polynomial P (3, x).
P (3, x) = x3 − (2 σ(7) − 1) x2 + 2 ρ(7) x − 1 ,
(7)
(7)
(6)
with ρ(7) = R2 and σ(7) := R3 (see also the introduction). The three positive zeros are, written in
terms of the DSRs, ρ(7), σ(7) − 1 and σ(7) − ρ(7). Here we used the general translation formulae eq.
(4) to rewrite the positive zeros of S(n − 1, x) in terms of DSRs. One should also write this in terms of
powers of
the
help of eq. (1) (see eq. (7) below).
ρ(7) with
n−1
+ 1)-term recurrence
, x can be considered as characteristic polynomial for a ( n−1
Each P
2
2
∞
sequence {fk }k=0 in the integral domain of Q(ρ(n)) (represented by a δ(n)-tuple of ordinary (rational)
integers (A1,k , A2,k , ..., Aδ(n),k ). Here δ(n) is the degree of the minimal polynomial of ρ(n) which will be
such δ(n)-tuple sequences corresponding to the independent
discussed in sect. 3. In fact, one has n−1
2
inputs. As a simple example take the n = 5 pair (δ(5) = 2) of sequences corresponding
√ to P (2, x) =
5) = A1,k 1 +
x2 − (2 ϕ − 1) x + 1, with√
the golden section
ϕ
:=
ρ(5)
and
the
simplest
input.
This
is
S(k,
√
k
A2,k ϕ with input S(−1, 5) = 0, S(0, 5) = 1. See A005013(k + 1) (−1) , and 2 A147600(k − 1), for
A1,k and A2,k , respectively.
3
For later
it is also useful to give a dictionary between the positive zeros of S(n − 1, x), i.e. those
n−1purposes
of P ( 2 , x), and powers of ρ(n). One direction follows from the above given formula, eq. (2), for
(n−1)
xk
in terms of the monic integer polynomials t̂(k, ρ) = 2 T (k, ρ/2). The dependence on n has here
been omitted.
k
⌊X
2⌋
(n−1)
l k−l
(−1)
=
Positive zeros of S(n − 1, x) from ρ(n)-powers:
xk
ρ(n)k−2l .
(7)
l
l=0
The coefficients are given in the triangle A127672. E.g., x5 = 5 ρ − 5 ρ3 + ρ5 (omitting n). For
the heptagon this shows that the three positive zeros of S(6, x) are ρ(7), σ(7) − 1 = ρ2 (7) − 2 and
σ(7) − ρ(7) = ρ(7)2 − ρ(7) − 1 . In this case these three zeros could be used as a Q-vector space
basis. Note that the n dependence is via the t̂-polynomial variable ρ(n) (which satisfies C(n, ρ(n)) = 0 by
definition of the minimal polynomial C). For example, it is true for all n that the second largest positive
(n−1)
(n)
zero of S(n − 1, x), namely x2
, is always ρ(n)2 − 2. Hence, from above, R3 − 1 = ρ(n)2 − 2; or
(n)
if one calls R3 =: σ(n) then ρ(n)2 = σ(n) + 1. This has been noted for n = 7 above but is holds
in general, showing that σ(n) is algebraically dependent on ρ(n). This fact has been observed already in
[29] where it appeared as a special product formula (see the later DPF eq. (13), m = k = 2, or eq. (11)
with k = 2).
The inverse formulae (DSRs in terms of ρ-powers) have already been given above in eq. (1). Here one
(n)
uses the S-triangle A049310 for the translation. E.g., R4 = −2 ρ(n) + ρ3 (n) = ρ(n) (ρ(n)2 − 2)
(n)
= ρ(n) (σ(n) − 1) . Of course, R4 is interesting only for n ≥ 9 because only then the corresponding
diagonal lies in the upper half-plane (including the negative real axis).
, is less or equal to n2 , the number
Note that the number of positive zeros of S(n − 1, x), i.e., n−1
2
of DSRs for diagonals in the upper half plane including the negative real axis. The zero x = 1 never
appears. As can be seen in the n = 7 case these zeros can nevertheless be used as Q-vector space basis (in
sect. 3 it will become clear that the degree δ(7) is also 3). In the case n = 9 the two numbers are both 4,
but the degree δ(9) is 3, hence only three of the zeros and three of the DSRs are rationally independent.
(8)
(8)
(8)
(9)
(9)
(9)
To wit: x(8) = −5 x3 + 3 x1 − 4 x2 and R4 = R1 + R2 = 1 + ρ(9).
Because of symmetry
k DSRs for diagonals of the upper half plane (including the negative real axis),
j nonly
} , are of interest. The reduction for other k values, also negative ones, is
i.e., k ∈ {0, 1, ...,
2
accomplished by the rules
jnk
(n)
(n)
(n)
(n)
(n)
(n)
o) Rn+k = −Rk , i) R−|k| = −R|k| , ii) for k ∈ {
+ 1, ..., n} : Rk = Rn−k .
(8)
2
These rules follow from eq. (1) with the trigonometric definition (with the specific value of ρ(n)) of the
S-polynomials, and therefore also negative values for the DSRs show up (the interesting DSRs are, of
course, positive). i) follows also from S(−|n|, x) = −S(|n| − 2, x) which derives from a backward use of
the recurrence (given later in eq. (12)). This rule does therefore not depend on the special choice of the
variable x, in contrast to the rules o) and ii). With negative k one counts the diagonals in the clockwise
direction. ii) is used to translate from the lower to the upper half-plane.
The product formula for Chebyshev S-polynomials of different degree, but with the same argument, is
well known. See e.g., [1], p.782, 22.7.25, for the U -polynomials, and replace the T -polynomials by the
S-polynomials via the trace formula given in eqs (2) and (3).
S(m − 1, x) S(n − 1, x) =
1
4 (( x2 )2 − 1)
[S(n + m, x) − S(n + m − 2, x) −
S(n − m, x) + S(n − m − 2, x) ] n ≥ m .
4
(9)
(n)
For our purpose x = ρ(n) = R2 6= ±2 but its special form is for this formula not of interest . It is in
fact symmetric under exchange of n with m if the rules for negative indices on the S-polynomials, stated
above, are employed. Here we restrict to n ≥ m in order to have non-negative indices. Thus the product
formula for the DSRs, now with x = ρ(n), but without using its specific value, is
(n)
(n)
(n)
(4 − (R2 )2 ) Rm
Rk
(n)
(n)
(n)
(n)
= Rk−m+1 − Rk−m−1 − Rk+m+1 + Rk+m−1 , k ≥ m > 0 .
(10)
For m = 0 this becomes trivial. This is not yet the DPF given by Steinbach in [29] which linearizes
(n) (n)
the product of Rm Rk . In order to eliminate the pre-factor (4 − ρ(n)2 ) one can use the three term
recurrence relation of the orthogonal S-polynomials, written for the DSRs by eq. (1) as a special product
formula
(n)
(n)
(n) (n)
(11)
[Rec, k](n) :
R2 Rk = Rk+1 + Rk−1 , k ≥ 0 .
From now on we also use the specific form of ρ(n). Thus also rules o) and ii) of eq. (8) will be
(n)
(n)
applicable. For k = 0 one uses (see ii)) R−1 = − R1 = −1, and it becomes trivial. For k = 2
(n)
it shows that ρ(n)2 = σ(n) + 1 for all n, if one uses σ(n) := R3 . This recurrence can now be
(n)
(n) (n)
used twice as R2 (R2 Rk ) in eq. (10) to produce the following recurrence for two step differences of
(n,m)
(n) (n)
Rm
Rk =: pk
≡ pk
(pk+2 − pk ) − (pk − pk−2 ) = ck − ck−2 ,
(n,m)
(12)
(n)
(n)
:= Rk+m+1 − Rk−m+1 . This shows that pk+2 − pk − ck
where we used the abbreviation ck ≡ ck
is k independent. From the inputs p2 − p0 − c0 = 0, due to the recurrence relation eq. (11), and
p1 − p−1 − c−1 = 0 with the help of the rule i) from eq. (8), this leads to the recurrence pk+2 − pk = ck ,
(n)
with the inputs p0 = 0 and p1 = −p1 = Rm . The solution for pk is found for even and odd k separately,
where again the rules for negative indices are employed. Both solutions can then be combined as
(n)
DPF [m, k]
(n) (n)
Rm
Rk
:
=
k−1
X
j=0
(n)
Rm+k−(2j+1) , 1 ≤ m ≤ k .
(13)
This is finally the DPF found by Steinbach in [29] and gives a linearization of the DSR products. We
need only to consider m ≥ 2 (m = 1 becomes trivial), and due to the symmetry of this formula
n under
the transformation k → n − k (using
the rules of eq. (8)) is suffices to consider k ∈ {1, 2, ... 2 }. Hence
n
2
one has only to consider the
products for 2 ≤ m ≤ k ≤ n2 .
2
The DPF formula looks un-symmetric with respect to m ↔ k, but it is, in fact, symmetric because
m−1
X (n)
Rk+m−(2j+1) for k < m reduces, due to cancellations after using the rule i) from eq (8), to the
j=0
expected sum with only k terms. One can see this for even and odd m > k separately, remembering
(n)
that R0 = 0 in the former case.
The idea is to work out, for a
given
n ≥ 4, all DPFs of interest, using the rules from eq. (8), especially
n
2
products as linear combinations of the DSRs.
2
Example 1: n = 7 heptagon, treated in detail in reference [29]. Here we recapitulate and link to OEIS
[27] sequences (the analoga of Fibonacci numbers in the pentagon case). Superscripts and arguments 7
are suppressed. [2, 2]: ρ2 = σ + 1. [2, 3]: ρ σ = σ + ρ + 0. [3, 3]: σ 2 = ρ + σ + 1. Here, R5 = R2
was used. This shows that the Q-vector space basis is at most < 1, ρ, σ >. It will be shown to be indeed
the heptagon basis in sect. 4, where the power basis will be used instead.
With these DPFs one can compute all powers of interest in the heptagon basis < 1, ρ, σ > (this has
already been done explicitely for σ in [29], p. 28).
i) and ii) in order to write all
5
ρk = A052547(k − 2) 1 + A052547(k − 1) ρ + A006053(k) σ, k ≥ 0,
ρ−k = A077998(k) 1 + A077998(k − 1) ρ − A006054(k + 1) σ, k ≥ 0 ,
σ k = A106803(k − 1) 1 + A006054(k − 1) ρ + A106803(k) σ, k ≥ 0,
σ −k = (σ − ρ)k = A052547(k) 1 − A006053(k + 1) ρ − A052547(k − 1) σ, k ≥ 0,
(ρ σ)k = (ρ + σ)k = A120757(k) 1 + —A006054(k − 1)| ρ + 4 A181879(k) σ, k ≥ 0,
(ρ σ)−k = (ρ + σ)−k = A085810(k) 1 + (−1)k A181880(k − 2) ρ + (−1)k+1 A116423(k + 1) σ, k ≥ 0.
1
= A 1 + B ρ + σ and find with N (a, b, c) = a3 − b3 − c3 −
One can also compute
a1 + bρ + cρ
2 a b2 − a b c + a2 b + b2 c + 2 a2 c − a c2 + 2 b c2
A
=
C
=
1
1
(a2 − b2 + a b + 2 a c − b c) , B =
(b2 − c2 + a b) ,
N (a, b, c)
N (a, b, c)
1
(c2 − b2 + a c − b c).
N (a, b, c)
(9)
(9)
(14)
(15)
(9)
Example 2: n = 9 (Enneagon), where we use ρ = R2 , σ = R3 , τ = R4 . [2, 2] : ρ2 = 1 + σ,
[2, 3] : ρσ = ρ + τ , [2, 4] : ρτ = σ + τ , [3, 3] : σ 2 = 1 + τ + σ, [3, 4] : σ τ = ρ + σ + τ ,
(n)
[4, 4] : τ 2 = 1 + ρ + σ + τ . This DSR-algebra (over Q) shows that not all ratios (including R1 = 1)
are linear independent: τ (ρ − 1) = σ = ρ2 − 1 = (ρ + 1) (ρ − 1), i.e., τ = ρ + 1. Therefore,
ρ3 = ρ (σ + 1) = 2 ρ + τ + 2 = 3 ρ + 1 . In sect. 3 we will see that the algebraic number ρ(9)
has degree 3, and its minimal polynomial is indeed C(9, x) = x3 − 3 x − 1. The enneagon basis is
thus < 1, ρ, σ > which can be related to the power basis < 1, ρ, ρ2 > (remember that we use the same
notation for different bases). The reduced DSR-algebra (the algebra modulo C(9, ρ) = 0 ) is ρ2 = 1 + σ,
ρ σ = 1 + 2 ρ, and σ 2 = 2 + ρ + σ.
(n)
(n)
Remark 1: [2, 2] in eq. (13) becomes (R2 )2 = R3 + 1, showing that for all interesting values n ≥ 4
(n)
one has σ(n) := R3 = ρ(n)2 − 1 . This is known already from eq (1) for k = 3, and it will be used in
(n)
sect. 3 as R3 rewritten in terms of the power basis of Q(ρ(n)).
(n)
Remark 2: For n = 4 and n = 5 the second diagonal (defining R3 ) is not of interest because it is
(5)
(5)
in the lower half plane. According to eq. (8), ii) R2 = R3 , hence σ(5) = ρ(5) and the general
relation [2, 2] from eq.√(13) between σ(n) and ρ(n) leads to ρ(5) = ρ(5)2 − 1, the golden section formula
5)/2. In sect. 2 it will be seen that the minimal polynomial for ρ(5) is indeed
ρ(5) = ϕ := (1 +
2
C(5, x) = x − x − 1.
Instead of deriving the DPF from the product formula for S-polynomials and the rules of eq. (8) one
can prove it directly by induction from the S−recurrence including negative indices in accordance with
these rules.
Proposition 1:
The DPFs [m, k](n) , eq. (13) but now for all 2 ≤ m, 2 ≤ k, follow from the recurrence, eq. (11), and
the rules i) and ii) from eq. (8).
Proof: This is shown by double induction. First one shows this for given m ≥ 2 for all k ≥ 2 by
induction over k. Then by induction over m for all k ≥ 2.
The k-induction uses as starter [2, 2](n) which is the recurrence. Assume that [2, k′ ](n) is true for all
k′ = 2, 3, ..., k − 1. Now (we omit the superscripts) R2 Rk = R2 (R2 Rk−1 − Rk−2 ) from the recurrence.
Then use for both terms the induction hypothesis. Note that one obtains for the first term k − 1, and
for the second one k − 2 terms. However, due to rule ii) only two terms survive in each case. (The same
type of cancellation was at work when we remarked above on the symmetry of the DPF formula in m
and k.) Therefore, one obtains R2 (Rk − Rk−2 + 0) − (Rk−1 − Rk−3 + 0) which becomes, after use of
the recurrence applied twice, Rk+1 + Rk−1 . This is indeed the desired result for [2, k] if one uses again
rule ii) to get a truncation of the sum after two terms.
6
The m-induction uses as starter [2, k] for all k ≥ 2, which has just been established. Then assume that
[m′ , k] is true for all m′ = 2, 3, ..., m − 1. Rm Rk = R2 Rm−1 Rk − Rm−2 Rk from the recurrence.
Assume the induction hypothesis for each term, obtaining, after use of the recurrence for the first sum,
k−1
k−1
X
X
(Rm−2+k−(2j+1) which is indeed the assertion after cancellation
(Rm+k−(2j+1) + Rm+k−2−(2j+1) ) −
j=0
j=0
of the last two sums.
As mentioned in the example 2 the DSR-algebra turns sometimes out to be reducible because some DSRs
can be expressed as rational linear
combinations of other ones. Later it will become clear that this
n
happens precisely whenever 2 − δ(n) > 0, and this is the number of linear dependent DSRs. See
Table 1 for details for n = 3, ..., 12.
We shall see in sects. 3 and 4 that it is simpler to use the power basis of Q(ρ(n)) and the minimal C(n, x)polynomials of the algebraic number ρ(n) instead of the DSR-algebra. Then one obtains automatically
the reduced algebra.
3
Minimal polynomial of ρ(n), its zeros, absolute term and factorization
The minimal polynomial of an algebraic number α of degree dα is the monic, minimal degree
rational polynomial which has as root, or as one of its roots, α. This degree dα is 1 iff α is rational, and
the minimal polynomial in this case is p(x) = x − α. For the notion ‘minimal polynomial of an algebraic
number’ see, e.g., [22], p. 28 or [25] p. 13.
π
ϕ(2 n)
For the algebraic number ρ(n) := 2 cos
, for n ∈ N, the degree is δ(1) = 1, and δ(n) =
n
2
for n ≥ 2 , with Euler’s totient function ϕ(n) = A000010(n). This is the sequence A055034(n). The
sequence of allowed δ values is given in A207333. The array with the indices of the polynomials for given
allowed δ values is shown in A207334. Of course, δ is not multiplicative,
e.g., 2 = δ(6) 6= δ(3) δ(3) =
2π
1 · 1 = 1. These polynomials can be obtained from the ones of cos
which are found, e.g., under
n
A181875/ A181876, and they have been called there Ψ(n, x). See also [17], and [22],Theorem
3.9, p. 37,
π
2π
for the degree d(n) of these polynomials. From the trivial identity cos
= cos
one finds the
n
2n
π
, called here C(n, x), from
minimal polynomial of 2 cos
n
x
C(n, x) = 2δ(n) Ψ 2 n,
.
(16)
2
Therefore the above formula for δ(n) derives from d(2 n). These polynomials are given for n = 1, ..., 30
in Table 2, and the first 15 rows of their coefficient array are shown in Table 3. Concerning the parity of
the degree δ one has the following lemma.
Lemma 1: Parity of the degree δ
δ(n) is odd iff n = 1, 2, and pe , with an odd prime p of the form 4 k + 3, k ∈ N0 , and e ≥ 1.
Proof: The case n = 1 is clear by definition. For n = 2 one uses the prime number factorization n =
N
Y
Q
e −1
ej
β−1
(pj − 1) pj j .
2β N
p
,
with
β
≥
0
and
e
=
1.
It
follows
from
the
definition
that
δ(n)
=
2
j
j=1 j
j=1
For β ≥ 2 this is always even. For β = 0 one needs N = 1, otherwise this will be even. For N = 1 one
p−1
needs
to be odd, i.e., p = 4 k + 3 with k ∈ N0 . If β = 1 one needs N = 0 in order to have an
2
odd value. Thus n = 2.
7
In order to relate to sect. 2 we are interested here in n ≥ 4. It turns out that these polynomials are in
fact integer (not only rational) polynomials. The proof will use the following lemma based on our divisor
product representation paper [14].
Lemma 2: Minimal polynomial C(n, x) written as a rational function.
C(n, x) =
p(dp (n), x)
,
q(dq (n), x)
(17)
with monic integer polynomials p and q with a certain degree dp (n) and dq (n) = δp (n) − δ(n), respectively.
Proof: Ψ 2 n, x2 is obtained from the unique divisor product representation dpr(2 n) defined in [14]
by replacing each a(k) in the numerator, as well as in the denominator, by t(k, x2 ) which is given as a
difference of monic integer polynomials t̂ which have been defined already in eq. (2), multiplied with a
certain prefactor.
1
k
k
t̂(
+
1,
x)
−
t̂(
−
1,
x)
if k is even ,
k +1
2
2
22
x
t(k, ) =
(18)
2
k+1
1
k−1
k−1
t̂(
,
x)
−
t̂(
,
x)
if
k
is
odd
.
2
2
+1
2
2
The monic Ψ(2 n, x) polynomials have degree d(2 n) (see e.g., [13]) which implies that all the prefactors
in this replacement of the a(k)s in dpr(2 n) have to become 1/2d(2 n) . (This could be formulated as a
a(1)
18+1
separate lemma). E.g., n = 34 with dpr(34) = a(34)
a(17) a(2) has from the numerator the factor 1/2
and from the denominator 29+2 , fitting with 1/2d(34) because d(34) = 8. Therefore, one may in the
calculation of C(n, x), found above from Ψ(2 n, x2 ), forget about these prefactors in the replacements
altogether. Thus the numerator, resp. denominator, is a product of monic integer polynomials t̂ leading
to the monic integer polynomials p, resp. q, of a certain degree dp (n), resp. dq (n). (One could give
more details on these degrees but this is not important here. Trivially, dp (n) − dq (n) = d(2 n). For
the given example n = 34: (18 + 1) − (9 + 2) = 8 from the degrees of the t̂ polynomials.) Because
C(n,x) is a minimal polynomial this rational function allows polynomial division without remainder.
Proposition 2: C ∈ Z[x]
π
, is an integer monic polynomial.
n
Proof: From lemma 2 we have q(dq (n), x) C(n, x) = p(dp (n), x) which leads by induction to the result that the integer coefficients of the monic polynomials p and
integer coefficients
for the
PN q imply
PM
l , C(M, x) :=
monic polynomial C. Call these monic polynomials q(N, x) :=
q
x
c
xk and
l
k
l=0
k=0
PN +M
p(N + M, x) :=
pl xl with qN = 1, cM = 1 and pN +M = 1. Collecting terms for xN +M −j , for j =
l=0
(0), 1, 2, ..., M , one obtains the formula for the C-coefficient cM −j in terms of lower indexed ones: cM −j =
j
X
pN +M −j −
cM −j−k qN −k . With this the inductive proof on j, using the integer coefficients of q, p and
C(n, x), the minimal polynomial of ρ(n) = 2 cos
k=1
the integer higher C-coefficients due to the induction hypothesis becomes obvious. The starting point is
the trivial j = 0 case.
Next we give all the zeros of C(n, x). This follows directly from the knowledge of all zeros of Ψ(2 n, x2 )
(see, e.g., [13], from p. 473 of [30])
Proposition 3: All zeros of C
C(n, x) =
n−1
Y
k=1
gcd(k,2 n)=1
x − 2 cos
8
πk
n
, n ∈ N.
(19)
In the product the index k starts with 1 because gcd(0, 2 n) = 2n, and it stops at n − 1 for all n ≥ 2
because gcd(n, 2 n) = n. One has to omit the even k values > 0 which leads to (for n = 1 one takes
only l = 0 in the following product)
C(n, x) =
⌊ n−2
2 ⌋
Y
l=0
gcd(2 l+1,n)=1
x − 2 cos
π (2 l + 1)
n
.
(20)
Proof: This is a simple consequence of the mentioned known results for Ψ(2 n, x2 ). It implies, en passant,
a formula for the known degree δ(n) (see A055034) in terms of the number of factors of this product.
The following proposition lists the nonnegative and negative zeros of C for prime n. A vanishing zero
(which will be counted later as positive) appears only for n = 2.
Proposition 4: Non-negative and negative zeros of C(p, x)
i) If n = 2 then C(2, 0) = 0 , and this is the only case with a vanishing zero.
ii) If n is an odd prime 1 (mod 4) then the number
of positive and negative zeros coincides, and this
k+1 2 cos π k , with k values 1, 2, ..., n−1 = δ(n).
.
These
zeros
are
(−1)
number is n−1
4
n
2
iii) If n is an odd prime 3 (mod 4) then the number of positive zeros exceeds the number
of negative
k+1 2 cos π k with k values
ones by one, and the number of negative ones is n−3
.
These
zeros
are
(−1)
4
n
n−1
1, 2, ..., n−3
2 , and an extra positive zero appears for k = 2 = δ(n) .
Proof:
i) Vanishing zeros require n ≡ 2 (mod 4) from nk = 12 and k odd in eq. (20). A vanishing zero appears
for n = 2. For other such n values, n = 4 K + 2, K ≥ 1, the odd k = n2 > 1 divides n, hence it does
not appear in the product of eq. (20).
p−1
For odd primes n = p = 2 q + 1 the product in eq. (20) is unrestricted, and there are (q−1) + 1 =
2
ϕ(2 p)
factors, in accordance with the degree δ(p) =
. The cos function will produce negative zeros for
2
1
2l + 1
> , i.e., for 2 l ≥ q.
p
2
ii) n = p = 4 K + 1 means that l = 0, 1, ..., K − 1 lead to positive zeros, and l = K, K + 1, ..., 2K − 1
n−1
to negative ones. In both cases there are K =
such zeros. For the l values leading to negative
4
π
π
(2 l + 1) = − cos
(p − (2 l + 1)) . In this way even multiples
zeros one can use the formula cos
p
p
π
of
appear, and one produces, counted backwards, beginning with the largest l value, the new values
p
2, 4, ..., 2 K. This proves ii). As a test consider n = 13, K = 3: the 2 l + 1 values for the positive
zeros are 1, 3, 5, and the ones for the negative ones are 7, 9, 11. The latter become (we use underlining
to indicate that they come with a minus sign in the final formula) 6, 4, 2. Rearranged these values are
1, 2, 3, 4, 5, 6, as given in ii).
iii) For n = p = 4 K + 3 a similar analysis leads to l = 0, 1, ..., K, and l = K + 1, ..., 2 K for positive,
p− 3
, and the
and negative zeros, respectively. This shows that the number of negative zeros is K =
4
number of positive ones exceeds the one for negative zeros by one. Again, the odd values leading to negative zeros are transformed to even ones 2, 4, ..., 2 K with a minus sign in front of 2 cos. Take, e.g., n = 19,
K = 4, with the new k values of iii) 1, 2, 3, 4, 5, 6, 7, 8, 9.
For general values n one can also find the number of positive and negative zeros of C(n, x). In the
non-prime or non-power-of-2 case the gcd restriction in the product excludes certain l values. Therefore
one has to eliminate from the unrestricted 2 l + 1 values in the product of eq. (20) all odd multiples
9
for each odd prime dividing n. Multiples of 2 are irrelevant for even n , therefore only the odd primes
dividing n are of interest. This shows that the prime factors of the squarefree kernel of n, denoted here by
sqf k(n), i.e., the largest squarefree number dividing n (see A007947) will be of interest. This squarefree
kernel is also known as radical of n, denoted by rad(n). We will denote the set of primes of this kernel
by sqf kset(n), E.g., sqf kset(23 · 32 · 11) = {2, 3, 11}, sqf k(23 · 32 · 11) = 66 (strip off all exponents in
the prime number factorization of n). Because one may encounter multiple counting (e.g., 15 is hit by
the multiples of 3 as well as 5 for any n which has in its squarefree kernel set the primes 3 and 5, and
which is larger than 16) one can employ PIE, the principle of inclusion-exclusion, e.g., [5], p. 134, to get
a correct counting.
Proposition 5: Number of positive and negative zeros of C(n, x)
i) If n = 2m , m ≥ 1, one has for m = 1 a vanishing zero (see proposition 4 i)). For m ≥ 2 the number
n
of positive zeros, which is , coincides with the one for negative ones. This is in accordance with the
4
π
n
n
m
m−1
degree δ(2 ) = 2
= . These zeros lie symmetric: ± 2 cos
(2 l + 1) , l = 0, 1, ..., − 1.
2
n
4
ii) If n = 1 then there is only the negative zero −2. If n ≥ 3 is not a power of 2 then the number of
positive zeros, called δ+ (n), is
M (n)
δ+ (n) = K(n) +
X
X
r
(−1)
r=1
<i1 ,i2 ,...,ir >
1
2
L(n)
− 1
,
p i1 · · · p ir
(21)
where M (n) (sometimes called ω(n)) is in the odd n case the number of elements (cardinality) of the
set sqf kset(n), denoted by |sqf kset(n)|, (see A001221). M (n) is for even n the number of odd primes
in sqf kset(n), i.e., |sqf kset(n)|
− 1.
This is because multiples of 2 are irrelevant here. The sum
M
< i1 , i2 , ..., ir > extends over the
combinations 1 ≤ i1 < i2 < ... < ir ≤ M . Only the odd
r
primes from the set sqf kset(n) enter. The values K(n) and L(n) depend on the parity of n and they are
given by
n−2
α) n even :
K(n) =
, L(n) = 2 K(n) + 1 ,
(22)
4
n−3
n−5
,
L(n) = 2 K(n) + 1 =
,
(23)
β1) n odd, 1(mod 4) : K(n) =
4
2
n−1
n−3
,
L(n) = 2 K(n) + 1 =
.
(24)
β2) n odd, 3(mod 4) : K(n) =
4
2
Note that also negative values for the floor function may appear. In this way the pure prime case from
proposition 4 is also included.
iii) If n ≥ 3 is not a power of 2 then the number of negative zeros, called δ− (n), is
M (n)
δ− (n) = N (n) +
X
r=1
r
(−1)
X
<i1 ,i2 ,...,ir >
1
2
1
P (n)
L(n)
− 1
−
− 1
, (25)
p i1 · · · p ir
2 p i1 · · · p ir
where M (n) and L(n) are as above, and
n−2
−
α) n even :
N (n) =
2
n−1
,
β1) n odd, 1(mod 4) : N (n) =
4
β2) n odd, 3(mod 4) : N (n) = K(n) =
10
n−2
4
, P (n) = n − 1 ,
P (n) = n − 2,
n−3
,
4
P (n) = n − 2.
(26)
(27)
(28)
Proof:
n−2
i) All l = 0, ..,
= 2m−1 − 1 values contribute to the product of eq. (20), compatible with the
2
n
ϕ(2m )
= 2m−1 = . The negative zeros appear for 2 l + 1 = 2 2m−2 + 1, ..., n − 1, hence
degree δ(2m ) =
2
2
there are n4 of them. With the cos formula given in the proof of proposition 4 ii) with p → n, they can
be rewritten in the notation, also used in the above context, when read backwards, as 1, 3, ..., 2m−1 − 1.
Therefore one obtains the same cos arguments like for the n4 positive zeros, only the overall sign is
different.
ii) For n ≥ 3, not a power of 2, one uses PIE to count the positive zeros. For this consider the odd
multiples of some odd number a, say (2 k + 1) a, (only such multiples come up as 2 l + 1 values in
the product (20)) up to kmax such that all positive zeros are reached.If the largest 2 l + 1 value in the
1 n̄
− 1 , and there are kmax + 1 of these
product which leads to a positive zero is n̄, then kmax =
2 a
odd multiples of a. In the following application of PIE a will be taken as any odd prime or product of
odd primes from the set sqf kset(n).
α) Case even n, not a power of 2: The counting of the positive
starts at level r = 0 with all
zeros
n−2
+ 1 of them (the l values are
positive ones in the unrestricted product eq. (20). There are
4
n−2
0, 1, ...,
). In the next step, r = 1, all odd multiples of every odd prime pi1 in sqf kset(n) not
4
n−2
+ 1 are discarded (there are M (n) such primes, where M (n) := |sqf kset(n)| − 1).
exceeding 2
4
M (n)
X 1 L(n)
− 1
+ 1 with L(n) given in the proposition
This leads to a subtraction of the form −
2
p i1
i1 =1
ii) α). Now double subtractions may have appeared and one adds, at step r = 2, all odd multiples of
the product of two odd primes from sqf kset(n), call them pi1 pi2 with i1 < i2 (because
no problem
of
X
L(n)
1
− 1
+ 1.
over-subtraction appeared for i1 = i2 ). This leads to the term +
2 p i1 p i2
1 ≤ i1 <i2 ≤ M (n)
Now one has to correct for triple products in the step r = 3 , etc., up to level r = M (n). This leads
to the formula given in part ii) α) if one observes that the +1 term in each r-sum becomes a M (n)
r
term because of the number of terms of this sum. Then, because the alternating sum over row M (n) in
M (n) = 0), all these terms add up to −1 (the missing
Pascal’s triangle (see A007318) vanishes ((1 − 1)
n−2
r = 0 term). Therefore the formula starts with
, the K(n) value given in ii) α), and only the
4
floor term remains for each r-sum .
β1) The counting of the positive zeros in the odd n case distinguishes the two cases 1 (mod 4) and
3 (mod 4). Here we discuss the former one. The procedure is the same as the one given for case α).
One has just to determine the boundary values for the lmax value leading still to a positive zero. This is
n−5
. Hence there are lmax + 1 such l values to start with at level
lmax = K − 1, if n = 4 K + 1, i.e.,
4
r = 0. Again the extra +1 can later be taken as r = 0 term M 0(n) for the vanishing alternating sum
over row No. M (n) in Pascal’s triangle. Here M (n) := |sqf kset(n)|. Therefore the PIE formula starts
n−5
n−3
with the K(n) given in β1) of the proposition. The L(n) of the PIE sum is now 2
+ 1 =
,
4
2
the largest 2 l + 1 value leading to a positive zero.
β2) In the case M (n) := |sqf kset(n)|, n = 4 K + 3 the maximal 2 l + 1 value leading to a positive
n−1
zero, the L(n) in the formula, is 2 K + 1 =
, coming from the largest l value which is K. This K
2
is the K(n) in the formula claimed for this case β2).
11
iii) For the number of negative zeros the counting is done by finding all the odd multiples of odd primes,
or products of them, which cover the 2 l + 1 range for the negative zeros. This is the difference of the
number of such multiples for the whole range and the range for the positive zeros.
α) If n is even, not a power of 2, the whole range is determined by lmax + 1 = n − 1. From the above
ii) α) case one knows the corresponding
upper bound for the positive zeros, therefore the PIE formula
n−2
n−2
has N (n) = (
+ 1) − (
+ 1), the number of factors in the unrestricted product which
2
4
lead to negative zeros,
which is the value claimed in the proposition iii) α). Accordingly, P (n) = n − 1
n−2
+ 1 as given in eq. (26).
and L(n) = 2 4
β1) In this n ≡ 1 (mod 4) case the maximal number 2 l + 1 in the product is n − 2, determining
n−3
. There are
P (n). The corresponding number L(n) for the positive zeros is taken from above as
2
n−3
n−5
n−1
N (n) =
+ 1 −
+ 1 =
unrestricted factors in the product with negative
2
4
4
zeros.
β2) In this n ≡ 3 (mod 4) case one has also P (n) = n − 2 like
for any odd
n. L(n) is taken
from the
n−3
n−3
n−3
n−1
, and N (n) is obtained from
+ 1 −
+ 1 =
.
corresponding ii) case as
2
2
4
4
We give the first entries of the δ+ (n) and δ− (n) sequences. Remember that the vanishing zero for n = 2
is here counted as positive. The A-numbers given for δ− (2 k) and δ− (2 k) are conjectured.
n even:
δ+ (2 k), k = 1, 2, ...; A055034 : [1, 1, 1, 2, 2, 2, 3, 4, 3, 4, 5, 4, 6, 6, 4, 8, 8, 6, 9, 8, 6, 10, ...],
δ− (2 k), k = 1, 2, ...; A055034 with first entry 0 : [0, 1, 1, 2, 2, 2, 3, 4, 3, 4, 5, 4, 6, 6, 4, 8, 8, 6, 9, 8, 6, 10, ...],
n odd, 1(mod 4):
δ+ (4 k + 1), k = 0, 1, ...; [1, 5, 9, 13, 17, 21, 25, 29, 33, 37, 41, 45, 49, 53, 57, 61, 65, 69, 73, ...],
δ− (4 k + 1), k = 0, 1, ...; [1, 1, 2, 3, 4, 4, 5, 7, 6, 9, 10, 6, 11, 13, 10, 15, 12, 12, 18, 16, 14, ...],
n odd, 3(mod 4):
δ+ (4 k + 3), k = 0, 1, ...; [1, 2, 3, 2, 5, 6, 5, 8, 6, 6, 11, 12, 8, 10, 15, 10, 17, 18, 10, 20, ...],
δ− (4 k + 3), k = 0, 1, ...; [0, 1, 2, 2, 4, 5, 4, 7, 6, 6, 10, 11, 8, 10, 14, 8, 16, 17, 10, 19, 20, ...].
The start of the sequences for odd n is therefore
δ+ (2 k + 1), k = 0, 1, ...; [0, 1, 1, 2, 1, 3, 3, 2, 4, 5, 2, 6, 5, 5, 7, 8, 4, 6, 9, 6, 10, 11, ...],
δ− (2 k + 1), k = 0, 1, ...; [1, 0, 1, 1, 2, 2, 3, 2, 4, 4, 4, 5, 5, 4, 7, 7, 6, 6, 9, 6, 10, 10, 6, 11, ...].
Finally, combining for all n:
δ+ (n), n = 1, 2, ...; A193376 : [0, 1, 1, 1, 1, 1, 2, 2, 1, 2, 3, 2, 3, 3, 2, 4, 4, 3, 5, 4, 2, 5, 6, 4, 5, ...],
δ− (n), n = 1, 2, ...; A193377 : [1, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 2, 4, 4, 3, 4, 4, 4, 5, 5, 4, 5, 6, ...].
On can check that δ(n) − (δ+ (n) + δ− (n)) vanishes in each case.
Observe that from these examples it seems that δ+ (2 k) = δ− (2 k) = A055034(k) = δ(k) for k ≥ 2.
ϕ(4 k)
= 2 ϕ(2 k) for these k values. The latter eq., which
This is compatible with δ(2 k) = 2 δ(k), i.e.,
2
holds also for k = 1, can be checked by considering the two cases k = 2e(1) · (odd number) and k odd.
ϕ(2)
).
Recall that δ(2) = δ(1) = 1, because δ(1) was special (it is not
2
Zeros of C(n, x) in the power basis < ρ(n)0 = 1, ρ(n)1 , ..., ρ(n)δ(n)−1 >
As will be explained in the next section, the minimal polynomial C(n, x) with degree δ(n) leads to the
splitting field Q(ρ(n)) which is a simple field extension of the rational field Q, and the degree of Q(ρ(n))
12
over Q, the dimension of Q(ρ(n)), considered as a vector space over Q, is just the degree of C. In standard
notation [Q(ρ(n)) : Q] = δ(n). Therefore it is interesting to write the zeros of C
in the powerbasis
π
(n)
(2 l + 1) , l ∈
for this Q-vector space. This task is accomplished by using for the zeros ξ̃l = cos
n
n−2
(n)
{0, 1, ...,
} with gcd(2 l + 1, n) = 1 (see eq. (20)),the formula ξ̃l = t̂(2 l + 1, ρ(n)) = 2 T (2 l +
2
ρ(n)
) (compare with eq. (2)). Remember that the integer coefficient array for these t̂-polynomials is
1,
2
shown in A127672. Of course, one has to reduce this integer polynomials using C(n, ρ(n)) = 0, i.e., one
has to replace ρ(n)δ(n) (and higher powers) by an integer polynomial of lower degree. This can be done on
a computer and Table 4 has been found with the help of Maple13 [18]. For example, the zeros of C(15, x)
(15)
(15)
(15)
(15)
are ξ̃1
= ρ(15), ξ̃2
= −2 + 3 ρ(15) + ρ(15)2 − ρ(15)3 , ξ˜3
= −1 − 4 ρ(15) + ρ(15)3 , ξ̃4
=
2
4
3
2
2 − ρ(15) . Here the reduction has been performed with ρ(15) → −ρ(15) + 4 ρ(15) + 4 ρ(15) − 1.
One can check these zeros which have approximate values 1.956295201, 0.209056928, −1.338261216 and
−1.827090913, respectively. We will call the zeros with increasing value ξ1 , ξ2 , ..., ξδ(n) .
Our next objective is to compute C(n, 0) (the sign and the number multiplying x0 , also called the absolute
term) and relate it first to the cyclotomic polynomial cy(n, x) (the minimal polynomial of the complex
2πi
algebraic number e n , see e.g., [30] or [9], p. 149, Exercise 50 a and b), evaluated at x = −1, which is
n−1
Y
ϕ(n)
cy(n, −1) = (−1)
k
(1 + e2π i n ) , n = 2, 3, ...
(29)
k=1
gcd(k,n)=1
We used the fact that the product in the definition of cy(n, x) has ϕ(n) factors (the degree as minimal
polynomial). For n = 1 one has cy(1, −1) = −2, which fits this formula if one defines ϕ(1) := 1 and
πi
takes the undefined product as 1. One may rewrite this for n ≥ 2 by extracting e n , using for the sum
of the ϕ(n) terms the formula
2
n2
n−1
X
k =
k=1
gcd(k,n)=1
ϕ(n)
, n = 2, 3, ...
n
(30)
This is a formula listed by R. Zumkeller under A023022 which is found in[2], p. 48, exercises 15 and 16,
written such that both sides depend only on the distinct primes in the prime number factorization of n.
M (n)
Y
X
1
n
ϕ(n)
1−
=
=
µ(d) , if n
For the r.h.s. there is the well known formula [2], p. 27,
n
pj
d
j=1
d|n
has the distinct prime factors pj , j = 1, 2, ..., M (n) = A001221(n) = |sqf kset(n)| (see above for the
notion ‘squarefree kernel’). The Möbius function µ entered here (see A008683). The PIE proof of eq.
(30), given in the appendix A, uses this form of ϕ(n). If one uses also ei π = −1 the rewritten eq. (29)
becomes
n−1
Y
ϕ(n)
k
2
, n = 2, 3, ....
(31)
2 cos π
cy(n, −1) = (−1)
n
k=1
gcd(k,n)=1
This can also be related to a special Sylvester sequence. In general the complex cyclotomic Sylvesternumbers Sy(a, b; n) are related to the sequence with three term recurrence fn = √a fn−1 + b fn−2 ,
2
2
which has characteristic
√ polynomial x − a x − b with zeros α ≡ α(a, b) = (a + a + 4 b)/2 and
β ≡ β(a, b) = (a − a2 + 4 b)/2. The definition is (see e.g., [19])
Sy(a, b; n) :=
n−1
Y
k=1
gcd(k,n)=1
k
(α − β e2π i n ) , n = 2, 3, ...
13
(32)
For n = 1 one takes Sy(a, b; 1) = α − β. For our purpose the values are (a, b) = (0, −1), i.e.,
(α, β) = (i, −i), and
Sy(0, −1; n) := (−1)
n−1
Y
ϕ(n)
2
k
(1 + e2π i n ) , n = 2, 3, ...,
(33)
k=1
gcd(k,n)=1
with Sy(0, −1; 1) = 2 i. This can be rewritten, using again eq. (30) and ei π = −1, as
ϕ(n)
Sy(0, −1; n) = (−1)
n−1
Y
k=1
gcd(k,n)=1
Therefore
Sy(0, −1; n) := (−1)
ϕ(n)
2
k
2 cos π
n
cy(n, −1) , n = 2, 3, ....
One could also include the n = 1 case, with ϕ(1) := 1 if one takes
After these preliminaries back to the number C(n, 0).
Proposition 6: C(2 m, 0), m ∈ N
C(2 m, 0) = (−1)
ϕ(4 m)
2
+ ϕ(2 m)
, n = 2, 3, ....
Sy(0, −1; 2 m) = (−1)
ϕ(4 m)
2
+
1
iϕ(n)
ϕ(2 m)
2
(34)
(35)
instead of the prefactor.
cy(2 m, −1) , m ∈ N .
(36)
ϕ(4 m)
, the degree
2
δ(2 m). Note that the restriction in the Sy(0, −1; 2 m) product is gcd(k, 2 m) = 1, while in the C(2 m, 0)
product it is gcd(k, 4 m) = 1, but this just says that only odd ks can contribute in both cases, and those
odd numbers ≤ 2 m − 1 dividing 4 m or 2 m have to be omitted. Both restrictions exclude the same odd
numbers.
Corollary 1: C(2 p, 0), p an odd a prime
Proof: Use eqs. (19) and (35). The number of factors in eq. (19) with n = 2 m is
For odd prime p : C(2 p, 0) = (−1)
p−1
2
p.
(37)
Proof:
This follows immediately from eqs. (19) and (31) if one uses the known formulae cy(2 p, x) = cy(p, −x),
cy(p,
Y 1) = p, nand evaluates the ϕ functions to obtain the sign. This formula follows from cy(n, x) =
(xd − 1)µ( d ) where the multiplicative Möbius function (with µ(p) = −1) enters (see, e.g., [9], exercise
d|n
50 b., solution p.506).
Proposition 7: C(2m , x), m ∈ N0
C(2m , x) = t̂(2m−1 , x) , m ∈ N,
(38)
with the integer polynomials t̂ defined in eqs. (2) and (3) in terms of Chebyshev T - or S−polynomials.
Proof:
Use eq. (16)
to the link [13], where it was shown in eq. (8) that
and the
Notexadded
m+1 x
2m−1
m−1
Ψ 2
,
2
= 2T 2
,
=: t̂(2m−1 , x), for m = 1, 2, ... . It is clear that C(1, x) = x + 2.
2
2
This proposition will later be generalized in theorem 1A.
Corollary 2: C(2m , 0), m ∈ N0
From the proposition 7 and the known fact that T (2 n, 0) = (−1)n follows that
C(1, 0) = +2, C(2, 0) = 0, C(22 , 0) = −2, C(2m , 0) = +2, m ≥ 3 .
(39)
Note that a standard formula for the cyclotomic polynomials cy(2m , x) which involves divisors (see e.g.,
[9], p. 149, Exercise 50 b, with Ψm (x) ≡ cy(m, x)) leads to undetermined expressions for cy(2m , −1) if
m ≥ 1.
14
For the following theorem 1A on a formula for C(n, x) for even n, based on the divisor product representation (dpr) of numbers [14], we need two lemmata.
Lemma 3: Pairing in dpr(2 m)
In the divisor product representation of an even number 2 m, m ∈ N, for each a-factor in the numerator
(resp. denominator) one finds exactly one a-factor in the denominator (resp. numerator) with argument
ratio either 1 : 2 or 2 : 1.
Proof:
Due to the ∗-multiplication property of dprs (see the theorem in [14]) it is sufficient to prove this
lemma for dpr(2 p1 · · · pN ) with odd primes p1 , ..., pN . Indeed, if the prime factorization of 2 m is
2k+1 pe11 · · · peNN , with k ∈ N0 and odd primes pj and positive exponents ej , for j = 1, ..., N , then
dpr(2 m) = (2k pe11 −1 · · · peNN −1 ) ∗ dpr(2 p1 · · · pN ) and the pairing property of the latter dpr will be preserved after multiplication of each argument. Now the proof of the pairing property for dpr(2 p1 · · · pN )
becomes elementary, once one splits the products in eq. (7) of [14] (note that N in this reference is now
N + 1) into those a-factors which are even (i.e., contain the even prime 2) and those which are odd (i.e.,
those composed of only odd primes). Indicate a-factors with odd arguments, being the product of k odd
primes, by ao (.(k).). Then any a-factor in the numerator (resp. denominator) with some even argument
2 n is found in the denominator (resp. numerator) exactly once with the argument
n. This is because for
N
}, say, a(2 pi1 · · · piN−2 j ),
each factor in the numerator product Π a(2 .(N − 2 j).), with j ∈ {1, ...,
2
there is exactly one factor in the denominator product Π ao (.(N − 2 j).), namely the one which uses
the same N − 2 j odd primes a(pi1 · · · piN−2 j ). Similarly, for each a-factor in the denominator product
N −1
Π a(2 .(N − (2 j + 1)).), with j ∈ {0, ...,
}, there is exactly one factor in the numerator product
2
Π ao (.(N − (2 j + 1)).). The first factor in the numerator, a(2 p1 · · · pN ) (j = 0) is paired with the single
odd argument a-factor of the first product in the denominator Π ao (.(N ).) = a(p1 · · · pN ) .
Note, that the example for the primorial A002110(5) = 2310, given in [14], table 2 shows that the
pairing occurs in general not between numerator and denominator factors at the same position, given the
ordering prescription indicated by O (falling arguments) in eq. (7) in [14]. See the fifth position in the
N = 5 case there.
Recall also, for later purposes, that because dpr(2 m) has the same number of factors in the numerator
and in the denominator, viz 2N ( this balance holds true for all dpr(n), n ≥ 2, due to proposition 5 in
[14]), after the split into even and odd arguments the number of pairs with even argument in the numerator matches the one with even argument in the denominator. This will later lead to a balanced formula for
C(2 m, x) in terms of t̂-polynomials (the same number of t̂s in the numerator and in the denominator, viz
2N −1 ).
Lemma 4:
n
n x
S( n2 + 1, x)
t̂( n2 + 1, x) − t̂( n2 − 1, x)
=
=
t̂
,
, x , for n ≡ 0 (mod 4) .
=
2
T
S( n4 − 1, x)
4 2
4
t̂( n4 + 1, x) − t̂( n4 − 1, x)
(40)
This identity involves the integer polynomials t̂ introduced earlier in eqs. (2) and (3).
Proof:
Use the known identity for Chebyshev polynomials, [20], p. 261, first line, specialized, T (m + 1, x) −
x
x
T (m − 1, x) = 2 (x2 − 1) U (m − 1, x), written for x replaced by , with S(m − 1, x) = U (m − 1, )
2
2
and the definition of the t̂-polynomials from eq. (2). For n ≡ 0 (mod 4) this identity can be used
x
in the numerator as well as in the denominator. The n−independent factor ( )2 − 1 drops out if
2
x 6= ±2. However, the lemma holds also for these x−values as can be seen after applying l’Hôpital’s
rule, using the well known identity T ′ (x, n) = n U (n − 1, x). In the second to last step the well known
15
x
S(m − 1, x) = S(2 m − 1, x) has been employed.
identity, [20] p. 260, last line, written as 2 T m,
2
Theorem 1A: C(2 m, x), m ≥ 1, in terms of t̂-polynomials
With the prime number factorization of 2 m = 2k pe11 · · · peNN with k ∈ N, odd primes pj and positive
exponents ej , j = 1, ..., N , one has
C(2 m, x) =
t̂(2k−1 pe11 · · · peNN , x) Π t̂(2k−1 ∗ (N − 2)., x) Π tˆ(2k−1 ∗ (N − 4)., x) · · ·
.
Π t̂(2k−1 ∗ (N − 1)., x) Π t̂(2k−1 ∗ (N − 3)., x) · · ·
(41)
with t̂(2k−1 ∗ (0)., x) = t̂(2k−1 pe11 −1 pe22 −1 · · · peNN −1 , x), and the products Π t̂(2k−1 ∗ (K)., x) are over the
N factors with K primes from the set {p , p , ..., p } multiplied by pe1 −1 pe2 −1 · · · peN −1 , i.e., K1 2
N
1
2
N
K
eN −1
products of the form t̂(2k−1 pe11 −1 pe22 −1 · · · pN
pi1 pi2 · · · piK , x). We used ∗ (k). instead of .(k). for the
indices of the t̂ polynomials to remind one of the extra factor (without the powers of 2 because the even
prime has been extracted everywhere) due to the ∗-multiplication.
Before we give the proof an example will illustrate this theorem.
Example 3: k = 1, N = 2
C(1350, x) = C(2 · 33 52 , x) =
t̂(33 52 , x) t̂(32 5, x)
t̂(675, x) tˆ(45, x)
=
.
t̂(33 5, x) t̂(32 52 , x)
t̂(135, x) tˆ(225, x)
Proof:
(42)
This is based on [14] applied for the minimal polynomials Ψ 4 m, x2 (see the definition for C(2 m, x)
given in eq. (16)). Let 2k+1 pe11 · · · peNN be the prime number factorization for 4 m. Only the squarefree
kernel 2 p1 · · · pN is important due to the theorem in [14] for the divisor product representations (dprs)
and the ∗-multiplication. This means that one has to multiply after the computation of Ψ(2 p1 · · · pN ),
using proposition 1 of [14], the first argument, the index in conventional notation, of each factor t(ni , x)
in the the numerator, and of each t(mi , x) in the denominator (see [14], eq. (1)) with the number
2k pe11 −1 pe22 −1 · · · peNN −1 . Here the pairing lemma 3 is crucial which carries over to the indices of the tfactors in the numerator and denominator of [14], eq. (1). Because there is always at least one factor of
2 in the number which multiplies every t index after the ∗-multiplication, the pairing will occur always
between
which are 0 (mod 4) and 0 (mod 2). Then only the n even alternative in the definition
xindices
from eq. (2) of [14], rewritten here as eq. (18), is relevant. Observe that the prefactors
of t n,
2
n
n
x
x
x
are not of interest, provided we take 2 T
− T
+ 1,
− 1,
in this definition of t n,
2
2
2
2
2
n
n
which is the monic integer polynomial t̂
+ 1, x − t̂
− 1, x . See eq. (18), and the discussion in
2
2
connection with lemma
C(2m, x)is written here as a rational function of monic integer polynomials.
x2. .
n x
t
(either in the numerator or denominator, depending on where the
For each replaced t n,
,
2
2 2
n
larger index appears), one can apply lemma 4. This is how for each such t-quotient one obtains t̂
,x .
4
Here the fact that the larger index of a pair is always equivalent to 0 (mod 4) is important. From the
structure of the numerator and denominator of the original dpr(2 m) with the separation of the even
and odd a-arguments (which carries over to the t-indices), discussed in the proof of lemma 3, one now
finds the numerators and denominators of the theorem. Just search the products for the even t-indices.
E.g., Π t̂(2k−1 ∗ (N − 2)., x) in the numerator of the theorem originates from the quotient of products
Π t(2 .(N − 2)., x2 )
before the ∗-multiplication has to be applied. This leads, with eq. (18) and lemma
Π to (.(N − 2)., x2 )
k+1
2
k e1 −1 pe2 −1 · · · peN −1 ,
4, to Π t̂
∗ (N − 2)., x after multiplying each t̂ index in the product with 2 p1
2
N
4
where the ∗ reminds one to multiply each index with this number divided by 2k .
16
As announced at the end of the proof of lemma 3 the number of t̂ polynomials in the numerator is
the same as the one for the denominator, viz 2N −1 , which is due to the sum in the Pascal-triangle A007318
row No. N over even, resp. odd numbered positions.
Note that this theorem 1A, evaluated for x = 0 leads in general to undetermined expressions, remembern
ing that t̂(n, 0) = 0 if n is odd, and 2 (−1) 2 if n is even. However, a correct evaluation (using l’Hôpital’s
rule) has to reproduce the result known from corollary 2.
The following factorization of the monic integer t̂-polynomials is related to theorem 1A.
Theorem 1B: Factorization of t̂-polynomials in terms of the minimal C-polynomials
Y
Y
t̂(n, x) =
C(2 n/d, x) =
C(2k+1 d, x) ,
d|op(n)
(43)
d|op(n)
with op(n) = A000265(n), the odd part of n, and 2k is the largest power of 2 dividing n. The exponents
are k = k(n) = A007814(n), k ∈ N0 .
Before the proof we give an example.
Example 4: n = 10
−2 + 25 x2 − 50 x4 + 35 x6 − 10 x8 + x10 = t̂(10, x) = C(20, x) C(4, x) =
(x8 − 8 x6 + 19 x4 − 12 x2 + 1) (x2 − 2) .
(44)
Proof: This is modeled after a similar proof in [30].
It is clear that both sides are monic integer polynomials. See the definition of t̂ in eqs. (2), (3) and
proposition 2.
In order to check the degree we consider the even and odd n cases separately. If n = 2k(n) op(n)
X
X ϕ(2k+2 d)
from the known degree δ of the
δ(2k+1 d) =
with k(n) = A007814(n) ≥ 1 then
2
d|op(n)
d|op(n)
C-polynomials, here for index > 1. Due to well known properties of the Euler totient function (see e.g.,
2k(n)+2 − 2k(n)+1
[2], Theorem 2.5, (a) and (c), p. 28 and Theorem 2.2., p. 26) this leads to
op(n) = n,
2
X ϕ(4 d)
X
the degree of t̂(n, x). In the odd case, if n = op(n) then
=
d = n, again the degree of
2
d|n
d|n
t̂(n, x).
π
(n)
For the proof one compares the zeros of both sides. The zeros of t̂(n, x) are x̂l = 2 cos (2 l + 1)
2n
for l = 0, 1, ..., n − 1. This is known from the zeros of the Chebyshev T -polynomials.
The zeros
π
for
of the C(2k+1 d, x)-polynomials are used in the form given by eq. (20): 2 cos (2 l′ + 1) k+1
2
d
l′ = 0, 1, ..., 2k d − 1, where gcd(2 l′ + 1, 2k+1 d) = 1, i.e., gcd(2 l′ + 1, d) = 1. Because the degrees
match, it is sufficient to show that each zero of t̂(n, x) occurs on the r.h.s.. For n = 2k (n) op(n), with
k(n) = A007814(n), consider gcd(2 l + 1, 2 n) = gcd(2 l + 1, 2k+1 op(n)) = gcd(2 l + 1, op(n)) = g (some
2l + 1
=
odd number). Hence 2 l+1 = (2 l′ +1) g, with some l′ and op(n) = d g, i.e., d|op(n). Therefore
2n
′
2l + 1
π
appears on the r.h.s. as one of the C(2k+1 d, x) zeros.
. Thus the t̂(n, x) zero 2 cos (2 l + 1)
k+1
2
d
2n
.
Remark 3: Derivation of Theorem 1A from Theorem 1B
Theorem 1B can be used as recurrence for the C-polynomials in terms of the t̂-polynomials. This is
2π
(see A181875/A181876 for their
similar to the case treated in [30] for the minimal polynomials Ψ of
n
17
coefficients). The solution of this recurrence has been given in [14]. Indeed, theorem 1A has been derived
above from this solution. Originally we found theorem 1B starting from theorem 1A building up iteratively
a formula for t̂(2k−1 pe11 · · · peNN , x) in terms of the C-polynomials. We give this formula as a corollary.
Corollary 3: t̂-polynomials in terms of C-polynomials
eN
e2
e1 Y
Y
Y
···
C(2k pq11 pq22 · · · pqnN , x) , k = N, N ∈ N0 .
t̂(2k−1 pe11 · · · peNN , x) =
q1 =0 q2 =0
(45)
qN =0
In order to find a simplified expression for C(2 m + 1, x), m ∈ N0 we need the following lemma in order
t(even, x2 )
to rewrite the quotient
, given the definition eq. (18).
t(odd, x2 )
Lemma 5:
for n = 2 M + 1, M ∈ N0 one has
x
x
t 2 n, x2
1 (x2 − 4) S(n − 1, x)
M −n t̂(n + 1, 2 ) − t̂(n − 1, 2 )
√
=
=
2
2M +1 (x − 2) S(2 M, 2 + x )
t n, x2
t̂(M + 1, x2 ) − t̂(M, x2 )
x + 2 S(n − 1, x)
√
.
=
2M +1 S(n − 1, 2 + x )
(46)
(47)
Here new important monic integer polynomials enter the stage:
Definition 1: q-polynomials
With n ∈ N0 define
q(n, x) :=
S(2 n, x)
√
.
S(2 n, 2 + x )
(48)
That this defines indeed monic integer polynomials of degree n is shown by the next lemma.
Lemma 6:
q(n, x) = (−1)n S(2 n ,
√
2 − x ) = S(n, x) − S(n − 1, x) .
(49)
Proof:
It is known from the o.g.f. of the Chebyshev
S-polynomials that the bisection yields S(2 n, y) = S(n, y 2 −
√
2
2) + S(n
S(2 n , x + 2 ) = S(n, x) + S(n − 1, x). With x replaced by −x one has
√ − 1, y − 2), or
n
S(2 n, 2 − x ) = (−1) (S(n, x) − S(n − 1, x)) which explains the second equation of the lemma.
Therefore one has to prove S(2 n, x) = S(n, x)2 − S(n − 1, x)2 . This identity can, for example, be
proved using the o.g.f. for the square of the S-polynomials (for their coefficient table see A181878,
also for the paper [15] given there as a link). This computation was based on the Binet-de Moivre
1
1+ z
. The o.g.f.
formula for the S−polynomials. This o.g.f. was found to be
1 − z 1 + (2 − x2 ) z + z 2
2
2
for {S(2 n, x)}∞
n=0 is (1 + z)/(1 + (2 − x ) z + z ) (from the bisection). Because the o.g.f. for
2
2
S(n, x) − S(n − 1, x) is (1 + z) times the one for S(n, x)2 the completion of the proof is then obvious.
Remark 4 : O.g.f. and coefficient array for the q-polynomials
∞
X
q(n, x) z n = (1 − z)/(1 − x z + z 2 ) from the known o.g.f.
From lemma 6 the o.g.f. Q(z, x) :=
n=0
of the S-polynomials. This shows that the q(n, x) coefficients constitute aRiordan-array (infinite
lower
1− x
x
triangular ordinary convolution matrix), which is in standard notation
,
, meaning
2
2
m 1 + x 1 + x
x
1− x
. This is the triangle A130777
that the o.g.f. of the column No. m sequence is
2
1+ x
1 + x2
18
where moreinformation
can be found. For example, the explicit form for the coefficients is Q(n, m) =
n+m
n−m+1
2
(−1) 2
if n ≥ m ≥ 0 and 0 otherwise.
m
The pairing lemma 3 will also be used in the proof of the following theorem.
Theorem 2A: C(n, x), n ≥ 3, odd, in terms of q-polynomials
With the prime number factorization of n = pe11 · · · peNN with odd primes p1 , ..., pN and positive exponents
ej , j = 1, ..., N , one has
Q
Q
n/(pi1 pi2 )−1
n/(pi1 pi2 pi3 pi4 )−1
,
x
q
,
x
,
x
···
q
q n−1
i1 <i2
i1 <i2 <i3 <i4
2
2
2
Q
.
(50)
C(n, x) = Q
n/pi1 −1
n/(pi1 pi2 pi3 )−1
q
,
x
,
x
·
·
·
q
i1
i1 <i2 <i3
2
2
Here each index ij of the products runs from 1 to N .
Before we give the proof an examples will illustrates this theorem.
Example 5: n = 32 · 5 = 45
C(45, x) = C(32 · 5, x) =
q(22, x) q(1, x)
.
q(7, x) q(4, x)
(51)
This checks.
Proof: This is analogous to the proof of theorem 1A. Again the dpr(2 n) representation, from which one
derives Ψ(2 n, x), with the prime number factorization for the odd n, is considered. The ∗-multiplication
property allows to consider dpr(2 p1 · · · pN ) with a subsequent multiplication of all arguments in the
numerator and denominator by pe11 −1 · · · peNN −1 . Because of the paired numerator/denominator structure
t(2 k, x2 )
due to lemma 3 one finds, either in the numerator or in the denominator quotients of the type
t(k, x2 )
k+1
n−1
, x . Factors of powers of 2 are irrelevant
which are up to a factor 2 2 (x + 2) equal to q
2
(they have to cancel) because on both sides of theorem monic polynomials appear. The factors of x − 2
also cancel because the number of q-polynomials in the numerator and denominator have also to match
(see the remark at the end of the proof of lemma 3). This number is 2N−1 . The
structure of the
x
polynomials in the
numerator of the theorem originates from products with the even indexed t 2k,
2
x
numerator after ∗- multiplication. Before this multiplication on has in the numerator t 2 p1 · · · pN ,
2
and t−products over all possibilities
xto leave out 2, 4, ... of the odd primes from the set {p1 , ..., pN }. The
and t−products over all possibilities to divide n by these 2, 4, ...
∗-multiplication then leads to t n,
2
odd primes. Together with the pairing partners from the denominator this leads to the q-polynomials
given in the numerator of the theorem. A similar argument produces the denominator q-polynomials.
The number of these q−polynomials in the numerator, viz 2N −1 , matches the one for the denominator.
Theorem 2B: Factorization of q-polynomials in terms of C-polynomials
Y
q(n, x) =
C(d, x) , n ∈ N .
(52)
1 < d|(2 n+1)
Example 6: n = 17
q(17, x) = C(5, x) C(7, x) C(35, x) .
This checks.
Proof:
19
(53)
This is again modeled after a similar proof in [30]. Compare this with the proof of theorem 1B.
X ϕ(d)
It is clear that both sides are monic integer polynomials, and the degree fits due to n =
−
2
d|(2 n+1)
1
ϕ(1). See the properties of the Euler totient function mentioned above in the proof of theorem 1B.
2
2l + 1
(n)
The zeros of q(n, x) are xl = 2 cos π
, for l = 0, 1, ..., n − 1. This follows from definition 1
√ 2n + 1
and lemma 6 with the zeros of S(2 n, 2 − x ) which are known from those of Chebyshev S-polynomials.
To show
each
that
of these zeros appears on the r.h.s. for C(d, x) with some d|(2 n + 1), d 6= 1, i.e., as
2l′ + 1
d−3
2 cos π
} and gcd(2l′ + 1, d) = 1 (see eq. 20), let gcd(2l + 1, 2n +
for some l′ ∈ {0, 1, ...,
q
2
1) = g, with some odd g with 2 n + 1 = d g. Then 2 l + 1 = (2 l′ + 1) g = (2 l′ + 1) 2n+1
d , and for each
}.
l ∈ {0, 1, ..., n − 1} there is one l′ ∈ {0, 1, ... , d−3
2
Remark 5: Derivation of Theorem 2A from Theorem 2B
Theorem 2B can be used as recurrence for the C-polynomials in terms of the q-polynomials. See the
remark 3. We give the solution in the following corollary.
Corollary 4: q-polynomials in terms of C-polynomials
With the prime number factorization of 2 n + 1 = pe11 · · · peNN , with odd primes, n ∈ N, one has
q(n, x) =
e1
Y
q1 =0
···
eN
Y
C(pq11 · · · pqNN , x)/C(1, x) .
qN =0
(54)
The division by C(1, x) = x + 2 was necessary because not all qj -indices were originally allowed to
vanish.
These rational representations of C(n, x) do not lend itself to extraction of the value C(n, 0) because
0
undetermined quotients appear. In the following we give the absolute term of C for prime indices.
0
Proposition 8: C(p, 0)
For n = p, a prime, one has
0
p−1
(−1) 4
C(p, 0) =
p+1
(−1) 4
if p = 2 ,
if p ≡ 1 (mod 4) ,
if p ≡ 3 (mod 4) .
(55)
Proof: In eq. (31) with n = p there is no gcd-restriction on the product. Therefore, one can use a
known formula (see appendix B for a proof)
n−1
Y
k=1
2 cos π
k
n
=
n−1
(−1) 2
0
if n is odd,
(56)
if n is even,
to obtain cy(2, −1) = 0 (which is also clear from the definition: cy(2, x) = x + 1), and for odd primes
p−1 = +1. Because all k contribute in the product, one can use for even k the formula
cy(p,
−1) = (−1)
p − 2K
2K
= − cos π
which shows that one generates again all odd k contributions, howcos π
p
p
p−3
2
2
Y
2l + 1
p−1
ever each with a minus sign. This leads to cy(p, −1) = (−1)
= C(p, 0)2 .
2 cos π
p
l=0
20
Thus C(p, 0) = ±1. The sign of C(p, 0) is (−1)δ+ (p) from eq. (19) for n = p. This number δ+ (p) of
p+1
positive zeros has been found in proposition 5 to be p−1
4 if p ≡ 1 (mod 4), and 4 if p ≡ 3 (mod 4). For
C(2, 0) = 0, the sign is, of course, irrelevant.
We close this section with a conjecture on the discriminant of the C−polynomials. The discriminant
of a monic polynomial p of degree n can be written as the square of the determinant of an n × n Van(n)
(n)
(n)
dermonde matrix Vn (x1 , ..., xn ) with elements (Vn )i,j := (xi )j , = 1, ..., n and j = 0, ..., n − 1
(n)
with the zeros xi of p . Here the δ(n) zeros of Cn are given in eq. (20). Another formula for the
discriminant (see e.g., [26], Theorem 5.1, p. 218) is in terms of the derivative C ′ (n, x) and the zeδ(n)
δ(n) (δ(n)−1) Y
(n)
2
ros of C: (−1)
C ′ (n, xi ). The result is the sequence Discr(C(n, x)) = A193681(n) =
i=1
[1, 1, 1, 8, 5, 12, 49, 2048, 81, 2000, 14641, 2304, 371293, ...] for n ≥ 1. The following conjecture is on the senδ(n)
quence q(n) =
which is A215041, [1, 2, 3, 2, 5, 3, 7, 2, 9, 5, 11, 9, 13, 7, 45, 2, 17, 27, 19, 25, ...].
Discr(C(n, x))
Conjecture: explicit form of the q-sequence
o) q(1) = 1 (clear).
!
i) If n = 2k for k ∈ N then q(n) = 2.
!
ii) If n = pk for odd prime p and k ∈ N then q(n) = p(p
k−1 +1)/2
.
kiN
ki1
iii) if n = 2k2 p(i1 ) · · · p(iN )
with k2 ∈ N0 , the ij − th odd primes p(ij ), where 2 ≤ i1 < i2 <
... < iN , with N ∈ N if k2 6= 0 and N ≥ 2 if k2 = 0, then
q(n)
!
=
N
Y
k2 −1
p(ij )2
ki −1
1
p(i1 )
ki −1
N
··· p(iN )
P (N,j)
, with P (N, j) =
j=1
=
Y
N
Y
l=1,l6=j
p
δ(n)
p−1
.
(p(il ) − 1) ,
(57)
odd p|n
ϕ(2 n)
. This last formula does, however, not work in the
2
cases i) and ii). One can compare this formula with the proven one for the discriminant of the cyclotomic
1
polynomials (the minimal polynomials of exp(2 π ) (or any of the primitive n-th root of 1), as given in
n
[25], eq. (1) p. 297. For this (slightly rewritten) formula see also A004124 and A193679.
The last eq. follows from the degree δ(n) =
Example 7:
2 +1)/2
ii) p = p(4) = 7, k = 3: q(73 ) = q(343) = 7(7
iii) n =
23 ·32 ·7,
q(n) =
2
2
32 ·3·6 72 ·3·2
= 1341068619663964900807.
= 4316018525852839090954658176626149564980915348463203041.
These values have been checked with the help of Maple13 [18].
Note added: The conversion of the C polynomials in terms of Chebyshev S polynomials (A049310) has
been initiated by Ahmet Zahid Küçük and was finished together with the author. See A255237 for the
details.
4
Splitting field Q(ρ(n)) for C(n, x), field extension and Galois group
π
, n ∈ N, with minimal polynomial C(n, x) over Q of degree
n
δ(n), has been studied in sect. 3. Each of these polynomials, being minimal, is irreducible. All roots
have been given in eq. (19) (or eq. (20)). C(n, x) is also separable because all of its roots are distinct.
The algebraic number ρ(n) = 2 cos
21
πk
= t̂(k, ρ(n)) (see eq. (2), and the coefficient array A127672), with the monic
n
integer t̂-polynomials, each zero can be written as integer linear combination in the vector space basis
< 1, ρ(n), ρ(n)2 , ..., ρ(n)δ(n)−1 >, called the power basis. One has to reduce in t̂(k, ρ(n)) all powers ρ(n)p ,
p ≥ δ(n) with the help of the equation C(n, ρ(n)) = 0. See Table 4 for the zeros of C(n, x), n = 1, ..., 30
written in this power basis.
Because 2 cos
Example 8: n = 8 (octogon)
p
√
The δ(8)p
= 4 zeros of C(8, x) = x4 − 4 x2 + 2 are, with ρ ≡ ρ(8) = 2 + 2, ± ρ and ±( − 3 ρ +
√
ρ3 ) = ± 2 − 2. In this case the degree 4 coincides with the number of DSRs in the upper half plane
and the negative real axis.
This shows that the extension of the rational field Q, called Q(ρ(n)), obtained by adjoining just one
algebraic element (called a simple field extension) is the splitting field (Zerfällungskörper in German)
for the polynomial C(n, x). Note that even though the polynomial C(n, x) is from the ring Z[x] one
Pδ(n)−1
needs Q(ρ(n)) with rational coefficients rj for the general element α =
rj ρ(n)j . For example,
j=0
1
ρ(8)−1 = 2ρ(8) − ρ(8)3 . Some references for field extensions and Galois Theory are [6], [12], chpts. V
2
and VI, [3], and the on-line lecture notes [16]. The dimension of of Q(ρ(n)) as a vector space over Q is
δ(n). This is the degree of the extension, denoted usually by [Q(ρ(n)) : Q], and it coincides with the
degree of the minimal polynomial for ρ(n). Of course, it is a proper extension only if δ(n) ≥ 2, i.e., for
n ≥ 4. This extension of Q is separable, i.e., the minimal polynomial for the general algebraic number
α given above is separable (has only distinct zeros). It is a normal field extension, meaning that every
irreducible rational polynomial with one root in Q(ρ(n)) splits completely over Q(ρ(n)). See, e.g., [6],
Theorem 5.24, p. 108.
We now consider a subgroup of the group of automorphisms of Q(ρ(n)), called Aut(Q(ρ(n)), which
consists of those elements σ which leave the subfield Q pointwise invariant (fixed point field Q):
σ : Q(ρ(n)) → Q(ρ(n)), α 7→ σ(α), with σ(β) = β for all β ∈ Q. The subgroup of Aut(Q(ρ(n)) of these
so-called Q-automorphisms is called the Galois group of Q(ρ(n)) over Q, denoted by Gal(Q(ρ(n))/Q).
Occasionally we abbreviate this with Gn . In order to find the elements σ (we omit the label n) of this
subgroup it is sufficient to know σ(ρ(n)), because of the usual rules for automorphisms: (i) σ(α + β) =
σ(α) + σ(β), (ii) σ(α β) = σ(α) σ(β), and (iii) α 6= 0 ⇒ σ(α) 6= 0. Indeed, because of σ(1) = σ(12 ) =
σ(1) σ(1), one has σ(1) = 1, and the images of products of ρ(n) are obtained from products of σ(ρ(n)).
Applying σ on the equation C(n, ρ(n)) = 0 (minimal polynomial), leads to C(n, σ(ρ(n))) = 0, because
the integer (rational) coefficients and 0 are invariant under the Q-automorphism σ we are looking for.
Therefore we have exactly δ(n) distinct Q-automorphisms σj , j = 0, ..., δ(n) − 1, determined from the
(n)
distinct roots of C(n, σ(ρ(n))), viz σj (ρ(n)) = ξ̃j+1 with the zeros of C(n, x) ordered like in eq. (19) with
(n)
increasing k values (see the Table 4). By the same token C(n, σ(ξ̃j )) = 0, for j = 1, ..., δ(n) for every
σ. Because all roots are distinct (separable C) this leads to an isomorphism between Gal(Q(ρ(n))/Q)
and a subgroup of the symmetric group Sδ(n) . The Q-automorphisms σ can therefore be identified with
(n)
permutations of the roots of C (see e.g., [6] ch. 6.3, pp. 132 ff. One identifies the roots ξ˜j with j, and
1
2 ... δ(n)
σj with the permutation
≡ [σj (1) σj (2) ... σj (δ(n))]. This subgroup of Sδ(n) is
σj (1) σj (2) ... σj (δ(n))
Abelian because only products of powers of ρ(n) appear and due to the automorphism property of σ this
carries over to the Galois group.
Example 9: n = 5 (pentagon) Q-automorphisms
α = r0 1 + r1 ρ(5) with
√ Q(ϕ) is as quadratic number
√ ρ(5) = ϕ, the golden section. In this case
field usually called Q( 5) with the basis < 1, ϕ > for integers in Q( 5) (see e.g., [10], ch. 14.3, p.
22
207, where τ = (ϕ − 1)). The two (δ(5) = 2) Q-automorphisms are obtained from the solutions
1
σ(ϕ) = ϕ or − = 1 − ϕ of C(5, n) = x2 − x − 1. Hence σ0 = id : σ0 (1) = 1, σ0 (ϕ) = ϕ and
ϕ
σ1 : σ1 (1) = 1, σ1 (ϕ) = 1 − ϕ. Because σ12 = σ0 , G5 is generated by σ1 , hence the Galois group
Gal(Q(ϕ)/Q) is the cyclic group Z2 (also known as additive group Z/2 Z or C2 , but we reserve C for the
minimal polynomials). The fixed field for G5 is Q; for the trivial subgroup with element σ0 it is Q(ϕ).
Example 10: n = 7 (heptagon) Q-automorphisms
(7)
(7)
(7)
The three zeros (δ(7) = 3) of C(7, n) are ξ̃1 = ρ(7), ξ˜2 = −1 − ρ(7) + ρ(7)2 and ξ˜3 = 2 − ρ(7)2
(see Table 4). In the sequel we omit the argument 7. Computing powers of ρ modulo C(7, ρ) = 0 one
finds the Q-automorphisms σ0 = id, σ1 (ρ = ξ˜1 ) = ξ̃2 , σ1 (ξ̃2 ) = ξ̃3 , σ1 (ξ̃3 ) = ξ̃1 and σ2 (ρ = ξ̃1 ) = ξ˜3 ,
.
σ2 (ξ̃2 ) = ξ̃1 , σ2 (ξ̃3 ) = ξ˜2 . Therefore, the identification with S3 permutations is σ0 = [1 2 3] = e,
.
.
σ1 = [2 3 1] and σ2 = [3 1 2]. Each σj permutation can be depicted in a circle diagram with vertices
labeled 1, 2 and 3 and directed edges, also allowing for loops. The only (Abelian) group of order three is
the cyclic Z3 subgroup of S3 .
We can also characterize these Galois groups Gn by giving their cycle structure and depict them as cycle
graphs. For cycle graphs see, e.g., [31] “Cycle graph” and “List of small groups” with all cycle graphs for
groups (also non-Abelian ones) of order 1, ... , 16. In order to manage powers of Galois group elements
we first need some fundamental identities of Chebyshev T -polynomials, whence t̂-polynomials.
We will need iterations of t̂ polynomials governed by the following well known identity.
Lemma 7: Iteration of t̂-polynomials
t̂(n, t̂(m, x)) = t̂(n m, x), n, m ∈ N0 .
(58)
Proof: See [26], Exercise 1.1.6, p. 5, first with the trigonometric definition of Chebyshev T-polynomials
which then carries over to the general polynomials defined by their recurrence relation. This identity is
then rewritten for t̂-polynomials.
Lemma 8: mod n reduction of t̂-polynomials in the variable ρ(n)
k
t̂(k, ρ(n)) = (−1)⌊ n ⌋ t̂(k(mod n), ρ(n)), n ∈ N, k ∈ Z .
(59)
k
(−1)⌊ n ⌋ =: pn (k), the parity of nk , will become important in the following. Sometimes nk is called
quotient and denoted also by k\n.
π
(l n + r) =
Proof: This follows, with k = l n + r, trivially from the trigonometric identity cos
n
π
π r
= (−1)l cos
cos π l +
r , with r ∈ {0, 1, ..., n−1}.
n
n
To simplify notation we will use also t̂n (x) for t̂(n, x). From lemmata 7 and 8 we have e.g., t̂3 (t3 (ρ(7))
=
π
2
t̂9 (ρ(7)) = −t̂2 (ρ(7)) = −(ρ(7) − 2). This is also +t̂5 (ρ(7)) from the identity cos
(n ± l) =
n
π
− cos
l . This proves the following lemma.
n
Lemma 9: Symmetry relation of t̂ polynomials
t̂(n − l, ρ(n)) = −t̂(l, ρ(n)) , n ∈ N, l ∈ {0, 1, ..., n} .
(60)
Lemma 7 which used the trigonometric ρ(n) definition can be rewritten as a congruence for the t̂polynomials with indeterminate x. This is because all what is needed is that ρ(n) is a zero of C(n, x).
Corollary 5: Congruence for t̂ polynomials modulo C-polynomials
k
t̂(k, x) ≡ (−1)⌊ n ⌋ t̂(k(mod n), x) (mod C(n, x)) , n ∈ N, k ∈ Z .
(61)
An example will illustrate this before we give an another proof of this corollary based on known T polynomial identities.
23
Example 11: Congruence for n = 7, k = 9
t̂(9, x) = x9 − 9 x7 + 27 x5 − 30 x3 + 9 x and C(7, x) = x3 − x2 − 2 x + 1. Polynomial
division shows that t̂(9, x) = (x6 + x5 − 6 x4 − 5 x3 + 9 x2 + 5 x − 2) C(7, x) + (−x2 + 2), hence
t̂(9, x) ≡ −t̂(2, x) (mod C(7, x)).
The alternative proof of corollary 5 is based on the following
factorization of Chebyshev S-polynomials
2π
in terms of the minimal polynomials Ψ(n, x) of cos
over Q (for these polynomials see A049310
n
and A181875/A181876).
Proposition 9: Factorization of S-polynomials in terms of Ψ-polynomials
x
Y
n−1
,n ∈ N.
Ψ d,
S(n − 1, x) = 2
2
(62)
2<d | 2n
As usual, the undefined (empty) product is defined to be 1.
Proof: Start with the eq. (3) of [30], p. 471, written for n → 2 nand for the t̂-polynomials instead of
Y
x
. Then use the identity [20], p.
Ψ d,
the T -polynomials. This is t̂(n + 1, x) − t̂(n − 1, x) = 2n+1
2
d|2 n
261, first line, with m → n, n → m and U -polynomials replaced by S−polynomials. This leads to the
identity
t̂(n + m, x) − t̂(n − m, x) = (x2 − 4) S(n − 1, x) S(m − 1, x) , m ≤ n ∈ N .
(63)
Here we only need the case m = 1. Then dividing out 2 Ψ(1, x2 ) = x − 2 and 2 Ψ(2, x2 ) = x + 2 leads
to the claimed identity.
To end the preparation for an independent proof of corollary 5 we state:
Corollary 6: C-polynomial divides some family of S-polynomials
C(n, x) | S(l n − 1, x) , n ≥ 2, l ∈ N .
(64)
This is clear from the definition of C in eq. (16) and the fact that 2 < 2 n | 2 l n, for l ≥ 1. E.g.,
C(10, x) = x4 − 5 x2 + 5 divides the family {S(9, x), S(19, x), S(29, x), ...}
Proof of corollary 5: From the corollary 6 and the identity eq. (63) with n → l n and m → r one sees
that t̂(k, x) = t̂(l n + r, x) ≡ t̂(l n − r, x) (mod C(n, x)). Now t̂(l n − r, x) = t̂((l − 1) n + (n − r), x),
and one can use l1 := l − 1 and r1 := n − r (remember that r ∈ {0, 1, ..., n − 1}, hence r1 has identical
range) as new l and r variables in this congruence, to get t̂(l1 n + r1 , x) ≡ t̂(l1 n − r1 , x) (mod C(n, x)) =
t̂((l − 2) n + r, x) (mod C(n, x)). This can be continued until one ends up with
t̂(0 n + k(mod n), x) (mod C(n, x)).
Now we are in a position to compute powers of elements of the Galois group Gn . The subset of odd
numbers 2 l + 1 < n entering the product in eq. (20) will be denoted by M(n). There are δ(n) (degree
of C(n, x)) such odd numbers.
Definition 2: The fundamental set M(n)
n−2
M(n) := 2 l + 1 l ∈ 0, ....,
and gcd(2 l + 1, n) = 1 = {m1 (n), ..., mδ(n) (n)},
(65)
2
with m1 (n) = 1 and we use the order mi (n) < mj (n) if i < j. For n = 1 one takes M(1) = {1}.
Example 12: |M(2)| = δ(2) = 1, M(2) = {1}; |M(14)| = δ(14) = 6 , M(14) = {1, 3, 5, 9, 11, 13}.
For M(n) see the row No. n of the array A216319.
c
For later purpose we define here, for odd n, the extended fundamental set M(n)
and its first difference
c
set △ M(n)
c
Definition 3: The fundamental extended set M(n)
for odd n
c
For n odd, ≥ 1: M(n)
:= {0, m1 (n) = 1, ..., mδ(n) (n) = n − 2, n + 2} .
24
c
Thus |M(n)|
= δ(n) + 2. Note that gcd(n ± 2, n) = 1 for odd n (proof by assuming the contrary:
gcd(n + 2, n) = d > 2 because n and n + 2 are odd. Then d | (n + 2) and d | n, hence d | ((n + 2) − n),
d | 2, implying d = 2 or d = 1, but d > 2.) The reason for defining this extended set is that for odd n the
c
first difference set △ M(n)
= {1, △m2 (n), ..., △mδ(n) (n), 4} with △mj (n) := mj (n) − mj−1 (n), will
become important later on. △m1 (n) = 1 and △mδ(n)+1 (n) = 4 for each odd n.
With definition 2 eq. (20) implies,
σj (ρ(n))
=
π
(n)
(n)
σj (ξ̃1 ) = ξ˜j+1 = 2 cos
mj+1 (n) =
n
t̂(mj+1 (n), ρ(n)) for j ∈ {0, 1, ..., δ(n) − 1} .
(66)
E.g., n = 7: δ(7) = 3, M(7) = {1, 3, 5}, σ0 (ρ(7)) = ρ(7), σ1 (ρ(7)) = t̂(3, ρ(7)), and σ2 (ρ(7)) =
t̂(5, ρ(7)).
Because t̂ is a rational integer polynomial
σj2 (ρ(n)) = σj (σj (ρ(n))) = t̂(mj+1 (n), σj (ρ(n))) = t̂(mj+1 (n), t̂(mj+1 , ρ(n))) ,
(67)
and with lemma 6 this becomes σj2 (ρ(n)) = t̂((mj+1 (n))2 , ρ(n)). In general we have
σjk (ρ(n)) = t̂((mj+1 (n))k , ρ(n)) , for j ∈ {0, 1, ..., δ(n) − 1} , n, k ∈ N .
(68)
Instead of powers of σj (ρ(n)) we can therefore consider powers of mj+1 (n). Lemmata 8 and 9 are now
employed to prove that a t̂ polynomial with a product of elements from M(n) as its first argument (or
index) is again a t̂ polynomial with first argument from M(n). In this way one can build sequences of
powers, starting from any element of M(n). Trivially, 1k = 1. Before proving this closure of M(n)
under powers, provided the rules for t̂ polynomials are taken into account, we give two examples, and
n.
then define a new equivalence relation on the integers, called (Modd n), denoted by ∼
Example 13: Cycle structure for n = 12 (dodecagon)
δ(12) = 4, M(12) = {1, 5, 7, 11}. 52 = 25 ≡ 1 (mod 12), reflecting lemma 8. The sign pn (25) in eq.
(59) is here +. (If later the sign will be − , the mod n result will be underlined.) The first 2-cycle is
therefore [5, 1]. Similarly, 72 = 49 ≡ 1 (mod 12) (sign +), whence the second 2-cycle is [7, 1], and finally,
112 = 121 ≡ 1 (mod 12) (sign +), producing the third 2-cycle [11, 1]. This result appears as the n = 12
entry in Table 6. In this example the Galois group is not generated by one element. hence it is non-cyclic.
In fact, Gal(Q(ρ(12))/Q) = Z2 × Z2 = Z22 (see first entry in Table 8). The corresponding cycle graph is
shown in Figure 4 as the first entry, where the shaded (colored) vertex stands for 1 and the open vertices
should here be labeled with 5, 7 and 11. The cycle structure is 23 (three 2-cycles). In this example it was
not necessary to employ lemma 9 because the signs were always +.
Example 14: Cycle structure for n = 7 (heptagon)
δ(7) = 3, M(7) = {1, 3, 5}. 32 ≡ 2 (mod 7), where now the sign − in eq. (59) is remembered by
the underlining. 2 ∈
/ M(7), and now lemma 9 is used to rewrite this 2 as 7 − 2 = 5. Therefore,
2
7
7 (or M odd 7) is used for the congruence in the new
3 ∼ 5 ∈ M(7) (or 32 ≡ 5 (Modd 7)). The symbol ∼
sense, due to lemmata 8 and 9. Then 3 · 5 = 15 ≡ 1 (mod n) (sign +). The first 3-cycle is therefore
[3, 5, 1]. Here 3 generates all the elements of M(7), and Gal(Q(ρ(7))/Q) = Z3 , the cyclic group of order
3 = δ(7). The corresponding cycle graph is a circle with three vertices, one of them, labeled 1, is shaded
(colored) and the other two open ones are labeled by 3 and 5.
n (or Modd n) over the integers Z. Remember
This brings us to the definition of an equivalence relation ∼
that the floor function for negative arguments is defined as ⌊−x⌋ = − ⌊x⌋ if x ∈ N0 , and ⌊−x⌋ =
−(⌊x⌋ + 1) if 0 < x ∈
/ N.
n on Z
Definition 4: Equivalence relation ∼
25
n l ⇔ an (k) = an (l), with the map an : Z → In := {0, 1, ..., n − 1}, k 7→ an (k),
For k, l ∈ Z, n ∈ N: k ∼
where
if pn (k) = +1 ,
rn (k)
an (k) =
(69)
rn (−k) if pn (k) = −1 ,
where we used the division algorithm to write
k = qn (k) n + rn (k), with the quotient qn (k) ∈ Z and
k
k
the residue rn (k) ∈ In . Note that qn (k) =
. The sign pn (k) = (−1)⌊ n ⌋ = (−1)qn (k) corresponding
n
to the parity of qn (k), appeared already in lemma 8. an (1) = 0 because −1 ≡ 0 (mod 1).
n l we also write k ≡ l (Modd n) (this should not to be confused with mod n). Therefore
Instead of k ∼
the sequence an could also be called Modd n. The first of these 2 n-periodic sequences an are found in
A000007(n + 1), n ≥ 0, (the 0-sequence), A000035, A193680, A193682, A203571, A203572 and A204453,
for n = 1, ..., 7, respectively.
The smallest non-negative residue system mod n, viz 0, 1, ..., n − 1 is used here. For the residue of k
modulo n Maple [18] uses rn (k) = modp(k, n). We also use k(mod n) for rn (k). The reader should
n is indeed an equivalence relation satisfying reflexivity, symmetry and transitivity. The
verify that ∼
disjoint and exhaustive equivalence classes are given by {[0], [1], .., [n − 1]}, called the smallest nonnegative complete representative classes (or residue classes) Modd n (we omit the index n at the classes:
[m] = n [m], written this way in order to distinguish this class from the ordinary one [m]n used in the
n m} = {l ∈ Z | an (l) = m}. Because
arithmetic mod n). These classes are defined by [m] = {l ∈ Z |l ∼
rn (−k) = 0 if k ≡ 0 (mod n) and rn (−k) = n − rn (k) if k 6≡ 0 (mod n) (later listed as lemma 17) one
can characterize these residue classes also in the following way.
Lemma 10: Complete residue classes Modd n
l ∈ [0] ⇔ l ≡ 0 (mod n). Equivalently, [0] = [0]n ,
l ∈ [1] ⇔ l ≡ 1 (mod 2 n) or ≡ −1 (mod 2 n). Equivalently, [1] = [1]2 n ∪ [2 n − 1]2 n ,
..
.
l ∈ [n − 1] ⇔ l ≡ n − 1 (mod 2 n) or ≡ −(n − 1) (mod 2 n).
Equivalently, [n − 1] = [n − 1]2 n ∪ [n + 1]2 n .
(70)
For example, take n = 7, then [3] = 7[3] = [3]14 ∪ [14 − 3]14 = {..., −25, −11, 3, 17, ...} ∪
{... − 17, −3, 11, 25, ...} = {..., −25, −17, −11, −3, 3, 11, 17, 25, ...}. If n = 2 one has 2 [0] = [0]2 and
2[1] = [1]2 . The first differences in the class [0] are, of course, n, and in the class [m], for m = 1, 2, ..., n−1,
they alternate between △1 = 2 (n − m) and △2 = 2 m, e.g., n = 7, m = 3, △1 = 8, △2 = 6 :
3, 11, 17, 25, 31, .... This difference alternation invalidates certain theorems known for mod n. E.g.,
Theorem 53 of reference [10], p. 50, is no longer true: take m = 5, n = 7, a = 3, b = 17. a and b
belong both to the class 5[3] as well as 7[3] but they do obviously not both belong to the class 35 [3] . The
following lemma 11 shows that is is sufficient to know the positive values, and append the negative of
these values for each class [m], for m ∈ {0, 1, ..., n − 1}.
Lemma 11: Antisymmetry of the classes [m], for m ∈ {0, ..., n − 1}
For n ∈ N and every m ∈ {0, ..., n − 1} the elements of the equivalence class [m] are antisymmetric
around 0.
Proof: This is obvious for the class [0] = {...−2 n, −n, 0, n, 2 n, ...}. For m > 0 the negative of every positive numbers of m (mod 2 n) appears as a negative number of (2 n−m) (mod 2 n), and vice versa: the negative of every positive numbers of [2 n−m]2 n appears as a negative one of [m]2 n . This is true because −(m+
l 2 n) = (2 n − m) − (l + 1) 2 n, for every l ∈ N, and similarly, −((2 n − m) + l 2 n) = m − (l + 1) 2 n, for
every l ∈ N.
This leads immediately to the following corollary.
26
Corollary 7: Non-negative elements of the classes [m]
For n ∈ N and every m ∈ {0, ..., n − 1} one has
[m]≥ = [m]2 n,≥ ∪ [2 n − m]2 n,> , and [m] = [m]≥ ∪ −([m]> ) .
(71)
Here we used the notations [m]≥ and [ · ]2 n,≥ or [ · ]2 n,> to denote the subset of non-negative numbers
of [m] and the non-negative or positive numbers of the ordinary residue classes mod 2n, respectively. Of
course 0 appears only in the class [0]. −([m]> ) is obtained from the set [m]> (excluding 0 in the case of
class [0]) by taking all elements negative. Note that [−m] is not used here.
With g(n, m) := gcd(m, 2 n−m) one has, for m ∈ {1, 2, ..., n−1}, [m]2n,> = g(n, m) [m/g(n, m)]2 n/g(n,m),>
and [2 n − m]2n,> = g(n, m) [(2 n − m)/g(n, m)]2 n/g(n,m),> . This is. of course, only of interest if
g(n, m) 6= 1. E.g., n = 6, m = 3, g(n, m) = 3: [3]12,> = 3 ∗ [1]4,> and [9]12,> = 3 ∗ [3]4,> , where
again k ∗ [p]q,> is the set with all members of the set [p]q,> multiplied with k. This is obvious.
The trivial formula for the members of the residue classes [m]≥ , considered as sequences of increasing
numbers called {c(n, m; k)}∞
k=1 , is
if m = 0 ,
(k − 1) n
(72)
c(n, m; k) =
k
k
2 2 n + (−1) m if m ∈ {1, 2, ..., n − 1} .
The nonnegative members of the complete residue classes (Modd n) for n = 3, 4, 5, 6, and 7 are found in
A088520, A203575, A090298, A092260, and A113807. Sometimes 0 has to be added, in order to obtain
the class [0]. Of course, these complete residue classes can be recorded as a permutation sequence of the
non-negative integers.
We now list several lemmata (the trivial lemma 17 has already been used) in order to prepare for the
proof of the multiplicative structure of these equivalence classes.
Lemma 12: Parity of Modd n residue classes
For even n the parity of the members of the residue class n [m], m ∈ {0, 1, ..., n − 1} coincides with the
one of m. If n is odd this is also true for the classes with m ∈ {1, 2, ... , n − 1}, and for m = 0 the parity
of the elements alternates, starting with + (for even).
Proof: The case of the residue class [0] is clear for even or odd n because of its members 0 mod n (see
lemma 10). Similarly, for the other m values, because then mod 2n applies.
Lemma 13: Periodicity of the parity sequence pn with period length 2 n
pn (k) = pn (k + 2 n l) for l ∈ Z, i.e., pn (k) = pn (r2 n (k)) = pn (k (mod 2 n)), for k ∈ Z, n ∈ N. (73)
Proof: This is obvious from the definition of pn (k) given in lemma 8, eq. (59), with definition 4, eq. (69).
For the second part use k = 2 n q2 n (k) + r2 n (k).
Lemma 14: (A)symmetry of sequence pn around k = 0
if k ≡ 0 (mod n)
+pn (k)
for n ∈ N, k ∈ N0 : pn (−k) =
(74)
−pn (k) , if k 6≡ 0 (mod n) .
Proof: This follows immediately from the property of the ⌊−x⌋ function mentioned above before definition 4.
Lemma 15: Product formula for the residue rn
For n ∈ N and k, l ∈ Z one has: rn (k l) = rn (k) rn (l) (mod n).
Proof: Just multiply k = qn (k) n + rn (k) with l = qn (l) n + rn (l).
Lemma 16: Residue r2 n from rn
iff r2 n (k) ∈ {0, 1, ..., n − 1},
rn (k)
For n ∈ N and k ∈ Z one has: r2 n (k) =
rn (k) + n iff r2 n (k) ∈ {n, n + 1, ..., 2 n − 1}.
Proof: Obvious for r2 n (k) ∈ {0, ..., n−1} as well as ∈ {n, ..., 2 n−1}.
27
Lemma 17: Residue for negative numbers
if k ≡ 0 (mod n) ,
0
For n ∈ N and k ∈ N0 one has: rn (−k) =
n − rn (k) if k 6≡ 0 (mod n) .
Proof: Obvious.
Lemma 18: Symmetry of the an (or Modd n) sequence around k = 0
For n ∈ N and k ∈ N0 one has: an (−k) = an (k).
Proof: This is proved for the two cases k ≡ 0 (mod n) and k 6≡ 0 (mod n) separately. In the first case
rn (k) = 0, as well as rn (−k) = 0 , hence an (k ≡ 0 (mod n)) = 0, which is symmetric around k = 0. In
the other case, we employ lemma 13, noting that r2 n (k) 6= 0, n (otherwise k ≡ 0 (mod n)). The two cases
r2 n (k) ∈ {1, 2, ..., n − 1} and r2 n (k) ∈ {n + 1, n + 2, ..., 2n − 1} have pn (k) = pn (r2 n (k)) equal +1 or
−1, respectively. Then with the second alternative of lemma 14 one finds an (−k) = rn (−k) if pn (−k) =
−pn (k) = +1 and an (−k) = rn (k) if pn (−k) = −pn (k) = −1 which coincides with the definition of
an (+k).
Now we turn to the arithmetic structure of the Modd n residue classes. It is clear from the following
counter-example that addition cannot be done class-wise. Consider n = 6, k = 2 and l = 7. Then
6 2 and 7 ∼
6 5, but
a6 (2 + 7) = 3 but a6 (a6 (2) + a6 (7)) = a6 (2 + 5) = a6 (7) = 5. In other words, 2 ∼
6
6
n0
2 + 7 = 9 ∼ 3 but (2 + 5) = 7 ∼ 5, and 3 is not equivalent to 5 (Modd 6). Similarly, from (a − 1) ∼
n 1, because for n > 2 in the latter case a can also be of the form −1 + k 2 n
does in general not follow a ∼
if it belongs to a class [m] with positive m, whereas in the first case it has to be of the form 1 + k′ n,
which can only match for n = 1 and 2. However, it turns out that multiplication can be done class-wise.
This is the content of the following proposition.
Proposition 10: Modd n residue classes are multiplicative
For n ∈ N and k, l ∈ Z one has:
n an (k) an (l) , i.e., kl ≡ an (k) an (l) (Modd n) .
an (k l) = an (an (k) an (l)) i.e., k l ∼
(75)
Before giving the proof, consider the example n = 6, k = 2 and l = 7. Now a6 (2 · 7) = 2 and
6 2∼
6 2 · 5∼
6 a6 (2) a6 (7).
a6 (a6 (2) · a6 (7)) = a6 (2 · 5) = a6 (10) = 2. Or stated equivalently, 2 · 7 = 14 ∼
Proof:
i) Due to the symmetry of an (see lemma 18) it is clear that it is sufficient to consider only non-negative
k and l.
ii) Consider first the cases k ≡ 0 (mod n) or l ≡ 0 (mod n). If m ≡ 0 (mod n) then an (m) = 0. This
follows for both alternatives in eq. (69). Therefore, if k ≡ 0 (mod n), an (k l) = 0 because k l ≡ 0 (mod n)
for every l, and an (0 · an (l)) = an (0) = 0, proving the assertion. In the other case, l ≡ 0 (mod n), the
proof is done analogously.
iii) Now rn (k) and rn (l) are non-vanishing, and k and l are positive. Four cases are distinguished
according to the signs of (pn (k), pn (l)), viz (+, +), (−, −), (+, −) and (−, +).
(+,+): In this case, due to the 2 n-periodicity of pn (see lemma 13), r2 n (k) and r2 n (l) are both from
{1, 2, ..., n − 1}, hence
(+, +) :
r2 n (k) = rn (k) and r2 n (l) = rn (l) .
(76)
Also, from eq. (69), one obtains in this case an (k) an (l) = rn (k) rn (l). There is an alternative for
an (rn (k) rn (l)), depending on pn (rn (k) rn (l)) being +1 or −1. Both cases are possible as the following
examples for n = 6 show: p6 (3 · 4) = p6 (12 (mod 12)) = p6 (0) = +1 and p6 (3 · 7) = p6 (21 (mod 12)) =
p6 (9) = −1. First the argument of pn is rewritten with the help of eq. (76). pn (rn (k) rn (l)) =
pn (r2 n (k) r2 n (l)). Due to the (2 n)-periodicity (lemma 13) this is pn (r2 n (k) r2 n (l) mod(2 n)) = pn (r2 n (k l)) =
28
p(k l), due to lemma 15 with n → 2 n, and again the (2 n)-periodicity. In the first alternative, the +1 case,
an (an (k) an (l)) = rn (k) rn (l) (mod n) = rn (k l), again from lemma 15. This is just a(k l) if pn (k l) =
+1, proving the assertion for this alternative. In the other case, pn (k l) = −1, an (rn (k) rn (l)) =
−(rn (k) rn (l))(mod n) which is rewritten with mod n-arithmetic and lemma 15 as −rn (k l)(mod n) =
−(k l)(mod n) = rn (−k l). This coincides with a(k l) for this alternative, proving the assertion.
(-,-): Now we have from lemma 16
(−, −) :
r2 n (k) = n + rn (k) and r2 n (l) = n + rn (l) .
(77)
Here pn ((rn (−k) rn (−l)) is rewritten with lemma 17, eq. (77), the 2 n periodicity, and lemma 15,
as follows. pn ((n − rn (k)) (n − rn (l))) = pn ((2 n − r2 n (k)) (2 n − r2 n (l))) = pn (r2 n (k) r2 n (l)) =
pn (r2 n (r2 n (k) r2 n (l))) = pn (r2 n (k) r2 n (l) (mod 2 n)). With lemma 15 (with n → 2 n this becomes
pn (r2 n (k l)) = pn (k l), again from the (2 n)-periodicity. In the first alternative pn (k l) = +1 and
an (rn (−k) rn (−l)) = rn (rn (−k) rn (−l)). With lemma 17 and mod n-arithmetic this is rn ((−rn (k)) (−rn (l))) =
rn (rn (k) rn (l)), and with lemma 15 this becomes rn (k l), coinciding with an (k l) for this alternative. For
the other alternative, pn (k l) = −1, an (rn (−k) rn (−l)) = −(rn (−k) rn (−l))(mod n). This becomes, with
lemma 17, mod n-arithmetic and lemma 15
−(rn (k) rn (l)) (mod n) = −rn (k l)(mod n). This vanishes if rn (k l) = 0 which means k l ≡ 0 (mod n)
(which is possible, e.g., n = 6, k = 2, l = 3), and then this coincides with the claim which is for this
alternative a(k l) = rn (−(k l)) = 0. If rn (k l) 6= 0 then −rn (k l) (mod n) = n − rn (k l) which also
coincides with the claim a(k l) = n − rn (k l) if k l 6≡ 0 (mod n).
We skip the proofs of the other two cases, (+, −) and (−, +), which run along the same line. Here one
arrives first at pn (−(k l)), and in order to compare it with p(k l) both alternatives in lemma 17 have to be
considered like in the just considered second alternative.
For the computation of the cycle structure of the Galois group Gn = Gal(Q(ρ(n))/Q) we are only
interested, due to eq. (20), in odd numbers relatively prime to n. Contrary to ordinary mod n-arithmetic
where the set of odd numbers O := {2l + 1 | l ∈ Z} is in general not closed under multiplication
(e.g., 5 · 5 = 25 ≡ 4 (mod 7)), it will be shown that O is closed under M odd n multiplication. Of
interest are the units, the elements which have inverses, in order to see the expected group structure.
First consider the reduced set O∗n , given by the odd numbers relatively prime to n. The negative odd
numbers in this set are just the negative of the positive odd numbers, therefore it will suffice to consider
O∗n,> := {2l + 1 | l ∈ N0 , gcd(2l + 1, n) = 1}. If n is a power of 2 the set O∗2 will be O, with the o.g.f.
x
(1 + x) for the sequence of positive odd numbers {o∗2,> (k) := 2 k + 1}∞
G(x) =
k=0 . For the
(1 − x)2
other even numbers n only the odd numbers relatively prime to the squarefree kernel of n, called sqf k(n),
(see A007947, encountered already several times), will enter the discussion. Therefore, besides the just
considered (trivial) case of the even prime 2, the set O∗n is only relevant for squarefree odd moduli, either
prime or composite. We consider first the case of odd primes n = p, and give the o.g.f. of the sequences
of numbers from O∗p,> , called o∗p,> (k), as well as an explicit formula in terms of floor functions. These
are the positive odd integers without odd multiples of the odd p.
Proposition 11: Odd prime moduli, o.g.f. and explicit formula for O∗p,> elements
∞
X
With Gp (x) :=
o∗p,> (k) xk , for odd primes p, one has
k=0
p−3
2
X
p−1
p−1
x
p−1
k
2 ) + 4 2
1
+
2
+
x
,
x
(1
+
x
Gp (x) =
(1 − xp−1 ) (1 − x)
k=1
29
(78)
and
o∗p,> (k) =
0
if k = 0 ,
2k − 1 + 2
k+ p−3
2
p−1
(79)
if k ∈ N .
We have used the even number 0 for k = 0 such that the Gp (x) sum can start with k = 0.
Proof: This proposition will become a corollary to the later treated general case of odd squarefree moduli
n in proposition 13.
Example 15: O.g.f.s and formula for reduced odd numbers for modulus p = 7.
G7 (x)
=
o∗7,> (n)
=
x
1 + 2 (x + x2 ) + 4 x3 + 2 (x4 + x5 ) + x6
(1 − x6 ) (1 − x)
n+2
2n − 1 + 2
, n ≥ 1.
6
,
(80)
(81)
The instances for p = 3, 5, 7, 11, 13 and 17 are found under A007310, A045572, A162699, A204454,
A204457, and A204458, respectively.
In order to prepare for the general case of odd squarefree modulus n, we state a proposition on the
structure of the reduced odd numbers set O∗n,> .
Proposition 12: Mirror symmetry and modular periodicity of O∗n,> for odd n
i) Mirror symmetry. For k ∈ {1, 2, ..., δ(n)} one has:
o∗n,> (2 δ(n) − (k − 1)) = 2 n − o∗n,> (k),
(82)
where o∗n,> (k) = mk (n) from M(n), given in definition 2, eq. (65).
ii) mod 2 n periodicity: For k ∈ N one has:
o∗n,> (k) = o∗n,> (k + 2 δ(n)) (mod 2 n).
(83)
Written as a relation between neighboring fundamental units, numbered by N ≥ 1, this becomes the
following statement. For k ∈ {2 (N − 1) δ(n) + 1, ..., 2 N δ(n)}, with N ∈ {2, 3, 4, ...}, one has:
o∗n,> (k) = (N − 1) 2 n + o∗n,> (k − (N − 1) 2 δ(n)) .
(84)
δ(n) is the degree of the minimal polynomial C(n, x) for the algebraic number ρ(n) introduced in section 2.
ω(n)
ω(n)
Y
Y
(pj − 1).
Note that if n =
pj with distinct odd primes p, and ω(n) = A001221(n), then 2 δ(n) =
j=1
j=1
L(n) := 2 δ(n) is the length of the fundamental N -units.
Before we give the proof consider figure 3 for the case n = 3 · 5 = 15 with δ(n) = 4. The second
statement ii) concerns the relation of the numbers of the second fundamental unit (N = 2) to the one
in the N = 1 unit. E.g., the odd number 37 for k = 2 · 4 + 2 = 10 is equal to 30 + o∗15,> (10 − 2 · 4) =
30 + o∗15,> (2) = 30 + 7, which checks. The statement i) shows the mirror symmetry within the first
(and any other) unit of length L(15) = 2 · 4 = 8. In figure 3 this symmetry is indicated by the brackets
below the first unit, and it is a symmetry around the missing number n = 15. Missing numbers have
been indicated by a dot. It is the pattern of missing odd numbers which is mirror symmetric, not the
one of the actual values of the odd numbers. But the relation between the odd numbers in the second
half of a unit and the first one then follows, and is given by the statement of the proposition. In figure
3 P (n) := 2 n = 30 is the shift for the o∗>,n values from the N = 1 to the N = 2 unit (or any of the
neighboring units), and p(n) := n + 1 = 16 is the shift for these values from the first half of every unit
to the second half.
30
Figure 3: Structure of the sequence of mod n reduced odd numbers, n=15
N=2
N=1
δ (n)
k
1 2 3 4 5 6 7
2 δ (n)
3 δ (n)
4 δ (n)
8 9 10 11 12 13 14 15 16 17...
o*>,n(k) 1 .. 7 . 11 13 .17 19 . 23 .. 29 31 .. 37 . 41 43 .47 49 . 51 .. 57 61 ...
n−2 n+2
+p(n)=n+1=16
+ P(n) = 2n = 30
Proof: i) It is clear from the degree of the minimal polynomial and eq. (20) that the number of mod n
reduced positive odd numbers smaller than 2 n is δ(2 n) = 2 δ(n) (from the definition of δ(n) in terms of
Euler’s ϕ function). We now determine δ(n) reduced odd numbers mod n which lie between n and 2 n,
by mirroring the δ(n) elements of M(n) around the position where the missing number n is situated,
which is between the position k = δ(n) and the next one. The mirror symmetry refers to the gaps. The
number mk (n) from M(n) will be mapped to o∗n,> (2 δ(n) − (k − 1)) at the mirrored position in the second
half of the later defined first fundamental unit (N = 1). To find this odd number the first difference
c
c
set △ M(n),
introduced in connection with M(n)
from the definition 3, becomes important in order to
count the gaps in the sequence of odd numbers when they are reduced mod n. o∗n,> (2 δ(n) − (k − 1)) is
found by adding to mk (n) twice the value of the sum of the gaps from the position k to the center, the
mirror axis. This is 2 (△mk+1 (n) + △mk+2 (n) + ... + △ mδ(n) (n) + 2). The 2 in this sum is half the
gap-length from the value n − 2 at the position δ(n) and n + 2 at the next position, the mirror-position
δ(n) + 1 of δ(n). This is a telescopic sum which becomes 2 (−mk (n) + mδ(n) (n) + 2). Therefore, the
value of o∗n,> (2 δ(n) − (k − 1)) is mk + 2 (−mk (n) + mδ(n) (n) + 2) = 4 + 2 mδ(n) (n) − mk (n) =
2 n − mk (n), because always mδ(n) (n) = n − 2 holds. See figure 3, n = 15 with the values 30 − 1 = 29,
30 − 7 = 23, 30 − 11 = 19 and 30 − 13 = 17 for k = 1, 2, 3 and 4, respectively. Now it is
clear that this mirroring leads to the correct number of reduced odd numbers for the second half of the
fundamental N = 1 unit, because gcd(mk (n), n) = 1, as member of M(n), implies for the mirror image
also gcd(o∗n,> (2 δ(n) − (k − 1)), n) = 1. All positive odd numbers relatively prime to n and not exceeding
2 n − m1 (n) = 2 n − 1 have thus be found.
Proof of ii): From above we know that the number at position k = 2 δ(n) + 1 is 2 n + 1 because the
one for k = 2 δ(n) has been shown to be 2 n − 1 and gcd(2 n + 1, n) = 1 (indirect proof by assuming the
c
contrary, using odd n; this is similar to the proof given in connection with definition 3 of M(n)).
Now it
∗
is clear that a shift in the on,> numbers with P (n) := 2 n leads from the fundamental unit No. N = 1
to the second one, N = 1, by putting o∗n,> (2 δ(n) + k) = o∗n,> (k) + P (n), for k ∈ {1, 2, ..., δ(n)}.
This is obvious because the gcd value 1 is not changed by adding 2 n. This process can be iterated
to find the mod 2 n periodicity structure stated in ii). See figure 3, n = 15 with k = 14, N = 2:
o∗15,> (14) = 49 = 2 · 15 + o∗15,> (14 − 1 · 2 · 4) = 30 + o∗15,> (6) = 30 + 19.
∗
The sequences {on,> } for n = 15 and n = 21 are found in A007775 and A206547, respectively.
The sequences of the M odd n residues of the numbers o∗n,> (k), for k = 1, 2, ..., for prime moduli
n = p = 3, 5, 7, 11, 13, 17, and for the first odd composed ones for n = 15, 21 are found in A000012,
A084101, A110551, A206543, A206544, A206545, and A206546, A206548, respectively.
31
For the following formula for the nonnegative odd numbers reduced mod n involving floor functions we
c
need the following list (increasingly ordered set) F(n) of length L(n) = 2 δ(n) derived from the list △M.
Definition 5: List F(n)
F(n) = {f (n, 1), ..., f (n, 2 δ(n)} with
f (n, j)
=
f (n, δ(n))
=
f (n, 2 δ(n))
=
△mj+1 (n) − 2
, for j ∈ {1, 2, ..., δ(n) − 1}
2
1, f (n, δ(n) + j) = f (n, δ(n) − j), for j ∈ {1, 2, ..., δ(n) − 1}, and
0.
(85)
c
Here △mk (n) := mk (n) − mk−1 (n), (the kth element of △M(n)).
This list F(n) is obtained from first
c
enlarging △M by mirroring the first δ(n) entries at the last element 4, to obtain a list of order 2 δ(n) + 1.
Then the first and last element of this new list is put to zero and all the other elements are diminished
by 2, thus obtaining a set of only even numbers. Then one divides by 2, omits the final 0, and reverses
the remaining list.
Example 16: F(15)
c
c
n = 15 = 3 · 5, 2 δ(15) = 2 · 4 = 8, M(15) = {1, 7, 11, 13}, M(15)
= {0, 1, 7, 11, 13, 17}, △M(15)
=
{1, 6, 4, 2, 4}, the mirror extension is {1, 6, 4, 2, 4, 2, 4, 6, 1}, the reduction step leads to {0, 4, 2, 0, 2, 0, 2, 4, 0},
and finally, dividing by 2, omitting the last 0 and reverting, leads to F(15) = {2, 1, 0, 1, 0, 1, 2, 0}. Except
for the last 0 there is a mirror-symmetry around the fourth entry 1. See the first row of table 5.
c
For the odd squarefree composite numbers n = A024556(m), m = 1, 2, ... , 17, see table 5 for M(n),
c
△ M(n)
and F(n).
Proposition 13: O.g.f. and formula for O∗n,> elements.
∞
X
o∗n,> (k) xk , for odd n, one has
With Gn (x) :=
k=1
x
Gn (x) =
1+
(1 − x2 δ(n) ) (1 − x)
2 δ(n)−1
X
k=1
△o∗n,> (k + 1) xk + x2 δ(n)
with the first differences △o∗n,> (j) := o∗n,> (j) − o∗n,> (j − 1) .
This generates
2 δ(n)
X
k + (j − 1)
∗
.
on,> (k) = 2 k − 1 + 2
f (n, j)
2 δ(n)
(86)
(87)
j=1
Note that the numerator polynomial in Gn (x) needs the first differences of the sequence members of the
fundamental N = 1 unit, which due to the mirror symmetry of proposition 12 can be reduced to the
first differences of {m2 (n), m3 (n), .., mδ(n) , n + 2} with mδ(n) = n − 2.
One could use, like in proposition 11, on,> (0) = 0 and let the sum in Gn (x) start with k = 0.
Proof: We start with the periodicity of the sequence due to proposition 12 ii). Taking the difference
o∗n,> (k) − o∗n,> (k − 1) = o∗n,> (k − 2 δ(n)) − o∗n,> (k − 1 − 2 δ(n)) leads to the recurrence (we omit
all unnecessary indices) o(k) = o(k − 1) + o(k − 2 δ(n)) − o(k − 1 − 2 δ(n)), for k ≥ 2 δ(n) + 1.
Here one needs also o∗n,> (0) := −1. For some of these recurrences the o.g.f.s were determined by
2 δ(n)
R. J. Mathar in e.g.,
A045572 and A162699. I general Gn (x) =
X
k=1
1) x
k−1
+ x
2 δ(n)
∞
X
k=2 δ(n)+1
o(n − 2 δ(n)) − x
2 δ(n)+1
∞
X
k=2 δ(n)+1
32
k
o(k) x + x
∞
X
k=2 δ(n)+1
o(k −
o(n − 2 δ(n) − 1) xn−2 δ(n)−1 , where the
recurrence has been used (and the infinite sum has been reordered, not bothering about absolute con2 δ(n)
X
vergence, in the sense of formal power series). Shifting the indices, one arrives at Gn (x) =
o(k) xk +
k=1
o
n
2 δ(n)
2 δ(n)+1
2 δ(n)−1)
+x
Gn (x) − x
(Gn (x) + (−1)), where
x Gn (x) − (o(1) x + .... + o(2 δ(n) − 1) x
the −1 resulted from putting o(0) = −1 (see above). This rearranges into (1 − x − x2 δ(n) +
x2 δ(n)+1 ) G n (x) = x 1 + △o(2) x + ... + △o(2 δ(n)) x2 δ(n)−1 + x2 δ(n) . Factorizing the bracket on
the l.h.s. and division leads to the claimed form for Gn (x). Of course, the denominator factor (1 − x2 δ(n) )
can be factorized into cyclotomic polynomials.
For the proof of the second part we use L := 2 δ(n). Given Gn (x) with the L + 1 input coefficients of the
numerator polynomial one derives the claimed explicit form for o∗n,> (k), by defining first the sequence with
x
k+L−1
(partial sums of the characteristic sequence
, generated by
entries bL (k) :=
L
L
(1 − x ) (1 − x)
for multiples of L, then shifted). The numerator polynomial leads for o∗n,> (k) to a sum of L + 1 floor
k−L+L−1
k−0+L−1
, ending with
=
functions with decreasing arguments, starting with
L
L
k−1
, and corresponding coefficients. In order to find a standard form for this sum we use the following
L
floor-identity
L−1
X k + j
= k, for L ∈ N, and k ∈ Z ,
(88)
L
j=0
which can be seen from
L−1
X
j=0
(k + j) (mod L) =
L (L − 1)
, k ∈ Z, which is trivial (just add the L terms
2
which
are 0, 1, ... , L − 1 in a certain cyclic order), and the relation between mod L and the floor function
k
1
(k − k (mod L)). This identity allows us to lower L consecutive coefficients of this sum of
=
L
L
L + 1 terms by 1, producing
k if one uses the identity for the first L terms because the second to
a term
,
and
the identity is read backwards. If one uses the identity for the last
last floor-argument is then k+0
L
L terms one produces a k − 1, because the last term has k−1
L . When we apply this identity twice in the
described way we pick up n + (n − 1) = 2 n − 1 and the first and last coefficient, which were originally
1, become 0, and all other coefficients are diminished by 2 because they participate in both applications
of the identity. It is guaranteed that all coefficients are now even and ≥ 0, because the coefficients
except the first and last one were even and ≥ 2, because these numerator polynomial coefficients resulted
from first differences of the sequence of odd numbers. Therefore one
extract a factor 2 and if the
can
k
(the second to last term in the
floor-functions are written with increasing arguments, starting with
L
original order), one has exactly the coefficients given by the list F(n) of length L of definition 5. This
proves the explicit form for o∗n,> (k).
Example 17: O.g.f. G8 (x)
(1 + 2 (x + x2 + x3 + x4 + x5 + x6 + x7 ) + x8
1+ x
G8 (x) = x
= x
(1 − x8 ) (1 − x)
(1 − x)2
(89)
generating, with offset 0, the odd numbers A004273.
Before coming to the multiplicative group Modd n we define reduced residue systems Modd n as well as
reduced odd residue systems Modd n. See e.g., [2], p. 113 for the mod n case.
33
Definition 6a: Reduced residue system Modd n
A reduced residue system Modd n (RRSn) is any set of ϕ(n) pairwise incongruent Modd n numbers, each
of which is relatively prime to n.
E.g., n = 15, ϕ(15) = 8, {1, 2, 4, 7, 8, 11, 13, 14} or {29, 32, 26, 23, 22, 41, 17, 16}, etc. The first one is the
smallest positive one. These systems will not play a rôle later on.
Definition 6b: Reduced odd residue system Modd n
A reduced odd residue system Modd n (RoddRSn) is any set of δ(n) odd pairwise incongruent Modd n
numbers, each of which is relatively prime to n
E.g., n = 15, δ(15) = 4, {1, 7, 11, 13} or {29, 27, 19, 17}. The first one coincides with M(15) from
eq. (65) and is the smallest positive reduced residue system Modd n. Remember that the parity of the
members in each Modd n residue class, not in class [0], is the same (see lemma 12). Later on we will
restrict ourselves mostly to the system M(n).
We next study the multiplicative group M odd n. The elements are the residue classes [mj ] , for j =
1, 2, ..., δ(n), corresponding to the members of the reduced residue system M(n). In fact, we will take
these representatives, multiplying Modd n, as we have done above.
In order to see the group structure one first convinces oneself that this set M(n) is closed under Modd n
multiplication. This follows from proposition 10 and lemma 12 which showed that these classes have
only odd numbers, and every odd numbers appear exactly once because of the definition of these classes.
The associativity, commutativity and the identity element 1 are also clear. We do not have a formula
how to find the inverse element m−1
j of mj but coming to this Modd n multiplication from the study of
automorphisms of the splitting field for the minimal polynomial C(n, x) for the algebraic number ρ(n),
it is clear that these inverses have to exist (the invariance property defines a group). We are dealing here
with the Galois groups for these polynomials. The cycle structure for n = 1, 2, ..., 40 is given in table
6. Because each group is Abelian one has for every group of prime order a cyclic group, which can be
checked for the examples given in table 6.
Remark 6: The multiplicative group Modd n is cyclic if δ(n) is prime.
This is a corollary on the fundamental theorem on finite Abelian groups (see e.g., [28], p. 49, see also
[6], p. 511, Cauchy’s Theorem A.1.5), or use Lagrange’s theorem on the order of subgroups, and the fact
that the powers of an element, not the identity, generate a cyclic subgroup, to prove that in fact every
group of prime order is cyclic. The values n for which the order δ(n) is prime are given in A215046.
This is the sequence [4, 5, 6, 7, 9, 11, 23, 47, 59, 83, 107, 167, 179, 227, 263, 347, 359, ...]. Of course, there are
other values n with cyclic Modd n group, like n = 2, 3, 8, 10, .... For n = 1 one has the trivial case with
cycle structure [[0]], also a cyclic group, viz Z1 . One can give a more general sequence of n numbers
with cyclic Gn ∼
= multiplicative Modd n group, namely A210845.
Remark 7: The multiplicative group Modd n is cyclic if δ(n) is squarefree.
The squarefree numbers are given in A005117, namely [1, 2, 3, 4, 5, 6, 7, 9, 11, 13, 14, 18, 21, 22, 23, 25, 29,
31, ...]. This remark 7 results from the fact that A000688, giving the number of Abelian groups of order
n, is 1 exactly for the squarefree numbers A005117. See the formula, based on the H.-E. Richert reference
quoted there. Because for each order there is at least the cyclic group these values lead necessarily to a
cyclic group. The above given A215046 values are a proper subset of those from A210845. There are.
however, still more values n with a cyclic Modd n group. Missing are e.g., 8, 10, 15, 16, 17, .... All the n
values with cyclic Modd n group are in A206551. The complementary sequence is A206552, giving the n
values with non-cyclic Modd n group. See Table 8 for all values n ≤ 100.
As an aside we remark that special squarefree numbers are the so called cyclic numbers A003277. It is
known that if the order of a (finite) group is a cyclic number then there is only one group, the cyclic one.
See a comment on A003277. Therefore A000001(n), the number of groups of order n is 1 if n is a cyclic
number, in fact, the reverse also holds: if A000001(n) = 1 then n is a cyclic number. This can be taken
as an alternative definition for cyclic numbers because then there is only one (non-isomorphic) group of
34
this order which has to be the cyclic group. See also Yimin Ge’s Math Blog [YiminGe], where the ‘only
if’ statement in the proposition may be misleading but in the proof the given statement is correct. E.g.,
for order 6 (not a cyclic number) there are groups other than Z6 . In fact A000001(6) = 2, and there
is the (non-Abelian) group D(3) (dihedral group). (It is clear that cyclic numbers are not the numbers
n for which the multiplicative group mod n, which is Abelian of order ϕ(n), is cyclic. These numbers
are given in A033948). Because we are dealing with Abelian groups the squarefree numbers are more
interesting here.
Recall the situation for the Galois group for the cyclotomic polynomials (the minimal polynomials for
ζ(n), an n-th root of unity) which is isomorphic to the multiplicative (Abelian) group of integers modulo
n. See Table 7 were we have listed these non-cyclic groups for n ∈ {1, ... , 100}. See also A282624 for a
W. Lang link with a table for n ∈ {1, ... , 130} also with generators. This n−value sequence is known as
A033949. (The case n = 15 is an exercise in [7], p. 159, Ex. 9.6. 3)a)). The values n with a cyclic Galois
group coincide with the moduli n which possess primitive roots r = r(n), i.e., the order of r modulo n
is ϕ: r ϕ(n) ≡ 1 (mod n), and no smaller positive exponent k satisfies this congruence. These moduli n
′
are known to be exactly pe , 2 pe , 1 and 2, with some positive powers e and e′ . See e.g., [23], Theorem
2.41, p. 104. All other n lead to non-cyclic multiplicative groups modulo n . See A033949. For the
smallest primitive roots in this case see [1], pp. 864-869, the column called g. We do not know a similar
characterization of the non-cyclic numbers n for the Modd n group which are shown in Table 8. However,
these numbers have also to appear in A033949. It is clear that we should determine primitive roots
r ≡ r(n) for the Modd n multiplication, i.e., find those r from M(n) which have order δ(n): the smallest
positive k which satisfies r k ≡ +1 (Modd n) is k = δ(n). See A206550 for these smallest positive primitive
roots Modd n, where a 0 entry, except for n = 1, indicates that there exists no such primitive root. This
sequence starts, with offset 1, as [0, 1, 1, 3, 3, 5, 3, 3, 5, 3, 3, 0, 7, 5, 7, 3, 3, 5, 3, 0, 11, 3, 3, 0, 3, 7, ...]. In general
one does not expect a formula for these primitive roots r = r(n), because also in the multiplication
modulo n case there is no one available.
Proposition 14: Number of Modd n primitive roots
If a primitive root r for the multiplicative group Modd n exists there are ϕ(δ(n)) of them.
This is the sequence A216322. For example, n = 13, δ(13) = 6, ϕ(6) = 2 with r1 = 7 and r2 = 11;
n = 14, δ(14) = 6, ϕ(6) = 2 with r1 = 5 and r2 = 11.
Proof: If a Modd n primitive root exists then the multiplicative group Modd n is cyclic. The order of this
group is δ(n), Then the number of pairwise incongruent primitive roots is obtained exactly like in the
case of primitive m-roots of unity on the unit circle by the number of relatively prime numbers less than
m which is Euler’s ϕ(m). Here m = δ(n).
Now the question whether the multiplicative group Modd p with p a prime is cyclic is answered. Up to
p−1
squarefree.
now we know from remark 7 the positive answer only for those primes with δ(p) =
2
These are the primes given in A066651. The case p = 2 with δ(2) = 1 is trivially cyclic. For general
prime p one wants to show that there exists a primitive root in the multiplicative Modd p group. This is
analog to the mod p case, where a proof can be found in [23], Theorem 2.36, p. 99, or in [2], ch. 10.4,
pp 206 ff. One has, however, to be careful to use only the multiplicative group structure in the proof.
Already in connection with lemma 10 we have given a warning to use theorems valid in the modular
arithmetic mod n in our case. Another failure occurs
for the theorem [23],
Theorem 2.3(c), p. 49, e.g.,
m1 m2
x ≡ y (mod m1 ), x ≡ y (mod m2 ) if and only if x = y mod
. For the Modd n case one has
gcd(m1 , m2 )
the counterexample 19 = 5 (M odd 3) and 19 = 5 (M odd 7) but 19 is obviously not congruent 5 (Modd 21).
Several lemmata, the analoga of mod n facts, are collected before stating proposition 15.
Lemma 19: Analogon of the Fermat-Euler Theorem
If a is odd and gcd(a, n) = 1 then aδ(n) ≡ +1 (M odd n), with delta(n) = A055034(n) (see the start of
section 3).
35
Proof: Analog to e.g., [2], Theorem 5.17, p. 113, with m → n, ϕ → δ, the order of the set (reduced
odd residue system) M(n) of eq. (65), and modulo → Modd n (we avoid the term Moddulo). Instead
of M(n) one can use any other reduced odd residue system Modd n (see definition 6b), with mj (n), j ∈
{1, 2, ... , δ(n)}, replaced by any member of the residue class [mj (n)]. The cancellation used at the end of
the proof in [2] can be done here by multiplying with the existing inverses of the bi s there. We are dealing
with the multiplicative group Modd n.
Next follows the definition of the order of an element from a reduced odd residue system Modd n.
Definition 7: Modd n order of a from a RoddRSn
The Modd n order of a positive (odd) integer a from a reduced odd residue system Modd n is the smallest
positive integer h such that ah ≡ 1 (Modd n).
E.g., n = 10, a = 9, h = 2 because 91 = 9 (Modd 10), and 92 = 1 (Modd 10). See the table A216320
corresponding to the Modd n orders h of the smallest positive RoddRSn members A216319.
Lemma 19 guarantees the existence of a Modd n order h ≤ δ(n) for each positive odd number a with
gcd(a, n) = 1.
Example 18: n = 12, δ(12) = 4, M(12) = {1, 5, 7, 11}, 14 ≡ 1 (Modd 12), 52 ≡ 1 (Modd 12) ⇒ 54 ≡
(Modd 12), similarly for 7 and 11. See the cycle structure for n = 12 in table 6.
Lemma 20: The Modd n order divides δ(n)
If h is the Modd n order of a then h|δ(n). Moreover, aj ≡ ak (Modd n), w.l.o.g. j > k, if and only if
h|(j − k).
Proof: The first part is analog to [23], Corollary 2.32, p. 98, with ϕ → δ. It uses lemma 19, the
Fermat-Euler analogon. Remember that always gcd(a, n) = 1 from the order definition. The second
part, for which one can use the group property, follows from the analog of [23], Lemma 2.31, p. 98. (In
the older German version of this book, vol. I, the analog of this lemma 20 appears as Satz 22.3 on p. 63.)
Lemma 21: On the Modd p order of a, p a prime
If h is the Modd p order of a, with p a prime, one has (ak )h ≡ 1 (Modd p) for all k, and 1 =
a0 , a1 , a2 , ... , ah−1 are pairwise incongruent Modd p.
Proof: First part: (ak )h = (ah )k , ah ≡ 1 (Modd p) and Modd p respects multiplication (see proposition
10). Second part: assume the contrary, i.e., ai ≡ aj (Modd p), 1 ≤ j < i ≤ h − 1. Apply lemma 20 for
n = p, showing that h|(i − j), but i − j ≤ i ≤ h − 1 leading to a contradiction.
h
Lemma 22: X ≡ 1 (Modd p) has at most h incongruent solutions
The congruence X h ≡ 1 (Modd p), h a positive integer, has at most h pairwise incongruent solutions
Modd p.
h
X
is even then X h ≡ 1 (mod p). and due to the mod p theorem, e.g., [21], Theorem
Proof: If
p
h
X
42, p. 80, there are at most h incongruent solutions. If
is odd then −X h ≡ 1 (mod p), i.e.,
p
X h ≡ (p−1) (mod p) which again has at most h incongruent solutions.
This is a weak statement, but sufficient for the following. One could try to prove that the number of
p−1
) if p is odd. This
Modd n incongruent solutions of the congruence X h ≡ 1 (Modd p) is gcd(h,
2
number is trivially 1 for p = 2.
Lemma 23: On the Modd n order of ak
h
If h is the Modd n order of a, then
is the Modd n order of ak .
gcd(h, k)
Proof: This is the analogon of [23] lemma 2.33, p. 98, using the present lemma 20, part 2.
Now we are ready for the following proposition.
36
Proposition 15: Existence of Modd p primitive roots
For every prime p there exists a primitive root for the multiplicative group Modd p.
Proof: One shows that there are precisely ϕ(δ(p)) primitive roots and because this number is always
≥ 1 the claim will then follow.
Following the proof of [23], Theorem 2.36, p. 99, we infer from the present lemma 19 (with n → p) that
p−1
,
every positive odd integer a with gcd(a, p) = 1 has a Modd p order h ≡ hp (a), 1 ≤ h ≤ δ(p) =
2
h
k
1
2
h−1
and from lemma 20 h|δ(p). Moreover, (a ) ≡ 1(Modd p) for all k, and 1, a , a ... a
are h pairwise
incongruent odd numbers Modd p. From lemma 22 these are all the solutions of the congruence X h ≡ 1
(Modd p). Lemma 23 shows that there are ϕ(h) numbers ak , with k ∈ {1, ... h − 1} which have Modd p
order h ≡ hp (a) because ϕ(h) is the number of such k with gcd(k, h) = 1. The case h = 1 is covered
with ϕ(1) := 1.
)
(
With [2], p. 207, we define for each h|δ(p) the set Ap (h) :=
a
a ∈ M(p) and Modd p order of a is h .
There are τ (δ(p)) (τ = A000005, the number of divisors) such disjoint sets. This is the sequence
[1, 1, 2, 2, 2, 4, 4, 3, 2, 4, 4, 6, ...] = A216326. E.g., p = 7, M(7) = {1, 3, 5}, τ (3) = 2 (see also
row No. n = p = 7 of the Modd n order table A21630). Call ψp (h) := |Ap (h)|, the number of elements
of this set. We have seen that ψp (h) is 0 or ϕ(h) because for each h|δ(p) the set Ap (h) is either empty
(no
X a ∈ M(p) has Modd p order h) or this set has ϕ(h) elements. Thus ψp (h) ≤ ϕ(h). It is clear that
ψp (h) = δ(p) because each a ∈ M(p) belongs to one of these sets Ap (h). Now a standard result is
h|δ(h)
X
h|δ(h)
ϕ(h) = δ(p) (see e.g., [2], Theorem 2.2, p. 26, wit hn → δ(p)), therefore,
X
h|δ(h)
(ϕ(h) − ψp (h)) = 0,
but because ϕ(h) − ψp (h) ≥ 0 it follows that ψp (h) = ϕ(h) > 1, not 0, for each h|δ(p). This holds
especially for h = δ(p). Hence the number of primitive Modd p roots is ψp (δ(p)) = ϕ(δ(p)) > 1, and the
proof is complete.
For the sequence δ(p(n)) n ≥ 1, see A130290. For the smallest positive primitive Modd n roots see the
sequence A206550. Here for prime n.
Corollary 8: Cyclic Modd p group
The multiplicative (Abelian) Modd p group, for p a prime, is the cyclic group Zδ(p) .
We now concentrate on those groups Modd n which are cyclic and notice that whenever δ(n) is even there
exists a unique smallest positive odd number > 1 which solves the congruence x2 ≡ 1 (Modd n). This is
the nontrivial solution of this congruence. The trivial one is x = 1 (standing for the class of solutions
1 (Modd n) which also includes −1). We prove this first for even n.
Proposition 16: Nontrivial square-root of 1 (Modd 2 k), k ≥ 2
If n = 2 k, for k ∈ {2, 3, 4, ...}, and if the Galois group Gal(Q(ρ(n))/Q) is cyclic, a unique smallest
positive solution r > 1 of the congruence x2 ≡ 1 (Modd 2 k) exists, and it is r = n − 1.
Proof: For even n = 2 k, k ≥ 2, the cyclic group is then Zδ(2 k) and δ(2 k) is even because with the
Y ej −1
pj
standard prime factorization δ(2e pe11 · · · peNN ) = 2e−1
(pj − 1), and if N = 0 then e ≥ 2, and
j=1..N
if N ≥ 1 then e ≥ 1. In both cases δ is even, because in the second case there is at least one j and
pj − 1 is even. Even order δ(2 k) of the cyclic group means that powers of the generator c of the δ-cycle,
the smallest positive primitive root (of 1) (Modd 2 k), come in pairs which are mutually inverse, except
δ(2 k)
for 1 = cδ(2 k) ands := c 2 . This s > 1 is unique and solves of the congruence x2 ≡ 1 (Modd 2 k).
(n − 1)2
= n − 2 + 0 is, with even n, even, and in order to compute (n − 1)2 (Modd n)
Because
n
37
one has to determine (n − 1)2 (mod n) , which is 1. Therefore, the unique s for even n is s = n − 1.
Example 19: Nontrivial square-root of 1 (Modd 8)
49
s(8) = 7: 72 = 49 ≡ 1 (M odd 8), because
= 6, hence p8 (49) = +1, and 49 ≡ +1 (mod 8),
8
therefore 49 ≡ +1 (Modd 8).
p−1
p−1
Similarly, for n = p, p an odd prime, the cycle length is δ(p) =
. In the event that
is even,
2
2
i.e., if p is of the form 4 k + 1, i.e., p ≡ 1 (mod 4), (see A002144) a smallest positive nontrivial solution
of the quadratic congruence x2 ≡ 1 (Modd n) exists. We will numerate the primes congruent 1 (mod 4)
by defining p̂(n) := A002144(n), n ≥ 1. E.g., for n = p̂(2) = 13 one has, besides the trivial solution
x = 1 (Modd 13) (this residue class includes −1), x ≡ 5 (Modd 13) as a solution. This means that in the
group table the row for the element 5 has a 1 in the diagonal, like in the row for the identity element 1.
These elements are self-inverse like the identity element 1. Note that in the ordinary reduced mod n case
there is always a 1 in the diagonal for row p − 1, besides the 1 in the row for the identity element, but in
that case p − 1 ≡ −1 (mod n), and hence is a trivial solution. See A206549 for these non-trivial solutions
in our case. The other inverses in these cyclic groups can be read off the cycle structure. One pairs the
first entry, the generator c of the cycle, with the second to last (the one before the 1, the second with
the third to last, etc. E.g., for n = 13 one has the inverses 7−1 ≡ 11 (Modd 13), 11−1 ≡ 7 (Modd 13),
and 3−1 ≡ 9 (Modd 13) , 9−1 ≡ 3 (Modd 13), and the left over 5 is self-inverse, as mentioned above. This
pairing of inverses is clear from the cyclic structure.
The next objective is to find some method to compute this uniquely existing smallest positive nontrivial solution s(p) of the congruence x2 ≡ 1 (M odd p) for given prime p = p̂(n) = A002144(n) for
the k(n) := p̂(n)4− 1 values given in A005098. This k-sequence starts with [1, 3, 4, 7, 9, 10, 13, 15, 18, ...].
The following l-algorithm will determine all these non-trivial
solutions s(p̂(n)) of the congruence x2 ≡
p
1 (M odd p̂(n)). We start with the Ansatz s(p) =
(o + 1) p − 1, with some odd number o. This
2
s + 1
means that we are looking for primes of the form
, and we are interested only in primes of the
o+1
2
form 4 k + 1, i.e., p̂(n). Because s(p̂) = (o + 1) p̂ − 1, s(p̂) is odd, say 2 K + 1. This implies
o+1
ˆ
p̂, which shows that o+1
2 K (K + 1) + 1 =
2 has to be odd, i.e., o = 4 l + 1. Now we compute
2
p−1
(2 K + 1)2 − (4 ˆl + 1) p̂
the even number 2 k =
=
= 4 T (K) − 2 ˆl − 2 k (4 ˆl + 1), where
2
2
, given in A000217. This implies that
we have introduced the triangular numbers T (K) := K (K+1)
2
2 k (2 ˆl + 1) = 2 T (K) − ˆl . Therefore, ˆl is even, say 2 l, and we have, with inputs k, l and K:
p = 4 k + 1, s2 (p̂) = (o + 1) p − 1, o = 8 l + 1 and s(p̂) = 2 K(p̂) + 1 . Because now
2 T (K) = 2 k + 2l (4 k + 1) = 2 l + 2k (4 l + 1), we find the ‘nice equation’
4 T (K) + 1 = (4 k + 1) (4 l + 1) .
(90)
For given prime p̂ of the form 4 k + 1, i.e. for k = p −4 1 (see A005098), we determine the minimal l ∈ N0 ,
such that k (1 + 4 l) +
√ l produces a triangular number, namely T (K) . The index of this triangular
8 T (K) + 1 − 1
number is then K =
. We have uses Maple [18] to compute these minimal l values. It
2
will be seen from the later proposition 15 that there is always such a minimal l, and not all non-negative
values appear. Precisely the entries of A094178 and 0 occur as minimal l values. We will from now on
denote this minimal l also by l (hoping to cause no confusion). If l = 0 then k is a triangular number,
and conversely. The non-vanishing l values are characterized by having as elements of the squarefree
kernel set of 4 l + 1 only distinct primes congruent 1 (mod 4). E.g., l = 2 will not appear because the
squarefree kernel of 9 consists of the prime 3 which is not congruent 1 (mod 4). In order to proof that
only numbers from A094178 can appear as positive l values, we use the ‘nice equation’ and observe that
38
4 T (K) + 1 = K 2 + (K + 1)2 , i.e., it is the sum to two neighboring squares with gcd(K, K + 1) = 1
(shown by indirect proof, see the remark following definition 3 ). Now a theorem ensures that all prime
factors of 4 T (K) + 1 are congruent 1 (mod 4). See the theorem 3.20 on p. 164 of the reference [23].
From the ‘nice equation’ we have either 4 l + 1 = 1, i.e., l = 0, or, if l is non-vanishing, then the
squarefree kernel set of 4 l + 1 has only primes congruent 1 (mod 4). Therefore l can be 0 or it is from
A094178, i.e., from [0, 1, 3, 4, 6, 7, 9, 10, 13, 15, 16, 18, 21, 22, 24, 25, 27, 28, ...]. The l = 0 value appears
for the primes [5, 13, 41, 61, 113, 181, 313, 421, ...], which is the sequence A027862. In this case k = p̂−1
4
is a triangular number. l = 1 for the primes [17, 29, 53, 73, 109, 137, 281, 397, 449, 593, 757, ...] This is the
sequence A207337(n), n ≥ 2. These are the primes p̂ such that 5 p̂ = 4 T (K) + 1 for some K = K(p̂).
2
E.g., 5 17 = 85 = 4 21 = 1, thus K(17) = 6. These are also the primes of the form
p (m +1)/10. Just use
2 (4 l + 1) p̂ + 1 − 1
T (K) − l
, (see eq. (90) and K =
.
m = m(K) = 2 K + 1. In general, k =
4l + 1
2
E.g., p̂ = 53 belongs to l = 1 and K = 11, and T (K) = 66. This checks with the ‘nice equation’. We
do not give here the tables of the sequences l(n) = l(p̂(n)) and K(n) = K(l(n)) (this incorrect renaming
of arguments should not lead to confusion) for given p̂(n) because after the next proposition 17 we will
find another way to obtain the K(n) numbers directly, hence the corresponding l(n) numbers, and from
these the desired s(p̂(n)) solutions from s(p̂(n)) = 2 K(n) + 1.
Proposition 17: Congruence 4 T(X) + 1 = X2 + (X + 1)2 ≡ 0 (mod p)
There are precisely two incongruent solutions of the congruence f (X) := 2 X 2 + 2 X + 1 ≡ 0 (mod p),
provided p = p̂(n) = A002144(n), n ≥ 1. The smallest positive representative will be called K(n), and
the next larger incongruent one is then K2(n) := p̂(n) − 1 − K(n).
Proof: Because of the degree 2 of f this congruence has at most two incongruent solutions. We shall
see that in fact there are two. This congruence is reduced to a problem of quadratic residues, following
a standard prescription (see, e.g., [21], pp. 132-3). The discriminant of f is D = 22 − 4 2 1 = −4.
Multiplying f by 8 yields (4 X + 2)2 + 4 ≡ 0 (mod 8 p). With Y = 4 X + 2 this is Y 2 = D (mod n),
with a composite modulus n = 23 p. This is a quadratic residue problem, but gcd(D, n) is not 1, but 4.
Theorem 77 of [21] is applied with d = 4, e = 2, f = 1, a + 1 = −1 and n1 = 2 p Thus the problem is
reduced to the quadratic residue problem Z 2 ≡ −1 (mod 2 p). This is solved by studying the congruences
for the powers of primes, here just 2 and p, separately (see e.g., [21], sect. 26, pp. 83-5). The congruence
modulo 2 has only
thesolution +1 (because −1 ≡ +1 (mod 2)), and modulo p one can consult the
p−1
−1
Legendre symbol
which is (−1) 2 (see e.g., [23], Theorem 3.2 (1), p. 132). Therefore −1 is a
p
quadratic residue modulo p if and only if this symbol is +1, demanding that p ≡ 1 (mod 4), i.e., p = p̂
from A002144. Call the smallest positive solution x0 , then p̂ − x0 is also an incongruent solution modulo
p̂, and two is the maximal number of solutions because of the degree 2 of this congruence. This implies
that there are 1 · 2 = 2 incongruent solutions of this congruence modulo 2 p (see e.g., [21], Theorem 46,
p. 84). Returning to the original problem this proves that there are also two incongruent solutions. If the
smallest positive solution for p̂(n) = A207337(n) is called K(n), then the next larger incongruent solution
of 4 T (X) + 1 = X 2 + (X + 1)2 ) ≡ 0 (mod p̂(n)) is K2(n) := p̂(n) − 1 − K(n), which is obvious.
The pair of sequences of all positive solutions modulo 5, 13, and 17 are given in A047219, A212160 and
A212161, respectively, where in each case the even indexed members are the positive solutions congruent
to K(n) and the odd indexed ones are the positive solutions congruent to K2(n), i.e., a(2 k) = k p̂(n) +
K(n) and a(2 k + 1) = k p̂(n) + K2(n), k ≥ 0. For the three given examples n = 1, 2, 3 respectively.
The companions (K(n), K2(n)) have been computed with Maple [18] and they can be found as A212353
and A212354. The first entries are for K: [1, 2, 6, 8, 15, 4, 11, 5, 13, 27, 37, 45, 16, 7, 18, 52, 64, ...] and for
K2: [3, 10, 10, 20, 21, 36, 41, 55, 59, 61, 59, 55, 92, 105, 118, 96, 92, 126, ...] corresponding to the primes p̂:
[5, 13, 17, 29, 37, 41, 53, 61, 73, 89, 97, 101, 109, 113, 137, 149, 157, 173, ...]. Note that proposition 17 yields
directly the searched for nontrivial solution of s(p̂(n))2 ≡ +1 (Modd p̂(n)) via s(p̂(n)) = 2 K(n) + 1.
Compare this with the s values given in A206549 with first entries [3, 5, 13, 17, 31, 9, 23, 11, 27, 55, 75, 91, 33, 15, 37, 10
39
p̂ − 3
b
iff s2(n)
:= K2(n) + 1 ≥ p̂(n)
Because ŝ(n) := s(p̂(n)) = 2 K(n) + 1 ≤ p̂(n) − 2 , i.e., K(n) ≤
2
b
and ŝ(n) ≤ s2(n),
n ≥ 1. Thus only ŝ(n) ∈ M(p̂), the restricted odd residue class Modd p̂, and the
b
solution s2(n)
is discarded.
This proposition and the ‘nice equation’ eq. (90) show that the l−algorithm
pro
from above will indeed
4 T (K(n)) + 1
− 1 /4. The
duce a solution l, related to the existing K(n) for given p̂(n), via l(n) =
p̂(n)
smallest positive non-trivial solution s(p̂(n)) of the congruence x2 ≡ +1 (M odd
p̂(n)) is then
2 K(n) + 1.
4 120, + 1
See K(n) = A212353(n), n ≥ 1. E.g., n = 5, p̂(5) = 37, K(5) = 15, l =
− 1)/4 = 3,
37
2
31
ŝ(n) = 31, with
= 25 = o and 312 = 961 = −1 (mod 37) = 36.
37
Q
As a application involving s(p̂(n)) we derive the analog of Wilson’s theorem
R(p) = (p − 1)! ≡
(p − 1) (mod p) ≡ −1 (mod p), for each prime p (see e.g., [2], Theorem 5.24, p. 116, or [10], Theorem 80,
p. 68), where R(n)is the set of the representatives of the smallest positive reduced residue system mod n
for n ≥ 2 which is of order ϕ(n). In the Modd p case we have to replace the set R(n) by the set M(n) of
eq. (65) with order δ(n).
Proposition 18: Analog of Wilson’s theorem for Modd p
is odd
if p−1
1
Y
2
(Modd p) ,
M(p) = (p − 2)!! ≡
p−1
ŝ(n) if 2 is even, p = p̂(n), n ∈ N
(91)
where ŝ(n) = s(p̂(n)), p̂(n) = A002144(n), stands for the above treated nontrivial root Modd n, i.e., the
solution of the congruence x2 ≡ 1 (Modd n) which is not 1, if it exists. Note that −1 ≡ +1 (Modd n),
n ≥ 2 (see lemma 10). For the double factorials (p − 2)!! see A207332.
Proof: The multiplicative group group M odd p, p a prime, is the cyclic group Zδ(p) from proposition
14. The order δ(p) of this group is even if and only if p = p̂, i.e., a prime 1 (mod 4). In this case we
know from above that the non-trivial ŝ(n) = s(p̂(n)) exists, and an Q
algorithm for finding it has been
given. Besides the unit element 1 and this ŝ(n) all other factors in
M(p̂) can be paired such that
their product is 1 (Modd p̂) (see the discussion after proposition 16), leaving only s(p̂). In the other case,
when p ≡ 3 (mod p), the group order is odd. Then all numbers besides 1 can be paired in the product to
produce 1 (Modd p), and the result is therefore 1.
For the cyclic Galois groups belonging to C(n, x) the list of the smallest positive primitive roots r(n)
(i.e., the smallest element from M(n) which generate these cyclic groups) are found under A206550. As
mentioned above, we do not have a formula for those n > 3 values whose cycle does not start with 3 as
smallest positive primitive root, like for n = [6, 9, 13, 14, 15, 18, 21, 26, 27, 33, 37, 38, 39, ...].
Open problems:
• Proof of the conjecture on the q−sequence related to the discriminant of the minimal C(n, x) polynomials; or find a counterexample.
• Characterization of the values n for which the Galois group Gn is non-cyclic.
• Characterization of the values n for which the cycle of the cyclic multiplicative group Modd n ∼
= Gn is
generated by 3.
• More theorems on multiplicative Modd n arithmetic.
40
Appendix A
The proof of eq. (30), with ϕ(n)
replaced by the product-formula given there a bit later, is a (nice)
n
application of PIE (the principle of inclusion and exclusion. See e.g., [5], Theorem 4.2, pp. 134 ff).
As mentioned above, this formula uses only the distinct prime factors of n, the elements of the set
sqf kset(n), the set of primes of the squarefree kernel of n. The product formula for ϕ(n)
n can be read
as the generating function for the elementary symmetric functions (here polynomials) in the variables
1
pj , j = 1, 2, ..., M (n). M (n), also called ω(n), is given in [27] as A001221(n).
M (n)
M (n)
Y
X
ϕ(n)
1
1
1
r
=
, ...,
,
=
(−1) σr
1−
n
pj
p1
pM (n)
r=0
(92)
j=1
with σ0 = 1, and symbolically σr =
X
r
1
, where the sum extends over the
p. · · · p.
M
r
terms with r
with increasing indices. E.g., M = 3 with σ2 = p11p2 + p11p3 + p21p3 .
n−1
X
k one starts, at the zeroth step (r = 0), with the unrestricted sum which is
To calculate
factors
1
p.
n
2
(n−1),
k=1
gcd(k,n)=1
and subtracts, in the next step (r = 1), the sum of all multiples of each prime dividing n which are
n
− 1
pj
X
n n
− 1 , for j = 1, 2, ..., M (n). This
k =
≤ (n − 1). For the pj -multiples this sum is pj
2 pj
k=1
n X n
leads, in step r = 1 to the subtraction of
− 1 . Now in this subtraction all multiples of the
2
pj
j
product of two different pj s appeared twice, therefore one has, in the next step (r = 2), to add them
n
− 1
pi pj
X
X
n X
n
k =
once. This is done by +1
pi pj
− 1 . In step r = 3 one concentrates
2
p. p.
i<j
i<j
k=1
on products of three different primes pi1 pi2 pi3 with i1 < i2 < i3 . Now such a 3−product appeared
once in step r = 0 ( in the unrestricted sum), −3 times in step r = 1, originating from the multiples
(pi1 pi2 ) · pi3 , (pi1 pi3 ) · pi2 and (pi2 pi3 ) · pi1 , and +3 times in step r = 2, from the multiples pi1 · ( pi2 pi3 ),
pi2 · ( pi1 pi3 ), and pi3 · ( pi1 pi2 ). Therefore, up to this stage, each such 3−product appeared once too
much, and it is subtracted in this step r = 3 when all these 3−products terms are summed. Now the
pattern starts to become clear. Up to, and including, step r = 3 one has for each 4−product the counting
1 − 4 + 6 − 4 = −1. Hence in step r = 4 one adds once the sum over all multiples of each such
4−product. In general, in step r this counting will produce (−1)r+1 (from the alternating row r = 4 in
the Pascal triangle A007318), and one will therefore add (−1)r times the sum over all multiples of each
such r−product. This yields
"
n−1
X
X n
X n
n
k =
(n − 1) −
− 1 +
− 1 ... − +
2
p i1
p i1 p i2
k=1
i1
gcd(k,n)=1
... (−1)M (n)
i1 <i2
X
i1 <...<iM (n)
n
p1 p2 · · · pM (n)
− 1 .
(93)
Now the ±1 terms all cancel if one takes into account the number of terms of each r− sum, which is M
r .
This is due to the vanishing alternating sum over Pascal triangle’s row M . Thus the result becomes the
alternating sum over the elementary symmetric polynomials with the reciprocal distinct prime factors of
41
n as variables.
n−1
X
k=1
gcd(k,n)=1
M (n)
n2 X
1
1
r
k =
(−1) σr
, ...,
.
2
p1
pM (n)
(94)
r=0
As mentioned above this is written as the generating function of these polynomials and the final result
is with the second part of eq. (92)
n−1
X
k =
n−1
X
k is indeed
k=1
gcd(k,n)=1
(95)
j=0
k=1
gcd(k,n)=1
2
This shows that s(n) := 2
n
M (n)
1
n2 Y
1−
.
2
pj
ϕ(n)
, which appears in [27] as A076512(n)/A109395(n).
n
s(n) is identical for all n with the same distinct prime factors, independent of their multiplicity, i.e., it
depends only on the elements of sqf set(n). From the multiplicativity of ϕ one sees that this scaled sum
s(n) is also multiplicative. E.g., s(12) = s(6) = s(2) s(3).
Appendix B
For the proof of eq. (56) we follow [24], p.33, and consider the rewritten version
n−1
1 if n is odd ,
Y
k
1 − e2 π i n =
k=1
0 if n is even .
(96)
k
In order to see that this is identical to eq. (56), just extract eπ i n , leading to the (2 cos) factors under
n−1
1
the product, and in front summing in the exponent leads to the factor e 2 π i (n−1) = in−1 = (−1) 2 .
The l.h.s. of eq. (96) is then computed using (for the last step see e.g., [9], Exercise 50, p. 149)
2
1 + z + z + ... + z
n−1
n−1
Y
k
zn − 1
(z − e2 π i n ) ,
=
=
z − 1
(97)
k=1
where one specializes to z = −1. (By putting z = +1 one finds the result
with 1 for n = 1.)
42
n−1
Y
k=1
k
2 sin π
n
= n , n ≥ 2,
References
[1] Abramowitz and I. A. Stegun, eds., Handbook of Mathematical Functions, National Bureau of Standards Applied Math. Series 55, Tenth Printing, reprinted as Dover publication, seventh printing
1968, New York, 1972
[2] T. Apostol, Introduction to Analytic Number Theory, Springer, 1986
[3] E. Artin, with a section by N. A, Milgram, Galoissche Theorie, 3. Auflage, Harri Deutsch, 1988.
English version: Galois theory. Lectures delivered at the University of Notre Dame, 1966
[4] Chan-Lye Lee and K. B. Wong, On Chebyshev’s polynomials and certain combinatorial identities,
2009, Bull. Malaysian Sciences Soc.
http://www.emis.de/journals/BMMSS/accepted_papers.htm.
[5] Ch. A. Charalambides, Enumerative Combinatorics, Chapman &Hall/CRC, 2002
[6] D. A. Cox, Galois Theory, Wiley, 2004
[7] J.-P. Escofier, Galois Theory, Springer, 2001
[8] Yimin Ge’s Math Blog, 2009,
http://yiminge.wordpress.com/2009/01/22/all-groups-of-order-n-are-cyclic-iff/
[9] R. L. Graham, D. E. Knuth and O. Patashnik, Concrete Mathematics, Addison-Weseley, Reading,
Massachusetts, 1991
[10] G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed,. Oxford Science
publications, 2003
[11] The On-Line Encyclopedia of Integer SequencesT M
http://oeis.org, 2011
(OEIS), published electronically at
[12] S. Lang, Algebra, revised third ed., Springer, 2002
2π
, a link under [27] A181875.
[13] W. Lang, Minimal Polynomials for cos
n
[14] W. Lang, Divisor Product Representation for Natural Numbers, a link under [27] A007955.
[15] W. Lang, A181878, Coefficient array for square of Chebyshev S-polynomials, a link under
[27] A181878.
[16] A. Lazarev, Galois Theory, lecture notes with problems and solutions, University of Leicester,
http://www2.le.ac.uk/departments/mathematics/extranet/staff-material/staff-profiles/al179
[17] D. H. Lehmer, A Note on Trigonometric Algebraic Numbers, Am. Math. Monthly 40,8 (1933) 165-6.
[18] MapleT M , http://www.maplesoft.com/
[19] Wolfram
Mathworld,
http://mathworld.wolfram.com/SylvesterCyclotomicNumber.html,
Sylvester Cyclotomic Number.
[20] W. Magnus, F. Oberhettinger, R. P. Soni, Formulas and theorems for the special functions of mathematical physics. 3rd enlarged ed., 1966, Springer, Berlin.
[21] T. Nagell, Introduction to Number Theory, Chelsea Pub. Comp., N.Y., 2nd ed., 1964.
43
[22] I. Niven, Irrational Numbers, The Math. Assoc. of America, second printing, distributed by John
Wiley and Sons, 1963.
[23] I. Niven, H. S. Zuckerman and H. L. Montgomery, An Introduction to the Theory Of Numbers, Fifth
Edition, John Wiley and Sons, Inc., NY, 1991.
[24] R. Remmert, Funktionentheorie 2, Springer, 1991
[25] P. Ribenboim, Classical Theory of Algebraic Numbers, Springer, 2001.
[26] Th. J. Rivlin, Chebyshev Polynomials. From Approximation Theory to Algebra and Number Theory,
second edition, Wiley-Interscience, 1990.
[27] The On-Line Encyclopedia
http://oeis.org.
of
Integer
Sequences
(2010),
published
electronically
at
[28] A. Speiser, Die Theorie der Gruppen endlicher Ordnung, Vierte Auflage, 1956, Birkhäuser, Basel
[29] P. Steinbach, Golden Fields: A Case for the Heptagon, Mathematics Magazine, 70,1 (1997) 22-31,
http://www.jstor.org/pss/2691048.
[30] W. Watkins and J. Zeitlin, The Minimal Polynomial of cos(2π/n), Am. Math. Monthly 100,5 (1993)
471-4, http://www.jstor.org/pss/2324301.
[31] Wikipedia, Cycle graph (algebra) http://en.wikipedia.org/wiki/Cycle_graph_(algebra) and
List of small groups http://en.wikipedia.org/wiki/List_of_small_groups.
Keywords: regular n-gons, algebraic number, minimal polynomial, zeros, factorization, congruences,
Galois theory, cycle graphs.
AMS MSC numbers: 11R32, 11R04, 08B10, 13F20, 12D10, 13P05
OEIS A-numbers: A000001, A000005, A000007, A000010, A000012, A000035, A000265, A000668, A000688,
A001221, A002110, A002144, A003277, A004124, A005013, A005098, A005117, A006053, A006054,
A007310, A007318, A007775, A007814, A007947, A007955, A008683, A023022, A024556, A027862,
A033949, A045572, A049310, A0503384, A052547, A047219, A053120, A055034, A066651, A076512,
A077998, A084101, A085810, A088520, A090298, A092260, A094178, A106803, A109395, A110551,
A113807, A116423, A120757, A127672, A128672, A130290, A130777, A147600, A162699, A181875,
A181876, A181878, A181879, A181880, A193376, A193377, A193679, A193680, A193681, A193682,
A203571, A203572, A203575, A204453, A204454, A204457, A204458, A206543, A206544, A206545,
A206546, A206547, A206548, A206549, A206550, A206551, A206552, A207333, A207334, A207337,
A210845, A212160, A212161, A212353, A212354, A215041, A215046, A216319, A216320, A216322,
A216326, A255237. A282624.
44
Table 1: Reduced DSR-algebras (over Q), n = 3, ..., 12.
n
ρ ≡ ρ(n)
3
1
4
√
5
ϕ =
6
√
7
2 cos
8
p
9
2 cos
10
ϕ
11
2 cos
12
p
..
.
reduced DSR-algebra
2
√
1
(1 + 5)
2
3
2+
√
π
7
√
π
9
2
3− ϕ
2+
π
11
√
3
δ(n)
DSR − basis
ρ2 = 1
1
<1>
ρ2 = 2
2
< 1, ρ >
ρ2 = ρ + 1
2
< 1, ρ >
ρ2 = 3, [σ = ρ2 − 1 = 2]
2
< 1, ρ >
ρ2 = 1 + σ, σ 2 = 1 + ρ + σ, ρ σ = ρ + σ
3
< 1, ρ, σ >
ρ2 = 1 + σ, σ 2 = 2 σ + 1, τ 2 = 2 (1 + σ),
ρσ = ρ + √
τ, ρ τ = √
2 σ, σ τ = 2 ρ + τ ,
(σ = 1 + 2, τ = 2 ρ)
4
< 1, ρ, σ, τ >
ρ2 = 1 + σ, σ 2 = 2 + ρ + σ, ρ σ = 2 ρ + 1,
(σ = ρ2 − 1), [τ = ρ (σ − 1) = 1 + ρ]
3
< 1, ρ, σ >
ρ2 = 1 + σ, σ 2 = −1 + 3 σ, τ 2 = −1 + 4 σ,
ρ σ = ρ + τ, ρ τ = 3 σ − 2, σ τ = ρ + 2 τ ,
(σ = ρ2 − 1 = 1 + ϕ, τ = ρ (σ − 1),
√
= (1 + ϕ) 3 − ϕ)
[ω = 2 (−1 + σ) = 2 ϕ]
4
< 1, ρ, σ, τ >
ρ2 = 1 + σ, σ 2 = 1 + σ + ω,
τ 2 = 1 + σ + τ + ω, ω 2 = 1 + ρ + σ + τ + ω,
ρ σ = ρ + τ, ρ τ = σ + ω, ρ ω = τ + ω,
σ τ = ρ + τ + ω, σ ω = σ + τ + ω,
τ ω = ρ + σ + τ + ω,
(σ = ρ2 − 1, τ = ρ (σ − 1), ω = σ (σ − 1) − 1)
5
< 1, ρ, σ, τ, ω >
ρ2 = 1 + σ, σ 2 = 2 (1 + τ ), τ 2 = 3 (1 + σ),
ρ σ = ρ + τ, ρ τ = 1 + 2 σ, σ τ = 3 ρ + τ ,
[ω = 1 + σ, χ = 2 ρ],
√
√
√
2
2
(σ = ρ − 1 = 1 + 3, τ =
(3 + 3 ))
2
4
< 1, ρ, σ, τ >
π
jnk
= S(k − 1, ρ(n)), k = 1, ...,
,
n
2
R1 = 1, R2 ≡ ρ, R3 ≡ σ, R4 ≡ τ, R5 ≡ ω, R6 ≡ χ (dependence on n suppressed).
ρ(n) := 2 cos
(n)
, Rk
In round brackets the values for the basis elements are given in terms of ρ. In square brackets the linear
dependent DSRs are given. Boxed n-numbers indicate linear dependent DSRs.
45
Table 2: Minimal polynomials of 2 cos
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
..
.
π
n
for n = 1, 2, ..., 30.
C(n, x)
x+2
x
x−1
x2 − 2
x2 − x − 1
x2 − 3
x3 − x2 − 2 x + 1
x4 − 4 x2 + 2
x3 − 3x − 1
x4 − 5 x2 + 5
x5 − x4 − 4 x3 + 3 x2 + 3 x − 1
x4 − 4 x2 + 1
x6 − x5 − 5 x4 + 4 x3 + 6 x2 − 3 x − 1
x6 − 7 x4 + 14 x2 − 7
x4 + x3 − 4 x2 − 4 x + 1
x8 − 8 x6 + 20 x4 − 16 x2 + 2
x8 − x7 − 7 x6 + 6 x5 + 15 x4 − 10 x3 − 10 x2 + 4 x + 1
x6 − 6 x4 + 9 x2 − 3
x9 − x8 − 8 x7 + 7 x6 + 21 x5 − 15 x4 − 20 x3 + 10 x2 + 5 x − 1
x8 − 8 x6 + 19 x4 − 12 x2 + 1
x6 + x5 − 6 x4 − 6 x3 + 8 x2 + 8 x + 1
x10 − 11 x8 + 44 x6 − 77 x4 + 55 x2 − 11
x11 − x10 − 10 x9 + 9 x8 + 36 x7 − 28 x6 − 56 x5 + 35 x4 +
35 x3 − 15 x2 − 6 x + 1,
x8 − 8 x6 + 20 x4 − 16 x2 + 1
x10 − 10 x8 + 35 x6 − x5 − 50 x4 + 5 x3 + 25 x2 − 5 x − 1
x12 − 13 x10 + 65 x8 − 156 x6 + 182 x4 − 91 x2 + 13
x9 − 9 x7 + 27 x5 − 30 x3 + 9 x − 1
x12 − 12 x10 + 53 x8 − 104 x6 + 86 x4 − 24 x2 + 1
x14 − x13 − 13 x12 + 12 x11 + 66 x10 − 55 x9 − 165 x8 +
120 x7 + 210 x6 − 126 x5 − 126 x4 + 56 x3 + 28 x2 − 7 x − 1
x8 − 7 x6 + 14 x4 − 8 x2 + 1
46
Table 3: A187360(n, m) coefficient
array of
π
minimal polynomials of 2 cos
, rising powers
n
n/m
0
1
2
3
1
2
1
2
0
1
3
-1
1
4
-2
0
1
5
-1
-1
1
6
-3
0
1
7
1
-2
-1
1
8
2
0
-4
0
9
-1
-3
0
1
10
5
0
-5
0
1
11
-1
3
3
-4
-1
12
1
0
-4
0
1
13
-1
-3
6
4
-5
-1
1
14
-7
0
14
0
-7
0
1
15
..
.
1
-4
-4
1
1
47
4
5
6
1
1
...
Table 4: Zeros of C(n, x) in power basis (rising powers of ρ(n)) for n = 1, 2, ..., 30.
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
..
.
coefficients of C-zeros in power basis < ρ0 , ..., ρδ(n)−1 >
[[−2]]
[[0]]
[[1]]
[[0, 1], [0, −1]]
[[0, 1], [1, −1]]
[[0, 1], [0, −1]]
[[0, 1], [−1, −1, 1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 3, 0, −1], [0, −1]]
[[0, 1], [−2, −1, 1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 3, 0, −1], [0, −1]]
[[0, 1], [0, −3, 0, 1], [1, 2, −3, −1, 1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, 4, 0, −1], [0, −4, 0, 1], [0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [1, −3, −3, 4, 1, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −5, 0, 5, 0, −1], [0, 3, 0, −1], [0, −1]]
[[0, 1], [−2, 3, 1, −1], [−1, −4, 0, 1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 7, 0, −14, 0, 7, 0, −1], [0, −5, 0, 5, 0, −1],
[0, 3, 0, −1], [0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [−1, 4, 6, −10, −5, 6, 1, −1],
[2, 0, −9, 0, 6, 0, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, 5, 0, −5, 0, 1], [0, −4, 0, 5, 0, −1], [0, 4, 0, −5, 0, 1], [0, −5, 0, 5, 0, −1], [0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [1, 4, −10, −10, 15, 6, −7, −1, 1],
[−2, 0, 16, 0, −20, 0, 8, 0, −1], [2, 0, −9, 0, 6, 0, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 8, 0, −18, 0, 8, 0, −1], [0, −8, 0, 18, 0, −8, 0, 1],
[0, 7, 0, −14, 0, 7, 0, −1], [0, 3, 0, −1], [0, −1]]
[[0, 1], [0, 5, 0, −5, 0, 1], [2, 3, −4, −1, 1], [−3, −9, 1, 6, 0, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 9, 0, −30, 0, 27, 0, −9, 0, 1],
[0, −9, 0, 30, 0, −27, 0, 9, 0, −1], [0, 7, 0, −14, 0, 7, 0, −1], [0, −5, 0, 5, 0, −1], [0, 3, 0, −1], [0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 9, 0, −30, 0, 27, 0, −9, 0, 1],
[−1, −5, 15, 20, −35, −21, 28, 8, −9, −1, 1], [2, 0, −25, 0, 50, 0, −35, 0, 10, 0, −1],
[−2, 0, 16, 0, −20, 0, 8, 0, −1], [2, 0, −9, 0, 6, 0, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, −8, 0, 6, 0, −1], [0, 8, 0, −6, 0, 1],
[0, 7, 0, −14, 0, 7, 0, −1], [0, −5, 0, 5, 0, −1], [0, −1]]
[[0, 1], [0, −3, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 9, 0, −30, 0, 27, 0, −9, 0, 1],
[0, −10, 5, 30, −5, −27, 1, 9, 0, −1], [0, 10, −15, −15, 20, 7, −8, −1, 1], [−2, 0, 16, 0, −20, 0, 8, 0, −1],
[2, 0, −9, 0, 6, 0, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 9, 0, −30, 0, 27, 0, −9, 0, 1],
[0, −11, 0, 55, 0, −77, 0, 44, 0, −11, 0, 1], [0, 11, 0, −55, 0, 77, 0, −44, 0, 11, 0, −1],
[0, −9, 0, 30, 0, −27, 0, 9, 0, −1], [0, 7, 0, −14, 0, 7, 0, −1], [0, −5, 0, 5, 0, −1], [0, 3, 0, −1], [0, −1]]
[[0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [−2, 7, 1, −14, 0, 7, 0, −1], [2, −5, −4, 5, 1, −1],
[2, −1, −16, 0, 20, 0, −8, 0, 1], [−2, 0, 16, 0, −20, 0, 8, 0, −1], [−2, 0, 4, 0, −1], [2, 0, −1]]
continued on next page
Note: the higher power coefficients not shown are all zero.
Example: n = 27: the first zero of C(27, x) is ρ(27), the second zero is 5 ρ(27) − 5 ρ(27)3 −
1 ρ(27)5 , etc.
48
Table 4 continued: Zeros of C(n, x) in power basis (rising powers of ρ(n)) for
n = 28, ..., 30.
n
28
29
30
..
.
power basis coefficients for < ρ0 , ..., ρδ(n)−1 >
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, 9, 0, −30, 0, 27, 0, −9, 0, 1], [0, −11, 0, 55, 0, −77, 0, 44, 0, −11, 0, 1],
[0, 12, 0, −67, 0, 96, 0, −52, 0, 12, 0, −1], [0, −12, 0, 67, 0, −96, 0, 52, 0, −12, 0, 1],
[0, 11, 0, −55, 0, 77, 0, −44, 0, 11, 0, −1], [0, −9, 0, 30, 0, −27, 0, 9, 0, −1], [0, −5, 0, 5, 0, −1], [0, 3, 0, −1],
[0, −1]]
[[0, 1], [0, −3, 0, 1], [0, 5, 0, −5, 0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, 9, 0, −30, 0, 27, 0, −9, 0, 1],
[0, −11, 0, 55, 0, −77, 0, 44, 0, −11, 0, 1], [0, 13, 0, −91, 0, 182, 0, −156, 0, 65, 0, −13, 0, 1],
[1, −7, −21, 56, 70, −126, −84, 120, 45, −55, −11, 12, 1, −1], [−2, 0, 36, 0, −105, 0, 112, 0, −54, 0, 12, 0, −1],
[2, 0, −25, 0, 50, 0, −35, 0, 10, 0, −1], [−2, 0, 16, 0, −20, 0, 8, 0, −1], [2, 0, −9, 0, 6, 0, −1], [−2, 0, 4, 0, −1],
[2, 0, −1]]
[[0, 1], [0, −7, 0, 14, 0, −7, 0, 1], [0, −7, 0, 22, 0, −13, 0, 2], [0, 4, 0, −13, 0, 7, 0, −1], [0, −4, 0, 13, 0, −7, 0, 1],
[0, 7, 0, −22, 0, 13, 0, −2], [0, 7, 0, −14, 0, 7, 0, −1], [0, −1]]
49
c
c
Table 5: Extended set M(n),
first differences △ M(n),
and floor-pattern F(n) for composed odd squarefree
modulus n.
c
△ M(n)
n(m)
1
2
3
15
21
33
8
12
20
[0, 1, 7, 11, 13, 17]
[0, 1, 5, 11, 13, 17, 19, 23]
[0, 1, 5, 7, 13, 17, 19, 23, 25, 29, 31, 35]
[1, 6, 4, 2, 4]
[1, 4, 6, 2, 4, 2, 4]
[1, 4, 2, 6, 4, 2, 4, 2, 4, 2, 4]
4
35
5·7
39
3 · 13
51
3 · 17
55
5 · 11
24
[0, 1, 3, 9, 11, 13, 17, 19, 23, 27, 29, 31, 33, 37]
[1, 2, 6, 2, 2, 4, 2, 4, 4, 2, 2, 2, 4]
24
[0, 1, 5, 7, 11, 17, 19, 23, 25, 29, 31, 35, 37, 41]
[1, 4, 2, 4, 6, 2, 4, 2, 4, 2, 4, 2, 4]
32
[0, 1, 5, 7, 11, 13, 19, 23, 25, 29,
31, 35, 37, 41, 43, 47, 49, 53]
[0, 1, 3, 7, 9, 13, 17, 19, 21, 23, 27, 29,
29, 31, 37, 39, 41, 43, 47, 49, 51, 53, 57]
[1, 4, 2, 4, 2, 6, 4, 2, 4, 2, 4, 2, 4, 2, 4, 2, 4]
8
57
3 · 19
36
[0, 1, 5, 7, 11, 13, 17, 23, 25, 29, 31,
31, 35, 37, 41, 43, 47, 49, 53, 55, 59]
[1, 4, 2, 4, 2, 4, 6, 2, 4,
2, 4, 2, 4, 2, 4, 2, 4, 2, 4]
9
65
5 · 13
48
[0, 1, 3, 7, 9, 11, 17, 19, 21, 23, 27, 29, 31, 33,
33, 37, 41, 43, 47, 49, 51, 53, 57, 59, 61, 63, 67]
[1, 2, 4, 2, 2, 6, 2, 2, 2, 4, 2, 2,
2, 4, 4, 2, 4, 2, 2, 2, 4, 2, 2, 2, 4]
10
69
3 · 23
44
[0, 1, 5, 7, 11, 13, 17, 19, 25, 29, 31, 35, 37,
41, 43, 47, 49, 53, 55, 59, 61, 65, 67, 71]
[1, 4, 2, 4, 2, 4, 2, 6, 4, 2, 4, 2,
4, 2, 4, 2, 4, 2, 4, 2, 4, 2, 4]
11
77
7 · 11
60
[0, 1, 3, 5, 9, 13, 15, 17, 19, 23, 25, 27,
29, 31, 37, 39, 41, 43, 45, 47, 51, 53,
57, 59, 61, 65, 67, 69, 71, 73, 75, 79]
[1, 2, 2, 4, 4, 2, 2, 2, 4, 2, 2, 2, 2, 6, 2, 2,
2, 2, 2, 4, 2, 4, 2, 2, 4, 2, 2, 2, 2, 2, 4]
5
6
50
7
continued
2 δ(n)
c
M(n)
m
40
[1, 2, 4, 2, 4, 4, 2, 2, 2, 4, 2,
2, 6, 2, 2, 2, 4, 2, 2, 2, 4]
F(n)
[2, 1, 0, 1, 0, 1, 2, 0]
[1, 2, 0, 1, 0, 1, 0, 1, 0, 2, 1, 0]
[1, 0, 2, 1, 0, 1, 0, 1, 0, 1,
0, 1, 0, 1, 0, 1, 2, 0, 1, 0]
[0, 2, 0, 0, 1, 0, 1, 1, 0, 0, 0, 1,
0, 0, 0, 1, 1, 0, 1, 0, 0, 2, 0, 0]
[1, 0, 1, 2, 0, 1, 0, 1, 0, 1, 0, 1,
0, 1, 0, 1, 0, 1, 0, 2, 1, 0, 1, 0]
[1, 0, 1, 0, 2, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 0, 1, 0, 1, 0]
[0, 1, 0, 1, 1, 0, 0, 0, 1, 0, 0, 2, 0, 0,
0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 2, 0,
0, 1, 0, 0, 0, 1, 1, 0, 1, 0, 0]
[1, 0, 1, 0, 1, 2, 0, 1, 0, 1, 0, 1, 0, 1, 0,
1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0,
2, 1, 0, 1, 0, 1, 0]
[0, 1, 0, 0, 2, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1,
0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1,
0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 1, 0, 0]
[1, 0, 1, 0, 1, 0, 2, 1, 0, 1, 0, 1, 0, 1, 0,
1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
0, 1, 0, 1, 0, 1, 2, 0, 1, 0, 1, 0, 1, 0]
[0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0, 2, 0, 0,
0, 0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1,
0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0,
0, 2, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 0, 0]
c
c
Table 5 ctnd.: Extended set M(n),
first differences △ M(n),
and floor-pattern F(n) for composed odd squarefree
modulus n.
c
M(n)
c
△ M(n)
51
m
n(m)
2 δ(n)
12
85
5 · 17
64
[0, 1, 3, 7, 9, 11, 13, 19, 21, 23, 27, 29, 31,
33, 37, 39, 41, 43, 47, 49, 53, 57, 59, 61,
63, 67, 69, 71, 73, 77, 79, 81, 83, 87]
[1, 2, 4, 2, 2, 2, 6, 2, 2, 4, 2,
2, 2, 4, 2, 2, 2, 4, 2, 4, 4, 2,
2, 2, 4, 2, 2, 2, 4, 2, 2, 2, 4]
13
87
3 · 29
56
[0, 1, 5, 7, 11, 13, 17, 19, 23, 25, 31
35, 37, 41, 43, 47, 49, 53, 55, 59, 61,
65, 67, 71, 73, 77, 79, 83, 85, 89]
[1, 4, 2, 4, 2, 4, 2, 4, 2, 6, 4, 2, 4, 2, 4,
2, 4, 2, 4, 2, 4, 2, 4, 2, 4, 2, 4, 2, 4]
14
91
7 · 13
72
[1, 2, 2, 4, 2, 4, 2, 2, 4, 2, 2, 2, 2,
2, 4, 4, 2, 2, 2, 4, 2, 2, 2, 2, 2, 6,
2, 2, 2, 2, 4, 2, 2, 2, 2, 2, 4]
15
93
3 · 31
60
[0, 1, 3, 5, 9, 11, 15, 17, 19, 23, 25,
27, 29, 31, 33, 37, 41, 43, 45, 47,
51, 53, 55, 57, 59, 61, 67, 69, 71,
73, 75, 79, 81, 83, 85, 87, 89, 93]
[0, 1, 5, 7, 11, 13, 17, 19, 23, 25, 29, 35,
37, 41, 43, 47, 49, 53, 55, 59, 61, 65, 67,
71, 73, 77, 79, 83, 85, 89, 91, 95]
16
95
5 · 19
72
[0, 1, 3, 7, 9, 11, 13, 17, 21, 23, 27,
29, 31, 33, 37, 39, 41, 43, 47, 49,
51, 53, 59, 61, 63, 67, 69, 71, 73,
77, 79, 81, 83, 87, 89, 91, 93, 97]
[1, 2, 4, 2, 2, 2, 4, 4, 2, 4, 2, 2, 2,
4, 2, 2, 2, 4, 2, 2, 2, 6, 2, 2, 4, 2,
2, 2, 4, 2, 2, 2, 4, 2, 2, 2, 4]
17
105
3·5·7
48
[0, 1, 11, 13, 17, 19, 23, 29, 31, 37,
41, 43, 47, 53, 59, 61, 67, 71, 73,
79, 83, 89, 97, 101, 103, 107]
[1, 10, 2, 4, 2, 4, 6, 2, 6,
4, 2, 4, 6, 6, 2, 6, 4, 2,
6, 4, 6, 8, 4, 2, 4]
..
.
[1, 4, 2, 4, 2, 4, 2, 4, 2, 4, 6,
2, 4, 2, 4, 2, 4, 2, 4, 2, 4, 2,
4, 2, 4, 2, 4, 2, 4, 2, 4]
F(n)
[0, 1, 0, 0, 0, 2, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
1, 0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1,
0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 1, 0, 1, 0,
0, 0, 1, 0, 0, 0, 1, 0, 0, 2, 0, 0, 0, 1, 0, 0]
[1, 0, 1, 0, 1, 0, 1, 0, 2, 1, 0, 1, 0, 1,
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
0, 1, 0, 1, 2, 0, 1, 0, 1, 0, 1, 0, 1, 0]
[0, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1,
0, 0, 0, 1, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 1,
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0,
0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0]
[1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 0, 1, 0, 1, 0,
1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0,
1, 0, 1, 0, 2, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0]
[0, 1, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 1, 0, 0,
0, 1, 0, 0, 0, 2, 0, 0, 1, 0, 0, 0, 1, 0, 0,
0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0,
0, 0, 1, 0, 0, 2, 0, 0, 0, 1, 0, 0, 0, 1, 0,
0, 0, 1, 0, 1, 1, 0, 0, 0, 1, 0, 0]
[4, 0, 1, 0, 1, 2, 0, 2, 1, 0, 1, 2,
2, 0, 2, 1, 0, 2, 1, 2, 3, 1, 0, 1,
0, 1, 3, 2, 1, 2, 0, 1, 2, 0, 2, 2,
1, 0, 1, 2, 0, 2, 1, 0, 1, 0, 4, 0]
Table 6: Cycle structure of Gal (Q(ρ(n))/Q) for n = 1, 2, ..., 40.
n
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
..
.
cycles
[[0]] = [[1]]
[[1]]
[[1]]
[[3, 1]]
[[3, 1]]
[[5, 1]]
[[3, 5, 1]]
[[3, 7, 5, 1]]
[[5, 7, 1]]
[[3, 9, 7, 1]]
[[3, 9, 5, 7, 1]]
[[5, 1], [7, 1], [11, 1]]
[[7, 3, 5, 9, 11, 1]]
[[5, 3, 13, 9, 11, 1]]
[[7, 11, 13, 1]]
[[3, 9, 5, 15, 13, 7, 11, 1]]
[[3, 9, 7, 13, 5, 15, 11, 1]]
[[5, 11, 17, 13, 7, 1]]
[[3, 9, 11, 5, 15, 7, 17, 13, 1]]
[[3, 9, 13, 1], [7, 9, 17, 1], [11, 1], [19, 1]]
[[11, 5, 13, 17, 19, 1]]
[[3, 9, 17, 7, 21, 19, 13, 5, 15, 1]]
[[3, 9, 19, 11, 13, 7, 21, 17, 5, 15, 1]]
[[5, 23, 19, 1], [7, 1], [11, 23, 13, 1], [17, 1]]
[[3, 9, 23, 19, 7, 21, 13, 11, 17, 1]]
[[7, 3, 21, 9, 11, 25, 19, 23, 5, 17, 15, 1]]
[[5, 25, 17, 23, 7, 19, 13, 11, 1]]
[[3, 9, 27, 25, 19, 1], [5, 25, 13, 9, 11, 1], [17, 9, 15, 25, 23, 1]]
[[3, 9, 27, 23, 11, 25, 17, 7, 21, 5, 15, 13, 19, 1]]
[[7, 11, 17, 1], [13, 11, 23, 1], [19, 1], [29, 1]]
[[3, 9, 27, 19, 5, 15, 17, 11, 29, 25, 13, 23, 7, 21, 1]]
[[3, 9, 27, 17, 13, 25, 11, 31, 29, 23, 5, 15, 19, 7, 21, 1]]
[[5, 25, 7, 31, 23, 17, 19, 29, 13, 1]]
[[3, 9, 27, 13, 29, 19, 11, 33, 31, 25, 7, 21, 5, 15, 23, 1]]
[[3, 9, 27, 11, 33, 29, 17, 19, 13, 31, 23, 1]]
[[5, 25, 19, 23, 29, 1], [7, 23, 17, 25, 31, 1], [11, 23, 35, 25, 13, 1]]
[[5, 25, 23, 33, 17, 11, 19, 21, 31, 7, 35, 27, 13, 9, 29, 3, 15, 1]]
[[13, 17, 7, 15, 33, 27, 29, 3, 37, 25, 21, 31, 23, 5, 11, 9, 35, 1]]
[[7, 29, 31, 17, 37, 25, 19, 23, 5, 35, 11, 1]]
[[3, 9, 27, 1], [7, 31, 23, 1], [11, 39, 29, 1], [13, 9, 37, 1], [17, 31, 33, 1], [19, 39, 21, 1]]
Boxed and colored n-numbers indicate non-cyclic Galois groups. See Table 8.
Boldface numbers are nontrivial square roots Modd n, denoted by s in the text.
52
Table 7: Non-cyclic Galois groups Gal (Q(ζ(n))/Q) , n ≤ 100
n
ϕ(n)
cycle structure
no. of cycles
8
12
15
16
20
21
24
28
30
32
33
35
36
39
40
42
44
45
48
51
52
55
56
57
60
63
64
65
66
68
69
70
72
75
76
77
78
80
4
4
8
8
8
12
8
12
8
16
20
24
12
24
16
12
20
24
16
32
24
40
24
36
16
36
32
48
20
32
44
24
24
40
36
60
24
32
23
23
42 22
42 22
42 22
63
27
63
42 22
82 41 22
103
122 62
63
122 62
44 26
63
103
122 62
44 26
162 81 41 22
122 62
202 102
67
183
44 26
612
162 81 41 22
126
103
162 81 41 22
223
122 62
67
202 102
183
303
122 62
412 24
3
3
4
4
4
3
7
3
4
5
3
4
3
4
10
3
3
4
10
6
4
4
7
3
10
12
6
6
3
6
3
4
7
4
3
3
4
16
Continued on the next page.
53
Galois group
Z2
Z2
Z4
Z4
Z4
Z3
× Z2
× Z2
× Z2
× Z2
× Z2
× Z22
Z23
Z3 × Z22
Z4 × Z2
Z8 × Z2
Z5 × Z22
Z4 × Z3 × Z2
Z3 × Z22
Z4 × Z3 × Z2
Z4 × Z22
Z3 × Z22
Z5 × Z22
Z4 × Z3 × Z2
Z4 × Z22
Z16 × Z2
Z4 × Z3 × Z2
Z5 × Z4 × Z2
Z3 × Z23
Z9 × Z22
Z4 × Z22
Z32 × Z22
Z16 × Z2
Z42 × Z3
Z5 × Z22
Z16 × Z2
Z11 × Z22
Z4 × Z3 × Z2
Z3 × Z23
Z5 × Z4 × Z2
Z9 × Z22
Z5 × Z3 × Z22
Z4 × Z3 × Z2
Z42 × Z2
Table 7 continued: Non-cyclic Galois groups Gal (Q(ζ(n))/Q) , n ≤ 100
n
ϕ(n)
cycle structure
no. of cycles
Galoisgroup
84
85
87
88
90
91
92
93
95
96
99
100
..
.
24
64
56
40
24
72
44
60
72
32
60
40
67
164 82 44
282 142
107
122 62
128 68
223
303
362 182
84 43 26
303
202 102
7
10
4
7
4
16
3
3
4
13
3
4
Z3 × Z23
Z16 × Z4
Z7 × Z4 × Z2
Z5 × Z23
Z4 × Z3 × Z2
Z4 × Z32 × Z2
Z11 × Z22
Z5 × Z3 × Z22
Z9 × Z4 × Z2
Z8 × Z22
Z5 × Z3 × Z22
Z5 × Z4 × Z2
The cyclic group of order m is denoted by Zm . For all other values n ≤ 100 the Galois group is the
cyclic group Zϕ(n) .
Only independent cycles are counted, i.e., cycles which appear as sub-cycles of the given ones have been
omitted.
The notation, e.g., 162 81 41 22 , means that there are 2 cycles of order (length) 16, one cycle of order 8,
one cycle of order 4 and two cycles of order 2.
Direct products of identical cyclic groups are sometimes written in exponent form, e.g., Z22 stands for
Z2 × Z2 .
Boxed and colored n-numbers indicate where some non-cyclic Galois group appears for the first time.
Some of the cycle graphs are shown in Fig. 4.
54
Table 8: Non-cyclic Galois groups Gal (Q(ρ(n))/Q) , n ≤ 100
n
δ(n)
cycle structure
no. of cycles
Galoisgroup
12
20
24
28
30
36
40
42
44
48
52
56
60
63
65
66
68
70
72
76
78
80
84
85
88
90
91
92
96
100
..
.
4
8
8
12
8
12
16
12
20
16
24
24
16
18
24
20
32
24
24
36
24
32
24
32
40
24
36
44
32
40
23
42 22
42 22
63
42 22
63
46
63
103
82 41 22
122 62
122 62
44 26
64
122 62
103
162 81 41 22
122 62
122 62
183
122 62
84 44
67
162 81 41 22
202 102
122 62
124
223
162 81 41 22
202 102
3
4
4
3
4
3
6
3
3
5
4
4
10
4
4
3
6
4
4
3
4
8
7
6
4
4
4
3
6
4
Z2 × Z2
Z4 × Z2
Z4 × Z2
Z3 × Z22
Z4 × Z2
Z3 × Z22
Z4 × Z4
Z3 × Z22
Z5 × Z22
Z8 × Z2
Z4 × Z3 × Z2
Z4 × Z3 × Z2
Z4 × Z22
Z32 × Z2
Z4 × Z3 × Z2
Z5 × Z22
Z16 × Z2
Z4 × Z3 × Z2
Z4 × Z3 × Z2
Z9 × Z22
Z4 × Z3 × Z2
Z8 × Z4
Z3 × Z23
Z16 × Z2
Z5 × Z4 × Z2
Z4 × Z3 × Z2
Z4 × Z32
Z11 × Z22
Z16 × Z2
Z5 × Z4 × Z2
The cyclic group of order m is denoted by Zm . For all other n ≤ 100 cases the Galois group is the
cyclic group Zδ(n) . The n values are given in A206552.
Only independent cycles are counted, i.e., cycles which appear as sub-cycles of the given ones have been
omitted.
The notation, e.g., 162 81 41 22 , means that there are 2 cycles of order (length) 16, one cycle of order 8,
one cycle of order 4 and two cycles of order 2.
Direct products of identical cyclic groups are sometimes written in exponent form, e.g., Z22 stands for
Z2 × Z2 .
Boxed and colored n-numbers indicate where some non-cyclic Galois group appears for the first time.
For the cycle graphs see Fig. 4.
55
δ=4: Z_2 x Z_2
δ=8: Z_4 x Z_2
δ=12: Z_3 x Z_2 x Z_2
δ=16: Z_4 x Z_4
δ=20: Z_5 x Z_2 x Z_2
δ=16: Z_8 x Z_2
δ=24: Z_4 x Z_3 x Z_2
δ=16: Z_4 x Z_2 x Z_2
δ=18: Z_3 x Z_3 x Z_2
δ=32: Z_16 x Z_2
δ=36: Z_9 x Z_2 x Z_2
Figure 4: Cycle graphs for non-cyclic Galois groups Gal (Q(ρ(n))/Q) appearing for n = 1..100. δ is the
degree of C(n, x), the maximal polynomial of ρ(n), hence the order of the Galois group. See Table 8.
Continued on next page.
56
δ=32: Z_8 x Z_4
δ=24: Z_3 x Z_2 x Z_2 x Z_2
δ=40: Z_5 x Z_4 x Z_2
δ=36: Z_4 x Z_3 x Z_3
Figure 4 continued: Cycle graphs for non-cyclic Galois groups Gal (Q(ρ(n))/Q) appearing for n = 1..100.
See Table 8.
57
| 4math.GR
|
Guided Grammar Convergence
Vadim Zaytsev, [email protected]
arXiv:1503.08476v1 [cs.SE] 29 Mar 2015
SWAT/CWI and IvI/UvA, Amsterdam, The Netherlands
Abstract. Relating formal grammars is a hard problem that balances
between language equivalence (which is known to be undecidable) and
grammar identity (which is trivial). In this paper, we investigate several milestones between those two extremes and propose a methodology
for inconsistency management in grammar engineering. While conventional grammar convergence is a practical approach relying on human
experts to encode differences as transformation steps, guided grammar
convergence is a more narrowly applicable technique that infers such
transformation steps automatically by normalising the grammars and
establishing a structural equivalence relation between them. This allows
us to perform a case study with automatically inferring bidirectional
transformations between 11 grammars (in a broad sense) of the same
artificial functional language: parser specifications with different combinator libraries, definite clause grammars, concrete syntax definitions,
algebraic data types, metamodels, XML schemata, object models.
1
Introduction
Modern grammar theory has shifted its focus from general purpose programming
languages to a broader scope of software languages that comprise programming
languages, domain specific languages, markup languages, API libraries, interaction protocols, etc [12]. Such software languages are specified by grammars in
a broad sense that still rely on the familiar infrastructure of terminals, nonterminals and production rules, but specify general commitment to grammatical
structure found in software systems. In that sense, a type safe program commits
to a particular type system; a program that uses a library, commits to using its
exposed interface; an XML document commits to the structure defined by its
schema — failure to commit in any of these cases would mean errors in interpretation of the language entity. These, and many other, scenarios can be expressed
and resolved in terms of grammar technology, but not all structural commitments profit from grammatical approach (as the most remarkably problematic
ones we can note indentation policies and naming conventions).
One of the problems of multiple implementations of the same language, which
is known for many years, is having an abstract syntax definition and a concrete
syntax definition [25]. Basically, the abstract syntax defines the kind of entities
that inhabit the language and must be handled by semantics specification. A
concrete syntax shows how to write down language entities and how to read
them back. It is not uncommon for a programming language to have several
possible concrete syntaxes: for example, any binary operation may use prefix,
infix or postfix notation, without any changes to the language semantics. Indeed,
we have seen infix dialects of postfix Forth (Forthwrite, InfixForth) and prefix
dialects of infix REBOL (Boron). For software languages, the problem is broader:
we can speak of one intended language specification and a variety of abstract
and concrete syntaxes, data models, class dictionaries, metamodels, ontologies
and similar contracts that conform to it.
Our definition of the intended language relies on bidirectional transformations [1,17,22,28] and in particular on their notation by Meertens [17], which we
redefine here for the sake of completeness and clarity:
Definition 1. For a relation R ⊆ S × T , a semi-maintainer is a function
⊲ : S × T → T , such that ∀x ∈ S, ∀y ∈ T, hx, x ⊲ yi ∈ R, and ∀x ∈ S, ∀y ∈
T, hx, yi ∈ R ⇒ x ⊲ y = y.
The first property is called correctness and ensures that the update caused by
the semi-maintainer restores the relation. The second property is hippocraticness
and states that an update has no effect (“does no harm”), if the original pair is already in the relation [22]. Other properties of bidirectional transformations such
as undoability are often unachievable. A maintainer is a pair of semi-maintainers
⊲ and ⊳. A bidirectional mapping is a relation and its maintainer.
Definition 2. A grammar G conforms to the language intended by the master
grammar M , if there exists a bidirectional mapping between instances of their
G |= L(M ) ⇐⇒ ∃R ⊆ L(G) × L(M )
languages.
∃⊲ : L(G) × L(M ) → L(M )
∃⊳ : L(G) × L(M ) → L(G)
Naturally, for any grammar holds G |= L(G).
For example, consider a concrete syntax Gc of a programming language used
by programmers and an abstract syntax M = Ga used by a software reengineering tool. We would need ⊲ to produce abstract syntax trees from parse trees
and ⊳ to propagate changes done by a reengineering tool, back to parse trees. If
those can be constructed — examples of algorithms have been seen [10,23,25], —
then Gc conforms to the language intended by Ga . As another example, consider
an object model used in a tool that stores its objects in an external database
(XML or relational): the existence of a bidirectional mapping between entries
(trees or tables) in the database and the objects in memory, means that they
represent the same intended language, even though they use very different ways
to describe it and one may be a superlanguage of the other. For a more detailed
formalisation and discussion of bidirectional mappings and grammars, a reader
is redirected elsewhere [29,31].
Roadmap. In the following sections, we will briefly present the following milestones
of relationships between languages:
§2. Grammar identity: structural equality of grammars
§3. Nominal equivalence: name-based equivalence of grammars
§4. Structural equivalence: name-agnostic footprint-matching equivalence
§5. Abstract normalisation: structural equivalence of normalised grammars
Then, §6 summarises the proposed method and discusses its evaluation.
Finally, §7 concludes the paper by establishing context and contributions.
2
Grammar identity
Let us assume that grammars are traditionally defined as quadruples G =
hN , T , P, Si where their elements are respectively the sets of nonterminal symbols, terminal symbols, production rules and starting nonterminal symbols.
Definition 3. Grammars G and G′ are identical, if and only if all their components are identical: G = G′ ⇐⇒ N = N ′ ∧ T = T ′ ∧ P = P ′ ∧ S = S ′ .
The definition is trivial, and in practice is commonly weakened somehow.
For example, many metalanguages allow the right hand sides of rules from P to
contain disjunction (inner choice), which is known to be commutative, so it is
natural to disregard the order of disjunctive clauses when comparing grammars:
gdt, the “grammar diff tool” used in convergence case studies [15,32] implements
that. However, many grammar manipulation technologies such as PEG [8] or
TXL [2], use ordered choices, so this optimisation can be perceived as premature.
For this reason, we will explicitly abandon disjunction in later sections.
3
Nominal equivalence
Since identity can be seen as a trivial bijection, a disciplined weakening of Def. 3
that works across all grammars in a broad sense, is this:
Definition 4. Grammars G and G′ are nominally equivalent, if there is a
bijection β between their production rules:
G ≅ G′ ⇐⇒ N = N ′ ∧ T = T ′ ∧ S = S ′ ∧ ∃β : P → P ′ ,
∀q ∈ P ′ , ∃p ∈ P, q = β(p);
∀p1 , p2 ∈ P, β(p1 ) = β(p2 ) ⇒ p1 = p2
Algorithms that are used to construct β can be different. For example, in
Popart the metalanguage is designed in such a way that it contains enough information to generate both abstract and concrete syntaxes [25]. In TIF-grammars,
a concrete syntax specification is annotated with directions on which nodes need
to be folded/unfolded or removed when constructing an abstract syntax tree
(AST) [10]. In Rascal, the implode function that maps parse trees to ASTs uses
names of nonterminals and subexpressions to direct the automatic construction of β: for example, an optional nonterminal occurrence can be mapped to a
string, a list or a Boolean; a Kleene star can be imploded to either a list or a
set, etc [13]. Similar techniques can be spotted in OOP [4], in MDE [6], in data
binding frameworks [7], etc.
Once the bijective β is agreed on the level of grammars, we need to construct
a coupled maintainer on the language instance level. If it cannot be constructed,
then β is useless for us. However, there are many cases when the maintainer can
be constructed to be bidirectional (R from Def. 2 can be partial and ⊲ and/or
⊳ can be only injective but not surjective). An example of that is matching
different representations of lists/sets: “one or more” and “zero or more” Kleene
repetitions are commonly used in syntactic notations, but one can always be
bidirectionally matched to the other. In our prototype implementation, we disregard unreachable nonterminals, treat built-in and user-defined nonterminals
equally, allow sequence element permutations and desugar metasyntactic constructs like “separator lists”, as well as match different kinds of lists/sets. Such
strategy set was chosen to be valid for nominal matching techniques, but also
useful for the following sections.
4
Structural equivalence
In order to reuse the methods from the nominal equivalence approach in the
case when nonterminal names do not match, we need to construct an additional
mapping between the nonterminals, and use that instead of nominal identity.
We shall refer to this mapping as nominal resolution. To construct it, we generalise permutations and construct signatures that express general structure of
a production rule. Such signatures depend heavily on the expressiveness of the
chosen metalanguage (in particular, when the formal grammar is defined with
just terminals and nonterminals, their use is severely limited), but most common
metalanguages contain enough functionality to allow signatures to work.
Definition 5. A footprint πn (x) of a
nonterminal n in an expression x is defined as a multiset of presence indicadef
tors.
πn (x) =
Definition 6. Two
footprints
are
equivalent, if they are equal modulo
repetition kinds: π ≈ ξ ⇐⇒ π = ξ ∨ π ′ =
ξ ′ , where ζ ′ is ζ with all + elements
replaced by ∗ elements.
{1}
{?}
{+}
{∗}
πn (y)
πn (e1 ) ∪ πn (e2 )
∅
if x = n
if x = n?
if x = n+
if x = n∗
if x = name: y
if x = e1 e2
otherwise
Note how disjunction is missing from the definition of a footprint. We will
see later how it can be removed from any grammar by factoring and folding.
The identity of footprints follows the standard definition of identity of multisets
(which naturally subsumes abstraction over permutations). We also define equivalence of them that generalises the treatment of lists discussed in the previous
section. Footprints together form a signature.
Definition 7. A prodsig, or a signature of a production rule p = (m ::= x) is
defined as a set of tuples with nonterminals used in its right hand side and their
footprints:
σ(m ::= x) = {hn, πn (x)i | n ∈ N , πn (x) 6= ∅}.
For example, the prodsig of a production rule P ::= F + is {hF, {+}i} and
the prodsig of F ::= SS ∗ E is {hE, {1}i, hS, {1, ∗}i}.
Definition 8. Two production rules are prodsig-equivalent, if and only if
there is an equivalent match between tuple ranges of their signatures:
p ≎ q ⇐⇒ ∀hn, πi ∈ σ(p), ∃hm, ξi ∈ σ(q), π ≈ ξ
Consider a simple case of exactly one production rule taken from each of
the grammars: pm from the master grammar and ps from the servant grammar.
Suppose that the left hand sides of them are assumed to match, and we want
to see if the right hand sides are matched nominally as well, and whether they
deliver any new information with respect to nominal resolution. When prodsigs
σ(em ) and σ(es ) are constructed, we have effectively built relations that bind
nonterminal names to their occurrences. By subsequently matching the ranges
of them with either strong or weak prodsig-equivalence, we can infer nominal
matching of the nonterminals by matching the domains of the relations.
Definition 9. For any two prodsig-equivalent production rules p and q, p ≎ q,
there is (at least one) nominal resolution relationship p ⋄ q that satisfies:
∀ha, bi ∈ p ⋄ q :
a = ω ∨ b = ω ∨ ∃π, ∃ξ, π ≈ ξ, ha, πi ∈ σ(p), hb, ξi ∈ σ(q)
∀hc, di ∈ p ⋄ q :
a = c 6= ω ⇒ b = d, where ω denotes unmatched nonterminals.
For two arbitrarily provided grammars (presumably of the same intended
software language, but not necessarily admitting any kind of equivalence), we
cannot claim the existence of only one nominal resolution that works across all
their production rules, but we can attempt to construct a minimum possible one:
Definition 10. Given two grammars G1 and G2 , a nominal resolution between
them is a relation between their nonterminals ♦ ∈ N1 × N2 such that ∀pi ∈ P1 ,
if ∃qj ∈ P2 , pi ≎ qj , then ∃⋄ij ⊂ ♦, such that pi ⋄ij qj .
In our case study, we have used various definitions of the same toy functional
language FL taken from [15]. For instance, some grammars were extracted from
object models by analysing Java code. Since the Java implementation of FL
used List<Expr> to represent lists, the production rules for function declaration
and function call assumed zero or more arguments, while the master grammar
assumed one or more. Hence, production rules F1 ::= S1 S1∗ E1 and E1 ::= S1 E1∗
were matched with their prodsig-equivalent counterparts F2 ::= S2 S2+ E2 and
E2 ::= S2 E2+ . All the various grammars of FL, their prodsigs and nominal matching reports are exposed for inspection in the full report on the case study [27].
5
Abstract normalisation
In order to apply the methodology based on nonterminal footprints, production
signatures and their equivalence relations, we need the input grammars to comply with some assumptions that have been left informal so far. In particular, we
can foresee possible problems with names/labels for production rules and subexpressions, terminal symbols (often not a part of the abstract syntax), disjunction
(inner choices, also non-factored), separator lists and other metasyntactic sugar,
non-connected nonterminal call graph, inconsistent style of production rules, etc.
If by Pn ⊂ P we denote the subset of production rules concerning one particular
nonterminal: Pn = {p ∈ P | p = n ::= α, α ∈ (N ∪ T )∗ }, then we can define the
Abstract Normal Form as follows:
Definition 11. A grammar G = hN , T , P, Si, where T = ∅ and S = {s}, is
said to be in Abstract Normal Form, if and only if:
• N is decomposable to disjoint sets, such that N = N+ ∪ N− ∪ N⊥
• One of them is not empty and includes the root: s ∈ N+ ∪ N−
• Nonterminals from one subset are undefined: n ∈ N⊥ ⇒ Pn = ∅
• Nonterminals from one other subset are defined with exactly one rule:
n ∈ N− ⇒ |Pn | = 1, Pn = {n ::= α}, α ∈ N +
• Nonterminals from the other subset are defined with chain rules:
n ∈ N+ ⇒ ∀p ∈ Pn , p = (n ::= x), x ∈ N
In fact, any grammar can be rewritten to assume this form: in our prototype implementation, this is done by programming a grammar mutation [28,30].
(A grammar mutation is a general intentional grammar change that cannot be
expressed independently of the grammar to which it will be applied. Thus, if
“rename” is a parametric grammar transformation operator, then “rename A
to B” is a transformation, but “rename all nonterminals to uppercase” is a mutation that is equivalent to transformations like “rename a to A” and “rename
b to B” depending on the input grammar). Our prototype, normal::ANF, is a
metaprogram in Rascal [13] that is available for inspection as open source [32]. It
is in fact a superposition of mutations that address the items from the definition
individually: remove labels, desugar separator lists, fold/unfold chain production
rules, etc.
All the rewritings performed by transforming a grammar to its ANF, are assumed to be monadic in the sense of not only normalising the grammar, but also
yielding a bidirectional grammar transformation chain which execution would
normalise the grammar. (In our implementation, these steps are specified in
the ΞBGF language, primarily because no other bidirectional grammar transformation operator suite exists [30]). The bidirectional grammar transformation
chain can then be coupled to the bidirectional mapping between language instances from Def. 2, with the methodology described by [28]. This is required
for traceability: the conversion to ANF is one of the steps to achieve automated
convergence, not a one-way preprocessing.
6
Discussion
To summarise, grammar convergence is a technique of relating different grammars in a broad sense of the same intended software language [15]. It relies on
the transformations being programmed by an experienced grammar engineer:
even beside the required expertise, the process is not incremental — the transformation steps need to be considered carefully and constructed for each new
grammar added to the mix. With the definitions from previous sections, we have
described the process of guided grammar convergence, where the master grammar of the intended language is constructed once, and the transformations are
inferred for any directly available grammars as well for the ones possibly added
in the future. The process works as follows.
•
•
•
•
Extract pure grammatical knowledge from the grammar source.
Use grammar mutations to preprocess your grammars, if necessary.
Normalise the grammar by removing all problematic/ambiguous constructs.
Start by matching the roots of the connected normalised grammar.
• Match multiple production rules by prodsig-equivalence; infer new nominal
matches by matching equivalent prodsigs. Repeat for all nonterminals.
• If several matches are possible, explore all and fallback in case of failure. If
global nominal resolution scheme was impossible to infer, fail.
• Resolve structural differences in the production rules that matched nominally.
To evaluate the method of guided grammar convergence, we have applied
it to a case study of 11 different grammars of the same intended functional
language that was defined and used earlier in order to demonstrate the original
grammar convergence method that converged 5 of these grammars. The following grammar sources were used (all of them are available in the repository of
SLPS [32] together with their evolution history and authorship attribution):
adt: an algebraic data type1 in Rascal [13];
antlr: a parser description in the input language of ANTLR [19], with semantic
actions (in Java) intertwined with EBNF-like productions;
dcg: a logic program written in the style of definite clause grammars [20];
emf: an Ecore model, automatically generated by Eclipse [5] from the XML
Schema of the xsd source;
jaxb: an object model obtained by a data binding framework, generated automatically by JAXB [7] from the XML schema for FL;
om: a hand-crafted object model (Java classes) for the abstract syntax of FL;
python: a parser specification in a scripting language, using the PyParsing
library [16];
rascal: a concrete syntax specification in the metaprogramming language of
Rascal language workbench [13];
sdf: a concrete syntax definition in the notation of SDF [11] with scannerless
generalized LR parsing as parsing model.
txl: a concrete syntax definition in the notation of TXL (Turing eXtender Language) transformational framework [2], which, unlike SDF, uses a combination of pattern matching and term rewriting).
xsd: an XML schema [9] for the abstract syntax of FL.
The complete case study is too big to be presented here, interested readers
are redirected to a 40+ page long report [27], containing all production rules,
signatures, matchings and transformations. The case study was successful: on
average ANF was achieved after 20–30 transformation steps, nominal resolution
took up to 9 (proportional to the number of nonterminals) and structural resolution needed 0–5 more steps, after which all 11 grammars were converged.
ANTLR, DCG and PyParsing used layered definitions and therefore were the
only three grammars to require mutation (another 2–6 steps). The case study is
available for investigation and replication both in the form of Rascal metaprograms at http://github.com/grammarware/slps [32] (the main algorithm is
located in the converge::Guided module which can be observed and modified
1
http://tutor.rascal-mpl.org/Rascal/Declarations/AlgebraicDataType/AlgebraicDataType.html.
at shared/rascal/src/converge/Guided.rsc) and as a PDF report with all
grammars and transformations pretty-printed automatically [27].
Guided grammar convergence is a methodology stemming from the grammarware “technological space” [14]. When looking for similar techniques in other
spaces (engaging in “space travel”), the obvious candidates are schema matching
and data integration in the field of data modeling and databases [21]; comparison
of UML models or metamodels in the context of model-driven engineering [6,26];
model weaving for product line software development [3,24]; computation of
refactorings from different OO program versions [4,18]; etc. For example, Del
Fabro and Valduriez [3] utilised metamodel properties for automatically producing weaving models. The core difference is that (meta)model weaving ultimately
aims at incorporating all the changes into the resulting (meta)model, while
guided grammar convergence also makes complete sense when some changes in
the details are disregarded. The lowest limit in this process is needed (otherwise
additional claims on the minimality of inferred transformations are required),
and we specify this lowest limit as the master grammar. Another difference is
that model weaving rarely involves a number of models bigger than two, and even
our little case study of guided convergence had 10+ grammars in it. In general,
prodsig-based matching is more lightweight than those methods, since it in fact
compares straightforwardly structured prodsigs, thus easily wins in performance
and implementability but loses in applicability to complex scenarios.
7
Conclusion
We knew that language equivalence is undecidable and that grammar identity is
trivial. In this paper, we have attempted to reach a useful level of reasoning about
language relationships by departing from grammar identity as the “easy” side of
the spectrum. This was done in the scope of grammar convergence, when several
implementations of the same software language are inspected for compatibility.
A definition of an intended language was provided (Def. 2) based on a bidirectional transformation between language entities. Then we revisited existing
and possible techniques of structural matching that assumed nominal identity
(Def. 3). In order to automatically infer nominal matching, we introduced nonterminal footprints (Def. 5), production signatures (Def. 7) and various degrees
of equivalence among them. An extensive normalisation scheme (Def. 11) was
proposed to transform any given grammar into the form most suitable for nominal and then structural matching. It has been explained that when such normalisation is not enough, a more targeted yet still automated approach is needed
with grammar mutation strategies making the method robust with respect to
different grammar design decisions, such as the use of layers instead of priorities
or recursion instead of iteration. Just as all other parts of the proposed process,
these mutations operate automatically and do not require human intervention.
A case study was used to evaluate the proposed method of guided grammar convergence. The experiment concerned several implementations of a simple
functional language in ANTLR, DCG, Ecore, Java, Python, Rascal, SDF, TXL,
XML Schema. The diversity in language processing frameworks — metaprogramming languages, declarative specifications, syntax definitions, algebraic data
types, parsing libraries, transformation frameworks, software models, parser definitions — was intentional and aimed at stressing the definition of the intended
language and the guided convergence method. Casting all grammars from our
case study to ANF allowed us to make inference quicker and with less obstacles,
as well as to explain the process more clearly.
All artifacts discussed on the pages of this paper, are transparently available
to the public through a GitHub repository [32]. For each of the sources of the case
study, one could inspect the original file, the extracted grammar, the extractor
itself, the mutations that have been derived and applied, the normalisations
to ANF, the normalised grammar, the nominal resolution and reasons for each
match, as well as the structural resolution steps. One could also investigate the
implementation of the method of guided grammar convergence, the algorithm
for calculating prodsigs and the process of convergence. Supplementary material
contains 40+ pages of the full report, also generated by our prototype [27].
On the practical side, guided grammar convergence provides a balanced
method of grammar manipulation, positioned right between unstructured inline
editing (which makes grammar development very much like software development but lacks important properties such as traceability and reproducibility)
and strictly exogenous functional transformation (which requires substantially
more effort but is robust, repeatable and exposes the intended semantics). Its
future role can be seen as a support for grammar product lines that allows
both steady adaptation plans for deriving secondary artifacts from the reference
grammar, and occasional inline editing of the derived artifacts with subsequent
automated restoration of the adaptation scripts. This is a contribution to the
field of engineering discipline for grammarware [12].
References
1. K. Czarnecki, J. Foster, Z. Hu, R. Lämmel, A. Schürr, and J. Terwilliger. Bidirectional Transformations: A Cross-Discipline Perspective. In TPMT, volume 5563 of
LNCS, pages 260–283. Springer, 2009.
2. T. R. Dean, J. R. Cordy, A. J. Malton, and K. A. Schneider. Grammar Programming in TXL. In SCAM. IEEE, 2002.
3. M. D. Del Fabro and P. Valduriez. Semi-Automatic Model Integration Using
Matching Transformations and Weaving Models. In SAC, pages 963–970, 2007.
4. D. Dig, C. Comertoglu, D. Marinov, and R. Johnson. Automated Detection of
Refactorings in Evolving Components. In ECOOP, volume 4067 of LNCS, pages
404–428. Springer, 2006.
5. Eclipse.
Eclipse Modeling Framework Project (EMF 2.4), 2008.
http://www.eclipse.org/modeling/emf.
6. J.-R. Falleri, M. Huchard, M. Lafourcade, and C. Nebut. Metamodel Matching
for Automatic Model Transformation Generation. In MoDELS, volume 5301 of
LNCS, pages 326–340. Springer, 2008.
7. J. Fialli and S. Vajjhala. Java Specification Request 31: XML Data Binding Specification, 1999.
8. B. Ford. Parsing Expression Grammars: a Recognition-Based Syntactic Foundation. In POPL, January 2004.
9. S. Gao, C. M. Sperberg-McQueen, and H. S. Thompson. W3C XML Schema
Definition Language (XSD) 1.1. W3C Recommendation, Apr. 2012.
10. A. Johnstone and E. Scott. Tear-Insert-Fold grammars. In LDTA, pages 6:1–6:8.
ACM, 2010.
11. P. Klint. Meta-Environment for Generating Program- ming Environments.
TOSEM, 2(2):176–201, 1993.
12. P. Klint, R. Lämmel, and C. Verhoef. Toward an Engineering Discipline for Grammarware. TOSEM, 14(3):331–380, 2005.
13. P. Klint, T. van der Storm, and J. Vinju. EASY Meta-programming with Rascal.
In GTTSE, volume 6491 of LNCS, pages 222–289. Springer, January 2011.
14. I. Kurtev, J. Bézivin, and M. Akşit. Technological Spaces: an Initial Appraisal. In
CoopIS, DOA’2002, Industrial track, 2002.
15. R. Lämmel and V. Zaytsev. An Introduction to Grammar Convergence. In iFM,
volume 5423 of LNCS, pages 246–260. Springer, February 2009.
16. P. McGuire. Getting Started with Pyparsing. O’Reilly, first edition, 2007.
17. L. Meertens. Designing Constraint Maintainers for User Interaction. Manuscript,
June 1998.
18. M. O’Keeffe and M. O. Cinnéide. Search-Based Refactoring: an Empirical Study.
JSME, 20(5):345–364, 2008.
19. T. Parr. ANTLR—ANother Tool for Language Recognition, 2008.
20. F. Pereira and D. Warren. Definite Clause Grammars for Language Analysis. In
Readings in Natural Language Processing, pages 101–124. MKP, 1986.
21. E. Rahm and P. A. Bernstein. A Survey of Approaches to Automatic Schema
Matching. The VLDB Journal, 10:334–350, December 2001.
22. P. Stevens. Bidirectional Model Transformations in QVT: Semantic Issues and
Open Questions. In MODELS, volume 4735 of LNCS, pages 1–15. Springer, 2007.
23. T. v. d. Storm, W. R. Cook, and A. Loh. Object Grammars: Compositional &
Bidirectional Mapping Between Text and Graphs. In SLE, volume 7745 of LNCS,
pages 4–23. Springer, July 2013.
24. M. Völter and I. Groher. Product line implementation using aspect-oriented and
model-driven software development. In SPLC, pages 233–242. IEEE, 2007.
25. D. S. Wile. Abstract Syntax from Concrete Syntax. In ICSE, pages 472–480, 1997.
26. Z. Xing and E. Stroulia. Refactoring Detection based on UMLDiff Change-Facts
Queries. In WCRE, pages 263–274. IEEE, 2006.
27. V. Zaytsev. Guided Grammar Convergence. Full Case Study Report. Generated
by converge::Guided. ACM CoRR, 6541:1–44, July 2012.
28. V. Zaytsev. Language Evolution, Metasyntactically. EC-EASST, BX, 49, 2012.
29. V. Zaytsev. Case Studies in Bidirectionalisation. In Pre-proceedings of TFP’14,
pages 51–58, May 2014. Extended Abstract.
30. V. Zaytsev. Software Language Engineering by Intentional Rewriting. EC-EASST,
SQM, 65, Mar. 2014.
31. V. Zaytsev and A. H. Bagge. Parsing in a Broad Sense. In MoDELS, volume 8767
of LNCS, pages 50–67. Springer, Oct. 2014.
32. V. Zaytsev, R. Lämmel, T. van der Storm, L. Renggli, R. Hahn, and G. Wachsmuth.
Software Language Processing Suite2 , 2008–2013. http://slps.github.io.
2
The authors are given according to the list of
http://github.com/grammarware/slps/graphs/contributors.
contributors
at
| 6cs.PL
|
1
Coexistence of RF-powered IoT and a Primary
Wireless Network with Secrecy Guard Zones
arXiv:1705.07237v1 [cs.IT] 20 May 2017
Mustafa A. Kishk and Harpreet S. Dhillon
Abstract
This paper studies the secrecy performance of a wireless network (primary network) overlaid with
an ambient RF energy harvesting Internet of Things (IoT) network (secondary network). The nodes
in the secondary network are assumed to be solely powered by ambient RF energy harvested from
the transmissions of the primary network. We assume that the secondary nodes can eavesdrop on
the primary transmissions due to which the primary network uses secrecy guard zones. The primary
transmitter goes silent if any secondary receiver is detected within its guard zone. Using tools from
stochastic geometry, we derive the probability of successful connection of the primary network as
well as the probability of secure communication. Two conditions must be jointly satisfied in order
to ensure successful connection: (i) the signal-to-interference-plus-noise ratio (SINR) at the primary
receiver is above a predefined threshold, and (ii) the primary transmitter is not silent. In order to ensure
secure communication, the SINR value at each of the secondary nodes should be less than a predefined
threshold. Clearly, when more secondary nodes are deployed, more primary transmitters will remain
silent for a given guard zone radius, thus impacting the amount of energy harvested by the secondary
network. Our results concretely show the existence of an optimal deployment density for the secondary
network that maximizes the density of nodes that can harvest an amount of energy above a predefined
threshold. Furthermore, we show the dependence of this optimal deployment density on the guard zone
radius of the primary network. In addition, we show that the optimal guard zone radius selected by
the primary network is a function of the deployment density of the secondary network. The interesting
coupling between the performance of the two networks is studied using tools from game theory. In
addition, we propose an algorithm that can assist the two networks to converge to Nash equilibrium.
The convergence of this algorithm is verified using simulations. Overall, this work is one of the few
concrete works that symbiotically merge tools from stochastic geometry and game theory.
M. A. Kishk and H. S. Dhillon are with Wireless@VT, Department of ECE, Virginia Tech, Blacksburg, VA. Email:
{mkishk, hdhillon}@vt.edu. The support of the U.S. NSF (Grants CCF-1464293, CNS-1617896, and IIS-1633363) is gratefully
acknowledged. This paper was presented in part at the IEEE WCNC 2017, San Francisco, CA [1]. Last updated: May 23, 2017.
2
Index Terms
Stochastic geometry, wireless power transfer, physical layer security, game theory, coverage probability, Poisson Point Process, Poisson hole process.
I. I NTRODUCTION
Owing to rapid technological advances, wireless communication networks are undergoing
unprecedented paradigm shifts. One of the most interesting amongst them is the attempt to make
wireless networks virtually self-perpetual in terms of their energy requirements. This is especially
gaining importance in internet-of-things (IoT) realm where it may not be economical to charge
or replace batteries periodically in billions of devices worldwide [2]. In order to achieve this
vision of self-perpetual operation, it is necessary to provide energy harvesting capability to such
networks in addition to reducing energy expenditures through energy-efficient communication
policies [3], [4] and energy efficient hardware [5]. While, in principle, we can use any available
source of energy to power these networks, ambient RF energy harvesting is considered a preferred
option due to its ubiquity [6]. Further, due to the need of deploying these ambient RF energy
harvesting nodes/devices (referred to as energy receivers or ERs in the rest of this paper) close
to the sources of RF signals, concerns on the secrecy of these RF signals were recently raised
in the literature [7]–[13]. While the RF signal source (referred to as primary transmitter or
PT in this paper) transmits confidential messages to legitimate information receivers (IRs), the
existence of ERs close to the PT may enable them to decode these messages. Consequently,
careful study of physical layer security in such systems is required to glean insights on the
new secrecy performance limitations in the presence of ERs. The importance of physical layer
security lies in ensuring the ability of legitimate receivers to decode the confidential messages
while preventing illegitimate receivers from decoding these messages [14]. Many solutions were
proposed to enhance physical layer security including: (i) using protected zones in order to ensure
an eavesdropper-free regions around the transmitters [15], [16], (ii) using artificial noise in order
to degrade the confidential signal’s SNR at the energy receiver [9], [17], (iii) beamforming in
multi-antenna systems [10]–[12], and (iv) using guard zones to stop information transmission
whenever an eavesdropper is detected within a specific region around the transmitter [18], [19].
In this paper, we limit our attention to single-antenna transmitters and receivers, which means
beamforming is not applicable to our setup. In addition, using protected zones assumes some
physical control over the eavesdroppers, which will not be considered in this paper. Thus, we have
3
two options to choose from: either use artificial noise or guard zones. In [20], it was recently
shown that the guard zone technique outperforms the artificial noise technique in the noise
limited regime when the link distance is higher than a specific threshold. No such performance
comparison between the two techniques is known when interference is taken into account.
Regardless, we focus on the guard zone technique in this paper while leaving the artificial
noise technique as a promising direction of future work. In particular, modeling the interaction
between a primary network using guard zones to enhance secrecy and an IoT network using RF
signals transmitted by the primary network to harvest energy is the main focus of this paper.
A. Related Work
In this subsection, we will provide a brief summary of the related works in three general
directions of interest to this paper: (i) stochastic geometry-based analysis of secrecy and/or
energy harvesting wireless networks, (ii) game theory-based analysis of secrecy and/or energy
harvesting wireless networks, and (iii) analysis of secure wireless power transmission.
Stochastic geometry-based work. Stochastic geometry has emerged as a powerful mathematical
tool for the modeling and analysis of variety of wireless networks [21]–[24]. Out of numerous
aspects of wireless networks that have been studied using stochastic geometry, two that are
most relevant to this paper are: (i) secrecy [18], [25], [26], and (ii) energy harvesting [27]–
[31]. We first discuss the related works on secrecy. The work presented in [18] quantified the
loss in network throughput that results from ensuring a specific level of secrecy performance.
In addition, possible performance enhancement using guard zones was also studied. Authors
in [25] studied the physical layer security of downlink cellular networks assuming that all the
existing users in the network are potential cooperating eavesdroppers. In [26], authors studied
the secrecy of downlink transmission when the transmitter adopts transmit antenna selection.
In particular, they derived the secrecy outage probability for the cases of independent and
cooperating eavesdroppers, considering both half and full duplex legitimate receivers.
Stochastic geometry has also been applied to analyze the performance of energy harvesting cellular networks with emphasis on either the downlink channel [27]–[29] or the uplink
channel [30]. The work presented in [27] provided a comprehensive framework for analyzing
heterogeneous cellular networks with energy harvesting base stations (BSs). The primary focus
was on characterizing optimal transmission policies as well as quantifying the availability of
different tiers of BSs. In [28], [29], the joint analysis of the downlink signal quality and the
4
amount of energy harvested at an RF-powered user was performed. In [30], [31], the uplink
counterpart of this problem was explored in which the RF-powered node first harvests energy
from ambient RF signals and then uses it to perform data transmission. The work presented
in [32] focused on the secrecy analysis of wireless networks where the legitimate transmitters
are powered by energy harvesting. To the best of our knowledge, our work provides the first
stochastic geometry-based secrecy analysis of wireless power transmission with energy receivers
acting as potential eavesdroppers. More details will be provided shortly.
Game theory-based work. Tools from game theory have been widely used in the analysis
of secrecy in wireless networks [33]–[36]. Using these tools, in [33] authors modeled the
interaction between cognitive networks and eavesdroppers, where a channel selection algorithm
was proposed to reach the Nash equilibrium. In [34], authors studied the problem of optimizing
the uplink path in multi-hop networks in order to maximize the secrecy rate. In [35], authors
used matching theory to develop an algorithm that enhances the secrecy of source-destination
pairs using jamming nodes. In [36], authors proposed a distributed algorithm that enhances the
achievable secrecy rates in wireless networks with cooperative wireless nodes.
These tools have also been used in the analysis of energy harvesting wireless networks. For
instance, authors in [37] provided different approaches to manage energy trading among energy
harvesting small cell BSs in order to minimize the consumption of non-renewable energy. Authors
in [38] used theses tools to model the relay interference channels where the relay divides the
received power from the source into two parts: (i) the first part is used to charge its own battery,
and (ii) the rest of the received power is used to forward the received packet to its destination.
In [39], these tools were used to determine optimal probability of switching from listen to active
modes and from sleep to active modes for a solar powered wireless sensor network.
Secure wireless power transmission. While the idea of having the energy receiver as a potential
eavesdropper has not been studied yet in the stochastic geometry literature (with randomly located
ERs and legitimate transmitters), recent works have explored this idea for the deterministic
system setups [7], [8], [40]–[42]. These works assume that the transmitter aims to maximize
secrecy performance with the constraint of providing ERs with the required wireless power.
In [7], [8], authors focused on a single point-to-point link with one ER (potential eavesdropper)
in the system. An artificial noise-based solution was proposed to improve secrecy without
reducing the amount of wireless power received by the ER. In [40], authors studied the use of a
friendly jammer to enhance the performance of the point-to-point system. The friendly jammer
5
increases the amount of received power by the ER. In addition, it reduces the decodability of the
confidential message by the ER. The single link system was extended in [41] to consider one
multi-antenna transmitter, one legitimate receiver, and K ERs. Optimal beamforming schemes
were provided to either maximize secrecy subject to constraints on harvested energy by the ERs
or to maximize harvested energy by the ERs subject to secrecy constraints. Similar approach
was adopted in [42] for a more general system model of one macro BS serving M users, N
femto BSs each serving K users, and L ERs. Unlike all these works, we will model the locations
of ERs and legitimate transmitters using point processes, which will enable us to draw general
conclusions that will not be restricted to particular topologies.
In this paper, we study the secrecy performance of a primary network that consists of PTs and
information receivers (which will also be referred to as primary receivers or PRs), overlaid with
an IoT network that consists of RF-powered devices (energy receivers or ERs). Guard zones are
assumed to be present around PTs in order to improve secrecy. The only sources of ambient RF
signals for ERs are the PT transmissions. The PT transmits information (becomes active) to the
PRs only when the guard zone is free of ERs, otherwise it remains silent. On one hand, the IoT
network would prefer a dense deployment of ERs so as to increase the overall energy harvested
by the IoT network. However, on the other hand, more dense deployment of ERs will mean that
more PTs will stay silent (due to the higher likelihood of ERs lying in the guard zones of the
PTs) that will ultimately reduce the amount of ambient RF energy harvested, and hence degrade
energy harvesting performance. Modeling and analysis of this setup is the main focus of this
paper. We summarize the contributions of this paper next.
B. Contributions
Primary network performance analysis. Modeling the locations of PTs and ERs by two
independent Poisson point processes (PPPs), we show that the locations of active PTs (PTs
with ER-free guard zones) follow Poisson hole process. For this setup, we define the probability
of successful connection (Pcon ) between the PT and its associated PR by the joint probability
of the PT being active and the SINR at the PR being above a predefined threshold. We derive
Pcon as a function of the density of ERs and the guard zone radius rg . We concretely derive
a threshold on the density of the PTs below which Pcon is a decreasing function of rg . Above
this threshold, we prove the existence of an optimal value of rg that maximizes Pcon . For the
secrecy analysis, we derive the probability of secure communication (defined by the probability
6
of having the SINR value at any ER less than a predefined threshold). Referring to this metric
as Psec , we provide several useful insights on the effect of the PT transmission power as well
as the density of ERs on the value of Psec .
IoT network performance analysis. For the IoT network (secondary network), we derive the
probability of harvesting a minimum amount of energy Emin by the ERs. We define the density
of ERs that satisfy this condition as the density of successfully powered ERs. We prove the
existence of an optimal deployment density of ERs that maximizes this density of successfully
powered ERs. In order to capture the relation between this optimal density and the guard zone
radius, we derive a useful lower bound on the the density of successfully powered ERs. We show
that the optimal deployment density that maximizes this lower bound is a decreasing function
of rg . Although this conclusion is drawn using a lower bound, we use numerical results and
simulations to demonstrate that it holds for the exact expressions as well.
Modeling the interaction between the two networks. Building on the above results, we show
that the interaction between the two networks can be modeled using tools from game theory.
In particular, we show that this system can be modeled by a two player non-cooperative static
game. The first player is the primary network with the guard zone radius representing its strategy.
Its utility function is modeled to capture the successful connection probability as well as the
probability of secure communication. The second player is the IoT network with the deployment
density representing its strategy. Its utility function is modeled to capture the main performance
metric of this network, which is the density of successfully powered ERs. We propose a best
response-based learning algorithm that helps in achieving the Nash equilibrium of this game.
II. S YSTEM M ODEL
We consider a system that is composed of two wireless networks: (i) a primary network,
and (ii) an IoT network (will be referred to as secondary network in the rest of this paper).
The primary network is constructed of PTs and primary receivers (will be referred to as either
PRs or legitimate receivers interchangeably throughout the paper). The locations of the PTs are
modeled by a homogeneous PPP ΦP ≡ {xi } ⊂ R2 with density λP . In order to enhance secrecy
performance, each PT is surrounded by a circular guard zone of radius rg centered at the PT. We
assume that each PT is able to detect the presence of any illegitimate receiver within its guard
zone [43]. Various detecting devices can be used for this purpose including metal detectors and
leaked local oscillator power detectors [15]. The benefits of using secrecy guard zones and their
7
Guard
Zone
Active
PT
ER
PR
Silent
PT
Information
Transmission
Guard
Zone
Wireless
Charging
Fig. 1. The PT stops transmission (becomes silent) if any ER is detected within its guard zone.
effect on secrecy performance were discussed in [18]. As shown in Fig. 1, before the PT transmits
data to its associated PR, it scans the guard zone for any illegitimate receivers. If the guard zone
is clear, the PT transmits the confidential data, otherwise, the PT stops transmission (becomes
silent). The secondary network is constructed of RF-powered nodes that harvest RF energy from
the signals transmitted by the primary network. We refer to these nodes as energy receivers
(will be referred to as either ERs or eavesdroppers interchangeably throughout the paper). The
locations of the ERs are modeled by a homogeneous PPP ΦS ≡ {yi } ⊂ R2 with density λS .
From the primary network’s perspective, the ERs are considered illegitimate receivers. Hence,
applying the guard zone scheme, the PT stops transmission whenever at least one ER exists
within its guard zone. We assume that ERs are the only potential eavesdroppers in the system.
We focus our analysis on the typical PT whose intended PR is located at a given distance r1
from the PT. Drawing analogy from the Poisson bipolar model, we call this the typical link and
its constituent PT and PR as typical PT and typical PR, respectively. Note that this terminology
is indeed rigorous if we assume a Poisson bipolar model in which each PT has an associated
PR at a fixed distance. However, since the same setup can be used for cellular networks by
treating r1 as the conditional value of the serving distance, we leave the setup general. Overall,
the assumption of fixed r1 enables us to gain several useful insights. In particular, this enables
us to better understand how the rest of the system parameters (e.g. rg , λS , and PT’s transmission
power) affect the interaction between the two networks. Due to the stationarity of PPP, the
typical PR can be assumed to be located at the origin without loss of generality. The typical PT
is located at x1 at distance r1 = kx1 k from the typical PR. In case the guard zone is clear of ERs,
the typical PT transmits the confidential message to the typical PT. In that case, the received
power at the typical PR is Pt h1 r1−α , where Pt is the PT transmission power, h1 ∼ exp(1) models
8
TABLE I
TABLE OF N OTATIONS
Notation
Description
ΦP ; λP
PPP modeling the locations of the PTs; density of the PTs
ΦS ; λS
PPP modeling the locations of ERs; density of the ERs
EH
Amount of energy harvested by ER (in Joules)
Pcon
The probability of successful connection between a PT and its associated PR
Psec
The probability of secure communication
Penergy
The density of successfully powered ERs
h; g; w
Fading gains for the link between the typical PR and an arbitrary PT; the typical PT and an arbitrary ER; an ER and its nearest active PT
rg ; R e
Guard zone radius; distance between a PT and its nearest ER
1(Ξ)
An indicator function such that 1(Ξ) = 1 if the condition Ξ is satisfied and equals zero otherwise
Pt ; σP2 ; σS2
PT transmission power; noise power at the PR; noise power at the ER
; α
Minimum required value for Psec ; path loss exponent (α > 2)
βP ; βS
Minimum required SINR value at the PR to ensure successful connection; threshold on SINR at ERs to ensure perfect secrecy
Rayleigh fading gain, and r1−α models power law path-loss with exponent α > 2.
A. Primary Network Modeling
According to Wyner’s encoding scheme [14], [44] and the approach used in [18], the PT
defines the rate of codewords and the rate of confidential messages, Ct and Cs respectively.
The difference, Ce = Ct − Cs , can be interpreted as the cost paid to secure the confidential
messages. Let the mutual information between PT’s channel input and PR’s channel output be
It and between PT’s channel input and ER’s channel output be Ie . Then, the objective is to
ensure that the following two conditions are satisfied: (i) Ct ≤ It to ensure successful decoding
at the PR, and (ii) Ce ≥ Ie to ensure perfect secrecy. Equivalently, we can define two SINR
thresholds at the legitimate receiver and at the eavesdropper. To ensure successful decoding of
the received confidential message at the PR, the condition SINRP ≥ βP should be satisfied,
where βP = 2Ct − 1. Similarly, to ensure perfect secrecy, the SINR at any eavesdropper should
satisfy the condition SINRS ≥ βS , where βS = 2Ct −Cs − 1. In order to mathematically define the
instantaneous value of SINR at the legitimate receiver, we need to capture the effect of adopting
the guard zone scheme on the interference level at the typical PR. According to the guard zone
scheme, an interfering PT located at xi is active only if its guard zone is clear of ERs. Hence,
denoting the random variable representing the distance between the typical PT and its nearest
9
After each PT scans its guard zone for ERs:
Locations of retained (active) PTs are modeled by PHP
Before each PT scans its guard zone for ERs:
Locations of PTs are modeled by PPP
PTs inside
holes are
removed
(turn silent)
Arbitrary PT
ER
Active PT
Hole with
radius 𝑟𝑔
Silent PT
Fig. 2. The locations of active PTs is modeled by a PHP where the locations of ERs represent the centers of the holes.
eavesdropper by Re , the probability of a typical PT being active is
2
Pactive = P(Re ≥ rg ) = e−λS πrg .
(1)
Next, we formally define the value of SINR at the typical PR as
Pt h1 r1−α
,
−α + σ 2
P
xi ∈ΦP \x1 δi Pt hi kxi k
SINRP = P
(2)
where hi ∼ exp(1) models Rayleigh fading gain for the link between the typical PR and the PT
located at xi , and σP2 is the noise power at the PRs. The indicator function δi is used to capture
only the active interfering nodes. Hence, δi = 1 if the PT located at xi is active and equals zero
otherwise. The expected value of δi is E[δi ] = Pacitve .
The locations of the active PTs can be modeled using Poisson hole process (PHP) Ψ [45],
[46]. A PHP is constructed using two independent PPPs: (i) Baseline PPP Φ1 , and (ii) the PPP
modeling the locations of the hole centers Φ2 . All points of Φ1 that are within distance D from
any point in Φ2 are carved out. The remaining points of Φ1 represent the PHP. In our system, as
shown in Fig. 2, we can use the same approach to model the locations of the active PTs. Any
PT in ΦP that is within a distance rg from any ER in ΦS is inactive. Hence, the locations of
the remaining (active) PTs are modeled by PHP. This can be formally defined as follows
(
)
[
Ψ = x ∈ ΦP : x ∈
/
B(y, rg ) ,
y∈ΦS
(3)
10
where B(y, rg ) is a ball of radius rg centered at y. Since the probability of retention of any point
in ΦP is Pactive , the density of the resulting PHP Ψ is λP Pacitve .
The primary network tries to jointly optimize two main performance metrics: (i) successful
connection probability, and (ii) secure communication probability. While the former is related
to the successful delivery of the data from the PT to the PR, the latter is related to the security
of the transmitted data when the PT is active. Both metrics are formally defined below.
Definition 1 (Probability of successful connection). In order to ensure successful connection
between the typical PT and PR, two conditions need to be satisfied: (i) the typical PT is active,
and (ii) the SINR at the typical PR is greater than the threshold βP . Therefore, the probability
of successful connection is
Pcon (rg , λS ) = P(Re ≥ rg , SINRP ≥ βP ).
(4)
When a PT is transmitting confidential data (active PT), the condition of SINRS ≤ βS needs
to be satisfied at each ER in order to ensure perfect secrecy. Focusing on the signal transmitted
by the typical PT, the SINR value at an ER located at yj is
Pt g1,j kx1 − yj k−α
,
−α + σ 2
S
xi ∈ΦP \x1 δi Pt gi,j kxi − yj k
SINRS (yj ) = P
(5)
where gi,j ∼ exp(1) models Rayleigh fading gain for the link between the PT located at xi
and the ER located at yj , and σS2 is the noise power at the ERs. We now define the second
performance metric for the primary network, the secure communication probability, next.
Definition 2 (Secure communication probability). Given that a PT is active, the probability that
its transmitted data is perfectly secured is
Psec (rg , λS ) = E 1
[
SINRS (yj ) ≤ βS Re ≥ rg .
(6)
yj ∈ΦS
For a given value of λS , the primary network selects the value of rg = rg∗ that maximizes the
value of Pcon while ensuring Psec ≥ , where 0 < ≤ 1. The selection of rg∗ is mathematically
formulated next.
Definition 3 (Guard zone radius selection). The primary network selects the guard zone radius
rg∗ that satisfies the following
rg∗ = arg max Pcon (rg , λS ),
rg ∈G(λS )
11
G(λS ) = {rg : Psec (rg , λS ) ≥ }.
(7)
Remark 1. By observing Eq. 7, we note that the value of rg∗ selected by the primary network
is a function of the density of the secondary network λS . In addition, while Pactive is obviously
a decreasing function of rg , the exact effect of rg on either Pcon or Psec is harder to describe
than one expects. It might seem intuitive to expect that as rg increases the average distance
between any PT and its nearest ER will increase, leading to an increase in the probability of
secure communication. However, this is actually not always correct. The reason for that is the
effect of rg on the density of the PHP Ψ, λP Pacitve , which is the density of active interferers.
As the guard zone radius increases, the density of interferers decreases, which means that the
interference level at any ER will decrease on average leading to a larger SINR value.
To better understand the behavior of all the aforementioned performance metrics, we will
derive some insightful expressions that will, with the help of the numerical results, provide a
complete picture for the effect of the system parameters on these metrics.
B. Secondary Network Modeling
For an ambient RF energy harvesting device, the distance to the nearest source is critical and
has major effect on the average harvested energy value, as shown in [29]. The energy harvested
from the nearest source is frequently used in the literature to study the performance of ambient
RF energy harvesting wireless devices. For instance, in [47], authors proposed the concept of
harvesting zone where an ER is able to activate its power conversion circuit and harvest energy
only if it is within a specific distance from the active PT. Consequently, we focus our analysis of
the energy harvested from the nearest PT since it is the dominant source of ambient RF energy.
Hence, the energy harvested by the ER located at yj is
EH = ηT Pt kxj,1 − yj k−α wj ,
(8)
where η models the RF-DC efficiency of the ER, T is the duration of the transmission slot
(assumed to be unity in the rest of the paper), xj,1 is the location of the nearest active PT to
the ER located at yj , and wj ∼ exp(1) models Rayleigh fading gain for the link between the
ER located at yj and its nearest active PT. Most of the relevant existing works use one of two
performance metrics for the anlysis of energy harvesting wireless networks: (i) the expected
value of the harvested energy E[EH ] as in [31], or (ii) the energy coverage probability P(EH ≥
12
Emin ) as in [30], where Emin is the minimum required value of EH at each ER. We will use a
modified version of the latter metric. In order to maximize the density of ERs that harvest the
minimum amount of energy Emin , the secondary network selects the network density λS = λ∗S
that maximizes
Penergy = λS P(EH ≥ Emin ).
(9)
The above metric represents the density of the successfully powered ERs.
Remark 2. Note that the distance between an ER and its nearest active PT increases, on
average, as the density of active PTs, λP Pacitve , decreases. Recalling, from Eq. 1, that Pactive
is a decreasing function of both rg and λS , we can expect P(EH ≥ Emin ) to be a decreasing
function of both rg and λS . This implies two things: (i) the performance metric defined in Eq. 9
has a local maximum at λS = λ∗S , and (ii) the value of the selected density λ∗S that maximizes
Eq. 9 is a function of rg .
From Remarks 1 and 2 we can get some initial observations on the coupling between the
two networks. To better understand the relation between the parameters selected by each of the
networks (rg for the primary network, and λS for the secondary network), we first need to derive
an expression for each of Pcon and Psec for the primary network, and Penergy for the secondary
network.
III. P ERFORMANCE A NALYSIS
A. Primary Network
1) Probability of Successful Connection: As stated earlier, for a given value of λS , the primary
network selects the optimal guard zone radius rg∗ that maximizes Pcon while ensuring that Psec ≥
. The process of rg selection was mathematically formulated in Definition 3. In the following
Theorem we derive Pcon .
Theorem 1 (Probability of successful connection). For a given value of λS , the probability of
successful connection defined in Definition 1 is
"
2
2π 2 λP Pactive βPα r12
σ2
Pcon (rg , λS ) = exp − λS πrg2 + βP P r1α +
Pt
α sin( α2 π)
Proof: See Appendix A.
#!
.
(10)
13
Remark 3. According to Definition 1 of Pcon , the effect of rg on Pcon is two fold. On one hand,
increasing the value of rg decreases the probability of the PT being active, which decreases Pcon .
This effect appears in the first term in the exponent in Eq. 10. On the other hand, increasing
the value of rg decreases the density of active interferers. Hence, the probability of having SINR
higher than βP increases, which increases Pcon . This effect appears in the third term in the
exponent in Eq. 10 implicitly in the value of Pactive .
To better understand the effect of rg on Pcon , we derive the value of r̂g that maximizes Pcon
in the following Theorem.
2
Theorem 2. Defining A1 =
•
•
2π 2 λP βPα r12
2
α sin( α
π)
then:
If A1 ≤ 1, then Pcon is a decreasing function of rg , and r̂g = 0.
q
1)
If A1 > 1, then r̂g = ln(A
.
πλS
Proof: See Appendix B.
Remark 4. From the above Theorem, we conclude that when A1 ≤ 1 the effect of rg on
the density of active interferers is negligible. This is consistent with intuition. Observing the
expression of A1 in Theorem 2, we note that it is an increasing function of λP , βP , and r1 .
Hence, A1 ≤ 1 results from one or more of the following conditions: (i) λP is too small to
take the effect of interference into consideration (the system is noise limited), (ii) r1 is too small
such that the received signal power from the intended PT is hardly attenuated by the path-loss,
and (iii) βP is a very relaxed threshold. These three consequences of having A1 ≤ 1 make the
condition of Re ≥ rg in Definition 1 of Pcon dominate the other condition of SINRP ≥ βP . This
eventually makes Pcon a decreasing function of rg , similar to P(Re ≥ rg ).
In the following corollary, an upper bound on the value of Pcon (rg∗ , λS ) is provided.
Corollary 1. Using the result in Theorem 2, the value of Pcon (rg∗ , λS ) is upper bounded as
follows
s
Pcon (rg∗ , λS ) ≤ Pcon
where A1 is defined in Theorem 2.
ln (max{A1 , 1})
, λS ,
πλS
(11)
14
Proof: The proof follows by observing that r̂g is the value of rg that maximizes Pcon without
any constraints, while rg∗ is the value of rg that maximizes Pcon with the constraint of rg ∈ G(λS ),
as explained in Definition 3.
2) Secure Communication Probability: The secure communication probability, formally defined in Definition 2, is derived in the following theorem.
Theorem 3 (Secure communication probability). For a given value of rg and λS , the probability
of secure communication is
Z
Psec (rg , λS ) = exp −2πλS
rg
where LI2 (s) = exp −2πλP Pactive
R srg−α
0
!
2
σS βS ryα
LI2 (βS ryα )ry dry ,
exp −
Pt
2
sα
.
2 dz
∞
(12)
α(1+z)z α
Proof: See Appendix C.
Remark 5. The intuitive observations provided in Remark 1 can now be verified using Theorem 3.
For instance, the effect of rg on the distance between the PT and its nearest ER is captured
in the integration limits in the above equation. Obviously, as we increase rg , the integration
interval will decrease, which increases the value of Psec . On the other hand, the effect of rg on
the density of active interferers at the ER is captured in the term LI2 (βS rxα ). As we mentioned
earlier, increasing the value of rg decreases the interference levels at the ER, which increases
the value of SINRS , which eventually reduces Psec . This is verified here by observing that LI2 (s)
is an increasing function of rg .
Due to the complexity of the expression provided in Theorem 3 for Psec , it is not straightforward to derive expressions for either G(λS ) or rg∗ . However, using the results for Pcon and Psec
in Theorems 1 and 3, it is easy to compute rg∗ numerically. In order to better understand the
behavior of the system, we will provide expressions for Pcon and Psec in both noise-limited as
well as interference-limited regimes. In the discussion section, results for both regimes will be
compared with the results of the system at different values of rg resulting in several meaningful
insights. In the case of a noise limited regime, the interference terms in the expressions of both
Pcon and Psec will be removed leading to the following corollary.
15
Corollary 2. In the noise limited regime, Pcon will be a decreasing function of rg , while Psec
will be an increasing function of rg as follows
σP2 α
2
Noise limited
,
= exp − λS πrg + βP r1
Pcon
Pt
!
2
Z ∞
α
σ
β
r
S
y
S
Noise limited
Psec
= exp −2πλS
exp −
ry dry
Pt
rg
α2
!
Pt
2 rgα βS σS2
2πλS
Γ
,
.
= exp −
α
σS2 βS
α
Pt
And the value of rg∗ is presented by the following equation
1
∗ α
2
αLn
2 (rg ) βS σS
2
Γ
= min
,
α2 , Γ α .
α
Pt
2πλ
Pt
S σ 2 βS
(13)
(14)
(15)
S
Proof: Eq. 13 and 14 follow directly by removing the interference terms from the results in
Noise
Theorem 1 and Theorem 3. Since Pcon
limited
Noise
is a decreasing function of rg , and Psec
limited
is an increasing function of rg , the primary network picks the minimum value of rg that satisfies
the condition Psec ≥ . Substituting for Psec using Eq. 14 in the inequality Psec ≥ leads to
Eq. 15.
In the case of an interference-limited regime, the noise terms in the expressions of both Pcon
and Psec will be ignored leading to the following corollary.
Corollary 3. In the interference limited regime, expressions for Pcon and Psec are provided below.
#!
"
2
2
α 2
2π
λ
P
β
r
P
active
Int. limited
P 1
,
(16)
Pcon
= exp − λS πrg2 +
α sin( α2 π)
!
Z ∞
Int. limited
(17)
Psec
= exp −2πλS
LI2 (βS rxα )rx drx .
rg
Proof: These results follow by substituting for σP2 = σS2 = 0 in Theorems 1 and 3.
The main take-away from this subsection is that each of Pcon and Psec are functions of λS ,
which consequently implies that rg∗ is a function of λS . In the next subsection, we will show
that optimal density λ∗S selected by the secondary network is also a function of rg .
B. Secondary Network
The objective of the secondary network is to maximize the density of successfully powered
ERs Penergy , introduced in Eq. 9, which is derived in the following Theorem.
16
Theorem 4 (Density of successfully powered ERs). For a given value of rg and λS , the density
of successfully powered ERs is
Z ∞
Emin rpα
2
2
2πλP Pactive rp exp −πλP Pactive (rp − rg ) −
Penergy (rg , λS ) = λS
drp .
Pt η
rg
(18)
Proof: See Appendix D.
The above equation is a product of two terms: (i) the density of ERs λS , and (ii) the integral
term, which is the expression derived for the probability P(EH ≥ Emin ). Consequently, we can
claim the existence of an optimal value of λS that maximizes the density of successfully powered
ERs Penergy . The reason behind that is the dependence of the density of active sources of RF
energy (active PTs) on the density of ERs. This interesting trade-off arises due to the use of
secrecy guard zones by the primary network. Obviously, the value of Penergy is a function of rg .
Hence, the value of λ∗S , that maximizes Penergy , is also a function of rg . Unfortunately, λ∗S can
not be computed from the above expression due to its complexity. Hence, to get more insights
on the relation between λ∗S and rg , we derive a lower bound on Penergy in the following corollary.
We use this lower bound to compute λ∗S as a function of rg .
Corollary 4. The value of Penergy is lower bounded as follows
Z ∞
Emin rpα
2
2
drp .
Penergy ≥ λS
2πλP Pactive rp exp −πλP (rp − rg ) −
Pt η
rg
(19)
The value of λS that maximizes the lower bound is
λ∗S =
1
.
πrg2
(20)
Proof: The lower bound follows by simply replacing Pactive inside the exponent in Eq. 18
with unity. The value of λ∗S follows by differentiating Eq. 19 with respect to λS .
Remark 6. It can be noted from Eq. 20 that the value of λ∗S is a decreasing function of rg .
This agrees with intuition since when the value of rg increases, the secondary network needs to
decrease λS in order to maintain the density of active PTs, which are the sources of RF energy.
The mutual coupling between rg∗ and λS as well as λ∗S and rg can be best modeled using tools
from game theory. This will be the core of the next section.
17
IV. G AME T HEORETICAL M ODELING
Building on all the insights and comments given in the previous section, we can model the
interaction between the two networks using tools from game theory. For a given value of λS ,
the primary network selects rg∗ as presented in Definition 3, which can be rewritten as
rg∗ = arg max Pcon (rg , λS )1(Psec (rg , λS ) ≥ ),
rg ≥0
(21)
where 1(Ξ) = 1 if the condition Ξ is satisfied, and equals zero otherwise. On the other hand,
for a given value of rg , the secondary network selects λ∗S as follows
λ∗S = arg max Penergy (rg , λS ).
λS ≥0
(22)
Observing Eq. 21 and 22, it is fairly straightforward to see that the system setup can be modeled
as a non-cooperative static game with two players: (i) the primary network is player 1, and (ii)
the secondary network is player 2. The utility function of player 1 is
U1 (rg , λS ) = Pcon (rg , λS )1(Psec (rg , λS ) ≥ ),
(23)
while the utility function of player 2 is
U2 (rg , λS ) = Penergy (rg , λS ).
(24)
For this game, the main objective is to find the values of rg∗ and λ∗S where each of the two networks
has no tendency to change its tuning parameter. This is the definition of Nash equilibrium (NE).
This can be mathematically modeled as
rg∗ = arg max Pcon (rg , λ∗S )1(Psec (rg , λ∗S ) ≥ ),
rg ≥0
λ∗S = arg max Penergy (rg∗ , λS ).
λS ≥0
(25)
Unfortunately, due to the complexity of the expressions of Pcon in Theorem 1 and Psec in
Theorem 3, it is challenging to provide a closed-form solution for NE. However, based on the
comments given in Remark 2 and the result in Theorem 2, we provide a learning algorithm that
assists both networks to find NE.
The proposed learning algorithm is a simple best-response based algorithm [48]–[50]. In each
iteration, each network updates its parameters according to the opponent’s network parameter in
the previous iteration. This can be mathematically presented as follows:
(n−1)
rg(n) = rg∗ λS
,
18
(n)
λS = λ∗S rg(n−1) ,
(26)
(n)
(n)
(n−1)
where rg and λS are the outputs of the algorithm in the n-th iteration, rg∗ λS
is computed
(n−1)
(n−1)
(n−1)
is computed using Eq. 22 for rg = rg
.
using Eq. 21 for λS = λS , and λ∗S rg
Assuming that the secondary network has a maximum possible deployment density, λS ≤ λS,max ,
the algorithm is provided next.
Algorithm: Proposed algorithm for finding the NE network parameters rg∗
and λ∗S
input: λP , λS,max , βP , βS , Pt , α, η, and Emin .
output: rg∗ , λ∗S .
(0)
(0)
Initialization: rg = 0 for all PTs in ΦP , λS = λS,max , n = 1
1: Repeat
3:
(n−1)
(n)
rg = rg∗ λS
(n−1)
(n)
∗
λS = λS rg
4:
n=n+1
2:
(n)
(n−1)
5: Until rg = rg
(n)
(n−1)
& λS = λS
Note that we assume the ability of each network to estimate perfectly the opponent’s action
in the previous iteration. Including the possibility of the estimation error and its effect on
convergence is left as a promising direction of future work. In the next section, we will show
using simulations the convergence of the proposed algorithm to the NE of the modeled game.
V. R ESULTS AND D ISCUSSION
In this section, we will use both theoretical and simulation results (obtained from Monte-Carlo
trials) to analyze the performance of both primary and secondary networks. The values of the
system parameters used for the simulation setup are: λP = 1, η = 0.75, Emin = 10−4 Joules,
α = 4, λS,max = 2, Pt = 1, βP = 3 dB, βS = 0 dB, r1 = 0.1, and = 0.8. The values of the
rest of the parameters will be defined either on the figures or in their captions. We also refer to
the SNR values at the primary and secondary receivers as γP =
Pt
2
σS
and γS =
In addition, we define the ratio of λS to its maximum value by δS =
λS
.
λS,max
Pt
2
σS
respectively.
Our main goals
in this section are: (i) to validate the approximations used in our derivations, (ii) to get further
Probability of successful connection Pcon(rg,λS )
19
1
Theorem 1
Simulation
0.8
Increasing λS : [0.5, 0.65, 0.8, 0.95]
0.6
0.4
0.2
0
0
0.5
1
1.5
Guard zone radius rg
2
Fig. 3. The effect of λS on the behavior of Pcon against different values of rg . The SNR value at the legitimate receiver is
γP = 7 dB.
insights on the behavior of both the networks, and (iii) to verify the comments provided in
the Remarks throughout the paper. We first study the successful connection probability Pcon in
Fig. 3. First note that the simulation parameters chosen above are such that we get A1 < 1,
where A1 is defined in Theorem 2. The results in Fig. 3 show that increasing the density of
ERs decreases Pcon . Although increasing λS decreases the density of active interferers (which
should increase Pcon ), it also decreases the probability of the PT being active (which decreases
Pcon ). Ultimately, for this setup, Pcon is a decreasing function of λS because of having A1 < 1.
As explained in Remark 4, A1 < 1 leads to a noise limited system from the perspective of the
successful connection probability, which means that the effect of λS on the density of active
interferers is negligible compared to its effect on the probability of the PT being active. Further,
note that because of having A1 < 1, Pcon is also a decreasing function of rg . As a result, rg∗
will be the minimum value of rg that ensures Psec ≥ . The value of Psec as a function of rg is
plotted in Fig. 4 for different values of γS . Consistent with the intuition, increasing γS decreases
Psec . We also note that at low values of rg the effect of γS is hardly noticeable, while at higher
values of rg , the effect of γS becomes significant. Furthermore, we note that rg∗ is an increasing
function of γS . To better understand the effect of γS on the behavior of Psec , we plot in Fig. 5 the
value of Psec for both the noise limited and the interference limited regimes, which are derived
in Corollary 2 and 3, respectively. These results lead to the following key insights:
Secure communication probability Psec(rg,λS)
Secure communication probability Psec(rg,λS)
20
1
Increasing γS :
[1, 3, 5, 8] dB
0.8
0.6
r*g
0.4
Theorem 3
Simulation
0.2
0
0.5
1
1.5
2
2.5
3
1
Increasing γS :
[1, 3, 5, 8] dB
0.8
0.6
0.4
Theorem 3
Interference limited
Noise limited
0.2
0
0
0.5
1
1.5
2
2.5
3
Guard zone radius rg
Guard zone radius rg
NoiseLimited
Int.Limited
Fig. 5. Comparing Psec , Psec
, and Psec
Fig. 4. The effect of γS on the behavior of Psec against
as functions of rg for different values of γS . The density
different values of rg . The density of ERs is λS = 0.6.
of ERs is λS = 0.6.
•
At lower values of rg , the density of active interferers is relatively high. Hence, we observe
that from the secure communication probability perspective, the system is interference
limited in this regime. Furthermore, increasing the SNR value γS has a negligible effect.
This is a direct consequence of being in the interference-limited regime.
•
At higher values of rg , the density of active interferers decreases, which drives the system
to the noise limited regime. As a result, the noise power has more noticeable effect on Psec .
Equivalently, increasing value of γS decreases the value of Psec at higher values of rg .
•
Int.
Obviously, Psec
Limited
Noise
is a decreasing function of rg , while Psec
Limited
is an increasing
function of rg . However, the behavior of Psec is harder to describe. We note the existence
Int.
of a local minimum of Psec below which it behaves more like Psec
Noise
this local minimum, Psec behaves similar to Psec
Limited
Limited
, whereas above
.
In Fig. 6, we study the effect of λS on the behavior of Psec against different values of rg . We
note that rg∗ increases with λS . We also note that, unlike γS , the effect of λS is more prominent
at the lower values of rg . As explained above, the reason is that the system is interference
limited at lower values of rg , where increasing λS decreases the density of interferers, which in
Int.
turn decreases Psec
Limited
Int.
. We plot Psec , Psec
Limited
Noise
, and Psec
Limited
in Fig. 7 to verify these
arguments. We also note that at higher values of rg , as we stated earlier, the system is noise
Noise
limited, where λS has less effect on Psec
Limited
Int.
compared to Psec
Limited
. This can be verified
1
Increasing λS :
[0.5, 0.65, 0.8, 0.95]
0.8
0.6
r*g
0.4
Theorem 3
Simulation
0.2
0
0.5
1
1.5
2
Guard zone radius rg
Secure communication probability Psec(rg,λS)
Secure communication probability Psec(rg,λS)
21
1
0.8
Increasing λS : [0.5, 0.8]
0.6
0.4
Theorem 3
Interference limited
Noise limited
0.2
0
0
0.5
1
1.5
2
Guard zone radius rg
Fig. 6. The effect of λS on the behavior of Psec against
NoiseLimited
Int.Limited
Fig. 7. Comparing Psec , Psec
, and Psec
different values of rg . The SNR value at ERs is γS = 4.8
as functions of rg for different values of λS . The SNR
dB.
value at ERs is γS = 4.8 dB.
from Fig. 6 by observing that the gaps between the Psec curves for different values of λS get
tighter as rg increases.
Now that we have better understanding of the behavior of Pcon and Psec , we study the energy
harvesting performance of the secondary network. In Fig. 8, we illustrate the behavior of Penergy
against λS for different values of rg . We note that at lower values of rg , the optimal value of
λS is its maximum value λS,max . This is consistent with intuition that at lower values of rg , the
density of active PTs (sources of RF energy) will hardly get affected by increasing λS . As rg
increases, the impact of λS on the density of active PTs becomes noticeable. As shown in Fig. 8,
this eventually leads to the existence of a local maximum of Penergy . We also note that as the
value of rg increases, the optimal value λ∗S decreases. This is also consistent with our comments
in Remark 6 that as rg increases, the secondary network needs to decrease its density in order
to maintain the density of active PTs that provide the RF energy to the ERs.
Taking a second look at Figs. 6 and 8, we can easily verify our earlier comments that
the parameters selected by each of the networks (rg for the primary network and λS for the
secondary network) affect the value of the optimal parameter of the other network. Using the
procedure provided in Sec. IV, we simulate the interaction between both networks by simulating
the relations between rg∗ and δS as well as δS∗ and rg , where δS =
λS
λS,max
is the normalized value
of the density of ERs. In Fig. 9, we plot rg∗ (on the y-axis) for different values of δS (on the
Density of successfully powered ERs Penergy(rg,λS)
22
2
Theorem 4
Simulation
1.5
Increasing r g:
λ*S =2
[0.6, 0.8, 1, 1.2]
1
λ*S =0.4
λ*S =1.4
0.5
λ*S =0.85
0
0.5
1
1.5
2
Density of ERs λS
Fig. 8. The density of successfully powered ERs Penergy against λS for different values of rg .
Output of the proposed algorithm
2
r*g/rg
1.5
1
NE: r*g=1.48, δ*s=0.33
0.5
r*g using Eq. 21
λ*S
using Eq. 22
0
0.2
0.4
0.6
0.8
1
2
r(n)
g
δ(n)
S
1.5
Algorithm converged to
1
r(n)
=1.48,δ(n)
=0.33
g
S
0.5
0
2
δS/δ*S
4
6
8
10
12
14
Number of iterations n
Fig. 9. Plotting both rg∗ against δS and δS∗ against rg on
Fig. 10.
the same plot to compute the NE values.
equilibrium in a finite number of iterations.
The proposed algorithm converges to Nash
x-axis). On the same figure, we plot δS∗ (on the x-axis) for different values of rg (on the y-axis).
The intersection of both curves represents the value of NE where each of the two networks
has no intention to deviate as explained in Eq. 25. In Fig. 10, we evaluate the performance of
the algorithm proposed in Sec. IV. The results demonstrate the convergence of the proposed
algorithm to the NE found from Fig. 9 after less than 13 iterations.
23
VI. C ONCLUSIONS
The emergence of IoT regime is characterized by the deployment of billions of devices some
of which may be equipped with energy harvesting capability. Due to the ubiquity of RF signals,
harvesting energy from the ambient RF signals is perhaps the most attractive option for such
devices. This raises the possibility of some of these devices acting as eavesdroppers, which
motivates the need to study secure wireless power transfer to energy receivers acting as potential
eavesdroppers. While this problem received some attention in the literature, the existing works
are limited to simple point-to-point or simple deterministic topologies, which are not sufficient
to accurately model the massive scale of IoT. In this paper, we developed the first comprehensive
stochastic geometry-based model to study the performance of an ambient RF energy harvesting
network (secondary network) when the sources of RF signals are transmitters with secrecy guard
zones (primary network). First, using tools from stochastic geometry, we derived the successful
connection and secure communication probabilities of the primary network. Next, we derived the
density of successfully powered nodes in the secondary network. Furthermore, we showed that the
performance metrics of the two networks are coupled. In particular, we showed that the optimal
guard zone radius (that maximizes the successful connection probability while maintaining the
secure communication probability above a predefined threshold) is a function of the deployment
density of the secondary network. In addition, we showed that the optimal deployment density of
the secondary network (that maximizes the density of successfully powered nodes) is a function
of the guard zone radius of the primary network. Hence, we used tools from game theory to model
this interesting coupling between the two networks. In particular, we showed that such system
can be modeled as a two player non-cooperative game. In addition, we proposed a best-response
based algorithm and demonstrated with simulations its convergence to Nash equilibrium.
This paper is one of the few concrete works that symbiotically merge tools from stochastic
geometry and game theory. It can be extended in many directions. From the secrecy perspective,
it will be useful to extend the proposed model to incorporate other secrecy enhancing techniques,
such as beamforming and artificial noise technique. From the modeling perspective, it is worthwhile to investigate other meaningful morphologies, such as the one in which ERs are clustered
around the RF sources. From game theory perspective, it will be interesting to consider the effect
of imperfect estimation of the opponent’s action on the performance of the proposed algorithm.
24
A PPENDIX A
P ROOF OF T HEOREM 1
Recalling the expression for SINRP given in Eq. 2, specifically the indicator function δi that
indicates which interferer is active and which is silent, we concluded that the locations of active
PTs can be modeled by PHP Ψ in Eq. 3. However, before using Ψ in our analysis, we need to
make it clear that δi for different xi ∈ ΦP are correlated. This implicit correlation arises from
the dependence of δi for all i on the PPP ΦS . However, capturing this correlation in our analysis
will significantly reduce the tractability of the results. Hence, this correlation will be ignored in
our analysis for the sake of tractability. The tightness of this approximation will be verified in
the Numerical Results section. Now revisiting the expression of Pcon in Eq. 4, we note that the
correlation between δ1 at the typical PT and δi values at each of the interferers in the expression
of SINRP is the only source of correlation between the events (Re ≥ rg ) and (SINRP ≥ βP ).
Hence, ignoring this correlation will lead to the following
Pcon = P(Re ≥ rg )P(SINRP ≥ βP ).
(27)
The first term in the above expression represents Pactive = exp −πλS rg2 (please recall Eq. 1
where Pactive was derived). To derive the second term in the above expression, characterizing
the statistics of the interference from a PHP modeled network at a randomly located reference
point (the typical PR) is required. However, ignoring the correlation between {δi }, for the sake
of tractability as explained above, is equivalent to approximating the PHP Ψ with a PPP ΨP of
P
equivalent density λ̃P = λP Pactive . Defining I = xi ∈ΨP hi kxi k−α , then
!
σP2
h1 r1−α
(a)
= EI exp −βP I +
r1α
P(SINRP ≥ βP ) = P
2 ≥ βP
σP
P
t
I + Pt
σP2 α
σP2 α
α (b)
= exp −βP r1 EI [exp (−βP Ir1 )] = exp −βP r1 LI (βP r1α ) , (28)
Pt
Pt
where h1 ∼ exp(1) leads to step (a), and in step (b) we use the definition of Laplace transform
of I which is LI (s) = E [exp (−sI)]. The Laplace transform of the interference in PPP is a well
established result in the literature [23]. For completeness, the derivation of LI (s) is provided
next.
"
LI (s) = EΨP ,{hi } exp −s
!#
X
xi ∈ΨP
hi kxi k−α
"
= EΨP ,{hi }
#
Y
xi ∈ΨP
exp −shi kxi k
−α
25
"
(c)
Y
= EΨP
xi ∈ΨP
Z
(e)
= exp −2π λ̃P
0
∞
#
1
= exp −λ̃P
1−
dx
1 + skxk−α
x∈R2
!
2
2π 2 λ̃P s α csc 2π
srx−α
(f )
α
,
rx drx = exp −
1 + srx−α
α
1
1 + skxi k−α
(d)
Z
(29)
where knowing that the set of fading gains hi are i.i.d with hi ∼ exp(1) leads to step (c), step (d)
results from using the probability generating function (PGFL) of PPP [24], step (e) results from
converting to polar co-ordinates, and step (f) follows after some mathematical manipulations.
Substituting Eq. 29 in Eq. 28 and then in Eq. 27 leads to the final result in Theorem 1.
A PPENDIX B
P ROOF OF T HEOREM 2
Observing the expression of Pcon in Theorem 1, we note that it can be rewritten as a function
of Pactive = exp(−λS πrg2 ) as follows
σP2 α
Pcon = exp −βP r1 Pactive exp (−Pactive A1 ) ,
Pt
(30)
2
where A1 =
2π 2 λP βPα r12
.
2
α sin( α
π)
To get more information about the behavior of Pcon against Pactive , we
compute the first derivative (with respect to Pactive ). Given that Pactive is a decreasing function
of rg (recall Eq. 1), we conclude the following
1) If 1 − A1 Pactive ≥ 0, then Pcon is a decreasing function of rg ,
2) If 1 − A1 Pactive ≤ 0, then Pcon is an increasing function of rg .
Consequently, we can infer that, since 0 ≤ Pacitve ≤ 1, Pcon is a decreasing function of rg as
long as A1 ≤ 1. In the case of A1 ≥ 1, the relation between Pcon and rg can be explained as
q
1)
), and
follows: (i) Pcon is an increasing function of rg as long as Pactive ≥ A11 (or rg ≤ ln(A
λS π
q
1)
(ii) Pcon is a decreasing function of rg as long as Pactive ≤ A11 (or rg ≥ ln(A
). This concludes
λS π
the proof.
A PPENDIX C
P ROOF OF T HEOREM 3
From Definition 2 of Psec , we observe that we need to jointly analyze the values of SINRS (yj )
at all the locations yj ∈ ΦS . Despite the usual assumption throughout most of the stochastic
geometry-based literature on secrecy analysis that these values are uncorrelated, this is actually
not precise. The reason for that is the dependence of SINRS (yj ), by definition, on the PPP
26
ΦP for all yj ∈ ΦS . Some recent works started working on characterizing the correlation
between intereference levels at different locations [51]. However, most of these works focus on
characterizing the correlation between only two locations assuming the knowledge of the distance
between them. Unfortunately, these results will not be useful for our analysis. Hence, aligning
with the existing literature, we will ignore this correlation in our analysis with the knowledge
that this will provide an approximation. Furthermore, the accuracy of this approximation is
expected to get worse as the value of λS increases. This is due to the fact that the distances
between ERs decrease as λS increases, which was shown in [51] to increase the correlation.
For notational simplicity, and without any loss of generality due to the stationarity of PPP, we
will assume that the typical PT is placed at the origin, i.e. x1 = o, in the rest of this proof.
All the analysis provided in this section is conditioned on the event Re ≥ rg . Following the
same approach as in Appendix A of approximating the PHP Ψ with a PPP ΨP , and defining
P
I2 (yj ) = xi ∈ΨP \x1 gi,j kxi − yj k−α , Psec can be derived as follows
!
−α
Y
[ g1,j kyj k−α
g ky k
(g)
= EΦS ,I2 ,{g1,j }
1 1,j j σ2 ≤ βS
Psec = EΦS ,I2 ,{g1,j } 1
2 ≤ βS
σS
I2 (yj ) + PSt
yj ∈ΦS
yj ∈ΦS I2 (yj ) + Pt
2
σS
β
I
(y
)
+
Y
S
2 j
Pt
(h)
1 − exp −
= EΦS ,I2
−α
ky
jk
yj ∈ΦS
2
σS
β
Y
S
Pt
(i)
1 − exp −
E exp − βS I2 (yj ) ,
= EΦS
(31)
−α
−α
ky
ky
jk
jk
y ∈Φ
j
S
where step (g) (and step (i)) follow from assuming that the values of SINRS (yj ) (and I2 (yj ))
are uncorrelated, as we discussed earlier in this Appendix. Step (h) is due to assuming the set
of fading gains {g1,j } to be i.i.d with g1,j ∼ exp(1). Defining the Laplace transform of I2 (yj )
by LI2 (yj ) (s) = E[exp(−sI2 (yj ))], we note that there is only one difference in the derivation of
LI2 (yj ) (s) compared to that of LI (s) in Eq. 29. The difference is in the reference point from
where we are observing the interference. In Appendix A, the reference point was the PR, which
does not have a minimum distance from any active interfering PT. In the current derivation, the
reference point is an ER, which has a minimum distance of rg from any active interfering PT.
Hence, the derivation of LI2 (yj ) (s) will be exactly the same as in Eq. 29 until step (e), where
the minimum distance effect will appear in the lower limit of the integral as follows
!
Z ∞
srx−α
r drx .
LI2 (yj ) (s) = exp −2π λ̃P
−α x
rg 1 + srx
(32)
27
Note that the above expression is not a function of yj , so we drop it from the notation of
Laplace transform. The final expression for LI2 (s) as provided in Theorem 3 follows after
simple mathematical manipulations. Substituting Eq. 32 in Eq. 31, we get
2
Y
σS
Psec = EΦS
1 − exp −βS
kyj kα LI2 (βS kyj kα )
Pt
y ∈Φ
j
(k)
S
= exp −2πλS
!
2
σS
exp −βS
kykα LI2 (βS kykα ) dy ,
P
2
t
y∈R ∩B(o,rg )
Z
(33)
where step (k) results from applying PGFL of PPP, and the integration is over y ∈ R2 ∩ B(o, rg )
because the analysis in this section is conditioned on the event Re ≥ rg , which means that
the typical PT is active. Since we assumed that the typical PT is placed at the origin in this
derivation, the ball B(o, rg ) is clear of ERs. Converting from Cartesian to polar co-ordinates
leads to the final result in Theorem 3.
A PPENDIX D
P ROOF OF T HEOREM 4
The density of successfully powered ERs can be derived as follows
(l)
Emin Rpα
−α
,
Penergy = λS P ηPt Rp w ≥ Emin = ERp exp −
ηPt
(34)
where Rp is the distance between the ER and its nearest active PT, and step (l) is due to
w ∼ exp(1). The distance Rp represents the contact distance of a PHP observed from a hole
center. Unfortunately, the exact distribution of this distance is unknown. However, the approach
of approximating the PHP Ψ with a PPP ΨP is known to provide fairly tight approximation of
the contact distance distribution of PHP [52]. Given that the nearest active PT to the ER is at a
distance of at least rg , the distribution of Rp is
2
fRp (rp ) = 2π λ̃P exp −π λ̃P (rp − rg ) ,
rp ≥ rg .
(35)
Using this distribution to compute the expectation in Eq. 34 leads to the final result in Theorem 4.
R EFERENCES
[1] M. A. Kishk and H. S. Dhillon, “Modeling and analysis of ambient RF energy harvesting in networks with secrecy guard
zones,” in Proc., IEEE Wireless Commun. and Networking Conf. (WCNC), March 2017.
[2] H. S. Dhillon, H. Huang, and H. Viswanathan, “Wide-area wireless communication challenges for the Internet of Things,”
IEEE Commun. Mag., vol. 55, no. 2, pp. 168–174, Feb. 2017.
28
[3] H. S. Dhillon, H. C. Huang, H. Viswanathan, and R. A. Valenzuela, “Power-efficient system design for cellular-based
machine-to-machine communications,” IEEE Trans. on Wireless Commun., vol. 12, no. 11, pp. 5740 – 5753, Nov. 2013.
[4] ——, “Fundamentals of throughput maximization with random arrivals for M2M communications,” vol. 62, no. 11, pp.
4094 – 4109, Nov. 2014.
[5] V. Jelicic, M. Magno, D. Brunelli, V. Bilas, and L. Benini, “Analytic comparison of wake-up receivers for WSNs and
benefits over the wake-on radio scheme,” in Proc., ACM workshop on Perform. Monit. and Measurement of Heterog.
Wireless and Wired Netw., Oct. 2012, pp. 99–106.
[6] P. Kamalinejad, C. Mahapatra, Z. Sheng, S. Mirabbasi, V. C. Leung, and Y. L. Guan, “Wireless energy harvesting for the
internet of things,” IEEE Commun. Magazine, vol. 53, no. 6, pp. 102–108, 2015.
[7] H. Xing, L. Liu, and R. Zhang, “Secrecy wireless information and power transfer in fading wiretap channel,” IEEE Trans.
on Veh. Technology, vol. 65, no. 1, pp. 180–190, Jan. 2016.
[8] Q. Li, W. K. Ma, and A. M. C. So, “Robust artificial noise-aided transmit optimization for achieving secrecy and energy
harvesting,” in Proc., IEEE Intl. Conf. on Acoustics, Speech, and Sig. Proc. (ICASSP), May 2014, pp. 1596–1600.
[9] M. Zhang, Y. Liu, and R. Zhang, “Artificial noise aided secrecy information and power transfer in OFDMA systems,”
IEEE Trans. on Wireless Commun., vol. 15, no. 4, pp. 3085–3096, 2016.
[10] M. R. Khandaker and K.-K. Wong, “Robust secrecy beamforming with energy-harvesting eavesdroppers,” IEEE Wireless
Commun. Letters, vol. 4, no. 1, pp. 10–13, 2015.
[11] D. W. K. Ng, E. S. Lo, and R. Schober, “Robust beamforming for secure communication in systems with wireless
information and power transfer,” IEEE Trans. on Wireless Commun., vol. 13, no. 8, pp. 4599–4615, 2014.
[12] J. Xu, L. Liu, and R. Zhang, “Multiuser miso beamforming for simultaneous wireless information and power transfer,”
IEEE Trans. on Signal Processing, vol. 62, no. 18, pp. 4798–4810, 2014.
[13] K. Banawan and S. Ulukus, “Mimo wiretap channel under receiver-side power constraints with applications to wireless
power transfer and cognitive radio,” IEEE Trans. on Commun., vol. 64, no. 9, pp. 3872–3885, 2016.
[14] A. D. Wyner, “The wire-tap channel,” The Bell System Technical Journal, vol. 54, no. 8, pp. 1355–1387, Oct. 1975.
[15] S. H. Chae, W. Choi, J. H. Lee, and T. Q. S. Quek, “Enhanced secrecy in stochastic wireless networks: Artificial noise
with secrecy protected zone,” IEEE Trans. on Information Forensics and Security, vol. 9, no. 10, pp. 1617–1628, Oct.
2014.
[16] W. Liu, Z. Ding, T. Ratnarajah, and J. Xue, “On ergodic secrecy capacity of random wireless networks with protected
zones,” IEEE Trans. on Vehicular Technology, vol. 65, no. 8, pp. 6146–6158, Aug. 2016.
[17] X. Zhou and M. R. McKay, “Secure transmission with artificial noise over fading channels: Achievable rate and optimal
power allocation,” IEEE Trans. on Vehicular Technology, vol. 59, no. 8, pp. 3831–3842, 2010.
[18] X. Zhou, R. K. Ganti, J. G. Andrews, and A. Hjorungnes, “On the throughput cost of physical layer security in decentralized
wireless networks,” IEEE Trans. on Wireless Commun., vol. 10, no. 8, pp. 2764–2775, Aug. 2011.
[19] P. C. Pinto, J. Barros, and M. Z. Win, “Secure communication in stochastic wireless networks–Part I: Connectivity,” IEEE
Trans. on Info. Forensics and Security, vol. 7, no. 1, pp. 125–138, Feb. 2012.
[20] M. A. Kishk and H. S. Dhillon, “Stochastic geometry-based comparison of secrecy enhancement techniques in D2D
networks,” IEEE Wireless Commun. Letters, to appear.
[21] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for modeling, analysis, and design of multi-tier and cognitive
cellular wireless networks: A survey,” IEEE Commun. Surveys and Tutorials, vol. 15, no. 3, pp. 996 – 1019, 2013.
[22] H. ElSawy, A. Sultan-Salem, M. S. Alouini, and M. Z. Win, “Modeling and analysis of cellular networks using stochastic
geometry: A tutorial,” IEEE Commun. Surveys Tutorials, vol. 19, no. 1, pp. 167–203, Firstquarter 2017.
29
[23] J. G. Andrews, A. K. Gupta, and H. S. Dhillon, “A primer on cellular network analysis using stochastic geometry,” available
online: arxiv.org/abs/1604.03183.
[24] M. Haenggi, Stochastic geometry for wireless networks.
Cambridge University Press, 2012.
[25] G. Geraci, H. S. Dhillon, J. G. Andrews, J. Yuan, and I. B. Collings, “Physical layer security in downlink multi-antenna
cellular networks,” IEEE Trans. on Commun., vol. 62, no. 6, pp. 2006–2021, 2014.
[26] G. Chen, J. P. Coon, and M. Di Renzo, “Secrecy outage analysis for downlink transmissions in the presence of randomly
located eavesdroppers,” 2017, available online: arxiv.org/abs/1701.00982.
[27] H. S. Dhillon, Y. Li, P. Nuggehalli, Z. Pi, and J. G. Andrews, “Fundamentals of heterogeneous cellular networks with
energy harvesting,” IEEE Trans. on Wireless Commun., vol. 13, no. 5, pp. 2782 – 2797, May 2014.
[28] M. Di Renzo and W. Lu, “System-level analysis and optimization of cellular networks with simultaneous wireless
information and power transfer: Stochastic geometry modeling,” IEEE Trans. on Veh. Technology, vol. 66, no. 3, pp.
2251–2275, March 2017.
[29] M. A. Kishk and H. S. Dhillon, “Downlink performance analysis of cellular-based IoT network with energy harvesting
receivers,” in IEEE GLOBECOM, Dec. 2016.
[30] A. H. Sakr and E. Hossain, “Analysis of k-tier uplink cellular networks with ambient RF energy harvesting,” IEEE Journal
on Sel. Areas in Commun., vol. 33, no. 10, pp. 2226–2238, 2015.
[31] I. Flint, X. Lu, N. Privault, D. Niyato, and P. Wang, “Performance analysis of ambient RF energy harvesting with repulsive
point process modeling,” IEEE Trans. on Wireless Commun., vol. 14, no. 10, pp. 5402–5416, 2015.
[32] Y. Liu, L. Wang, S. A. R. Zaidi, M. Elkashlan, and T. Q. Duong, “Secure D2D communication in large-scale cognitive
cellular networks: A wireless power transfer model,” IEEE Trans. on Commun., vol. 64, no. 1, pp. 329–342, Jan. 2016.
[33] A. Houjeij, W. Saad, and T. Basar, “A game-theoretic view on the physical layer security of cognitive radio networks,” in
Proc., IEEE Intl. Conf. on Commun. (ICC), June 2013, pp. 2095–2099.
[34] W. Saad, X. Zhou, B. Maham, T. Basar, and H. V. Poor, “Tree formation with physical layer security considerations in
wireless multi-hop networks,” IEEE Trans. on Wireless Commun., vol. 11, no. 11, pp. 3980–3991, Nov. 2012.
[35] S. Bayat, R. H. Y. Louie, Z. Han, B. Vucetic, and Y. Li, “Physical-layer security in distributed wireless networks using
matching theory,” IEEE Trans. on Information Forensics and Security, vol. 8, no. 5, pp. 717–732, May 2013.
[36] W. Saad, Z. Han, T. Başar, M. Debbah, and A. Hjørungnes, “Distributed coalition formation games for secure wireless
transmission,” Mobile Networks and Applications, vol. 16, no. 2, pp. 231–245, 2011.
[37] N. Reyhanian, B. Maham, V. Shah-Mansouri, W. Tusher, and C. Yuen, “Game-theoretic approaches for energy cooperation
in energy harvesting small cell networks,” 2017, IEEE Trans. on Veh. Technology, to appear.
[38] H. Chen, Y. Li, Y. Jiang, Y. Ma, and B. Vucetic, “Distributed power splitting for SWIPT in relay interference channels
using game theory,” IEEE Trans. on Wireless Commun., vol. 14, no. 1, pp. 410–420, Jan. 2015.
[39] D. Niyato, E. Hossain, M. M. Rashid, and V. K. Bhargava, “Wireless sensor networks with energy harvesting technologies:
a game-theoretic approach to optimal energy management,” IEEE Wireless Commun., vol. 14, no. 4, pp. 90–96, Aug. 2007.
[40] M. Liu and Y. Liu, “Power allocation for secure SWIPT systems with wireless-powered cooperative jamming,” IEEE
Commun. Letters, to appear.
[41] L. Liu, R. Zhang, and K. C. Chua, “Secrecy wireless information and power transfer with MISO beamforming,” IEEE
Trans. on Signal Processing, vol. 62, no. 7, pp. 1850–1863, April 2014.
[42] Y. Ren, T. Lv, H. Gao, and Y. Li, “Secure wireless information and power transfer in heterogeneous networks,” 2017,
IEEE Access, to appear.
[43] Y. Zhou, Z. Z. Xiang, Y. Zhu, and Z. Xue, “Application of full-duplex wireless technique into secure MIMO communication:
Achievable secrecy rate based optimization,” IEEE Signal Processing Letters, vol. 21, no. 7, pp. 804–808, July 2014.
30
[44] A. Thangaraj, S. Dihidar, A. R. Calderbank, S. W. McLaughlin, and J. M. Merolla, “Applications of LDPC codes to the
wiretap channel,” IEEE Trans. on Info. Theory, vol. 53, no. 8, pp. 2933–2945, Aug. 2007.
[45] Z. Yazdanshenasan, H. S. Dhillon, M. Afshang, and P. H. J. Chong, “Poisson hole process: Theory and applications to
wireless networks,” IEEE Trans. on Wireless Commun., vol. 15, no. 11, pp. 7531–7546, Nov. 2016.
[46] C. h. Lee and M. Haenggi, “Interference and outage in Poisson cognitive networks,” IEEE Trans. on Wireless Commun.,
vol. 11, no. 4, pp. 1392–1401, April 2012.
[47] S. Lee, R. Zhang, and K. Huang, “Opportunistic wireless energy harvesting in cognitive radio networks,” IEEE Trans. on
Wireless Commun., vol. 12, no. 9, pp. 4788–4799, Sept. 2013.
[48] W. H. Sandholm, Population games and evolutionary dynamics.
MIT press, 2010.
[49] A. Cortes and S. Martinez, “Self-triggered best-response dynamics for continuous games,” IEEE Trans. on Automatic
Control, vol. 60, no. 4, pp. 1115–1120, April 2015.
[50] E. Barron, R. Goebel, and R. Jensen, “Best response dynamics for continuous games,” Proceedings of the American
Mathematical Society, vol. 138, no. 3, pp. 1069–1083, 2010.
[51] S. Krishnan and H. S. Dhillon, “Spatio-temporal interference correlation and joint coverage in cellular networks,” available
online: arxiv.org/abs/1606.05332.
[52] M. A. Kishk and H. S. Dhillon, “Tight lower bounds on the contact distance distribution in Poisson hole process,” IEEE
Wireless Commun. Letters, to appear.
| 7cs.IT
|
Nearly Optimal Robust Subspace Tracking
Praneeth Narayanamurthy and Namrata Vaswani
{pkurpadn, namrata}@iastate.edu
Department of Electrical and Computer Engineering,
Iowa State University, Ames, IA ∗
arXiv:1712.06061v3 [cs.IT] 25 Mar 2018
March 28, 2018
Abstract
In this work, we study the robust subspace tracking (RST) problem and obtain one of the first
two provable guarantees for it. The goal of RST is to track sequentially arriving data vectors that
lie in a slowly changing low-dimensional subspace, while being robust to corruption by additive sparse
outliers. It can also be interpreted as a dynamic (time-varying) extension of robust PCA (RPCA), with
the minor difference that RST also requires a short tracking delay. We develop a recursive projected
compressive sensing algorithm that we call Nearly Optimal RST via ReProCS (ReProCS-NORST)
because its tracking delay is nearly optimal. We prove that NORST solves both the RST and the
dynamic RPCA problems under weakened standard RPCA assumptions, two simple extra assumptions
(slow subspace change and most outlier magnitudes lower bounded), and a few minor assumptions.
Our guarantee shows that NORST enjoys a near optimal tracking delay of O(r log n log(1/)). Its
required delay between subspace change times is the same, and its memory complexity is n times
this value. Thus both these are also nearly optimal. Here n is the ambient space dimension, r is
the subspaces’ dimension, and is the tracking accuracy. NORST also has the best outlier tolerance
compared with all previous RPCA or RST methods, both theoretically and empirically (including for
real videos), without requiring any model on how the outlier support is generated. This is possible
because of the extra assumptions it uses.
1
Introduction
Principal Components Analysis (PCA) is one of the most widely used dimension reduction techniques.
It finds a small number of orthogonal basis vectors, called principal components, along which most of
the variability of the dataset lies. According to its modern definition [3], robust PCA (RPCA) is the
problem of decomposing a given data matrix into the sum of a low-rank matrix (true data) and a sparse
matrix (outliers). The column space of the low-rank matrix then gives the desired principal subspace (PCA
solution). A common application of RPCA is in video analytics in separating a video into a slow-changing
background image sequence (modeled as a low-rank matrix) and a foreground image sequence consisting
of moving objects or people (sparse) [3]. Robust Subspace Tracking (RST) can be simply interpreted as
a time-varying extension of RPCA. It assumes that the true data lies in a low-dimensional subspace that
can change with time, albeit slowly. The goal is to track this changing subspace over time in the presence
of additive sparse outliers. The offline version of this problem can be called dynamic (or time-varying)
RPCA. RST requires the tracking delay to be small, while dynamic RPCA does not. Time-varying
subspace is a more appropriate model for long data sequences, e.g., long surveillance videos, since if a
∗
A shorter version of this manuscript [1] is submitted to a double-blind conference. Another small part, Corollary 5.18,
will appear in [2].
1
single subspace model is used the resulting matrix may not be sufficiently low-rank. Moreover the RST
problem setting (short tracking delay) is most relevant for applications where real-time or near real-time
estimates are needed, e.g., video-based surveillance (object tracking) [4], monitoring seismological activity
[5], or detection of anomalous behavior in dynamic social networks [6].
In recent years, RPCA has since been extensively studied. Many fast and provably correct approaches
now exist: PCP introduced in [3] and studied in [3, 7, 8], AltProj [9], RPCA-GD [10] and NO-RMC
[11]. There is much lesser work on provable dynamic RPCA and RST: original-ReProCS [12, 13, 14]
for dynamic RPCA and simple-ReProCS [15] for both. The subspace tracking (ST) problem (without
outliers), and with or without missing data, has been studied for much longer in [16, 17, 18, 19]. However,
all existing guarantees for it only consider the statistically stationary setting of data being generated from
a single unknown subspace. Of course, the most general nonstationary model that allows the subspace
to change at each time is not even identifiable since at least r data points are needed to compute an
r-dimensional subspace even in the no noise or missing entries case.
In this work, we make the subspace tracking problem identifiable by assuming a piecewise constant
model on subspace change. We show that it is possible to track the changing subspace to within accuracy
as long as the subspace remains constant for at least O(r log n log(1/)) time instants, and some other
assumptions hold. This is more than r by only log factors. Here n is the ambient space dimension.
Notation. We use the interval notation [a, b] to refer to all integers between a and b, inclusive, and
we use [a, b) := [a, b − 1]. k.k denotes the l2 norm for vectors and induced l2 norm for matrices unless
specified otherwise, and 0 denotes transpose. We use MT to denote a sub-matrix of M formed by its
columns indexed by entries in the set T . For a matrix P we use P (i) to denote its i-th row. In our
algorithm statements, we use L̂t;α := [`ˆt−α+1 , · · · , `ˆt ] and SV Dr [M ] to refer to the matrix of top of r left
singular vectors of the matrix M . A matrix P with mutually orthonormal columns is referred to as a
basis matrix and is used to represent the subspace spanned by its columns. For basis matrices P1 , P2 , we
use SE(P1 , P2 ) := k(I − P1 P1 0 )P2 k as a measure of Subspace Error (distance) between their respective
subspaces. This is equal to the sine of the largest principal angle between the subspaces. It is also called
“projection distance” [20]. If P1 and P2 are of the same dimension, SE(P1 , P2 ) = SE(P2 , P1 ).
We reuse the letters C, c to denote different numerical constants in each use.
Robust Subspace Tracking (RST) and Dynamic RPCA Problem Setting. At each time t,
we get a data vector yt ∈ Rn that satisfies
yt := `t + xt + νt , for t = 1, 2, . . . , d
where νt is small unstructured noise, xt is the sparse outlier vector, and `t is the true data vector that
lies in a fixed or slowly changing low-dimensional subspace of Rn , i.e., `t = P(t) at where P(t) is an n × r
basis matrix with r n and with k(I − P(t−1) P(t−1) 0 )P(t) k small compared to kP(t) k = 1. We use Tt to
denote the support set of xt . Given an initial subspace estimate, P̂0 , the goal is to track span(P(t) ) and
`t either immediately or within a short delay. A by-product is that `t , xt , and Tt can also be tracked
on-the-fly. The initial subspace estimate, P̂0 , can be computed by applying any of the existing RPCA
solutions, e.g., PCP or AltProj, for the first roughly r data points, i.e., for Y[1,ttrain ] , with ttrain = Cr.
Dynamic RPCA is the offline version of the above problem. Define matrices L, X, W , Y with L =
[`1 , `2 , . . . `d ] and Y , X, W similarly defined. The goal is to recover the matrix L and its column space
with error. We use rL to denote the rank of L. The maximum fraction of nonzeros in any row (column)
of the outlier matrix X is denoted by max-outlier-frac-row (max-outlier-frac-col).
Identifiability and other assumptions. The above problem definition does not ensure identifiability since either of L or X can be both low-rank and sparse. Moreover, if the subspace changes at every
time, it is impossible to correctly estimate all the subspaces. One way to ensure that L is not sparse is
by requiring that its left and right singular vectors are dense (non-sparse) or “incoherent” w.r.t. a sparse
vector [3, 8, 9].
2
Definition 1.1. An n × r basis matrix P is µ-incoherent if maxi=1,2,..,n kP (i) k22 ≤ µr/n. Here µ is called
the coherence parameter. It quantifies the non-denseness of P .
A simple way to ensure that X is not low-rank is by imposing upper bounds on max-outlier-frac-row
and max-outlier-frac-col [8, 9]. One way to ensure identifiability of the changing subspaces is to assume
that they are piecewise constant:
P(t) = Pj for all t ∈ [tj , tj+1 ), j = 1, 2, . . . , J,
and to lower bound tj+1 − tj . Let t0 = 1 and tJ+1 = d. With this model, rL = rJ in general (except if
subspace directions are repeated). The union of the column spans of all the Pj ’s is equal to the span of
the left singular vectors of L. Thus, assuming that the Pj ’s are µ-incoherent implies their incoherence.
We also assume that the subspace coefficients at are mutually independent over time, have identical and
diagonal covariance matrices denoted by Λ, and are element-wise bounded. Element-wise bounded-ness
of at ’s, along with the statistical assumptions, is similar to incoherence of right singular vectors of L
(right incoherence); see Remark 3.5. Because tracking requires an online algorithm that processes data
vectors one at a time or in mini-batches, we need these statistical assumptions on the at ’s. For the same
reason, we also need to re-define max-outlier-frac-row as the maximum fraction of nonzeroes in any row of
any α-consecutive-column sub-matrix of X. Here α is the mini-batch size used by the RST algorithm.
We will refer to it as max-outlier-frac-rowα to indicate this difference.
Contributions. (1) We develop a recursive projected compressive sensing (ReProCS) algorithm
for RST that we call Nearly Optimal RST via ReProCS (ReProCS-NORST) because its tracking delay
is nearly optimal. We will refer to it as just “NORST” in the sequel. NORST has a significantly
improved (and simpler) subspace update step compared to all previous ReProCS-based methods. Our
most important contribution is one of the first two provable guarantees for RST, and the first that ensures
near optimal tracking delay, needs a near optimal lower bound on how long a subspace should remain
constant, and needs minimal assumptions on subspace change. Moreover, our guarantee also shows that
NORST is online (after initialization), fast (has the same complexity as vanilla r-SVD), and has memory
complexity of order nr log n log(1/). Here “online” means the following. After each subspace change, the
algorithm detects the change in at most 2α = Cf 2 r log n time instants, and after that, it improves the
subspace estimate every α time instants1 . The improvement after each step is exponential and thus, one
can get an -accurate estimate within K = C log(1/) such steps. Offline NORST also provably solves
dynamic RPCA.
(2) Our guarantees for both NORST and offline NORST (Theorem 2.2) essentially hold under “weakened” standard RPCA assumptions and two simple extra assumptions: (i) slow subspace change and (ii)
a lower bound on most outlier magnitudes. (i) is a natural assumption for static camera videos (with
no sudden scene changes) and (ii) is also easy because, by definition, an “outlier” is a large magnitude
corruption. The small magnitude ones get classified as νt . Besides these, we also need that at ’s are
mutually independent, have identical and diagonal covariance matrix Λ, and are element-wise bounded.
Element-wise bounded-ness, along with the statistical assumptions on at , is similar to right incoherence of
L; see Remark 3.5. For the initial Cr samples, NORST needs the outlier fractions to be O(1/r) (needed
to apply AltProj). As explained in Sec. 2.3, the extra assumptions help ensure that, after initialization,
NORST can tolerate a constant maximum fraction of outliers per row in any α-column-sub-matrix of
the data matrix without assuming any outlier support generation model. This statement assumes that
the condition number of the covariance of at is a constant (with n). As is evident from Table 1, this
1
The reason that just O(r log n) samples suffice for each update is because we assume that the at ’s are element-wise
bounded, νt is very small and with effective dimension r or smaller (see Theorem 2.2). These, along with the specific
structure of the PCA problem we encounter (noise/error seen by the PCA step depends on the `t ’s and thus has “effective
dimension” r), is why so few samples suffice.
3
Video Frame
NORST (16.5ms)AltProj(26.0ms)RPCA-GD(29.5ms)GRASTA (2.5ms) PCP (44.6ms)
Figure 1: Background Recovery. NORST gives the best background estimate. All the other algorithms exhibit
some artifacts. Only NORST does not contain the person or even his shadow. NORST is also faster than all except
GRASTA (which does not work). The GRASTA output also slightly lags the actual frame. Time taken per frame
is shown in parentheses.
is better than what all existing RPCA approaches can tolerate. For the video application, this implies
that NORST tolerates slow moving and occasionally static foreground objects much better than all other
approaches. This is also corroborated by our experiments on real videos, e.g., see Fig 1 and Sec. 6.
(3) Unlike simple-ReProCS [15] or original-ReProCS [12, 13, 14], NORST needs only a coarse initialization which can be computed using just C log r iterations of any batch RPCA method such as AltProj
applied to Cr initial samples. In fact, if the outlier magnitudes were very large for an initial set of
O(r log n log r) time instants, or if the outliers were absent for this much time, even a random initialization would suffice. This simple fact has two important implications. First, NORST with the subspace
change detection step removed also provides an online, fast, memory-efficient, and provably correct approach for static RPCA (our problem with J = 1, i.e., with `t = P at ). The other online solution for such
a problem is ORPCA which comes with only a partial guarantee [21] (the guarantee requires intermediate
algorithm estimates to be satisfying certain properties). Moreover, a direct corollary of our result is a
guarantee that a minor modification of NORST-random (NORST with random initialization) also solves
the subspace tracking with missing data (ST-missing) and the dynamic matrix completion (MC) problems. All existing guarantees for ST-missing [18, 19] hold only for the case of a single unknown subspace
and are only partial guarantees. From the MC perspective, NORST-random does not assume any model
on the set of observed entries. However, the tradeoff is that it needs many more observed entries. Both
these results are given in Sec. 5.
Paper Organization. The rest of the paper is organized as follows. In Sec. 2 we explain the main
ideas of the NORST algorithm and present our main result for it (Theorem 2.2). We also discuss the
implications of our guarantee, and provide detailed comparison with related work. In Sec. 3, we give the
complete NORST algorithm and carefully explain the subspace change detection approach. In Sec. 4, we
give the proof outline, the three main lemmas leading to the proof of Theorem 2.2, then also prove the
lemmas. In Sec. 5, we provide useful corollaries for (a) Static RPCA, (b) Subspace Tracking with missing
entries and (c) a simple extension to recover the guarantee of s-ReProCS from [15]. Empirical evaluation
on synthetic and real-world datasets is described in Sec. 6. The complete proof of Theorem 2.2, of two
auxiliary lemmas, and of the extensions is given in the Appendix.
2
NORST Algorithm and Main Result
2.1
NORST: Nearly-Optimal RST
NORST starts with a “good” estimate of the initial subspace. This can be obtained by C log r iterations2
of AltProj applied to Y[1,ttrain ] with ttrain = Cr. It then iterates between (a) Projected Compressive
2
√
Using C log r iterations helps ensure that the initialization error is O(1/ r).
4
Sensing (CS) / Robust Regression3 in order to estimate the sparse outliers, xt ’s, and hence the `t ’s,
and (b) Subspace Update to update the subspace estimate P̂(t) . Projected CS proceeds as follows. At
time t, if the previous subspace estimate, P̂(t−1) , is accurate enough, because of slow subspace change,
projecting yt onto its orthogonal complement will nullify most of `t . We compute ỹt := Ψyt where
Ψ := I − P̂(t−1) P̂(t−1) 0 . Thus, ỹt = Ψxt + Ψ(`t + νt ) and kΨ(`t + νt )k is small due to slow subspace
change and small νt . Recovering xt from ỹt is now a CS / sparse recovery problem in small noise [22]. We
compute x̂t,cs using noisy l1 minimization followed by thresholding based support estimation to obtain T̂t .
A Least Squares (LS) based debiasing step on T̂t returns the final x̂t . We then estimate `t as `ˆt = yt − x̂t .
The `ˆt ’s are then used for the Subspace Update step which involves (i) detecting subspace change, and
(ii) obtaining improved estimates of the new subspace by K steps of r-SVD, each done with a new set of
α samples of `ˆt . While this step is designed under the piecewise constant subspace assumption (needed
for identifiability of P(t) ’s), if the goal is only to get good estimates of `t or xt , the method works even
when this assumption may not hold, e.g., for real videos. For ease of understanding, we present a basic
version of NORST in Algorithm 1. This assumes the change times tj are known. The actual algorithm,
that we study and implement, detects these automatically. It is given as Algorithm 2 in Sec. 3.
2.2
Main Result
Before stating the result, we precisely define max-outlier-frac-col and max-outlier-frac-rowα . Since NORST
is an online approach that performs outlier support recovery one data vector at a time, it needs different
bounds on both. Let max-outlier-frac-col := maxt |Tt |/n. We define max-outlier-frac-rowα as the maximum
fraction of outliers (nonzeros) per row of any sub-matrix of X with α consecutive columns. To understand
P
this precisely, for a time interval, J , define γ(J ) := maxi=1,2,...,n |J1 | t∈J 1{i∈Tt } where 1S is the indicator
P
function for statement S. Thus, t∈J 1{i∈Tt } counts the number of outliers (nonzeros) in row i of XJ ,
and so γ(J ) is the maximum outlier fraction in any row of the sub-matrix XJ of X. Let J α denote a
time interval of duration α. Then max-outlier-frac-rowα := maxJ α ⊆[1,d] γ(J α ).
We use t̂j to denote the time instant at which the j-th subspace change time is detected by Algorithm
2.
Theorem 2.2. Consider Algorithm 2 given in the next section. Let α := Cf 2 r log n, Λ := E[a1 a1 0 ],
λ+ := λmax (Λ), λ− := λmin (Λ), f := λ+ /λ− , and let xmin := mint mini∈Tt (xt )i denote the minimum
outlier magnitude. Pick an ε ≤ min(0.01, 0.03 minj SE(Pj−1 , Pj )2 /f ). Let K := C log(1/ε). If
1. Pj ’s are µ-incoherent; and at ’s are zero mean, mutually independent over time t, have identical
covariance matrices, i.e. E[at at 0 ] = Λ, are element-wise uncorrelated (Λ is diagonal), are elementwise bounded (for a numerical constant η, (at )2i ≤ ηλi (Λ)), and are independent of all outlier
supports Tt ;
2. kνt k2 ≤ crkE[νt νt 0 ]k, kE[νt νt 0 ]k ≤ cε2 λ− , νt ’s are zero mean, mutually independent, and independent of xt , `t ;
3. max-outlier-frac-col ≤ c1 /µr, max-outlier-frac-rowα ≤ b0 :=
c2
;
f2
4. subspace change: let ∆ := maxj SE(Pj−1 , Pj ), assume that
3
Robust Regression (with a sparsity model on the outliers) assumes that observed data vector y satisfies y = P̂ a + x + b
where P̂ is a tall matrix (given), a is the vector of (unknown) regression coefficients, x is the (unknown) sparse outliers, b is
(unknown) small noise/modeling error. An obvious way to solve this is by solving mina,x λkxk1 + ky − P̂ a − xk2 . In this,
one can solve for a in closed form to get â = P̂ 0 (y − x). Substituting this, the minimization simplifies to minx λkxk1 + k(I −
P̂ P̂ 0 )(y − x)k2 . This is equivalent to the Lagrangian version of the projected CS problem that NORST solves (see line 7 of
Algorithm 1).
5
(a) tj+1 − tj > (K + 2)α, and
√
(b) ∆ ≤ 0.8 and C1 rλ+ (∆ + 2ε) ≤ xmin ;
√
5. initialization4 : SE(P̂0 , P0 ) ≤ 0.25, C1 rλ+ SE(P̂0 , P0 ) ≤ xmin ;
and (6) algorithm parameters are set as given in Algorithm 2; then, with probability (w.p.) at least
1 − 10dn−10 ,
if t ∈ [tj , t̂j + α),
(ε + ∆)
k−1
SE(P̂(t) , P(t) ) ≤ (0.3) (ε + ∆) if t ∈ [t̂j + (k − 1)α, t̂j + kα),
ε
if t ∈ [t̂j + Kα + α, tj+1 ).
Treating f as a numerical constant, the memory complexity is O(nα) = O(nr log n) and time complexity
is O(ndr log(1/ε)).
Corollary 2.3. Under Theorem 2.2 assumptions, the following also hold:
1. kx̂t − xt k = k`ˆt − `t k ≤ 1.2(SE(P̂(t) , P(t) ) + ε)k`t k with SE(P̂(t) , P(t) ) bounded as above,
2. at all times, t, T̂t = Tt ,
3. tj ≤ t̂j ≤ tj + 2α,
of f line
f line
f line
4. Offline-NORST (last few lines of Algorithm 2): SE(P̂(t)
, P(t) ) ≤ ε, kx̂of
− xt k = k`ˆof
−
t
t
`t k ≤ εk`t k at all t. Its memory complexity is O(Knα) = O(nr log n log(1/ε)).
Remark 2.4 (Relaxing outlier magnitudes lower bound). The assumption on xmin (outlier magnitudes)
required by Theorem 2.2 can be significantly relaxed to the following which only requires that most outlier
magnitudes are lower bounded. Assume that the outlier magnitudes are such that the following holds: xt
can be split as xt = (xt )small + (xt )large with √
the two components being such that, in the k-th subspace
5
k−1
+
update interval , k(xt )small
√ k ≤ 0.3 (ε + ∆) rλ and the smallest nonzero entry of (xt )large is larger
k−1
than C1 · 0.3 (ε + ∆) rλ+ . For the case of j = 0, we need the bound to hold with ∆ replaced by
∆init = SE(P̂0 , P0 ), and ε replaced by zero.
If there were a way to bound the element-wise error of the CS step (instead of the l2 norm error), one
could relax the above requirement even more.
Discussion.
This discussion assumes that f is a constant (does not increase with n), i.e., it is
O(1). Theorem 2.2 shows that, with high probability (whp), when using NORST, the subspace change
gets detected within a delay of at most 2α = Cf 2 (r log n) time instants, and the subspace gets estimated
to ε error within at most (K + 2)α = Cf 2 (r log n) log(1/ε) time instants. The same is also true for the
recovery error of xt and `t . Both the detection and tracking delay are within log factors of the optimal
since r is the minimum delay needed even in the noise-free, i..e, xt = νt = 0, case. The fact that NORST
can detect subspace change within a short delay can be an important feature for certain applications,
e.g., this feature is used in [6] to detect structural changes in a dynamic social network. Moreover, if
offline processing is allowed, we can guarantee recovery within normalized error ε at all time instants.
This implies that offline-NORST solves the dynamic RPCA problem.
Observe that Theorem 2.2 allows a constant maximum fraction of outliers per row (after initialization),
without making any assumption on how the outlier support is generated, as long as the extra assumptions
4
This can be satisfied by applying C log r iterations of AltProj [9] on the first Cr data samples and assuming that these
have outlier fractions in any row or column bounded by c/r.
5
k-th subspace update interval refers to Jk := [t̂j + (k − 1)α, t̂j + kα) for k > 1 and J1 = [tj , t̂j + α) for k = 1. The first
interval also includes the subspace detection interval, [tj , t̂j ), since the analysis of the projected CS step for this interval is
the same as for [t̂j , t̂j + α).
6
Algorithm 1 Basic-NORST (with tj known). The actual algorithm that detects tj automatically is Algorithm 2.
Obtain P̂0 by C(log r) iterations of AltProj on Y[1,ttrain ] with ttrain = Cr followed by SVD on the output L̂.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
Input: yt , Output: x̂t , `ˆt , P̂(t) , T̂t
Parameters: K ← C log(1/ε), α ← Cf 2 r log n, ωsupp ← xmin /2, ξ ← xmin /15, r
Initialize: j ← 1, k ← 1 P̂(ttrain ) ← P̂0
for t > ttrain do
Ψ ← I − P̂(t−1) P̂(t−1) 0
ỹt ← Ψyt .
Projected-CS
x̂t,cs ← arg minx̃ kx̃k1 s.t. kỹt − Ψx̃k ≤ ξ.
(Robust
T̂t ← {i : |x̂t,cs | > ωsupp }.
Regression).
x̂t ← IT̂t (ΨT̂t 0 ΨT̂t )−1 ΨT̂t 0 ỹt .
`ˆt ← yt − x̂t .
if t = tj + kα − 1 then
P̂j,k ← SV Dr [L̂t;α ], P̂(t) ← P̂j,k , k ← k + 1.
else
P̂(t) ← P̂(t−1) .
Subspace
end if
Update.
if t = tj + Kα − 1 then
P̂j ← P̂(t) , k ← 1, j ← j + 1
end if
end for
discussed below hold. We explain why this is possible in Sec. 2.3. Of course, for the initial Cr samples,
NORST needs max-outlier-frac-rowCr ∈ O(1/r) (needed to apply AltProj). Also, the memory complexity
guaranteed by Theorem 2.2 is nearly d/r times better than that of all existing RPCA solutions; see Table
1. The time complexity is worse than that of only NO-RMC6 , but NO-RMC needs d ≥ cn (unreasonable
requirement for videos which often have much fewer frames d than the image size n). Finally, NORST also
needs outlier fraction per column to be O(1/r) instead of O(1/rL ). If J is large, e.g. if J = d/(r log n), it
is possible that rL r.
We should clarify that NORST allows max-outlier-frac-rowα ∈ O(1) but this does not necessarily imply
that the number of outliers in each row can be this high. The reason is it only allows the fraction per
column to only be O(1/r). Thus, for a matrix of size n × α, it allows the total number of outliers to be
O(min(nα, nα/r)) = O(nα/r). Thus the average fraction allowed is only O(1/r).
NORST needs the following extra assumptions. The main extra requirement is that xmin be lower
bounded as given in the last two assumptions of Theorem 2.2, or as stated in Remark 2.4. The lower bound
on xmin is reasonable7 as long as the initial subspace estimate is accurate enough and the subspace changes
√
slowly enough so that both ∆ and SE(P̂0 , P0 ) are O(1/ r). This requirement may seem restrictive on
first glance but actually is not. The reason is that SE(.) is only measuring the largest principal angle.
This bound on SE still allows the chordal distance between the two subspaces to be O(1). Chordal
distance [20] is the l2 norm of the vector containing the sine of all principal angles. The second related
extra requirement is an upper bound on ∆ (slow subspace change) which depends on the value of xmin .
We discuss this point next. Other than these two, NORST only needs simple statistical assumptions on
6
NO-RMC is so fast because it is actually a robust matrix completion solution and it deliberately undersamples the entire
data matrix Y to get a faster
√ RPCA algorithm.
7
requires xmin to be C λ+ or larger.
7
Table 1: Comparing RPCA and RST solutions. All algorithms also require left and right incoherence or left
incoherence and at ’s element-wise bounded (this, along with the i.i.d. assumption on at ’s, is similar to right
incoherence), and hence these are not compared. The incoherence parameter µ and the condition numbers are
treated as constants in this table. In general rL = rJ. Strong or unrealistic assumptions are shown in red.
PCP(C) [3], PCP(H) [7, 8], mod-PCP [23], AltProj [9], RPCA-GD [10], NO-RMC [11], orig-ReProCS [13, 14],
s-ReProCS [15].
Algorithm
Outlier tolerance
Other Assumptions
Memory, Time,
# params
PCP(C) [3]
max-outlier-frac-row ∈ O(1)
outlier support: unif. random,
Memory: O(nd)
zero
max-outlier-frac-col ∈ O(1)
rL ≤ c min(n, d)/log2 n
Time: O(nd2 1 )
max-outlier-frac-row ∈ O (1/rL )
d ≥ crL
Memory: O(nd)
PCP(H) [8]
AltProj [9]
max-outlier-frac-row ∈ O (1/rL )
d ≥ crL
Memory: O(nd)
max-outlier-frac-row ∈ O(1/rL1.5 )
d ≥ crL
Memory: O(nd)
max-outlier-frac-row ∈ O (1/rL )
c2 n ≥ d ≥ cn
Memory: O(nd)
max-outlier-frac-row ∈ O(1)
outlier support: unif. random,
Memory: O(nr log2 n)
max-outlier-frac-col ∈ O(1)
slow subspace change (unrealistic),
Time: O(ndr log2 n 1 )
rL ≤
orig-ReProCS [13, 14]
max-outlier-frac-rowα
tracking delay
max-outlier-frac-col ∈ O(1/rL )
∈ O(1)
too large
4
Time: O(nrL3 log2 n log2 1 )
max-outlier-frac-col ∈ O(1/rL )
mod-PCP [23]
5
Time: O(ndrL log 1 )
max-outlier-frac-col ∈ O(1/rL1.5 )
NO-RMC [11]
2
Time: O(ndrL2 log 1 )
max-outlier-frac-col ∈ O (1/rL )
RPCA-GD [10]
1
Time: O(nd2 1 )
max-outlier-frac-col ∈ O (1/rL )
c min(n,d)
log2 n
Delay: ∞
outlier support: moving object model,
Memory: O(nr2 /2 )
unrealistic subspace change model,
Time: O(ndr log 1 )
5
changed eigenvalues small for some time,
outlier mag. lower bounded,
init data: AltProj assu’s,
at ’s independent, d ≥ Cr2 /2
s-ReProCS: [15]
max-outlier-frac-rowα
solves
max-outlier-frac-col ∈ O(1/r)
∈ O(1)
subspace change: only 1 direc at a time,
Memory: O(nr log n log 1 )
outlier mag. lower bounded,
Time: O(ndr log 1 )
4
at ’s i.i.d., d ≥ Cr log n log 1 .
RST with
sub-optimal delay
init data: AltProj assumptions
NORST
max-outlier-frac-rowα ∈ O(1)
subspace change: mild,
Memory: O(nr log n log 1 )
(this work): solves
max-outlier-frac-col ∈ O(1/r)
outlier mag. lower bounded,
Time: O(ndr log 1 )
RST with
at ’s i.i.d., d ≥ Cr log n log 1 ,
near-optimal delay
first Cr samples: AltProj assumptions
4
at ’s. The zero-mean assumption is a minor one. The assumption that Λ be diagonal is also minor8 .
In the video setting, zero-mean can be ensured by subtracting the empirical average of the background
images computing using the first ttrain frames. Mutual independence of at ’s holds if the changes in each
background image w.r.t. a “mean” background are independent, when conditioned on their subspace.
This is valid, for example, if the background changes are due to illumination variations or due to moving
curtains (see Fig. 5). Moreover, by using the approach of [14], it is possible to relax this to just requiring
that the at ’s satisfy an autoregressive model over time. Element-wise boundedness, along with the above,
is similar to right incoherence (see Remark 3.5).
Outlier v/s Subspace Assumptions. When there are fewer outliers in the data or when outliers
are easy to detect, one would expect to need weaker assumptions on the true data’s subspace and/or
on its rate of change. This is indeed true. The max-outlier-frac-col bound relates max-outlier-frac-col to µ
(not-denseness parameter) and r (subspace dimension). The upper bound on ∆ implies that, if xmin is
8
It only implies that Pj is the matrix of principal components of E[Lj L0j ] where Lj := [`tj , `tj +1 , . . . , `tj+1 −1 ].
8
larger (outliers are easier to detect), a larger amount of subspace change ∆ can be tolerated. The relation
of max-outlier-frac-row to rate of subspace change is not evident from the way the guarantee is stated above
because we have assumed max-outlier-frac-row ≤ b0 := c/f 2 with c being a numerical constant, and used
c∆
this to get a simple expression for K. If we did not do this, we would get K = Cd − log(1√b f ) log( 0.8ε
)e,
0
see Remark A.1. Since we need tj+1 − tj ≥ (K + 2)α, a smaller b0 means a larger ∆ can be tolerated for
the same delay, or vice versa.
Algorithm Parameters. Algorithm 2 assumes knowledge of 4 model parameters: r, λ+ , λ− and
xmin to set the algorithm parameters. The initial dataset used for estimating P̂0 (using AltProj) can be
used to get an accurate estimate of r, λ− and λ+ using standard techniques. Thus one really only needs to
set xmin . If continuity over time is assumed, we can let it be time-varying and set it as mini∈T̂t−1 |(x̂t−1 )i |
at t.
Related Work.
For a summary of comparisons, see Table 1. In terms of other solutions for
provably correct RST or dynamic RPCA, there is very little work. For RST, there is only one other
provable algorithm, simple-ReProCS (s-ReProCS) [15]. This has the same tracking delay and memory
complexity as NORST, however, it assumes that only one subspace direction can change at each change
time. This is a more restrictive model than ours. Moreover, it implies that the tracking delay of s-ReProCS
is r-times sub-optimal. Also, s-ReProCS uses a projection-SVD step for subspace update (as opposed to
simple SVD in NORST). These two facts imply that it needs an -accurate subspace initialization in order
to ensure that the later changed subspaces can be tracked with -accuracy. Thus, it does not provide a
static RPCA or subspace tracking with missing data solution.
For dynamic RPCA, the earliest result was a partial guarantee (a guarantee that depended on intermediate algorithm estimates satisfying certain assumptions) for the original reprocs approach (originalReProCS) [12]. This was followed up by two complete guarantees for reprocs-based approaches with
minor modifications [13, 14]. For simplicity we will still call these “original-ReProCS”. These guarantees
needed very strong assumptions and their tracking delay was O(nr2 /2 ). Since can be very small, this
factor can be quite large, and hence one cannot claim that original-ReProCS solves RST. Our work is
a very significant improvement over all these works. (i) The guaranteed memory complexity, tracking
delay, and required delay between subspace change times of NORST are all r/2 times lower than that
of original-ReProCS. (ii) All the original-ReProCS guarantees needed a very specific assumption on how
the outlier support could change. They required an outlier support model inspired by a video moving
object that moves in one direction for a long time; and whenever it moves, it must move by a fraction of
s := maxt |Tt |. This is very specific model with the requirement of moving by a fraction of s being the
most restrictive. Our result removes this model and replaces it with just a bound on max-outlier-frac-row.
We explain in the last para of Sec. 4.1 why this is possible. (iii) The subspace change model assumed
in [13, 14] required a few new directions, that were orthogonal to Pj−1 , to be added at time tj and some
others to be removed. This is an unrealistic model for slow subspace change, e.g., in 3D, it implies that
the subspace needs to change from the x-y plane to the y-z plane. Moreover because of this model, their
results needed the “energy” (eigenvalues) along the newly added directions to be small for a period of
time after each subspace change. This is a strong (and not easy to interpret) requirement. Our result
removes all these requirements and replaces them with a bound on SE(Pj−1 , Pj ) which is much more
realistic. Thus, in 3D, our result allows the x-y plane to change to a slightly tilted x-y plane.
An approach called modified-PCP (mod-PCP) was proposed to solve the problem of RPCA with partial
subspace knowledge [23]. A corollary of its guarantee shows that it can also be used to solve dynamic
RPCA [23]. However, since it adapted the PCP proof techniques from [3], its pros and cons are similar
to those of PCP, e.g., it also needs a uniformly randomly generated outlier support. As can be seen from
Table 1, its pros and cons are similar to those of the PCP result by [3] (PCP(C)) discussed below.
We also provide a comparison with provably correct RPCA approaches in Table 1. In summary,
9
NORST has significantly better memory complexity than all of them, all of which are batch; it has the
best outlier tolerance (after initialization), and the second-best time complexity, as long as its extra
assumptions hold. It can also detect subspace change quickly, which can be a useful feature. Consider outlier tolerance. PCP(H), AltProj, RPCA-GD, and NO-RMC need both max-outlier-frac-row and
max-outlier-frac-col to be O(1/rL ); PCP(C) [3] and modified-PCP [23] need the outlier support to uniformly random (strong requirement: for video it implies that objects are very small sized and jumping
around randomly); and original-ReProCS needs it to satisfy a very specific moving object model described above (restrictive). Instead, after initialization, NORST only needs max-outlier-frac-rowα ∈ O(1)
and max-outlier-frac-col ∈ O(1/r).
2.3
The need for extra assumptions
As noted in [9], the standard RPCA problem (that only assumes left and right incoherence of L and
nothing else) cannot tolerate a bound on outlier fractions in any row or any column that is larger than
1/rL 9 . The reason NORST can tolerate a constant max-outlier-frac-rowα bound is because it uses extra
assumptions. We explain the need for these here. It recovers xt first and then `t and does this at each
time t. When recovering xt , it exploits “good” knowledge of the subspace of `t (either from initialization
or from the previous subspace’s estimate and slow subspace change), but it has no way to deal with the
residual error, bt := (I − P̂(t−1) P̂(t−1) 0 )`t , in this knowledge. Since the individual vector bt does not have
any structure that can exploited10 , the error in recovering xt cannot be lower than Ckbt k. This means
that, to correctly recover the support of xt , xmin needs to be larger than Ckbt k. This is where the xmin
lower bound comes from. As we will see in Sec. 4, correct support recovery is needed to ensure that the
subspace estimate can be improved with each update. In particular, it helps ensure that the error vectors
et := xt − x̂t in a given subspace update interval are mutually independent, when conditioned on the yt ’s
from all past intervals. This step also uses element-wise boundedness of the at ’s along with their mutual
independence and identical covariances.
3
Automatic NORST
We present the actual NORST algorithm (automatic NORST) in Algorithm 2. The main idea why
automatic NORST works is the same as that of the basic algorithm with the exception of the additional
subspace detection step. The subspace detection idea is borrowed from [15], although its correctness proof
has differences because we assume a much simpler subspace change model. In Algorithm 2, the subspace
update stage toggles between the “detect” phase and the “update” phase. It starts in the “update” phase
with t̂0 = ttrain . We then perform K r-SVD steps with the k-th one done at t = t̂0 + kα − 1. Each such
step uses the last α estimates, i.e., uses L̂t;α . Thus at t = t̂0 + Kα − 1, the subspace update of P0 is
complete. At this point, the algorithm enters the “detect” phase.
For any j, if the j-th subspace change is detected at time t, we set t̂j = t. At this time, the algorithm
enters the “update” (subspace update) phase. We then perform K r-SVD steps with the k-th r-SVD step
done at t = t̂j + kα − 1 (instead of at t = tj + kα − 1). Each such step uses the last α estimates, i.e., uses
9
The reason is this: let b0 = max-outlier-frac-row, one can construct a matrix X with b0 outliers in some rows that
has rank equal to 1/b0 . A simple way to do this would be to let the support and nonzero entries of X be constant for b0 d
columns before letting either of them change. Then the rank of X will be d/(b0 d). A similar argument can be used for
max-outlier-frac-col.
10
However the bt ’s arranged into a matrix do form a low-rank matrix whose approximate rank is r or even lower (if not
all directions change). If we try to exploit this structure we end up with a modified-PCP [23] type approach. This needs
the uniform random support assumption (used in its guarantee). Or, if the [8] approach were used for its guaratee, for
identifiability reasons similar to the one described above, it will still not tolerate outlier fractions larger than 1/rnew where
rnew is the (approximate) rank of the matrix formed by the bt ’s.
10
Algorithm 2 Automatic-NORST.
Obtain P̂0 by C(log r) iterations of AltProj on Y[1,ttrain ] with ttrain = Cr followed by SVD on the output
L̂.
1: Input: P̂0 , yt , Output: x̂t , `ˆt , P̂(t)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
Parameters: K ← C log(1/ε), α ← Cf 2 r log n, ωsupp ← xmin /2, ξ ← xmin /15, ωevals ← 2ε2 λ+ , r.
P̂(ttrain ) ← P̂0 ; j ← 1, k ← 1
phase ← update; t̂0 ← ttrain ;
for t > ttrain do
Lines 5 − 10 of Algorithm 1
Projected CS.
if phase = detect and t = t̂j−1,f in + uα then
Φ ← (I − P̂j−1 P̂j−1 0 ).
B ← ΦL̂t,α
Subspace
if λmax (BB 0 ) ≥ αωevals then
Detect Phase.
phase ← update, t̂j ← t,
end if
end if
if phase = update then
if t = t̂j + uα − 1 for u = 1, 2, · · · , then
Subspace
P̂j,k ← SV Dr [L̂t;α ], P̂(t) ← P̂j,k , k ← k + 1.
Update Phase.
else
P̂(t) ← P̂(t−1)
end if
if t = t̂j + Kα − 1 then
t̂j,f in ← t, P̂j ← P̂(t)
k ← 1, j ← j + 1, phase ← detect.
end if
end if
end for
Offline NORST: At t = t̂j + Kα, for all t ∈ [t̂j−1 + Kα, t̂j + Kα − 1],
offline ← [P̂
0
P̂(t)
j−1 , (I − P̂j−1 P̂j−1 )P̂j ]
offline P̂ offline 0
Offline
Ψ ← I − P̂(t)
(t)
NORST.
x̂offline ← I (Ψ 0 Ψ )−1 Ψ 0 y
t
T̂t
T̂t
T̂t
30: `ˆoffline
← yt − x̂offline
.
t
t
T̂t
t
L̂t;α Thus, at t = t̂j,f in = t̂j + Kα − 1, the update is complete. At this time, the algorithm enters the
“detect” phase again.
To understand the change detection strategy, consider the j-th subspace change. Assume that the
previous subspace Pj−1 has been accurately estimated by t = t̂j−1,f in = t̂j−1 + Kα − 1 and that t̂j−1,f in <
tj . Let P̂j−1 denote this estimate. At this time, the algorithm enters the “detect” phase in order to detect
the next (j-th) change. Let Bt := (I − P̂j−1 P̂j−1 0 )L̂t;α . For every t = t̂j−1,f in + uα − 1, u = 1, 2, . . . , we
√
detect change by checking if the maximum singular value of Bt is above a pre-set threshold, ωevals α, or
not.
We claim that, whp, under assumptions of Theorem 2.2, this strategy has no “false subspace detections” and correctly detects change within a delay of at most 2α samples. The former is true because,
for any t for which [t − α + 1, t] ⊆ [t̂j−1,f in , tj ), all singular values of the matrix Bt will be close to
11
√
√
zero (will be of order ε λ+ ) and hence its maximum singular value will be below ωevals α. Thus,
whp, t̂j ≥ tj . To understand why the change is correctly detected within 2α samples, first consider
t −t̂
in
t = t̂j−1,f in + d j j−1,f
eα := tj,∗ . Since we assumed that t̂j−1,f in < tj (the previous subspace update is
α
complete before the next change), tj lie in the interval [tj,∗ − α + 1, tj,∗ ]. Thus, not all of the `t ’s in this
interval lie in the new subspace. Depending on where in the interval tj lies, the algorithm may or may
not detect the change at this time. However, in the next interval, i.e., for t ∈ [tj,∗ + 1, tj,∗ + α], all of the
`t ’s lie in the new subspace. We can prove that Bt for this time t will have maximum singular value that
is above the threshold. Thus, if the change is not detected at tj,∗ , whp, it will get detected at tj,∗ + α.
Hence one can show that, whp, either t̂j = tj,∗ , or t̂j = tj,∗ + α, i.e., tj ≤ t̂j ≤ tj + 2α (see Appendix A).
√
Time complexity. Consider initialization. To ensure that SE(P̂0 , P0 ) ∈ O(1/ r), we need to use
C log r iterations of AltProj. Since there is no lower bound in the AltProj guarantee on the required
number of matrix columns (except the trivial lower bound of rank) [9], we can use ttrain = Cr frames for
√
initialization. Thus the initialization complexity is O(nttrain r2 log( r) = O(nr3 log r) [9]. The projectedCS step complexity is equal to the cost of a matrix vector multiplication with the measurement matrix
times negative logarithm of the desired accuracy in solving the l1 minimization problem. Since the
measurement matrix for the CS step is I − P̂(t−1) P̂(t−1) 0 , the cost per CS step (per frame) is O(nr log(1/))
[24] and so the total cost is O((d−ttrain )nr log(1/)). The subspace update involves at most ((d−ttrain )/α)
rank r-SVD’s on n×α matrices all of which have constant eigen-gap (this is proved in the proof of tTheorem
4.14 from [25] which we use to show correctness of this step). Thus the total time for subspace update
steps is at most ((d − ttrain )/α) ∗ O(nαr log(1/)) = O((d − ttrain )nr log(1/)) [26]. Thus the running
time of the complete algorithm is O(ndr log(1/) + nr3 log r). As long as r2 log r ≤ d log(1/), the time
complexity of the entire algorithm is O(ndr log(1/)).
Remark 3.5 (Relating our assumptions to right incoherence of Lj := L[tj ,tj+1 ) [8]). From our assumptions, Lj = Pj Aj with Aj := [atj , atj +1 , . . . atj+1 −1 ], the columns of Aj are zero mean, mutually independent, have identical covariance Λ, Λ is diagonal, and are element-wise bounded as specified by Theorem
P
2.2. Let dj := tj+1 − tj . Define a diagonal matrix Σ with (i, i)-th entry σi and with p
σi2 := t (at )2i /dj .
Define a dj × r matrix Ṽ with the t-th entry of the i-th column being (ṽi )t := (at )i /(σi dj ). Then, Lj =
Pj ΣṼ 0 and each column of Ṽ is unit 2-norm. Also, from the bounded-ness assumption, (ṽi )2t ≤ η σλ2i · d1j
i
where η is a numerical constant.
Observe that Pj ΣṼ 0 is not exactly the SVD of Lj since the columns of Ṽ are not necessarily exactly
mutually orthogonal. However, if dj is large enough, one can argue using any law of large numbers’ result
(e.g., Hoeffding inequality), that the columns of Ṽ are approximately mutually orthogonal whp. Also, whp,
σi2 ≥ 0.99λi . This also follows using Hoeffding11 . Thus, our assumptions imply that, whp, (ṽi )2t ≤ C/dj .
If one interprets Ṽ as an “approximation” to the right singular vectors of Lj , this is the right incoherence assumed by [8] and slightly stronger than what is assumed by [3, 9] and others (these require that
the squared norm of each row of the matrix of right singular vectors be bounded by Cr/dj ).
The claim that “Ṽ can be interpreted as an “approximation” to the right singular vectors of Lj ” is
not rigorous. But it is also not clear how to make it rigorous since our work uses statistical assumptions
SVD
on the at ’s. To get the exact SVD of Lj , we need the SVD of Aj . Suppose Aj = RΣV 0 , then
SVD
Lj = (Pj R)ΣV 0 . Here R will be an r × r orthonormal matrix. Now it is not clear how to relate the
element-wise bounded-ness assumption on at ’s to an assumption on entries of V , since now there is no
easy expression for each entry of V or of the entries of Σ in terms of at (since R is unknown).
11
The first claim uses all the four assumptions on at ; the second claim uses all assumptions except diagonal Λ
12
4
Proof Outline and (most of the) Proof
In this section we first give the main ideas of the proof (without formal lemmas). We then state the three
main lemmas and explain how they help prove Theorem 2.2. After this, we prove the three lemmas.
4.1
Main idea of the proof
It is not hard to see that the “noise” bt := Ψ(`t + νt ) seen by the projected CS step is proportional
the error between the subspace estimate from (t − 1) and the current subspace. Moreover, incoherence
(denseness) of the P(t) ’s and slow subspace change together imply that Ψ satisfies the restricted isometry
property (RIP) [12]. Using this, a result for noisy l1 minimization [22], and the lower bound assumption
on outlier magnitudes, one can ensure that the CS step output is accurate enough and the outlier support
Tt is correctly recovered. With this, we have that `ˆt = `t + νt − et where et := xt − x̂t satisfies
et = ITt (ΨTt 0 ΨTt )−1 ITt 0 Ψ0 `t and ket k ≤ Ckbt k. Consider subspace update. Every time the subspace
changes, one can show that the change can be detected within a short delay. After that, the K SVD
steps help get progressively improved estimates of the changed subspace. To understand this, observe
that, after a subspace change, but before the first update step, bt is the largest and hence, et , is also the
largest for this interval. However, because of good initialization or because of slow subspace change and
previous subspace correctly recovered (to error ε), neither is too large. Both are proportional to (ε + ∆),
or to the initialization error. Using the idea below, we can show that we get a “good” first estimate of
the changed subspace.
The input to the PCA step is `ˆt and the noise seen by it is et . Notice that et depends on the true
data `t . Hence this is a setting of PCA in data-dependent noise [27, 25]. From [25], it is known that
the subspace recovery error of the PCA step is proportional to the ratio between the time averaged noise
P
P
power plus time-averaged signal-noise correlation, (k t E[et et 0 ]k + k t E[`t et 0 k)/α, and the minimum
signal space eigenvalue, λ− . The instantaneous value of noise power is (∆ + ε)2 times λ+ while that
of signal-noise correlation is of order (∆ + ε) times λ+ . However, using the fact that et is sparse with
α is bounded, one can argue (using
support Tt that changes enough over time so that max-outlier-frac-row
√
α
Cauchy-Schwartz) that their time averaged values are max-outlier-frac-row
result,
√ times smaller. Asα a +
after the first subspace update, the subspace recovery error is at most 4 max-outlier-frac-row (λ /λ− )
times (∆ + ε). Since max-outlier-frac-rowα (λ+ /λ− )2 is bounded by a constant c2 < 1, this means that,
√
after the first subspace update, the subspace error is at most c2 times (∆ + ε).
√
This, in turn, implies that kbt k, and hence ket k, is also c2 times smaller in the second subspace
update interval compared to the first. This, along with repeating the above argument, helps show that
√
√
the second estimate of the changed subspace is c2 times better than the first and hence its error is ( c2 )2
√
times (∆ + ε). Repeating the argument K times, the K-th estimate has error ( c2 )K times (∆ + ε).
Since K = C log(1/ε), this is an ε accurate estimate of the changed subspace.
A careful application of the result of [25] is the reason why we are able to remove the moving object
model assumption on the outlier support needed by the earlier guarantees for original-ReProCS [13, 14].
P
Applied to our problem, this result requires k t∈J α ITt ITt 0 /αk to be bounded by a constant less than
P
one. It is not hard to see that maxJ α ∈[1,d] k t∈J α ITt ITt 0 /αk = max-outlier-frac-rowα . To understand
P
0
this simply, the matrix
t∈J α ITt ITt is diagonal, and the i-th diagonal entry counts the number of
time the index i appears in the support set Tt in the interval J α which is precisely the definition of
max-outlier-frac-rowα · α. This is also why a constant bound on max-outlier-frac-rowα suffices for our setting.
On the other hand the guarantees of [13, 14] required that, for any sequence of positive semi-definite
P
(p.s.d.) matrices, At , k t∈J α ITt At ITt 0 /αk, be bounded by a constant less than one. This is a much
more stringent requirement; one way to satisfy it is using the moving object model on outlier supports
assumed there.
13
4.2
Main Lemmas
For simplicity, we give the proof for the νt = 0 case. The changes with νt 6= 0 are minor, see Appendix
A.
First consider the simpler case when tj ’s are known, i.e., consider Algorithm 1. In this case, t̂j = tj .
Definition 4.6. Define
√
1. the constants used in Theorem 2.2: c1 = 0.01, c2 = 0.01, and C1 = 15 η
2. s := max-outlier-frac-col · n
3. φ+ = 1.2
4. bound on max-outlier-frac-rowα : b0 := 0.01/f 2 .
5. q0 := 1.2(ε + SE(Pj−1 , Pj )), qk = (0.3)k q0
6. et := x̂t − xt . Since νt = 0, et = `t − `ˆt
7. Events: Γ0,0 := {assumed bound on SE(P̂0 , P0 )},
Γ0,k := Γ0,k−1 ∩ {SE(P̂0,k , P0 ) ≤ 0.3k SE(P̂0 , P0 )},
Γj,0 := Γj−1,K , Γj,k := Γj,k−1 ∩ {SE(P̂j,k , Pj ) ≤ qk−1 /4} for j = 1, 2, . . . , J and k = 1, 2, . . . , K.
8. Using the expression for K given in the theorem, and since P̂j = P̂j,k (from the Algorithm), it
follows that Γj,K implies SE(P̂j , Pj ) = SE(P̂j,K , Pj ) ≤ ε.
Observe that, if we can show that Pr(ΓJ,K |Γ0,0 ) ≥ 1 − dn−10 we will have obtained all the subspace
recovery bounds of Theorem 2.2. The next two lemmas applied sequentially help show that this is true
for Algorithm 1 (tj known). The correctness of the actual algorithm (Algorithm 2) follows using these,
Corollary 4.10, and Lemma 4.12. The Theorem’s proof is in Appendix A.
Lemma 4.7 (first subspace update interval). Under the conditions of Theorem 2.2, conditioned on Γj,0 ,
p
1. for all t ∈ [t̂j , t̂j + α), kΨ`t k ≤ (ε + ∆) ηrλ+ < xmin /15, kx̂t,cs − xt k ≤ 7xmin /15 < xmin /2,
T̂t = Tt , and the error et := x̂t − xt = `t − `ˆt satisfies
−1
(1)
ITt 0 Ψ`t ,
et = ITt ΨTt 0 ΨTt
and ket k ≤ 1.2(ε + ∆)
p
ηrλ+ .
2. w.p. at least 1 − 10n−10 , the first subspace estimate P̂j,1 satisfies SE(P̂j,1 , Pj ) ≤ (q0 /4), i.e., Γj,1
holds.
Lemma 4.8 (k-th subspace update interval). Under the conditions of Theorem 2.2, conditioned on Γj,k−1 ,
1. for all t ∈ [t̂j p
+ (k − 1)α, t̂j + kα − 1), all claims of the
p first part of Lemma 4.7 holds, kΨ`t k ≤
0.3k−1 (ε + ∆) ηrλ+ , and ket k ≤ (0.3)k−1 · 1.2(ε + ∆) ηrλ+ .
2. w.p. at least 1 − 10n−10 the subspace estimate P̂j,k satisfies SE(P̂j,k , Pj ) ≤ (qk−1 /4), i.e., Γj,k holds.
Remark 4.9. For the case of j = 0, in both the lemmas above, ∆ gets replaced by SE(P̂0 , P0 ) and ε by
zero.
Corollary 4.10. Under the conditions of Theorem 2.2 the following hold
14
1. For all t ∈ [tj , t̂j ), conditioned on Γj−1,K , all claims of the first item of Lemma 4.7 hold.
2. For all t ∈ [t̂j + Kα, tj+1 ), conditioned on Γj,K , the first item of Lemma 4.8 holds with k = K.
Thus, for all t, by the above two claims and Lemmas 4.7, 4.8, under appropriate conditioning, et satisfies
(1).
We prove these lemmas in the next few subsections. The projected CS proof (item one of both lemmas)
uses the following lemma from [12] that relates the s-Restricted Isometry Constant (RIC), δs (.), [22] of a
projection matrix to the incoherence of its orthogonal complement.
Lemma 4.11. [[12]] For an n × r basis matrix P , (1) δs (I − P P 0 ) = max|T |≤s kIT 0 P k2 ; and (2)
max|T |≤s kIT 0 P k2 ≤ s maxi=1,2,...,n kIi 0 P k2 ≤ sµr/n.
The last bound of the above lemma is a consequence of Definition 1.1. We apply this lemma with
s = max-outlier-frac-col · n. The subspace update step proof (item 2 of both the above lemmas) uses a
guarantee for PCA in sparse data-dependent noise, Theorem 4.14, due to [25]. Notice that et = `t − `ˆt is
the noise/error seen by the subspace update step. By (1), this is sparse and depends on the true data `t .
Consider the actual tj unknown case. The following lemma is used to show that, whp, we can detect
subspace change within 2α time instants. This lemmas assumes detection threshold ωevals = 2ε2 λ+ (see
Algorithm 2).
Lemma 4.12 (Subspace Change Detection). Consider an α-length time interval J α ⊂ [tj , tj+1 ] (so that
`t = Pj at ).
1. If Φ := I − P̂j−1 P̂j−1 0 and SE(P̂j−1 , Pj−1 ) ≤ ε, with probability at least 1 − 10n−10 ,
!
1 X ˆ ˆ0
λmax
Φ`t `t Φ ≥ 0.28λ− SE2 (Pj−1 , Pj ) > ωevals
α
α
t∈J
2. If Φ := I − P̂j P̂j 0 and SE(P̂j , Pj ) ≤ ε, with probability at least 1 − 10n−10 ,
!
1 X ˆ ˆ0
Φ`t `t Φ ≤ 1.37ε2 λ+ < ωevals
λmax
α
α
t∈J
4.3
Proof of Lemma 4.7: projected CS and subspace update in the first update
interval
We first state a simple lemma. This is proved in Appendix B.
Lemma 4.13. Let Q1 , Q2 and Q3 be r-dimensional subspaces in Rn such that SE(Q1 , Q2 ) ≤ ∆1 and
SE(Q2 , Q3 ) ≤ ∆2 . Then, SE(Q1 , Q3 ) ≤ ∆1 + ∆2 .
Proof of Lemma 4.7. Recall from Definition 4.6 that s := max-outlier-frac-col · n and φ+ = 1.2. Recall also
that, for simplicity, we are considering the νt = 0 case.
Proof of item 1. First consider j > 0. We have conditioned on the event Γj,0 := Γj−1,K . This implies
that SE(P̂j−1 , Pj−1 ) ≤ ε.
We consider the interval t ∈ [t̂j , t̂j + α). For this interval, P̂(t−1) = P̂j−1 and thus Ψ = I − P̂j−1 P̂j−1 0
(from Algorithm). For the sparse recovery step, we first need to bound the 2s-RIC of Ψ. To do this, we
first obtain bound on max|T |≤2s kIT 0 P̂j−1 k as follows. Consider any set T such that |T | ≤ 2s. Then,
IT 0 P̂j−1 ≤ IT 0 (I − Pj−1 Pj−1 0 )P̂j−1 + IT 0 Pj−1 Pj−1 0 P̂j−1
≤ SE(Pj−1 , P̂j−1 ) + IT 0 Pj−1 = SE(P̂j−1 , Pj−1 ) + IT 0 Pj−1
15
Using Lemma 4.11, and the bound on max-outlier-frac-col from Theorem 2.2,
max kIT 0 Pj−1 k2 ≤ 2s max kIi 0 Pj−1 k2 ≤
i
|T |≤2s
2sµr
≤ 0.01
n
(2)
Thus, using SE(P̂j−1 , Pj−1 ) ≤ ε,
max IT 0 P̂j−1 ≤ ε + max IT 0 Pj−1 ≤ ε + 0.1
|T |≤2s
|T |≤2s
Finally, using Lemma 4.11, δ2s (Ψ) ≤ 0.112 < 0.15. Hence
Ψ Tt 0 Ψ Tt
−1
≤
1
1
1
≤
≤
< 1.2 = φ+ .
1 − δs (Ψ)
1 − δ2s (Ψ)
1 − 0.15
When j = 0, there are some minor changes. From the initialization assumption, we have SE(P̂0 , P0 ) ≤
0.25. Thus, max|T |≤2s IT 0 P̂0 ≤ 0.25 + 0.1 = 0.35. Thus, using Lemma 4.11, δ2s (Ψ0 ) ≤ 0.352 < 0.15.
The rest of the proof given below is the same for j = 0 and j > 0.
Next we bound norm of bt := Ψ`t . This and the RIC bound will then be used to bound kx̂t,cs − xt k.
kbt k = kΨ`t k = (I − P̂j−1 P̂j−1 0 )Pj at ≤ SE(P̂j−1 , Pj ) kat k
(a)
p
≤ (ε + SE(Pj−1 , Pj )) ηrλ+ := bb
where (a) follows from Lemma 4.13 with Q1 = P̂j−1 , Q2 = Pj−1 and Q3 = Pj . Under the assumptions
of Theorem 2.2, bb < xmin /15. This is why we have set ξ = xmin /15 in the Algorithm. Using these facts,
and δ2s (Ψ) ≤ 0.15, the CS guarantee from [22, Theorem 1.3] implies that
kx̂t,cs − xt k ≤ 7ξ = 7xmin /15 < xmin /2
Consider support recovery. From above,
|(x̂t,cs − xt )i | ≤ kx̂t,cs − xt k ≤ 7xmin /15 < xmin /2
The Algorithm sets ωsupp = xmin /2. Consider an index i ∈ Tt . Since |(xt )i | ≥ xmin ,
xmin − |(x̂t,cs )i | ≤ |(xt )i | − |(x̂t,cs )i | ≤ |(xt − x̂t,cs )i | <
Thus, |(x̂t,cs )i | >
(xt )j = 0 and so
xmin
2
xmin
2
= ωsupp which means i ∈ T̂t . Hence Tt ⊆ T̂t . Next, consider any j ∈
/ Tt . Then,
|(x̂t,cs )j | = |(x̂t,cs )j )| − |(xt )j | ≤ |(x̂t,cs )j − (xt )j | ≤ bb <
xmin
2
which implies j ∈
/ T̂t and T̂t ⊆ Tt implying that T̂t = Tt .
Finally, we get an expression for et and bound it. With T̂t = Tt and since Tt is the support of xt ,
xt = ITt ITt 0 xt , and so
−1
−1
x̂t = ITt ΨTt 0 ΨTt
ΨTt 0 (Ψ`t + Ψxt ) = ITt ΨTt 0 ΨTt
ITt 0 Ψ`t + xt
since ΨTt 0 Ψ = IT0 t Ψ0 Ψ = ITt 0 Ψ. Thus et = x̂t − xt satisfies
−1
et = ITt ΨTt 0 ΨTt
ITt 0 Ψ`t and
−1
ket k ≤ ΨTt 0 ΨTt
ITt 0 Ψ`t ≤ φ+ ITt 0 Ψ`t ≤ 1.2bb
Proof of Item 2 : We will use the following result from [25, Remark 4.18].
16
Theorem 4.14 (PCA in sparse data-dependent noise (PCA-SDDN)). We are given data vectors yt :=
`t + wt with wt = ITt Ms,t `t , t = 1, 2, . . . , α, where Tt is the support set of wt , and `t = P at with at
satisfying the assumptions of Theorem 2.2. Pick an εSE > 0. Assume that maxt kMs,t P k2 ≤√q < 1,
the fraction of non-zeroes in any row of the noise matrix [w1 , w2 , . . . , wα ] is bounded by b, and 3 bqf ≤
0.9εSE /(1 + εSE ). Define
α0 := Cη
q2f 2
r log n
ε2SE
For an α ≥ α0 , let P̂ be the matrix of top r eigenvectors of D :=
1 − 10n−10 , SE(P̂ , P ) ≤ εSE .
1
α
Pα
0
t=1 yt yt .
With probability at least
Since `ˆt = `t − et with et satisfying (1), updating P̂(t) from the `ˆt ’s is a problem of PCA in sparse
data-dependent noise (SDDN), et . To analyze this, we use Theorem 4.14 given above. Recall from Item
1 of this lemma that, for t ∈ [t̂j , t̂j + α), et satisfies (1). Recall from the Algorithm that we compute
Pt̂j +α−1 ˆ ˆ 0
the first estimate of the j-th subspace, P̂j,1 , as the top r eigenvectors of α1 t=
`t `t . In the notation
t̂j
−1
0
0
of Theorem 4.14, yt ≡ `ˆt , wt ≡ et , `t ≡ `t , Ms,t = − (ΨT ΨT ) ΨT , P̂ = P̂j,1 , P = Pj , and so
t
t
t
kMs,t P k = k (ΨTt 0 ΨTt )−1 ΨTt 0 Pj k ≤ 1.2(ε + SE(Pj−1 , Pj )) := q0 . Also, b ≡ b0 which is the upper
bound on max-outlier-frac-rowα (see Definition 4.6). Applying Theorem 4.14 with q ≡ q0 , b ≡ b0 and using
εSE = q0 /4, observe that we require
p
0.9(q0 /4)
b0 q0 f ≤
.
1 + (q0 /4)
Since q0 = 1.2(ε + SE(Pj−1 , Pj )) < 1.2(0.01 + 0.8) < 0.98 (follows from the bounds on ε and on ∆
√
given in Theorem 2.2), this holds if b0 f ≤ 0.18. This is true since we have assumed b0 = 0.01/f 2 (see
Definition 4.6). Thus, from Theorem 4.14, with probability at least 1 − 10n−10 , SE(P̂j,1 , Pj ) ≤ q0 /4.
Thus, conditioned on Γj,0 , with this probability, Γj,1 holds.
Remark 4.15 (Clarification about conditioning). In the proof above we have used Theorem 4.14 which
assumes that, for t ∈ J α , the at ’s are mutually independent and the matrices Ms,t are either nonrandom or are independent of the at ’s for this interval. When we apply the theorem for our proof, we are
conditioning on Γj,0 . This does not cause any problem since the event Γj,0 is a function of the random
variable yold := {y1 , y2 , . . . , yt̂j −1 } where as our summation is over J α := [t̂j , t̂j + α). Also, by Theorem
2.2, at ’s are independent of the outlier supports Tt .
To be precise, we are applying Theorem 4.14 conditioned on yold , for any yold ∈ Γj,0 . Even conditioned
on yold , clearly, the matrices Ms,t used above are independent of the at ’s for this interval. Also, even
conditioned on yold , the at ’s for t ∈ [t̂j , t̂j + α) are clearly mutually independent. Thus, the theorem can be
applied. Its conclusion then tells us that, for any yold ∈ Γj,0 , conditioned on yold , with probability at least
1 − 10n−10 , SE(P̂j,1 , Pj ) ≤ q0 /4. Since this holds with the same probability for all yold ∈ Γj,0 , it also holds
with the same probability when we condition on Γj,0 . Thus, conditioned on Γj,0 , with this probability, Γj,1
holds.
An analogous argument will also apply to the following proofs.
4.4
Proof of Lemma 4.8: lemma for projected CS and subspace update in k-th update
interval
Proof of Lemma 4.8. We first present the proof for the k = 2 case and then generalize it for an arbitrary
k.
(A) k = 2: We have conditioned on Γj,1 . This implies that SE(P̂j,1 , Pj ) ≤ q0 /4.
17
Proof of Item 1 : We consider the interval t ∈ [t̂j + α, t̂j + 2α). For this interval, P̂(t−1) = P̂j,1 and
thus Ψ = I − P̂j,1 P̂j,1 0 (from Algorithm). For the sparse recovery step, we need to bound the 2s-RIC of
Ψ. Consider any set T such that |T | ≤ 2s. We have
IT 0 P̂j,1 ≤ IT 0 (I − Pj Pj 0 )P̂j,1 + IT 0 Pj Pj 0 P̂j,1
≤ SE(Pj , P̂j,1 ) + IT 0 Pj = SE(P̂j,1 , Pj ) + IT 0 Pj
The equality holds since SE is symmetric for subspaces of the same dimension. Using SE(P̂j,1 , Pj ) ≤ q0 /4,
(2),
max IT 0 P̂j,1 ≤ q0 /4 + max IT 0 Pj ≤ q0 /4 + 0.1.
|T |≤2s
|T |≤2s
Finally, from using the assumptions of Theorem 2.2, q0 ≤ 0.96. Using this and Lemma 4.11,
δ2s (Ψj ) = max IT 0 P̂j,1
|T |≤2s
2
≤ 0.352 < 0.15.
From, this
ΨTt 0 ΨTt
−1
≤
1
1
1
≤
≤
< 1.2 = φ+ .
1 − δs (Ψ)
1 − δ2s (Ψ)
1 − 0.15
Consider kbt k.
p
kbt k = kΨ`t k = (I − P̂j,1 P̂j,1 0 )Pj at ≤ SE(P̂j,1 , Pj ) kat k ≤ (q0 /4) ηrλ+
(a)
p
≤ 0.3(ε + SE(Pj−1 , Pj )) ηrλ+ := 0.3bb
We have 0.3bb < bb < xmin /15 as in the proof of Lemma 4.7. The rest of the proof is the same too. Notice
here that, we could have loosened the required lower bound on xmin for this interval.
Proof of Item 2 : Again, updating P̂(t) using `ˆt ’s is a problem of PCA in sparse data-dependent
noise (SDDN), et . We use the result of Theorem 4.14. Recall from Item 1 of this lemma that, for
Pt̂j +2α−1 ˆ ˆ 0
t ∈ [t̂j + α, t̂j + 2α), et satisfies (1). We compute P̂j,2 as the top r eigenvectors of 1
`t `t . In
α
t=t̂j +α
notation of Theorem 4.14, yt ≡ `ˆt , wt ≡ et , `t ≡ `t , P ≡ Pj , P̂ ≡ Pj,2 , and Ms,t = − (ΨTt 0 ΨTt )−1 ΨTt 0 .
So kMs,t Pj k = k (ΨTt 0 ΨTt )−1 ΨTt 0 Pj k ≤ (φ+ /4)q0 := q1 . Applying Theorem 4.14 with q ≡ q1 , b ≡ b0 (b0
bounds max-outlier-frac-rowα ), and setting εSE = q1 /4, observe that we require
p
b0 q1 f ≤
0.9(q1 /4)
1 + (q1 /4)
√
which holds if b0 f ≤ 0.18. This is ensured since b0 = 0.01/f 2 . Thus, from Theorem 4.14, with probability
at least 1 − 10n−10 , SE(P̂j,2 , Pj ) ≤ (q1 /4) = 0.25 · 0.3q0 . Thus, with this probability, conditioned on Γj,1 ,
Γj,2 holds.
(B) General k: We have conditioned on Γj,k−1 . This implies that SE(P̂j,k−1 , Pj ) ≤ qk−1 /4.
Proof of Item 1 : Consider the interval [t̂j + (k − 1)α, t̂j + kα). In this interval, P̂(t−1) = P̂j,k−1 and
thus Ψ = I − P̂j,k−1 P̂j,k−1 0 . Using the same idea as for the k = 2 case, we have that for the k-th interval,
qk−1 = (φ+ /4)k−1 q0 . Pick εSE = (qk−1 /4). From this it is easy to see that
2
0
δ2s (Ψ) ≤ max IT P̂j,k−1
≤ (SE(P̂j,k−1 , Pj ) + max IT 0 Pj )2
|T |≤2s
|T |≤2s
(a)
≤ (SE(P̂j,k−1 , Pj ) + 0.1)2 ≤ ((φ+ /4)k−1 (ε + SE(Pj−1 , Pj ) + 0.1)2 < 0.15
18
where (a) follows from (2). Using the approach Lemma 4.7,
p
kΨ`t k ≤ SE(P̂j,k−1 , Pj ) kat k ≤ (φ+ /4)k−1 (ε + SE(Pj−1 , Pj )) ηrλ+
(a)
p
≤ (φ+ /4)k−1 (ε + ∆) ηrλ+ := (φ+ /4)k−1 bb
Proof of Item 2 : Again, updating P̂(t) from `ˆt ’s is a problem of PCA in sparse data-dependent noise
given in Theorem 4.14. From Item 1 of this lemma that, for t ∈ [t̂j + (k − 1)α, t̂j + kα], et satisfies
Pt̂j +kα−1
(1). We update the subspace, P̂j,k as the top r eigenvectors of α1 t=
`ˆ `ˆ 0 . In the setting above
t̂j +(k−1)α t t
yt ≡ `ˆt , wt ≡ et , `t ≡ `t , and Ms,t = − (ΨT 0 ΨT )−1 ΨT 0 , and so kMs,t Pj k = k (ΨT 0 ΨT )−1 ΨT 0 Pj k ≤
t
t
t
t
t
t
(φ+ /4)k−1 q0 := qk−1 . Applying Theorem 4.14 with q ≡ qk−1 , b ≡ b0 (b0 bounds max-outlier-frac-rowα ),
and setting εSE = qk−1 /4, we require
p
0.9(qk−1 /4)
b0 qk−1 f ≤
1 + (qk−1 /4)
√
which holds if b0 f ≤ 0.12. This is true by our assumption. Thus, from Theorem 4.14, with probability
at least 1 − 10n−10 , SE(P̂j,k , Pj ) ≤ (φ+ /4)k−1 q1 . Thus, with this probability, conditioned on Γj,k−1 , Γj,k
holds.
4.5
Proof of Lemma 4.12: subspace change detection lemma
Proof of Lemma 4.12. The proof uses the following lemma. It is proved in Appendix B. The proof
uses Cauchy-Schwartz for sums of matrices, followed by either matrix Bernstein [28] or Vershynin’s subGaussian result [29].
Lemma 4.16 (Concentration Bounds). Assume that the assumptions of Theorem 2.2 hold. For this
P
lemma assume that `t = P at , Φ := I − P̂ P̂ 0 , et = M2,t M1,t `t , with k α1 t∈J α M2,t M2,t 0 k ≤ b0 and
kM1,t P k ≤ q. Assume that event E0 holds. Then,
!
1X
Pr
at at 0 − Λ ≤ 0 λ− ≥ 1 − 10n−10
α t
!
p
1X
Pr
Φ`t et 0 Φ ≤ (1 + 1 )SE(P̂ , P ) b0 qλ+ ≥ 1 − 10n−10
α t
!
p
1X
Pr
Φet et 0 Φ ≤ (1 + 2 ) b0 q 2 λ+ ≥ 1 − 10n−10
α t
!
1X
0
−
≥ 1 − 2n−10
Pr
Φ`t νt Φ ≤ l,v λ
α t
!
1X
2
−
0
Pr
≥ 1 − 2n−10
Φνt νt Φ ≤ (ε + v,v )λ
α t
Proof of Item (a): First from Corollary 4.10, note that for t ∈ [tj , t̂j ], the error et satisfies (1). We
19
have
!
1 X
λmax
Φ[Pj at at 0 Pj 0 + et et 0 + `t et 0 + et `t 0 ]Φ
α
t∈J α
!
!
X
(a)
1 X
1
≥ λmax
ΦPj at at 0 Pj 0 Φ + λmin
Φ[et et 0 + `t et 0 + et `t 0 ]Φ
α
α
α
α
t∈J
t∈J
!
1 X
1 X
1 X
ΦPj at at 0 Pj 0 Φ −
Φet et 0 Φ − 2
Φ`t et 0 Φ
≥ λmax
α
α
α
α
α
α
t∈J
t∈J
t∈J
1 X
1 X
:= λmax (T ) −
Φet et 0 Φ − 2
Φ`t et 0 Φ
α
α
α
α
t∈J
(3)
t∈J
where (a) follows from Weyl’s Inequality. Now we bound the second and third terms by invoking Lemma
4.16 with E0 := {SE(P̂j−1 , Pj−1 ) ≤ ε}, P̂ ≡ P̂j−1 , P ≡ Pj , M2,t ≡ ITt and M1,t ≡ (ΨTt 0 ΨTt )−1 ΨTt 0 ,
where Ψ = I − P̂j−1 P̂j−1 0 . Thus, b0 ≡ b0 , q ≡ q0 . Thus, with probability at least 1 − 4n−10 ,
p
p
1 X
1 X
Φet et 0 Φ + 2
Φ`t et 0 Φ ≤ b0 q02 λ+ (1 + 2 ) + 2 b0 q0 (SE(Pj−1 , Pj ) + ε)λ+ (1 + 1 )
α
α
α
α
t∈J
t∈J
(4)
The above equation uses the fact that SE(P̂j−1 , Pj ) ≤ ε + SE(Pj−1 , Pj ) which is a direct consequence
QR
of using Lemma 4.13. We bound the first term of (3) as follows. Let ΦPj = Ej Rj be its reduced QR
decomposition. Thus Ej is an n × r matrix with orthonormal columns and Rj is an r × r upper triangular
matrix. Let
!
1 X
0
at at Rj 0 .
A := Rj
α
α
t∈J
Observe that T can also be written as
A0
Ej 0
T = Ej Ej,⊥
0 0 Ej,⊥ 0
(5)
and thus λmax (A) = λmax (T ). We work with λmax (A) in the sequel. We will use the following simple
claim.
Claim 4.17. If X 0 (i.e., X is a p.s.d matrix), where X ∈ Rr×r , then RXR0 0 for all R ∈ Rr×r .
Proof. Since X is p.s.d., y 0 Xy ≥ 0 for any vector y. Use this with y = R0 z for any z ∈ Rr . We get
z 0 RXR0 z ≥ 0. Since this holds for all z, RXR0 0.
Using Lemma 4.16, it follows that
Pr
1X
at at 0 − (λ− − 0 )I 0
α t
!
≥ 1 − 2n−10 .
Using Claim 4.17, with probability 1 − 2n−10 ,
Rj
1X
at at 0 − (λ− − 0 )I
α t
=⇒ λmin
Rj
!
Rj 0 0
1X
at at 0 − (λ− − 0 )I
α t
20
!
!
Rj 0
≥0
Using Weyl’s inequality [30], with the same probability,
!
!
1X
λmin Rj
at at 0 − (λ− − 0 )I Rj 0 ≤ λmax Rj
α t
1X
at at 0
α t
!
!
Rj 0
− (λ− − 0 )λmax Rj Rj 0
and so,
λmax (A) ≥ (λ− − 0 )λmax (Rj Rj 0 ).
We now obtain a lower bound on the second term in the rhs above.
λmax (Rj Rj 0 ) = λmax (Pj 0 (I − P̂j−1 P̂j−1 0 )Pj )) = λmax (Pj 0 (I − Pj−1 Pj−1 0 + Pj−1 Pj−1 0 − P̂j−1 P̂j−1 0 )Pj ))
≥ λmax (Pj 0 (I − Pj−1 Pj−1 0 )Pj ) + λmin (Pj 0 (Pj−1 Pj−1 0 − P̂j−1 P̂j−1 0 )Pj )
2
= σmax
((I − Pj−1 Pj−1 0 )Pj ) + λmin (Pj 0 (Pj−1 Pj−1 0 − P̂j−1 P̂j−1 0 )Pj )
≥ SE2 (Pj−1 , Pj ) − Pj 0 (Pj−1 Pj−1 0 − P̂j−1 P̂j−1 0 )Pj
(a)
≥ SE2 (Pj−1 , Pj ) − Pj−1 Pj−1 0 − P̂j−1 P̂j−1 0 ≥ SE2 (Pj−1 , Pj ) − 2ε
(6)
where we have used [12, Lemma 2.10]. Thus, combining (3), (4), (6), and using 0 = 0.01, 1 = 2 = 0.01,
with probability at least 1 − 10n−10 ,
!
p
1 X ˆ ˆ0
λmax
Φ`t `t Φ ≥ 0.99λ− (SE2 (Pj−1 , Pj ) − 2ε) − λ+ [ b0 q0 (1.01q0 + 2.02(ε + SE(Pj−1 , Pj )))]
α
α
t∈J
(a)
h
i
p
≥ λ− 0.91SE2 (Pj−1 , Pj ) − 2.7 b0 f (ε + SE(Pj−1 , Pj )2
(b)
≥ λ− 0.91SE2 (Pj−1 , Pj ) − 0.54(ε2 + SE2 (Pj−1 , Pj )
≥ λ− SE2 (Pj−1 , Pj )(0.91 − 0.54 · 1.16) ≥ 0.28λ− SE2 (Pj−1 , Pj )
where (a) uses q0 = 1.2(ε + SE(Pj−1 , Pj )) and ε ≤ 0.03SE2 (Pj−1 , Pj )/f 2 < 0.4SE2 (Pj−1 , Pj ), (b) uses
√
b0 f = 0.1 and (a + b)2 ≤ 2(a2 + b2 ). and the last inequality again uses ε ≤ 0.03SE2 (Pj−1 , Pj )/f 2 .
Proof of Item (b): First, we recall that from Corollary 4.10, for t ∈ [t̂j + Kα, tj+1 ), the error et satisfies
(7).
!
!
!
X
1 X ˆ ˆ0
1 X
1
λmax
Φ`t `t Φ ≤ λmax
Φ`t `t 0 Φ + λmax
Φ[`t et 0 + et `t 0 + et et 0 ]Φ
α
α
α
α
α
α
t∈J
t∈J
t∈J
!
1 X
1 X
1 X
≤ λmax
Φ`t `t 0 Φ +
Φet et 0 Φ + 2
Φ`t et 0 Φ
α
α
α
α
α
α
t∈J
t∈J
t∈J
1 X
1 X
:= λmax (T ) +
Φet et 0 Φ + 2
Φ`t et 0 Φ
α
α
α
α
t∈J
t∈J
To obtain bounds on the second and third terms in the equation above we invoked Lemma 4.16 with
E0 := {SE(P̂j , Pj ) ≤ ε}, P̂ ≡ P̂j , P ≡ Pj , M2,t ≡ ITt , M1,t ≡ (ΨTt 0 ΨTt )−1 ΨTt 0 , where Ψ = I − P̂j P̂j 0
and b0 ≡ b0 , ≡ qK , . Thus, with probability at least 1 − 10n−10 ,
p
1 X
1 X
Φet et 0 Φ + 2
Φ`t et 0 Φ ≤ b0 qK λ+ (qK (1 + 2 ) + 2(1 + 1 )ε)
α
α
α
α
t∈J
t∈J
21
QR
The above equation also uses SE(P̂j , Pj ) ≤ ε. Proceeding as before to bound λmax (T ), define ΦPj =
Ej Rj , define A as before, we know λmax (T ) = λmax (Ej 0 T Ej ) = λmax (A). Further,
!
!
!
X
(a)
1 X
1
λmax (A) = λmax Rj
at at 0 Rj 0 ≤ λmax
at at λmax (Rj Rj 0 )
α
α
α
α
t∈J
t∈J
where (a) uses Ostrowski’s theorem [30, Theorem 5.4.9]. We have
2
2
λmax (Rj Rj 0 ) = σmax
(Rj ) = σmax
((I − P̂j P̂j 0 )Pj ) ≤ ε2
P
and we can bound λmax ( α1 t∈J α at at 0 ) using the first item of Lemma 4.16 with 0 = 0.01. Combining
all of the above, and setting 1 = 2 = 0.01, when the subspace has not changed, with probability at least
1 − 10n−10 ,
!
(a)
p
1 X ˆ ˆ0
λmax
Φ`t `t Φ ≤ λ+ [1.01ε2 + b0 qK (1.01qK + 2.01ε)] ≤ 1.37ε2 λ+
α
α
t∈J
where (a) uses qK ≤ ε and b0 f 2 = 0.01. Under the condition of Theorem 2.2, recall that ωevals = 2ε2 λ+
and thus, with high probability, 1.37ε2 λ+ < ωevals < 0.28λ− SE2 (Pj−1 , Pj ).
5
5.1
Extensions
Static Robust PCA
A useful corollary of our result for RST is that NORST is also the first online algorithm that provides
a provable finite sample guarantee for the static Robust PCA problem. Static RPCA is our problem
setting with J = 1, or in other words, with `t = P at . A recent work, [21], developed an online stochastic
optimization based reformulation of PCP, called ORPCA, to solve this. Their paper provides only a
partial guarantee because the guarantee assumes that the intermediate algorithm estimates, P̂(t) , are fullrank. Moreover the guarantee is only asymptotic. Instead our result given below is a complete guarantee
and is non-asymptotic.
Corollary 5.18 (Static RPCA). Consider Algorithm 1 with t2 = ∞. Theorem 2.2 holds with following
modification: replace the slow subspace change assumption with a fixed subspace P . Everything else
remains as is, but with r ≡ rL . Under the assumptions of Theorem 2.2, all the conclusions hold with same
probability. The time and memory complexity are O(ndr log(1/)) and O(nr log n log(1/)).
In applications such as “robust” dimensionality reduction [31, 32] where the objective is to just obtain
the top-r directions along which the variability of data is maximized, we only need the first Kα =
Cf 2 r log n log(1/ε) samples to obtain an ε-accurate subspace estimate. If only these are used, the time
complexity reduces to O(nKαr log(1/ε)) = O(nr2 log n log2 (1/)). This is faster than even NO-RMC [11]
and does not require d ≈ n; of course it requires the other extra assumptions discussed earlier.
5.2
Subspace Tracking with Missing Data (ST-missing) and Dynamic Matrix Completion
Another useful corollary of our result is a guarantee for the ST-missing problem. Consider the subspace
tracking with missing data (ST-missing) problem. By setting the missing entries at time t to zero, and
by defining Tt to be the set of missing entries at time t, we observe n-dimensional vectors that satisfy
yt := `t − ITt ITt 0 `t , for t = 1, 2, . . . , d.
22
(7)
Algorithm 3 NORST-Random for subspace tracking with missing data (ST-missing)
Algorithm 2 with the following changes
1. Replace line 3 with: compute P̂0 ← P̂init ← Generate an n × r basis matrix from the random orthogonal model;
j ← 1, k ← 1
2. Replace line 6 with the following
• Ψ ← I − P̂(t−1) P̂(t−1) 0 ; ỹt ← Ψyt ; x̂t ← ITt (ΨTt 0 ΨTt )−1 ΨTt 0 ỹt ; `ˆt ← yt − x̂t .
with xt ≡ ITt ITt 0 `t . This can be interpreted as a special case of the RST problem where the set Tt is
known. Because there are no sparse corruptions (outliers), there is no xmin . Thus the initialization error
√
need not be O(1/ r) (needed in the RST result to ensure a reasonable lower bound on xmin ) and so
one can even use random initialization. We assume that the initialization is obtained using the Random
Orthogonal Model described in [33]. As explained in [33], a basis matrix generated from this model is
already µ-incoherent. We have the following corollary. The only change in its proof is the proof of the
first subspace update interval for the j = 0 case.
Corollary 5.19 (ST-missing). Consider NORST-Random (Algorithm 3). If the assumptions of Theorem
2.2 on at and νt hold, Pj ’s are µ-incoherent, tj+1 − tj > (K + 2)α, ∆ < 0.8, the outlier fraction bounds
given in Theorem 2.2 hold, and if, for t ∈ [t0 , t1 ], max-outlier-frac-col ≤ c/(log n), then all conclusions of
Theorem 2.2 on P̂(t) and on `t hold with the same probability.
To our knowledge, the above is the first complete non-asymptotic guarantee for ST-missing; and the
first result that allows changing subspaces. All existing guarantees are either partial guarantees (make
assumptions on intermediate algorithm estimates), e.g., [19], or provide only asymptotic results [16, 18].
Moreover, from a dynamic matrix completion viewpoint, it is also giving a matrix completion solution
without assuming that the set of observed entries is generated from a uniform or a Bernoulli model. Of
course the tradeoff is that it needs many more observed entries. All these points will be discussed in detail
in follow-up work where we will also numerically evaluate NORST-random for this problem.
5.3
Fewer than r directions change
It is possible to relax the lower bound on outlier magnitudes if not all of the subspace directions change
at a given subspace change time. Suppose that only rch < r directions change. When rch = 1, we recover
the guarantee of [15] but for NORST (which is a simpler algorithm than s-reprocs).
Let Pj−1,fix denote a basis for the fixed component of Pj−1 and let Pj−1,ch denote a basis for its
changing component. Thus, Pj−1 R = [Pj−1,fix , Pj−1,ch ], where R is a r × r rotation matrix. We have
Pj = [Pj−1,fix , Pj,chd ]
(8)
where Pj,chd is the changed component and has the same dimension as Pj−1,ch . Thus,
SE(Pj−1 , Pj ) = SE(Pj−1,ch , Pj,chd )
(9)
and so ∆ = maxj SE(Pj−1 , Pj ) = maxj SE(Pj−1,ch , Pj,chd ). Let λ+
ch denote the largest eigenvalue along
any direction in Pj,chd .
Corollary 5.20. In Algorithm 2, replace line 17 by P̂(t) ← basis(P̂j−1 , P̂j,k ). For basis matrices P1 , P2 ,
we use P = basis(P1 , P2 ) to mean that P is a basis matrix with column span equal to the column span of
23
log(SE(P̂(t) , P(t) ))
GRASTA
ORPCA
s-ReProCS
NORST
Offline-NORST
0
−2
−4
2,000
4,000
6,000
8,000
10,000
2,000
t
4,000
6,000
t
8,000
10,000
Figure 2: Left plot illustrates the `t error for outlier supports generated using Moving Object Model and
right plot illustrates the error under the Bernoulli model. The values are plotted every kα−1 time-frames.
[P1 , P2 ]. Assume that (8) and (9) hold. Also assume that the conditions
of Theorem 2.2 holds with the
q
p
+
lower bound on xmin relaxed to xmin ≥ C(ε η(r − rch )λ + (ε + ∆) ηrch λ+
ch ). Then, all conclusions of
Theorem 2.2 hold.
6
Empirical Evaluation
In this section we present the results for extensive numerical experiments on synthetic and real data to
R
validate our theoretical claims. All time comparisons are performed on a Desktop Computer with Intel
Xeon E3-1240 8-core CPU @ 3.50GHz and 32GB RAM and all synthetic data experiments are averaged
over 100 independent trials. The codes are available at https://github.com/praneethmurthy/NORST.
6.1
Synthetic Data
We perform three experiments on synthetic data to corroborate our theoretical claims.
Experiment 1.
We compare the results of NORST and Offline-NORST with static RPCA algorithms, and Robust Subspace Tracking/Online RPCA methods proposed in literature. For our first
experiment, we generate the changing subspaces using Pj = eγj Bj Pj−1 as done in [34] where γj controls
the subspace change and Bj ’s are skew-symmetric matrices. In the first experiment we used the following
parameters. n = 1000, d = 12000, J = 2, t1 = 3000, t2 = 8000, r = 30, γ1 = 0.001, γ2 = γ1 and
the matrices B1 and B2 are generated as B1 = (B̃1 − B̃1 ) and B2 = (B̃2 − B̃2 ) where the entries of
B̃1 , B̃2 are generated independently from a standard normal distribution. We set α = 300. This gives us
the basis matrices P(t) for all t. To obtain the low-rank matrix L from this we generate the coefficients
i.i.d
at ∈ Rr as independent zero-mean, bounded random variables. They are (at )i ∼ unif [−qi , qi ] where
√
√
qi = f − f (i − 1)/2r for i = 1, 2, · · · , r − 1 and qr = 1. thus the condition number is f and we
selected f = 50. For the sparse supports, we considered two models according to which the supports are
generated. First we use Model G.24 [15] which simulates a moving object pacing in the video. For the
first ttrain = 100 frames, we used a smaller fraction of outliers with parameters s/n = 0.01, b0 = 0.01. For
t > ttrain we used s/n = 0.05 and b0 = 0.3. Secondly, we used the Bernoulli model to simulate sampling
uniformly at random, i.e., each entry of the matrix, is independently selected with probability ρ or not
selected with probability 1 − ρ. We generate the sparse supports using the Bernoulli model using ρ = 0.01
for the first ttrain frames and ρ = 0.3 for the subsequent frames. The sparse outlier magnitudes for both
support models are generated uniformly at random from the interval [xmin , xmax ] with xmin = 10 and
xmax = 20.
24
Table 2: Comparison of kL̂ − LkF /kLkF for Online and offline RPCA methods. Average time for the
Moving Object model is given in parentheses. The offline (batch) methods are performed once on the
complete dataset.
Outlier Model
GRASTA
(0.02 ms)
s-ReProCS
(0.9 ms)
ORPCA
(1.2ms)
NORST
(0.9 ms)
RPCA-GD
(7.8ms)
AltProj
(4.6ms)
Offline-NORST
(1.7ms)
Moving Object
Bernoulli
0.630
6.163
0.598
2.805
0.861
1.072
4.23 × 10−4
0.002
4.283
0.092
4.441
0.072
8.2 × 10−6
2.3 × 10−4
We initialized the s-ReProCS and NORST algorithms using AltProj applied to Y[1,ttrain ] with ttrain =
100. For the parameters to AltProj we used used the true value of r, 15 iterations and a threshold of 0.01.
This, and the choice of γ1 and γ2 ensure that SE(P̂init , P0 ) ≈ SE(P1 , P0 ) ≈ SE(P2 , P1 ) ≈ 0.01. The other
algorithm parameters are set as mentioned in the theorem, i.e., K = dlog(c/ε)e = 8, α = Cr log n = 300,
ω = xmin /2 = 5 and ξ = 7xmin /15 = 0.67, ωevals = 2ε2 λ+ = 7.5 × 10−4 . For l1 minimization we used
the YALL-1 toolbox [35] and set the tolerance to 10−4 . For the least-squares step we use the Conjugate
Gradient Least Squares instead of the well-known “backslash” operator in MATLAB since this is a well
conditioned problem. For this we set the tolerance as 10−10 and the number of iterations as 10. We have
not done any code optimization such as use of MEX files for various sub-routines to speed up our algorithm.
For the other online methods we implement the algorithms without
modifications. The regularization
√
√
parameter for ORPCA was set as with λ1 = 1/ n and λ2 = 1/ d according to [21]. Wherever possible
we set the tolerance as 10−6 and 100 iterations to match that of our algorithm. As shown in Fig. 2,
NORST is significantly better than all the RST methods - s-ReProCS [15], and two popular heuristics
from literature - ORPCA [21] and GRASTA [34].
We also provide a comparison of offline techniques in Table 2. We must mention here that we implemented the static RPCA methods once on the entire data matrix, Y . We do this to provide a roughly
equal comparison of the time taken. In principle, we could also implement the static techniques on disjoint
batches of size α, but we observed that this did not yield significant improvement in terms of reconstruction accuracy, while being considerably slower, and thus we report only the latter setting. As can be
seen, offline NORST outperforms all static RPCA methods, both for the moving object outlier support
model and for the commonly used random Bernoulli support model. For the batch comparison we used
√
PCP, AltProj and RPCA-GD. We set the regularization parameter for PCP 1/ n in accordance with [3].
The other known parameters, r for Alt-Proj, outlier-fraction for RPCA-GD, are set using the true values.
Furthermore, for all algorithms (the IALM solver in case of PCP) we set the threshold as 10−6 and the
number of iterations to 100 as opposed to 10−3 and 50 iterations which were set as default to provide a
fair comparison with NORST and Offline-NORST. All results are averaged over 100 independent runs.
Experiment 2.
Next we perform an experiment to validate our claim of NORST admitting a
higher fraction of outliers per row than the state of the art. In particular, since AltProj has the highest
tolerance to max-outlier-frac-row, we only compare with it. The experiment proceeded as follows. We chose
10 different values of each of r and b0 (we slightly misuse notation here to let b0 := max-outlier-frac-row for
this section only). For each pair of b0 and r we implemented NORST and ALtProj over 100 independent
trials and computed the relative error in recovering L, i.e., we computed kL̂ − LkF /kLkF for each run.
We computed the empirical probability of success, i.e., we enumerated the number of times out of 100
the error seen by each algorithm was less than a threshold, 0.5.
For each pair of {b0 , r} we used the Bernoulli model for sparse support generation, the low rank matrix
is generated exactly as done in the previous experiments with the exception that again to provide an equal
footing, we increased the “subspace change” by setting γ1 and γ2 to 10 times the value that was used in
the previous experiment. For the first ttrain frames we used b0 = 0.02. We provide the phase transition
25
(a) Phase Transition for Alt Proj
(b) Phase Transition for NORST
30
30
(a)
(b)
20
r
r
20
10
10
0.1
0.2
0.3
b0
0.4
0.5
0.6
0.7
0.1
0.2
0.3
b0
0.4
0.5
0.6
0.7
Figure 3: Numerically computed kL̂ − Lk2F /kLk2F for AltProj and for Offline NORST. Note that NORST
indeed has a much higher tolerance to outlier fraction per row as compared to AltProj. Black denotes 0
and white denotes 1.
log(SE(P̂(t) , P(t) ))
AltProj
Offline NORST
xmin = 0.5
xmin = 5
xmin = 10
0
−2
−4
2,000
4,000
6,000
8,000
10,000
2,000
t
4,000
6,000
8,000
10,000
t
Figure 4: In the above plots we show the variation of the subspace errors for varying xmin . In particular,
we set all the non-zero outlier values to xmin . The results are averaged over 100 independent trials.
plots for both algorithm in Fig. 2. Here, white represents success while black represents failure. As can
be seen, NORST is able to tolerate a much larger fraction of outlier-per-row as compared to AltProj.
Experiment 3. Finally we perform an experiment to analyze the effect of the lower bound on the
outlier magnitude xmin with the performance of NORST and AltProj. We show the results in Fig. 4.
In the first stage, we generate the data exactly as done in the Moving Object sparse support model of
the first experiment. The only change to the data generation parameters is that we now choose three
different values of xmin = {0.5, 5, 10}. Furthermore, we set all the non-zero entries of the sparse matrix
to be equal to xmin . This is actually harder than allowing the sparse outliers to take on any value since
for a moderately low value of xmin the outlier-lower magnitude bound of Theorem 2.2 is violated. This is
indeed confirmed by the numerical results presented in Fig. 4. (i) When xmin = 0.5, NORST works well
since now all the outliers get classified as the small unstructured noise νt . (ii) When xmin = 10, NORST
still works well because now xmin is large enough so that the outlier support is mostly correctly recovered.
(iii) But when xmin = 5 the NORST reconstruction error stagnates around 10−3 .
All AltProj errors are much worse than those of NORST because the outlier fraction per row is the
same as in the first experiment. What can be noticed though is that the variation with varying xmin is
not that significant.
6.2
Real Data
In this section we provide empirical results on real video for the task of Background Subtraction. For the
AltProj algorithm we set r = 40. The remaining parameters were used with default setting. For NORST,
we set α = 60, K = 3, ξt = kΨ`ˆt−1 k2 . We found that these parameters work for most videos that we
26
(a) original
(b) NORST
(c) AltProj
(d) RPCA-GD
(e) GRASTA
(f) PCP
Figure 5: Comparison of visual performance in Foreground Background separation for the MR dataset.
The images are shown at t = ttrain + 10, 140, 630, 760.
(a) original
(b) NORST
(c) AltProj
(d) RPCA-GD
(e) GRASTA
(f) PCP
Figure 6: Comparison of visual performance in Foreground Background separation for the Lobby (LB)
dataset. The recovered background images are shown at t = ttrain + 260, 545, 610.
verified our algorithm on. For RPCA-GD we set the “corruption fraction” α = 0.2 as described in their
paper.
Meeting Room (MR) dataset: The meeting room sequence is set of 1964 images of resolution 64 × 80.
The first 1755 frames consists of outlier-free data. Henceforth, we consider only the last 1209 frames. For
NORST, we used ttrain = 400. In the first 400 frames, a person wearing a black shirt walks in, writes
something on the board and goes back. In the subsequent frames, the person walks in with a white shirt.
This is a challenging video sequence because the color of the person and the color of the curtain are hard
to distinguish. NORST is able to perform the separation at around 43 frames per second. We present
the results in Fig. 5
Lobby (LB) dataset: This dataset contains 1555 images of resolution 128 × 160. The first 341 frames
are outlier free. Here we use the first 400 “noisy” frames as training data. The Alt Proj algorithm is used
to obtain an initial estimate with rank, r = 40. The parameters used in all algorithms are exactly the
same as above. NORST achieves a “test” processing rate of 16 frames-per-second. We present the results
in Fig. 6
27
Acknowledgments
The authors would like to thank Praneeth Netrapalli and Prateek Jain of Microsoft Research India for
fruitful discussions on strengthening the guarantee by removing assumptions on subspace change model.
References
[1] P. Narayanamurthy and N. Vaswani, “Nearly optimal robust subspace tracking and dynamic robust
pca,” submitted to a double blind conference, 2018.
[2] P. Narayanamurthy and N. Vaswani, “A Fast and Memory-Efficient Algorithm for Robust PCA
(MERoP),” in IEEE Intl. Conf. Acoustics, Speech, Sig. Proc. (ICASSP), 2018.
[3] E. J. Candès, X. Li, Y. Ma, and J. Wright, “Robust principal component analysis?,” J. ACM, vol.
58, no. 3, 2011.
[4] J.P. Costeira and T. Kanade, “A multibody factorization method for independently moving objects,”
International Journal of Computer Vision, vol. 29, no. 3, pp. 159–179, 1998.
[5] G. S. Wagner and T. J. Owens, “Signal detection using multi-channel seismic data,” Bulletin of the
Seismological Society of America, vol. 86, no. 1A, pp. 221–231, 1996.
[6] A. Ozdemir, E. M. Bernat, and S. Aviyente, “Recursive tensor subspace tracking for dynamic brain
network analysis,” IEEE Transactions on Signal and Information Processing over Networks, 2017.
[7] V. Chandrasekaran, S. Sanghavi, P. A. Parrilo, and A. S. Willsky, “Rank-sparsity incoherence for
matrix decomposition,” SIAM Journal on Optimization, vol. 21, 2011.
[8] D. Hsu, S. M. Kakade, and T. Zhang, “Robust matrix decomposition with sparse corruptions,” IEEE
Trans. Info. Th., Nov. 2011.
[9] P. Netrapalli, U N Niranjan, S. Sanghavi, A. Anandkumar, and P. Jain, “Non-convex robust pca,”
in NIPS, 2014.
[10] X. Yi, D. Park, Y. Chen, and C. Caramanis, “Fast algorithms for robust pca via gradient descent,”
in NIPS, 2016.
[11] Y. Cherapanamjeri, K. Gupta, and P. Jain, “Nearly-optimal robust matrix completion,” ICML,
2016.
[12] C. Qiu, N. Vaswani, B. Lois, and L. Hogben, “Recursive robust pca or recursive sparse recovery in
large but structured noise,” IEEE Trans. Info. Th., pp. 5007–5039, August 2014.
[13] B. Lois and N. Vaswani, “Online matrix completion and online robust pca,” in IEEE Intl. Symp.
Info. Th. (ISIT), 2015.
[14] J. Zhan, B. Lois, H. Guo, and N. Vaswani, “Online (and Offline) Robust PCA: Novel Algorithms
and Performance Guarantees,” in Intnl. Conf. Artif. Intell. Stat. (AISTATS), 2016.
[15] P. Narayanamurthy and N. Vaswani, “Provable dynamic robust pca or robust subspace tracking,”
arXiv:1705.08948 (submitted to IEEE Trans. Info. Theory), 2017.
[16] B. Yang, “Projection approximation subspace tracking,” IEEE Trans. Sig. Proc., pp. 95–107, 1995.
28
[17] T. Adali and S. Haykin, Eds., Adaptive Signal Processing: Next Generation Solutions, Wiley & Sons,
2010.
[18] Y. Chi, Y. C. Eldar, and R. Calderbank, “Petrels: Parallel subspace estimation and tracking by
recursive least squares from partial observations,” IEEE Trans. Sig. Proc., December 2013.
[19] L. Balzano and S. Wright, “Local convergence of an algorithm for subspace identification from partial
data,” Found. Comput. Math., vol. 15, no. 5, 2015.
[20] K. Ye and L. H. Lim, “Schubert varieties and distances between subspaces of different dimensions,”
SIAM Journal on Matrix Analysis and Applications, vol. 37, no. 3, pp. 1176–1197, 2016.
[21] J. Feng, H. Xu, and S. Yan, “Online robust pca via stochastic optimization,” in NIPS, 2013.
[22] E. Candes, “The restricted isometry property and its implications for compressed sensing,” C. R.
Math. Acad. Sci. Paris Serie I, 2008.
[23] J. Zhan and N. Vaswani, “Robust pca with partial subspace knowledge,” IEEE Trans. Sig. Proc.,
July 2015.
[24] Lin Xiao and Tong Zhang, “A proximal-gradient homotopy method for the l1-regularized leastsquares problem,” in ICML, 2012.
[25] N. Vaswani and P. Narayanamurthy, “Finite sample guarantees for pca in non-isotropic and datadependent noise,” arXiv:1709.06255, 2017.
[26] Cameron Musco and Christopher Musco, “Randomized block krylov methods for stronger and faster
approximate singular value decomposition,” in Advances in Neural Information Processing Systems,
2015, pp. 1396–1404.
[27] N. Vaswani and H. Guo, “Correlated-pca: Principal components’ analysis when data and noise are
correlated,” in Adv. Neural Info. Proc. Sys. (NIPS), 2016.
[28] J. A. Tropp, “User-friendly tail bounds for sums of random matrices,” Found. Comput. Math., vol.
12, no. 4, 2012.
[29] R. Vershynin, “Introduction to the non-asymptotic analysis of random matrices,” Compressed
sensing, pp. 210–268, 2012.
[30] R. Horn and C. Johnson, Matrix Analysis, Cambridge Univ. Press, 1985.
[31] E. Keogh, K. Chakrabarti, M. Pazzani, and S. Mehrotra, “Dimensionality reduction for fast similarity
search in large time series databases,” Knowledge and information Systems, vol. 3, no. 3, pp. 263–286,
2001.
[32] J. C. Harsanyi and C. I. Chang, “Hyperspectral image classification and dimensionality reduction:
an orthogonal subspace projection approach,” IEEE Transactions on geoscience and remote sensing,
vol. 32, no. 4, pp. 779–785, 1994.
[33] E. J. Candes and B. Recht, “Exact matrix completion via convex optimization,” Found. of Comput.
Math, , no. 9, pp. 717–772, 2008.
[34] J. He, L. Balzano, and A. Szlam, “Incremental gradient on the grassmannian for online foreground
and background separation in subsampled video,” in IEEE Conf. on Comp. Vis. Pat. Rec. (CVPR),
2012.
29
[35] J. Yang and Y. Zhang, “Alternating direction algorithms for l1 problems in compressive sensing,”
Tech. Rep., Rice University, June 2010.
A
Proof of Theorem 2.2
We divide the proof into 3 parts for better clarity. We first prove the νt = 0 case for NORST (Algorithm
2), then prove the correctness of Offline NORST, and finally explain the changes needed when νt 6= 0.
A.1
Proof with νt = 0
Remark A.1 (Deriving the long expression for K given in the discussion). We have used max-outlier-frac-rowα ≤
b0 with b0 = 0.01/f 2 throughout the analysis in order to simplify the proof. If we were not to do this, and
√
if we used [25], it is possible to show that the “decay rate” qk is of the form qk = (c2 b0 f )k q0 from which
it follows that to obtain an ε-accurate approximation of the subspace it suffices to have
log (c1 ∆/ε)
√
K=
.
− log(c2 b0 f )
We first prove Theorem 2.2 for the case when tj ’s are known, i.e., correctness of Algorithm 1.
Proof of Theorem 2.2 with assuming tj known. In this case t̂j = tj . The proof is an easy consequence of
Lemmas 4.7 and 4.8. Recall that Γj,K ⊆ Γj,K−1 ⊆ · · · Γj,0 and ΓJ,K ⊆ ΓJ−1,K ⊆ · · · ⊆ Γ1,K . To show
that the conclusions of the Theorem hold, it suffices to show that Pr(ΓJ,K |Γ0,0 ) ≥ 1 − 10dn−10 . Using the
chain rule of probability,
Pr(ΓJ,K |Γ1,0 ) = Pr(ΓJ,K , ΓJ−1,K , · · · , Γ1,K |Γ1,0 )
=
J
Y
Pr(Γj,K |Γj,0 ) =
j=1
=
J Y
K
Y
J
Y
Pr(Γj,K , Γj,K−1 , · · · , Γj,1 |Γj,0 )
j=1
(a)
Pr(Γj,k |Γj,k−1 ) ≥ (1 − 10n−10 )JK ≥ 1 − 10JKn−10 .
j=1 k=1
where (a) used Pr(Γj,1 |Γj,0 ) ≥ 1 − 10n−10 from Lemma 4.7 and Pr(Γj,k |Γj,k−1 ) ≥ 1 − 10n−10 from Lemma
4.8.
Proof of Theorem 2.2. Define
t̂j−1,f in := t̂j−1 + Kα, tj,∗
tj − t̂j−1,f in
= t̂j−1,f in +
α
α
Thus, t̂j−1,f in is the time at which the (j − 1)-th subspace update is complete; w.h.p., this occurs before
tj . With this assumption, tj,∗ is such that tj lies in the interval [tj,∗ − α + 1, tj,∗ ].
Recall from the algorithm that we increment j to j + 1 at t = t̂j + Kα := t̂j,f in . Define the events
Ptj,∗
1. Det0 := {t̂j = tj,∗ } = {λmax ( α1 t=t
Φ`ˆt `ˆ0t Φ) > ωevals } and
j,∗ −α+1
P
tj,∗ +α
Det1 := {t̂j = tj,∗ + α} = {λmax ( 1
Φ`ˆt `ˆ0 Φ) > ωevals },
α
t
t=tj,∗ +1
2. SubUpd := ∩K
k=1 SubUpdk where SubUpdk := {SE(P̂j,k , Pj ) ≤ qk },
P
3. NoFalseDets := {for all J α ⊆ [t̂j,f in , tj+1 ), λmax ( α1 t∈J α Φ`ˆt `ˆ0t Φ) ≤ ωevals }
30
4. Γ0,end := {SE(P̂0 , P0 ) ≤ 0.25},
5. Γj,end := Γj−1,end ∩ (Det0 ∩ SubUpd ∩ NoFalseDets) ∪ (Det0 ∩ Det1 ∩ SubUpd ∩ NoFalseDets) .
Let p0 denote the probability that, conditioned on Γj−1,end , the change got detected at t = tj,∗ , i.e.,
let
p0 := Pr(Det0|Γj−1,end ).
Thus, Pr(Det0|Γj−1,end ) = 1 − p0 . It is not easy to bound p0 . However, as we will see, this will not be
needed. Assume that Γj−1,end ∩Det0 holds. Consider the interval J α := [tj,∗ , tj,∗ +α). This interval starts
at or after tj , so, for all t in this interval, the subspace has changed. For this interval, Φ = I − P̂j−1 P̂j−1 0 .
Applying the first item of Lemma 4.12, w.p. at least 1 − 10n−10 ,
!
1 X ˆ ˆ0
λmax
Φ`t `t Φ ≥ ωevals
α
α
t∈J
and thus t̂j = tj,∗ + α. In other words,
Pr(Det1|Γj−1,end ∩ Det0) ≥ 1 − 10n−10 .
Conditioned on Γj−1,end ∩ Det0 ∩ Det1, the first SVD step is done at t = t̂j + α = tj,∗ + 2α and the
subsequent steps are done every α samples. We can prove Lemma 4.7 with Γj,0 replaced by Γj,end ∩ Det0 ∩
Det1 and Lemma 4.8 with Γj,k−1 replaced by Γj,end ∩ Det0 ∩ Det1 ∩ SubUpd1 ∩ · · · ∩ SubUpdk−1 and with
the k-th SVD interval being Jk := [t̂j + (k − 1)α, t̂j + kα). Applying Lemmas 4.7, and 4.8 for each k, we
get
Pr(SubUpd|Γj−1,end ∩ Det0 ∩ Det1) ≥ (1 − 10n−10 )K+1 .
We can also do a similar thing for the case when the change is detected at tj,∗ , i.e. when Det0 holds. In
this case, we replace Γj,0 by Γj,end ∩ Det0 and Γj,k by Γj,end ∩ Det0 ∩ SubUpd1 ∩ · · · ∩ SubUpdk−1 and
conclude that
Pr(SubUpd|Γj−1,end ∩ Det0) ≥ (1 − 10n−10 )K .
Finally consider the NoFalseDets event. First, assume that Γj−1,end ∩ Det0 ∩ SubUpd holds. Consider
any interval J α ⊆ [t̂j,f in , tj+1 ). In this interval, P̂(t) = P̂j , Φ = I − P̂j P̂j 0 and SE(P̂j , Pj ) ≤ ε. Using the
second part of Lemma 4.12 we conclude that w.p. at least 1 − 10n−10 ,
!
1 X ˆ ˆ0
λmax
Φ`t `t Φ < ωevals
α
α
t∈J
t
−t
−Kα−α
Since Det0 holds, t̂j = tj,∗ . Thus, we have a total of b j+1 j,∗α
c intervals J α that are subsets of
tj+1 −tj,∗ −Kα−α
tj+1 −tj −Kα−α
tj+1 −tj
[t̂j,f in , tj+1 ). Moreover, b
c≤b
c ≤ b α c − (K + 1) since α ≤ α. Thus,
α
α
Pr(NoFalseDets|Γj−1,end ∩ Det0 ∩ SubUpd) ≥ (1 − 10n−10 )b
tj+1 −tj
α
c−(K)
On the other hand, if we condition on Γj−1,end ∩ Det0 ∩ Det1 ∩ SubUpd, then t̂j = tj,∗ + α. Thus,
Pr(NoFalseDets|Γj−1,end ∩ Det0 ∩ Det1 ∩ SubUpd) ≥ (1 − 10n−10 )b
tj+1 −tj
α
c−(K+1)
We can now combine the above facts to bound Pr(Γj,end |Γj−1,end ). Recall that p0 := Pr(Det0|Γj−1,end ).
Clearly, the events (Det0∩SubUpd∩NoFalseDets) and (Det0∩Det1∩SubUpd∩NoFalseDets) are disjoint.
31
Thus,
Pr(Γj,end |Γj−1,end )
= p0 Pr(SubUpd ∩ NoFalseDets|Γj−1,end ∩ Det0)
+ (1 − p0 ) Pr(Det1|Γj−1,end ∩ Det0) Pr(SubUpd ∩ NoFalseDets|Γj−1,end ∩ Det0 ∩ Det1)
≥ p0 (1 − 10n−10 )K (1 − 10n−10 )b
tj+1 −tj
α
c−(K)
+ (1 − p0 )(1 − 10n−10 )(1 − 10n−10 )K (1 − 10n−10 )b
= (1 − 10n−10 )b
tj+1 −tj
α
c
tj+1 −tj
α
c−(K+1)
≥ (1 − 10n−10 )tj+1 −tj .
Since the events Γj,end are nested, the above implies that
Pr(ΓJ,end |Γ0,end ) =
Y
Pr(Γj,end |Γj−1,end ) ≥
j
Y
(1 − 10n−10 )tj+1 −tj = (1 − 10n−10 )d
j
≥ 1 − 10dn−10 .
A.2
Proof of Offline NORST
We now provide the proof of the Offline Algorithm (lines 26-30 of Algorithm 2).
Proof of Offline NORST. The proof of this follows from the conclusions of the online counterpart. Note
that the subspace estimate in this case is not necessarily r dimensional. This is essentially done to ensure
that in the time intervals when the subspace has changed, but has not yet been updated, the output of the
algorithm is still an ε-approximate solution to the true subspace. In other words, for t ∈ [t̂j−1 + Kα, tj ],
the true subspace is Pj−1 and so in this interval
offline
SE(P̂(t)
, Pj−1 ) = SE([P̂j−1 , (I − P̂j−1 P̂j−1 0 )P̂j ], Pj−1 )
(a)
= [I − (I − P̂j−1 P̂j−1 0 )P̂j P̂j 0 (I − P̂j−1 P̂j−1 0 )][I − P̂j−1 P̂j−1 0 ]Pj−1
≤ [I − (I − P̂j−1 P̂j−1 0 )P̂j P̂j 0 (I − P̂j−1 P̂j−1 0 )] SE(P̂j−1 , Pj−1 ) ≤ ε
where (a) follows because for orthogonal matrices P1 and P2 ,
I − P1 P1 0 − P2 P2 0 = (I − P1 P1 0 )(I − P2 P2 0 ) = (I − P2 P2 0 )(I − P1 P1 0 )
Now consider the interval t ∈ [tj , t̂j + Kα]. In this interval, the true subspace is Pj and we have
back propagated the ε-approximate subspace P̂j in this interval. We first note that span([P̂j−1 , (I −
P̂j−1 P̂j−1 0 )P̂j ]) = span([P̂j , (I − P̂j P̂j 0 )P̂j−1 ]). And so we use the latter to quantify the error in this
interval as
offline
SE(P̂(t)
, Pj ) = SE([P̂j , (I − P̂j P̂j 0 )P̂j−1 ], Pj )
= [I − (I − P̂j P̂j 0 )P̂j−1 P̂j−1 0 (I − P̂j P̂j 0 )][I − P̂j P̂j 0 ]Pj
≤ [I − (I − P̂j P̂j 0 )P̂j−1 P̂j−1 0 (I − P̂j P̂j 0 )] SE(P̂j , Pj ) ≤ ε
32
A.3
Proof with νt 6= 0
In this section we analyze the “stable” version of RST, i.e., we let νt 6= 0.
Proof. The proof is very similar to that of the noiseless case but there are two differences due to the
additional noise term. The first is the effect of the noise on the sparse recovery step. The approach to
address this is straightforward. We note that the error now seen in the sparse recovery step is bounded by
kΨ(`t + νt )k and using the bound on kνt k, we observe
p that the error only changes by a constant factor.
In particular, we can show that ket k ≤ 2.4(2ε + ∆) ηrλ+ . The other crucial difference is in updating
subspace estimate. To deal with the additional uncorrelated noise, we use the following result.
Remark 4.18 of [25] states the following for the case where the data contains unstructured noise νt
2 +
that satisfies the assumptions of Theorem 2.2. Thus, in the notation of [25], λ+
v ≤ cε λ and rv = r. The
following result also assumes r, n large enough so that (r + log n) ≤ r log n.
Corollary A.2 (Noisy PCA-SDDN). Given data vectors yt := `t + wt + zt = `t + ITt Ms,t `t + zt ,
t = 1, 2, . . . , α, where Tt is the support set of wt , and `t satisfying the model detailed above. Furthermore,
maxt kMs,t P k2 ≤ q < 1. zt is small uncorrelated noise such that E[zt zt 0 ] = Σz , maxt kzt k2 := b2z < ∞.
2
+
Define λ+
z := λmax (Σz ) and rz as the “effective rank” of zt such that bz = rz λz . Then for any α ≥ α0 ,
where
C
b2z
2 2
α0 := 2 max ηq f r log n, − f log n
λ
εSE
the fraction of nonzeroes in any row of the noise matrix [w1 , w2 , . . . , wα ] is bounded by b, and
√
0.9εSE
−
3 bqf + λ+
z /λ ≤
1 + εSE
P
For an α ≥ α0 , let P̂ be the matrix of top r eigenvectors of D := α1 t yt yt0 . With probability at least
1 − 10n−10 , SE(P̂ , P ) ≤ εSE .
We illustrate how applying Corollary A.2 changes the subspace update step. Consider the first subspace estimate, i.e., we are trying to get an estimate P̂j,1 in the j-th subspace change time interval. Define
(e` )t = ITt 0 (ΨTt 0 ΨTt )−1 ΨTt 0 `t and (eν )t = ITt 0 (ΨTt 0 ΨTt )−1 ΨTt 0 νt . We estimate the new subspace, P̂j,1
Pt̂j +α−1 ˆ ˆ 0
as the top r eigenvectors of 1
`t `t . In the setting above, yt ≡ `ˆt , wt ≡ (e` )t , zt ≡ (eν )t , `t ≡ `t
α
t=t̂j
and Ms,t = − (ΨTt 0 ΨTt )−1 ΨTt 0 and so kMs,t P k = k (ΨTt 0 ΨTt )−1 ΨTt 0 Pj k ≤ φ+ (ε + SE(Pj−1 , Pj )) := q0 .
Applying Corollary A.2 with q ≡ q0 , and recalling that the support, Tt satisfies the assumptions similar
to that of the noiseless case and hence b0 ≡ max-outlier-frac-rowα . Now, setting εSE,1 = q0 /4, observe that
we require
p
0.5 · 0.9εSE,1
0.5 · 0.9εSE,1
λ+
z
, and, (ii) −
≤
.
(i) b0 q0 f ≤
1 + εSE,1
λ
1 + εSE,1
√
which holds if (i) b0 f ≤ 0.12, and (ii) is satisfied as follows from using the assumptions on νt as follows.
−
2
It is immediate to see that λ+
z /λ ≤ ε ≤ .2εSE,1 . Furthermore, the sample complexity term remains
unchanged due to the choice of νt . To see this, notice that the only extra term in the α0 expression
is b2z f log n/(ε2SE λ− ) which simplifies to ε2 f 2 r log n/ε2SE which is what was required even in the noiseless
case. Thus, from Corollary A.2, with probability at least 1 − 10n−10 , SE(P̂j,1 , Pj ) ≤ εSE,1 = q0 /4. The
argument in other subspace update stages will require the same changes and follows without any further
differences.
The final difference is in the subspace detection step. Notice that here too, in general, there will be
some extra assumption required to provably detect the subspace change. However, due to the bounds
assumed on kνt k and the bounds on using l,v = v,v = 0.01ε, we see that (i) the extra sample complexity
term is the same as that required in the noiseless case.
33
B
Proof of Lemmas 4.13 and 4.16
Proof of Lemma 4.13. The proof follows from triangle inequality as
SE(Q1 , Q3 ) = (I − Q1 Q1 0 )Q3 = (I − Q1 Q1 0 )(I − Q2 Q2 0 + Q2 Q2 0 )Q3
≤ (I − Q1 Q1 0 )(I − Q2 Q2 0 )Q3 + (I − Q1 Q1 0 )Q2 Q2 0 Q3
≤ (I − Q1 Q1 0 ) SE(Q2 , Q3 ) + SE(Q1 , Q2 ) Q2 0 Q3 ≤ ∆1 + ∆2
We need the following results for proving Lemma 4.16.
Theorem B.1 (Cauchy-Schwartz for sums of matrices [12]). For matrices X and Y we have
1X
Xt Yt 0
α t
2
≤
1X
Xt Xt 0
α t
1X
Yt Yt 0
α t
(10)
The following theorem is adapted from [28].
Theorem B.2 (Matrix Bernstein [28]). Given an α-length sequence of n1 × n2 dimensional random matrices and a r.v. X. Assume the following holds. For all X ∈ C, (i) conditioned on X, the matrices Zt are
P
P
mutually independent, (ii) P(kZt k ≤ R|X) = 1, and (iii) max α1 t E[Zt 0 Zt |X] , α1 t E[Zt Zt 0 |X] ≤
σ 2 . Then, for an > 0 and for all X ∈ C,
!
−α2
1X
1X
P
Zt ≤
E[Zt |X] + X ≥ 1 − (n1 + n2 ) exp
.
(11)
α t
α t
2 (σ 2 + R)
The following theorem is adapted from [29].
Theorem B.3 (Sub-Gaussian Rows [29]). Given an N -length sequence of sub-Gaussian random vectors
wi in Rnw , an r.v X, and a set C. Assume the following holds. For all X ∈ C, (i) wi are conditionally independent given X; (ii) the sub-Gaussian norm of wi is bounded by K for all i. Let W := [w1 , w2 , . . . , wN ]0 .
Then for an ∈ (0, 1) and for all X ∈ C
1
1 0
c2 N
0
P
W W − E W W |X ≤ X ≥ 1 − 2 exp nw log 9 −
.
(12)
N
N
4K 4
Proof of Lemma 4.16. The proof approach is similar to that of [15, Lemma 7.17] but the details are
different since we use a simpler subspace model.
√
Item 1 : Recall that the (at )i are bounded r.v.’s
at are
p satisfying |(at )i | ≤ ηλi . Thus, the vectors,p
sub-Gaussian with kat kψ2 = maxi k(at )i kψ2 = ηλ+ . We now apply Theorem B.3 with K ≡ ηλ+ ,
= 0 λ− , N ≡ α and nw ≡ r to conclude the following: For an α ≥ α(0) := C(r log 9 + 10 log n)f 2 ,
!
1X
−
0
≥ 1 − 10n−10
Pr
at at − Λ ≤ 0 λ
α t
The Lemma statement assumes α = Cf 2 r log n. For large r, n, this α > α(0) = Cf 2 (r + log n). Thus, the
above holds under the Lemma statement.
Item 2 : For the second term, we proceed as follows. Since kΦk = 1,
1X
1X
Φ`t et 0 Φ ≤
Φ`t et 0 .
α t
α t
34
To bound the RHS above, we will apply Theorem B.2 with Zt = Φ`t et 0 . Conditioned on {P̂∗ , Z}, the
Zt ’s are mutually independent. We first bound obtain a bound on the expected value of the time average
of the Zt ’s and then compute R and σ 2 . By Cauchy-Schwartz,
"
1X
E
Φ`t et 0
α t
2
#
1X
=
ΦP ΛP 0 M1,t 0 M2,t 0
α t
2
(a)
1X
ΦP ΛP 0 M1,t 0 M1,t P ΛP 0 Φ
α t
h
i
(b)
2
≤ b0 max ΦP ΛP 0 M1,t 0
≤
1X
M2,t M2,t 0
α t
t
2
≤ b0 SE (P̂ , P )q 2 (λ+ )2
(13)
where (a) follows by Cauchy-Schwartz (Theorem B.1) with Xt = ΦP ΛP 0 M1,t 0 and Yt = M2,t , (b) follows
from the assumption on M2,t . To compute R
kZt k ≤ kΦ`t k ket k ≤ SE(P̂ , P )qηrλ+ := R
Next we compute σ 2 . Since wt ’s are bounded r.v.’s, we have
1X
1X
1
E[Zt Zt 0 ] =
E Φ`t et 0 et `t 0 Φ =
E[ket k2 Φ`t `t 0 Φ]
α t
α t
α
1X
2
E Φ`t `t 0 Φ
≤ max ket k
et
α t
≤ q 2 SE2 (P̂ , P )ηr(λ+ )2 · := σ 2
it can also be seen that
1
α
P
Pr
0Z ]
t
evaluates to the same expression. Thus, applying Theorem B.2
!
p
1X
Φ`t et 0 ≤ SE(P̂ , P ) b0 qλ+ +
α t
−α
n
o .
≥ 1 − 2n exp
σ2 R
4 max 2 ,
t E[Zt
Let = 1 λ− , then, σ 2 /2 = cηf 2 r and R/ = cηf r. Hence, for the probability to be of the form 1 − 2n−10
we require that α ≥ α(1) := C · ηf 2 (r log n).
Item 3 : We use Theorem B.2 with Zt := Φet et 0 Φ. The proof is analogous to the previous item. First
we bound the norm of the expectation of the time average of Zt :
X
1X
1
0
E
Φet et Φ =
ΦM2,t M1,t P ΛP 0 M1,t 0 M2,t 0 Φ
α
α
1X
≤
M2,t M1,t P ΛP 0 M1,t 0 M2,t 0
α
!
h
i 1/2
(a)
1X
0
0
0 2
≤
M2,t M2,t max M2,t M1,t P ΛP M1,t (·)
t
α t
i p
(b) p h
≤ b0 max M1,t P ΛP 0 M1,t 0 M2,t 0 ≤ b0 q 2 λ+
t
35
where (a) follows from Theorem B.1 with Xt = M2,t and Yt = M1,t P ΛP 0 M1,t 0 M2,t 0 and (b) follows
from the assumption on M2,t . To obtain R,
kZt k = Φet et 0 Φ ≤ max kΦMt P at k2 ≤ q 2 rηλ+ := R
t
To obtain σ 2 ,
i
1X
1X h
E Φet (Φet )0 (Φet )et 0 Φ =
E Φet et 0 Φ kΦet k2
α t
α t
2
ΦMt P ΛP 0 Mt 0 Φ
≤ max kΦet k
et
2
≤ q rηλ+ · q 2 λ+ := σ 2
Applying Theorem B.2, we have
Pr
!
p
1X
−α2
0
2 +
Φet et Φ ≤ b0 q λ + ≥ 1 − n exp
α t
2(σ 2 + R)
Letting = 2 λ− we get R/ = cηrf and σ 2 /2 = cηrf 2 . For the success probability to be of the form
1 − 2n−10 we require α ≥ α(2) := Cη · 11f 2 (r log n).
The proof of the last two items follow from using [25, Lemma 7.19].
C
Proof of Extensions
In this section we present the proof of the extensions stated in Sec. 5.
C.1
Static Robust PCA
The proof follows directly from Theorem 2.2 by setting J = 1.
C.2
Subspace Tracking with missing data and dynamic Matrix Completion
Here we present the proof of the subspace tracking with missing data problem. The only changes needed
for this proof are in the initialization step, i.e., for j = 0. For this we use the following lemma.
Lemma C.1 (Lemma 2.1, [33]). Set r̄ = max(r, log n). Then there
exist constants C and c such
that
2
the random orthogonal model with left singular vectors P̂init obeys Pr maxi Ii 0 P̂init ≤ C r̄/n ≥ 1 −
cn−β log n.
Thus,
2
0
Pr max Ii P̂init
i
≤ µr̄/n ≥ 1 − n−10
Consider the two scenarios (i) if r ≥ log n, then everything discussed before remains true, whereas, if (ii)
0.01
r ≤ log n, we redefine µ2 = C log n/r and thus in the interval [t0 , t1 ] we require max-outlier-frac-col ≤ log
n.
Further, using the bound on max-outlier-frac-col it follows from triangle inequality that
max IT 0 P̂init
T ≤2s
2
≤ 2s max Ii 0 P̂init
i
36
2
≤
2sµr
< 0.01
n
We only mention the changes needed for Lemma 4.7 for when j = 0 since the initialization is different.
The rest of the argument of recursively applying the lemmas hold exactly as before. First, ttrain = 1 since
we use random initialization. Thus from the Algorithm, t̂0 = ttrain = 1.
Proof. Proof of item 1. Since the support of xt is known, the LS step gives
−1
−1
x̂t = ITt ΨTt 0 ΨTt
ΨTt 0 (Ψ`t + Ψxt ) = ITt ΨTt 0 ΨTt
ITt 0 Ψ`t + xt
Thus et = x̂t − xt satisfies
et = ITt ΨTt 0 ΨTt
−1
ITt 0 Ψ`t
Now, from the incoherence assumption on P̂init , Lemma 4.11, the bound on max-outlier-frac-col, and
recalling that in this interval, Ψ = I − P̂init P̂init 0 we have
2sµr
≤ 0.09 =⇒ δ2s (Ψ) ≤ 0.32 < 0.15,
n
1
1
1
≤
≤
< 1.2 = φ+ .
≤
1 − δs (Ψ)
1 − δ2s (Ψ)
1 − 0.15
max IT 0 P̂init
|T |≤2s
Ψ Tt 0 Ψ Tt
−1
2
≤
Secondly,
ITt 0 ΨP0 ≤
ITt 0 P0 + ITt 0 P̂init
≤ 0.3 + 0.3 = 0.6
(14)
p
Thus, kITt 0 Ψ`t k ≤ 0.6 ηrλ+ .
Proof of Item 2 : Since `ˆt = `t − et with et satisfying the above equation, updating P̂(t) from the `ˆt ’s
is a problem of PCA in sparse data-dependent noise (SDDN), et . To analyze this, we use the PCA-SDDN
result of Theorem 4.14 (this is taken from [25]). Recall from above that, for t ∈ [t̂0 , t̂0 + α], T̂t = Tt , and
`ˆt = `t − et . Recall from the algorithm that we compute the first estimate of the j-th subspace, P̂j,1 ,
Pt0 +α−1 ˆ ˆ 0
as the top r eigenvectors of α1 t=t
`t `t . In the notation of Theorem 4.14, yt ≡ `ˆt , wt ≡ et , `t ≡ `t
0
−1
0
0
and Ms,t = − (ΨTt ΨTt ) ΨTt and so kMs,t P k = k (ΨTt 0 ΨTt )−1 ΨTt 0 P0 k ≤ φ+ · 0.6 = 0.72 := q0 . This
follows using (14). Also, b0 ≡ max-outlier-frac-rowα .
Applying Theorem 4.14 with q ≡ q0 , b0 ≡ max-outlier-frac-rowα and setting εSE = q0 /4, observe that
we require
p
0.9(q0 /4)
b0 q0 f ≤
.
1 + (q0 /4)
√
This holds if b0 f ≤ 0.12 as provided by Theorem 2.2. Thus, from Corollary 4.14, with probability at
least 1 − 10n−10 , SE(P̂j,1 , P0 ) ≤ q0 /4.
C.3
Fewer than r directions change
Proof of Corollary 5.20. The proof is very similar to that of Theorem 2.2. The only changes occur
in the
at,fix
(a) Projected CS step. With the subspace change model, we define `t = P(t) at := Pj−1,fix Pj,chd
at,chd
where at,fix is a (r − rch ) × 1 dimensional vector and at,chd is a rch × 1 dimensional vector. In the first α
frames after the j-th subspace changes (or, the j-th subspace change is detected in the automatic case),
recall that P̂(t) = P̂j−1 . Thus, SE(P̂(t) , Pj−1,fix ) = SE(P̂j−1 , Pj−1,fix ) ≤ SE(P̂j−1 , Pj−1 ) ≤ ε and so, the
error can be bounded as
q
p
kΨ`t k ≤ kΨPj−1,fix at,fix k + kΨPj,chd at,chd k ≤ ε η(r − rch )λ+ + (ε + SE(Pj−1 , Pj )) ηrch λ+
ch
37
In the second α frames, have P̂(t) = basis(P̂j−1 , P̂j,1 ). Thus SE(P̂(t) , Pj−1,fix ) ≤ SE(P̂j−1 , Pj−1,fix ) ≤
SE(P̂j−1 , Pj−1 ) ≤ ε and SE(P̂(t) , Pj,chd ) ≤ SE(P̂j,1 , Pj,chd ) ≤ SE(P̂j,1 , Pj ) ≤ 0.3 · (ε + SE(Pj−1 , Pj )).
Thus, the error in the sparse recovery step in the interval after the first subspace update is performed is
given as
q
p
kΨ`t k ≤ ε η(r − rch )λ+ + 0.3 · (ε + SE(Pj−1 , Pj )) ηrch λ+
ch
The rest of the proof follows as before. The error after the k-th subspace update is also bounded using
the above idea.
(b) Subspace Detection step: The proof of the subspace detection step follows exactly analogous to
Lemma 4.12. One minor observation is noting that SE(Pj−1 , PJ ) = SE(Pj−1,ch , Pj,chd ) in the proof of
part (a) of Lemma 4.12. The rest of the argument is exactly the same.
38
| 7cs.IT
|
Private and Secure Coordination of
Match-Making for
Heavy-Duty Vehicle Platooning ?
Farhad Farokhi ∗ Iman Shames ∗ Karl H. Johansson ∗∗
∗
arXiv:1702.08582v1 [cs.SY] 28 Feb 2017
Department of Electrical and Electronic Engineering,
University of Melbourne, Parkville, VIC 3010, Australia
(e-mails: {ffarokhi,ishames}@unimelb.edu.au)
∗∗
ACCESS Linnaeus Center, School of Electrical Engineering,
KTH Royal Institute of Technology, SE-100 44 Stockholm, Sweden
(e-mail: [email protected])
Abstract: A secure and private framework for inter-agent communication and coordination is
developed. This allows an agent, in our case a fleet owner, to ask questions or submit queries in
an encrypted fashion using semi-homomorphic encryption. The submitted query can be about
the interest of the other fleet owners for using a road at a specific time of the day, for instance, for
the purpose of collaborative vehicle platooning. The other agents can then provide appropriate
responses without knowing the content of the questions or the queries. Strong privacy and
security guarantees are provided for the agent who is submitting the queries. It is also shown
that the amount of the information that this agent can extract from the other agent is bounded.
In fact, with submitting one query, a sophisticated agent can at most extract the answer to
two queries. This secure communication platform is used subsequently to develop a distributed
coordination mechanisms among fleet owners.
Keywords: Privacy; Security; Coordination; Homomorphic Encryption; Vehicle Platooning.
1. INTRODUCTION
The advances in communication technology has created
new opportunities regarding shared economy. An example
is collaborative driving, such as heavy-duty vehicle platooning or ride sharing, with the aim of reducing costs or
carbon footprint of commuters or fleets (Besselink et al.,
2016). This has motivated studies for creating appropriate
incentives for the fleet owners to collaborate (Farokhi
et al., 2015). Although promising, heavy-duty vehicle platooning has not yet been adopted by the larger population
of vehicles. Beside technological and legal barriers, this
could be partly motivated by that commercial, often competing, entities are unwilling to share their entire private
data, e.g., the routes of their vehicles and their travel
times, with each other even if doing so results in lower operative costs. This could be because of privacy constraints
by the customers or secretive nature of marketing agencies.
Therefore, it is of interest to create private and secure
match-making services for effective coordination among
competing entities to facilitate these new technologies. It
goes without saying that the use for such services is not
limited to vehicle platooning and can be justified in many
other setups, such as ride sharing, collaborative logistics,
energy markets, and even online dating services.
? The work of F. Farokhi was supported by a McKenzie Fellowship
and an Early Career Grant from Melbourne School of Engineering.
The work of K. H. Johansson was supported by Knut och Alice
Wallenbergs Foundation (KAW), Swedish Foundation for Strategic
Research (SSF), and Swedish Research Council (VR).
Following this motivation, in this paper, we create a secure
and private framework for communication between two
agents, fleet owners within the context of the heavy-duty
vehicle platooning. In the presented framework, it is possible for an agent to ask a question or submit a simple query
regarding the interest of the other agents about using a
road at a specific time of the day (for forming platoons) in
an encrypted fashion. This is done in such a way that the
other agents can provide their responses without knowing
the content of the questions or the queries. The framework
is developed with the aid of semi-homomorphic encryption,
which allows algebraic manipulation of the plain data
without the need for decryption using appropriate computations over the encrypted data; see, e.g., (Yi et al., 2014a)
about homomorphic encryption. This category of encryption techniques makes it possible for the second agent
(i.e., the fleet owner receiving the encrypted question)
to respond to it using appropriate manipulations of the
encrypted data. Strong privacy and security guarantees
are provided for the agent who submits the query. It is
also shown that the amount of the information that the
questioning agent can extract from all the other agents
is bounded (in fact with one question a fleet owner can
at most extract the answer to two questions about the
interests of the other fleet owners, which could be negligible in comparison to the number of possible questions).
This secure communication platform is used subsequently
to develop distributed coordination mechanisms for heavyduty vehicle platooning. Note that although the platform
is developed in the context of match-making for heavy-
duty vehicle platooning, the outcomes are more general
and can be readily used in other examples as well (specifically if the questions or the queries are regarding selecting
an element of a finite discrete set).
This paper is in essence close to the problem of private
searching on streaming data (Ostrovsky and Skeith, 2007;
Yi and Bertino, 2011; Yi et al., 2014b; Boneh and Waters,
2007), where it is of interest to determine if certain
important keywords are used within private encrypted
messages, such as text messages and e-mails, while the
content of the messages itself is not of special interest
(at least not if the keywords are not utilized). However,
it should be noted that the privacy guarantees of those
studies are one sided (with the aim of protecting the
privacy of the messages). However, in this paper, we would
like to provided guarantees to both sides, i.e., both the
agents posing the question and the ones responding to it.
Recently, homomorphic encryption have been utilized to
solve security and privacy issues in networked control and
estimation (Farokhi et al., 2016; Kogiso and Fujita, 2015).
Those studies are however involved with the intricacies
of dynamical systems and using encryption for closing the
control loop rather than creating a private/secure communication framework among multiple (possibly competing)
agents. Further, they do not provide two-sided privacy
guarantees (as it is not of special interest to preserve the
privacy of a malicious agent if it communicates with the
controller).
The rest of the paper is organized as follows. First,
the problem formulation and some necessary background
material are presented in Section 2. A secure and private
communication framework between two agents is provided
in Section 3. Section 4 uses the framework to construct
a distributed mechanism for coordination among several
fleet owners. Finally, numerical examples are presented in
Section 5 and the paper is concluded in Section 6.
2. PROBLEM FORMULATION
In this paper, the problem formulation and the results are
provided within the context of coordination among fleet
owners for heavy-duty vehicle platooning. However, the
results can be readily used in other classes of problems involving match-making and coordination, such as ride sharing and electricity markets. Investigating the platooning
coordination problem allows us to pose concrete questions
and provide meaningful privacy and security guarantees.
Consider the case with F fleet owners, each owning Ii ,
i ∈ F := {1, . . . , F }, heavy-duty vehicles. These vehicles
can operate over various roads on a transportation network
in set P (based on the requests of the customers of the fleet
owners) and at various time intervals of the day (based on
logistical constraints and customer preferences) in set T
such that |T | < ∞. Here, we discretize the time of the
day (e.g., half an hour windows of the day) so that the
number of time windows of interest is finite. We use the
set of integers W := {1, . . . , |P||T |} to refer to all the
possible combinations of roads and time windows (in which
a fleet owner might be interested or one of its heavy-duty
vehicles might be operating) captured by the entries of the
product set P × T . In fact, it can be said that the set W
is isomorph to the set P × T . Our goal is to develop a
secure and private communication framework for the fleet
owners to identify potential heavy-duty vehicles that can
form platoons. This is clearly possible if the fleet owners
exchange the time and the roads over which all their heavyduty vehicles operate; however, that would violate the
privacy of the customers and the fleet owners (possibly
jeopardizing their competitive advantages). Therefore, we
would like to create a communication platform so that the
fleet owners can only enquire about each other’s interests
and logistical constraints under strict privacy constraints.
In fact, we show that the enquiring fleet owner does not
leak any private information (i.e., the other fleet owners
cannot realize the road and the time window of interest
of the enquiring fleet owner). In addition, we show that
even with most sophisticated manoeuvres the enquiring
fleet owner can only extract information on the interests
of the other fleet owners about at most two pairs of
roads and time windows (which considering the sheer
number of roads and time windows is negligible). To do
so, we use the Paillier’s encryption method, which is a
semi-homomorphic encryption technique. The encryption
method is introduced in the following subsection.
2.1 Background Material
In this subsection, the Paillier’s encryption technique
is briefly introduced (Paillier, 1999). The method (or
rather its security) relies on the Decisional Composite
Residuosity Assumption, i.e., for given integers N ∈ Z
and x ∈ ZN 2 , it is “hard” to decide whether there exists
y ∈ ZN 2 such that x ≡ y N mod N . Here, the notation
ZN denotes the set of integers modulo N for all N ∈
N. More information regarding the assumption can be
found in (Paillier, 1999; Yi et al., 2014a). The encryption
scheme is as follows. First the public and private keys are
generated. To do so, large prime numbers p and q are
selected randomly and independently of each other such
that gcd(pq, (1 − p)(1 − q)) = 1, where gcd(a, b) refers
to the greatest common divisor of a and b. The public
key (which is shared with all the parties and is used for
encryption) is N = pq. The private key (which is only
available to the entity that needs to decrypt the data) is
(λ, µ) with λ = lcm(p−1, q−1) and µ = λ−1 mod N , where
lcm(a, b) is the least common multiple of a and b. The
ciphertext of a plain message t ∈ ZN can be constructed
using E(t; r) = (N + 1)t rN mod N 2 , where r is randomly
selected with uniform probability from Z∗N := {x ∈
ZN | gcd(x, N ) = 1}. Finally, to decrypt any ciphertext
c ∈ ZN 2 , we may use D(c) = L(cλ mod N 2 )µ mod N ,
where L(x) = (x − 1)/N . The correctness of the Paillier’s
encryption technique implies that D(E(t; r)) = t for all
r ∈ Z∗N and all t ∈ ZN (Paillier, 1999). The following
important property shows that the Paillier’s encryption
is a semi-homomorphic encryption scheme, i.e., algebraic
manipulation of the plain data is possible without the need
for decryption using appropriate computations over the
encrypted data.
Proposition 1. The following identities hold:
(1) For all r, r0 ∈ Z∗N and t, t0 ∈ ZN such that t + t0 ∈ ZN ,
E(t; r)E(t0 ; r0 ) mod N 2 = E(t + t0 ; rr0 );
(2) For all r ∈ Z∗N and t, t0 ∈ ZN such that tt0 ∈ ZN ,
0
0
E(t; r)t mod N 2 = E(t0 t; rt ).
Algorithm 1 Procedure SubmitQuery for the first fleet
owner.
1: procedure SubmitQuery(w)
2:
# Computed by the first fleet owner
3:
for i ∈ W do
4:
if i = w then
5:
xi ← E(1; ri )
6:
else
7:
xi ← E(0; ri )
8:
end if
9:
end for
10:
return x
11: end procedure
Proof. The proof follows from simple algebraic manipulations and can be found in (Paillier, 1999).
Proposition 1 shows that summation and multiplication
can be performed on the encrypted data. This is used
subsequently for creating a secure and private method for
coordination between fleet owners for heavy-duty vehicle
platooning.
3. SECURE AND PRIVATE COMMUNICATION
FRAMEWORK
In this section, we restrict ourselves to two fleet owners. We
subsequently generalize the setup to develop a distributed
coordination mechanism. Let us assume that the first fleet
owner would like to know if the second fleet owner has any
heavy-duty vehicle that uses the path and the time of the
day associated with w ∈ W. However, it does not wish
to let the second fleet owner to know w perhaps due to
the competitive nature of the fleet owners or the privacy
constraints imposed by the end user. Therefore, the first
fleet owner construct an encrypted vector x ∈ Z|W| such
that the i-th element of x is given by
E(1; ri ), i = w,
xi =
E(0; ri ), otherwise.
Note that the presence of the random element ri ensures
that with a high probability cyphertexts associated with 0
are different 1 . Then the first fleet owner transmits x to the
second fleet owner. The second fleet owner computes the
set W ⊆ W as the set of all w0 ∈ W with which at least
one of its heavy-duty vehicles is associated. The second
fleet owner then computes
Y
v
y=
xj j mod N 2 mod N 2 ,
j∈W
where vj is randomly selected from the set {1, . . . , N − 1}
with a uniform probability. We summarize the procedures
of the first and the second fleet owner in Algorithms 1
and 2. We can prove the following useful result pointing
to the correctness of our proposed methodology.
Proposition 2. If the first and the second fleet owner,
respectively, use Algorithms 1 and 2, then D(y) 6= 0 if
any of the heavy-duty vehicles owned by the second fleet
1 In fact, for industry standards that assume both q and q are of the
order of 1024 bits, we have N = O(22048 ). Thus the probability of
selecting the same r twice even in a vector of millions of elements is
smaller than 10−1000 , which is practically zero.
Algorithm 2 Procedure ReturnResponse for the first
fleet owner.
1: procedure SubmitQuery(x,W)
2:
# Computed by the second fleet owner
3:
y←1
4:
for i ∈ W do
5:
Select vi randomly from {1, . . . , N − 1}
6:
y ← y(xvi i mod N 2 ) mod N 2
7:
end for
8:
return y
9: end procedure
owner uses the path and time window associated with w
and D(y) = 0 otherwise.
Proof. The proof follows from the construction of the
vector x and the application of Proposition 1.
To be able to assess the security and privacy of the
proposed method from the perspective of the first fleet
owner, we present the following definition. This notion is
derived from what is known in the encryption literature
as semantic security. To present this definition, we need
to define the notion of negligible functions. A function f :
N → R≥0 is called negligible if, for any c ∈ N, there exists
nc ∈ N such that f (n) ≤ 1/nc for all n ≥ nc (Ostrovsky
and Skeith, 2007).
Definition 1. Let the second fleet owner propose w1 , w2 ∈
W. The first fleet owner chooses at random w from
{w1 , w2 } with equal probability and sends x constructed
using Algorithm 1. The second fleet owner can based
on its knowledge of x guess w. This guess is denoted
by w0 . The second fleet owner’s advantage 2 is given by
Adv(k) := |P{w = w0 }−1/2|, where k denotes the security
parameter, e.g., min(p, q) in the Paillier’s technique. The
proposed strategy is defined to be secure and private if
Adv is negligible.
Definition 1, although long and cumbersome, says that the
proposed strategy is secure and private if, essentially, the
performance of the second fleet owner in guessing the first
fleet owner’s preference w is not better than a pure random
number generator.
Proposition 3. Under the Decisional Composite Residuosity Assumption, Algorithm 1 is secure and private in the
sense of Definition 1.
Proof. The proof follows from the application of the
results of (Paillier, 1999).
Unfortunately, the privacy and security guarantees are
weaker for the second fleet owner. This is because by definition if the first and the second fleet owner, respectively, use
Algorithms 1 and 2, the first fleet owner can successfully
determine if the second fleet owner is in possession of
any heavy-duty vehicles that operate on the path and in
the time window associated with w (therefore some private information is leaked even in the best of situations).
However, one might be able to argue that the first fleet
owner can potentially extract more information by not
following Algorithm 1. This is in fact true (pointing to
2
The advantage refers to its superior performance to that of a “dart
throwing monkey.”
a deeper erosion of privacy). However, in what follows, we
prove that the amount of information the first fleet owner
can optimally extract is limited (with the bound being
extremely small). Assume that the first fleet owner does
not use Algorithm 1. Instead, it constructs the vector x
such that xi = E(x̃i ; ri ) for some integer x̃i ∈ ZN . We can
prove the following lemma.
Lemma 1. If, for all i, xi = E(x̃i ; ri ) for some integer
x̃i ∈ ZN , then
X
D(y) =
x̃i vi zi mod N,
(1)
w∈W
where zi = 1 if any of the heavy-duty vehicles owned by
the second fleet owner uses the path and the time window
associated with i ∈ W and zi otherwise.
Proof. The proof follows from the construction of the
vector x, x̃i , ∀i, and the application of Proposition 1.
Following
P Lemma 1, the first fleet owner needs to solve
D(y) = w∈W x̃i vi zi mod N to find zi for all i. Note that
the first fleet owner can introduce the change of variable
ξi = vi zP
i and instead solve the linear equation modulo
D(y) = w∈W x̃i ξi mod N . Evidently, zi = 1 if ξi 6= 0 and
zi = 0 otherwise. Let us form the set
(
)
X
|W|
Ξ := ξ ∈ ZN D(y) =
x̃i ξi mod N ,
(2)
w∈W
which represents the set ofPall solutions of the linear
equation modulo D(y) =
w∈W x̃i ξi mod N . We can
prove the following useful result.
Proposition 4. Let t = |{i | x̃i 6= 0}| > 1. Then |Ξ| ≥ (N −
1)t−1 if there exists i such that gcd(x̃i , N ) = 1.
Proof. For i such that gcd(x̃i , N ) = 1, there exists
x̃−1
i mod N . Thus we get
X
ξi = D(y) −
x̃−1
x̃
ξ
j j mod N.
i
j6=i
Therefore, all (ξj )j6=i are free variables, i.e., for any selection of (ξj )j6=i , there exists
P ξi that satisfies the linear
equation modulo D(y) = w∈W x̃i ξi mod N . This points
to that the number of solutions of the linear equation
modulo (which is equal to |Ξ|) is equal to the number of
all the possible choices of (ξj )j6=i .
This shows that even if only two x̃iPare non-zero, the
number of possible solutions of D(y) = w∈W x̃i ξi mod N ,
i.e., |Ξ|, is larger than N − 1. The situation worsens as
more x̃i become non-zero because t (the number of the
non-zero elements) appear as an exponent. Noting that
the encryption relies on N being extremely large 3 , the
first fleet owner needs to check a huge number of possible
solutions, which is numerically impractical.
Proposition 5. Let t = |{i | x̃i 6= 0}| > 2. Then |Ξ| ≥
2(N −1)t−2 if there does not exist i such that gcd(x̃i , N ) =
1.
Proof. If there does not exist i such that gcd(x̃i , N ) = 1,
we can construct two sets where in the first one x̃i is
divisible by q and in the second one x̃i is divisible by p
3
Routinely, p, q are selected as prime numbers with the length of
1024 bits pointing to that N = O(22048 ).
(note that x̃i cannot be divisible by both as otherwise it
will be larger than lcm(p, q) = pq = N ). Let the sets be
denoted by J1 and J2 , respectively. In this case, we can
write
X
X
x̃j
x̃j
ξj
D(y) = q
ξj
+p
mod N.
q
p
j∈J1
j∈J
2
| {z }
| {z }
x̃0j
x̃0j
Noting that gcd(p, q) = 1 (since p and q are prime
numbers), this equation can be separated into
X
ξj x̃0j mod N,
(3a)
α=
j∈J1
β=
X
ξj x̃0j mod N,
(3b)
j∈J2
where α = D(y)ᾱ and β = D(y)β̄ with ᾱ and β̄ denoting
Bézout coefficients, i.e., ᾱq + β̄p = 1. There are only two
Bézout coefficients that satisfy |ᾱ| < p and |β̄| < q (Vialar,
2015, Proposition 13, p. 60). The number of the solutions
of (3a) can be lower bounded with the same line of
reasoning as in Proposition 4 by (N − 1)|J1 |−1 . Similarly,
the number of the solutions of (3b) can be lower bounded
by (N − 1)|J2 |−1 . This concludes the proof.
Proposition 5 shows that, by smart planning and sophisticated manoeuvres, the first fleet owner can realize if the
second fleet owner has any heavy-duty vehicles on the
roads and the time windows associated with two entries
of W instead of one by following Algorithm 1. Assuming
that |W| is large, this might not matter in practice.
Remark 1. (Computational Complexity). Before moving
to more complex communication structures, the computational complexity of the proposed algorithms should be
discussed. Algorithm 1 involves |W| encryption operations.
Each encryption operation has a cost that is a (nonlinear)
function of N . Noting that the size of N is often a constant
set by standards of the industry, the cost of the encryption
is also constant (albeit a large one). Therefore, the computational complexity of Algorithm 1 is O(|W|). Algorithm 2,
on the other hand, involves |W| ≤ maxi Ii exponentiations and multiplications. Therefore, the computational
complexity of Algorithm 2 is O(maxi Ii ). Therefore, the
total computational complexity of Algorithms 1 and 2
is O(|W|) because, in practice, maxi Ii |W|. Finally,
since the first fleet owner needs to submit O(maxi Ii )
questions or queries to solve the platooning coordination
problem, the computational complexity of the whole task
is O(|W| maxi Ii ).
4. DISTRIBUTED COORDINATION
Now, we use the results of the previous section to develop
a distributed mechanism for the fleet owners to coordinate
their efforts. Let the undirected graph G with the vertex
set F (i.e., the vertices are the fleet owners) and the
edge set E ⊆ F × F capture the communication structure
among the agents. A walk over G (not to be mistaken with
roads over which the heavy-duty vehicles travel on real
transportation network) is a sequence of (not necessarily
unique) vertices L = (v0 , . . . , vk ) such that (vi , vi+1 ) ∈ E
for all 0 ≤ i ≤ k − 1. A k-connected graph is a graph that
after removing any k vertices (and all the edges connected
Algorithm 3 Procedure DistResponse for the fleet
owners in the walk L responding to the query of the fleet
owner ` distributedly.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
procedure DistResponse(x,L,(Wj )j∈L\{`} )
# Computed by the fleet owners in L except i
for j = v1 , . . . , vk−1 do
for i ∈ W j do
Select ωi randomly in {1, . . . , bN/(|L|−2)c}
2
i
xi ← xω
i mod N
end for
end for
y←1
for i ∈ W vk do
Select ωi randomly in {1, . . . , bN/(|L| − 2)c}
2
2
i
y ← y(xω
i mod N ) mod N
end for
return y
end procedure
focus on the privacy guarantees of the other fleet owners
in L \ {`}. Now, we can prove the following lemma.
Lemma 2. If, for all i, xi = E(x̃i ; ri ) for some integer
x̃i ∈ ZN , then
X
X
D(y) =
x̃i vi
zij
mod N,
(4)
w∈W
= 1 if any of the heavy-duty vehicles owned by
where
fleet owner j ∈ L \ {`} uses the path and the time window
associated with i ∈ W and zij = 0 otherwise.
Proof. The proof is similar to that of Lemma 1.
Assumption 1 states that even if one of the fleet owners
is removed, all the remaining ones can still communicate
with each other successfully. This allows us to develop an
algorithm for the fleet owners to collaboratively respond to
queries by avoiding communication with fleet owners that
have submitted the queries. Let us consider the case where
fleet owner ` ∈ F would like to find out if there exists any
other fleet owner such that one of its heavy-duty vehicles
operate over the path and the time window associated with
w ∈ W. Further, let there be a walk L = (v0 , . . . , vk ) over
G such that v0 = vk = ` (therefore the walk is a loop)
while vj 6= ` for all 1 ≤ j ≤ k − 1. Existence of such
a walk is guaranteed by Assumption 1. Similar, to the
previous section, fleet owner i can follow Algorithm 1 to
construct the vector x, thus submitting an encrypted query
for coordination. After that all the other fleet owners in
the walk L can follow the procedure in Algorithm 3 to
respond to the query of fleet owner `. In this algorithm,
Wj ⊆ W denotes the set of all w ∈ W with which at
least one of its heavy-duty vehicles of fleet owner j ∈ F
is associated. The following proposition shows that if fleet
owner ` uses Algorithm 1 and all the fleet owners in L\{`}
use Algorithm 3, the provided response is correct.
Proposition 6. If fleet owner ` uses Algorithm 1 and all
the fleet owners in L \ {`} use Algorithm 3, then D(y) 6= 0
if any of the heavy-duty vehicles owned by the fleet owner
in L \ {`} uses the path and time window associated with
w, and D(y) = 0 otherwise.
Proof. The proof follows from the application of Proposition 1.
Note that, if fleet owner ` is interested in figuring out the
possibility of forming a platoon with all the other fleet
owners (and not a select few), it should find a walk L that
spans all the vertices of the graph.
A similar result as in Proposition 3 can be proved for the
enquiring fleet owner in this case as well. Therefore, we
Similarly, following Lemma 2, the enquiring fleet owner `
must solve the linear equation modulo
X X
D(y) =
x̃i ξij mod N,
w∈W j∈L\{`}
6= 0 and zij = 0 otherwise. Let us
= 1 if
where
construct the set of all possibilities
|W|(|L|−2)
Ξ := (ξij )j∈L\{`} ∈ ZN
X X
j
D(y) =
x̃i ξi mod N . (5)
zij
to those vertices) is still a connected graph. We make the
following standing assumption.
Assumption 1. G is 2-connected.
j∈L\{`}
zij
ξij
w∈W j∈L\{`}
We can prove the following useful result regarding the size
of the set Ξ.
Proposition 7. The following two statements hold:
• Let t = |{i | x̃i 6= 0}| > 1. Then |Ξ| ≥ (|L| − 2)(N −
1)t−1 if there exists i such that gcd(x̃i , N ) = 1.
• Let t = |{i | x̃i 6= 0}| > 2. Then |Ξ| ≥ 2(|L| −
2)2 (N − 1)t−2 if there does not exist i such that
gcd(x̃i , N ) = 1.
Proof. The proof follows a similar line of reasoning as in
Propositions 4 and 5.
This proposition shows that the privacy guarantees of the
fleet owners of the walk is stronger than those in the case
of two agents as the responses of all the agents gets mixed.
Therefore, even if the fleet owner can extract the aggregate
answers to two questions, it would not know which one
of the fleet owners from the set L \ {`} has responded
positively. In practice, however, when the fleet owners
show up for forming a platoon the set of all possibilities
gets further narrowed down.
5. NUMERICAL EXAMPLE
In this section, we review some practical aspects of the
developed framework. Specifically, we review the computation time and the communication burden associated with
executing the proposed algorithms.
Let us consider an example, where two fleet owners need to
coordinate their efforts for organizing heavy-duty vehicle
platoons. Assume that there are |P| = 10 roads and time
of the day is discretized into |T | = 24 one-hour windows. Therefore, |W| = 240. The second fleet owner has
heavy-duty vehicles on roads and time windows associated
with w = 1, 6, 21, 50. Figure 1 illustrates the outcome
of Algorithms 1 and 2 for various queries submitted by
computational complexity grows polynomially with the
key length (the linear appearance is due to the logarithmic
scaling of both axes in Figure 2). In fact, upon fitting an
appropriate curve, we get that the computational time (in
sec) grows as O(k 2.44 ) with k denoting the key length.
On the other hand, the computational burden is a linear
function of the key length (the slope of the line in Figure 2
is equal to one) as the size of the integers that needs to be
transmitted grows linearly with the key length.
D(y) > 0
1
6. CONCLUSIONS AND FUTURE WORK
0
16
21
50
w
Fig. 1. The outcome of Algorithms 1 and 2 for various w.
2
15
A secure and private framework for communication between two fleet owners was proposed. This secure communication platform was generalized to create distributed coordination mechanisms among fleet owners for heavy-duty
vehicle platooning. The future work will focus on developing a centralized coordination mechanism for reducing
the computational complexity or outsourcing burden to
cloud computing services as well as the application of the
framework to other services.
64
Computation Time (sec)
1 hour
2
10
16
1 min
25
4
20
1
2-5
16
32
64
129
256
512
Communication Complexity (KBytes)
REFERENCES
1024
Key Length (bits)
Fig. 2. The computation time (
) and the communication burden ( × ) associated with executing
Algorithms 1 and 2 versus the key length.
the first fleet owner with a key length of 128 bits. The
vertical axis of Figure 1 is equal to one if D(y) > 0 and
is equal to zero otherwise. Evidently, D(y) > 0 for only
w = 1, 6, 21, 50. Therefore, as expected, upon following the
proposed algorithms, the second fleet owner can correctly
respond to the query of the first fleet owner without even
knowing the content of the query. However, this secure
communication channel comes at a price. Figure 2 shows
the computation time and the communication burden associated with executing Algorithms 1 and 2 versus the key
length, measured in bits. The computation is done with
Python programming language on Windows 7 over a PC
with Intel(R) i7-4770 CPU at 3.40GHz and 16GB of RAM.
The computation time rapidly increases with increasing
the key length. The amount of data the fleet owners need to
communicate also increases with the key length. Note that
the security of the encryption is related to the key length.
In fact, the computational complexity of a brute-force attack that requires the test of all the keys of a specific length
grows exponentially with the key length. Note that the
Besselink, B., Turri, V., van de Hoef, S.H., Liang, K.Y.,
Alam, A., Mrtensson, J., and Johansson, K.H. (2016).
Cyber-physical control of road freight transport. Proceedings of the IEEE, 104(5), 1128–1141.
Boneh, D. and Waters, B. (2007). Conjunctive, subset, and
range queries on encrypted data. In S.P. Vadhan (ed.),
Theory of Cryptography: 4th Theory of Cryptography
Conference, TCC 2007, Amsterdam, The Netherlands,
February 21-24, 2007. Proceedings, 535–554. Springer
Berlin Heidelberg, Berlin, Heidelberg.
Farokhi, F., Shames, I., and Batterham, N. (2016).
Secure and private cloud-based control using semihomomorphic encryption. In Proceedings of the 6th
IFAC Workshop on Distributed Estimation and Control
in Networked Systems.
Farokhi, F., Liang, K.Y., and Johansson, K.H. (2015).
Cooperation patterns between fleet owners for transport
assignments. In Proceedings of the IEEE Conference on
Control Applications, 1124–1129. IEEE.
Kogiso, K. and Fujita, T. (2015). Cyber-security enhancement of networked control systems using homomorphic
encryption. In Proceedings of the 54th IEEE Conference
on Decision and Control, 6836–6843.
Ostrovsky, R. and Skeith, E.W. (2007). Private searching
on streaming data. Journal of Cryptology, 20(4), 397–
430.
Paillier, P. (1999). Public-key cryptosystems based on
composite degree residuosity classes. In J. Stern (ed.),
Advances in Cryptology — EUROCRYPT ’99: International Conference on the Theory and Application of
Cryptographic Techniques Prague, Czech Republic, May
2–6, 1999 Proceedings, 223–238. Springer Berlin Heidelberg, Berlin, Heidelberg.
Vialar, T. (2015). Handbook of mathematics. Books on
Demand.
Yi, X., Paulet, R., and Bertino, E. (2014a). Homomorphic
Encryption and Applications. SpringerBriefs in Computer Science. Springer International Publishing.
Yi, X. and Bertino, E. (2011). Private searching for
single and conjunctive keywords on streaming data.
In Proceedings of the 10th Annual ACM Workshop on
Privacy in the Electronic Society, 153–158.
Yi, X., Bertino, E., Vaidya, J., and Xing, C. (2014b).
Private searching on streaming data based on keyword
frequency. IEEE Transactions on Dependable and Secure Computing, 11(2), 155–167.
| 3cs.SY
|
1
Simulated Annealing Algorithm for Graph Coloring
Alper Köse, Berke Aral Sönmez, Metin Balaban, Random Walks Project
arXiv:1712.00709v1 [cs.AI] 3 Dec 2017
Abstract—The goal of this Random Walks project is to code and
experiment the Markov Chain Monte Carlo (MCMC) method for
the problem of graph coloring. In this report, we present the plots
of cost function H by varying the parameters like q (Number of
colors that can be used in coloring) and c (Average node degree).
The results are obtained by using simulated annealing scheme,
where the temperature (inverse of β) parameter in the MCMC
is lowered progressively.
Keywords—Graph Coloring, Simulated Annealing, MCMC
Method.
I. I NTRODUCTION
A. Graph Coloring
Graph coloring is one of the most important concepts in
graph theory and is used in many real time applications in
computer science. It can be defined as a problem of how
to assign colors to certain elements of a graph given some
constraints. It is especially used in research areas of science
such as data mining, image segmentation, clustering, image
capturing, networking etc. Vertex coloring is the most common
application of graph coloring. In vertex coloring, given m
colors, the aim is to find the coloring of vertices of the graph
such that no two adjacent vertices have the same color (Here,
adjacent means that two vertices are directly connected by an
edge). If this goal is achieved, the graph is said to have proper
coloring. However, it is not always possible to find the proper
coloring of a graph. In this situation, the coloring with the
minimum cost function is searched:
H(x) =
X
1xv =xw
(1)
(v,w)∈E
This energy function assigns a unit cost to every edge which
has the same colors on two vertices bounded to it. Therefore,
one can state that this function counts the total number of
such bad edges in a graph. Our goal here is to minimize the
energy function and to see if there exist minimizers that yield
zero cost (proper coloring). On the other hand, if the global
minimum is strictly positive, then there does not exist a proper
coloring for the given graph for given number of colors.
To minimize the energy function, we use the simulated
annealing method in this project.
B. Simulated Annealing
Finding the optimal solution for some optimization problems
can be an incredibly difficult task because when a problem gets
sufficiently large we need to search through an enormous number of possible solutions to find the optimal one. Therefore,
there are often too many possible solutions to consider. In such
cases, since it is not possible to find the optimal solution within
a reasonable length of time, one has to try to find a solution
that is close enough to global optimum.
Most minimization strategies find the nearest local minimum
but we need a different strategy not to get stuck in a local minimum. Namely, we should sometimes iterate to new colorings
that do not improve the current situation. In such a situation,
we consider Simulated Annealing, since it is a probabilistic
technique for approximating the global optimum of a given
function which works as described above.
In Simulated Annealing, there is a temperature parameter
that has to be tuned, in order to get close enough to the
optimal solution. It has an initial value and it should be slowly
decreased. One must be careful for finding such a schedule
because if the cooling is too slow, it may not be possible to
reach the minimizer in a reasonable amount of time, on the
other hand, if the cooling is too fast, it is likely to end up in
a local minimum.
For now, practical deterministic algorithms for finding
optimal solution are not known. As a probabilistic model,
Simulated Annealing method can do quite well. In other
words, given sufficient time at each temperature, it leads to
equilibrium state and in the end it can converge to a solution
which is close enough to the optimal one.
C. Metropolis Algorithm
In the problem, we want to find the coloring vector that
minimizes the cost function H. In order to achieve this
goal, we use simulated annealing which is an adaptation of
Metropolis Algorithm and the Monte Carlo method. Metropolis
Algorithm that is given in the project description is used in
optimization with only modifying β parameter in some steps
of the iterations.
The Metropolis algorithm is given below (Think β as fixed
for now since we do not change it in every iteration) :
1. At time t = 0, initialize with a coloring x0 taken
uniformly at random.
2. At times t ≥ 0, make a transition xt → xt+1 according
to the following rules:
• Select a vertex v ∈ V uniformly at random, with current
color xv . Pick a color different from xv at random and recolor
the vertex v. Consider the new coloring xnew and compute
4 = H(xnew ) − H(xt ).
• if 4 ≤ 0, then accept the new color (with probability one)
and set xt+1 = xnew .
• if 4 > 0, then accept the new color with probability
p(H(xnew ))
= exp(−β4)
p(H(xt ))
(2)
and set xt+1 = xnew ; reject the move with the complementary probability and set xt+1 = xt .
3. Iterate for n time steps.
2
Fig. 1. Results on an Erdos-Renyi graph with 1000 vertices and average
node degree c = 5. The algorithm is iterated 106 times for every q value of
3,5 and 7. For these q values, Hmin is obtained as 34, 0, and 0 respectively.
Fig. 2. Results on an Erdos-Renyi graph with 1000 vertices and average
node degree c = 10. The algorithm is iterated 106 times for every q value of
3,5 and 7. For these q values, Hmin is obtained as 395, 0,and 0 respectively.
II.
program. Therefore, we also changed initial β = 0.8 and
0
trials0 = 1.5 ∗ N which we get good results for this number
of iterations. The final β was 23 which makes the acceptance
probability close enough to zero as to be considered as the
chain is frozen. As we expected, results were not as good as
it is in 1 billion iterations, but still close to it.
We plotted the curves of H(xt ) as a function of time for
three different values of c = 5, 10, 20 and three different values
of q = 3, 5, 7. The plots can be seen in Figures 1,2 and 3.
In general, as we desired, there are significant oscillations
in the H values at the beginning (β is small) of iterations and
they become more stable with the time and at the end (β is
large) they are frozen.
From the perspective of c values, the number of edges in
the graph is proportional to the c value. Thus, with higher
connectivity, it is harder to obtain a proper coloring scheme
for the graph. This can be verified from the Figures 1,2 and
3, because it is seen that when c = 5, q = 5 and q = 7
give proper coloring where q = 3 gives a coloring scheme
which has a relatively small cost. When c = 10, q = 5 and
q = 7 again give proper coloring where q = 3 gives a coloring
scheme which has a higher cost than c = 5 case. Finally, when
c = 20, only q = 7 is able to give a proper coloring, q = 5 no
longer provides a proper coloring scheme and q = 3 has the
highest cost among all schemes.
From the perspective of q values, it is straight-forward to
state that when q increases, we obtain smaller costs in every
iteration, and it is intuitively obvious since higher number of
available color means higher possibility of having different
colors on two ends of an edge. Although there is a possibility
that in some iterations we can have smaller cost in colorings
while also having smaller q values, this probability is very
small and even gets smaller when the vertex number or
iterations (with same modification scheme of β) increases.
In our second experiment, we aim to show the relation
E XPERIMENTAL S ETUP AND D ISCUSSION OF R ESULTS
As recommended in the lab sheet, we coded the algorithm
in Matlab firstly. We know that there is not a certain optimal
point that is aimed to be reached. We want to make Hmin
as small as possible. As known, β must be increased slow
enough not to converge to a local minimum and fast enough
to complete the simulation in a reasonable amount of time. By
trials we saw that increasing the number of iterations allows
us to reach lower energies and sometimes proper colorings for
some parameters. Therefore, we chose our iteration number
to be as big as possible and also as small as allowing as to
do computations in a reasonable amount of time. We experimented with 1000x1000 Erdos-Renyi random graph to see if
the execution time is reasonable. When we do 105 iterations, it
takes approximately 50 seconds to run the program in Matlab.
We thought that we can complete this task in a time which is
much shorter than this, so we coded the algorithm in C++. As
a result, it took 0.04 seconds to run the same algorithm which
is 1250 times faster.
Then, in the optimization step, we thought that we can
increase β slowly not to get stuck in a local minima since
our computation time is really low. With the trials by hand,
we estimated the modification scheme as (0.2+N/n)
where N
0.2
is the number of vertices and n is the number of iterations. To
find a good initial β and how often to apply the modification,
we ran our code many times by taking random samples
for the initial β and the 0 trials0 (modifying in every that
number of iteration). As a result, we obtained the best results
when we used 1 billion iterations with initial β = 0.98 and
0
trials0 = 3.4 ∗ N . This simulation of 1 billion iterations is
completed in just 100 seconds. Also, when we increase the
number of iterations, we could not observe a better Hmin .
On the other hand, since we are obliged to give the graphs
of H and Hmin in the report, we decided to use 1 million
iterations for plotting them in Matlab without crashing the
3
Fig. 3. Results on an Erdos-Renyi graph with 1000 vertices and average
node degree c = 20. The algorithm is iterated 106 times for every q value
of 3,5 and 7. For these q values, Hmin is obtained as 1510, 302, and 5
respectively.
between the node degree of Erdos-Renyi graph and the minimum of the cost function (Hmin ). For the experiment setup,
we created 100 random graphs whose c parameters range
from 1 to 100. On each of these graphs, we attempted to
get the best coloring for q = 3, 5, 7 using our simulated
annealing algorithm implementation. We kept the parameters
of the algorithm the same as we set in the first experiment.
As a result, we got the plot in Figure 4. As expected, when
the node degree is very low, q = 3, 5, 7 all gives proper
coloring for the problem. When we gradually increase the node
degree, minimum value of the cost function (Hmin ) starts to
be non-zero for q = 3 at first, then for q = 5 and lastly for
q = 7. This is a normal behavior because we can color a graph
easier if we have more colors to use, namely we have more
chance to get closer to proper coloring by using more colors.
Also, after being non-zero, linear increase of Hmin can be
seen from the graphs. This can be explained like that: The
expectation of edge number will be c(N − 1)/2, when the
node degree c is increased linearly, edge number will increase
in the same pattern, and as edge number increases and since
we have random coloring, we will also have a linear increase
in the minimum of the cost function Hmin . But, the slope of
the linearly increasing functions are bigger when q is smaller.
This is because the increase in the total number of possible
color combinations will grow more while q is increasing. For
instance, think of a graph which has 3 vertices and all of the
vertices are connected. In this case, we have 6 different proper
colorings if q = 3, 24 if q = 4 and 60 if q = 5. So, as one
can see, we have increasingly growing options while q value
increases which explains the different values of the slopes of
the Hmin s.
III. C ONCLUSION
1) The more the iteration the larger the probability of having
better coloring.The reason for this is we have the possibility
Fig. 4. The plot showing the relation between average node degree (c) and
the minimum potential. c values ranges from 1 − 100 and the algorithm is
iterated 106 times for every q value of 3,5 and 7.
of making more iterations for each β value which makes us
closer to mixing time for that β value.
2) Different β modification schemes give similar results
given the required conditions are satisfied (small initial β,
modifying slow enough not to get stuck in a local minimum).
3) Final β value must be large enough to guarantee that the
chain is frozen(1/β goes to 0).
4) The simulated annealing algorithm on large instances of
graph coloring problem does not assure the optimal coloring
assignment. The mixing time for the chain to converge to
the stationary distribution is larger than the time required for
exploring the state space of the problem. That state space is
quite large, for example, when the number of vertices is 1000
and number of colors are 5, the size is 51000 .
5) Since the simulated annealing algorithm is a randomized
algorithm and it does not guarantee reaching the optimal value,
different runs produce different results. In order to obtain the
minimum potential in limited time, the best strategy is to split
the time interval into several runs and return the minimum
value among the runs. Another advantage of this strategy is
that different runs can be parallelizable due to independence
of the runs. We run the algorithm concurrently on the same
graph instance, achieving higher computational power.
6) C++ is efficient in time but harder to code.
R EFERENCES
[1]
Project Description. http://lthiwww.epfl.ch/~leveque/Random_Walks/
project_RW_2016.pdf
[2] Ahmed, Shamim. "Applications of Graph Coloring in Modern Computer
Science." International Journal of Computer and Information Technology.() (2012).
[3] http://csg.sph.umich.edu/abecasis/class/2006/615.19.pdf
[4] Johnson, David S., et al. "Optimization by simulated annealing: an
experimental evaluation; part II, graph coloring and number partitioning."
Operations research 39.3 (1991): 378-406.
| 2cs.AI
|
Application of a Semi-automatic Algorithm for Identification
of Molecular Components in SBML Models
Andrea Maggiolo-Schettini
Paolo Milazzo
Giovanni Pardini
Dipartimento di Informatica
Università di Pisa, Italy
[email protected]
Dipartimento di Informatica
Università di Pisa, Italy
[email protected]
Dipartimento di Informatica
Università di Pisa, Italy
[email protected]
Reactions forming a pathway can be rewritten by making explicit the different molecular components
involved in them. A molecular component represents a biological entity (e.g. a protein) in all its states
(free, bound, degraded, etc.). In this paper we show the application of a component identification
algorithm to a number of real-world models to experimentally validate the approach. Components
identification allows subpathways to be computed to better understand the pathway functioning.
1
Introduction
A typical aspect of a biochemical pathway is that the process described involves mainly a few chains
of reactions, in which the occurrence of a reaction produces some intermediate molecule which is then
transformed subsequently by other reactions. Just a few basic biological entities are often involved in
a pathway, for example a simple protein undergoes a series of transformations, starting from its initial
synthesized form, which can then be activated and also become part of different complexes. Therefore,
intermediate molecular species can actually be seen as different states of the same initial biological
entities. In accordance with this view, in the modelling of biochemical pathways we can consider a
notion of molecular component [9, 8] that is the formal counterpart of the notion of biological entity.
A molecular components will hence be associated with the set of species mentioned in the model and
corresponding to different states of the same biological entity.
This paper discusses the application of a semi-automatic algorithm for the identification of molecular components in pathways which is presented in [18]. The algorithm infers components from the
interactions of molecular species. In order to make components involved in each reaction explicit, the
algorithm replaces each compound species (which is usually identified by a distinct name) with a fresh
name for each one of its components, and transforms reactions into a normal form in which each complex is represented by as many species as are the biological entities involved in it. The algorithm is
actually semi-automatic, since there are cases in which the components involved in a reaction cannot be
univocally determined from the context; these cases must be solved with the help of the user.
In this paper, we present the results of the application of such an algorithm to a number of realworld pathways from the BioModels repository [12], an online database of machine-readable models
of biological processes formalized in the well-known Systems Biology Markup Language (SBML). In
particular, we have analyzed all of the models in the curated section, which contains only models from
the literature, for a total of 436 models. In this paper, some of the models are also discussed individually
in order to experimentally assess the validity of our approach when applied to real-world models.
The identification of components by the transformation of the pathway reactions into normal form
allows different kinds of analysis to be performed. For example, syntactic transformations, such as
the projection of the pathway over a subset of components, would allow the user to obtain insights on
A. Graudenzi, G. Caravagna, G. Mauri, M. Antoniotti (Eds.):
Wivace 2013 - Italian Workshop on
Artificial Life and Evolutionary Computation
EPTCS 130, 2013, pp. 43–52, doi:10.4204/EPTCS.130.7
c A. Maggiolo-Schettini, P. Milazzo & G. Pardini
This work is licensed under the
Creative Commons Attribution License.
44
Identification of Molecular Components in SBML Models
the functioning of the pathway. In particular, by focusing on the reactions in which they are involved,
the mechanisms underlying the pathway dynamics could be more easily identified. Moreover, from a
theoretical point of view, the normalization would allow the automatic translation of a pathway into a
set of automata or terms of process calculi [19, 1, 2, 5, 7], thus enabling the use of any tool available for
their analysis.
2
Algorithm for Component Identification
In this section we recall the algorithm for component identification presented in [18]. The algorithm
is based on the idea that a species is a “state” of a more abstract biological entity, and a reaction is
a synchronized state change of a set of such entities. The algorithm assumes that biological entities
involved in a reaction cannot appear from nothing and/or disappear (i.e. degradations should be modelled
by using a species representing the degraded state of the biological entity). As a consequence, every
entity involved in a reaction should be mentioned both among reactants and among products, and hence
in a reaction there should be as many reactants as products.
Let a component be a set of species representing all possibile states of a given biological entity. The
component identification algorithm transforms a given pathway into a normal-form pathway, namely
such that, for all its reactions, there are as many reactants as products and there is a one-to-one positional
correspondence between reactants and products which are part of the same component. The fundamental
operation of the algorithm is to split a species into new subspecies, denoted by newly-introduced symbols. A split of a species is performed any time the algorithm can infer that, in a reaction, the occurrence
of such a species must actually denote a complex composed of multiple bound molecules. For example,
this is the case of a reaction A, B → C, where C can be split into subspecies CA ,CB , obtaining the normalform reaction A, B → CA ,CB , with components {A,CA } and {B,CB }. In general, this process may not be
completely automatic since some ambiguities may arise. For example, a reaction such as A, B → C, D, E
is ambiguous since it is not clear which of the two reactants has to be split into two in order to obtain as
many reactants as products.
Another important aspect is that, in principle, the number of components of a pathway is not univocally determined. For example, it is possible to split the reactants/products of a reaction A → B into
A1 , . . . , An → B1 , . . . , Bn , thus identifying any number of components, one for each pair Ai , Bi . However,
the idea is to identify only those components which can be inferred from the context, namely from the
reactions in which a component is involved. For this reason, the algorithm performs a split only if it is
needed to match the components of reactants and products of a reaction.
Structure of the algorithm Let us denote (in an abstract manner) by comp(S) the component in which
a species S occurs at a given step. Initially, the algorithm assumes that each species occurring in the
input pathway is part of a different component. It then iteratively performs two alternating phases, until a
normal-form pathway is obtained. In the first phase, it tries to transform the set of reactions into reactions
having the same number of reactants and products. This entails inspecting each reaction in turn, in order
to both (i) split species into subspecies as needed, and (ii) refine the information about components by
collapsing different components into single ones, according to what can be inferred from reactions.
In case of ambiguities, the algorithm may fail to generate a normal-form pathway since either the
number of reactants and products differ, or the correspondence between reactants and products is not
completely specified for all the reactions. In such a case, the algorithm performs a second phase, which
demands user intervention to resolve a single ambiguous reaction. Then the two phases are repeated to
propagate the new information derivable from the resolved reaction to the others, as it may be useful to
A. Maggiolo-Schettini, P. Milazzo & G. Pardini
45
resolve other ambiguous reactions, and possibly all of them.
As regards the first phase, the algorithm examines each reaction singularly. Let us consider a reaction
having reactants R = {A1 , . . . , Aγ } and products P = {B1 , . . . , Bδ }. Moreover, assume that the components of each of the first k < γ, δ pairs of reactant/product match, namely ∀i ∈ {1, . . . , k}. comp(Ai ) =
comp(Bi ). Let n = γ − k and m = δ − k denote the number of remaining unmatched reactants/products
of the reaction, respectively. The algorithm then distinguishes five kinds of reactions, as follows:
• Case n = 0, m = 0: the reaction is resolved, with complete correspondence between each reactant
and product. This case is not necessarily definitive, since the algorithm, due to some split, may
subsequently modify the reaction by replacing a species occurring in it with new symbols.
• Case n = 1, m = 1: since there is just one species remaining in each side, this means that both of
them are part of the same component. Thus, the components of the reactant and product are joint
into a single component by taking their union.
• Case (n = 0 ∧ m > 0) or (n > 0 ∧ m = 0): if either of reactants or products are empty, there is at
least a component which appears in either side of the reaction which does not occur in the other
side. In other words, such a component either appears or disappears in between the reaction, which
is not allowed by our approach. Thus this case is regarded as an error, which cannot be resolved.
• Case n = 1 ∧ m > 1 (or n > 1 ∧ m = 1): in this case there is a single reactant which is actually
formed from m subspecies. Thus, m new symbols Aγ -Bk , . . . , Aγ -Bδ are introduced, by combining
the name of Aγ with the Bi ’s. Then Aγ is replaced by the new symbols in all the reactions, and
the components of each pair of species (Aγ -Bk , Bk ), . . . , (Aγ -Bδ , Bδ ) are matched, i.e. ∀k ≤ j ≤
δ . comp(Aγ -B j ) = comp(B j ). (The converse case, with one product and n reactants is analogous.)
• Case n > 1, m > 1: if there are multiple unmatched reactants and products, the reaction is currently
ambiguous. This means that the current information regarding components does not allow the
algorithm to decide which species must be split, thus this reaction is skipped.
The first phase continues until there are only resolved, erroneous and ambiguous reactions. In case of
ambiguous reactions remaining, the algorithm asks the user to resolve one reaction by specifying the
species to split and rewriting it in normal-form. The procedure is repeated from the first phase in order
to propagate the updated information about components to the other reactions.
As an example, let us consider pathway P = {r1 : Lig, rcpt → C1 ; r2 : GDP, Gα → C2 ; r3 : GT P, Gα →
C3 ; r4 : C3 → C2 ; r5 : C2 , Gβ γ → C4 ; r6 : C4 ,C1 → C5 }, which models a small fragment of the well-known
G protein signalling pathway, in which a ligand Lig (representing the signal) binds the receptor rcpt on
cellular membrane, triggering the internal process. The algorithm identifies 5 components (representative
species of which are Lig, rcpt, Gα , Gβ γ and GDP) and produces the follwing normal-form pathway
(where species names are abbreviated):
P0 ={r01 : Lig, rcpt → C1 -Lig,C1 -rcpt;
r02 : GDP, Gα → C2 -GDP,C2 -Gα ;
r04 : C3 -GT P,C3 -Gα → C2 -GDP,C2 -Gα ;
r03 : GT P, Gα → C3 -GT P,C3 -Gα ;
r05 : C2 -GDP,C2 -Gα , Gβ γ → C4 -GDP,C4 -Gα ,C4 -Gβ γ ;
r06 : C4 -GDP,C4 -Gα ,C4 -Gβ γ ,C1 -Lig,C1 -rcpt → C5 -GDP,C5 -Gα ,C5 -Gβ γ ,C5 -Lig,C5 -rcpt}.
The sets of species forming each component are as follows: {Lig,C1 -Lig,C5 -Lig}, {rcpt,C1 -rcpt,C5 -rcpt},
{Gα ,C2 -Gα ,C4 -Gα ,C5 -Gα ,C5 -Gα ,C3 -Gα }, {Gβ γ ,C4 -Gβ γ ,C5 -Gβ γ }, {GDP,C2 -GDP,C4 -GDP,C5 -GDP, GT P,
C3 -GT P}. Now, we can compute subpathways describing the activity of a subset of the molecular components. For instance, by assuming that we are not interested in the role of Lig, rcpt and GDP, we can compute the subpathway dealing only with components Gα and Gβ γ as follows: P00 = {r002 : Gα → C2 -Gα ; r003 :
Gα → C3 -Gα ; r004 : C3 -Gα → C2 -Gα ; r005 : C2 -Gα , Gβ γ → C4 Gα ,C4 -Gβ γ ; r006 : C4 -Gα ,C4 -Gβ γ → C5 -C2 -Gα ,C5 -Gβ γ }.
Moreover, the subpathways describing each component individually can be trivially translated into a
finite state automaton or into a process algebra term to enable the application of formal analysis tools.
46
3
Identification of Molecular Components in SBML Models
Applications
In order to test our component identification algorithm on a relevant number of real pathways we downloaded all of the pathway descriptions available in the BioModels database [12] under the category
“curated models”. So, our testbed consisted of 436 different SBML models of pathways.
SBML [10] is a well-established XML-based language for pathway description. A SBML pathway
model includes (among others) the following elements:
• Species: Proteins, genes, ions and other molecules that can participate in reactions.
• Compartments: Well-stirred containers in which species can be located. A SBML model may
contain multiple compartments (at least one) and each species must be located in one of them.
• Reactions: Statements describing transformation, transport or binding processes that can change
the amount of one or more species. A reaction consists of reactants, products and modifiers.
Moreover, the kinetic law of a reaction can be expressed by using arbitrary mathematical functions.
• Rules: Mathematical expressions describing how some variable values (e.g. species amounts) can
be calculated from other variables, or used to define the rate of change of variables. Rules in a
model can be used either together with (or in place of) reactions to determine the model dynamics.
• Events: A statement describing an instantaneous, discontinuous change in a set of variables when
a triggering condition is satisfied.
We developed a simple translator of SBML models into CSV files accepted by our implementation of
the algorithm. The translator considers only the reactions of a SBML model and transforms them into
the format expected by the algorithm. SBML models in which the dynamics is governed only by rules,
and not by reactions, are translated into empty CSV files, and hence are unusable by our tool. Moreover,
we excluded from our test models containing reactions with fractional stoichiometry, that are meaningful
only when the dynamics is described by means of ODEs. In the considered testbed, unusable models
turn out to be 59 out of 436 and excluded models are 23 out of 436, hence we have 354 usable CSV files.
The 354 usable models consisted on average of 22.07 species and 31.05 reactions, with 108 models
dealing with more than 20 species, 31 dealing with more than 50 species, and 6 dealing with more than
100 species. We executed our algorithm batch on all of the models. The execution (on a standard laptop)
takes a few seconds for each model. The execution on a model can terminate either successfully with an
automatically generated normal-form pathway (OK), or with an error message if an erroroneous reaction
is encountered, or with a human intervention request if an ambiguous reaction is encountered. In the
latter two cases, the execution is interrupted and the number of erroneous and ambiguous reactions,
respectively, is printed. The results we obtained on the 354 usable models are shown in the table below,
on line “Batch execution”. Most of the models (244 out of 354) encountered some errors. By inspecting
some of them we discovered that in most cases the error was due to reactions describing either synthesis
or degradation of some species, in which either products or reactants were absent, respectively.
Batch execution
Preprocessing & Batch execution
Preprocessing & Batch execution with dynamic correction
Ok
96
241
318
Erroneous
244
93
0
Ambiguous
14
20
36
We decided to solve the problem of synthesis/degradation reactions by preprocessing the SBML
models. The preprocessing was performed on the CSV translation and consisted in inserting a dummy
species in each empty set of reactants or products encountered. Note that every time an empty set is
encountered by the preprocessor, a new fresh species is generated and used as dummy. Hence, each
dummy species will appear only once in the pathway. Since dummy species added during preprocessing
do not interact with other species, error situations are solved without affecting component identification.
A. Maggiolo-Schettini, P. Milazzo & G. Pardini
47
The results of executions of the algorithm after preprocessing are in the table, on line “Preprocessing
& Batch execution”. Most of the previously erroneous situations can now be handled by the algorithm,
and in the vast majority of the cases they did not encountered ambiguous reactions (151 errors are solved
and only 6 of them turn out to need human intervention).
By inspecting some of the models still giving an erroneous result we discovered that in some cases
the error was caused by a reaction describing the degradation of a part of a complex. For example, let
us consider a pathway consisting of reactions A, B → C and C → A. The first reaction is the formation
of a complex composed by A and B, and the second is the degradation of B when it bound to A. In this
case the algorithm transforms the first reaction into A, B → CA ,CB . Consequently, in the second reaction
it replaces C by CA ,CB obtaining CA ,CB → A. Now, in the second reaction the algorithm associates CA
with A (as this follows from the first reaction) and it has nothing to associate with CB (error situation).
To solve this second kind of errors we choose to add fresh dummy objects also at runtime when errors
are encountered. In the example, the second reaction could be corrected by adding a fresh dummy object
DB as follows: CA ,CB → A, DB . This is not always correct since the dynamic insertion of dummy species
may lead to different components being identified, depending on the order of processing of reactions.
The results of executions of the algorithm after preprocessing and with insertion of dummy objects at
runtime are in the table, on line “Preprocessing & Batch execution with dynamic correction”. All of the
error situations are now solved and only 36 of the 354 models turn out to need human intervention.
The component identification algorithm completed automatically its execution in the 89.83% of the
cases. In the remaining 36 cases we had an average of 4.67 ambiguous reactions, that means that at most
4.67 questions are asked to the user in an average execution of a model with ambiguous reaction (with a
maximum of 16 and only two models over 10). The average computed over all of the 354 cases is 0.53.
The algorithm, although semi-automatic, turns out to require very limited human intervention in
practice. Let us now assess the quality of the computed results by comparing the molecular components
identified by the algorithm with the biological entities the modelled pathways deal with according to the
referenced literature. We consider eight “randomly” chosen SBML models (numbers 50, 100, 150, 200,
250, 300, 350 and 400 in the BioModels database), three randomly chosen big models (numbers 88,
235 and 293) and three randomly chosen models with ambiguities (numbers 82, 143 and 165). In the
following paragraphs a brief summary of the analysis of each of these SBML models is given.
Analysis of model 50 (BIOMD0000000050.xml). This SBML model describes a kinetic model of N(1-deoxy-D-fructos-1-yl)-glycine (DFG) thermal decomposition [14]. The SBML model includes 14
species and 16 reactions. During its execution the algorithm encountered one erroneous reaction, that
was solved by dynamically inserting a dummy object in it. In the model DFG can be degraded into
several different ways obtaining a number of different substances. The model includes some species
representing unidentified intermediate components (E1 and E2 ) and unidentified carbohydrate fragments
(Cn ). Most of the species involved in the pathway can be transformed into these unidentified molecules,
thus causing all of these species to be included in the same molecular component by the algorithm. As a
consequence, the algorithm identifies only 3 components: two including all and only the species related
with Glycine and methyulglyoxal, respectively, and one including all of the other species. The quality
of the result in this case are hence only partially satisfactory, and the cause of unsatisfaction is that the
model includes ambiguities (unidentified species).
Analysis of model 82 (BIOMD0000000082.xml). This SBML model describes the formation of an
inhibitor of the Adenylate Cyclase enzyme [24]. The SBML model includes 10 species and 6 reactions.
During its execution the algorithm encountered no erroneous reactions, but two ambiguous reactions. The
two ambiguous reactions are the following: DR, G GDP → DRG GDP and DRG GDP → GDP, DRG.
48
Identification of Molecular Components in SBML Models
These reactions describe the passage (in two steps) of a G-protein from GDP to DR. This is a typical
case of ambiguity (see [18]) that can be solved by making it explicit the involvement of the “hidden”
component representing the G protein, namely by replacing the ambiguous reactions with the following
ones: DR, GG GDP , GDPG GDP → DRG GDP and DRG GDP → GDP, GDRG , DRDRG . So, in this case the
algorithm needs human intervention, after which 4 components are correctly identified.
Analysis of model 88 (BIOMD0000000088.xml). This model describes the Rho-kinase pathway in
order to study thrombin-dependent in vivo transient responses of Rho activation and Ca2+ increase [13].
It includes 104 species and 110 reactions, and it is one of the biggest models we have considered. During
its execution the algorithm encountered no erroneous and ambiguous reactions. The algorithm identified
28 components. Species in the SBML file are represented by numbers, hence checking the correctness of
the identified components was not trivial. We checked a few randomly chosen components and each of
them turned out to be composed of species representing different states of the same molecule. Although
we do not have a complete proof of the correctness of the result, we consider this case satisfactory.
Analysis of model 100 (BIOMD0000000100.xml). This SBML model is used to study the effects of
cytosolic calcium oscillations on activation of glycogen phosphorylase [21]. The SBML model includes
5 species and 10 reactions. During its execution the algorithm encountered five errors, that were solved
by preprocessing. The species involved in the model are: external calcium (EC), cytosolic calcium (Z),
intravescicular calcium (Y), inositol 1,4,5-trisphosphate IP3 (A), and glycogen phosphorlylase (GP). The
reactions described in the model are: calcium influx, transportation of calcium between compartments,
and IP3 and glycogen phosphorylase syntheses and degradations. The algorithm identifies 3 components:
one including all of the species representing calcium (EC, Z and Y) and the other two including IP3 and
glycogen, respectively. The quality of the result in this case is hence satisfactory.
Analysis of model 143 (BIOMD0000000143.xml). This SBML model describes the metabolism of
activated neutrophils in which oscillatory behaviours have been observed [17]. The model includes 20
species and 20 reactions. During its execution the algorithm encountered three errors, solved by preprocessing. The algorithm encountered also 9 ambiguous reactions. After inspecting the model we
discovered that thee reactions in particular were problematic, the solution of which solved also the ambiguities in the other 6 reactions. The three problematic reactions describe the Peroxidase Cycle in which
an enzyme is activated by a H2 O2 molecule and then transforms in two steps two Melatonin molecules
into Melatonin-free-radical. During this cycle two molecules of water are released, and also some hydrogen ions are involved. Water and hydrogen are not mentioned in the model, and this creates ambiguities
in the reactions. Such ambiguities can be solved by a human intervention aimed at clarifying the three
problematic reactions. After this, the algorithm completes by correctly identifiying 5 components.
Analysis of model 150 (BIOMD0000000150.xml). This is a very simple SBML model describing formation and activation of Cdk/Cyclin Complexes [15]. The model includes 4 species and 4 reactions. The
algorithm encountered no erroneous reactions. The modelled reactions are trivial (complex formation
from two species, activation of the complex and inverse reactions). The algorithm correctly identifies 2
components, one for each molecule involved in the complex. The result in this case is hence satisfactory.
Analysis of model 165 (BIOMD0000000165.xml). This model deals with intracellular signalling through
cAMP and its cAMP-dependent protein kinase (PKA) [22]. The SBML model includes 37 species and
30 reactions. During its execution the algorithm encountered three erroneous reactions, two of which
solved by preprocessing and one by dynamically inserting a dummy object. The model included one ambiguous reaction similar to those encountered in the case of model 82. The reaction can be disambiguated
by means of a single human intervention after which 16 components are correctly identified.
Analysis of model 200 (BIOMD0000000200.xml). This SBML model describes binding reactions
leading to a transmembrane receptor-linked multiprotein complex involved in bacterial chemotaxis [4].
A. Maggiolo-Schettini, P. Milazzo & G. Pardini
49
The SBML model includes 21 species and 34 reactions. During its execution the algorithm encountered
no erroneous reactions. The biochemical entities involved in the pathway are the dimeric aspartatebinding receptor Tar (TT) and four cytoplasmic proteins CheW (W), CheA (AA), CheY (Y), CheB
(B) and CheZ (Z). TT, W and AA are involved in a number of reactions leading to the formation of a
complex TTWWAA that activates proteins Y, B, Z in different ways . The algorithm identifies exactly 6
components out of the 21 species, and such components correspond exactly to the 6 biochemical entities
involved in the pathway. The quality of the result in this case is hence satisfactory.
Analysis of model 235 (BIOMD0000000235.xml). This SBML model describes the sea urchin endomesoderm network [11], a very big gene regulation network. The model includes 618 species and 778
reactions. During its execution the algorithm encountered only 3 erroneous reactions, solved by preprocessing. The algorithm identified only 47 components. By inspecting them we discovered that actually
one of the components included the vast majority of the species. By inspecting reactions we discovered
that actually they consisted mostly of syntheses and degradations in which species “none” was used in all
the case of empty reactants or products. The algorithm associated all of the species involved in a synthesis or degradation reaction in the same component. This was due to the fact that all of such species are in
relation with the same species “none”. We modified the model by replacing the unique “none” by fresh
dummy species (as in the case of preprocessing). After this change the algorithm identified 406 components. We checked some of them (randomly chosen) and they turned out to be correct representations
of biochemical entities involved in the network. We consider hence the quality of the result satisfactory,
although this model needed to be slightly modified in order to let the algorithm work as expected.
Analysis of model 250 (BIOMD0000000250.xml). This SBML model is used to study how epidermial
growth factor (EGF) and heregulin (HRG) generate distinct responses of the transcription factor c-fos
[16]. The SBML model includes 49 species and 78 reactions. During its execution the algorithm encountered 19 erroneous reactions solved by preprocessing. The algorithm identifies 18 components that,
after an analysis of the description of the pathway in the paper, seem to correspond to the biochemical
entities involved in the pathway. The quality of the result in this case is hence satisfactory.
Analysis of model 293 (BIOMD0000000293.xml). This SBML model describes an ubiquitin-proteasome
system [20]. The SBML model includes 136 species and 316 reactions. During its execution the algorithm encountered 114 erroneous reactions, all solved by dynamic insertion of dummy objects. The
algorithm identified only 12 components. By inspecting the model we discoverd that is suffered from the
same problem of model 235, namely the same dummy species “source” and “AggP Proteasome” were
used in many syntheses and inhibition reactions, causing the most of the species to be assigned to the
same component by the algorithm. We modified this model as we did for model 235. The number of erroneous reactions encounterd decreased to 77, and the number of components identified by the algorithm
decreased to 16. Components seems to be rather correct, since they seems to represent a reasonable
partition of the species set. However, the complexity of the model and the high number of erroneous
reactions solved only at runtime does not allow us to be sure about the correctness of the result. We leave
the assessment of the quality of this result as a future work.
Analysis of model 300 (BIOMD0000000300.xml). This SBML model is used to study how the activity
of the heterodimeric transcription factor hypoxia inducible factor (HIF) if affected by the interaction
of factors inhibiting HIF (FIH) with ankyrin-repeat domain (ARD) proteins [23]. The SBML model
includes 9 species and 10 reactions. Reactions are all degradations and syntheses of species. All of the
10 reactions become erroneous reactions that are solved by preprocessing. The algorithm identifies 9
components. However, by inspecting the models it emerged that most of the dynamics is described by
means of SBML rules, hence the models turns out to be unusable.
Analysis of model 350 (BIOMD0000000350.xml). This SBML model is used to study the circadian
50
Identification of Molecular Components in SBML Models
network of higher plants. In particular, it is a model of the clock of the picoeukaryotic alga Ostreococcus
tauri as a feedback loop between the genes TOC1 and CCA1 [25]. The SBML model includes 14 species
and 30 reactions. Actually, most reactions are syntheses and degradations. Indeed, 27 reactions out of
30 become erroneous reactions that are solved by preprocessing. Only three reactions describe transformation of species, and consequently the algorithm correctly identifies 11 components. The quality of the
result in this case is hence satisfactory.
Analysis of model 400 (BIOMD0000000400.xml). This SBML model belongs to the set of unusable
models since it does not include any reaction.
4
Discussion
The component identification algorithm we proposed in [18] turned out to work pretty well. In the
majority of the cases the execution of the algorithm has been completely automatic, and the computed
molecular components correctly represented the biological entities involved in the pathway. In the cases
in which human intervention was necessary, it usally consisted in answering very few questions on how
resolve ambiguous reactions. After the detailed analysis of some of the models (numbers 235 and 293)
a recurrent “error” in the modelling of pathways (from the viewpoint of the algorithm) is to use the
same special species to represent the pre-synthesis or the degraded form of many different biological
entities. This causes different molecular components to be erroneously merged into one. Moreover, it
also emerged that ambiguities often are of the same kind, namely they are include some “hidden” species
that is always bound to some other species. In these cases the algorithm cannot identify the component
corresponding to such species and the reactions turn out to be ambiguous.
Most of the erroneous and ambiguous situations could be prevented by a more accurate construction
of models. However, it could be an interesting further development of our work the definition of some
rule or heuristics able to solve these situations when encountered. Otherwise, the prevention of erroneous
and ambiguous situations could be approached by developing of a model repair preprocessing routine
based for instance on static checking of conservation laws or P-invariants [6].
References
[1] Roberto Barbuti, Giulio Caravagna, Andrea Maggiolo-Schettini, Paolo Milazzo & Giovanni Pardini (2008):
The calculus of looping sequences. In: Formal Methods for Computational Systems Biology, Springer, pp.
387–423. doi:10.1007/978-3-540-68894-5 11
[2] Roberto Barbuti, Andrea Maggiolo-Schettini, Paolo Milazzo & Angelo Troina (2006): A calculus of looping
sequences for modelling microbiological systems. Fundamenta Informaticae 72(1), pp. 21–35.
[3] Michael L Blinov, James R Faeder, Byron Goldstein & William S Hlavacek (2004): BioNetGen: software for
rule-based modeling of signal transduction based on the interactions of molecular domains. Bioinformatics
20(17), pp. 3289–3291. doi:10.1093/bioinformatics/bth378
[4] D. Bray & R.B. Bourret (1995): Computer analysis of the binding reactions leading to a transmembrane
receptor-linked multiprotein complex involved in bacterial chemotaxis. Molecular biology of the cell 6(10),
pp. 1367–1380.
[5] Federica Ciocchetta & Jane Hillston (2009): Bio-PEPA: A framework for the modelling and analysis of
biological systems. Theoretical Computer Science 410(33), pp. 3065–3084. doi:10.1016/j.tcs.2009.02.037
[6] Allan Clark, Vashti Galpin, Stephen Gilmore, Maria Luisa Guerriero & Jane Hillston (2012): Formal methods
for checking the consistency of biological models. In: Advances in Systems Biology, Springer, pp. 461–475.
doi:10.1007/978-1-4419-7210-1 27
A. Maggiolo-Schettini, P. Milazzo & G. Pardini
51
[7] Vincent Danos & Cosimo Laneve (2004): Formal molecular biology. Theoretical Computer Science 325(1),
pp. 69–110. doi:10.1016/j.tcs.2004.03.065
[8] Peter Drábik, Andrea Maggiolo-Schettini & Paolo Milazzo (2012): On conditions for modular verification
in systems of synchronising components. Fundamenta Informaticae 120(3), pp. 259–274.
[9] Peter Drábik, Andrea Maggiolo-Schettini & Paolo Milazzo (2012): Towards modular verification of pathways: fairness and assumptions. In: Proceedings of MeCBIC 2012, EPTCS 100, pp. 63–81.
[10] M. Hucka, A. Finney, H.M. Sauro, H. Bolouri, J.C. Doyle, H. Kitano & the rest of the SBML Forum (2003):
The systems biology markup language (SBML): a medium for representation and exchange of biochemical
network models. Bioinformatics 19(4), pp. 524–531. doi:10.1093/bioinformatics/btg015
[11] C. Kühn, C. Wierling, A. Kühn, E. Klipp, G. Panopoulou, H. Lehrach & A. Poustka (2009): Monte carlo
analysis of an ode model of the sea urchin endomesoderm network. BMC systems biology 3(1), p. 83.
doi:10.1186/1752-0509-3-83
[12] C. Li, M. Donizelli, N. Rodriguez, H. Dharuri, L. Endler, V. Chelliah, L. Li, E. He, A. Henry, M.I. Stefan, J.L.
Snoep, M. Hucka, N. Le Novère & C. Laibe (2010): BioModels Database: An enhanced, curated and annotated resource for published quantitative kinetic models. BMC Systems Biology 4, p. 92. doi:10.1186/17520509-4-92
[13] A. Maeda, Y. Ozaki, S. Sivakumaran, T. Akiyama, H. Urakubo, A. Usami, M. Sato, K. Kaibuchi & S. Kuroda
(2006): Ca2+-independent phospholipase A2-dependent sustained Rho-kinase activation exhibits all-ornone response. Genes to Cells 11(9), pp. 1071–1083. doi:10.1111/j.1365-2443.2006.01001.x
[14] S.I.F.S. Martins & M.A.J.S. Van Boekel (2003): Kinetic modelling of Amadori N-(1-deoxy-d-fructos-1-yl)glycine degradation pathways. Part II–Kinetic analysis. Carbohydrate research 338(16), pp. 1665–1678.
doi:10.1016/S0008-6215(03)00174-5
[15] M.C. Morris, C. Gondeau, J.A. Tainer & G. Divita (2002): Kinetic Mechanism of Activation
of the Cdk2/Cyclin A Complex.
Journal of Biological Chemistry 277(26), pp. 23847–23853.
doi:10.1074/jbc.M107890200
[16] T. Nakakuki, M.R. Birtwistle, Y. Saeki, N. Yumoto, K. Ide, T. Nagashima, L. Brusch, B.A. Ogunnaike,
M. Okada-Hatakeyama & B.N. Kholodenko (2010): Ligand-specific c-Fos expression emerges from the spatiotemporal control of ErbB network dynamics. Cell 141(5), pp. 884–896. doi:10.1016/j.cell.2010.03.054
[17] L.F. Olsen, U. Kummer, A.L. Kindzelskii & H.R. Petty (2003): A model of the oscillatory metabolism of
activated neutrophils. Biophysical journal 84(1), pp. 69–81. doi:10.1016/S0006-3495(03)74833-4
[18] G. Pardini, P. Milazzo & A. Maggiolo-Schettini: An Algorithm for the Identification of Components in Biochemical Pathways. In: Proceedings of CS2Bio 2013, ENTCS. In press.1 .
[19] A. Phillips & L. Cardelli (2007): Efficient, correct simulation of biological processes in the stochastic picalculus. In: Computational Methods in Systems Biology, Springer, pp. 184–199. doi:10.1007/978-3-54075140-3 13
[20] C.J. Proctor, P.J. Tangeman & H.C. Ardley (2010): Modelling the Role of UCH-L1 on Protein Aggregation
in Age-Related Neurodegeneration. PLoS ONE 5(10), p. e13175. doi:10.1371/journal.pone.0013175.s013
[21] A. Rozi & Y. Jia (2003): A theoretical study of effects of cytosolic Ca2+ oscillations on activation of glycogen
phosphorylase. Biophysical chemistry 106(3), pp. 193–202. doi:10.1016/S0301-4622(03)00192-3
[22] J.J. Saucerman, J. Zhang, J.C. Martin, L.X. Peng, A.E. Stenbit, R.Y. Tsien & A.D. McCulloch (2006): Systems analysis of PKA-mediated phosphorylation gradients in live cardiac myocytes. Proceedings of the
National Academy of Sciences 103(34), pp. 12923–12928. doi:10.1073/pnas.0600137103
[23] B. Schmierer, B. Novák & C. Schofield (2010): Hypoxia-dependent sequestration of an oxygen sensor by
a widespread structural motif can shape the hypoxic response-a predictive kinetic model. BMC systems
biology 4(1), p. 139. doi:10.1186/1752-0509-4-139
1 Pre-proceedings
version available online at http://cs2bio13.di.unito.it/papers/Pardini.pdf
52
Identification of Molecular Components in SBML Models
[24] W.J. Thomsen, J.A. Jacquez & R.R. Neubig (1988): Inhibition of adenylate cyclase is mediated by the high
affinity conformation of the alpha 2-adrenergic receptor. Molecular pharmacology 34(6), pp. 814–822.
[25] C. Troein, F. Corellou, L.E. Dixon, G. van Ooijen, J.S. ONeill, F.-Y. Bouget & A.J. Millar (2011): Multiple
light inputs to a simple clock circuit allow complex biological rhythms. The Plant Journal 66(2), pp. 375–385.
doi:10.1111/j.1365-313X.2011.04489.x
| 5cs.CE
|
HyperENTM: Evolving Scalable Neural Turing Machines through HyperNEAT
Jakob Merrild and Mikkel Angaju Rasmussen and Sebastian Risi
arXiv:1710.04748v1 [cs.AI] 12 Oct 2017
IT University of Copenhagen, Denmark
{jmer, mang, sebr}@itu.dk
Abstract
Recent developments within memory-augmented neural networks have solved sequential problems requiring long-term
memory, which are intractable for traditional neural networks. However, current approaches still struggle to scale to
large memory sizes and sequence lengths. In this paper we
show how access to memory can be encoded geometrically
through a HyperNEAT-based Neural Turing Machine (HyperENTM). We demonstrate that using the indirect HyperNEAT
encoding allows for training on small memory vectors in a
bit-vector copy task and then applying the knowledge gained
from such training to speed up training on larger size memory vectors. Additionally, we demonstrate that in some instances, networks trained to copy bit-vectors of size 9 can be
scaled to sizes of 1,000 without further training. While the
task in this paper is simple, these results could open up the
problems amendable to networks with external memories to
problems with larger memory vectors and theoretically unbounded memory sizes.
Introduction
Memory-augmented neural networks are a recent improvement on artificial neural networks (ANNs) that allow them to
solve complex sequential tasks requiring long-term memory
(Sukhbaatar et al. 2015; Graves et al. 2016; Graves, Wayne,
and Danihelka 2014) . Here we are particularly interested in
Neural Turing Machines (NTM) (Graves, Wayne, and Danihelka 2014), which allow a network to use an external memory tape to read and write information during execution.
This improvement enables the ANN to be trained in fewer
iterations, for certain tasks, and also allows the network to
change behaviour on the fly.
However, scaling to large memory sizes and sequence
length is still challenging. Additionally, current algorithms have difficulties extrapolating information learned on
smaller problem sizes to larger once, thereby bootstrapping
from it. For example, in the copy tasks introduced by Graves
et al. (2014) the goal is to store and later recall a sequence
of bit vectors of a specific size. It would be desirable that
a network trained on a certain bit vector size (e.g. 8 bits)
would be able to scale to larger bit vector sizes without further training. However, current machine learning approaches
often cannot transfer such knowledge.
Recently, Greve et al. (2016a) introduced an evolvable
version of the NTM (ENTM), which did not rely on differen-
tiability and offered some unique advantages. First, in addition to the networks weights, the optimal neural architecture
can be learned at the same time. Second, a hard memory attention mechanism is directly supported and the complete
memory does not need to be accessed each time step. Third,
a growing and theoretically infinite memory is now possible.
Additionally, in contrast to the original NTM, the network
was able to perfectly scale to very long sequence lengths.
However, because it employed the direct genetic encoding
NEAT, which means that every parameter of the network is
described separately in its genotype, the approach had problems scaling to copy tasks with vectors of more than 8 bits.
To overcome these challenges, in this paper we combine
the ENTM with the indirect Hypercube-based NeuroEvolution of Augmenting Topologies (HyperNEAT) encoding
(Stanley, D’Ambrosio, and Gauci 2009). HyperNEAT provided a new perspective on evolving ANNs by showing that
the pattern of weights across the connectivity of an ANN can
be generated as a function of its geometry. HyperNEAT employs an indirect encoding called compositional pattern producing networks (CPPNs), which can compactly encode patterns with regularities such as symmetry, repetition, and repetition with variation (Stanley 2007). HyperNEAT exposed
the fact that neuroevolution benefits from neurons that exist
at locations within the space of the brain and that by placing
neurons at locations, evolution can exploit topography (as
opposed to just topology), which makes it possible to correlate the geometry of sensors with the geometry of the brain.
While lacking in many ANNs, such geometry is a critical
facet of natural brains (Sporns 2002). This insight allowed
large ANNs with regularities in connectivity to evolve for
high-dimensional problems.
In the new approach introduced in this paper, called HyperENTM, an evolved neural network generates the weights
of a main model, including how it connects to the external
memory component. Because HyperNEAT can learn the geometry of how the network should be connected to the external memory, it is possible to train a CPPN on a small bit
vector sizes and then scale to larger bit vector sizes without
further training.
While the task in this paper is simple it shows – for the
first time – that access to an external memory can be indirectly encoded, an insight that could directly benefit indirectly encoded HyperNetworks training through gradient de-
scent (Ha, Dai, and Le 2016) and could be applied to more
complex problem by employing recent advances in Evolutionary Strategies (Salimans et al. 2017).
1) Query each potential 2) Feed each coordinate pair into CPPN
connection on substrate
1
x1
Y
Backgorund
This section reviews NEAT, HyperNEAT, and Evolvable
Neural Turing Machines, which are foundational to the approach introduced in this paper.
Neuroevolution of Augmenting Topologies
The HyperNEAT method that enables learning from geometry in this paper is an extension of the original NEAT algorithm that evolves ANNs through a direct encoding (Stanley
and Miikkulainen 2002; Stanley and Miikkulainen 2004). It
starts with a population of simple neural networks and then
complexifies them over generations by adding new nodes
and connections through mutation. By evolving networks in
this way, the topology of the network does not need to be
known a priori; NEAT searches through increasingly complex networks to find a suitable level of complexity.
The important feature of NEAT for the purpose of this paper is that it evolves both the topology and weights of a network. Because it starts simply and gradually adds complexity, it tends to find a solution network close to the minimal
necessary size. The next section reviews the HyperNEAT extension to NEAT that is itself extended in this paper.
HyperNEAT
In direct encodings like NEAT, each part of the solution’s
representation maps to a single piece of structure in the final
solution (Floreano, Dürr, and Mattiussi 2008). The significant disadvantage of this approach is that even when different parts of the solution are similar, they must be encoded and therefore discovered separately. Thus this paper
employs an indirect encoding instead, which means that the
description of the solution is compressed such that information can be reused. Indirect encodings are powerful because
they allow solutions to be represented as a pattern of parameters, rather than requiring each parameter to be represented individually (Bongard 2002; Gauci and Stanley 2010;
Hornby and Pollack 2002; Stanley and Miikkulainen 2003).
HyperNEAT, reviewed in this section, is an indirect encoding extension of NEAT that is proven in a number of challenging domains that require discovering regularities (Clune
et al. 2009; Gauci and Stanley 2010; Stanley, D’Ambrosio,
and Gauci 2009). For a full description of HyperNEAT see
(Gauci and Stanley 2010).
In HyperNEAT, NEAT is altered to evolve an indirect
encoding called compositional pattern producing networks
(CPPNs (Stanley 2007)) instead of ANNs. CPPNs, which
are also networks, are designed to encode compositions of
functions, wherein each function in the composition loosely
corresponds to a useful regularity.
The appeal of this encoding is that it allows spatial patterns to be represented as networks of simple functions (i.e.
CPPNs), which means that NEAT can evolve CPPNs just
like ANNs. CPPNs are similar to ANNs, but they rely on
-1
-1
Substrate
X
1
1,0
1,1
...
-0.5,0 0,1
...
-1,-1 -0.5,0
...
-1,-1
- 1,0
...
y1
x2 y2
CPPN (evolved)
3) Output is weight
between (x 1,y1) and (x 2,y2)
Figure 1: Hypercube-based Geometric Connectivity Pattern
Interpretation. A collection nodes, called the substrate, is
assigned coordinates that range from −1 to 1 in all dimensions. (1) Every potential connection in the substrate is
queried to determine its presence and weight; the dark directed lines in the substrate depicted in the figure represent
a sample of connections that are queried. (2) Internally, the
CPPN (which is evolved) is a graph that determines which
activation functions are connected. As in an ANN, the connections are weighted such that the output of a function is
multiplied by the weight of its outgoing connection. For
each query, the CPPN takes as input the positions of the two
endpoints and (3) outputs the weight of the connection between them. Thus, CPPNs can produce regular patterns of
connections in space.
more than one activation function (each representing a common regularity) and are an abstraction of biological development rather than of brains. The indirect CPPN encoding can compactly encode patterns with regularities such as
symmetry, repetition, and repetition with variation (Stanley
2007). For example, simply by including a Gaussian function, which is symmetric, the output pattern can become
symmetric. A periodic function such as sine creates segmentation through repetition. Most importantly, repetition with
variation (e.g. such as the fingers of the human hand) is easily discovered by combining regular coordinate frames (e.g.
sine and Gaussian) with irregular ones (e.g. the asymmetric x-axis). The potential for CPPNs to represent patterns
with motifs reminiscent of patterns in natural organisms has
been demonstrated in several studies (Secretan et al. 2008;
Stanley 2007).
The main idea in HyperNEAT is that CPPNs can naturally encode connectivity patterns (Gauci and Stanley 2010;
Stanley, D’Ambrosio, and Gauci 2009). That way, NEAT
can evolve CPPNs that represent large-scale ANNs with
their own symmetries and regularities.
Formally, CPPNs are functions of geometry (i.e. locations in space) that output connectivity patterns whose nodes
are situated in n dimensions, where n is the number of dimensions in a Cartesian space. Consider a CPPN that takes
four inputs labeled x1 , y1 , x2 , and y2 ; this point in fourdimensional space also denotes the connection between the
two-dimensional points (x1 , y1 ) and (x2 , y2 ), and the output
of the CPPN for that input thereby represents the weight of
that connection (Figure 1). By querying every possible connection among a pre-chosen set of points in this manner, a
CPPN can produce an ANN, wherein each queried point is
a neuron position. Because the connections are produced by
a function of their endpoints, the final structure is produced
with knowledge of its geometry.
In HyperNEAT, the experimenter defines both the location and role (i.e. hidden, input, or output) of each such
node. As a rule of thumb, nodes are placed on the substrate to reflect the geometry of the task (Clune et al. 2009;
Stanley, D’Ambrosio, and Gauci 2009). That way, the connectivity of the substrate is a function of the task structure.
How to integrate this setup with an ANN that has an external memory component is an open question, which this
paper tries to address.
Evolvable Neural Turing Machines (ENTM)
Based on the principles behind the NTM, the recently introduced ENTM uses NEAT to learn the topology and weights
of the ANN controller (Greve, Jacobsen, and Risi 2016a).
That way the topology of the network does not have to be
defined a priori (as is the case in the original NTM setup)
and the network can grow in response to the complexity
of the task. As demonstrated by Greve et al., the ENTM
often finds compact network topologies to solve a particular task, thereby avoiding searching through unnecessarily
high-dimensional spaces. Additionally, the ENTM was able
to solve a complex continual learning problem (Lüders et
al. 2017). Because the network does not have to be differentiable, it can use hard attention and shift mechanisms, allowing it to generalize perfectly to longer sequences in a copy
task. Additionally, a dynamic, theoretically unlimited tape
size is now possible.
The ENTM has a single combined read/write head. The
network emits a write vector w of size M , a write interpolation control input i, a content jump control input j, and three
shift control inputs sl , s0 , and sr (left shift, no shift, right
shift). The size of the write vector M determines the size of
each memory location on the tape. The write interpolation
component allows blending between the write vector and the
current tape values at the write position, where Mh (t) is the
content of the tape at the current head location h, at time step
t, it is the write interpolation, and wt is the write vector, all
at time step t: Mh (t) = Mh (t − 1) · (1 − it ) + wt · it .
The content jump determines if the head should be moved
to the location in memory that most closely resembles the
write vector. A content jump is performed if the value of the
control input exceeds 0.5. The similarity between write vector
w and memory vector m is determined by: s(w, m) =
P
M
i=1
|wi −mi |
.
M
At each time step t, the following actions are
performed in order: (1) Record the write vector wt to the
current head position h, interpolated with the existing content according to the write interpolation it . (2) If the content
jump control input jt is greater than 0.5, move the head to
location on the tape most similar to the write vector wt . (3)
Shift the head one position left or right on the tape, or stay
at the current location, according to the shift control inputs
sl , s0 , and sr . (4) Read and return the tape values at the new
head position.
Approach: Hyper Neural Turing Machines
(HyperENTM)
In the HyperENTM the CPPN does not only determine the
connections between the task related ANN inputs and outputs but also how the information coming from the memory
is integrated into the network and how information is written
back to memory. Because HyperNEAT can learn the geometry of a task it should be able to learn the geometric pattern
in the information written to and read from memory.
The following section details the HyperENTM approach
on the copy task, which was first introduced by (Graves,
Wayne, and Danihelka 2014). In the copy task the network
is asked to store and later recall a sequence of bit vectors.
At the start of the task the network receives a special input,
which denotes the start of the input phase. Afterwards, the
network receives the sequence of bit-vectors, one at a time.
Once the sequence has been fully passed to the network it
receives another special input, signaling the end of the input
phase and the start of the output phase. For any subsequent
time steps the network does not receive any input.
In summary, the network has the following inputs: Start:
An input that is activated when the storing of numbers
should begin. Switch: An input that is activated when the
storing should stop and the network must start recalling the
bit vectors instead. Bit-vector input: Before the switch input
has been activated this input range is activated with the bits
that are to be recited later. Memory read input: The memory
vector that the TM read in the previous time step.
And following outputs: Bit-vector output: The bit vector
that the network outputs to the environment. During the input phase this output is ignored. Memory write output: The
memory vector that should be written to memory. TM controls: TM specific control outputs. Jump, interpolation, and
three shift controls (left, stay, and right).
Copy Task Substrate
The substrate for the copy task is shown in Figure 2. The
substrate is designed such that the bit-vector input nodes
share x-coordinates with the memory vector write nodes
and vice versa with memory vector read nodes and bitvector output nodes. Furthermore, the switch input shares
its x-coordinate with the jump output, thus encouraging the
network to jump in memory when it should start reciting.
In this paper, the size of the memory vector equals the
bit vector size. Furthermore, none of the substrates contain hidden nodes as it has been shown that it is possible to
solve even large versions of the problem without any hidden
nodes (Greve, Jacobsen, and Risi 2016b).
Following Verbancsics and Stanley (2011), in addition to
the CPPN output that determines the weights of each connection, each CPPN has an additional step-function output,
called the link-expression output (LEO), which determines
if a connections should be expressed. Potential connections
are queried for each input on layers y = 1 and y = −1
to each output on layers y = 1 and y = −1. The number
of inputs and outputs on the y = −1 layer (Figure 2b) are
scaled dependent on the size of the copy task bit vector. In
the shown example the bit-vector size is three. Neurons are
z
Inputs
y=1
z
1
Inputs
−1
1
Outputs
1
y = −1
x
−1
1
x
Outputs
(a)
Bit-vector input
Memory read vector
Switch
Start
Shift controls
Jump control
Interp control
Memory write vector
Bit-vector output
(b)
Figure 2: The HyperENTM substrate for the copy task. All inputs are in z = 1 and all outputs in z = −1. Figure (a) shows
all nodes in y = 1 which is the start/switch inputs and the tm controls. Notably the x-coordinate is the same for the switch
input and the jump control output. Figure (b) shows the nodes in y = −1 which are the bit-vector and memory vector input and
outputs. Bit-vector input nodes share x-coordinates with memory vector write nodes, while memory vector read nodes share
x-coordinates with bit-vector output nodes.
uniformly distributed in the x interval [−1.0, −0.2] for bit
vector inputs and memory write vector and in the interval
[0.2, 1.0] for the memory read vector and bit vector output.
The CPPN has an additional output that determines the
bias values for each nodes in the substrate. These values
are determine through node-centric CPPN queries (i.e. both
source and target neuron positions xyz are set to the location
of the node whose bias should be determined).
Experiments
A total of three different approaches are evaluated on bitsizes of 1, 3, 5, and 9. HyperNEAT is compared to the direct NEAT encoding, and a Seeded HyperNEAT treatment
that starts evolution with a CPPN seed that encourages locality on both the x- and y-coordinates (Figure 7a). A similar locality seed has been shown useful in HyperNEAT to
encourage the evolution of modular networks (Verbancsics
and Stanley 2011). This locality seed is then later adjusted
by evolution (e.g. adding/removing nodes and connections
and changing their weights).
The fitness function in this paper follows the one in the
original ENTM paper (Greve, Jacobsen, and Risi 2016a).
During training the network is given a sequence of random
bit vectors, between 1 and 10 vectors long, and asked to recite it. The network is tasked to do this with 50 random sequences and assigned a normalized fitness. The network is
evaluation the bit-vectors recited by the network are compared to those given to it during the input phase; for every
bit-vector the network is given a score based on how close
the output from the network corresponding to a specific bit
was to the target. If the bit was within 0.25 of the target,
the network is awarded a fitness of the difference between
the actual output and the target output. Otherwise, the nett|
work is not awarded for that specific bit: f = 1 − |x−x
0.25 if
|x−xt | < 0.2 and 0 otherwise. The fitness for any given bitvector is equal to the the sum of the fitness for each individual bit, normalized to the length of the bit-vector. Similarly,
the fitness for a complete sequence is the sum of the fitness
for each bit-vector normalized to the length of the sequence.
This results in a fitness score between 0 and 1. This fitness
function rewards the network for gradually getting closer to
the solution, but it does not actively reward the network for
using the memory to store the inputs.
Experimental Parameters
For the NEAT experiments, offspring generation proportions are 50% sexual (crossover) and 50% asexual (mutation). Following (Lüders et al. 2017), we use 98.8% synapse
weight mutation probability, 9% synapse addition probability, and 5% synapse removal probability. Node addition probability is set to 0.05%. The NEAT implementation SharpNEAT uses a complexity regulation strategy for
the evolutionary process, which has proven to be quite impactful on our results. A threshold defines how complex the
networks in the population can be (here defined as the number of genes in the genome and set to 10 in our experiments),
before the algorithm switches to a simplifying phase, where
it gradually reduces complexity.
For the HyperNEAT experiments the following parameters are used. Elitism proportion is 2%. Offspring generation
proportions are 50% sexual (crossover) and 50% asexual
(mutation). CPPN connection weights have a 98.8% probability of being changed, a 1% change of connection addition,
and 0.1% change of node addition and node deletion. The
activation functions available to new neurons in the CPPN
are Linear, Gaussian, Sigmoid, and Sine, each with a 25%
probability of being added. Both NEAT and HyperNEAT experiments run with a population size of 500 for a maximum
of 10,000 generations or until a solution is found. The code
is available from: https://github.com/kalanzai/
ENTM_CSharpPort.
Mean fitness over 10,000 generations
1
HyperNEAT
Seeded HyperNEAT
NEAT
Comparison of solution rate
0.7
0.6
1
0.8
Size 3
Fitness
0.9
0.7
0.6
Number of experiments solved
20
0.8
Size 1
Fitness
0.9
15
10
5
0
1
HyperNEAT
Seeded HyperNEAT
NEAT
1
3
5
9
Size
0.8
Size 5
Fitness
0.9
0.7
0.6
Figure 4: The number of solutions found by each configuration for the four different bit vector sizes.
1
0.8
Comparison of final fitness
Size 9
Fitness
0.9
0.7
Figure 3: Mean champion fitness for the different treatments
and bit-sizes, averaged over 20 independent evolutionary
runs.
Results
Figure 3 shows the mean champion fitness over 10, 000 generations for each of the different approaches and bit vector
sizes. While NEAT performs best on smaller bit vectors, as
the size of the vector grows to 9 bits, the seeded HyperNEAT
variant outperforms both NEAT and HyperNEAT. The numbers of solutions found (i.e. networks that reach a training
score ≥ 0.999) in regards to the bit vector size are shown
in Figure 4. For bit size 1 all approaches solve the problem equally well. However, as the size of the bit vector is
increased the configurations using HyperNEAT and locality
seed performs best and the only method that is able to find
any solution for size 9.
Testing Performance. To determine how well the champions from the last generation generalize, they were tested
on 100 random bit-vector sequences of a random lengths
between 1 and 10 (Figure 5). On sizes 1 and 5 there is
no statistical difference between either treatment (following a two-tailed Mann-Whitney U test). On size 3, NEAT
performs significantly better than the seeded HyperNEAT
(p < .00001). Finally, seeded HyperNEAT performs significantly better than NEAT on size 9 (p < .00001). The
main conclusions are that (1) while NEAT performs best
on smaller bit vectors it degrades rapidly with increased bit
sizes, and (2) the seeded HyperNEAT variant is able to scale
to larger sizes while maintaining performance better.
Generalizing to longer sequences. We also tested how
many of the solutions, which were trained on bit sequences
of length 10, generalize to sequences of length 100. The
training and generalisation results are summarized in Ta-
Fitness
0.6
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
3
5
9
Size
NEAT
HyperNEAT
Seeded HyperNEAT
Figure 5: The mean performance of the champion networks
from the last generation. NEAT does well on smaller sizes,
but degenerates quickly as the size goes up.
ble 1, which shows the number of solutions for each of the
three approaches, how many of those solutions generalized
to sequences of length 100, and the average number of generations it took to find a solution. For all three methods, most
solutions generalize perfectly to sequences that are longer
than the sequences encountered during training.
Transfer Learning
To test the scalability of the Seeded HyperNEAT solutions,
champion genomes from runs which found a solution for
a given size were used as a seed for evolutionary runs of
higher sizes. The specific runs and which seeds were used
can be seen in Table 2, which also contains the number of
solutions found, how many solutions generalized, and the
average number of generations needed to find the solutions.
Seeds denoted X → Y refer to champion genomes from a
run of size Y which was seeded with a champion from a run
of size X.
Because the number of solutions found varied between the
different sizes (see Table 1), the scaling experiments were
not run exactly 20 times. Instead, the number of runs was
Table 1: Generalisation Results. Shown are the number of
solutions, the number of solutions that generalize, together
with the average number of generations it took to find a solution and standard deviation.
HyperNEAT
NEAT
Size
1
3
5
9
1
3
5
9
1
3
5
9
#sol.
20
5
6
8
20
2
0
0
20
19
12
0
#gen
20
5
5
5
20
2
0
0
19
19
11
0
gens.
1055.8
4454.8
2695.5
1523.25
1481.45
4395.5
N/A
N/A
281
2140.5
3213.4
N/A
sd.
1147.4
3395.8
2666.5
2004.1
1670.8
388.2
N/A
N/A
336.3
1594.4
2254.2
N/A
Table 2: HyperNEAT Transfer Learning
Seed
3
3
5
3→5
5
9
5→9
9
9→17
Size
5
9
9
9
17
17
17
33
33
#sol
12/20
16/20
23/24
23/24
18/20
24/24
20/20
23/24
24/24
#general
6
2
11
8
12
12
17
17
19
gens.
434.1
1150.3
58.9
588
258.9
89.3
36.25
94
456.5
sd.
574.6
2935.9
129.2
1992
301.3
235.9
47.5
217
2002.4
the smallest number above or equal to 20 which allowed for
each champion to be seeded an equal number of times, e.g.
if there were 6 solutions 24 runs were made; 4 runs with the
champion from each solution. Figure 6 shows a comparison of HyperNEAT seeded with the locality seed and seeded
with champion genomes of smaller sizes on the size 9 problem. HyperNEAT yielded significantly better results when
seeded with size 5 and 3→5 champions compared to starting with the locality seed (p < .001), but not when seeded
with the champion from size 3.
Scaling without further training
The champions from runs which found a solution were
tested for scaling to larger bit-vector sizes without further
evolutionary training (i.e. new input and output nodes are
created and queried by the CPPN; see copy task substrate
section for details). Each genome was tested on 50 sequences of 100 random bit-vectors of size 1,000. Some of
the champions found using only the LEO size 9 configuration scaled perfectly to a bit-size of 1,000 without further
training, as seen in Table 3.
The main results is that it is possible to find CPPN that
perfectly scale to any size. The fact that evolution with HyperNEAT performs significantly better, when seeded with a
champion genome which solved a smaller size of the problem, together with the fact that evolution sometimes finds
0.9
Fitness
Seeded
HyperNEAT
Size 9
1
0.8
0.7
Seeded HyperNEAT
Seeded HyperNEAT from size 3 champion
Seeded HyperNEAT from size 5 champion
Seeded HyperNEAT from size 3 -> 5 champion
0.6
0
1000
2000
3000
4000
5000
6000
7000
8000
9000 10000
Generations
Figure 6: Comparison of the size 9 problem with normal locality seed and with champion seeds from a smaller size.
Table 3: Scaling using LEO without further evolution
Size
# of champions # which scaled to 1000
9
8
2
9 → 17
24
7†
†
6 of these can be traced back to the 2 champions
from size 9 which scaled perfectly.
solutions which scale without further training, demonstrates
that HyperNEAT can be used to scale the dimensionality of
the bit-vector in the copy task domain.
Solution Example
Here we take a closer look at one of the champion genomes
(trained on bit-vector size 9), which was able to scale perfectly to the size 1000 problem (Table 3). Figure 7 shows a
visualization of the champion genome, as well as the locality seed from which it was evolved. The champion genome
does resemble the seed but also evolved several additional
connections that are necessary to solve the problem.
Figure 8 shows two different ANNs generated by the the
same CPPN, which is shown in Figure 7b. It can be seen
that for non-bias connections to be expressed, the source and
destination nodes have to be located in the same position on
both the x position and y layer in the substrate. These results
suggest that the locality encouraging seed works as intended.
To further demonstrate the scalability of this evolved
CPPN, memory usage for size 9 and 17 are shown in Figure 9. The output of the networks is shown at the top, followed by the input to the network. Next follows the difference between the given input and the output, i.e. how well
the network recited the sequence given to it. The fourth section of the recording shows a heat map of the fitness score
based on the bit-vector in that position, where red indicates
a high fitness while blue indicates a low fitness.
Next follows the output from the network to the TM controller, followed by the interpolation output, and what was
written to the memory tape. Finally, the recording shows
the jump output from the network, the three different shift
outputs, what read from the tape, and the position of the
read/write head. As the recording shows, the network which
x1
bias
y1
weight
z1
x2
y2
bias
z2
x1
weight
y1
z1
y=1
Bias
y=1
LEO
x2
bias
y2
y = −1
(a)
(a)
bias
Bias
z2
LEO
(b)
Figure 7: (a) The CPPN seed that is used to promote locality
on the x and y axes. (b) A champion trained on the size 9
problem, which was able to scale without further evolution
to size 1,000. Blue connections have a positive weight, while
red connections have a negative weight.
solves the problem jumps exactly once when the switch input neuron is activated.
Conclusion
This paper showed that the indirect encoding HyperNEAT
makes it feasible to train ENTMs with large memory vectors for a simple copy task, which would otherwise be infeasible to train with an direct encoding such as NEAT. Furthermore, starting with a CPPN seed that encouraged locality, it
was possible to train solutions to the copy task that perfectly
scale with the size of the bit vectors which should be memorized, without any further training. Lastly, we demonstrated
that even solutions which do not scale perfectly can be used
to shorten the number of generations needed to evolve a
solution for bit-vectors of larger sizes. In the future it will
be interesting to apply the approach to more complex domains, in which the geometry of the connectivity pattern
to discover is more complex. Additionally, combining the
presented approach with recent advances in Evolutionary
Strategies (Salimans et al. 2017), which have been shown
to allow evolution to scale to problems with extremely highdimensionality is a promising next step.
References
[Bongard 2002] Bongard, J. C. 2002. Evolving modular
genetic regulatory networks. In Proceedings of the 2002
Congress on Evolutionary Computation.
y = −1
(b)
Figure 8: Networks produced by the champion CPPN (Figure 7b) for bit sizes 3 (a) and 17 (b).
[Clune et al. 2009] Clune, J.; Beckmann, B. E.; Ofria, C.;
and Pennock, R. T. 2009. Evolving coordinated quadruped
gaits with the HyperNEAT generative encoding. In Proceedings of the IEEE Congress on Evolutionary Computation (CEC-2009) Special Session on Evolutionary Robotics.
Piscataway, NJ, USA: IEEE Press.
[Floreano, Dürr, and Mattiussi 2008] Floreano, D.; Dürr, P.;
and Mattiussi, C. 2008. Neuroevolution: from architectures
to learning. Evolutionary Intelligence 1(1):47–62.
[Gauci and Stanley 2010] Gauci, J., and Stanley, K. O. 2010.
Indirect encoding of neural networks for scalable go. In
Schaefer, R.; Cotta, C.; Kołodziej, J.; and Rudolph, G., eds.,
Parallel Problem Solving from Nature, PPSN XI: 11th International Conference, Kraków, Poland, September 11-15,
2010, Proceedings, Part I, 354–363. Berlin, Heidelberg:
Springer Berlin Heidelberg.
[Graves et al. 2016] Graves, A.; Wayne, G.; Reynolds, M.;
Harley, T.; Danihelka, I.; Grabska-Barwińska, A.; Colmenarejo, S. G.; Grefenstette, E.; Ramalho, T.; Agapiou, J.;
et al. 2016. Hybrid computing using a neural network with
dynamic external memory. Nature 538(7626):471–476.
[Graves, Wayne, and Danihelka 2014] Graves, A.; Wayne,
G.; and Danihelka, I. 2014. Neural turing machines. CoRR
abs/1410.5401.
[Greve, Jacobsen, and Risi 2016a] Greve, R. B.; Jacobsen,
E. J.; and Risi, S. 2016a. Evolving neural turing machines
for reward-based learning. In Proceedings of the Genetic
and Evolutionary Computation Conference 2016, GECCO
’16, 117–124. New York, NY, USA: ACM.
[Greve, Jacobsen, and Risi 2016b] Greve, R. B.; Jacobsen,
E. J.; and Risi, S. 2016b. Evolving neural turing machines
(a)
(b)
Figure 9: Recordings of the activities for bit-sizes 9 (a) and 17 (b) networks. Both networks are produced by the same CPPN.
See main text for details.
for reward-based learning. In Proceedings of the 2016 on
Genetic and Evolutionary Computation Conference, 117–
124. ACM.
[Ha, Dai, and Le 2016] Ha, D.; Dai, A.; and Le, Q. V.
2016. Hypernetworks. arxiv preprint. arXiv preprint
arXiv:1609.09106 2.
[Hornby and Pollack 2002] Hornby, G. S., and Pollack, J. B.
2002. Creating high-level components with a generative representation for body-brain evolution. Artificial Life 8(3).
[Lüders et al. 2017] Lüders, B.; Schläger, M.; Korach, A.;
and Risi, S. 2017. Continual and one-shot learning through
neural networks with dynamic external memory. In European Conference on the Applications of Evolutionary Computation, 886–901. Springer.
[Salimans et al. 2017] Salimans, T.; Ho, J.; Chen, X.; and
Sutskever, I.
2017.
Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint
arXiv:1703.03864.
[Secretan et al. 2008] Secretan, J.; Beato, N.; D’Ambrosio,
D. B.; Rodriguez, A.; Campbell, A.; and Stanley, K. O.
2008. Picbreeder: Evolving pictures collaboratively online.
In CHI ’08: Proceedings of the twenty-sixth annual SIGCHI
conference on Human factors in computing systems, 1759–
1768. New York, NY, USA: ACM.
[Sporns 2002] Sporns, O. 2002. Network analysis, complexity, and brain function. Complexity 8(1):56–60.
[Stanley and Miikkulainen 2002] Stanley, K. O., and Miikkulainen, R. 2002. Evolving neural networks through aug-
menting topologies. Evolutionary Computation 10(2):99–
127.
[Stanley and Miikkulainen 2003] Stanley, K. O., and Miikkulainen, R. 2003. A taxonomy for artificial embryogeny.
Artificial Life 9(2):93–130.
[Stanley and Miikkulainen 2004] Stanley, K. O., and Miikkulainen, R. 2004. Competitive coevolution through evolutionary complexification. 21:63–100.
[Stanley, D’Ambrosio, and Gauci 2009] Stanley, K. O.;
D’Ambrosio, D. B.; and Gauci, J. 2009. A hypercube-based
encoding for evolving large-scale neural networks. Artificial
life 15(2):185–212.
[Stanley 2007] Stanley, K. O. 2007. Compositional pattern
producing networks: A novel abstraction of development.
Genetic Programming and Evolvable Machines 8(2):131–
162.
[Sukhbaatar et al. 2015] Sukhbaatar, S.; Weston, J.; Fergus,
R.; et al. 2015. End-to-end memory networks. In Advances
in neural information processing systems, 2440–2448.
[Verbancsics and Stanley 2011] Verbancsics, P., and Stanley,
K. O. 2011. Constraining connectivity to encourage modularity in hyperneat. In Proceedings of the 13th Annual Conference on Genetic and Evolutionary Computation, GECCO
’11, 1483–1490. New York, NY, USA: ACM.
| 9cs.NE
|
arXiv:1703.09332v3 [math.GR] 26 Dec 2017
ORDERINGS OF WITZEL-ZAREMSKY-THOMPSON GROUPS
TOMOHIKO ISHIDA
Abstract. We prove the orderability of the Witzel-Zaremsky-Thompson group
for a direct system of orderable groups under a certain compatibility assumption.
1. Introduction
Thompson’s groups are interesting countable groups and several kinds of their
generalizations have been developed and studied. As one of such examples, Witzel
and Zaremsky introduced the group T (G∗ ) associated to a direct system (Gn )
of groups satisfying certain axioms called the cloning system [6]. We call it the
Witzel-Zaremsky-Thompson group for (Gn ).
The Witzel-Zaremsky-Thompson group T (G∗ ) can be considered as not only a
generalization of Thompson’s groups but also a limit of the groups Gn . We prove
the group T (G∗ ) inherits the orderability of the groups Gn under the assumption
that there exists an ordering of Gn which are compatible with the cloning maps. A
precise definition of the compatibility is given in Section 2.
Proposition 1.1. Let (Gn ) be a cloning system.
(i) The Witzel-Zaremsky-Thompson group T (G∗ ) for (Gn ) is left-orderable if
the groups Gn admit left-orderings which are compatible with the cloning
maps.
(ii) Suppose that the cloning system is pure. The group T (G∗ ) is bi-orderable
if the groups Gn admit bi-orderings which are compatible with the cloning
maps.
The statement (ii) in Proposition 1.1 is proved by an argument which is a
straightforward generalization of that used in [3]. The author does not know
whether the assumption that the cloning system is pure is essential or not.
2. Preliminaries
Let (Gn ) be a direct system of groups. That is, an injective homomorphism
Gn ֒→ Gn+1 is given for each n ∈ N. The Witzel-Zaremsky-Thompson group
T (G∗ ) is simply a certain subgroup of the cancellative group of the Brin-ZappaSzép product G ⊲⊳ F , where G is the direct limit of the system (Gn ) and F is the
monoid of binary forests. However we recall the definition of T (G∗ ) in a fashion
describing its elements.
Date: December 27, 2017.
2010 Mathematics Subject Classification. Primary 20F60, Secondary 20B07, 20B27, 20E99,
20F36.
Key words and phrases. Thompson’s groups, Thompson-like groups, Witzel-ZaremskyThompson groups, braided Thompson group, orderings, braid groups, Dehornoy ordering.
1
2
TOMOHIKO ISHIDA
Let Sn be the symmetric group on the set {1, . . . , n}. Set the map ςnk : Sn →
Sn+1 by
gm
if m ≤ k, gm ≤ gk,
gm
+
1
if m < k, gm > gk,
ςnk (g)m =
g(m
−
1)
if
m > k, g(m − 1) < gk,
g(m − 1) + 1 if m > k, g(m − 1) ≥ gk.
Definition 2.1. Suppose that a group homomorphism ρn : Gn → Sn and an injective map κkn : Gn → Gn+1 is given for each n ∈ N and for each k = 1, . . . , n.
The triple (Gn ) = ((Gn ), (ρn ), (κkn )) is a cloning system if the following three
axioms are satisfied.
ρ (g)k
(h) for g, h ∈ Gn ,
(1) κkn (gh) = κkn (g)κnn
k
(2) κkn+1 ◦ κln = κl+1
◦
κ
n for 1 ≤ k < l ≤ n, and
n+1
(3) ρn (κkn (g)) = ςnk (ρn (g)) or sk ςnk (ρn (g)) for g ∈ Gn and for 1 ≤ k ≤ n, where
the symbol sk means the transposition (k, k + 1) ∈ Sn+1 .
The maps κkn : Gn → Gn+1 are called the cloning maps of the cloning system (Gn ).
Suppose that a cloning system (Gn ) = ((Gn ), (ρn ), (κkn )) is given. A tree diagram
is a triple (T− , g, T+ ), where g ∈ Gn for certain n ∈ N and T± are rooted planar
binary trees with n leaves. A simple expansion of (T− , g, T+ ) is a tree diagram
of the form (T−′ , κkn (g), T+′ ). Here, T−′ and T+′ are trees obtained by adjoining
carets to the k-th and (ρ(g)k)-th leaves of T− and T+ , respectively. An expansion
of (T− , g, T+ ) is a tree diagram obtained as an iterated simple expansions. Two
tree diagrams are equivalent if they have a common expansion. The equivalence
class represented by a tree diagram (T− , g, T+ ) will be denoted by [T− , g, T+ ]. The
Witzel-Zaremsky-Thompson group T (G∗ ) consists of equivalence classes of tree
diagrams as a set. For two elements [S− , f, S+ ], [T− , g, T+ ] ∈ T (G∗ ), their product
′
′
) and (T−′ , g ′ , T+′ )
, f ′ , S+
is defined as follows. There always exist tree diagrams (S−
′
representing [S− , f, S+ ] and [T− , g, T+ ], respectively and such that S+
= T−′ . The
′
, f ′ g ′ , T+′ ]. This is well-defined.
product [S− , f, S+ ][T− , g, T+ ] is defined to be [S−
The cloning system (Gn ) is pure if the homomorphism ρn : Gn → Sn is trivial for
each n ∈ N.
The subgroup of T (G∗ ) consisting of elements represented by the form (T− , 1, T+ )
is isomorphic to the Thompson’s group F . and we denote this subgroup again by
F . If we denote by P T (G∗ ) the kernel of natural projection p : T (G∗ ) → V , then
the group P T (G∗ ) consists of the elements of the form [T, g, T ], where g ∈ Kerρn
and n is the number of leaves of T . If the cloning system (Gn ) is pure, then the
image of the projection p : T (G∗ ) → V is equal to Thompson’s group F < V and
thus there exists an isomorphism T (G∗ ) ∼
= P T (G∗ ) ⋊ F .
When we fix a left-ordering or a bi-ordering <n of Gn for each n, we say the
orderings are compatible with the cloning maps κkn ’s if 1 <n+1 κkn (g) for any n and
any k whenever 1 <n g.
3. Proof of the main proposition
Proof of Proposition 1.1. Let (Gn ) = ((Gn ), (ρn ), (κkn )) be a cloning system. Suppose that a left-ordering <n of the group Gn is given for each n and the orderings
<n are compatible with the cloning maps. Since the Thompson’s group F is known
ORDERINGS OF WITZEL-ZAREMSKY-THOMPSON GROUPS
3
to be bi-orderable, we fix a left-ordering <F of F . Set the subset Π of the WitzelZaremsky-Thompson group T (G∗ ) for (Gn ) by
Π = {[T− , g, T+ ] ∈ T (G∗ ); g ∈ Gn , 1 <n g} ∐ {f ∈ F ; 1 <F f }.
Since the orders <n are compatible with the cloning maps, the set Π is well-defined.
Further it is easy to verify that Π is a positive cone and thus defines a left-ordering
of T (G∗ ). This completes the proof of the statement (i).
If the groups Gn admit bi-orderings which are compatible with the cloning maps,
then the bi-ordering on P T (G∗ ) is induced. Under the assumption that the cloning
system (Gn ) is pure, since both the groups P T (G∗ ) and F are bi-orderable, the
group T (G∗ ) which is isomorphic to the semi-direct product P T (G∗ ) ⋊ F is also.
This completes the proof of the statement (ii).
Remark 3.1. If we fix a left-ordering or bi-ordering on the group T (G∗ ), its restriction on P T (G∗ ) uniquely determines a family of left-ordering or bi-ordering
respectively on the groups Gn . In fact, if for each rooted planar binary tree T with
n leaves, we denote by GT the subgroup of P T (G∗ ) consisting of elements represented by tree diagrams of the form (T, g, T ), where g ∈ Gn , then Gn is isomorphic
to Gn . Furthermore, if T and T ′ are rooted planar binary trees with same number
of leaves, then in T (G∗ ) the subgroups group GT and G′T are conjugate to each
other by an element of F . Thus if we assume the cloning system (Gn ) is pure, the
converse of each statement of Proposition 1.1 always holds.
4. Examples
In this section we discuss examples of Witzel-Zaremsky-Thompson groups to
which Proposition 1.1 is applicable.
4.1. The braided Thompson group. Let Bn be the braid group of n strands and
σ1 , . . . , σn−1 ∈ Bn the Artin generators. Taking ρn : Bn → Sn to be the natural
projection and setting κkn : Bn → Bn+1 by
σi+1
if k < i,
σi+1 σi if k = i,
k
κn (σi ) =
σi σi+1 if k = i + 1,
σi
if i + 1 < k,
we have a well-defined cloning system on (Bn ). Then the Witzel-Zaremsky-Thompson
group T (Bn ) for (Bn ) is isomorphic to the group called the braided Thompson
group BV which was independently introduced by Brin and Dehornoy [2][4].
Recall that β ∈ Bn is positive with respect to the Dehornoy ordering if and only
if β is represented by a word w in the Artin generators which satisfies the following
condition (D):
±1
(D) σi occurs in w but σ1±1 , . . . , σi−1
, σi−1 does not for certain i.
It is not difficult to verify only by definition of the cloning maps that if w is
a representation of β satisfying the condition (D) then κkn (w) also. Hence the
following lemma holds:
Lemma 4.1. The Dehornoy orderings of the braid groups Bn are compatible with
the cloning maps of (Bn ) defined above.
By Proposition 1.1 and Lemma 4.1, we have in easy way the following theorem
which was first proved in [4].
4
TOMOHIKO ISHIDA
Theorem 4.2. The braided Thompson group BV is left-orderable.
4.2. The pure braided Thompson group. Let Pn be the pure braid group of
n strands. We also denote the restriction on Pn of the map κkn : Bn → Sn we
set in the previous subsection again by κkn . Take ρn : Pn → Sn to be the trivial
homomorphism. Then we have a well-defined cloning system on (Pn ) and the
Witzel-Zaremsky-Thompson group T (Pn ) for (Pn ) which is isomorphic to the pure
braided Thompson group BF introduced by Brady, Burillo, Cleary and Stein in
[1].
Lemma 4.3 ([3]). The orderings on the braid groups Pn induced from the Magnus
orderings of the free groups via the Artin combing are compatible with the cloning
maps on (Pn ) defined above.
Theorem 4.4 ([3]). The pure braided Thompson group BF is bi-orderable.
4.3. Witzel-Zaremsky-Thompson group for direct powers of a group. For
arbitrary group G, set Gn to be the n-th direct power Gn of G. Fix injective
homomorphisms φ1 , φ2 of G. If we define κkn : Gn → Gn+1 by κkn (g1 , . . . , gn ) =
(g1 , . . . , φ1 (gk ), φ2 (gk ), . . . , gn ) and set ρn : Gn → Sn to be the trivial homomorphism, then we have a cloning system on direct powers (Gn ) of the group G and the
Witzel-Zaremsky-Thompson group T (Gn ) which was introduced by Tanushevski
[5]. Suppose that G is left-orderable and fix a left-ordering < of G. If the homomorphism φ1 : G → G preserves the order of G, then the lexicographic orderings
of Gn ’s induced by < are compatible with the cloning maps of (Gn ). Further if
the order on G is bi-invariant, then the induced orders on Gn are also. Since the
identity map of G trivially preserves all the orderings of G, by Proposition 1.1 we
have the following Theorem:
Theorem 4.5. If the group G is left-orderable or bi-orderable, then the WitzelZaremsky-Thompson group T (Gn ) of the direct powers of G is also left-orderable
or bi-orderable, respectively.
Acknowledgments. The author is grateful to the referee for carefully reading the
manuscript and pointing out mistake in it. The author was supported by JSPS
Research Fellowships for Young Scientists (26·110).
References
[1] T. Brady, J. Burillo, S. Cleary, and M. Stein, Pure braid subgroups of braided Thompson’s
groups, Publ. Mat. 52 (2008), no. 1, 57–89. MR 2384840
[2] M. G. Brin, The algebra of strand splitting. I. A braided version of Thompson’s group V , J.
Group Theory 10 (2007), no. 6, 757–788. MR 2364825
[3] J. Burillo and J. González-Meneses, Bi-orderings on pure braided Thompson’s groups, Q. J.
Math. 59 (2008), no. 1, 1–14. MR 2392498
[4] P. Dehornoy, The group of parenthesized braids, Adv. Math. 205 (2006), no. 2, 354–409.
MR 2258261
[5] S. Tanushevski, A new class of generalized Thompson’s groups and their normal subgroups,
Comm. Algebra 44 (2016), no. 10, 4378–4410. MR 3508306
[6] S. Witzel and C. B. Zaremsky, Thompson groups for systems of groups, and their finiteness
properties, preprint, arXiv:1405.5491.
Department of Mathematics, Kyoto University, Kitashirakawa Oiwake-cho, Sakyoku, Kyoto 606-8502, Japan.
E-mail address: [email protected]
| 4math.GR
|
Evolutionary Multimodal Optimization: A Short Survey
Ka-Chun Wong (Department of Computer Science, University of Toronto)
arXiv:1508.00457v1 [cs.NE] 3 Aug 2015
August 4, 2015
Real world problems always have different multiple solutions. For instance, optical engineers need to tune the recording parameters to get as many optimal solutions as possible
for multiple trials in the varied-line-spacing holographic grating design problem. Unfortunately, most traditional optimization techniques focus on solving for a single optimal
solution. They need to be applied several times; yet all solutions are not guaranteed to
be found. Thus the multimodal optimization problem was proposed. In that problem,
we are interested in not only a single optimal point, but also the others. With strong
parallel search capability, evolutionary algorithms are shown to be particularly effective
in solving this type of problem. In particular, the evolutionary algorithms for multimodal
optimization usually not only locate multiple optima in a single run, but also preserve their
population diversity throughout a run, resulting in their global optimization ability on multimodal functions. In addition, the techniques for multimodal optimization are borrowed
as diversity maintenance techniques to other problems. In this chapter, we describe and
review the state-of-the-arts evolutionary algorithms for multimodal optimization in terms
of methodology, benchmarking, and application.
1
Introduction and Background
Since genetic algorithm was proposed by John H. Holland [9] in the early 1970s, researchers
have been exploring the power of evolutionary algorithms [47]. For instance, biological pattern discovery [48] and computer vision [51]. In particular, its function optimization capability was highlighted [6] because of its high adaptability to different non-convex function
landscapes, to which we cannot apply traditional optimization techniques.
Real world problems always have different multiple solutions [45, 46]. For instance,
optical engineers need to tune the recording parameters to get as many optimal solutions
as possible for multiple trials in the varied-line-spacing holographic grating design problem
because the design constraints are too difficult to be expressed and solved in mathematical
forms [31]. Unfortunately, most traditional optimization techniques focus on solving for a
single optimal solution. They need to be applied several times; yet all solutions are not
guaranteed to be found. Thus the multimodal optimization problem was proposed. In that
1
problem, we are interested in not only a single optimal point, but also the others. Given
an objective function, an algorithm is expected to find all optimal points in a single run.
With strong parallel search capability, evolutionary algorithms are shown to be particularly
effective in solving this type of problem [6]: Given function f : X → R, we would like to
find all global and local maxima (or minima) of f in a single run. Although the objective
is clear, it is not easy to be satisfied in practice because some problems may have too
many optima to be located. Nonetheless, it is still of great interest to researchers how
these problems are going to be solved because the algorithms for multimodal optimization
usually not only locate multiple optima in a single run, but also preserve their population
diversity throughout a run, resulting in their global optimization ability on multimodal
functions. Moreover, the techniques for multimodal optimization are usually borrowed as
diversity maintenance techniques to other problems [42, 50].
2
Problem Definition
The multimodal optimization problem definition depends on the type of optimization (minimization or maximization). They are similar in principle and defined as follows:
2.1
Minimization
In this problem, given f :X → R, we would like to find all global and local minimums of f
in a single run.
Definition 1 Local Minimum [41]: A (local) minimum x̂l ∈ X of one (objective) function f :X → R is an input element with f (x̂l ) ≤ f (x) for all x neighboring x̂l . If X ∈ RN ,
we can write: ∀x̂l ∃ > 0 : f (x̂l ) ≤ f (x) ∀x ∈ X, |x − x̂l | < .
Definition 2 Global Minimum [41]: A global minimum xˆg ∈ X of one (objective) function
f :X → R is an input element with f (xˆg ) ≤ f (x) ∀x ∈ X.
2.2
Maximization
In this problem, given f :X → R, we would like to find all global and local maximums of f
in a single run.
Definition 3 Local Maximum [41]: A (local) maximum x̂l ∈ X of one (objective) function
f :X → R is an input element with f (x̂l ) ≥ f (x) for all x neighboring x̂l . If X ∈ RN , we
can write: ∀x̂l ∃ > 0 : f (x̂l ) ≥ f (x) ∀x ∈ X, |x − x̂l | < .
Definition 4 Global Maximum [41]: A global maximum xˆg ∈ X of one (objective) function
f :X → R is an input element with f (xˆg ) ≥ f (x) ∀x ∈ X.
2
3
Methodology
In the past literature, there are different evolutionary methods proposed for multimodal
optimization. In this section, we discuss and categorize them into different methodologies.
3.1
Preselection
In 1970, the doctorial thesis by Cavicchio introduced different methods for genetic algorithms [3]. In particular, the preselection scheme was proposed to maintain the population
diversity. In this scheme, the children compete with their parents for survival. If a child
has a fitness (measured by an objective function) higher than its parent, the parent is
replaced by the child in the next generation.
3.2
Crowding
In 1975, the work by De Jong [13] introduced the crowding technique to increase the chance
of locating multiple optima. In the crowding technique, each child is compared to a random sub-population of cf members in the existing parent population (cf means crowding
factor). The parent member which is most similar to the child itself is selected (measured
by a distance metric). If the child has a higher fitness than the parent member selected,
then it replaces the parent member in the population. Besides genetic algorithm, Thomsen
has also incorporated crowding techniques [13] into differential evolution (CrowdingDE) for
multimodal optimization [40]. In his study, the crowding factor is set to the population size
and Euclidean distance is adopted as the dissimilarity metric. The smaller the distance,
the more similar they are and vice versa. Although an intensive computation is accompanied, it can effectively transform differential evolution into an algorithm specialized for
multimodal optimization. In 2012, CrowdingDE has been investigated and extended by
Wong et al , demonstrating competitive performance even when it is compared to the other
state-of-the-arts methods [49].
3.3
Fitness Sharing
In 1989, Goldberg and Richardson proposed a fitness-sharing niching technique as a diversity preserving strategy to solve the multimodal optimization problem [7]. They proposed
a shared fitness function, instead of an absolute fitness function, to evaluate the fitness of
a individual in order to favor the growth of the individuals which are distinct from the
others. The shared fitness function is defined as follows:
f 0 (xi ) = Shared F itness =
Actual F itness
f (xi )
= PN
Degree of Sharing
j=1 sh(d(xi , xj ))
where f 0 (xi ) is the shared fitness of the ith individual xi ; f (xi ) is the actual fitness of the
ith individual xi ; d(xi , xj ) is the distance function between the two individuals xi and xj ;
3
sh(d) is the sharing function. With this technique, a population can be prevented from
the domination of a particular type of individuals. Nonetheless, a careful adjustment to
the sharing function sh(d) definition is needed because it relates the fitness domain f (xi )
to the distance domain d(xi , xj ) which are supposed to be independent of each other.
3.4
Species Conserving
Species conserving genetic algorithm (SCGA) [17] is a technique for evolving parallel subpopulations for multimodal optimization. Before each generation starts, the algorithm
selects a set of species seeds which can bypass the subsequent procedures and be saved into
the next generation. The algorithm then divides a population into several species based
on a dissimilarity measure. The fittest individual is selected as the species seed for each
species. After the identification of species seeds, the population undergoes the usual genetic
algorithm operations: selection, crossover, and mutation. As the operations may remove
the survival of less fit species, the saved species seeds are copied back to the population at
the end of each generation.
To determine the species seeds in a population, the algorithm first sorts the population
in a decreasing fitness order. Once sorted, it picks up the fittest individual as the first
species seed and forms a species region around it. The next fittest individual is tested
whether it is located in a species region. If not, it is selected as a species seed and another
species region is created around it. Otherwise, it is not selected. Similar operations are
applied to the remaining individuals, which are subsequently checked against all existing
species seeds.
To copy the species seeds back to the population after the genetic operations have been
executed, the algorithms need to scan all the individuals in the current population and
identify to which species they belong. Once it is identified, the algorithm replaces the
worst individual (lowest fitness) with the species seed in a species. If no individuals can
be found in a species for replacement, the algorithm replaces the worst and un-replaced
individual in the whole population. In short, the main idea is to preserve the population
diversity by preserving the fittest individual for each species.
3.5
Covariance Matrix Adaptation
Evolution strategy is an effective method for numerical optimization. In recent years, its
variant CMA-ES (Covariance Matrix Adaptation Evolution Strategy) showed a remarkable
success [8]. To extend its capability, niching techniques have been introduced to cope
with multimodal functions [35]. For instance, a concept called adaptive individual niche
radius has been proposed to solve the niche radius problem commonly found in speciation
algorithms [34].
4
3.6
Multiobjective Approach
At this point, we would like to note that the readers should not confuse evolutionary
multimodal optimization (main theme of this chapter) with evolutionary multiobjective optimization. The former aims at solving a single function for multiple opima, while
the latter aims at solving multiple functions for pareto front solutions. Nonetheless, the
techniques involved are related. In particular, Deb and Saha demonstrated that, by decomposing a single multimodal objective function problem into a bi-objective problem, they
can solve a multimodal function using a evolutionary multiobjective optimization algorithm
[4]. Briefly, they keep the original multimodal objective function as the first objective. On
the other hand, they use the gradient information to define peaks in the second objective.
3.7
Ensemble
As mentioned in the previous section, different niching algorithms have been proposed
over the past years. Each algorithm has its own characteristics and design philosophy
behind. Although it imposes difficult conditions to compare them thoroughly, it is a
double-edged sword. Such a vast amount of algorithms can provide us a “swiss army
knife” for optimizations on different problems. In particular, Yu and Suganthan proposed
an ensemble method to combine those algorithms and form a powerful method called
Ensemble of Niching Algorithms (ENA) [52]. An extension work can also be found in [32].
3.8
Others
Researchers have been exploring many different ways to deal with the problem. Those
methods include: clearing [29], repeated iterations [1], species-specific explosion [43], traps
[15], stochastic automation [25], honey bee foraging behavior [39], dynamic niching [27],
spatially-structured clearing [5], cooperative artificial immune network [21], particle swarm
optimization [10, 19, 14, 22] , and island model [2]. In particular, Stoean et al. have proposed a topological species conservation algorithm in which the proper topological separation into subpopulations has given it an advantage over the existing radius-based algorithms
[38]. Comparison studies were conducted by Singh et al. [36], Kronfeld et al. [16], and
Yu et al. [53]. Though different methods were proposed in the past, they are all based
on the same fundamental idea: it is to strike an optimal balance between convergence
and population diversity in order to locate multiple optima simultaneously in a single run
[37, 44].
5
4
Benchmarking
4.1
Benchmark Functions
There are many multimodal functions proposed for benchmarking in the past literature.
In particular, the following five benchmark functions are widely adopted in literature:
Deb’s 1st function [43], Himmelblau function [1], Six-hump Camel Back function [24],
Branin function [24], and Rosenbrock function [33]. In addition, five more benchmark
functions (PP1 to PP5) can be found in [43, 23]. For more viogorous comparisons, the IEEE
Congress on Evolutionary Computation (CEC) usually releases a test suite for multimodal
optimization every year. More than 15 test functions can be found there [20].
4.2
Performance Metrics
Several performance metrics have been proposed in the past literature [23, 18, 17, 40].
Among them, Peak Ratio (PR) and Average Minimum Distance to the Real Optima (D)
[23, 43] are commonly adopted as the performance metrics.
• A peak is considered found when there exists a individual which is within 0.1 Euclidean distance to the peak in the last population. Thus the Peak Ratio is calculated
using the equation (3):
P eak Ratio =
N umber of peaks f ound
T otal number of peaks
(3)
• The average minimum distance to the optima (D) is calculated using the equation
(4):
n
X
min d(peaki , indiv)
i=1
indiv∈pop
(4)
n
where n is the number of peaks, indiv denotes a individual, peaki is the ith peak, pop
denotes the last population, and d(peaki , indiv) denotes the distance between peaki
and indiv.
D=
As different algorithms perform different operations in one generation, it is unfair to set the
termination condition as the number of generations. Alternatively, it is also unfair to adopt
CPU time because it substantially depends on the implementation techniques for different
algorithms. For instance, the sorting techniques to find elitists and the programming languages used. In contrast, fitness function evaluation is always the performance bottleneck
[28] 1 . Thus the number of fitness function evaluations is suggested to be adopted as the
running or termination condition for convergence analysis.
1
For instance, over ten hours are needed to evaluate a calculation in computational fluid dynamics [11]
6
4.3
Statistical Tests
Since evolutionary multimodal optimization is stochastic in nature, multiple runs are
needed to evaluate each method on each test function. The means and standard deviations
of performance metrics are usually reported for fair comparison. To justify the results,
statistical tests are usually adopted to assess the statistical significances. For instance,
t-tests, Mann-Whitney U-tests (MWU), and Kolmogorov-Smirnov test (KS).
5
Application
Holographic gratings have been widely used in optical instruments for aberration corrections. In particular, Varied-Line-Spacing (VLS) holographic grating is distinguished by the
high order aberration eliminating capability in diffractive optical systems. It is commonly
used in high resolution spectrometers and monochromaters. A recording optical system of
VLS holographic grating is outlined in [31].
5.1
Problem Modelling
The core component descriptions of the optical systems are listed as follows [30]:
M1 , M2 : Two spherical mirrors
C, D
: Two coherent point sources
G
: A grating blank
In this system, there are two light point sources C and D. They emit light rays which are
then reflected by mirrors M1 and M2 respectively. After the reflection, the light rays are
projected onto the grating blank G. More details are given in [31, 26]. The objective for
the design is to find several sets of design variables (or recording parameters [31]) to form
the expected groove shape of G (or the distribution of groove density [30]). The design
variables are listed as follows:
γ :
ηC :
δ :
ηD :
pC :
qC :
pD :
qD :
The
The
The
The
The
The
The
The
incident angle of the ray O1 O
incident angle of the ray CO1
incident angle of the ray O2 O
incident angle of the ray DO2
distance between C and M1 (CO1 )
distance between M1 and G (O1 O)
distance between D and M2 (DO2 )
distance between M2 and G (O2 O)
7
Mathematically, the goal is to minimize the definite integral of the square
R w error between
the expected groove density and practical groove density [31]: min J = −w0 0 (np − ne )2 dw
where w0 is the half-width of the grating, np is the practical groove density, and ne is
the expected groove density. These two groove densities are complicated functions of the
design variables. Ling et al. have further derived the above formula into a simpler one [30]:
min J = r12 +
w02 (2r1 r3 + r22 ) w04 (r32 + 2r2 r4 ) w06 r42
+
+
3
5
7
j20
− n0 b2
λ0
j40
r4 =
− n0 b4
2λ0
j10
− n0 ,
λ0
3j30
r3 =
− n0 b3 ,
2λ0
r2 =
r1 =
where j10 , j20 , j30 and j40 are the functions of the design variables, which are n10 , n20 ,
n30 and n40 respectively in [26]. Theoretically, the above objective is simple and clear.
Unfortunately, there are many other auxiliary optical components in practice, which constraints are too difficult to be expressed and solved in mathematical forms. An optimal
solution is not necessarily a feasible and favorable solution. Optical engineers often need to
tune the design variables to find as many optimal solutions as possible for multiple trials.
Multimodal optimization becomes necessary for this design problem.
5.2
Performance measurements
As the objective function is an unknown landscape, the exact optima information is not
available. Thus the previous performance metrics cannot be adopted. We propose two
new performance metrics in this section. The first one is the best fitness, which is the
fitness value of the fittest individual in the last population. The second one is the number
of distinct peaks, where a distinct peak is considered found when there exists a individual
which fitness value is below a threshold 0.0001 and there isn’t any other individual found as
a peak before within 0.1 Euclidean distance in the last population. The threshold is chosen
to 0.0001 because the fitness values of the solutions found in [31] are around this order of
magnitude. On the other hand, the distance is chosen to 0.1 Euclidean distance because
it has already been set for considering peaks found in peak ratio [43, 23]. Nonetheless, it
is undeniable that such a threshold may not be suitable for this application because the
landscape is unknown, although the value of 0.1 is the best choice we can adopt in this
study.
5.3
Parameter setting
CrowdingDE-STL [49], CrowdingDE-TL [49], CrowdingDE-SL [49], Crowding Genetic Algorithm (CrowdingGA) [13], CrowdingDE [40], Fitness Sharing Genetic Algorithm (SharingGA) [7], SharingDE [40], Species Conserving Genetic Algorithm (SCGA) [17], SDE
8
Table 1: Results for all algorithms tested on the VLS holographic grating design problem
(50 runs)
Measurement
Mean of Best Fitness
StDev of Best Fitness
Mean of Peaks Found
StDev of Peaks Found
Measurement
Mean of Best Fitness
StDev of Best Fitness
Mean of Peaks Found
StDev of Peaks Found
CrowdingDE-STL [49]
8.29E-08
2.91E-07
41.42
13.07
SharingGA [7]
1.87E+04
6.82E+04
0.0
0.0
CrowdingDE-TL [49]
7.17E-07
5.01E-06
45.54
9.00
SharingDE [40]
1.74E+02
2.65E+02
0.0
0.0
CrowdingDE-SL [49]
1.18E-07
4.04E-07
43.38
10.69
SDE [18]
1.13E+00
1.56E+00
0.06
0.31
CrowdingGA [13]
9.02E-06
3.40E-05
8.94
5.04
SCGA [17]
1.24E+02
4.59E+02
0.02
0.14
CrowdingDE [40]
3.66E-06
2.31E-05
41.98
14.26
UN [12]
9.19E-04
3.15E-03
7.22
3.92
[18], and UN [12] are selected for illustrative purposes in this application. All the algorithms were run up to a maximum of 10000 fitness function evaluations. The above performance metrics were obtained by taking the average and standard deviation of 50 runs.
The groove density parameters followed the setting in [31]: n0 = 1.400 × 103 (line/mm),
b2 = 8.2453 × 10−4 (1/mm), b3 = 3.0015 × 10−7 (1/mm2 ) and b4 = 0.0000 × 10−10 (1/mm3 ).
The half-width w0 was 90mm. The radii of spherical mirrors M1 and M2 were 1000mm.
The recording wavelength (λ0 ) was 413.1nm. The population size of all the algorithms
was set to 50. The previous settings remained the same, except the algorithm-specific
parameters: The species distance of SDE and SCGA was set to 1000. The scaling factor
and niche radius of SharingDE and SharingGA were set to 1 and 1000 respectively. The
discount factor of the temporal locality was set to 0.5. The survival selection method of
the non-crowding algorithms was set to binary tournament [12].
5.4
Results
The result is tabulated in Table 1. It can be observed that CrowdingDE-STL can achieve
the best fitness whereas CrowdingDE-TL can acheive the best number of peaks found. To
compare the algorithms rigorously, statistical tests have also been used. The results are
depicted in Figure 1. One can observe that there are some statistically significant performance differences among them. In particular, CrowdingDE-based methods are shown to
have the results statistically different from CrowdingGA, SharingGA, SharingDE, SDE,
and SCGA. Some configurations obtained after a run of CrowdingDE-STL on this problem
are depicted in Figure 2. It can be seen that they are totally different and feasible configurations with which optical engineers can feel free to perform multiple trials after the single
run.
6
Discussion
To conclude, we have briefly reviewed the state-of-the-arts methods of evolutionary multimodal optimization from different perspectives in this chapter. Different evolutionary
9
multimodal optimization methodologies are described. To compare them fairly, we described different benchmarking techniques such as performance metrics, test functions,
and statistical tests. An application to Varied-Line-Spacing (VLS) holographic grating
is presented to demonstrate the real-wold applicability of evolutionary multimodal optimization. Nonetheless, we would like to note several current limitations of evolutionary
multimodal optimization as well as the possible solutions at the end of this chapter.
First, most of the past studies just focus on low dimensional test functions for benchmarking. More high dimensional test functions should be incorporated in the future. Second, we would like to point out that evolutionary multimodal optimization is actually
far from just finding multiple optima because the algorithms for multimodal optimization
usually not only locate multiple optima in a single run, but also preserve their population
diversity throughout a run, resulting in their global optimization ability on multimodal
functions. Moreover, the techniques for multimodal optimization are usually borrowed as
diversity maintenance techniques to other problems. Third, the computational complexities of the methods are usually very high comparing with the other methods since they
involve population diversity maintenance which implies that the related survival operators
need to take into account the other individuals, resulting in additional time complexity.
References
[1] David Beasley, David R. Bull, and Ralph R. Martin. A sequential niche technique for
multimodal function optimization. Evol. Comput., 1(2):101–125, 1993.
[2] Mourad Bessaou, Alain Pétrowski, and Patrick Siarry. Island model cooperating with
speciation for multimodal optimization. In PPSN VI: Proceedings of the 6th International Conference on Parallel Problem Solving from Nature, pages 437–446, London,
UK, 2000. Springer-Verlag.
[3] Daniel Joseph Cavicchio. Adaptive search using simulated evolution. 1970.
[4] Kalyanmoy Deb and Amit Saha. Finding multiple solutions for multimodal optimization problems using a multi-objective evolutionary approach. In Proceedings of
the 12th Annual Conference on Genetic and Evolutionary Computation, GECCO ’10,
pages 447–454, New York, NY, USA, 2010. ACM.
[5] G. Dick. Automatic identification of the niche radius using spatially-structured clearing methods. In 2010 IEEE Congress on Evolutionary Computation (CEC), pages 1
–8, july 2010.
[6] David E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1989.
10
[7] David E. Goldberg and Jon Richardson. Genetic algorithms with sharing for multimodal function optimization. In Proceedings of the Second International Conference
on Genetic algorithms and their application, pages 41–49, Hillsdale, NJ, USA, 1987.
L. Erlbaum Associates Inc.
[8] Nikolaus Hansen and Andreas Ostermeier. Completely derandomized self-adaptation
in evolution strategies. Evol. Comput., 9:159–195, June 2001.
[9] John H. Holland. Adaptation in natural and artificial systems. MIT Press, Cambridge,
MA, USA, 1992.
[10] Zhen Ji, Huilian Liao, Yiwei Wang, and Q.H. Wu. A novel intelligent particle optimizer
for global optimization of multimodal functions. In CEC 2007. IEEE Congress on
Evolutionary Computation, 2007, pages 3272 –3275, sept. 2007.
[11] Yaochu Jin, M. Olhofer, and B. Sendhoff. A framework for evolutionary optimization
with approximate fitness functions. IEEE Transactions on Evolutionary Computation,
6(5):481–494, 2002.
[12] Kenneth A. De Jong. Evolutionary Computation. A Unified Approach. MIT Press,
Cambridge, MA, USA, 2006.
[13] Kenneth Alan De Jong. An analysis of the behavior of a class of genetic adaptive
systems. PhD thesis, University of Michigan, Ann Arbor, MI, USA, 1975.
[14] Yau-Tarng Juang, Shen-Lung Tung, and Hung-Chih Chiu. Adaptive fuzzy particle
swarm optimization for global optimization of multimodal functions. Information
Sciences, 181(20):4539 – 4549, 2011. Special Issue on Interpretable Fuzzy Systems.
[15] Naoya Karatsu, Yuichi Nagata, Isao Ono, and Shigenobu Kobayashi. Globally multimodal function optimization by real-coded genetic algorithms using traps. In 2010
IEEE Congress on Evolutionary Computation (CEC), pages 1–8, 2010.
[16] M. Kronfeld and A. Zell. Towards scalability in niching methods. In 2010 IEEE
Congress on Evolutionary Computation (CEC), pages 1 –8, july 2010.
[17] Jian Ping Li, Marton E. Balazs, Geoffrey T. Parks, and P. John Clarkson. A species
conserving genetic algorithm for multimodal function optimization. Evol. Comput.,
10(3):207–234, 2002.
[18] Xiaodong Li. Efficient differential evolution using speciation for multimodal function
optimization. In GECCO ’05: Proceedings of the 2005 conference on Genetic and
evolutionary computation, pages 873–880, New York, NY, USA, 2005. ACM.
11
[19] Xiaodong Li. Niching without niching parameters: Particle swarm optimization using
a ring topology. IEEE Transactions on Evolutionary Computation, 14(1):150 –169,
feb. 2010.
[20] Xiaodong Li, Ke Tang, Mohammad N. Omidvar, Zhenyu Yang, and Kai Qin. Benchmark functions for the cec 2013 special session and competition on large-scale global
optimization, 2013.
[21] Li Liu and Wenbo Xu. A cooperative artificial immune network with particle swarm
behavior for multimodal function optimization. In CEC 2008. (IEEE World Congress
on Computational Intelligence). IEEE Congress on Evolutionary Computation, 2008,
pages 1550 –1555, june 2008.
[22] Lili Liu, Shengxiang Yang, and Dingwei Wang. Force-imitated particle swarm optimization using the near-neighbor effect for locating multiple optima. Information
Sciences, In Press, Uncorrected Proof:–, 2010.
[23] Rodica I. Lung, Camelia Chira, and D. Dumitrescu. An agent-based collaborative
evolutionary model for multimodal optimization. In GECCO ’08: Proceedings of the
2008 GECCO conference companion on Genetic and evolutionary computation, pages
1969–1976, New York, NY, USA, 2008. ACM.
[24] Zbigniew Michalewicz. Genetic algorithms + data structures = evolution programs
(3rd ed.). Springer-Verlag, London, UK, 1996.
[25] K. Najim and A.S. Poznyak. Multimodal searching technique based on learning automata with continuous input and changing number of actions. IEEE Transactions
on Systems, Man, and Cybernetics, Part B: Cybernetics, 26(4):666 –673, aug. 1996.
[26] Takeshi Namioka and Masato Koike. Aspheric wave-front recording optics for holographic gratings. Appl. Opt., 34(13):2180–2186, 1995.
[27] A. Nickabadi, M.M. Ebadzadeh, and R. Safabakhsh. Evaluating the performance of
dnpso in dynamic environments. pages 2640 –2645, oct. 2008.
[28] Yew S. Ong, Prasanth B. Nair, and Andrew J. Keane. Evolutionary optimization of
computationally expensive problems via surrogate modeling. AIAA Journal, 41(4):687
–696, 2003.
[29] A. Petrowski. A clearing procedure as a niching method for genetic algorithms. In
Proceedings of IEEE International Conference on Evolutionary Computation, 1996,
pages 798–803, Nagoya, Japan, May 1996.
[30] Ling Qing, Wu Gang, and Wang Qiuping. Restricted evolution based multimodal
function optimization in holographic grating design. In The 2005 IEEE Congress
12
on Evolutionary Computation, 2005, volume 1, pages 789–794, Edinburgh, Scotland,,
September 2005.
[31] Ling Qing, Wu Gang, Yang Zaiyue, and Wang Qiuping. Crowding clustering genetic
algorithm for multimodal function optimization. Appl. Soft Comput., 8(1):88–95, 2008.
[32] Bo-Yang Qu and P.N. Suganthan. Novel multimodal problems and differential evolution with ensemble of restricted tournament selection. In 2010 IEEE Congress on
Evolutionary Computation (CEC), pages 1 –7, july 2010.
[33] Yun-Wei Shang and Yu-Huang Qiu. A note on the extended rosenbrock function.
Evol. Comput., 14(1):119–126, 2006.
[34] Ofer Shir and Thomas Back. Niche radius adaptation in the cma-es niching algorithm.
In Thomas Runarsson, Hans-Georg Beyer, Edmund Burke, Juan Merelo-Guervos,
L. Whitley, and Xin Yao, editors, Parallel Problem Solving from Nature - PPSN IX,
volume 4193 of Lecture Notes in Computer Science, pages 142–151. Springer Berlin /
Heidelberg, 2006.
[35] Ofer M. Shir, Michael Emmerich, and Thomas Bäck. Adaptive niche radii and niche
shapes approaches for niching with the cma-es. Evol. Comput., 18:97–126, March
2010.
[36] Gulshan Singh and Dr. Kalyanmoy Deb. Comparison of multi-modal optimization
algorithms based on evolutionary algorithms. In GECCO ’06: Proceedings of the 8th
annual conference on Genetic and evolutionary computation, pages 1305–1312, New
York, NY, USA, 2006. ACM.
[37] M. Srinivas and L.M. Patnaik. Adaptive probabilities of crossover and mutation in
genetic algorithms. IEEE Transactions on Systems, Man and Cybernetics, 24(4):656
–667, apr. 1994.
[38] C. Stoean, M. Preuss, R. Stoean, and D. Dumitrescu. Multimodal optimization by
means of a topological species conservation algorithm. IEEE Transactions on Evolutionary Computation, 14(6):842 –864, dec. 2010.
[39] K. Sundareswaran and V.T. Sreedevi. Development of novel optimization procedure
based on honey bee foraging behavior. pages 1220 –1225, oct. 2008.
[40] R. Thomsen. Multimodal optimization using crowding-based differential evolution.
In CEC2004. IEEE Congress on Evolutionary Computation, 2004, volume 2, pages
1382–1389, June 2004.
[41] Thomas Weise. Global Optimization Algorithms ? Theory and Application. Thomas
Weise, july 16, 2007 edition, July 2007. Online available at http://www.itweise.de/projects/book.pdf.
13
[42] Ka-Chun Wong, Tak-Ming Chan, Chengbin. Peng, Yue Li, and Zhaolei Zhang. DNA
motif elucidation using belief propagation. Nucleic Acids Res., 41(16):e153, Sep 2013.
[43] Ka-Chun Wong, Kwong-Sak Leung, and Man-Hon Wong. An evolutionary algorithm
with species-specific explosion for multimodal optimization. In GECCO ’09: Proceedings of the 11th Annual conference on Genetic and evolutionary computation, pages
923–930, New York, NY, USA, 2009. ACM.
[44] Ka-Chun Wong, Kwong-Sak Leung, and Man-Hon Wong. Effect of spatial locality
on an evolutionary algorithm for multimodal optimization. In EvoApplications 2010,
Part I, LNCS 6024. Springer-Verlag, 2010.
[45] Ka-Chun Wong, Kwong-Sak Leung, and Man Hon Wong. Protein structure prediction
on a lattice model via multimodal optimization techniques. In GECCO ’10: Proceedings of the 12th annual conference on Genetic and evolutionary computation, pages
155–162, New York, NY, USA, 2010. ACM.
[46] Ka-Chun Wong, Yue Li, Chengbin. Peng, and Zhaolei Zhang. SignalSpider: probabilistic pattern discovery on multiple normalized ChIP-Seq signal profiles. Bioinformatics,
Sep 2014.
[47] Ka-Chun Wong, Chengbin Peng, Yue Li, and Tak-Ming Chan. Herd clustering: A synergistic data clustering approach using collective intelligence. Applied Soft Computing,
23:61–75, 2014.
[48] Ka-Chun Wong, Chengbin Peng, Man-Hon Wong, and Kwong-Sak Leung. Generalizing and learning protein-dna binding sequence representations by an evolutionary
algorithm. Soft Computing - A Fusion of Foundations, Methodologies and Applications, 15:1631–1642, 2011. 10.1007/s00500-011-0692-5.
[49] Ka-Chun Wong, Chun-Ho Wu, Ricky KP Mok, Chengbin Peng, and Zhaolei Zhang.
Evolutionary multimodal optimization using the principle of locality. Information
Sciences, 194:138–170, 2012.
[50] Ka-Chun Wong and Zhaolei Zhang.
SNPdryad: predicting deleterious nonsynonymous human SNPs using only orthologous protein sequences. Bioinformatics,
Jan 2014.
[51] Chun-Ho Wu, Na Dong, Waihung Ip, Ching-Yuen Chan, Kei-Leung Yung, and
Zengqiang Chen. Chaotic hybrid algorithm and its application in circle detection.
In EvoWorkshops, pages 302–311, 2010.
[52] E. L. Yu and P. N. Suganthan. Ensemble of niching algorithms. Inf. Sci., 180:2815–
2833, August 2010.
14
[53] E. L. Yu and P.N. Suganthan. Empirical comparison of niching methods on hybrid
composition functions. In CEC 2008. (IEEE World Congress on Computational Intelligence). IEEE Congress on Evolutionary Computation, 2008, pages 2194 –2201, june
2008.
15
(a) Best Fitness
(b) Best Fitness
(c) Peaks Found
(d) Peaks Found
Figure 1: For Table 1, we depict the statistical significance test results for the pairwise
performance differences between all algorithms tested on the VLS holgraphic grating design
problem by Mann-Whitney U-test (MWU) and Two-sample Kolmogorov-Smirnov test (KS)
with p=0.05. Each sub-figure correspond to the performance comparison using a metric
by a statistical test. The vertical axis is the same as the horizontal axis. Each algorithm is
represented by a number on each axis. The numbering of the algorithms follows the order
in Table 1. For instance, 1 refers to CrowdingDE-STL, 2 refers to CrowdingDE-TL......10
refers to UN. The color of each block represents whether the algorithm indicated by the
horizontal axis shows a performance different from the algorithm indicated by the vertical
axis in a statistically significant way. The black color denotes the p-values higher than
0.05 whereas the white color denotes the p-values lower than 0.05. The even numbered
sub-figures are the results obtained by MWU, whereas the odd numbered sub-figures are
the results obtained by KS.
16
(a)
(b)
Figure 2: Configurations obtained by a single run of CrowdingDE-STL on the VLS holographic grating design problem. It can be seen that they are totally different and feasible
configurations with which optical engineers can feel free to perform multiple trials after
the single run.
17
| 9cs.NE
|
The Non-Uniform k-Center Problem
Deeparnab Chakrabarty
Prachi Goyal
Ravishankar Krishnaswamy
arXiv:1605.03692v2 [cs.DS] 13 May 2016
Microsoft Research, India
dechakr,t-prgoya,[email protected]
Abstract
In this paper, we introduce and study the Non-Uniform k-Center (NUkC) problem. Given a finite
metric space (X, d) and a collection of balls of radii {r1 ≥ · · · ≥ rk }, the NUkC problem is to find a
placement of their centers on the metric space and find the minimum dilation α, such that the union of
balls of radius α · ri around the ith center covers all the points in X. This problem naturally arises as a
min-max vehicle routing problem with fleets of different speeds.
The NUkC problem generalizes the classic k-center problem when all the k radii are the same (which
can be assumed to be 1 after scaling). It also generalizes the k-center with outliers (kCwO for short)
problem when there are k balls of radius 1 and ` balls of radius 0. There are 2-approximation and
3-approximation algorithms known for these problems respectively; the former is best possible unless
P=NP and the latter remains unimproved for 15 years.
We first observe that no O(1)-approximation is to the optimal dilation is possible unless P=NP,
implying that the NUkC problem is more non-trivial than the above two problems. Our main algorithmic
result is an (O(1), O(1))-bi-criteria approximation result: we give an O(1)-approximation to the optimal
dilation, however, we may open Θ(1) centers of each radii. Our techniques also allow us to prove a simple
(uni-criteria), optimal 2-approximation to the kCwO problem improving upon the long-standing 3-factor.
Our main technical contribution is a connection between the NUkC problem and the so-called
firefighter problems on trees which have been studied recently in the TCS community. We show NUkC is
as hard as the firefighter problem. While we don’t know if the converse is true, we are able to adapt ideas
from recent works [3, 1] in non-trivial ways to obtain our constant factor bi-criteria approximation.
1
Introduction
Source location and vehicle routing problems are extremely well studied [19, 23, 9] in operations research. Consider the following location+routing problem: we are given a set of k ambulances with speeds
s1 , s2 , . . . , sk respectively, and we have to find the depot locations for these vehicles in a metric space (X, d)
such that any point in the space can be served by some ambulance as fast as possible. If all speeds were the
same, then we would place the ambulances in locations S such that maxv∈X d(v, S) is minimized – this is
the famous k-center problem. Differing speeds, however, leads to non-uniformity, thus motivating the titular
problem we consider.
Definition 1.1 (The Non-Uniform k-Center Problem (NUkC)). The input to the problem is a metric space
(X, d) and a collection of k balls of radii {r1 ≥ r2 ≥ · · · ≥ rk }. The objective is to find a placement C ⊆ X
of the centers of these balls, so as to minimize the dilation parameter α such that the union of balls of radius
α · ri around the ith center covers all of X. Equivalently, we need to find centers {c1 , . . . , ck } to minimize
i)
maxv∈X minki=1 d(v,c
ri .
1
As mentioned above, when all ri ’s are the same (and equal to 1 by scaling), we get the k-center problem.
The k-center problem was originally studied by Gonzalez [10] and Hochbaum and Shmoys [13] as a clustering
problem of partitioning a metric space into different clusters to minimize maximum intra-cluster distances.
One issue (see Figure 1 for an illustration and refer to [11] for a more detailed explanation) with k-center
(and also k-median/means) as an objective function for clustering is that it favors clusters of similar sizes
with respect to cluster radii. However, in presence of qualitative information on the differing cluster sizes,
the non-uniform versions of the problem can arguably provide more nuanced solutions. One such extreme
special case was considered as the “clustering with outliers” problem [7] where some fixed number/fractions
of points in the metric space need not be covered by the clusters. In particular, Charikar et al [7] consider
(among many problems) the k-center with outlier problem (kCwO, for short) and show a 3-approximation
for this problem. It is easy to see that kCwO is a special case of the NUkC problem when there are k balls of
radius 1 and ` (the number of outliers) balls of radius 0.
Motivated by the aforementioned reasons (both from facility location as well as from clustering settings), we investigate the worst-case complexity of the NUkC problem. Gonzalez [10] and Hochbaum and
Shmoys [13] give 2-approximations for the k-center problem, and also show that no better factor is possible
unless P = NP. Charikar et al [7] give a 3-approximation for the kCwO problem, and this has been the best
factor known for 15 years. Given these algorithms, it is natural to wonder if a simple O(1)-approximation
exists for the NUkC problem. In fact, our first result shows a qualitative distinction between NUkC and these
problems: constant-approximations are impossible for NUkC unless P=NP.
Theorem 1.2. For any constant c ≥ 1, the NUkC problem does not admit a c-factor approximation unless
P = N P , even when the underlying metric is a tree metric.
The hardness result is by a reduction from the so-called resource minimization for fire containment
problem on trees (RMFC-T, in short), a variant of the firefighter problem. To circumvent the above hardness,
we give the following bi-criteria approximation algorithm which is the main result of the paper, and which
further highlights the connections with RMFC-T since our algorithms heavily rely on the recent algorithms
for RMFC-T [3, 1]. An (a, b)-factor bi-criteria algorithm for NUkC returns a solution which places at most
a balls of each type (thus in total it may use as many as a · k balls), and the dilation is at most b times the
optimum dilation for the instance which places exactly one ball of each type.
Theorem 1.3. There is an (O(1), O(1))-factor bi-criteria algorithm for the NUkC problem.
Figure 1: The left figure shows the dataset, the middle figure shows a traditional k-center clustering, and the
right figure depicts a non-uniform clustering
2
Furthermore, as we elucidate below, our techniques also give uni-criteria results when
√ the number of
distinct radii is 2. In particular, we get a 2-approximation for the kCwO problem and a (1+ 5)-approximation
when there are only two distinct types of radii.
Theorem 1.4. There is a 2-approximation for the kCwO problem.
√
Theorem 1.5. There is a (1 + 5)-approximation for the NUkC problem when the number of distinct radii
is at most 2.
1.1
Discussion on Techniques
Our proofs of Theorems 1.2 and 1.3 shows a strong connection between NUkC and the so-called resource
minimization for fire containment problem on trees (RMFC-T, in short). This connection is one of the main
findings of the paper, so we first formally define this problem.
Definition 1.6 (Resource Minimization for Fire Containment on Trees (RMFC-T)). Given a rooted tree T as
input, the goal is to select a collection of non-root nodes N from T such that (a) every root-leaf path has
at least one vertex from N , and (b) maxt |N ∩ Lt | is minimized, where Lt is the tth-layer of T , that is, the
vertices of T at exactly distance t from the root.
To understand the reason behind the name, consider a fire starting at the root spreading to neighboring
vertices each day; the RMFC-T problem minimizes the number of firefighters needed per day so as to prevent
the fire spreading to the leaves of T .
It is NP-hard to decide if the optimum of RMFC-T is 1 or not [8, 17]. Given any RMFC-T instance
and any c > 1, we construct an NUkC instance on a tree metric such that in the “yes” case there is always
a placement with dilation = 1 which covers the metric, while in the “no” case even a dilation of c doesn’t
help. Upon understanding our hardness construction, the inquisitive reader may wonder if the reduction
also works in the other direction, i.e., whether we can solve NUkC using a reduction to RMFC-T problem.
Unfortunately, we do not know if this is true even for two types of radii. However, as we explain below we
still can use positive results for the RMFC-T problem to design good algorithms for the NUkC problem.
Indeed, we start off by considering the natural LP relaxation for the NUkC problem and describe an
LP-aware reduction of NUkC to RMFC-T. More precisely, given a feasible solution to the LP-relaxation for
the given NUkC instance, we describe a procedure to obtain an instance of RMFC-T defined by a tree T ,
with the following properties: (i) we can exhibit a feasible fractional solution for the LP relaxation of the
RMFC-T instance, and (ii) given any feasible integral solution to the RMFC-T instance, we can obtain a
feasible integral solution to the NUkC instance which dilates the radii by at most a constant factor. Therefore,
an LP-based ρ-approximation to RMFC-T would immediately imply (ρ, O(1))-bicriteria approximation
algorithms for NUkC. This already implies Theorem 1.4 and Theorem 1.5 since the corresponding RMFC-T
instances have no integrality gap. Also, using a result of Chalermsook and Chuzhoy [3], we directly obtain
an (O(log∗ n), O(1))-bicriteria approximation algorithm for NUkC.
Here we reach a technical bottleneck: Chalermsook and Chuzhoy [3] also show that the integrality
gap of the natural LP relaxation for RMFC-T is Ω(log∗ n). When combined with our hardness reduction
in Theorem 1.2 , this also implies a (Ω(log∗ n), c) integrality gap for any constant c > 1 for the natural LP
relaxation for NUkC. That is, even if we allow a violation of c in the radius dilation, there is a Ω(log∗ n)integrality gap in terms of the violation in number of balls opened of each type.
However, very recently, Adjiashvili, Baggio and Zenklusen [1] show an improved O(1)-approximation
for the RMFC-T problem. At a very high level, the main technique in [1] is the following. Given an
3
RMFC-T instance, they carefully and efficiently “guess” a subset of the optimum solution, such that the
natural LP-relaxation for covering the uncovered leaves has O(1)-integrality gap. However, this guessing
procedure crucially uses the tree structure of T in the RMFC-T problem. Unfortunately for us though, we get
the RMFC-T tree only after solving the LP for NUkC, which already has an Ω(log∗ n)-gap! Nevertheless,
inspired by the ideas in [1], we show that we can also efficiently preprocess an NUkC instance, “guessing”
the positions of a certain number of balls in an optimum solution, such that the standard LP-relaxation for
covering the uncovered points indeed has O(1)-gap. We can then invoke the LP-aware embedding reduction
to RMFC-T at this juncture to solve our problem. This is quite delicate, and is the most technically involved
part of the paper.
1.2
Related Work and Open Questions
The k-center problem [10, 13] and the k-center with outliers [7] probems are classic problems in approximation algorithms and clustering. These problems have also been investigated under various settings such as the
incremental model [5, 22], streaming model [4, 22], and more recently in the map-reduce model [14, 21].
Similarly, the k-median [6, 15, 20, 2] and k-means [15, 16, 12, 18] problems are also classic problems studied
extensively in approximation algorithms and clustering. The generalization of k-median to a routing+location
problem was also studied recently [9]. It would be interesting to explore the complexity of the non-uniform
versions of these problems. Another direction would be to explore if the new non-uniform model can be
useful in solving clustering problems arising in practice.
2
Hardness Reduction
In this section, we prove Theorem 1.2 based on the following NP-hardness [17] for RMFC-T.
Theorem 2.1. [17] Given a tree T whose leaves are at the same distance from the root, it is NP-hard to
distinguish between the following two cases. YES: There is a solution to the RMFC-T instance of value 1.
NO: All solutions to the RMFC-T instance has value 2.
Given an RMFC-T instance defined by tree T , we now describe the construction of our NUkC instance.
Let h be the height of the tree, and let Lt denote the vertices of the tree at distance exactly t from the root. So,
the leaves constitute Lh since all leaves are at the same distance from the root. The NUkC instance, I(T ), is
defined by the metric space (X, d), and a collection of balls. The points in our metric space will correspond to
the leaves of the tree, i.e., X = Lh . To define the metric, we assign a weight d(e) = (2c + 1)h−i+1 for each
edge whose one endpoint is in Li and the other in Li−1 ; we then define d be the shortest-path metric on X
induced by this weighted tree. Finally, we set k = h, and define the k radii r1 ≥ r2 ≥ . . . ≥ rk iteratively as
follows: define rk := 0, and for k ≥ i > 1, set ri−1 := (2c + 1) · ri + 2(2c + 1). This completes the NUkC
instance. Before proceeding we make the simple observation: for any two leaves u and u0 with lca v ∈ Lt , we
have d(u, u0 ) = 2(2c + 1 + (2c + 1)2 + · · · + (2c + 1)h−t ) = rt . The following lemma proves Theorem 1.2.
Lemma 2.2. If T is the YES case of Theorem 2.1, then I(T ) has optimum dilation = 2. If T is the NO case
of 2.1, then I(T ) has optimum dilation ≥ 2c.
Proof. Suppose T is in the YES case, and there is a solution to RMFC-T which selects at most 1 node from
each level Lt . If v ∈ Lt is selected, then select a center cv arbitrarily from any leaf in the sub-tree rooted at v
and open the ball of radius rt . We now need to show all points in X = Lh are covered by these balls. Let u
4
be any leaf; there must be a vertex v in some level Lt in u’s path to the root such that a ball of radius rt is
opened at cv . However, d(u, cv ) ≤ d(u, v) + d(v, cv ) = 2rt and so the ball of radius 2rt around cv covers u.
Now suppose T is in the NO case, and the NUkC instance has a solution with optimum dilation < 2c.
We build a good solution for the RMFC-T instance N as follows: suppose the NUkC solution opens the
radius < 2c · rt ball around center u. Let v be the vertex on the u-root path appearing in level Lt . We then
pick this node in N . Observe two things: first, this ball covers all the leaves in the sub-tree rooted at v since
rt ≥ d(u, u0 ) for any such u0 . Furthermore, since the NUkC solution has only one ball of each radius, we get
that |N ∩ Lt | ≤ 1. Finally, since d(u, w) ≥ 2c · rt for all leaves w not in the sub-tree rooted at v, the ball of
radius rt around u doesn’t contain any leaves other than those rooted at v. Contra-positively, since all leaves
w are covered in some ball, every leaf must lie in the sub-tree of some vertex picked in N . That is, N is a
solution to RMFC-T with value = 1 contradicting the NO case.
3
LP-aware reduction from NUkC to RMFC-T
For reasons which will be apparent soon, we consider instances I of NUkC counting multiplicites. That
is, we consider
instance to be a collection of tuples (k1 , r1 ), . . . , (kh , rh ) to indicate there are ki balls of
Pan
h
radius ri . So t=1 kt = k. Intuitively, the reason we do this is that if two radii rt and rt+1 are “close-by”
then it makes sense to round up rt+1 to rt and increase kt , losing only a constant-factor loss in the dilation.
LP-relaxation for NUkC. We now state the natural LP relaxation for a given NUkC instance I. For each
point p ∈ X and radius type ri , we have an indicator variable xp,i ≥ 0 for whether we place a ball of radius ri
centered at p. By doing a binary search on the optimal dilation and scaling, we may assume that the optimum
dilation is 1. Then, the following linear program must be feasible. Below, we use B(q, ri ) to denote the set
of points within distance ri from q.
∀p ∈ X,
∀t ∈ 1, · · · , h
h
X
X
t=1
q∈B(p,rt )
X
xq,t
≥1
(NUkC LP)
≤ kt
xp,t
p∈X
LP-relaxation for RMFC-T. Since we reduce fractional NUkC to fractional RMFC-T, we now state the
natural LP relaxation for RMFC-T on a tree T of depth h + 1. In fact, we will work with the following
budgeted-version of RMFC-T (which is equivalent to the original RMFC-T problem — for a proof, see [1]):
Instead of minimizing the maximum number of “firefighters” at any level t (that is |N ∩ Lt | where N is
the chosen solution), suppose we specify a budget limit of kt on |N ∩ Lt |. The goal is the minimize the
maximum dilation of these budgets. Then the following is a natural LP relaxation for the budgeted RMFC-T
problem on trees. Here L = Lh is the set of leaves, and Lt are the layer t-nodes. For a leaf node v, let Pv
denote the vertex set of the unique leaf-root path excluding the root.
min α
X
yu ≥ 1
∀v ∈ L,
u∈Pv
X
∀t ∈ 1, · · · , h
u∈Lt
5
yu
≤ α · kt
(RMFC-T LP)
The LP-aware Reduction to Tree metrics. We now describe our main reduction algorithm, which takes
as input an NUkC instance I = {(X, d); (k1 , r1 ), . . . , (kh , rh )} and a feasible solution x to NUkC LP, and
returns a budgeted RMFC-T instance IT defined by a tree T along with budgets for each level, and a feasible
solution y to RMFC-T LP with dilation 1. The tree we construct will have height h + 1 and the budgeted
RMFC-T instance will have budgets precisely kt at level 1 ≤ t ≤ h, and the budget for the leaf level is 0.
For clarity, throughout this section we use the word points to denote elements of the metric space in I, and
the word vertices/nodes to denote the tree nodes in the RMFC-T instance that we construct. We build the
tree T in a bottom-up manner, where in each round, we pick a set of far-away representative points (the
distance scale increases as we move up the tree) and cluster all points to their nearest representative. This
is similar to a so-called clustering step in many known algorithms for facility location (see, e.g. [6]), but
whereas an arbitrary set of far-away representatives would suffice in the facility location algorithms, we need
to be careful in how we choose this set to make the overall algorithm work.
Formally, each vertex of the tree T is mapped to some point in X, and we denote the mapping of the
vertices at level t by ψt : Lt → X. We will maintain that each ψt will be injective, so ψt (u) 6= ψt (v) for
u 6= v in Lt . So, ψt−1 is well defined for the range of ψt .
The complete algorithm runs in rounds h + 1 to 2 building the tree one level per round. To begin with,
the ψh+1 mapping is an arbitrary bijective mapping between L := Lh+1 , the set of leaves of the tree, and the
points of X (so, in particular, |L| = |X|). We may assume it to be the identity bijection. In each round t, the
range of the mappings become progressively smaller, that is1 , P
ψt (Lt ) ⊇ ψt−1 (Lt−1 ). We call ψt (Lt ) as the
winners at level t. We now describe round t. Let Covt (p) := q∈B(p,rt ) xq,t denote
Pthe fractional amount
the point p is covered by radius rt balls in the solution x. Also define Cov≥t (p) := s≥t Covs (p) denoting
the fractional amount p is covered by radius rt or smaller balls. Let Covh+1 (p) = 0 for all p.
Algorithm 1 Round t of the LP-aware Reduction.
Input: Level Lt , subtrees below Lt , the mappings ψs : Ls → X for all t ≤ s ≤ h.
Output: Level Lt−1 , the connections between Lt−1 and Lt , and the mapping ψt−1 .
Define A = ψt (Lt ) the set of points who are winners at level t.
while A 6= ∅ do
(a) Choose the point p ∈ A with minimum coverage Cov≥t (p).
(b) Let N (p) := {q ∈ A : d(p, q) ≤ 2rt−1 } be the set of all nearby points in A to p.
(c) Create a new tree vertex w ∈ Lt−1 corresponding to p and set ψt−1 (w) := p. Call p a winner at
level t − 1, and each q ∈ N (p) ⊆ A a loser to p at this level.
(d) Create edge (w, v) for tree vertices v ∈ ψt−1 (N (p)) associated with N (p) at level t.
(e) Set A ← A \ (N (p)).
(f) Set yw = Covt−1 (p).
end while
Finally, we add a root vertex and connect it to all vertices in L1 . This gives us the final tree T and a
solution y which assigns a value to all non-leaf, non-root vertices of the tree T . The following claim asserts
well-definedness of the algorithm.
Lemma 3.1. The solution y is a feasible solution to RMFC-T LP on IT with dilation 1.
Proof. The proof is via two claims for the two different set of inequalities.
1
We are using the notation ψ(X) :=
S
x∈X
ψ(x).
6
Claim 3.2. For all 1 ≤ t ≤ h, we have
P
w∈Lt
yw ≤ kt .
Proof.
Fix t. Let
P
P Wt ⊆ X denote the winners at level t, that is, Wt = ψt (Lt ). By definition of the algorithm,
y
=
w∈Lt w
p∈Wt Covt (p). Now note that for any two points p, q ∈ Wt , we have B(p, rt ) ∩ B(q, rt ) = ∅.
To see this, consider the first point which enters A in the (t + 1)th round when Lt was being formed. If
this is p, then all points in the radius
t ball are deleted from A. Since the balls are disjoint, the second
P 2rP
inequality of NUkC LP implies p∈Wt q∈B(p,rt ) xq,t ≤ kt . The second summand in the LHS is precisely
Covt (p).
P
Claim 3.3. For any leaf node w ∈ L, we have v∈Pw yv ≥ 1.
Proof. We start with an observation. Fix a level t and a winner point p ∈ Wt . Let u ∈ Lt such that ψt (u) = p.
Since Wt ⊆ Wt+1 ⊆ · · · ⊆ Wh , there is a leaf v in the subtree rooted at u corresponding to p. Moreover, by
0
0
the
P way we formed our tree edges in step (d), we have that ψs (w ) = p for all w in the (u, v)-path and hence
w0 ∈[u,v]-path yw0 is precisely Cov≥t (p).
Now, for contradiction, suppose there is some leaf corresponding to, say point p, such that the root-leaf
path has total y-assignment less than 1. Then, pick the point, among all such unsatisfied points p, who appears
in a winning set Wt with t as small as possible.
By the preceding observation, the total y-assignment p receives on its path from level h to level t is
−1
exactly Cov≥t (p). Moreover, suppose p loses to q at level t − 1, i.e., ψt−1 (p) is a child of ψt−1
(q). In
particular, it means that q has also been a winner up to level t and so the total y-assignment on q’s path
−1
upto level t is also precisely Cov≥t (q). Additionally, since ψt−1
(q) became the parent node for ψt−1 (p), we
know that Cov≥t (q) ≤ Cov≥t (p) due to the way we choose winners in step (a) of the while loop. Finally,
by our maximality assumption on p, we know that q is fractionally satisfied by the y-solution. Therefore,
there is fractional assignment of at least (1 − Cov≥t (q)) on q’s path from nodes in level t − 1 to level 1.
Putting these observations together, we get that the total fractional assignment on p’s root-leaf path is at least
Cov≥t (p) + (1 − Cov≥t (q)) ≥ 1, which results in the desired contradiction.
The following lemma shows that any good integral solution to the RMFC-T instance IT can be converted
to a good integral solution for the NUkC instance I.
Lemma 3.4. Suppose there exists a feasible solution N to IT such that for all 1 ≤ t ≤ h, |N ∩ Lt | ≤ αkt .
Then there is a solution to the NUkC instance I that opens, for each 1 ≤ t ≤ h, at most αkt balls of radius
≤ 2r≥t ,where r≥t := rt + rt+1 + · · · + rh .
Proof. Construct the NUkC solution as follows: for level 1 ≤ t ≤ h and every vertex w ∈ N ∩ Lt , place
the center at ψt (w) of radius 2r≥t . We claim that every point in X is covered by some ball. Indeed, for any
p ∈ X, look at the leaf v = ψh+1 (p), and let w ∈ N be a node in the root-leaf path. Let w ∈ Lt for some t.
Now observe that d(p, ψt (w)) ≤ 2r≥t ; this is because for any edge (u0 , v 0 ) in the tree where u0 is in Lt and
is the parent of v 0 , we have that d(ψt+1 (v 0 ), ψt+1 (u0 )) < 2rt .
This completes the reduction, and we now prove a few results which follow easily from known results about
the firefighter problem.
Theorem 3.5. There is a polynomial time (O(log∗ n), 8)-bi-criteria algorithm for NUkC.
7
Proof. Given any instance I of NUkC, we first club the radii to the nearest power of 2 to get an instance I 0
with radii (k1 , r1 ), · · · , (kh , rh ) such that an (a, b)-factor solution for I 0 is an (a, 2b)-solution for I. Now,
by scaling, we assume that the optimal dilation for I 0 is 1; we let x be the feasible solution to the NUkC LP.
Then, using Algorithm 1, we can construct the tree IT0 and a feasible solution y to the RMFC-T LP. We can
now use the following theorem of Chalermsook and Chuzhoy [3]: given any feasible solution to the RMFC-T
LP, we can obtain a feasible set N covering all the leaves such that for all t, |N ∩ Lt | ≤ O(log∗ n)kt . Finally,
we can apply Lemma 3.4 to obtain a (O(log∗ n), 4) solution to I 0 (since r≥t ≤ 2rt ).
Proof of Theorem 1.4 and Theorem 1.5. We use the following claim regarding the integrality gap of RMFC-T
LP for depth 2 trees.
Claim 3.6. When h = 2 and kt ’s are integers, given any fractional solution to RMFC-T LP, we can find a
feasible integral solution as well.
Proof. Given a feasible solution y to RMFC-T LP, we need to find a set N such that |N ∩ Lt | ≤ kt for
t = 1, 2. There must exist at least one vertex w ∈ L1 such that yw ∈ (0, 1) for otherwise the solution y
is trivially integral. If only one vertex w ∈ L1 is fractional, then since k1 is an integer, we can raise this
yw to be an integer as well. So at least two vertices w and w0 in L1 are fractional. Now, without loss of
generality, let us assume that |C(w)| ≥ |C(w0 )|, where C(w) is the set of children of w. Now for some small
0 := y + , y 0 := y 0 − , ∀c ∈ C(w), y 0 := y − , and
constant 0 < < 1, we do the following: yw
w
c
w
c
w0
∀c ∈ C(w0 ), yc0 := yc + . Note that y(L1 ) remains unchanged, y(L2 ) can only decrease, and root-leaf paths
still add to at least 1. We repeat this till we rule out all fractional values.
To see the proof of Theorem 1.4, note that an instance of the k-center with outliers problem is an NUkC
instance with (k, 1), (`, 0), that is, r1 = 1 and r2 = 0. We solve the LP relaxation and obtain the tree and an
RMFC-T solution. The above claim implies a feasible integral solution to RMFC-T since h = 2, and finally
note that r≥1 = r1 for kCwO, implying we get a 2-factor approximation.
√
The proof of Theorem 1.5 is similar. If√r1 < θr2 where θ = ( 5 + 1)/2, then we simply run k-center
with k = k1√+ k2 . This gives a 2θ = 5 + 1-approximation. Otherwise, we apply Lemma 3.4 to get a
2(1 + 1θ ) = 5 + 1-approximation.
We end this section with a general theorem, which is an improvement over Lemma 3.4 in the case when
many of the radius types are close to each other, in which case r≥t could be much larger than rt . Indeed, the
natural way to overcome this would be to group the radius types into geometrically increasing values as we
did in the proof of Theorem 3.5. However, for some technical reasons we will not be able to bucket the radius
types in the following section, since we would instead be bucketing the number of balls of each radius type in
a geometric manner. Instead, we can easily modify Algorithm 1 to build the tree by focusing only on radius
types where the radii grow geometrically.
Theorem 3.7. Given an NUkC instance I = {M = (X, d), (k1 , r1 ), (k2 , r2 ), . . . , (kh , rh )} and an LP
solution x for NUkC LP, there is an efficient reduction which generates an RMFC-T instance IT and an LP
solution y to RMFC-T LP, such that the following holds:
(i) For any two tree vertices w ∈ Lt and v ∈ Lt0 where w is an ancestor of v (which means t ≤ t0 ),
suppose p and q are the corresponding points in the metric space, i.e., p = ψt (w) and q = ψt0 (v), then
it holds that d(p, q) ≤ 8 · rt .
8
(ii) Suppose there exists a feasible solution N to IT such that for all 1 ≤ t ≤ h, |N ∩ Lt | ≤ αkt . Then
there is a solution to the NUkC instance I that opens, for each 1 ≤ t ≤ h, at most αkt balls of radius
at most 8 · rt .
3.1
Proof of Theorem 3.7
Both the algorithm as well as the proof are very similar, and we now provide them for completeness. At a
high level, the only difference occurs when we identify and propagate winners: instead of doing it for each
radius type, we identify barrier levels where the radius doubles, and perform the clustering step only at the
barrier levels. We now present the algorithm, which again proceeds in rounds h + 1, h, h − 1, . . . , 2, but
makes jumps whenever there are many clusters of similar radius type. To start with, define rh+1 = 0.
Algorithm 2 Round t of the Improved Reduction.
Input: Level Lt , subtrees below Lt , the mappings ψs : Ls → X for all t ≤ s ≤ h.
Output: Level Lt−1 , the connections between Lt−1 and Lt , and the mapping ψt−1 .
Let t0 = mins s.t.rs ≤ 2rt−1 be the type of the largest radius smaller than 2rt−1 .
Define A = ψt (Lt ) the set of points who are winners at level t.
while A 6= ∅ do
(a) Choose the point p ∈ A with minimum coverage Cov≥t (p).
(b) Let N (p) := {q ∈ A : d(p, q) ≤ 2rt0 } denote all points in A within 2rt0 from p.
(c) Create new vertices wt−1 , . . . , wt0 −1 ∈ Lt−1 , . . . , Lt0 −1 levels respectively, all corresponding to p,
i.e., set ψi (w) := p for all t0 − 1 ≤ i ≤ t − 1. Connect each pair of these vertices in successive levels
with edges. Call p a winner at levels t − 1, . . . , t0 − 1.
(d) Create edge (wt−1 , v) for vertices v ∈ ψt−1 (N (p)) associated with N (p) at level t.
(f) Set A ← A \ (N (p)).
(g) Set ywi = Covi (p) for all t − 1 ≤ i ≤ t0 − 1.
end while
Jump to round t0 − 1 of the algorithm. Add t0 − 1 to the set of barrier levels
Our proof proceeds almost in an identical manner to those of Lemmas 3.1 and 3.4, but now our tree has
an additional property that for any two nodes u ∈ Li and v ∈ Li0 where u is an ancestor of v, the distance
between the corresponding points in the metric space p = ψi (u) and q = ψi0 (v) is at most d(p, q) ≤ 8ri ,
which was the property not true in the earlier reduction. This is easy to see because as we traverse a tree
path from u to v, notice that each time we change winners, the distance between the corresponding points in
the metric space decreases geometrically. This proves property (i) of Theorem 3.7. With this in hand, the
remaining proofs to prove the second property are almost identical to the ones in Section 3 and we sketch
them below for completeness.
Lemma 3.8. The solution y is a feasible solution to RMFC-T LP on IT with dilation 1.
Proof. The proof is via two claims for the two different set of inequalities.
P
Claim 3.9. For all 1 ≤ t ≤ h, we have w∈Lt yw ≤ kt .
Proof. Fix a barrier
Plevel t. Let W
Pt ⊆ X denote the winners at level t, that is, Wt = ψt (Lt ). By definition
of the algorithm, w∈Lt yw = p∈Wt Covt (p). Now note that for any two points p, q ∈ Wt , we have
B(p, rt ) ∩ B(q, rt ) = ∅. To see this, consider the first point which enters A in the round (corresponding to the
9
previous barrier) when Lt was being formed. If this is p, then all points in the
P radiusP2rt ball is deleted from
A. Since the balls are disjoint, the second inequality of NUkC LP implies p∈Wt q∈Bt (p) xq,t ≤ kt . The
second summand in the LHS is the definition of Covt (p). The same argument holds for all levels t between
two consecutive barrier levels t1 and t2 s.t. t1 > t2 , as the winner set remains the same, and the radius rt is
only smaller than the radius rt2 at the barrier t2 .
P
Claim 3.10. For any leaf node w ∈ L, we have v∈Pw yv ≥ 1.
Proof. This proof is identical to that of Claim 3.3, and we repeat it for completeness. Fix a level t and a
winner point p ∈ Wt . Let u ∈ Lt such that ψt (u) = p. Since Wt ⊆ Wt+1 ⊆ · · · ⊆ Wh , there is a leaf v in
the subtree rooted at u corresponding to p. Moreover, by the
Pway we formed our tree edges in step (d), we
have that ψs (w) = p for all w0 in the (u, v)-path and hence w0 ∈[u,v]-path yw0 is precisely Cov≥t (p).
Now, for contradiction, suppose there is some leaf corresponding to, say point p, such that the root-leaf
path has total y-assignment less than 1. Then, pick the point, among all such unsatisfied points p, who appears
in a winning set Wt with t as small as possible.
By the preceding observation, the total y-assignment p receives on its path from level h to level t is
−1
exactly Cov≥t (p). Moreover, suppose p loses to q at level t − 1, i.e., ψt−1 (p) is a child of ψt−1
(q). In
particular, it means that q has also been a winner up to level t and so the total y-assignment on q’s path up
−1
to level t is also precisely Cov≥t (q). Additionally, since ψt−1
(q) became the parent node for ψt−1 (p), we
know that Cov≥t (q) ≤ Cov≥t (p) due to the way we choose winners in step (a) of the while loop. Finally,
by our maximality assumption on p, we know that q is fractionally satisfied by the y-solution. Therefore,
there is fractional assignment of at least (1 − Cov≥t (q)) on q’s path from nodes in level t − 1 to level 1.
Putting these observations together, we get that the total fractional assignment on p’s root-leaf path is at least
Cov≥t (p) + (1 − Cov≥t (q)) ≥ 1, which results in the desired contradiction.
Finally, the following lemma shows that any good integral solution to the RMFC-T instance IT can be
converted to a good integral solution for the NUkC instance I.
Lemma 3.11. Suppose there exists a feasible solution N to IT such that for all 1 ≤ t ≤ h, |N ∩ Lt | ≤ αkt .
Then there is a solution to the NUkC instance I that opens, for each 1 ≤ t ≤ h, at most αkt balls of radius
at most 8rt .
Proof. Construct the NUkC solution as follows: for level 1 ≤ t ≤ h and every vertex w ∈ N ∩ Lt , place
the center at ψt (w) of radius 8 · rt . We claim that every point in X is covered by some ball. Indeed, for any
p ∈ X, look at the leaf v = ψh+1 (p), and let w ∈ N be a node in the root-leaf path which covers it in the
instance IT . By property (i) of Theorem 3.7, we have that the distance between ψt (w) and p is at most 8 · rt ,
and hence the ball of radius 8 · rt around ψt (w) covers p. The number of balls of radius type t is trivially at
most αkt .
4
Getting an (O(1), O(1))-approximation algorithm
In this section, we improve our approximation factor on the number of clusters from O(log∗ n) to O(1),
while maintaining a constant-approximation in the radius dilation. As mentioned in the introduction, this
requires more ideas since using NUkC LP one cannot get any factor better than (O(log∗ n), O(1))-bicriteria
10
approximation since any integrality gap for RMFC-T LP translates to a (Ω(log∗ n), Ω(1)) integrality gap
for NUkC LP.
Our algorithm is heavily inspired by the recent paper of Adjiashvili et al [1] who give an O(1)approximation for the RMFC-T problem. In fact, the structure of our algorithms follows the same three
“steps” of their algorithm. Given an RMFC-T instance, [1] first “compress” the input tree to get a new tree
whose depth is bounded; secondly, [1] give a partial rounding result which saves “bottom heavy” leaves, that
is, leaves which in the LP solution are covered by nodes from low levels; and finally, Adjiashvili et al [1] give
a clever partial enumeration algorithm for guessing the nodes from the top levels chosen by the optimum
solution. We also proceed in these three steps with the first two being very similar to the first two steps in [1].
However, the enumeration step requires new ideas for our problem. In particular, the enumeration procedure
in [1] crucially uses the tree structure of the firefighter instance, and the way our reduction generates the
tree for the RMFC-T instance is by using the optimal LP solution for the NUkC instance, which in itself
suffers from the Ω(log∗ n) integrality gap. Therefore, we need to devise a more sophisticated enumeration
scheme although the basic ideas are guided by those in [1]. Throughout this section, we do not optimize for
the constants.
4.1
Part I: Radii Reduction
In this part, we describe a preprocessing step which decreases the number of types of radii. This is similar to
Theorem 5 in [1].
Theorem 4.1. Let I be an instance of NUkC with radii {r1 , r2 , · · · , rk }. Then we can efficiently construct a
new instance Ib with radii multiplicities (k0 , rb0 ), ..., (kL , rbL ) and L = Θ(log k) such that:
(i) ki := 2i for all 0 ≤ i < L and kL ≤ 2L .
b
(ii) If the NUkC instance I has a feasible solution, then there exists a feasible solution for I.
b we can efficiently obtain a (3α, β)-bicriteria solution to I.
(iii) Given an (α, β)-bicriteria solution to I,
Proof. For an instance I, we construct the compressed instance Ib as follows. Partition the radii into Θ(log k)
b we simply round up all
classes by defining barriers at rbi = r2i for 0 ≤ i ≤ blog kc. Now to create instance I,
i
i+1
the radii rj for 2 ≤ j < 2
to the value rbi = r2i . Notice that the multiplicity of rbi is precisely 2i (except
maybe for the last bucket, where there might be fewer radii rounded up than the budget allowed).
Property (i) is just by construction of instance. Property (ii) follows from the way we rounded up the
radii. Indeed, if the optimal solution for I opens a ball of radius rj around a point p, then we can open a
cluster of radius rbi around p, where i is such that 2i ≤ j < 2i+1 . Clearly the number of clusters of radius rbi
is at most 2i because OPT uses at most one cluster of each radius rj .
For property (iii), suppose we have a solution Sb for Ib which opens α2i clusters of radius βb
ri for all
0 ≤ i ≤ L. Construct a solution S for I as follows. For each 1 ≤ i ≤ L, let Ci denote the set of centers
where Sb opens balls of radius βb
ri . In the solution S, we also open balls at precisely these centers with 2α
i−1
balls of radius rj for every 2
≤ j < 2i . Since |Ci | ≤ α · 2i , we can open a ball at every point in Ci ;
furthermore, since j < 2i , we have rj ≥ rbi and so we cover whatever the balls from Sb covered.
Finally, we also open the α clusters (corresponding to i = 0) of radius βr1 = βb
r0 at the respective
centers C0 where Sb opens centers of radius rb0 . Therefore, the total number of clusters of radius type is at
most 2α with the exception of r1 , which may have 3α clusters.
11
4.2
Part II: Satisfying Bottom Heavy Points
One main reason why the above height reduction step is useful, is the following theorem from [1] for RMFC-T
instances on trees; we provide a proof sketch for completeness.
Theorem 4.2 ([1]). Given a tree T of height h and a feasible solution y to (RMFC-T LP), we can find a
feasible integral solution N to RMFC-T such that for all 1 ≤ t ≤ h, |N ∩ Lt | ≤ kt + h.
Proof. Let y be a basic feasible solution of (RMFC-T LP). Call a vertex v of the tree loose if yv > 0 and the
sum of y-mass on the vertices from v to the root (inclusive of v) is < 1. Let VL be the set of loose vertices of
the tree, and let VI be the set of vertices with yv = 1. Clearly N = VL ∪VI is a feasible solution: every leaf-toroot path either contains an integral vertex or at least two fractional vertices with the vertex closer to root being
loose. Next we claim that |VL | ≤ h; this proves the theorem since |N ∩ Lt | ≤ |VI ∩ Lt | + |VL | ≤ kt + |VL |.
The full proof can be found in Lemma 6, [1] – here is a high level sketch. There are |L| + h inequalities
in (RMFC-T LP), and so the number of fractional variables is at most |L| + h. We may assume there are no
yv = 1 vertices. Now, in every leaf-to-root path there must be at least 2 fractional vertices, and the one closest
to the leaf must be non-loose. If the closest fractional vertex to each leaf was unique, then that would account
for |L| fractional non-loose vertices implying the number of loose vertices must be ≤ h. This may not be
true; however, if we look at linearly independent set of inequalities that are tight, we can argue uniqueness as
a clash can be used to exhibit linear dependence between the tight constraints.
Theorem 4.3. Suppose we are given an NUkC instance Ib with radii multiplicities
(k0 , rb0 ), (k1 , rb1 ), . . . , (kL , rbL ) with budgets ki = 2i for radius type rbi , and an LP solution x to (NUkC LP)
b Let τ = log log L, and suppose X 0 ⊆ X be the points covered mostly by small radii, that is, let
for I.
Cov≥τ (p) ≥ 12 for every p ∈ X 0 . Then, there is an efficient procedure round which opens at most O(kt ) balls
of radius O(b
rt ) for τ ≤ t ≤ L, and covers all of X 0 .
Proof. The procedure round works as follows: we partition the points of X 0 into two sets, one set XU in
which the points receive at least 14 of the coverage by clusters of radius rbi where i ∈ {log log L, log log L +
1, . . . , log L}, and another set XB in which the points receive 14 coverage from clusters of levels t ∈ {log L +
P L
0
1, log L + 2, . . . , L}. More precisely, XU := {p ∈ X 0 : log
t=τ Covt (p) ≥ 1/4}, and XB = X \ XB .
Now consider the following LP-solution to (NUkC LP) for Ib restricted to XU : we scale x by a factor 4
and zero-out x on radii type rbi for i ∈
/ {log log L, . . . , log L}. By definition of XU this is a feasible fractional
solution; furthermore, the LP-reduction algorithm described in Section 3 will lead to a tree T of height ≤ log L
and fractional solution y for (RMFC-T LP) on T were each ki ≥ 2log log L = log L. Applying Theorem 4.2,
we can find an integral solution N with at most O(ki ) vertices at levels i ∈ {log log L, . . . , log L}. We can
then translate this solution back using Theorem 3.7 to NUkC and find O(kt ) clusters of radius O(b
rt ) to cover
all the points XU . A similar argument, when applied to the smaller radius types rbt for t ∈ {log L, . . . , L}
can cover the points in XB .
We now show how we can immediately also get a (very weakly) quasi-polynomial time O(1)-approximation
for NUkC. Indeed, if we could enumerate the set of clusters of radii rbt for 0 ≤ t < log log L, we can then
explicitly solve an LP where all the uncovered points need to be fractionally covered by only clusters of
radius type rbt for t ≥ log log L. We can then round this solution using Corollary 4.3 to obtain the desired
O(1)-approximation for the NUkC instance. Moreover, the time complexity of enumerating the optimal
clusters of radii rbt for 0 ≤ t < log log L is nO(log L) = nO(log log k) , since the number of clusters of radius at
least rblog log L is at most O(2log log L ) = O(log L). Finally, there was nothing special in the proof of Theorem 4.3 about the choice of τ = log log L — we could set t = log(q) L to be the q th iterated logarithm of
12
L, and obtain an O(q)-approximation. As a result, we get the following corollary. Note that this gives an
alternate way to prove Theorem 3.5.
Corollary 4.4. For any q ≥ 1, there exists an (O(q), O(1))-factor bicriteria algorithm for NUkC which
(q)
runs in nO(log k) time.
4.3
Part III: Clever Enumeration of Large Radii Clusters
In this section, we show how to obtain the (O(1), O(1))-factor bi-criteria algorithm. At a high level, our
algorithm tries to “guess” the centers2 A of large radius, that is rbi for i ≤ τ := log log L = log log log k,
which the optimum solution uses. However, this guessing is done in a cleverer way than in Corollary 4.4.
In particular, given a guess which is consistent with the optimum solution (the initial “null set” guess is
trivially consistent), our enumeration procedure generates a list of candidate additions to A of size at most
2τ ≈ poly log logk (instead of n), one of which is a consistent enhancement of the guessed set A. This
reduction in number of candidates also requires us to maintain a guess D of points where the optimum
solution doesn’t open centers. Furthermore, we need to argue that the “depth of recursion” is also bounded by
poly log logk; this crucially uses the technology developed in Section 3. Altogether, we get the total time is
at most (poly log logk)poly log logk = o(k) for large k.
We start with some definitions. Throughout, A and D represent sets of tuples of the form (p, t) where
p ∈ X and t ∈ {0, 1, . . . , τ }. Given such a set A, we associate a partial solution SA which opens a ball
of radius 22b
rt at the point p. For the sake of analysis, fix an optimum solution OPT. We say the set A is
consistent with OPT if for all (p, t) ∈ A, there exists a unique q ∈ X such that OPT opens a ball of radius
rbt at q and d(p, q) ≤ 11b
rt . In particular, this implies that SA covers all points which this OPT-ball covers.
We say the set D is consistent with OPT if for all (q, t) ∈ D, OPT doesn’t open a radius rbt ball at q (it may
open a different radius ball at q though). Given a pair of sets (A,D), we define the minLevel of each point p
as follows
minLevelA,D (p) := 1 + arg max{(q, t) ∈ D for all q ∈ B(p, rbt )}
t
If (A, D) is a consistent pair and minLevelA,D (p) = t, then this implies in the OPT solution, p is covered by
a ball of radius rbt or smaller.
Next, we describe a nuanced LP-relaxation for NUkC. Fix a pair of sets (A, D) as described above. Let
XG be the subset of points in X covered by the partial solution SA . Fix a subset Y ⊆ X \ XG of points.
Define the following LP.
∀p ∈ Y,
∀t ∈ 1, · · · , h
L
X
X
t=minLevel(p)
q∈B(p,b
rt )
X
xq,t
≤
xq,t
≥1
(LPNUkC (Y, A, D))
kt
q∈Y
∀(p, t) ∈ A,
xp,t = 1
The following claim encapsulates the utility of the above relaxation.
Claim 4.5. If (A, D) is consistent with OPT, then (LPNUkC (Y, A, D)) is feasible.
2
Actually, we end up guessing centers “close” to the optimum centers, but for this introductory paragraph this intuition is
adequate.
13
Proof. We describe a feasible solution to the above LP using OPT. Given OPT, define O to be the collection
of pairs (q, t) where OPT opens a radius rbt ball at point q. Note that the number of tuples in O with second
term t is ≤ kt .
Since A is consistent with OPT, for every (p, t) ∈ A, there exists a unique (q, t) ∈ O; remove all such
tuples from O. Define xq,t = 1 for all other remaining tuples. By the uniqueness property, we see the second
inequality of the LP is satisfied. We say a point p ∈ X is covered by (q, t) if p lies in the rbt -radius ball
around q. Since Y ⊆ X \ XG , and since the partial solution SA contains all points p which is covered by all
the removed (q, t) tuples, we see that every point p ∈ Y is covered by some remaining (q, t) ∈ O. Since
D is consistent with OPT, for every point p ∈ Y and t < minLevelA,D (p), if q ∈ B(p, rbt ) then (q, t) ∈
/ O.
Therefore, the first inequality is also satisfied.
Finally, for convenience, we define a forbidden set F := {(p, i) : p ∈ X, 1 ≤ i ≤ τ } which if added to
D disallows any large radii balls to be placed anywhere.
Now we are ready to describe the enumeration Algorithm 3. We start with A and D being null, and thus
vacuously consistent with OPT. The enumeration procedure ensures that: given a consistent (A, D) tuple,
either it finds a good solution using LP rounding (Step 10), or generates candidate additions (Steps 18–20) to
A or D ensuring that one of them leads to a larger consistent tuple.
Algorithm 3 Enum(A, D, γ)
1: Let XG = {p : ∃ (q, i) ∈ A s.t d(p, q) ≤ 22b
ri } denote points covered by SA .
2: if there is no feasible solution to LPNUkC (X \ XG , A, D) then
3:
Abort.
//Claim 4.5 implies (A, D) is not consistent.
4: else
5:
x∗ be a feasible solution to LPNUkC (X \ XG , A, D).
6: end if
7: Let XB = {u ∈ X \ XG : Cov≥τ (u) ≥ 21 } denote bottom-heavy points in x∗
8: Let SB be the solution implied by Theorem 4.3.
// This solution opens O(kt ) balls of radius O(b
rt ) for τ ≤ t ≤ L and covers all of XB .
9: Let XT = X \ (XG ∪ XB ) denote the top heavy points in x∗
10: if LPNUkC (XT , A, F ∪ D) has a feasible solution xT then
11:
By definition of F , in xT we have Cov≥τ (u) = 1 for all u ∈ XT .
12:
ST be the solution implied by Theorem 4.3.
// This solution opens O(kt ) balls of radius O(b
rt ) for τ ≤ t ≤ L and covers all of XT .
13:
Output (SA ∪ SB ∪ ST ). //This is a (O(1), O(1))-approximation for the NUkC instance.
14: else
15:
for every level 0 ≤ t ≤ τ do
16:
Let Ct = {p ∈ XT s.t minLevelA,D (p) = t}, the set of points in XT with minLevel t.
17:
Use the LP-aware reduction from Section 3 using x∗ and the set of points Ct to create tree Tt .
18:
for every winner p at level t in Tt do
19:
Enum(A ∪ {(p,St)}, D, γ − 1)
20:
Enum(A, D ∪ p0 ∈B(p,11brt ) {(p0 , t)}), γ − 1)
21:
end for
22:
end for
23: end if
14
Define γ0 := 4 log log k · log log log k. The algorithm is run with Enum(∅, ∅, γ0 ). The proof that we get
a polynomial time (O(1), O(1))-bicriteria approximation algorithm follows from three lemmas. Lemma 4.6
shows that if Step 10 is true with a consistent pair (A, D), then the output in Step 13 is a (O(1), O(1))approximation. Lemma 4.7 shows that indeed Step 10 is true for γ0 as set. Finally, Lemma 4.8 shows with
such a γ0 , the algorithm runs in polynomial time.
Lemma 4.6. If (A, D) is a consistent pair such that Step 10 is true, then the solution returned is an
(O(1), O(1))-approximation algorithm.
Proof. Since A is consistent with OPT, SA opens at most kt centers with radius ≤ 22b
rt for all 0 ≤ t ≤ τ .
By design, SB and ST open at most O(kt ) centers with radius ≤ O(rt ) for τ ≤ t ≤ L.
Lemma 4.7. Enum(∅, ∅, γ0 ) finds consistent (A, D) such that Step 10 is true.
Proof. For this we identify a particular execution path of the procedure Enum(A, D, γ), that at every point
maintains a tuple (A, D) that is consistent with OPT. At the beginning of the algorithm, A = ∅ and D = ∅,
which is consistent with OPT.
Now consider a tuple (A, D) that is consistent with OPT and let us assume that we are within the
execution path Enum(A, D, γ). Let X \ XG be the points not covered by A and let x∗ be a solution to
LPNUkC (X \ XG , A, D). If OPT covers all top-heavy points XT using only smaller radii, then this implies
LPNUkC (XT , A, F ∪ D) has a feasible solution implyin Step 10 is true. So, we may assume, there exists at
least one top-heavy point q ∈ XT that OPT covers using a ball radii ≥ rbτ around a center oq . In particular,
minLevelA,D (q) ≤ τ . Let q ∈ Ct and hence q belongs to Tt for some 0 ≤ t ≤ τ . Let p ∈ Pt be the level t
winner in Tt s.t q belongs to the sub-tree rooted at p in Ti ; p may or may not be q. We now show that there is
at least one recursive call where we make non-trivial progress in (A, D). Indeed, we do this in two cases:
case (A) If OPT opens a ball of radius rbt at a point o such that d(o, p) ≤ 11b
rt . In this case, Step 19 maintains
consistency. Furthermore, we can “charge” (p, t) uniquely to the point o with radius rbt . To see this, for
contradiction, let us assume that before arriving to the recursive call where (p, t) is added to A, some other
0
0
0
tuple (u, t) ∈ A , in an earlier recursive call with (A , D ) as parameters charged to (o, t). Then by definition
we know that d(u, o) ≤ 11b
rt implying d(u, p) ≤ 22b
rt . Then p would be in XG in all subsequent iterations,
contradicting that p ∈ XT currently.
case (B): Tf there is no (o, t) ∈ OPT with d(o, p) ≤ 11b
rt , then for all points p0 ∈ B(p, 11b
rt ) we can add
0
(p , t) to D. In this case, we follow the recursive call in Step 20.
To sum, we can definitely follow the recursive calls in the consistent direction; but how do we bound
the depth of recursion. In case (A), the measure of progress is clear – we increase the size of |A|, and it can
be argued (we do so below) the maximum size of A is at most poly log logk. Case (B) is subtler. We do
increase size of D, but D could grow as large as Θ(n). Before going to the formal proof, let us intuitively
argue what “we learn” in Case (B). Recall q is covered in OPT by a ball around the center oq . Since
minLevel(q) = t ≤ τ , by definition there is a point v ∈ B(q, rbt ) such that (v, t) ∈
/ D, and d(q, oq ) ≤ rbt .
Together, we get d(v, oq ) ≤ 2b
rt , that is, v ∈ B(oq , 2b
rt ). Now also note, since q lies in p’s subtree in Tt , by
construction of the trees, d(p, q) ≤ 8b
rt by property (i) of Theorem 3.7. Therefore, d(p, oq ) ≤ 9b
rt and in case
(B), for all points u ∈ B(oq , 2b
rt ) we put (u, t) in the set D in the next recursive call. This is “new info” since
for the current D we know that at least one point v ∈ B(oq , 2b
rt ), we had (v, t) ∈
/ D.
Formally, we define the following potential function. Let Oτ denote the centers in OPT around which
balls of radius rbj , j ≤ τ have been opened. Given the set D, for 0 ≤ t ≤ τ and for all o ∈ Oτ , define the
15
(D)
indicator variable Zo,t which is 1 if for all points u ∈ B(o, 2b
rt ), we have (u, t) ∈ D and 0 otherwise.
Φ(A, D) := |A| +
τ
XX
(D)
Zo,t
o∈Oτ t=0
Note that Φ(∅, ∅) = 0. From the previous paragraph, we conclude that in both case (A) or (B), the
potential increases by at least 1.
for any consistent A, D we can upper bound Φ(A, D) as follows.
PFinally,
τ
t
τ +1 = log L = log log k. The second term in Φ is at most
Since A is consistent, |A| ≤
t=0 2 ≤ 2
τ
+1
2
· τ = log L log log L. Thus, in at most 2 log log k · log log log k < γ0 steps we reach a consistent pair
(A, D) with Step 10 true.
Lemma 4.8. Enum(∅, ∅, γ0 ) runs in polynomial time for large enough k.
Proof. Each single call of Enum is clearly polynomial time, and so we bound the number of recursive
calls. Indeed, this number will be o(k). We first bound the number of recursive calls in a single execution
of Enum(A, D, γ). For a fixed tuple (A, D), Algorithm 3, constructs trees T0 , . . . , Tτ in Step 17, using
the reduction algorithm from Section 3. Let Ltj represent the set of nodes at level j in the tree Tt . Then
Pt = ψt (Ltt ) represents the set of points that are winners at level t in Tt . NowSfor any tree Tt , Algorithm 3
makes two recursive calls for each winner in Pt (Step 19-20). Let PAD = τt=0 Pt be the set of all the
winners that the algorithm considers in a single call to Enum(A, D, γ). The total number of recursive calls
in a single execution is therefore 2|PAD |. Now we claim that for a fixed tuple (A, D), the total number of
winners is bounded.
Claim 4.9. |PAD | ≤ 4 log log k · log log log k
Proof. Consider a tree Tt and the corresponding set Pt as defined above. We use y (t) to denote the RMFC-T
LP solution given along with the tree Tt given by the reduction algorithm in Section 3. Let p ∈ Pt be a
winner at level t (and consequently, at level τ also), and suppose it is mapped to tree vertices w at level t
and w0 at level τ . Then, by the way the tree was constructed and because p ∈ XT is top-heavy, we have
P
(t)
1
u∈[w,w0 ]−path yu ≥ 2 (Refer to the proof of Claim 3.3 for more clarity). So each winner at level t has a
1
path down to level τ with fractional coverage
2 . But the total fractional coverage in the top part of
Patτ least
t
the tree is at most the total budget, which is t=1 2 ≤ 2 log L ≤ 2 log log k. Therefore, |Pt | ≤ 4 log log k.
Adding for all 1 ≤ t ≤ τ , gives |PAD | ≤ 4 log log k · t ≤ 4 log log k · log log log k.
Since the recursion depth γ0 , the total number of recursive calls made to the Enum is loosely upper
bounded by γ0poly log logk = o(k), thus completing the proof.
References
[1] D. Adjiashvili, A. Baggio, and R. Zenklusen. Firefighting on trees beyond integrality gaps. CoRR,
abs/1601.00271, 2016.
[2] J. Byrka, T. Pensyl, B. Rybicki, A. Srinivasan, and K. Trinh. An improved approximation for k-median,
and positive correlation in budgeted optimization. Proceedings, ACM-SIAM Symposium on Discrete
Algorithms (SODA), 2015.
16
[3] P. Chalermsook and J. Chuzhoy. Resource minimization for fire containment. Proceedings, ACM-SIAM
Symposium on Discrete Algorithms (SODA), 2010.
[4] M. Charikar, L. O’ Callaghan, and R. Panigrahy. Better streaming algorithms for clustering problems.
ACM Symp. on Theory of Computing (STOC), 2003.
[5] M. Charikar, C. Chekuri, T. Feder, and R. Motwani. Incremental clustering and dynamic infomation
retrieval. ACM Symp. on Theory of Computing (STOC), 1997.
[6] M. Charikar, S. Guha, D. Shmoys, and E. Tardos. A constant-factor approximation algorithm for the
k-median problem. ACM Symp. on Theory of Computing (STOC), 1999.
[7] M. Charikar, S. Khuller, D. M. Mount, and G. Narasimhan. Algorithms for facility location problems
with outliers. Proceedings, ACM-SIAM Symposium on Discrete Algorithms (SODA), 2001.
[8] S. Finbow, A. King, G. MacGillivray, and R. Rizzi. The firefighter problem for graphs of maximum
degree three. Discrete Mathematics, 307(16):2094–2105, 2007.
[9] I. L. Goertz and V. Nagarajan. Locating depots for capacitated vehicle routing. Proceedings, International Workshop on Approximation Algorithms for Combinatorial Optimization Problems, 2011.
[10] T. F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer
Science, 38:293–306, 1985.
[11] S. Guha, R. Rastogi, and K. Shim. CURE: An efficient clustering algorithm for large databases.
Proceedings of SIGMOD, 1998.
[12] S. Har-Peled and S. Mazumdar. Coresets for k-means and k-median clustering and their applications.
ACM Symp. on Theory of Computing (STOC), 2004.
[13] D. S. Hochbaum and D. B. Shmoys. A best possible heuristic for the k-center problem. Mathematics of
operations research, 10(2):180–184, 1985.
[14] S. Im and B. Moseley. Fast and better distributed mapreduce algorithms for k-center clustering.
Proceedings, ACM Symposium on Parallelism in Algorithms and Architectures, 2015.
[15] K. Jain and V. V. Vazirani. Approximation algorithms for metric facility location and k-median problems
using the primal-dual schema and lagrangian relaxation. J. ACM, 48(2):274 – 296, 2001.
[16] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman, and A. Y. Wu. : A local search
approximation algorithm for k-means clustering. 2002.
[17] A. King and G. MacGillivray. The firefighter problem for cubic graphs. Discrete Mathematics,
310(3):614–621, 2010.
[18] A. Kumar, Y. Sabharwal, and S. Sen. A simple linear time (1 + )-approximation algorithm for k-means
clustering in any dimensions. Proceedings, IEEE Symposium on Foundations of Computer Science
(FOCS), 2004.
[19] G. Laporte. Location routing problems. In B. L. Golden and A. A. Assad, editors, Vehicle Routing:
Methods and Studies, pages 163–198. 1998.
17
[20] S. Li and O. Svensson. Approximating k-median via pseudo-approximation. ACM Symp. on Theory of
Computing (STOC), 2013.
[21] G. Malkomes, M. J. Kusner, W. Chen, K. Q. Weinberger, and B. Moseley. Fast distributed k-center
clustering with outliers on massive data. Advances in Neur. Inf. Proc. Sys. (NIPS), 2015.
[22] R. McCutchen and S. Khuller. Streaming algorithms for k-center clustering with outliers and with
anonymity. Proceedings, International Workshop on Approximation Algorithms for Combinatorial
Optimization Problems, 2008.
[23] H. Min, V. Jayaraman, and R. Srivastava. Combined location-routing problems: A synthesis and future
research directions. European Journal of Operational Research, 108:1–15, 1998.
18
| 8cs.DS
|
Entropy production rate as a criterion for inconsistency in decision theory
Purushottam D. Dixit1
Department of Systems Biology, Columbia University
New York, NYa)
arXiv:1801.01733v1 [cs.AI] 5 Jan 2018
Evaluating pairwise comparisons breaks down complex decision problems into tractable ones. Pairwise comparison matrices (PCMs) are regularly used to solve multiple-criteria decision-making (MCDM) problems
using Saaty’s analytic hierarchy process (AHP) framework. There are two significant drawbacks of using
PCMs. First, humans evaluate PCM in an inconsistent manner. Second, PCMs of large problems often
have missing entries. We address these two issues by first establishing a novel connection between PCMs
and time-irreversible Markov processes. Specifically, we show that every PCM induces a family of dissipative
maximum path entropy random walks (MERW) over the set of alternatives. We show that only ‘consistent’
PCMs correspond to detailed balanced MERWs. We identify the non-equilibrium entropy production in the
induced MERWs as a metric of inconsistency of the underlying PCMs. Notably, the entropy production
satisfies all of the recently laid out criteria for reasonable consistency indices. We also propose an approach
to use incompletely filled PCMs in AHP. Potential future avenues are discussed as well.
Individuals and organizations regularly to choose from
a large number of available alternatives. Often, the alternatives have multiple attributes (for example, cost, quality, and durability) which makes the decision complex.
On the one hand, if only single attributes are considered
individual alternatives can be ranked on a one dimensional absolute preference scale. On the other hand, no
such scale may exist when all attributes are considered.
Nevertheless, it has been argued that it is possible for human agents to robustly compare pairs of alternatives in a
high dimensional attribute space1 . Pairwise comparison
matrices (PCMs) were first introduced in a nascent form
in psychophysics by Fechner in the 18602 and later rigorously defined by Thurstone in the 1920s3 . PCMs allow
agents to simplify complex decision making problems by
breaking them up into smaller tractable ones. Starting
from the 1970s, Saaty devised a framework to approximate the absolute preference scale from PCMs using his
analytic hierarchy process (AHP)1,4 and analytic network
process (ANP)4 .
Mathematically, PCMs are organized as follows. Consider that an agent has to choose from n > 1 alternatives
denoted by {a}. The entry Wab in a PCM W denotes the
preference of an agent for alternative a over alternative
b. We assume that Wba = 1/Wab . While PCMs simplify large complex problems, two drawbacks have been
identified.
First, agents may decide between pairs of alternative
using a combination of quantitative analysis and intuition. As a result, not all pairwise comparisons within a
matrix may be ‘consistent’ with each other1 . For example, if an agent prefers ‘a’ over ‘b’ by a factor of 2 and ‘b’
over ‘c’ by a factor of 2, in real PCMs, it is not guaranteed that the same agent will also prefer ‘a’ over ‘c’ by
a factor of Wab Wbc = 4. For a consistent PCM, we have
a) Email:
[email protected]
for any path over the alternatives a1 → a2 → . . . ak
log Wa1 ak =
k−1
X
log Wai ai+1 .
(1)
i=1
Consequently, the individual entries of a consistent PCMs
can be expressed as1
log Wab = log fa − log fb
(2)
for some absolute preference scale f¯ > 01 . In other
words, individual pairwise comparisons in a consistent
PCM can be represented by a ‘state function’ log f¯. Notably, the absolute preference scale is proportional to the
right Perron-Frobenius eigenvector ν̄ if the PCM is consistent (νa ∝ fa ). Based on the relationship between the
Perron-Frobenius eigenvector and the absolute preference
scale for consistent PCMs, Saaty in his AHP and ANP
argued that the same eigenvector also approximates the
absolute preference scale for inconsistent PCMs1,4 . The
AHP/ANP approach is now regularly used to infer absolute preference scales over alternatives in a broad range
of areas such as environmental sciences5 , organizational
studies6 , and public health7 .
In addition to their use in AHP, PCMs also allow
agents to identify the sources of departure from consistency in individual and group decision making8 . Over the
last three decades, several indices have been developed to
quantify consistency of PCMs8 . For example, a popular
index by Saaty1 quantifies the Perron-Frobenius eigenvalue η of the PCM. Saaty showed that η ≥ n for any
PCMs with equality holding iff the PCM is consistent.
Based on this observation, he defined the consistency index CI:
η−n
CI =
.
(3)
n−1
Other examples of consistency indices include the Harmonic consistency index9 and Geometric consistency index10 . Recently Brunelli et al.11,12 laid out a set of requirements for reasonable quantifiers of departures from
consistency of PCMs.
2
Second, several entries in a large PCM (n 1) may
be missing. The reasons are several
fold. The total num
ber of pairwise comparisons n2 increases as the square
of the total number of alternatives n. Indeed, psychological studies have shown that human agents are not
able to reliable estimate multiple pairwise comparisons
at the same time because of information overload or simply because they get bored and/or inattentive (see13 references). Moreover, not all comparisons may be realistically available (for example when rankingsports teams
or players with non-overlapping stints14,15 ). A number
of proposals fill up incomplete PCMs using the available
entries for subsequent analysis have been suggested16–20 .
A popular proposal by Harker16 is as follows. We assume
that the adjacency graph A of an incompletely filled PCM
W is fully connected. For any missing entry, say Wab , we
first enumerate all possible elementary paths {Γ} on A
between a and b. Next, we approximate the comparison
along each path Γ ≡ a → a1 → a2 → · · · → ak → b as if
the known entries in the PCM were described by a state
function,
!
k−1
X
log Wab (Γ) = log Waa1 +
log Wai ai+1
i=1
+ log Wak b .
(4)
The logarithm of the missing entry log Wab is then approximated as an arithmatic mean of log Wab (Γ) over all
possible elementary paths between a and b. All previously missing entries defined this way automatically satisfy Wab Wba = 1. Notably, most subsequent analyses of
PCMs assume that they are positive (Wab > 0 ∀ a and b).
Thus, it is not clear how particular filling up procedures
may influence analysis of PCMs.
In this work we address the following problem: is there
a way to analyze PCMs without relying on a process
to fill them up? We provide a physics-based answer.
First, we establish a novel connection between PCMs
and and time-irreversible statistical physics. Specifically,
we show that every PCM (incompletely filled or otherwise) induces a family of Markovian random walks over
the alternatives. The random walks are maximum path
entropy random walks constrained to reproduce a ‘flux’
J ∝ hlog Wab i21–23 . This connection allows us to bring
insights from recent work in stochastic thermodynamics24 to study of PCMs. Notably, we show that the entropy production rate in the induced random walks is
intricately related to the consistency of the underlying
PCM. Quantification of entropy production does not require filling up of the PCM as long as the adjacency graph
A of the PCM is connected. Moreover, the entropy production can be decomposed as either a sum over (a) alternatives or (b) pairwise comparisons which allows us
to directly identify the alternatives or the comparisons
that are inconsistent with the rest of the PCM. We provide physics-based explanations for independently laid
out conditions for reasonable inconsistency indices11,12 .
We also show that the absolute preference scale can be
extracted from an incompletely filled but otherwise consistent PCM by correcting for the effect of the adjacency
graph of the incomplete PCM. This allows us to generalize Saaty’s AHP for incompletely filled PCMs.
We numerically compare our consistency index with
previously developed ones. We illustrate our development by systematically examining the effect of filling up
incompletely filled matrices on the evaluation of consistency and the AHP. We show that filling up a PCM introduces systematic biases in evaluation of consistency.
Importantly, we believe that the connections established
in this work between two previously unrelated fields of
scientific inquiry will allow a greater understanding of
consistency of pairwise comparison matrices in the future.
I.
RESULTS
A.
PCM-induced random walk
Consider an incompletely filled PCM W . We assume
that the unquerried entries in W are set to zero and
Wab > 0 ⇒ Wab Wba = 1. We also assume that the graph
of adjacency A of W is connected. We have Aab = 1 if
Wab > 0. We define a microscopic ‘flux’ jab = log Wab
between vertices a and b of A that have an edge between
them (Wab > 0 in the PCM). The microscopic flux is
antisymmetric; jab + jba = 0.
We construct a discrete time Markov process with transition probabilities {kab } and a stationary distribution
{pa } on A that is consistent with a given ensemble average flux J = hjab i per unit time. The ensemble average
flux is given by
X
hjab i =
pa kab jab .
(5)
a,b
There are infinitely many Markov processes consistent
with a single path ensemble average. We seek the one
with the maximum path entropy. The path entropy is
given by23,25,26
X
S=−
pa kab log kab .
(6)
a,b
Maximization of S is a constrained problem23,25,26 because {pa } and {kab } are dependent of each other,
X
X
pa kab = pa ,
pa kab = 1, and
(7)
b
X
a,b
pa kab = pb
(8)
a
and
X
a,b
pa kab jab = hjab i = J
(9)
3
Eqs. 8 represent the constraint of probability conservaThe entropy production rate ṡ of the induced MERWs
tion and normalization and {pa } as the stationary distridefined in Eq. 14 serves as a physics-based quantifer of
bution respectively. Eq. 9 represents the imposed path
the inconsistency of the underlying PCM. We prove that
ensemble constraint of the flux J. We solve the conṡ = 0 iff the underlying PCM is consistent.
strained problem using the method of Lagrange MultiFirst, consider ṡ = 0. We show that all non-zero entries
pliers. We write the unconstrained Caliber27
of W are given by Wab = fa /fb for some absolute scale
!
! f¯ > 0. Detailed balance implies
X
X
X
X
νb
νa
C=S+
la
pa kab − pa +
mb
pa kab − pb
µa νa
Wab = µb νb
Wba
(15)
ηνa
ηνb
a
a
b
b
!
µb
νa
Wab
2
X
X
= Wab
=
×
(16)
⇒
Wba
νb
µa
+δ
pa kab − 1 + γ
pa kab jab − J
(10)
ab
a,b
fa
⇒ Wab =
(17)
fb
In Eq. 10, {la }, {mb }, and δ are the Lagrange multipliers
where fa = (νa /µa ).
that impose the constraints in Eq. 8. γ is the Lagrange
Next, consider a PCM W whose non-zero entries are
multiplier that imposes the dynamical flux constraint.
given by Wab = fa /fb where f¯ > 0 is a vector of posiThe transition probabilities of maximum path entropy
tive elements. We evaluate the entropy production rate
random walks (MERW) are given by23
of the induced MERWs. First, we derive the transition
νb (γ)
γ
probabilities kab . We write
W
(11)
kab (γ) =
η(γ)νa (γ) ab
W = Diag(f¯)ADiag(1/f¯)
(18)
where η(γ) is the largest eigenvalue of the modified PCM
¯
¯
⇒ W Diag(f ) = Diag(f )A
(19)
W γ , and ν̄(γ) is the corresponding right eigenvector.
In Eq. 19, Diag(x̄) is the diagonal matrix with entries
From here onwards, we recognize W γ as the element-wise
from
the vector x̄ and A is the adjacency matrix of W .
exponentiation and not the matrix exponentiation. AcLet
η
be the Perron-Frobenius eigenvalue of A and ν̄
cording to the Perron-Frobenius theorem, ν̄(γ) has posibe the corresponding Perron-Frobenius eigenvector. We
tive entries and η(γ) > 0. Finally, the stationary distrihave
bution {pa (γ)} is given by the outer product23
pa (γ) = νa (γ)µa (γ)
(12)
where µa (γ) is the left Perron-Frobenius eigenvector of
W γ . We call the γ-dependent family of Markov processes
described by Eq. 11 the maximum path entropy random
walks induced by the PCM. Notably, both the left and
the right Perron-Frobenius eigenvectors are used as an
approximate absolute preference scale in Saaty’s AHP1,4 .
We note that the Markov process at γ = 1 corresponds
to the original PCM W . From here onwards, unless specified otherwise we will assume that γ = 1. We omit the
dependence on γ for brevity.
B.
Entropy production as a metric of inconsistency
The entropy production rate of a Markov process quantifies the degree of irreversibility in it; the entropy production rate is zero iff the process is time-symmetric and
satisfies detailed balance pa kab = pb kba ∀ a and b24 . We
have23
X
kab
(13)
ṡ =
pa kab log
kba
a,b
X
νb2
=
pa kab log 2 + 2 log Wab
νa
a,b
2X
=
µa νb Wab log Wab ≥ 0.
η
(14)
Aν̄ = ην̄
¯
⇒ W Diag(f )ν̄ = Diag(f¯)Aν̄
= ηDiag(f¯)ν̄.
(20)
(21)
In other words, the Perron-Frobenius eigenvalue of W is η
and the corresponding eigenvector is ḡ where ga = νa fa .
The transition probabilities of the MERWs are given by
(see Eq. 11)
kab =
ν b fb
νb
× fa /fb Aab =
Aab
ηνa fa
ηνa
(22)
where Aab = 1 if a and b are connected by an edge
(if the comparison between a and b is available) and
zero otherwise. Since A is symmetric (Aab = Aba ∀ a
and b), the Markov process described by Eq. 22 satisfies detailed balance and has a zero entropy production
rate21–23 . In other words, the induce MERWs is detailed
balanced iff the underlying PCM is consistent. Notably,
Brunelli’s first requirement for any metric that measures
inconsistency is the ability to uniquely identify consistent
PCMs11,12 . As shown here, the entropy production rate
satisfies this requirement.
Why does ṡ quantify consistency? Let us consider the
detailed balanced Markov process induced by a consistent PCM. As we showed above, the MERW induced by
a consistent PCM is detailed balanced (pa kab = pb kba ∀ a
and b). An illustrative analogy is to imagine that the
MERW describes a system at thermodynamic equilibrium with a surrounding bath. Let us assume that
4
the stationary distribution of the MERW is given by
pa ∝ e−βEa where Ea is the ‘energy’ of the alternative a. Let us consider a path over the alternatives
Γ = a1 → a2 → · · · → am and the corresponding time
reversed path Γ0 = am → am−1 → · · · → a1 .
The log ratio of the forward and the reverse path probabilities is related to the total heat exchanged during the
trajectory. We have,
ka a × · · · × kam−1 am
p(Γ|a1 )
log
(23)
= log 1 2
p(Γ0 |am )
kam am−1 × · · · × ka2 a1
pa2
pam
= log
× ··· ×
(24)
pa1
pam−1
= ∆Q = −β (Em − E1 )
(25)
Notably, the heat exchange is independent of the path
only for detailed balance processes. Specifically, the heat
exchange is zero for all loops i.e. am = a1 ⇒ ∆Q = 0.
In contrast, the heat dissipation in the MERW induced
by an inconsistent PCM depends on the entire history of
the trajectory. We have23
log
p(Γ|a1 )
= 2γ
p(Γ0 |am )
m−1
X
jai ai+1 + 2 log
i=1
νam
.
νa1
(26)
To translate this observation in the language of PCMs
let us consider a Markovian random walker on a looped
trajectory of the induced MERW. We imagine that the
random walker exchanges ‘energy’ with the ‘surrounding’. Every time step when the walker goes to an alternative that is less favored compared to the current
one (Wai ai+1 > 1) she receives energetic renumeration
∝ log Wai ai+1 . However, she has to pay the same amount
of energy when she goes to an alternative that is more
favored. On the one hand, if the MERW is detailed balanced (if the PCM is consistent), the walker will end up
with no net change in energy over any loop. On the other
hand, if the PCM is inconsistent, there will exist loops
which have a net exchange of energy between the walker
and the surrounding.
The induced MERWs have few other notable properties. Consider a long path Γ = a1 → a2 → . . . of length
T 1 of the MERW for a fixed value of γ. From Eq. 11,
we write the probability p(Γ)23
1
p(Γ) ≈
e(T −1)γj(Γ)
(27)
η(γ)T −1
We note that η(−γ) is the Perron-Frobenius eigenvalue
γ
of W −γ = (Trans(W )) where Trans(W ) is the transpose
of W . Since the eigenvalues of W and Trans(W ) are the
same, we conclude that η(−γ) = η(γ) i.e. η(γ) is an even
function of γ. Consequently, it’s derivative hjab i(γ) is an
odd function of γ. Moreover,
X
d
1 d2
log
e(T −1)γj(Γ)
hjab i(γ) =
2
dγ
T − 1 dγ
(31)
2
= hjab
i − hjab i2 ≥ 0.
(32)
Γ
Thus, hjab i(γ) is a monotonic function of γ. As a result,
(1) ṡ(γ) is a monotonically increasing function of γ for
γ > 0 and since hjab i(γ) is an odd function of γ (2)
ṡ = γhjab i(γ) is an even function of γ. Notably, these
two observations directly correspond with requirement
(3) and (6) laid out by Brunelli’s11,12 . In appendix A1,
we show that ṡ(γ) satisfies all of Brunelli’s requirements.
C.
Using incomplete PCMs in the AHP
Can we perform Saaty’s AHP analysis on an incompletely filled PCM? As above, let us consider an incompletely filled but otherwise consistent PCM. We have
Wab > 0 ⇒ Wab = fa /fb .
(33)
Can we extract the absolute preference scale from W ?
Eq. 21 shows that the Perron-Frobenius eigenvector ḡ of
W is given by ga = νa fa where ν̄ is the Perron-Frobenius
eigenvector of the adjacency matrix A corresponding to
W . Notably, the absolute preference scale can be extracted from an incompletely filled but otherwise consistent PCM not as the reciprocal of the right PerronFrobenius eigenvector ḡ, but with a correction that accounts for the connectivity in the adjacency graph:
fa = ga /νa .
(34)
Similar in spirit to the original observation of the AHP,
we propose that for inconsistent and incompletely filled
PCMs, the absolute preference scale f¯ can be approximated using the Perron-Frobenius eigenvector ḡ of the
PCM W and the Perron-Frobenius eigenvector ν̄ of the
corresponding adjacency matrix A as shown in Eq. 34.
where
j(Γ) =
T −1
1 X
ja a
T − 1 i=1 i i+1
(28)
is the flux per unit time associated with the path Γ. We
recognize η(γ)T −1 as the partition function. Since p(Γ)
is normalized, we write (in the limit T 1)
X
η(γ)T −1 =
e(T −1)γj(Γ)
(29)
Γ
d
⇒ hjab i(γ) =
log η(γ).
dγ
(30)
II.
ILLUSTRATIVE EXAMPLES
A. Numerical comparison between ṡ and other
consistency indices
We illustrate ṡ as a quantifier of inconsistency by evaluating the inconsistency of multiple incompletely filled
pairwise comparison matrices. Specifically, we study
Saaty’s CI (see Eq. 3) and the harmonic consistency index (HCI)9 . We have already introduced Saaty’s CI, here
5
A
B
D
F
N
S
A
B
D
F
1 1.39 0 0.76
0.72 1
0
0
0
0
1 0.95
1.32 0 1.05 1
1.11 0 1.29 1.91
1.36 1.3 0 0.95
N
S
0.9 0.73
0 0.77
0.77 0
0.52 1.05
1
0
0
1
TABLE I. Pairwise comparison matrix among Tennis players
based on their performance against each other. Entries between players who did not play any matches between them
are set to zero.
FIG. 1. Panel (a) Comparison between Saaty’s CI and ṡ as
defined in Eq. 14 for randomly generated PCMs with varying
degree of inconsistency. Panel (b) Comparison between the
Harmonic consistency index HCI and ṡ as defined in Eq. 14
for randomly generated PCMs with varying degree of inconsistency.
we briefly introduce the HCI and the GCI. Since consistent PCMs W have rank(W ) = 1, the colums P
are proportional to each other.
Consequently,
if
s
=
a
a Wab ,
P −1
it was proven that
sa = 1 iff W is consistent9 .
The HCI P
quantifies deviations of the harmonic mean
HM = n/ s−1
a from n. We have
HCI =
(HM − n) (n + 1)
n(n − 1)
(35)
We construct an ensemble of PCMs with n = 5 of
varying degree of inconsistency. First, we construct an
absolute scale f¯ > 0 by drawing n = 5 uniformly distributed random numbers between [0, 1]. We construct
a PCM W with elements
Wab =
fa
exp (ρab α) for a > b.
fb
(36)
The parameter α controls the inconsistency. PCMs with
α = 0 are consistent (but incompletely filled) and the
inconsistency increases as α increases. We choose α to
be randomly distributed between [0, 4]. ρab are normally
distributed random numbers with zero mean and unit
standard deviation. The lower-diagonal entries of W are
filled to satisfy the reciprocal relationship Wab Wba = 1.
In Fig. 1, we compare our inconsistency index ṡ with
the CI and the HCI. The dashed red line shows x = y.
Notably, ṡ correlates extremely well with Saaty’s CI
FIG. 2. The graph of adjacency between Tennis players corresponding to the pairwise comparison matrix given in Table I.
(Pearson r2 = 0.98). This may be because ṡ directly depends on the Perron-Frobenius eigenvalue η (see Eq. 14).
ṡ also correlates well with the HCI. However the correlation is lower (Pearson r2 = 0.71) and there is a a large
scatter.
B.
Inferring preference scales from incomplete matrices
We now show how to infer the absolute preference scale
using an incompletely filled PCMs using Eq. 34 as proposed above. We compare our approach with the approach by Harker16 described above.
Recently, Bozóki et al.15 studied the problem of determining ranking among 25 Tennis players based on their
performance against each other. Notably, not all players played with each other, for example, Agassi never
played a match with Nadal. Consequently, the PCM
constructed using players’ performance is inherently incomplete. Here, we carry out an analysis on 6 of the
25 players; Agassi (A), Baker (B), Djokovic (D), Federer
(F), Nadal (N), and Samprass (S). The PCM is given in
Table I (see Bozóki et al.15 for details). We set to zero
all incompletely filled entries. The graph of connectivity
among the tennis players is shown in Fig. 2.
In Table II, we show the results of our calculations.
First, we estimate the absolute preference scale using
Harker’s method16 . The ‘filled’ PCM is given in Ta-
6
A
B
D
F
N
S
f¯H
0.150
0.122
0.166
0.161
0.232
0.170
ν̄
0.211
0.120
0.120
0.211
0.170
0.170
ḡ
0.188
0.083
0.116
0.208
0.233
0.173
f¯
0.150
0.117
0.164
0.166
0.231
0.172
TABLE II. Comparison of the estimated absolute preference
scale using Harker’s method16 (f¯H ) and Eq. 34 (f¯).
ble A1. The principal eigenvector of the filled PCM f¯H
is given in column 1. The absolute preference scale is
L1 normalized. In order to evaluate the absolute preference scale using Eq. 34, we first evaluate the right
Perron-Frobenius eigenvector ν̄ (column 2). Next, we
evaluate the right Perron-Frobeius eigenvector of the incompletely filled matrix in Table I ḡ (column 3). Finally,
the L1 normalized estimate of the absolute preference
scale using Eq. 34 is given by f¯ (column 4). Notably, f¯H
and f¯ are highly correlated with each other (Pearson’s
r2 = 0.99, p = 3.3 × 10−5 ).
III.
CONCLUSION
In this work, we established a novel connection between a popular tool in decision theory; pairwise comparison matrices (PCMs) and non-equilibrium statistical physics. Specifically, we showed that PCMs induce
a family of maximum path entropy random walks constrianed to reproduce a non-equilibrium flux. Notably,
only consistent PCMs (incompletely filled or otherwise)
induce detailed balanced random walks. Based on these
insights, we proposed the entropy production rate ṡ in
the induced MERWs as a quantifier of inconsistency. We
showed that ṡ satisfies all previously laid out criteria for
reasonable consistency indices. We also showed how to
use incompletly filled PCMs in Saaty’s AHP.
We hope that our work brings together two previously unrelated areas of scientific inquiry namely nonequilibrium Markov processes and pairwise comparison matrices. Notably, recent years have seen a renewed interest in the study of statistical physics of nonequilibrium Markov processes (reviewed by Seifert28 ).
For example, many new identities such as various ‘fluctuation theorems’ have been discovered across a wide range
of settings. We speculate that the connections established in the current work will allow a greater exchange
of ideas between the two previously unrelated fields of
inquiry and potentially refine our understanding of consistency in pairwise comparisons.
IV.
ACKNOWLEDGMENTS
I would like to Thank Dr. Jason Wagoner for stimulating discussions on non-equilibrium flow processes that
lead to an investigation into pairwise comparison matrices. I would also like to thank Dr. Matteo Brunelli and
Dr. Luis Vargas for a critical reading of the manuscript
and multiple valuable suggestions.
1
T. L. Saaty and L. G. VargasModels, methods, concepts & applications of the analytic hierarchy process Vol. 175 (Springer
Science & Business Media, 2012).
2 G. T. FechnerElemente der psychophysik Vol. 2 (, 2012).
3 L. L. Thurstone, Psychological review 34, 273 (1927).
4 T. L. Saaty, Journal of systems science and systems engineering
13, 1 (2004).
5 R. Ramanathan, Journal of environmental management 63, 27
(2001).
6 R. L. Nydick and R. P. Hill, Journal of supply chain management
28, 31 (1992).
7 M. J. Liberatore and R. L. Nydick, European Journal of Operational Research 189, 194 (2008).
8 M. Brunelli, L. Canal, and M. Fedrizzi, Annals of Operations
Research 211, 493 (2013).
9 W. E. Stein and P. J. Mizzi, European journal of operational
research 177, 488 (2007).
10 J. Aguaron and J. M. Moreno-Jiménez, European journal of
operational research 147, 137 (2003).
11 M. Brunelli and M. Fedrizzi, Journal of the Operational Research
Society 66, 1 (2015).
12 M. Brunelli, Annals of Operations Research 248, 143 (2017).
13 F. J. Carmone, A. Kara, and S. H. Zanakis, European journal of
operational research 102, 538 (1997).
14 L. Csató, Central European Journal of Operations Research , 1
(2013).
15 S. Bozóki, L. Csató, and J. Temesi, European Journal of Operational Research 248, 211 (2016).
16 P. T. Harker, Mathematical Modelling 9, 837 (1987).
17 P. T. Harker, Mathematical Modelling 9, 353 (1987).
18 W. W. Koczkodaj, M. W. Herman, and M. Orlowski, Knowledge
and Information Systems 1, 119 (1999).
19 M. Fedrizzi and S. Giove, European Journal of Operational Research 183, 303 (2007).
20 S. BozóKi, J. Fülöp, and L. RóNyai, Mathematical and Computer Modelling 52, 318 (2010).
21 Z. Burda, J. Duda, J. Luck, and B. Waclaw, Physical review
letters 102, 160602 (2009).
22 L. R. Frank and V. L. Galinsky, Physical Review E 89, 032142
(2014).
23 P. D. Dixit, Physical Review E 92, 042149 (2015).
24 U. Seifert, The European Physical Journal B-Condensed Matter
and Complex Systems 64, 423 (2008).
25 P. D. Dixit and K. A. Dill, J. Chem. Th. Comp. 10, 3002 (2014).
26 P. D. Dixit, A. Jain, G. Stock, and K. A. Dill, Journal of chemical
theory and computation 11, 5464 (2015).
27 P. D. Dixit et al., Journal of Chemical Physics (2017).
28 U. Seifert, Physica A: Statistical Mechanics and its Applications
(2017).
7
A1. ṡ(γ) SATISFIES REQUIREMENTS FOR
REASONABLE CONSISTENCY INDICES
Recently, Brunelli et al.11,12 laid out six requirements
for any index I(W ) that quantifies the inconsistency in
a PCM W . Here, we show that the entropy production
rate ṡ(γ) satisfies all of those requirements. They are as
follows
1. I(W ) uniquely identifies consistent PCMs. If W is
a family of cosistent PCMs then I(W ∈ W) = v ∗
for some v ∗ ∈ R. Conversely, I(W ) = v ∗ ⇒ W ∈
W.
2. I(P W Trans(P )) = I(W ) for any permutation matrix P .
3. I(W γ ) > I(W ) if γ > 1.
element-wise exponentiation.
Here, W γ denotes
4. We start with a consistent PCM W . We choose
δ
one entry Wab and transform it Wab → Wab
,
γ
δ 6=. We also transform Wba → Wba . The resultant PCM W 0 (δ) is not consistent. We require
I(W 0 (δ)) > I(W 0 ()) > if δ > > 1. We also
require I(W 0 (δ)) > I(W 0 ()) > if δ < < 1.
5. I(W ) is continuous with respect to entries in W .
6. I(W ) = I(Trans(W )).
We proved that ṡ(γ) satisfies requirements (1), (3), and
(6) in the main text. Requirement (2) implies that the entropy production rate in the Markov process is invariant
under permutation of vertex labels. ṡ(γ) trivially satisfies this requirement as the entropy production rate is the
global property of the entire Markov process. ṡ(γ) satisfies requirement (5) as well. Since the Perron-Frobenius
eigenvalue η(γ) and the eigenvector ν̄(γ) are continuous
with respect to the elements of W 1 , ṡ(γ) is also continuous with respect to elements of W .
While we couldn’t prove that ṡ satisfies requirement
(4), we provide evidence that it is true based on a conjecture that we numerically checked.
Consider two consistent pairwise comparison matrices
W and Q. We have Wab > 0 ⇒ Wab fb /fa for some scale
f¯ > 0 and Qab > 0 ⇒ Qab = gb /ga for some other unrelated scale ḡ > 0. We assume that the adjacency graph A
corresponding to W and Q is identical and is connected.
We create two new matrices where a specific entry Wab
(and Qab ) is changed to Wab → αWab (and Qab → αQab ).
We also change the corresponding reciprocal entry Wba
(and Qba ). Let us call the modified matrices W (α) and
W (α) respectively. Note that W (α) and Q(α) are not
consistent if α 6= 1.
Eq. 22 of the main text suggests that the induced
MERW of a consistent matrix W (and Q) only depends
on the properties of the adjacency graph A. Hence, the
MERW induced by consistent PCMs W and Q are identical. Surprisingly, based on our numerical calculations
we observe that the MERWs induced by W (α) and Q(α)
are also identical. We conjecture that this is true.
Requirement (4) follows from this conjecture. Consider
an incompletely filled but otherwise consistent PCM W .
Let A denote its adjacency graph. We note that A is also
a consistent PCM. As above, let us modify Wab → αWab
and Aab = αAab for some specific entry. We assume
that α 6= 1. Let us denote the two modified matrices by
W (α) and A(α). Based on our conjecture, the family of
MERWs induced by W (α) is identical to the family of
MERWs induced by A(α) ∀ α. Consequently ṡ(W (α)) =
ṡ(A(α)).
Next, let us consider W (α1 ) and W (α2 ) such that α2 >
α1 > 1. To prove that the entrop production rate ṡ satisfies requirement (4), we need to show that ṡ(W (α2 )) >
ṡ(W (α1 )). First, we note that Aab (α2 ) = Aab (α1 )τ where
τ = logα1 α2 > 1. Since ṡ satisfies requirement (3), we
have ṡ(A(α2 )) > ṡ(A(α1 )) ⇒ ṡ(W (α2 )) > ṡ(W (α1 )).
This proves that ṡ satisfies requirement (4).
A.
Harker’s method to fill an incomplete PCM
A
B
D
F
N
S
A
1
0.72
1.21
1.32
1.11
1.36
B
1.39
1
1.36
1.15
2.02
1.3
D
0.83
0.74
1
1.05
1.29
1.05
F
0.76
0.87
0.95
1
1.91
0.95
N
0.9
0.50
0.77
0.52
1
0.71
S
0.73
0.77
0.95
1.05
1.42
1
TABLE A1. The ‘filled’ PCM using Harker’s method. The
original PCM is given in Table I.
| 2cs.AI
|
1
Pearson information-based lower bound
on Fisher information
arXiv:1611.07712v1 [math.ST] 23 Nov 2016
Dave Zachariah and Petre Stoica
Abstract—The Fisher information matrix (FIM) plays an
important role in the analysis of parameter inference and system
design problems. In a number of cases, however, the statistical
data distribution and its associated information matrix are either
unknown or intractable. For this reason, it is of interest to develop
useful lower bounds on the FIM. In this lecture note, we derive
such a bound based on moment constraints. We call this bound
the Pearson information matrix (PIM) and relate it to properties
of a misspecified data distribution. Finally, we show that the
inverse PIM coincides with the asymptotic covariance matrix of
the optimally weighted generalized method of moments.
The Cramér-Rao bound (CRB) is a useful tool for the analysis of parameter inference problems, benchmarking estimators,
and system design [1]–[4]. Let y = [y1 · · · yN ]⊤ denote the
observed data from a system and let θ ∈ Rn be the parameters
of interest. The CRB exists under certain regularity conditions
and is given by the inverse of the Fisher information matrix
(FIM) J(θ), which is a function of the probability density
p(y; θ). More specifically, the FIM is defined as
#
"
⊤
∂ ln p(y; θ) ∂ ln p(y; θ)
0,
(1)
J(θ) , E
∂θ
∂θ
∂ ln p(y;θ)
∂θ
where the gradient
is known as the ‘score function’.
Under certain regularity conditions, the score function has zero
mean
∂ ln p(y; θ)
= 0.
(2)
E
∂θ
In many applications, J(θ) may not be obtainable. For
instance, p(y; θ) may be unknown in a practical problem. As
an example, let the observations be modeled as y = f (u),
where f (·) is a nonlinear function and u follows a probability
density function p(u; θ). Then a closed-form expression for
p(y; θ) is not available in general even if p(u; θ) is known.
It is possible, however, to derive tractable lower bounds on
the FIM. If E[y] is a function of θ and Cov[y] is independent
of θ, then using the Gaussian distribution in lieu of p(y; θ)
leads to the minimum FIM JG (θ) J(θ) (where G stands
for Gaussian) and therefore the ‘worst-case’ inference scenario
[5], [6]. A generalization to θ-dependent Cov[y] was given
in [7]. The minimum FIM can be used for robust system
design and estimator formulations. However, in some cases
the minimum information can be overly conservative.
In this lecture note, we derive a tighter lower bound
0 L(θ) J(θ) ∈ Rn×n ,
(3)
based on moment constraints. For this reason we call L(θ)
the Pearson information matrix (PIM) as an homage to the
This work has been partly supported by the Swedish Research Council
(VR) under contract 621-2014-5874.
inventor of ‘the method of moments’ [8]. As we will see, the
PIM is related to the generalized method of moments [9].
I. R ELEVANCE
The PIM is a tractable tool for analyzing parameter estimation and system design problems when the statistical data
distribution is unknown or intractable.
II. P REREQUISITIES
The reader needs basic knowledge about linear algebra,
elementary probability theory, and statistical signal processing.
III. P RELIMINARIES
We begin by constructing a function z(y) that contains M
statistics of y. We assume that z(y) has computable—either
analytically or numerically—mean and covariance
µ(θ) , E[z(y)] ∈ RM ,
⊤
Σ(θ) , E[(z(y) − µ) (z(y) − µ) ] ∈ RM×M ,
(4)
where M ≥ n. For instance, z may be constructed using
powers of the data, that is, its elements are made up of
empirical moments {yi } {yi yj }, {yi yj yk }, etc. We assume
that Σ(θ) is nonsingular. For notational simplicity, we drop the
argument θ in the next analysis and reinstate it when needed.
We also write pθ = p(y; θ).
IV. P EARSON
INFORMATION MATRIX
In Section IV-A, we begin with a step-by-step algebraic
derivation of L in (3) which will define the Pearson information matrix. As explained there, the PIM generalizes
the results in [5]–[7] and coincides with a bound recently
derived in [10] (by comparison this lecture notes provides
a simple textbook-style derivation of the bound as well as
further connections). In Section IV-B we go on to provide
an information-theoretic connection between the PIM and
misspecified data distributions using the principle of maximum
entropy [11]. Then we study the behaviour of PIM when M
increases in Section IV-C. Finally, in Section V, we establish a
relation between the PIM and generalized method of moments
that is analogous to the relation between the FIM and the
maximum likelihood method. The presented results enable a
tractable analysis of a wider class of data models that satisfy
certain moment constraints.
2
A. Algebraic derivation
Consider a linear combination of the centered statistics z −
µ:
D⊤ W(W⊤ ΣW)−1 W⊤ D D⊤ Σ−1 D,
⊤
W (z − µ),
Under regularity conditions that allow the interchanging of
integral and derivative operations [2], the following identity
holds:
∂ ln pθ
E
(z − µ)⊤
∂θ
Z
1 ∂pθ
(z − µ)⊤ pθ dy
=
pθ ∂θ
Z
∂µ⊤
∂pθ (z − µ)⊤
dy
+ pθ
=
(6)
∂θ
∂θ
Z
∂
∂µ⊤
=
E (z − µ)⊤ + pθ dy
∂θ |
∂θ
{z
}
| {z }
=0
=1
= D⊤ ,
D⊤ =
(12)
−1
where W⊤ ∈ Rn×M denotes a linear combiner matrix. This
vector has zero-mean similar to the score function, cf. (2). We
construct the following matrix
∂ ln p
⊤
∂ ln p
θ
θ
∂θ
∂θ
E
0.
(5)
W⊤ (z − µ) W⊤ (z − µ)
where
Using the Schur complement of the lower-right block of (11)
we obtain the upper bound
∂µ⊤
∈ Rn×M
∂θ
(7)
is the gradient of the mean vector. Using (4) and (6), the matrix
in (5) can be expressed as
J
D⊤ W
0.
(8)
W⊤ D W⊤ ΣW
It follows from the Schur complement of the lower-right block
of (8) that
0 D⊤ W(W⊤ ΣW)−1 W⊤ D J,
(9)
assuming that W⊤ ΣW has full rank [12]. Equation (9) yields
a nonnegative lower bound on the FIM that is dependent on
the choice of the linear combiner W.
The tightest lower bound (9) is found by solving the
problem
max D⊤ W(W⊤ ΣW)−1 W⊤ D.
(10)
W
The combiner that produces the tightest bound is W⋆ =
Σ−1 D. To show this, begin by constructing the following
positive semidefinite matrix,
⊤ −1
D Σ D
D⊤ W
W⊤ D
W⊤ ΣW
⊤ −1
I
0
D Σ D D⊤ I 0
=
0 W
0 W⊤
D
Σ
⊤ −1/2 ⊤ −1/2 ⊤
I 0
D Σ
I
0
D Σ
0.
=
0 W
0 W⊤
Σ1/2
Σ1/2
(11)
which is clearly attained at W⋆ = Σ D.
In conclusion, using (9) and (12), we have proved the
following theorem.
Theorem 1. The optimal lower bound (in the class of bounds
considered) is
L , D⊤ Σ−1 D J,
(13)
where Σ and D are either obtained analytically or computed
numerically. We call L 0 the Pearson information matrix
for reasons explained above.
Remark: Suppose y ∼ pθ can be modeled as
y = µ(θ) + w ∈ RN ,
where w is a zero-mean random variable. Let z(y) = y. Then
the corresponding PIM coincides with the FIM bounds in [5],
[6] and in [7], when the covariance matrix is fixed, Σ, and
variable, Σ(θ), respectively. The above algebraic derivation
of the PIM provides, moreover, a simple textbook-like proof
of the optimized FIM bound in [10] (which also contains
an illustrative example consisting of a nonlinear amplification
device).
B. Connection to misspecified data distributions
We now relate L to certain properties of misspecified data
distributions using the principle of maximum entropy. Instead
of the unknown or intractable distribution pθ , we will use
an alternative statistical model, denoted p⋆ , along with the
following identity:
ln pθ = ln p⋆ + ln δ⋆ ,
which holds for any choice of p⋆ , where δ⋆ = ppθ⋆ .
The uncertainty of the data y is quantified by the (differential) entropy which can be decomposed as
H(pθ ) , − E[ln pθ ] = − E[ln p⋆ ] − ∆(pθ ||p⋆ ),
where ∆(pθ ||p⋆ ) = E[ln δ⋆ ] ≥ 0 is the divergence of p⋆ from
the unknown distribution pθ [13], [14]. We decompose the
score function into
∂ ln pθ
∂ ln p⋆
∂ ln δ⋆
=
+
,
(14)
∂θ
∂θ
∂θ
where the terms correspond to a misspecified score and a
divergence score, respectively. The misspecified information
matrix is defined as
"
#
∂ ln p⋆ ∂ ln p⋆ ⊤
J⋆ , E
0,
(15)
∂θ
∂θ
where the expectation is taken with respect to pθ .
Lemma 1. A general lower bound on J is
e J,
J⋆ + J
(16)
3
where
"
e = E ∂ ln δ⋆ ∂ ln p⋆
J
∂θ
∂θ
⊤
#
+E
"
∂ ln p⋆ ∂ ln δ⋆
∂θ
∂θ
⊤
#
where
that
.
(17)
Proof: Inserting (14) into (1), we obtain the following
decomposition
#
"
∂ ln δ⋆ ∂ ln δ⋆ ⊤
e
(18)
J = J⋆ + J + E
∂θ
∂θ
and the result follows immediately.
We are concerned with misspecified data models p⋆ that
satisfy the given constraint E[z] = µ. That is, distributions
that satisfy
Z
zp⋆ dy = µ.
(19)
In particular, we let p⋆ correspond to the maximum uncertainty
of y. The distribution with the maximum (differential) entropy
is known to be
p⋆ ≡ arg max − E′ [ln p′ ] = exp(λ⊤ z − λ0 ),
p′ ∈P
where P is the set of valid probability distributions for y that
satisfy (19) and λ0 , λ are multipliers that are chosen to satisfy
the constraint (assuming that the problem is feasible) [14]. For
completeness, we prove this result by noting that the following
upper bound holds for any p′ :
H(p′ ) = − E′ [ln p′ ]
≤ − E′ [ln p⋆ ]
Z
= − (λ⊤ z − λ0 )p′ dy
Z
= − (λ⊤ z − λ0 )p⋆ dy
= H(p⋆ ).
The equality in the penultimate line follows since both p′
and p⋆ satisfy the constraint (19). The maximum entropy
distribution therefore belongs to the exponential family, that
is,
p⋆ (y; θ) = exp λ⊤ (θ)z(y) − λ0 (θ) ,
(20)
where
λ0 (θ) = ln
Z
is a normalizing constant.
exp λ⊤ (θ)z(y) dy
is n × M . The second line follows from the fact
Z
∂λ0
1
∂ exp(λ⊤ z)
= R
dy
∂θ
∂θ
exp(λ⊤ z)dy′
!
Z
exp(λ⊤ z)
∂λ⊤
dy
z R
=
∂θ
exp(λ⊤ z)dy′
Z
∂λ⊤
=
zp⋆ dy
∂θ
Z
∂λ⊤
=
zpθ dy
∂θ
∂λ⊤
=
µ.
∂θ
Next, from the proof of (16) we have that
#
"
∂ ln δ⋆ ∂ ln δ⋆ ⊤
e
.
J − (J⋆ + J) = E
∂θ
∂θ
Proof: For (20), we have that
(22)
(23)
(24)
Using (14), the divergence score can be written as the sum of
two random vectors,
∂ ln δ⋆
∂ ln pθ
⊤ −1
=
− D Σ (z − µ)
∂θ
∂θ
(25)
∂ ln p⋆
.
+ D⊤ Σ−1 (z − µ) −
∂θ
For the maximum entropy distribution, these two random
vectors are orthogonal, i.e.,
h ∂ ln p
θ
− D⊤ Σ−1 (z − µ)
E
∂θ
⊤ i
∂ ln p⋆
⊤ −1
× D Σ (z − µ) −
∂θ
∂λ
= D⊤ − D⊤ Σ−1 Σ Σ−1 D −
∂θ
= 0,
where the equality follows from (22) and (6). Finally, by
inserting (25) into (24), the right hand side of (24) equals
⊤
∂λ
∂λ
−1
−1
⊤ −1
Σ Σ D−
J−D Σ D+ Σ D−
∂θ
∂θ
and result (21) follows.
Theorem 2. The tightest bound (16) is the PIM:
e L J,
J⋆ + J
Lemma 2. Using the maximum entropy distribution (20), the
bound in (16) is given by
⊤
∂λ
e = L − Σ−1 D − ∂λ
(21)
Σ Σ−1 D −
J⋆ + J
∂θ
∂θ
∂λ⊤
∂λ0
∂ ln p⋆
=
z−
∂θ
∂θ
∂θ
∂λ⊤
(z − µ)
=
∂θ
∂λ⊤
∂θ
(26)
and the corresponding misspecified information matrix is J⋆ =
L.
Proof: It follows from (22) that for maximum entropy
distributions, the misspecified information matrix is
∂λ⊤ ∂λ
Σ .
∂θ
∂θ
Furthermore, in (21) it is readily seen that the tightest bound
is attained for
∂λ⊤
= D⊤ Σ−1 .
(27)
∂θ
e = 0 in (16).
Therefore (27) leads to J⋆ = L and J
J⋆ =
4
C. The PIM increases as M increases
The vector z employs M statistics, and to stress that we
−1
write LM = D⊤
M ΣM DM .
Theorem 3. Including more statistics in z can never worsen
the bound, i.e.,
0 Ln · · · LM LM+1 J.
(28)
Proof: Suppose we extend the vector z with an (M +1)th
statistic so that we can write
ΣM c
DM
ΣM+1 =
and
D
=
.
(29)
M+1
c⊤ κ
d⊤
Then calculating Σ−1
M+1 using [15, Lemma A.2], we obtain
⊤
ΣM c −1 DM
LM+1 = DM d
c⊤ κ
d⊤
I
= D⊤
d
Σ−1
M
M I 0
0
⊤ −1
DM
1
−Σ−1
M c
+
−c ΣM 1
d⊤
1
κ − c⊤ Σ−1
M c
= LM +
⊤ −1 ⊤
(d − D⊤ Σ−1
M c)(d − D ΣM c)
κ − c⊤ Σ−1
M c
LM .
An interesting research problem is to study the limit of LM
as M → ∞. Under what conditions will LM converge to J?
V. PIM AND G ENERALIZED METHOD OF MOMENTS
b exists if and only if the
An efficient unbiased estimator θ
following identity holds [3], [4]
∂ ln pθ
b − θ),
= J(θ
(30)
∂θ
which is satisfied only in rare cases. In more general scenarios,
the maximum likelihood estimator
b = arg min − ln p(y; θ).
θ
(31)
θ
is (asymptotically) unbiased with (asymptotic) covariance mab = J−1 under certain regularity conditions. It is
trix Cov[θ]
thus asymptotically efficient. Given an appropriate initializab0 , (31) can be solved iteratively using the Newtontion point θ
based scoring method:
bi+1 = θ
bi − J−1 ∂ − ln pθ
θ
∂θ
bi
θ=θ
∂ − ln pb⋆
bi+1 = θ
bi − J
b−1
θ
⋆
∂θ
bi
θ=θ
.
(32)
pθ
are unknown, we may use instead
When J and ∂ ln
∂θ
the best misspecified information matrix (15) and score (22),
in the sense of Theorem 2. Then the so-obtained scoring
method is related to a method that minimizes a certain class
of cost functions V (θ) which we characterize in what follows.
Consider evaluating the misspecified variables (15) and (22)
b instead of Σ.
obtained using (27) and a consistent estimate Σ
Then the corresponding scoring method, analogous to (32),
can be expressed as
.,
(33)
where we define (according to the above discussion about (15),
(22) and (27))
∂ ln pb⋆
b −1 (z − µ(θ))
= D⊤ (θ)Σ
∂θ
b⋆ = D⊤ (θ)Σ
b −1 D(θ).
J
(34)
One can verify that (33) is a scoring method for solving the
following problem
b = arg min 1 (z − µ(θ))⊤ Σ
b −1 (z − µ(θ)),
θ
2
θ
|
{z
}
(35)
,V (θ)
p
b⋆
b⋆ is an estimate of
by noting that ∂θ V (θ) = − ∂ ln
and that J
∂θ
2
the Hessian ∂θ V (θ). Eq. (35) is recognized as a generalized
method of moments, using an asymptotically optimal weight
b −1 [9], [15].
matrix Σ
The cost function V (θ) can be characterized around its
minimum, as follows
b ≃ ∂θ V (θ) + ∂ 2 V (θ)(θ
b − θ),
0 = ∂θ V (θ)
θ
(36)
where the right-hand side is a Taylor expansion. Using properb and obtain the following
ties of (34) in (36), we can solve for θ
approximation
b ≃ θ + (D⊤ Σ
b −1 D)−1 D⊤ Σ
b −1 (z − µ).
θ
Since the unknown distribution pθ satisfies (19), it follows that
b ≃ θ and
E[θ]
b ≃ (D⊤ Σ−1 D)−1 = L−1 J−1 .
Cov[θ]
The above expressions hold asymptotically as the number of
samples N in y increases [15].
In summary, using a scoring method analogous to (32) leads
to the generalized method of moments (35) with asymptotic
covariance given by the inverse PIM.
VI. C ONCLUSIONS
We have provided a direct, algebraic derivation of a tractable
lower bound on the Fisher information matrix which we
called the Pearson information matrix (for reasons explained
above). Furthermore, we presented an information-theoretic
link between the PIM and misspecified data distributions as
well as a connection to the generalized method of moments.
R EFERENCES
[1] C. Rao, “Information and accuracy attainable in the estimation of
statistical parameters,” Bulletin of the Calcutta Mathematical Society,
vol. 37, no. 3, pp. 81–91, 1945.
[2] H. Cramér, “A contribution to the theory of statistical estimation,”
Scandinavian Actuarial Journal, vol. 1946, no. 1, pp. 85–94, 1946.
[3] H. Van Trees and K. Bell, Detection Estimation and Modulation Theory,
Pt.I. Detection Estimation and Modulation Theory, Wiley, second ed.,
2013 [1968].
[4] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
theory. Prentice Hall, 1993.
[5] P. Stoica and P. Babu, “The Gaussian data assumption leads to the
largest Cramér-Rao bound,” IEEE Signal Processing Mag., vol. 28, no. 3,
pp. 132–133, 2011.
[6] S. Park, E. Serpedin, and K. Qaraqe, “Gaussian assumption: The least
favorable but the most useful,” IEEE Signal Processing Mag., vol. 30,
no. 3, pp. 183–186, 2013.
5
[7] M. Stein, A. Mezghani, and J. Nossek, “A lower bound for the Fisher
information measure,” IEEE Signal Processing Letters, vol. 21, pp. 796–
799, July 2014.
[8] K. Pearson, “Contributions to the mathematical theory of evolution,”
Philosophical Transactions of the Royal Society of London. A, vol. 185,
pp. 71–110, 1894.
[9] L. P. Hansen, “Large sample properties of generalized method of moments estimators,” Econometrica: Journal of the Econometric Society,
pp. 1029–1054, 1982.
[10] M. Stein, J. A. Nossek, and K. Barbé, “Fisher information bounds with
applications in nonlinear learning, compression and inference,” arXiv
preprint arXiv:1512.03473, 2015.
[11] E. T. Jaynes, “Information theory and statistical mechanics,” Physical
review, vol. 106, no. 4, p. 620, 1957.
[12] R. Horn and C. Johnson, Matrix Analysis. Cambridge University Press,
1990.
[13] S. Kullback, Information Theory and Statistics. Dover, 1997 [1959].
[14] T. M. Cover and J. A. Thomas, Elements of Information Theory. John
Wiley & Sons, 2006.
[15] T. Söderström and P. Stoica, System identification. Prentice-Hall, Inc.,
1988.
| 10math.ST
|
Published as a conference paper at ICLR 2018
OF GRID - LIKE REPRESENTATIONS BY
TRAINING RECURRENT NEURAL NETWORKS TO
PERFORM SPATIAL LOCALIZATION
arXiv:1803.07770v1 [q-bio.NC] 21 Mar 2018
E MERGENCE
Christopher J. Cueva∗, Xue-Xin Wei∗
Columbia University
New York, NY 10027, USA
{ccueva,weixxpku}@gmail.com
A BSTRACT
Decades of research on the neural code underlying spatial navigation have revealed a diverse set of neural response properties. The Entorhinal Cortex (EC) of
the mammalian brain contains a rich set of spatial correlates, including grid cells
which encode space using tessellating patterns. However, the mechanisms and
functional significance of these spatial representations remain largely mysterious.
As a new way to understand these neural representations, we trained recurrent
neural networks (RNNs) to perform navigation tasks in 2D arenas based on velocity inputs. Surprisingly, we find that grid-like spatial response patterns emerge in
trained networks, along with units that exhibit other spatial correlates, including
border cells and band-like cells. All these different functional types of neurons
have been observed experimentally. The order of the emergence of grid-like and
border cells is also consistent with observations from developmental studies. Together, our results suggest that grid cells, border cells and others as observed in EC
may be a natural solution for representing space efficiently given the predominant
recurrent connections in the neural circuits.
1
I NTRODUCTION
Understanding the neural code in the brain has long been driven by studying feed-forward architectures, starting from Hubel and Wiesel’s famous proposal on the origin of orientation selectivity in
primary visual cortex (Hubel & Wiesel, 1962). Inspired by the recent development in deep learning (Krizhevsky et al., 2012; LeCun et al., 2015; Hochreiter & Schmidhuber, 1997; Mnih et al.,
2015), there has been a burst of interest in applying deep feedforward models, in particular convolutional neural networks (CNN) (LeCun et al., 1998), to study the sensory systems, which hierarchically extract useful features from sensory inputs (see e.g., Yamins et al. (2014); Kriegeskorte (2015);
Kietzmann et al. (2017); Yamins & DiCarlo (2016)).
For more cognitive tasks, neural systems often need to maintain certain internal representations
of relevant variables in the absence of external stimuli- a process that requires more than feature
extraction. We will focus on spatial navigation, which typically requires the brain to maintain a
representation of self-location and update it according to the animal’s movements and landmarks
of the environment. Physiological studies done in rodents and other mammals (including humans,
non-human primates and bats) have revealed a variety of neural correlates of space in Hippocampus
and Entorhinal Cortex (EC), including place cells (O’Keefe, 1976), grid cells (Fyhn et al., 2004;
Hafting et al., 2005; Fyhn et al., 2008; Yartsev et al., 2011; Killian et al., 2012; Jacobs et al., 2013),
along with border cells (Solstad et al., 2008), band-like cells (Krupic et al., 2012) and others (see
Figure 1a). In particular, each grid cell only fires when the animal occupies a distinct set of physical
locations, and strikingly these locations lie on a lattice. The study of the neural underpinning of
spatial cognition has provided an important window into how high-level cognitive functions are
supported in the brain (Moser et al., 2008; Aronov et al., 2017).
∗
equal contribution
1
Published as a conference paper at ICLR 2018
How might the spatial navigation task be solved using a network of neurons? Recurrent neural networks (RNNs) (Hochreiter & Schmidhuber, 1997; Graves et al., 2013; Oord et al., 2016; Theis &
Bethge, 2015; Gregor et al., 2015; Sussillo et al., 2015) seem particularly useful for these tasks.
Indeed, recurrent-based continuous attractor networks have been one popular type of models proposed for the formation of grid cells (McNaughton et al., 2006; Burak & Fiete, 2009; Couey et al.,
2013) and place cells (Samsonovich & McNaughton, 1997). Such models have provided valuable
insights into one set of possible mechanisms that could support the formation of the grids. However, these models typically rely on fine-tuned connectivity patterns, in particular the models need a
subtle yet systematic asymmetry in the connectivity pattern to move the attractor state according to
the animal’s own movement. The existence of such a specific 2D connectivity in rodent EC remains
unclear. Additionally, previous models have mainly focused on grid cells, while other types of responses that co-exist in the Entorhinal Cortex have been largely ignored. It would be useful to have
a unified model that can simultaneously explain different types of neural responses in EC.
Motivated by these considerations, here we present an alternative modeling approach for understanding the representation of space in the neural system. Specifically, we trained a RNN to perform
some spatial navigation tasks. By leveraging the recent development in RNN training and knowledge of the navigation system in the brain, we show that training a RNN with biologically relevant
constraints naturally gives rise to a variety of spatial response profiles as observed in EC, including
grid-like responses. To our knowledge, this is the first study to show that grid-like responses could
emerge from training a RNN to perform navigation.
Our result implies that the neural representation in EC may be seen as a natural way for the brain to
solve the navigation task efficiently (Wei et al., 2015). More generally, it suggests that RNNs can be
a powerful tool for understanding the neural mechanisms of certain high-level cognitive functions.
firing rate
a
grid cell
b
band cell
border cell
irregular cell
RNN
c
Input
Output
speed
x-position
direction
speed
Performance
Output: target
Output: RNN
direction
y-position
Figure 1: a) Example neural data showing different kinds of neural correlates underlying spatial navigation in EC. All figures are replotted from previous publications. From left to right: a “grid cell”
recorded when an animal navigates in a square environment, replotted from Krupic et al. (2012),
with the heat map representing the firing rate of this neuron as a function of the animal’s location
(red corresponds to high firing rate); a “band-like” cell from Krupic et al. (2012); a border cell from
Solstad et al. (2008); an irregular spatially tuned cell from Diehl et al. (2017); a “speed cell” from
Kropff et al. (2015), which exhibits roughly linear dependence on the rodent’s running speed; a
“heading direction cell” from Sargolini et al. (2006), which shows systematic change of firing rate
depending on animal’s heading direction. b) The network consists of N = 100 recurrently connected units (or neurons) which receive two external inputs, representing the animal’s speed and
heading direction. The two outputs linearly weight the neurons in the RNN. The goal of training is
to make the responses of the two output neurons accurately represent the animal’s physical location.
c) Typical trajectory after training. As shown, the output of the RNN can accurately, though not
perfectly, track the animal’s location during navigation.
2
Published as a conference paper at ICLR 2018
2
2.1
M ODEL
M ODEL DESCRIPTION
Our network model consists of a set of recurrently connected units (N = 100). The dynamics of
each unit in the network ui (t) is governed by the standard continuous-time RNN equation:
τ
Nin
N
X
X
dxi (t)
in
= −xi (t) +
Wijrec uj (t) +
Wik
Ik (t) + bi + ξi (t)
dt
j=1
(1)
k=1
for i = 1, . . . , N . The activity of each unit, ui (t), is related to the activation of that unit, xi (t),
through a nonlinearity which in this study we take to be ui (t) = tanh(xi (t)). Each unit receives
input from other units through the recurrent weight matrix W rec and also receives external input,
I(t), that enters the network through the weight matrix W in . Each unit has two sources of bias,
bi which is learned and ξi (t) which represents noise intrinsic to the network and is taken to be
Gaussian with zero mean and constant variance. The network was simulated using the Euler method
for T = 500 timesteps of duration τ /10.
To perform a 2D navigation task with the RNN, we linearly combine the firing rates of units in
the network to estimate the current location of the animal. The responses of the two linear readout
neurons, y1 (t) and y2 (t), are given by the following equation:
yj (t) =
N
X
out
Wji
ui (t)
(2)
i=1
2.2
I NPUT TO THE NETWORK
The network inputs and outputs were inspired by simple spatial navigation tasks in 2D open environments. The task resembles dead-reckoning (sometimes referred to as path integration), which is
ethologically relevant for many animal species (Darwin, 1873; Mittelstaedt & Mittelstaedt, 1980;
Etienne & Jeffery, 2004; McNaughton et al., 2006). To be more specific, the inputs to the network
were the animal’s speed and direction at each time step. Experimentally, it has been shown that the
velocity signals exist in EC (Sargolini et al., 2006; Kropff et al., 2015; Hinman et al., 2016), and
there is also evidence that such signals are necessary for grid formation (Winter et al., 2015a;b).
Throughout the paper, we adopt the common assumption that the head direction of the animal coincides with the actual moving direction. The outputs were the x- and y-coordinates of the integrated
position. The direction of the animal is modeled by modified Brownian motion to increase the
probability of straight-runs, in order to be consistent with the typical rodent’s behavior in an open
environment. The usage of such simple movement statistics has the advantage of having full control
of the simulated trajectories. However, for future work it would be very interesting to test the model
using different animals’ real movement trajectories to see how the results might change.
Special care is taken when the animal is close to the boundary. The boundary of the environment
will affect the statistics of the movement, as the animal cannot cross the boundary. This fact was
reflected in the model by re-sampling the angular input variable until the input angle did not lead
the animal outside the boundary. In the simulations shown below, the animal always starts from the
center of the arena, but we verified that the results are insensitive to the starting locations.
2.3
T RAINING
We optimized the network parameters W rec , W in , b and W out to minimize the squared error in
equation (3) between target x- and y-coordinates from a two dimensional navigation task (performed
in rectangular, hexagonal, and triangular arenas) and the network outputs generated according to
equation (2).
M,T,N
Xout
1
E=
(yj (t, m) − yjtarget (t, m))2
(3)
M T Nout m,t,j=1
Parameters were updated with the Hessian-free algorithm (Martens & Sutskever, 2011) using minibatches of size M = 500 trials. In addition to minimizing the error function in equation (3) we
3
Published as a conference paper at ICLR 2018
regularized the input and output weights according to equation (4) and the squared firing rates of the
units (referred to as metabolic cost) according to equation (5). In sum, the training aims to minimize
a loss function, that consists of the error of the animal, the metabolic cost, and a penalty for large
network parameters.
RL2
N,N
NX
out ,N
Xin
1
1
in 2
=
(W ) +
(Wijout )2
N Nin i,j=1 ij
N Nout i,j=1
RF R =
1
NTM
N,T,M
X
ui (t, m)2
(4)
(5)
i,t,m=1
We find that the results are qualitatively insensitive to the initialization schemes used for the recurrent weight matrix W rec . For the results presented in this paper, simulations in the hexagonal
environment were obtained by initializing the elements of W rec to be zero mean Gaussian random
variables with variance 1.52 /N , and simulations in the square and triangular environments were
initialized with an orthogonal W rec (Saxe et al., 2014). We initialized the bias b and output weights
W out to be zero. The elements of W in were zero mean Gaussian variables with variance 1/Nin .
a
grid-like
b
band-like
c
border
d
irregular
Figure 2: Different types of spatial selective responses of units in the trained RNN. Example simulation results for three different environments (square, triangular, hexagon) are presented. Blue
(yellow) represents low (high) activity. a) Grid-like responses. b) Band-like responses; c) Borderrelated responses; d) Spatially irregular responses. These responses can be spatially selective but
they do not form a regular pattern defined in the conventional sense.
3
R ESULTS
We run simulation experiments in arenas with different boundary shapes, including square, triangular and hexagonal. Figure 1c shows a typical example of the model performance after training; the
network (red trace) accurately tracks the animal’s actual path (black).
3.1
T UNING PROPERTIES OF THE MODEL NEURONS
We are mostly interested in what kind of representation the RNN has learned to solve this navigation
task, and whether such a representation resembles the response properties of neurons in EC (Moser
et al., 2008).
3.1.1
S PATIAL TUNING
To test whether the trained RNN developed location-selective representations, we plot individual
neurons’ mean activity level as a function of the animal’s location during spatial exploration. Note
4
Published as a conference paper at ICLR 2018
that these average response profiles should not be confused with the linear filters typically shown in
feedforward networks. Surprisingly, we find neurons in the trained RNN show a range of interesting
spatial response profiles. Examination of these response profiles suggests they can be classified into
distinct functional types. Importantly, as we will show, these distinct spatial response profiles can be
mapped naturally to known physiology in EC. The spatial responses of all units in trained networks
are shown in the Appendix.
Grid-like responses Most interestingly, we find some of the units in the RNN exhibit clear grid-like
responses (Figure 2a). These firing patterns typically exhibit multiple firing fields, with each firing
field exhibiting roughly circular symmetric or ellipse shape. Furthermore, the firing fields are highly
structured, i.e., when combined, are arranged on a regular lattice. Furthermore, the structure of the
response lattice depends on the shape of the boundary. In particular, training the network to perform
self-localization in a square environment tends to give rectangular grids. In hexagonal and triangular
environments, the grids are closer to triangular.
Experimentally, it is shown that (medial) EC contains so-called grid cells which exhibit multiple
firing fields that lie on a regular grid (Fyhn et al., 2004; Hafting et al., 2005). The grid-like firing
patterns in our simulation are reminiscent of the grid cells in rodents and other mammals. However,
we also notice that the the grid-like model responses typically exhibit few periods, not as many
as experimental data (see Figure 1a). It is possible that using a larger network might reveal finer
grid-patterns in our model. Nonetheless, it is surprising that the gird-like spatial representations can
develop in our model, given there is no periodicity in the input. Another potential concern is that,
experimentally it is reported that the grids are often on the corners of a triangular lattice (Hafting
et al., 2005) even in square environments (see Figure 1a), though the grids are somewhat influenced
by the shape of the environment. However, the rats in these experiments presumable had spatial
experience in other environments with various boundary shapes. Experimentally, it would be interesting to see if grid cells would lie on a square lattice instead if the rats are raised in a single square
environment - a situation we are simulating here.
Border responses Many neurons in the RNN exhibit selectivity to the boundary (Figure 2c). Typically, they only encode a portion of the boundary, e.g. one piece of wall in a square shaped environment. Such properties are similar to the border cells discovered in rodent EC (Solstad et al., 2008;
Savelli et al., 2008; Lever et al., 2009). Experimentally, border cells mainly fire along one piece of
wall, although some have been observed to fire along multiple borders or along the whole boundary of the environment; interestingly, these multi-border responses were also observed in some RNN
models. Currently, it is unclear how the boundary-like response profiles emerge (Solstad et al., 2008;
Savelli et al., 2008; Lever et al., 2009). Our model points to the possibility that the border cells may
emerge without the presence of tactile cues. Furthermore, it suggests that border cell formation may
be related to the movement statistics of the animals, i.e. due to the asymmetry of the movement
statistics along the boundary.
Band-like responses Interestingly, some neurons in the RNN exhibit band-like responses (Figure
2b). In most of our simulations, these bands tend to be parallel to one of the boundaries. For
some of the units, one of the bands overlaps the boundary, but for others, that is not the case.
Experimentally, neurons with periodic-like firing patterns have been recently reported in rodent EC.
In one study, it has been reported that a substantial portion of cells in EC exhibit band-like firing
characteristics (Krupic et al., 2012). However, we note that based on the reported data in Krupic
et al. (2012), the band pattern is not as clear as in our model.
Spatially-stable but non-regular responses Besides the units described above, most of the remaining units also exhibit stable spatial responses, but they do not belong to the above categories. These
response profiles can exhibit either one large irregular firing field; or multiple circular firing fields,
but these firing fields do not show a regular pattern. Experimentally these types of cells have also
been observed. In fact, it is recently reported that the non-grid spatial cells constitute a large portion
of the neurons in Layer II and III of rodent EC (Diehl et al., 2017).
3.1.2
S PEED TUNING AND HEAD DIRECTION TUNING
Speed tuning We next ask how neurons in the RNN are tuned to the inputs. Many of the model
neurons exhibit linear responses to the running speed of the animal, while some neurons show no
selectivity to speed, as suggested by the near-flat response functions. Example response profiles are
5
Published as a conference paper at ICLR 2018
a
d
g
b
e
h
c
f
i
direction
speed
direction
speed
direction
speed
Figure 3: Direction tuning and speed tuning for nine example units in an RNN trained in a triangular
arena. For each unit, we show the spatial tuning, (head) directional tuning, speed tuning respectively,
from left to right. a,b,c) The three model neurons show strong directional tuning, but the spatial
tuning is weak and irregular. The three neurons also exhibit linear speed tuning. d,e,f) The three
neurons exhibit grid-like firing patterns, and clear speed tuning. The strength of their direction
tuning differ. g,h) Border cells exhibit weak and a bit complex directional tuning and almost no
speed tuning. i) This band cell shows weak directional tuning, but strong speed tuning.
shown in Figure 3. Interestingly, we observe that the model border cells tend to have almost zero
speed-tuning (e.g., see Figure 3g,h).
Head direction tuning A substantial portion of the model neurons show direction tuning. There
are a diversity of direction tuning profiles, both in terms of the strength of the tuning and their preferred direction. Example tuning curves are shown in Figure 3, and the direction tuning curves of
a complete population are shown in the Appendix. Interestingly, in general model neurons which
show the strongest head direction tuning do not show a clear spatial firing pattern (see Figure 3a,b,c).
This suggests that there are a group of neurons which are mostly responsible for encoding the direction. We also notice that neurons with clear grid-like firing can exhibit a variety of direction tuning
strengths, from weak to strong (Figure 3d,e,f). In the Appendix, we quantify the relation between
these different tuning properties at the whole population level, which show somewhat complex dependence.
Experimentally, the heading direction tuning in EC is well-known, e.g., Sargolini et al. (2006). Both
the grid and non-grid cells in EC exhibit head direction tuning (Sargolini et al., 2006). Furthermore,
the linear speed dependence of the model neurons is similar to the properties of speed cells reported
recently in EC (Kropff et al., 2015). Our result is also consistent with another recent study reporting
that the majority of neurons in EC exhibit some amount of speed tuning (Hinman et al., 2016).
It remains an open question experimentally, at a population level, how different types of tuning
characteristics in EC relate to each other.
3.1.3
D EVELOPMENT OF THE TUNING PROPERTIES
We next investigate how the spatial response profiles evolve as learning/training progresses. We report two main observations. First, neurons that fire selectively along the boundary typically emerge
first. Second, the grid-like responses with finer spatial tuning patterns only emerge later in training. For visualization, we perform dimensionality reduction using the t-SNE algorithm (Maaten &
Hinton, 2008). This algorithm embeds 100 model neurons during three phases of training (early,
intermediate, and late) into a two-dimensional space according to the similarity of their temporal
responses. Here the similarity metric is taken to be firing rate correlation. In this 2D space as shown
in Figure 4a, border cell representations appear early and stably persist through the end of training.
Furthermore, early during training all responses are similar to the border related responses. In contrast, grid-like cells typically undergo a substantial change in firing pattern during training before
settling into their final grid-like representation (Figure 4b).
The developmental time line of the grid-like cells and border cells is roughly consistent with developmental studies in rodents. Experimentally, it is known that border cells emerge earlier in development, and they exist at about 2 weeks after the rat is born (Bjerknes et al., 2014). The grid
6
Published as a conference paper at ICLR 2018
a
Dimension 2
cells mature only at about 4 weeks after birth (Langston et al., 2010; Wills et al., 2010; Bjerknes
et al., 2014). Furthermore, our simulations suggest the reason why border cells emerge earlier in
development may be that computationally it is easier to wire-up a network that gives rise to border
cell responses.
b
0
-10
Early
Training iteration 100
-20
Intermediate
Training
500
-20 iteration
-10
Late
Training
iteration1000
0
Dimension 1
10
Figure 4: Development of border cells and grid-like cells. Early during training all responses are
similar to the border related responses, and only as training continues do the grid-like cells emerge.
We perform dimensionality reduction using the t-SNE algorithm on the firing rates of the neurons. Each dot represents one neuron (N = 100), and the color represents different training stages
(early/intermediate/late shown in blue/cyan/yellow). Each line shows the trajectory of a single highlighted neuron as its firing responses evolve during training. In panel a), we highlight the border
representation. It appears there are four clusters of border cells, each responding to one wall of a
square environment (spatial responses from four of these border cells are inset). These cells’ response profiles appear early and stably persist through training, illustrated by the short distance they
travel in this space. In b), we show that the neurons which eventually become grid cells initially have
tuning profiles similar to the border cells but then change their tuning substantially during learning.
As a natural consequence, they need to travel a long distance in this space between the early and late
phase of the training. Spatial responses are shown for four of these grid-like cells during the late
phase of training.
3.2
T HE IMPORTANCE OF REGULARIZATION
We find appropriate regularizations of the RNN to be crucial for the emergence of grid-like representations. We only observed grid-like representations when the network was encouraged to store
information while perturbed by noise. This was accomplished by setting the speed input to zero, e.g.
zero speed 90% of the time, and adding Gaussian noise to the network (ξi (t) in equation (1)); the
precise method for setting the speed input to zero and the value of the noise variance is not crucial
for our simulations to develop grid-like representations. The cost function which aims to capture
the penalization on the metabolic cost of the neural activity also acts as an important regularization. Our simulations show that the grid-like representation did not emerge without this metabolic
cost. In Figure 5, we show typical simulation results for a square environment, with and without
proper metabolic regularization. In the Appendix, we illustrate the effect of regularization further,
in particular the role of injecting noise into the RNN units.
Our results are consistent with the general notion on the importance of incorporating proper constraint for learning useful representations in neural networks (Bengio et al., 2013). Furthermore, it
suggests that, to learn a model with response properties similar to neural systems it may be necessary
to incorporate the relevant constraints, e.g., noise and metabolic cost.
3.3
E RROR CORRECTION AROUND THE BOUNDARY
One natural question is whether the trained RNNs are able to perform localization when the path
length exceeds the typical length used during training (500 steps), in particular given that noise in
7
Published as a conference paper at ICLR 2018
a
b
Figure 5: Complete set of spatial response profiles for 100 neurons in a RNN trained in a square
environment. a) Without proper regularization, complex and periodic spatial response patterns do
not emerge. b) With proper regularization, a rich set of periodic response patterns emerge, including
grid-like responses. Regularization can also be adjusted to achieve spatial profiles intermediate
between these two examples.
0.02
no
boundary interaction
boundary interaction
b
Squared error
Squared error
a
0.016
0.012
0 boundary interaction
1 boundary interaction
2 to 5 interactions
6 to 50 interactions
0.4
0
(in the next 50 time steps)
0.008
RNN output
constant output
0.8
-40
-20
0
20
40
Timesteps relative to first boundary interaction
0
2
4
6
Timesteps
8
10
10
5
Figure 6: Error-correction happens at the boundary and the error is stable over time. At the boundary,
the direction is resampled to avoid input velocities that lead to a path extending beyond the boundary
of the environment. These changing input statistics at the boundary, termed a boundary interaction,
are the only cue the RNN receives about the boundary. We find that the RNN uses the boundary
interactions to correct the accumulated error between the true integrated input and its prediction
based on the linear readout of equation (2). Panel a), the mean squared error increases when there
are no boundary interactions, but then decreases after a boundary interaction, with more boundary
interactions leading to greater error reduction. In the absence of further boundary interaction, the
squared error would gradually increase again (blue curve) at roughly a constant rate. b) The network
was trained using mini-batches of 500 timesteps but has stable error over a duration at least four
orders of magnitude larger. The error of the RNN output (mean and standard deviation shown in
black, computed based on 10000 timesteps) is compared to the error that would be achieved by an
RNN outputting the best constant values (red).
the network would gradually accumulate, leading to a decrease in localization performance. We test
this by simulating paths that are several orders of magnitude longer. Somewhat surprisingly, we find
the RNNs still perform well (Figure 6b). In fact, the squared error (averaged over every 10000 steps)
is stable. The spatial response profiles of individual units also remain stable. This implies that the
RNNs have acquired intrinsic error-correction mechanisms during training.
As shown earlier, during training some of the RNN units develop boundary-related firing (Figure 2c),
presumably by exploiting the change of input statistics around the boundary. We hypothesize that
8
Published as a conference paper at ICLR 2018
boundary interactions may enable error-correction through signals based on these boundary-related
activities. Indeed, we find that boundary interactions can dramatically reduce the accumulated error (Figure 6a). Figure 6a shows that, without boundary interactions, on average the squared error
grows roughly linearly as expected, however, interactions with the boundaries substantially reduce
the error, and more frequent boundary interactions can reduce the error further. Error-correction on
grid cells via boundary interactions has been proposed (Hardcastle et al., 2015; Pollock et al., 2017),
however, we emphasize that the model proposed here develops the grid-like responses, boundary
responses and the error-correction mechanisms all within the same neural network, thus potentially
providing a unifying account of a diverse set of phenomena.
4
D ISCUSSION
In this paper, we trained RNNs to perform path integration (dead-reckoning) in 2D arenas. We
found that after training RNNs with appropriate regularization, the model neurons exhibit a variety
of spatial and velocity tuning profiles that match neurophysiology in EC. What’s more, there is
also similarity in terms of when these distinct neuron types emerge during training/development.
The EC has long been thought to be involved in path integration and localization of the animal’s
location (Moser et al., 2008). The general agreement between the different response properties
in our model and the neurophysiology provide strong evidence supporting the hypothesis that the
neural population in EC may provide an efficient code for representation self-locations based on the
velocity input.
Recently, there has been increased interest in using complex neural network models to understand
the neural code. But the focus has been on using feedforward architectures, in particular CNNs (LeCun et al., 1998). Given the abundant recurrent connections in the brain, it seems a particularly
fruitful avenue to take advantage of the recent development in RNNs to help with neuroscience
questions (Mante et al., 2013; Song et al., 2016; Miconi, 2017; Sussillo et al., 2015). Here, we only
show one instance following this approach. However, the insight from this work could be general,
and potentially useful for other cognitive functions as well.
The finding that metabolic constraints lead to the emergence of grid-like responses may be seen
as conceptually related to the efficient coding hypothesis in visual processing (Barlow, 1961), in
particular the seminal work on the emergence of the V1-like Gabor filters in a sparse coding model
by Olshausen & Field (1996). Indeed, our work is partly inspired by these results. While there
are conceptual similarities, however, we should also note there are differences between the sparse
coding work and ours. First, the sparsity constraint in sparse coding can be naturally viewed as
a particular prior while in the context of the recurrent network, it is difficult to interpret that way.
Second, the grid-like responses are not the most sparse solution one could imagine. In fact, they are
still quite dense compared to a more spatially localized representation. Third, the grid-like patterns
that emerged in our network are not filters based on the raw input, rather the velocity inputs need to
be integrated first in order to encode spatial locations. Our work is also inspired by recent work using
the efficient coding idea to explain the functional architecture of the grid cells (Wei et al., 2015). It
has been shown that efficient coding considerations could explain the particular set of grid scales
observed in rodents (Stensola et al., 2012). However, in that work, the firing patterns of the neurons
are assumed to have a lattice structure to start with. Furthermore, our work is related to the study by
Sussillo and others (Sussillo et al., 2015), in which they show that regularization of RNN models are
important for generating solutions that are similar to the neural activity observed in motor cortex.
In Sussillo et al., a smoothness constraint together with others lead to simple oscillatory neural
dynamics that well matches the neural data. We have not incorporated a smoothness constraint into
our network.
Additionally, we note that there are a few recent studies which use place cells as the input to generate
grid cells (Dordek et al., 2016; Stachenfeld et al., 2016), which are fundamentally different from
our work. In these feedforward network models, the grid cells essentially perform dimensionality
reduction based on the spatial input from place cells. However, the main issue with these models
is that, it is unclear how place cells acquire spatial tuning in the first place. To the contrary, our
model takes the animal’s velocity as the input, and addresses the question of how the spatial tuning
can be generated from such input, which are known to exist in EC (Sargolini et al., 2006; Kropff
et al., 2015). In another related study (Kanitscheider & Fiete, 2016), the authors train a RNN with
9
Published as a conference paper at ICLR 2018
LSTM units (Hochreiter & Schmidhuber, 1997) to perform different navigation tasks. However, no
grid-like spatial firing patterns are reported.
Although our model shows a qualitative match to the neural responses observed in the EC, nonetheless it has several major limitations, with each offering interesting future research directions. First,
the learning rule we use seems to be biologically implausible. We are interested in exploring how a
more biologically plausible learning rule could give rise to similar results (Lillicrap et al., 2016; Miconi, 2017; Guerguiev et al., 2017). Second, the simulation results do not show a variety of spatial
scales in grid-like cells. Experimentally, it is known that grid cells have multiple spatial scales, that
scale geometrically with a ratio 1.4 (Stensola et al., 2012), and this particular scale ratio is predicted
by efficient coding of space (Wei et al., 2015). We are investigating how to modify the model to
get a hierarchy of spatial scales, perhaps by incorporating more neurons or modifying the regularization. Last but not least, we have focused on the representation produced by the trained RNN.
An equally important set of questions concern how the networks actually support the generation of
such a representation. As a preliminary effort, we have examined the connectivity patterns of the
trained network, and they do not seem to resemble the connectivity patterns required by standard
attractor network models. Maybe this should not be seen as too surprising. After all, the trained
networks can produce a diverse set of neural responses, while the previous models only led to grid
responses. It would be interesting for future work to systematically examine the questions related to
the underlying mechanisms.
5
ACKNOWLEDGEMENTS
We thank members of the Center for Theoretical Neuroscience at Columbia University for useful
discussions and three anonymous reviewers for constructive feedback. Research supported by NSF
NeuroNex Award DBI-1707398 and NIH training grant 5T32NS064929 (CJC).
R EFERENCES
Dmitriy Aronov, Rhino Nevers, and David W Tank. Mapping of a non-spatial dimension by the
hippocampalentorhinal circuit. Nature, 2017.
Horace B Barlow. Possible principles underlying the transformation of sensory messages. Sensory
communication, pp. 217–234, 1961.
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new
perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798–1828,
2013.
Tale L Bjerknes, Edvard I Moser, and May-Britt Moser. Representation of geometric borders in the
developing rat. Neuron, 82(1):71–78, 2014.
Yoram Burak and Ila R Fiete. Accurate path integration in continuous attractor network models of
grid cells. PLoS computational biology, 5(2):e1000291, 2009.
Jonathan J Couey, Aree Witoelar, Sheng-Jia Zhang, Kang Zheng, Jing Ye, Benjamin Dunn, Rafal
Czajkowski, May-Britt Moser, Edvard I Moser, Yasser Roudi, et al. Recurrent inhibitory circuitry
as a mechanism for grid formation. Nature neuroscience, 16(3):318–324, 2013.
Charles Darwin. Origin of certain instincts. Nature, 7:417–418, 1873.
Geoffrey W Diehl, Olivia J Hon, Stefan Leutgeb, and Jill K Leutgeb. Grid and nongrid cells in medial entorhinal cortex represent spatial location and environmental features with complementary
coding schemes. Neuron, 94(1):83–92, 2017.
Yedidyah Dordek, Daniel Soudry, Ron Meir, and Dori Derdikman. Extracting grid cell characteristics from place cell inputs using non-negative principal component analysis. eLife, 5:e10094,
2016.
Ariane S Etienne and Kathryn J Jeffery. Path integration in mammals. Hippocampus, 14(2):180–
192, 2004.
10
Published as a conference paper at ICLR 2018
Marianne Fyhn, Sturla Molden, Menno P Witter, Edvard I Moser, and May-Britt Moser. Spatial
representation in the entorhinal cortex. Science, 305(5688):1258–1264, 2004.
Marianne Fyhn, Torkel Hafting, Menno P Witter, Edvard I Moser, and May-Britt Moser. Grid cells
in mice. Hippocampus, 18(12):1230–1238, 2008.
Alex Graves, Abdel-Rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep recurrent neural networks. In Acoustics, speech and signal processing (icassp), 2013 ieee international
conference on, pp. 6645–6649. IEEE, 2013.
Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Jimenez Rezende, and Daan Wierstra. Draw: A
recurrent neural network for image generation. arXiv preprint arXiv:1502.04623, 2015.
Jordan Guerguiev, Timothy P Lillicrap, and Blake A Richards. Towards deep learning with segregated dendrites. eLife, 6, 2017.
Torkel Hafting, Marianne Fyhn, Sturla Molden, May-Britt Moser, and Edvard I Moser. Microstructure of a spatial map in the entorhinal cortex. Nature, 436(7052):801–806, 2005.
Kiah Hardcastle, Surya Ganguli, and Lisa M Giocomo. Environmental boundaries as an error correction mechanism for grid cells. Neuron, 86(3):827–839, 2015.
James R Hinman, Mark P Brandon, Jason R Climer, G William Chapman, and Michael E Hasselmo.
Multiple running speed signals in medial entorhinal cortex. Neuron, 91(3):666–679, 2016.
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
David H Hubel and Torsten N Wiesel. Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. The Journal of physiology, 160(1):106–154, 1962.
Joshua Jacobs, Christoph T Weidemann, Jonathan F Miller, Alec Solway, John F Burke, Xue-Xin
Wei, Nanthia Suthana, Michael R Sperling, Ashwini D Sharan, Itzhak Fried, et al. Direct recordings of grid-like neuronal activity in human spatial navigation. Nature neuroscience, 16(9):1188–
1190, 2013.
Ingmar Kanitscheider and Ila Fiete. Training recurrent networks to generate hypotheses about how
the brain solves hard navigation problems. arXiv preprint arXiv:1609.09059, 2016.
Tim Christian Kietzmann, Patrick McClure, and Nikolaus Kriegeskorte. Deep neural networks in
computational neuroscience. bioRxiv, pp. 133504, 2017.
Nathaniel J Killian, Michael J Jutras, and Elizabeth A Buffalo. A map of visual space in the primate
entorhinal cortex. Nature, 491(7426):761–764, 2012.
Nikolaus Kriegeskorte. Deep neural networks: a new framework for modeling biological vision and
brain information processing. Annual Review of Vision Science, 1:417–446, 2015.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,
2012.
Emilio Kropff, James E Carmichael, May-Britt Moser, and Edvard I Moser. Speed cells in the
medial entorhinal cortex. Nature, 523(7561):419–424, 2015.
Julija Krupic, Neil Burgess, and John O?Keefe. Neural representations of location composed of
spatially periodic bands. Science, 337(6096):853–857, 2012.
Rosamund F Langston, James A Ainge, Jonathan J Couey, Cathrin B Canto, Tale L Bjerknes,
Menno P Witter, Edvard I Moser, and May-Britt Moser. Development of the spatial representation system in the rat. Science, 328(5985):1576–1580, 2010.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
11
Published as a conference paper at ICLR 2018
Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444,
2015.
Colin Lever, Stephen Burton, Ali Jeewajee, John O’Keefe, and Neil Burgess. Boundary vector cells
in the subiculum of the hippocampal formation. The journal of neuroscience, 29(31):9771–9777,
2009.
Timothy P Lillicrap, Daniel Cownden, Douglas B Tweed, and Colin J Akerman. Random synaptic
feedback weights support error backpropagation for deep learning. Nature communications, 7,
2016.
Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine
Learning Research, 9(Nov):2579–2605, 2008.
Valerio Mante, David Sussillo, Krishna V Shenoy, and William T Newsome. Context-dependent
computation by recurrent dynamics in prefrontal cortex. Nature, 503(7474):78–84, 2013.
James Martens and Ilya Sutskever. Learning recurrent neural networks with hessian-free optimization. pp. 10331040, 2011.
Bruce L McNaughton, Francesco P Battaglia, Ole Jensen, Edvard I Moser, and May-Britt Moser.
Path integration and the neural basis of the ’cognitive map’. Nature Reviews Neuroscience, 7(8):
663–678, 2006.
Thomas Miconi. Biologically plausible learning in recurrent neural networks reproduces neural
dynamics observed during cognitive tasks. eLife, 6:e20899, 2017.
M-L Mittelstaedt and H Mittelstaedt.
senschaften, 67(11):566–567, 1980.
Homing by path integration in a mammal.
Naturwis-
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level
control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
Edvard I Moser, Emilio Kropff, and May-Britt Moser. Place cells, grid cells, and the brain’s spatial
representation system. Annu. Rev. Neurosci., 31:69–89, 2008.
John O’Keefe. Place units in the hippocampus of the freely moving rat. Experimental neurology, 51
(1):78–109, 1976.
Bruno A Olshausen and David J Field. Emergence of simple-cell receptive field properties by
learning a sparse code for natural images. Nature, 381(6583):607, 1996.
Aaron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks.
arXiv preprint arXiv:1601.06759, 2016.
Eli Pollock, Niral Desai, Xue-Xin Wei, and Vijay B Balasubramanian. A mechanism for selforganized error-correction of grid cells by border cells. CoSyNe abstract, 2017.
Alexei Samsonovich and Bruce L McNaughton. Path integration and cognitive mapping in a continuous attractor neural network model. Journal of Neuroscience, 17(15):5900–5920, 1997.
Francesca Sargolini, Marianne Fyhn, Torkel Hafting, Bruce L McNaughton, Menno P Witter, MayBritt Moser, and Edvard I Moser. Conjunctive representation of position, direction, and velocity
in entorhinal cortex. Science, 312(5774):758–762, 2006.
Francesco Savelli, D Yoganarasimha, and James J Knierim. Influence of boundary removal on the
spatial representations of the medial entorhinal cortex. Hippocampus, 18(12):1270, 2008.
Andrew M Saxe, James L McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. 2014.
Trygve Solstad, Charlotte N Boccara, Emilio Kropff, May-Britt Moser, and Edvard I Moser. Representation of geometric borders in the entorhinal cortex. Science, 322(5909):1865–1868, 2008.
12
Published as a conference paper at ICLR 2018
H Francis Song, Guangyu R Yang, and Xiao-Jing Wang. Training excitatory-inhibitory recurrent
neural networks for cognitive tasks: A simple and flexible framework. PLoS Comput Biol, 12(2):
e1004792, 2016.
Kimberly Lauren Stachenfeld, Matthew M Botvinick, and Samuel J Gershman. The hippocampus
as a predictive map. bioRxiv, pp. 097170, 2016.
Hanne Stensola, Tor Stensola, Trygve Solstad, Kristian Frøland, May-Britt Moser, and Edvard I
Moser. The entorhinal grid map is discretized. Nature, 492(7427):72–78, 2012.
David Sussillo, Mark M Churchland, Matthew T Kaufman, and Krishna V Shenoy. A neural network
that finds a naturalistic solution for the production of muscle activity. Nature neuroscience, 18(7):
1025–1033, 2015.
Lucas Theis and Matthias Bethge. Generative image modeling using spatial lstms. In Advances in
Neural Information Processing Systems, pp. 1927–1935, 2015.
Xue-Xin Wei, Jason Prentice, and Vijay Balasubramanian. A principle of economy predicts the
functional architecture of grid cells. Elife, 4:e08362, 2015.
B Willmore and DJ Tolhurst. Characterizing the sparseness of neural codes. Network, 12:255–270,
2001.
Tom J Wills, Francesca Cacucci, Neil Burgess, and John O’keefe. Development of the hippocampal
cognitive map in preweanling rats. Science, 328(5985):1573–1576, 2010.
Shawn S. Winter, Benjamin J. Clark, and Jeffrey S. Taube. Disruption of the head direction cell
network impairs the parahippocampal grid cell signal. Science, 347(6224):870–874, 2015a.
Shawn S. Winter, Max L. Mehlman, Benjamin J. Clark, and Jeffrey S. Taube. Passive transport
disrupts grid signals in the parahippocampal cortex. Current Biology, 25:2493–2502, 2015b.
Daniel LK Yamins and James J DiCarlo. Using goal-driven deep learning models to understand
sensory cortex. Nature neuroscience, 19(3):356–365, 2016.
Daniel LK Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon, Darren Seibert, and James J
DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual
cortex. Proceedings of the National Academy of Sciences, 111(23):8619–8624, 2014.
Michael M Yartsev, Menno P Witter, and Nachum Ulanovsky. Grid cells without theta oscillations
in the entorhinal cortex of bats. Nature, 479(7371):103–107, 2011.
13
Published as a conference paper at ICLR 2018
A
T RIANGULAR ENVIRONMENT
Noise 1, Metabolic 0
Noise 0, Metabolic 1
Noise and metabolic cost are important for grid-like representations. The figure on the left shows
the spatial responses for a network trained with noise and no metabolic cost. The figure on the right
shows the spatial responses for a network trained with no noise and the metabolic cost.
14
Published as a conference paper at ICLR 2018
Head direc&on tuning
Activity of unit (-1 to 1)
Ac&vity of unit (-1 to 1)
Training epoch 2900, 2000 trials, 500 time steps in simulation
errormain = 0.002977, normalized error overall = 1.3595%
Angular input (0 to 360 degrees)
Direc&on input (0 to 360 degrees)
Speed tuning
Activity of unit (-1 to 1)
Ac,vity of unit (-1 to 1)
Training epoch 2900, 2000 trials, 500 time steps in simulation
errormain = 0.002977, normalized error overall = 1.3595%
Speed input (0 to 0.199999)
Speed input
15
Published as a conference paper at ICLR 2018
R ECTANGULAR ENVIRONMENT
Head direc6on tuning
Activity of units (-1 to 1)
Ac6vity of unit (-1 to 1)
Training epoch 1000, 2000 trials, 450 time steps in simulation
errormain = 0.0047399, normalized error overall = 3.5752%
Angular input (0 to 360 degrees)
Angular input (0 to 360 degrees)
Speed tuning
Training epoch 1000, 2000 trials, 450 time steps in simulation
errormain = 0.0047399, normalized error overall = 3.5752%
Activity of unit (-1 to 1)
Ac,vity of unit (-1 to 1)
B
Speed input (0 to 0.2)
Speed input
16
Published as a conference paper at ICLR 2018
H EXAGONAL ENVIRONMENT
Head direc6on tuning
Activity of unit (-1 to 1)
Ac6vity of unit (-1 to 1)
Training epoch 900, 2000 trials, 450 time steps in simulation
errormain = 0.00403, normalized error overall = 4.2275%
Angular input (0 to 360 degrees)
Angular input (0 to 360 degrees)
Speed tuning
Activity of unit (-1 to 1)
Training epoch 900, 2000 trials, 450 time steps in simulation
errormain = 0.00403, normalized error overall = 4.2275%
Ac#vity of unit (-1 to 1)
C
Speed input (0 to 0.2)
Speed input
17
Published as a conference paper at ICLR 2018
D
R ELATION BETWEEN SPEED , DIRECTION ,
AND SPATIAL SELECTIVITY
To quantify the speed selectivity of each unit we first fit a line to the tuning curve of unit activity as
a function of speed. The speed selectivity is the absolute value of the slope. If the unit activity is not
modulated by speed then the speed selectivity is 0. To quantify the direction selectivity of each unit
we calculated the average unit activity as a function of direction input and then took the maximum
minus minimum of this tuning curve. If the unit activity is not modulated by direction then the
direction selectivity is 0. To quantify the spatial selectivity we used lifetime sparseness (Willmore
& Tolhurst, 2001). If the unit activity is not modulated by spatial location then the spatial selectivity
is 0. Each dot in the figures below show the selectivity for a single unit.
Rectangle
Hexagon
0.5
0.4
0.4
0.4
0.3
0.2
0.1
Spatial selectivity
0.5
Spatial selectivity
Spatial selectivity
Spa$al selec$vity
Triangle
0.5
0.3
0.2
0.1
0
0.3
0.2
0.1
0
0
0.5
1
1.5
2
0
0.5
Direction selectivity
1
1.5
0
2
0
Direction selectivity
0.5
Triangle
Rectangle
4
1
1.5
2
1.5
2
Direction selectivity
Direc$on selec$vity
Hexagon
4
3.5
2
1
Speed selectivity
Speed selectivity
Speed selectivity
Speed selec)vity
3
3
2.5
2
1.5
1
3
2
1
0.5
0
0
0
0.5
1
1.5
2
0
0.5
Direction selectivity
1
1.5
0
2
0
0.5
Direction selectivity
1
Direction selectivity
Direc)on selec)vity
Triangle
Rectangle
4
Hexagon
4
3.5
2
1
Speed selectivity
Speed selectivity
Speed selectivity
Speed selec$vity
3
3
2.5
2
1.5
1
3
2
1
0.5
0
0
0
0.1
0.2
0.3
Spatial selectivity
0.4
0.5
0
0.1
0.2
0.3
0.4
Spatial selectivity
Spa$al selec$vity
18
0.5
0
0
0.1
0.2
0.3
Spatial selectivity
0.4
0.5
Published as a conference paper at ICLR 2018
E
A DDITIONAL TRAINING DETAILS
During training we tried to balance all three terms we were minimizing (E, RL2 , and RF R ) so no
single term was neglected or dominated. At the beginning of training we weighted the regularization
term RL2 to be equal to the error function E and then decreased the weighting on RL2 according to
the schedule used by Martens & Sutskever (2011). We adaptively adjusted the weighting on RF R ,
starting from an initial value of E/10 and enforcing an upper bound of E/3 as training progressed.
We found this training procedure improved training performance and led to more interesting representations.
19
| 2cs.AI
|
arXiv:1706.02429v3 [stat.ME] 28 Nov 2017
Fast and General Best Subset Selection
using Data Depth and Resampling
Subhabrata Majumdar∗
University of Florida Informatics Institute
and
Snigdhansu Chatterjee†
School of Statistics, University of Minnesota
Abstract
We present a technique using data depth functions and resampling to perform
best subset variable selection for a wide range of statistical models. We do this by
assigning a score, called an e-value, to a candidate model, and use a fast bootstrap
method to approximate sample versions of these e-values. Under general conditions, e-values can separate statistical models that adequately explain properties of
the data from those that do not. This results in a fast algorithm that fits only a
single model and evaluates p + 1 models, p being the number of predictors under
consideration, as opposed to the traditional requirement of fitting and evaluating 2p
models. We illustrate in simulation experiments that our proposed method typically
performs better than an array of currently used methods for variable selection in
linear models and fixed effect selection in linear mixed models. As a real data application, we use our procedure to elicit climatic drivers of Indian summer monsoon
precipitation.
Keywords: Model selection, best subset selection, linear mixed model, generalized bootstrap, Indian monsoon.
∗
The author acknowledges the University of Minnesota Interdisciplinary Doctoral Fellowship program.
This research is partially supported by the National Science Foundation (NSF) under grants # IIS1029711 and # DMS-1622483 and by the National Aeronautics and Space Administration (NASA).
†
1
1
Introduction
Consider the linear model
Y = Xβ +
(1.1)
where Y ∈ Rn is the vector of responses, X ∈ Rn×p is the predictor matrix, β ∈ Rp is
the unknown vector of coefficients, and ∈ Rn is a vector of independent and identically
distributed random errors. Many important applications of this model demand the detection of important predictors affecting the response and thus obtaining a parsimonious
fit to the data in hand. Because of this, variable selection in linear models has been a
well-studied topic, including classics (Akaike, 1970; Schwarz, 1978; Shao, 1996; Tibshirani, 1996) and more recent advances (Bertsimas et al., 2016; Lai et al., 2015; Zou, 2006).
The two major avenues variable selection has been explored through are sparse penalized
regression and best subset selection. Sparse penalized regression has received abundant
attention in the past two decades due to its relative computational simplicity and ability
to tackle high-dimensional data. However, there are several inferential and algorithmic
issues associated with such methods: for example the bias and confidence intervals of
Lasso estimates (Zhang and Zhang, 2014) and use of convex relaxations to compute approximate local solutions of nonconvex penalized regression problems (Wang et al., 2013;
Zou and Li, 2008). A detailed discussion of the caveats of penalized regression procedures
can be found in Bertsimas et al. (2016).
On the other hand, the major challenge of performing best subset selection by checking multiple model fits is computational. In the setup of (1.1) there are 2p potential
models, and navigating through this model space is an NP-Hard problem (Natarajan,
1995). Greedy methods like forward selection or backward deletion give no theoretical
guarantee to select the correct model, and popular algorithms like leaps-and-bounds or
branch-and-bounds are not computationally scalable to even moderate data dimensions
(n ' 100, p ' 50). A recent paper by Bertsimas et al. (2016) proposed a tractable solution
to the best subset problem using techniques from the optimization literature. It requires
specifying the desired level of sparsity in the final model beforehand and works well in
practice when the true model has a small number of non-zero coefficients. However, this
is a largely unverifiable assumption, and there are more computational issues associated
with it in high-dimensional situations, which we shall discuss later in the paper.
Although there has been some progress in best subset model selection for linear regression, it has received lesser attention for more complex models due to its perceived
computational intractability and lack of intelligent search strategies. The fence methods
(Jiang et al., 2008; Nguyen and Jiang, 2014) are popular in performing model selection
in mixed effect models. Model selection criteria that expand the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) and Mallow’s Cp to mixed models
(Meza and Lahiri, 2005; Vaida and Blanchard, 2005) and robust regression (Ronchetti,
2000) have also been proposed. However, they still depend on multiple model fits and
suffer from the same scalability issues as linear regression best subset selection.
2
Our method In this paper, we propose a unified framework that is able to perform best
subset selection in the coefficient vector β for a range of statistical models much broader
than (1.1) by traversing a very small proportion of the space of candidate models. Even
though our proposal is not based on penalization, we are able to determine important
variables in a model by training a single model: the one with all covariates under consideration. We do this by obtaining a score for a candidate model M based on the amount
of similarity between the sampling distribution of the model coefficient estimate β̂M and
that of the least parsimonious estimate, say β̂. We use data depths, which are functions
that provide center-outward ranking for points in multivariate space with respect to a
probability distribution (Tukey, 1975; Zuo, 2003; Zuo and Serfling, 2000), to compute
this score- which we call the e-value of the candidate model.
Specifically, given some depth function D(., .) we define the e-value as
e(M) = ED(β̂M , [β̂])
i.e. the expected depth of the model estimates with respect to the sampling distribution
of β̂, which is denoted by [β̂]. In this setup, for large enough n the index set Sselect ⊆
{1, . . . , p} obtained using Algorithm 1 elicits all non-zero covariates in the true parameter
vector.
Algorithm 1. (Best subset selection using e-values)
1. Obtain e-value for the full model: e(Mf ull ) = ED(β̂, [β̂]).
2. Set Sselect = φ.
3. For j in 1 : p
Replace j th index of β̂ by 0, name it β̂−j .
Obtain e(M−j ) = ED(β̂−j , [β̂]).
If e(M−j ) < e(Mf ull ))
Set Sselect ← {Sselect , j}.
For a size-n finite sample, we use resampling to generate multiple copies of β̂ and subsequently estimate e(M) for each model.
The example in Figure 1.1 illustrates our motivation of using mean depth as a model
selection criterion. Suppose p = 2 and i ∼ N (0, σ 2 ), σ > 0, i = 1, . . . , n in (1.1), and
consider the following choice of models:
M1
M2
M3
M4
:
:
:
:
Y
Y
Y
Y
= X1 β1 + X2 β2 + ;
= X1 β1 + ;
= X2 β2 + ;
= ;
Θ1
Θ2
Θ3
Θ4
= R2
= R × {0}
= {0} × R
= (0, 0)T
where Θi denotes the domain of parameter values corresponding to model Mi ; i = 1, ..., 4.
Now assume that the true value of the coefficient vector is β0 = (5, 0)T , aT denoting
the transpose of a vector or matrix-valued a. The full model estimate has sampling
3
5
3
●
−4
−6 −
1
1
be
ta
1
7
●
M1
M2
M3
M4
●
2
ta2
0
−2
be
Figure 1.1: Illustration of the e-value method. The depth surface is obtained using
mahalanobis depth. The solid and dotted ellipses represent depth contours at some fixed
level α > 0 for sample sizes n1 and n2 , respectively, with n2 > n1 .
distribution N (β0 , σ 2 (X T X)−1 ). As n grows, this distribution concentrates around β0 ,
and the depths at points along the (red) line β2 = 0, as well as at the (red) point
(0,0) become smaller and smaller. Hence, the expected depths go to 0 for M3 and
M4 . However, owing to affine invariance of depth functions, the range of depth values
for points in M1 (blue line) and M2 (blue surface) stays the same. Moreover, depth
contours coincide with density contours for elliptical distributions (Liu and Singh, 1993),
thus e(M1 ) and e(M2 ) remain positive and do not converge to 0 as n → ∞.
For the purpose of covariate selection, the space of all possible models can be divided
into two classes: models with non-zero support that contains the non-zero support of
the true coefficient vector, and models that do not satisfy this. Because of the above
reasoning, e-values for models in the first class become clearly separated from e-values of
models in the second class. We further leverage the properties of depth functions to prove
that within the first class of ‘good’ models, a rank ordering of e-values of the models exists
based on the amount of parsimony in them: after which Algorithm 1 becomes immediate.
Structure of paper The paper is organized as follows. Section 2 provides background
for the development of our framework. We define e-values formally in Section 3, present
details about their estimation through resampling and their implementation through our
fast algorithm. Theoretical results are presented in Section 4. In Section 5 we present
illustrative examples on how our fast model selection algorithm is implemented, and its
performance relative to state-of-art methods in both classical n > p and high-dimensional
situations. This is followed in Section 6 with a study on the possible determinants of precipitation in the Indian sub-continent during the summer monsoon months. In Section 7
we discuss the potentials of our proposed method, directions of future research, as well as
some caveats and concluding comments. Proofs of all theoretical results are given in the
4
appendix.
Notations We denote the (i, j)th entry of a matrix M ∈ RA×B by (M )ij . The notations
IA and 0A indicate the identity matrix of dimension A × A and the vector of zeros of
dimension A, respectively.
For a vector v = (v1 , . . . , vA )T ∈ RA we denote its euclidean
qP
PA
A
2
norm by kvk =
i=1 I(vi 6= 0): I(.) being the indicator
i=1 vi , and l0 -norm by kvk0 =
function. We denote the non-zero support of v by supp(v), i.e. supp(v) = {i : vi 6=
0}. For random variables R, Rn , n = 1, 2, . . ., the notation Rn
R means convergence
P
in distribution, and Rn → R or plimn→∞ Rn = R means convergence in probability.
Expectations are denoted by E(·), and the variance of a scalar-valued random variable or
the covariance matrix of a vector-valued random variable is denoted by V(·).
2
Preliminaries
We introduce the theoretical framework necessary for our methodology in this section.
This includes specifications for the following three things: (1) an appropriate method of
estimation, (2) categorization of candidate models, and (3) the choice of a depth function.
Suppose at stage n there is a triangular array of functions
{Ψni (θn , Zni ) : 1 ≤ i ≤ kn , n ≥ 1}
(2.1)
where Zn = {Zn1 , . . . , Znkn } is an observable array of random variables, and θn ∈ Θn ⊆
Rp . Functions like Ψni (·) have been referred to as energy functions in optimization and
related literature, and have also been called contrast functions (Bose and Chatterjee, 2003;
Michel and Pfanzagl, 1971; Pfanzagl, 1969). This framework naturally includes different
kinds of regressions, mixed effect models, times series and other traditional topics where
covariate or model selection has been historically studied. Note that we use a triangular
array-based framework here, where the energy functions and the true parameter may
depend on n for precise theoretical study, and to capture the scientific reality that data,
models and goals of research evolve over time.
We assume that there is a true unknown vector of parameters θ0n , which is the unique
minimizer of
Ψn (θn ) = E
kn
X
Ψni (θn , Zni )
(2.2)
i=1
Also denote the common non-zero support of all estimable parameters as S∗n = ∪θn ∈Θn supp(θn ).
In this general setup, we associate a candidate model Mn with two quantities:
(a) The set of indices Sn ⊆ S∗n where the parameter values are unknown and estimated
from the data, and
5
(b) an ordered vector of known constants Cn = (Cnj : j ∈
/ Sn ) for parameters not
indexed by Sn .
The generic parameter vector corresponding to this model, denoted by θmn ∈ Θmn ⊆
Q
Θn := j Θnj , will thus have the structure
(
θmnj =
unknown θmnj ∈ Θnj
known Cnj ∈ Θnj
for j ∈ Sn ,
for j ∈
/ Sn .
(2.3)
So that in each model a part of the parameter vector consists of known constants, and
the other part is unknown.
2.1
Method of estimation
The estimator θ̂∗n of θ0n is obtained as a minimizer of the sample analog of (2.2). Thus
θ̂∗n = arg min Ψ̂n (θn ) = arg min
θn
θn
kn
X
Ψni θn , Yni
(2.4)
i=1
We assume an elliptical asymptotic distribution for θ̂∗n . Following Fang et al. (1990), the
density function of an elliptically distributed random variable takes the form:
h(x; µ, Σ) = |Σ|−1/2 g((x − µ)T Σ−1 (x − µ))
where µ ∈ Rp , Σ ∈ Rp×p is positive semi-definite, and g is a non-negative scalar-valued
density function that is continuous and strictly increasing, and is called the density generator of the elliptical distribution. We denote such an elliptical distribution by E(µ, Σ, g).
For the asymptotic parameter distribution we also assume the following conditions:
(A1) There exists a sequence of positive reals an ↑ ∞ such that an (θ̂∗n −θ0n )
E(0p , V, g),
p×p
for some positive-definite matrix V ∈ R
and density generator function g;
(A2) For almost every data sequence Zn , There exists a sequence of positive definite
matrices Vn ∈ Rp×p such that plimn→∞ Vn = V .
Note that under standard regularity conditions the limiting distribution of a suitably
centered and scaled parameter estimator will be Gaussian, in which case assumptions
(A1)-(A2) are naturally satisfied.
The estimator in (2.4) corresponds to the full model M∗n , i.e. the model where all
indices in S∗n are estimated. For any other model Mn , we simply augment entries of
θ̂∗n at indices in Sn with elements of Cn elsewhere to obtain a model-specific coefficient
estimate:
(
θ̂mnj =
θ̂∗nj
Cnj
for j ∈ Sn ,
for j ∈
/ Sn .
(2.5)
The logic behind this plug-in estimate is simple: for a candidate model Mn , a joint
distribution of its estimated parameters, i.e. [θ̂sn ], can actually be obtained from [θ̂∗n ] by
marginalizing at indices Sn .
6
2.2
Characterization of models
Given the formulation of models in (2.3), any model can be classified into two categories:
adequate and inadequate models.
Definition 2.1. A model Mn is called adequate if
lim
n→∞
X
|Cnj − θ0nj | = 0
(2.6)
j ∈S
/ n
A model that is not adequate, will be called an inadequate model.
We also consider a stronger version of adequacy below.
Definition 2.2. The model Mn is called strictly adequate if, for all n and j ∈
/ Sn ,
Cnj = θ0nj .
By definition the full model is always adequate, as is the model corresponding to the
singleton set containing the true parameter, i.e. {θ0n } (because supp(θ0n ) ⊆ S∗n ). Thus
the set of adequate models is non-empty by construction.
In classical variable selection where all elements of the constant terms Cn are set to 0
for all models, as in linear regression where a subset of covariates Xs is used in fitting the
expression Y = Xs βs + s , the above notion of model adequacy captures standard notions
of model ‘correctness’. Specifically, (2.6) reduces to limn→∞ EY −Xs βs = 0. This is weaker
than the traditional condition EY − Xs βs = 0 about model ‘correctness’ implicit in many
studies, which is captured in our condition for a model being strictly adequate. In addition,
our definition allows asymptotically similar models to be clubbed together, and has the
flexibility to accommodate local asymptotic characteristics. For instance, consider the
i.i.d.
data-generating model Yni = X1i β01 +X2i δn +i for some β01 ∈ R, δn = o(1), i = N (0, σ02 )
with σ0 > 0 and i = 1, . . . , kn . Evidently, here θ0n = (β01 , δn , σ0 )T . With our formulation,
the (constant) sequence of models Mn with Θmn = {(β1 , 0, σ 2 )T : β1 ∈ R, σ > 0} will
be adequate. Such situations can arise from prior choices in Bayesian variable selection
techniques, for example see Narisetty and He (2014); Roc̆kovà and George (2016).
Besides model adequacy, a notion that will be important in our framrwork is the one
of nested models.
Definition 2.3. We consider a model M1n to be nested in M2n , notationally M1n ≺
M2n , if S1n ⊂ S2n and C2n is a subvector of C1n .
ad
If a model Mad
n is adequate, then any model Mn such that Mn ≺ Mn is also adequate. In
the context of covariate selection in the linear model of (1.1), this obtains a linear ordering,
with the most parsimonious adequate model having Sn = supp(β0 ) and Cn = 0p−|S| nested
in all other adequate models. However all models nested within it shall be inadequate.
7
2.3
Depth functions
Let G denote the set of probability measures on Rp , and let D : Rp × G → [0, ∞) be a data
depth function (Zuo and Serfling, 2000). We consider data depth functions satisfying the
following properties:
(B1) The map D is invariant to affine transformations, i.e. for any non-singular matrix
A ∈ Rp×p , and b ∈ Rp and random variable Y having distribution G ∈ G,
D(x, G) = D(Ax + b, [AY + b])
(B2) The map D(., G) is Lipschitz continuous, where G ∈ G is a fixed, non-degenerate
probability distribution. That is, there exists δ > 0 and α ∈ (0, 1), possibly depending on G such that whenever |x − y| < δ, we have
|D(x, G) − D(y, G)| < |x − y|α
(B3) Assume that Yn ∈ Rp is a sequence of random variables converging in distribution
to some Y ∈ G. Then D(y, [Yn ]) converges uniformly to D(y, Y). In particular, if
Y ∼ Y, then
lim ED(Yn , [Yn ]) = ED(Y, Y)
n→∞
exists and is finite.
(B4) For any G ∈ G, limkxk→∞ D(x, G) = 0.
(B5) For any G ∈ G with a point of symmetry µ(G) ∈ Rp , we have for any t ∈ (0, 1) and
any x ∈ Rp
D(x, G) < D(µ(G) + t(x − µ(G)), G) < D(µ(G), G) = sup D(x, G) < ∞
x∈Rp
That is, the depth function takes a maximum value at µ(G), and is strictly decreasing along any ray connecting µ(G) to any point x ∈ Rp .
Conditions (B1), (B4) and (B5) are integral to the formal definition of data depth (Zuo and
Serfling, 2000), while (B2) and (B3) implicitly arise for several depth functions (Mosler,
2013). For most of the theoretical analysis in this paper we require only a subset of these
properties. However, we shall continue to use data depths throughout for simplicity.
3
The e-values methodology
We now associate with each model Mn a functional of the depth of the model estimate
with respect to the full model sampling distribution. We call this the e-value of the
model. An example of e-value is the mean depth:
en (Mn ) = ED(θ̂mn , [θ̂∗n ])
8
(3.1)
which we concentrate on for the rest of the paper. However, any other functional of
D(θ̂mn , [θ̂∗n ]) may also be used here, and a large proportion of our theoretical discussion
is applicable to any smooth functional of the distribution of D(θ̂mn , [θ̂∗n ]). Furthermore,
the distribution of D(θ̂mn , [θ̂∗n ]) is itself informative, and has an important role to play
in the study of uniform convergence. We defer all this discussion and analysis to future
research.
3.1
Resampling approximation of e-values
There are two random quantities involved in (3.1), namely θ̂mn and θ̂∗n . Typically, the
distribution of either of them is not known, and have to be elicited from the data. We use
resampling for this purpose. Because of the plugin method in (2.5), only [θ̂∗n ] needs to be
approximated, and we propose using Generalized Bootstrap (Chatterjee and Bose, 2005)
for this. We take an exchangeable array of non-negative random variables, independent
of the data, as resampling weights: Wrn = (Wrn1 , . . . , Wrnkn )T ∈ Rkn , and obtain the
resampling estimator θ̂r∗n as the minimizer of
Ψ̂rn (θn ) =
kn
X
Wrni Ψni θn , Zni .
(3.2)
i=1
We assume the following conditions on the resampling weights as n → ∞:
EWrn1
VWrn1
τn2
EWrn1 Wrn2
2
2
EWrn1
Wrn2
4
EWrn1
= 1,
= τn2 ↑ ∞,
= o(a2n ),
= O(kn−1 ),
→ 1,
< ∞.
(3.3)
(3.4)
(3.5)
(3.6)
(3.7)
(3.8)
Several resampling schemes can be described in the above format and are discussed in
Chatterjee and Bose (2005). Here we provide two examples.
Example 3.1 (The m-out-of-n or moon-bootstrap). In our framework, the moon-bootstrap
is identified with Wrn having a Multinomial distribution with parameters m and probabilities kn−1 (1, . . . , 1) ∈ Rkn , by a factor of kn /m. Thus we have EWrni = (m−1 kn )(m/kn ) = 1,
and VWrni = τn2 = (m−1 kn )2 (mkn−1 (1 − kn−1 ) = O(m−1 kn ). In typical applications of the
moon-bootstrap, we require that m → ∞ and m/kn → 0 as n → ∞. Thus we have
τn2 → ∞ as n → ∞, that is, the scale factor of the resampling weights Wrni tending to
infinity with n.
Example 3.2 (The scale-enhanced Bayesian bootstrap). A version of Bayesian bootstrap may be constructed by choosing Wrni to be independent and identically distributed
9
Gamma random variables, with mean 1 and variance τn2 → ∞ as n → ∞. The functionality of this resampling scheme and Bayesian interpretation remain similar to the standard
Bayesian bootstrap, however some convenient properties like conjugacy are lost.
Under fairly weak regularity conditions on the first two derivatives of Ψni (·), Chatterjee and Bose (2005) proved that (an /τn )(θ̂r∗n − θ̂∗n ) converges to the same weak limit in
probability as an (θ̂∗n − θ0n ). To estimate the e-value of any model we obtain two independent copies of the solution to (3.2), which we index by r and r1 . We use the first set
of resamples to obtain the distribution [θ̂r∗n ] conditional on the data that approximates
[θ̂∗n ], and the second set of resamples to obtain θ̂r1 mn corresponding to the model Mn :
(
θ̂r1 mnj =
for j ∈ Sn ;
for j ∈
/ Sn
θ̂r1 ∗nj
Cnj
(3.9)
Consequently, the resampling estimate of a model e-value is defined as
ern (Mn ) = Er1 D θ̂r1 mn , [θ̂r∗n ] ,
(3.10)
where Er1 is expectation, conditional on the data, of the resampling indexed by r1 .
The actual implementation of the resampling procedure is carried out by generating
independent copies W1n , . . . , WRn for some sufficiently large integer R. We do not need
to solve (3.2) explicitly to calculate θ̂r∗n . Rather, one can use already computed model
quantities in the following approximation to get them:
#−1 n
X
n
τn X
Wi Ψ00ni (θ̂∗n , Zni )
= θ̂∗n −
an i=1
"
θ̂r∗n
Wi Ψ0ni (θ̂∗n , Zni ) + Rrn ;
(3.11)
i=1
Er kRrn k2 = oP (1)
This is immediate from Theorem 3.5 in Chatterjee and Bose (2005). Consequently only
Monte Carlo sampling is required to obtain the resamples, resulting in significant computational speed gains.
3.2
Fast algorithm for best subset selection
For best subset selection we consider the class of candidate models
M0 = {M : Cj = 0 ∀ j ∈
/ S∗ }
(3.12)
dropping n in the subscripts of Mn , Cn , Sn , θn (and their variants). Note that all adequate
models are strictly adequate now, and their non-zero support contains supp(θ0 ). In this
setup our fast selection algorithm consists of only three stages:
(a) Fit the full model and estimate its e-value,
(b) Replace each covariate by 0 and compute e-value of all such reduced models, and
10
(c) Collect covariates dropping which causes the e-value to go down.
To compute the full model, we need to determine the estimable set of predictor indices
S∗ . In traditional n > p situations, the choice is simple: S∗ = {1, . . . , p}. In highdimnensional cases when p > n, we need to ensure that p0 = |S∗ | < n, so that θ̂∗ (properly
centered and scaled) has a unique asymptotic distribution, i.e. satisfies condition (A1).
We use fast screening methods available in the literature to achieve this, as implemented
previously by Lai et al. (2015). For high-dimensional linear regression we use the Sure
Independent Screening (SIS) procedure (Fan and Lv, 2008), while for more complex cases
like mixed models, single index models etc., we propose using the model-free screening
method of Zhu et al. (2011).
After obtaining S∗ and the full model estimate, for each of the p0 + 1 models under
consideration: the full model and all dropped-predictor models, we generate two sets of
resampling weights, use them in (3.11) to obtain bootstrapped coefficient estimates, and
finally calculate finite sample approximations of the bootstrap e-values in (3.10). This
gives us all the components for a sample version of the population level algorithm 1. We
present this as Algorithm 2.
Algorithm 2. (Best subset selection using e-values and bootstrap)
1. Fix resampling standard deviation τn .
2. Obtain bootstrap samples: T = {θ̂1∗ , ..., θ̂R∗ }, and T1 = {θ̂1∗ , ..., θ̂R1 ∗ }.
P 1
3. Calculate êrn (M∗ ) = R11 R
r1 =1 D(θ̂r1 ∗ , [T1 ]).
4. Set Ŝ0 = φ.
5. For j in 1 : p
For r1 in 1 : R1
Replace j th index of θ̂∗r1 by 0 to get θ̂r1 ,−j .
P 1
Calculate êrn (M−j ) = R11 R
r1 =1 D(θ̂r1 ,−j , [T1 ]).
If êrn (M−j ) < êrn (M∗ )
Set Ŝ0 ← {Ŝ0 , j}.
Remark Choice of the bootstrap standard deviation τn plays a central role in the performance of Algorithm 2. Its intermediate rate of divergence, i.e. τn → ∞, τn /an → 0,
is a necessary and sufficient condition for the consistency of the generalized bootstrap
procedure (see Chatterjee and Bose (2005) for details), as well as that of the bootstrap
approximation of population e-values (Theorem 4.2). Variable selection literature has
several examples of such quantities in action that result in model selection consistency:
for example the penalty term of BIC, or the tuning parameters of adaptive lasso (Zou,
2006) and Dantzig selector (Candes and Tao, 2007). We demonstrate the performance of
Algorithm 2 with several choices of τn in our simulations, and use validation on a test set
to select the optimal τn in our data example.
11
4
Theoretical results
We investigate theoretical properties of e-values in this section. Our first results are in
the most general setup of adequate and inadequate models. We then focus on strictly
adequate models, and finally on M0 , as a consequence deriving our fast algorithm.
4.1
Separation of adequate models
Our first result on the model elicitation process separates inadequate models from adequate models at the population level.
Theorem 4.1. Assume the conditions (A1)-(A2) on θ̂∗n , and that Ekθ̂∗n k4 < ∞. For
any data-depth function satisfying (B1)-(B4), the following hold as n → ∞:
1. For the full model M∗n , en (M∗n ) → e0 < ∞, where e0 = ED(Y, [Y ]) with Y ∼
E(0p , V, G),
2. For an adequate model Mn , |en (Mn ) − en (M∗n )| → 0,
3. For inadequate model Mn , en (Mn ) → 0.
This ensures that for large enough n, it is possible to find some threshold n ≤ en (M∗n )
such that all inadequate models have e-values less than the threshold, while e-values for
all adequate models fall above it.
For generalized bootstrap approximations of population e-values through (3.10), we
have the following equivalent result:
Theorem 4.2. Assume conditions of Theorem 4.1, and Ekθ̂∗n k8 < ∞. Additionally,
assume that the resampling weights Wrni are exchangeable random variables satisfying
the conditions (3.3)-(3.8). Then, conditional on the data, ern (Mn ) converges to zero in
probability if Mn is a sequence of inadequate models, and to e0 if it is a sequence of
adequate models.
Remark An examination of the proofs of the above two results, which follow similar
lines, reveals that they do not use all properties of data depths. Specifically, only location
and scale invariance is needed in place of the much stronger affine invariance in (B1),
and condition (B5) is not needed at all. We also do not use condition (A2) for θ̂∗n .
Consequently, theorems 4.1 and 4.2 continue to hold for depth-like functions in a divergingp setup under relaxed versions of conditions (B1)-(B5) and some technical conditions on
the energy functions, and possibly have broader implications than variable selection only.
We plan to explore this in detail in future work.
12
4.2
Derivation of algorithm
Our next result is concerned with strictly adequate models, and forms the foundation for
the fast model selection algorithm.
Theorem 4.3. Assume conditions (A1)-(A2), (B1)-(B5). Then, for a finite sequence of
strictly adequate models M1n ≺ . . . ≺ Mkn and any finite collection of inadequate models
Mk+1,n , . . . , MKn , we have
en (M1n ) > . . . > en (Mkn ) >
max
j∈{k+1,...K}
en (Mjn )
for large enough n.
We set the constant terms Cn to 0 now, and drop the subscript n from Mn , Sn , Cn , θn
and their variants. Recall that we have not yet used the actual data generating process as
a candidate model. Let us define the data generating model as M0 ≺ M∗ with associated
estimable index set S0 = supp(θ0 ) and C0 = 0p−|S0 | for j ∈
/ S0 . Then we have the following
result.
Corollary 4.1. Assume the conditions of Theorem 4.3. Consider the collection of candidate models M0 = {M : Cj = 0 ∀ j ∈
/ S∗ }. Then for large enough n,
M0 = arg max [en (M)]
M∈M0
(4.1)
Thus, within the restricted set of models M0 , the e-value maximizes at the true model.
However there are still 2p possible models. This is where the true advantage of using evalues comes through. It hinges on the following observation.
Corollary 4.2. Assume the conditions of Corollary 4.1. Consider the models M−j ∈ M0
with S−j = {1, . . . , p} \ {j} for j = 1, . . . , p. Then covariate j is a necessary component
of M0 , i.e. M−j is an inadequate model, if and only if
en (M−j ) < en (M∗ )
(4.2)
for sufficiently large n.
Dropping an essential predictor from the full model makes the model inadequate,
which has very small e-value for large enough n (Theorem 4.1), whereas dropping a nonessential predictor increases the e-value owing to Theorem 4.3. Thus, simply collecting
those predictors that cause decrease in the e-value on dropping them from the model suffices
for variable selection.
13
4.3
Model selection consistency
Consistency of bootstrap analogues of all model e-values follows from Theorem 4.2. This,
in conjunction with the results from Section 4.2 establishes model selection consistency
of the finite sample bootstrap procedure.
Corollary 4.3. Consider two sets of bootstrap estimates of θ̂∗ : T = {θ̂r∗ : r = 1, . . . , R}
and T1 = {θ̂r1 ∗ : r1 = 1, . . . , R1 }. Obtain sample e-value estimates as
R1
1 X
D(θ̂r1 m , [T ])
êrn (M) =
R1 r1 =1
(4.3)
where [T ] is the empirical distribution of the corresponding bootstrap samples, and θ̂r1 m
are obtained using (3.10). Consider the set of predictors Ŝ0 = {êrn (M−j ) < êrn (M∗ )}.
Then as n, R, R1 → ∞,
P2 (Ŝ0 = S0 ) → 1
(4.4)
where P2 is the probability conditional on the data and bootstrap samples.
Equation (4.4) is contingent on the fact that the the data generating model M0 is
indeed a member of M0 , i.e. S0 ⊆ S∗ . This is ensured trivially in classical n > p situations.
If p is larger than n, M0 is the set of all possible models on the predictor set selected by
an appropriate predictor screening procedure.
In high-dimensional linear models, we use the SIS procedure for this purpose. Given
that the estimable index set S∗ is selected using SIS, Fan and Lv (2008) proved under
fairly weak conditions that
exp[−Cn1−2κ ]
P (M0 ∈ M0 ) ≥ 1 − O
log n
!
(4.5)
where C > 0 and κ are constants that depend on the minimum signal in θ0 . For more
complex models we use the model-free screening method of Zhu et al. (2011) to obtain S∗ .
Under mild conditions on the predictor matrix, Theorem 3 in Zhu et al. (2011) ensures
the validity of their procedure. In our notations, the result therein states as
P (|S∗ ∩
S0c |
r
≥ r) ≤ 1 −
p+d
!d
(4.6)
for positive integers r and d: d = p being a good practical choice as per the numerical
studies in Zhu et al. (2011).
Combining (4.5) or (4.6) with Corollary 4.3 as needed establishes asymptotically accurate recovery of the support set S0 through Algorithm 2 in Section 3.2. For classical n > p
cases this holds for any statistical model satisfying the conditions (A1) and (A2), whereas
in high-dimensional situations it is contingent upon additional conditions required by the
screening method used.
14
Choice of τn
τn = log n
τn = n0.1
τn = n0.2
τn = n0.3
τn = n0.4
Choice of τn
τn = log n
τn = n0.1
τn = n0.2
τn = n0.3
τn = n0.4
Setting 1: n = 1000, p = 60
ρ = 0.5
ρ = 0.7
Sparsity
PE (×10−4 )
Sparsity
PE (×10−4 )
5.01
4.5
5.00
3.3
16.16
33.6
16.85
23.3
5.74
8.1
6.03
5.7
5.01
4.5
5.01
3.3
5.00
4.4
5.00
3.3
Setting 2: n = 60, p = 1000
ρ = 0.5
ρ = 0.7
Sparsity
PE (×10−2 )
Sparsity
PE (×10−2 )
6.63
4.5
6.34
3.4
7.57
4.6
7.24
3.0
7.38
4.6
7.23
3.0
6.94
4.6
6.58
3.0
6.08
4.1
5.66
3.9
Sparsity
5.06
17.89
6.75
4.96
3.14
ρ = 0.9
PE (×10−4 )
2.6
16.3
4.8
5.0
633.6
Sparsity
4.79
7.16
6.61
5.44
3.90
ρ = 0.9
PE (×10−2 )
3.7
2.1
2.3
3.0
6.0
Table 5.1: Model sparsity and prediction errors for different choices of τn
5
Simulation studies
We now present three simulation studies to examine the empirical performance of our
proposed method using model e-values, along with a host of other variable selection
methods for linear and linear mixed models. In terms of n and p, we consider both the
classical n > p case (Sections 5.1 and 5.3) and the high-dimensional case where p >> n
(Section 5.2). For all depth calculations, we use halfspace depth (Tukey, 1975).
5.1
Selecting covariates in linear regression (n > p)
We use the model Y = Xβ0 + , with ∼ Nn (0, In ), n = 1000 and p = 60. We generate
the rows of X independently from Np (0, ΣX ), with (ΣX )ij = ρ|i−j| . In the true coefficient
vector β0 , we set the first 5 coefficients to have non-zero values equal to 1, and the other
slope parameters to be zero. We repeat this setup for ρ ∈ {0.5, 0.7, 0.9}, and generate
100 independent datasets for each ρ. We use the scale-enhanced Bayesian bootstrap
of Example 3.2 with τn ∈ {log n, n0.1 , n0.2 , n0.3 , n0.4 }, and resampling Monte Carlo size
R = R1 = 1000 for the calculation of bootstrap e-values using (4.3).
Competing methods For comparative evaluation our technique, we consider methods
that come from different domains in the model selection literature: (a) The Mixed Integer
Optimization (MIO) method of Bertsimas et al. (2016), which is based on advanced
discrete optimization methods; (b) LASSO and (c) SCAD-penalized linear regression,
which represent the domain of penalized regressions; and (d) Stepwise regression using
the BIC and forward selection, which is a conventional favorite.
MIO has two versions: all subset and forward selection. The all subset method was
much more computationally intensive than other methods, so we used the forward selection MIO. This returns one model for each of the p + 1 possible sparsity levels of the
coefficient vector, following which we select the model with lowest BIC. For the penalized
regression methods we use ten-fold cross validation to select the optimal tuning parameter.
15
60
rho=0.9
60
rho=0.7
60
rho=0.5
●
evalue
MIO
Lasso
SCAD
Step
●
MIO
Lasso
SCAD
40
●
●
0
●
evalue
30
10
Model sparsity
50
●
20
50
40
30
Model sparsity
10
●
●
0
0
●
20
40
30
20
●
10
Model sparsity
50
●
Step
●
evalue
MIO
Lasso
SCAD
Step
●
●
●
●
●
2.0
2.0
2.0
●
●
●
●
●
●
●
●
●
●
●
●
●
1.5
1.0
●
●
●
●
●
Lasso
SCAD
Step
evalue
●
●
●
●
●
●
0.0
0.0
MIO
●
●
●
●
●
●
●
evalue
Prediction error
●
●
●
●
0.5
●
●
●
●
●
MIO
Lasso
●
●
0.0
●
●
●
0.5
●
●
●
●
●
●
●
1.0
●
Prediction error
1.5
●
1.5
1.0
●
●
0.5
Prediction error
●
●
SCAD
Step
●
evalue
MIO
Lasso
SCAD
Step
Figure 5.1: Model sparsity (top row) and prediction performance (bottom row) for all
methods in n = 60, p = 1000 setup.
Performance measures We use model sparsity and prediction error on an independently generated dataset (ytest , Xtest ) using the same setup as the training data to evaluate
each of the methods under consideration. For any estimate β̂, these metrics are defined
as
kXtest β̂ − Xtest β0 k
Sparsity(β̂) = kβ̂k0 ; PE(β̂) =
kXtest β0 k
These are plausible metrics of model evaluation when the features are highly correlated
(Bertsimas et al., 2016).
Choice of tuning parameter The top few rows of Table 5.1 give the average sparsity
and PE values of our e-value based method for different values of the bootstrap standard
deviation τn . Both the metrics take lowest values for ρ = 0.7 and ρ = 0.9 when τn = log n.
For ρ = 0.5, τn = n0.4 has marginally lower values, however it performs much worse than
τn = log n when ρ = 0.9.
Comparison with other methods Figure 5.2 presents boxplots of these performance
measures, with τn = log n assumed for our e-value based method. Our method gives the
sparsest models and lowest prediction errors among all methods across all values of ρ.
The second sparsest models are given by MIO, although its prediction errors are higher.
16
40
rho=0.9
40
rho=0.7
40
rho=0.5
●
30
30
30
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
20
Model sparsity
●
10
●
●
●
●
●
●
●
●
●
●
●
●
MIO
Lasso
SCAD
Step
MIO
Lasso
SCAD
Step
●
●
●
●
●
●
●
●
●
0.006
●
0.004
●
●
●
●
●
●
●
●
●
evalue
MIO
●
●
●
●
●
●
0.000
0.000
evalue
evalue
0.008
Step
●
●
0.000
SCAD
Prediction error
●
●
●
0.002
●
●
Lasso
0.006
Prediction error
0.006
●
●
●
●
MIO
0.008
evalue
●
0.004
0.002
Prediction error
Step
0.004
SCAD
0.002
Lasso
0.008
MIO
●
●
●
●
0
0
0
●
evalue
●
●
10
●
●
20
Model sparsity
20
●
10
Model sparsity
●
●
evalue
MIO
Lasso
SCAD
Step
Lasso
SCAD
Step
Figure 5.2: Model sparsity (top row) and prediction performance (bottom row) for all
methods in n = 1000, p = 60 setup. Dotted line in top row indicates true sparsity level.
On the other hand, prediction errors of SCAD are similar to our method, but our models
are more sparse. For all methods, higher values of ρ result in slightly sparser models and
higher prediction errors.
5.2
High-dimensional linear regression (p >> n)
Here we generate the data from the same setup as Section 5.1, but with n = 60, p = 1000.
To implement the e-values method, we first perform an initial screening of variables using
SIS, then apply our best subset selection procedure.
Among several choices of the tuning parameter, τn = log n and τn = n0.1 are good
choices (see second half of Table 5.1). The first choice provides sparser models, while the
second choice has lower prediction errors as the correlation among features increases.
To compare with the n > p case, we report the results corresponding to τn = log n
in the boxplots of Figure 5.1. As compared to the low-dimensional case in Section 5.1,
the e-values method provides much sparser estimates than other methods. MIO always
ends up selecting the least parsimonious model with 60 predictors. Our method has good
performance in terms of prediction error as well. Stepwise regression is the most unstable
among all five methods across all values of ρ and both performance measures used.
17
Method
Tuning
e-value based
τ =3
τ =4
τ =5
τ =6
τ =7
τ =8
BIC
AIC
GCV
p
log n/n
Peng and Lu Peng and Lu (2012)
FPR%
FNR%
Model size
Setting 1: ni = 5, m = 30
15.9
0.0
2.59
8.0
0.0
2.28
5.2
0.0
2.18
2.7
0.0
2.09
2.2
0.0
2.07
1.5
0.0
2.05
21.5
9.9
2.26
17
11.0
2.43
20.5
6
2.30
21
15.6
2.67
FPR%
FNR%
Model size
Setting 2: ni = 10, m = 60
5.2
0.0
2.17
2.8
0.0
2.09
2.0
0.0
2.06
0.7
0.0
2.02
0.3
0.0
2.01
0.3
0.0
2.01
1.5
1.9
2.10
1.5
3.3
2.20
1.5
3
2.18
1.5
4.1
2.26
Table 5.2: Comparison between our method and that proposed by Peng and Lu (2012)
through average false positive percentage, false negative percentage and model size
5.3
Best subset selection in presence of random effects
Here we use the simulation setup from Peng and Lu (2012). This is a random interceptonly model: Y = Xβ + ZU + . The data consists of independent groups of observations with multiple observations in each groups, Z being the within-group random
effects design matrix. We consider 9 fixed effects and 4 random effects, with true β0 =
(0, 1, 1, 0, 0, 0, 0, 0, 0) and random effect covariance matrix ∆ with elements (∆)11 = 9,
(∆)21 = 4.8, (∆)22 = 4, (∆)31 = 0.6, (∆)32 = (∆)33 = 1, (∆)4j = 0; j = 1, . . . , 4. The
error variance of is set at σ 2 = 1. The goal is to select covariates of the fixed effect. We
use two scenarios for our study: one where the number of subjects (m) considered is 30,
and the number of observations in the ith subject, i = 1, . . . , n, is ni = 5, and another
where m = 60, ni = 10.
ˆ for the resampling
Given the original restricted maximum likelihood estimates β̂, σ̂ 2 , ∆,
step we use the computational approximation
τn
β̂r ' β̂ + √ (X T V̂ −1 X)−1 Wr X T V̂ −1 (y − X β̂)
n
ˆ T . This adapts equation (3.11)
with Wr = diag(Wr1 I4 , . . . , Wr9 I4 ) and V̂ = σ̂ 2 Ip√+ Z ∆Z
in the mixed model setup. We consider τ = τn / n ∈ {1, . . . , 15} here, and independent
Gamma(1,1) random variables as the resampling weights {Wr1 + 1, . . . , Wr9 + 1}.
We compare our results with those obtained using the mixed effect model selection
method of Peng and Lu (2012), as well as two other methods, all of which are based on
sparse estimation of the fixed effect coefficients. In the method proposed by Peng and Lu
(2012), the tuning parameter can be selected using several different criteria. We present
the false positive percentage (FPR%), false negative percentage (FNR%) and model sizes
corresponding to four such criteria in Table 5.2. Our e-value based method handsomely
outperforms the method proposed by Peng and Lu (2012), especially in smaller sample
sizes, as long as τ ≥ 4. We also compare the percentages of times the correct model
was identified, and these results are presented in Table 5.3, along with the corresponding
results from two other papers. The proposed e-value based procedure performs best here
for τ ≥ 5 for the smaller sample setting, and for τ ≥ 7 for larger sample setting.
18
Method
e-value based
τ
τ
τ
τ
τ
τ
τ
=4
=5
=6
=7
=8
=9
= 10
Bondell et al. (2010)
Peng and Lu (2012)
Fan and Li (2012)
Setting 1
79
87
93
94
96
97
98
73
49
90
Setting 2
92
94
98
99
99
99
99
83
86
100
Table 5.3: Comparison of our method and three sparsity-based methods of mixed effect
model selection through accuracy of selecting correct fixed effects
6
Eliciting drivers of Indian monsoon precipitation
Various studies indicate that our knowledge about the physical drivers of precipitation in
India is incomplete; this is in addition to the known difficulties in modeling precipitation
itself (Knutti et al., 2010; Trenberth, 2011; Trenberth et al., 2003; Wang et al., 2005). For
example, Goswami et al. (2006) discovered an upward trend in frequency and magnitude
of extreme rain events, using daily central Indian rainfall data on a 10◦ × 12◦ grid, but a
similar study on a 1◦ ×1◦ gridded data by Ghosh et al. (2009) suggested that there are both
increasing and decreasing trends of extreme rainfall events, depending on the location.
Additionally, Krishnamurty et al. (2009) reported increasing trends in exceedances of the
99th percentile of daily rainfall; however, there is also a decreasing trend for exceedances
of the 90th percentile data in many parts of India. Significant spatial and temporal
variabilities at various scales have also been discovered by Dietz and Chatterjee (2014)
and Dietz and Chatterjee (2015).
Here we attempt to identify the driving factors behind precipitation during the Indian
monsoon season using our e-value based model selection criterion. Data are obtained from
the repositories of the National Climatic Data Center (NCDC) and National Oceanic and
Atmospheric Administration (NOAA), for the years 1978–2012. We obtained data on 35
potential covariates of the Indian summer precipitation based on existing knowledge and
conjectures from the actual physics driving Indian summer precipitations. Details are
provided in the supplementary material.
We consider annual medians of all covariates as fixed effects, the log yearly rainfall
at a weather station as response variable, and include year-specific random intercepts.
We use projection depth (Zuo, 2003) and resample Monte Carlo sizes R = R1 = 1000
to calculate bootstrap equivalents of the e-values. We train our model on data from
the years 1978-2002, run best subset selection for tuning parameter values τn ∈ nk ; k ∈
{0.01, 0.02, . . . , 0.16}, and select the predictor set corresponding to the τn that minimizes
fixed effect prediction errors on the samples from 2003–2012. Table 6.1 lists the estimated
e-values in increasing order for this full model, as well as all 35 models where a single
variable is dropped. Variables listed in the rows above none in Table 6.1 are considered
relevant by our e-value criterion.
19
3
Full model
Reduced model
2
●
●
●
●
0
●
●
●
●
●
●
●
●
●
−1
Bias
1
●
●
●
−2
●
●
−3
●
2004
2006
2008
2010
2012
year
8
e-value of M−j
0.1490772
0.2190159
0.2288938
0.2290021
0.2371846
0.2449195
0.2468698
0.2538924
0.2541503
0.2558397
0.2563105
0.2565388
0.2565687
0.2596766
0.2604214
0.2609039
0.261159
0.2624313
0.2626321
0.2630901
0.2633658
0.2649837
0.2660394
0.2663496
0.2677756
0.268231
0.2687891
0.2690791
0.2692731
0.2695297
0.2700276
0.2700579
0.2715089
0.2731708
0.2748239
0.2764488
Full model
●
Reduced model
6
●
4
MSE
●
●
●
2
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
●
2004
2006
2008
2010
2012
year
0.5
Year 2012
0.3
0.2
0.1
density
0.4
Truth
Full model pred
Reduced model pred
0.0
Predictor dropped
- TMAX
- X120W
- ELEVATION
- X120E
- del TT Deg Celsius
- X80E
- LATITUDE
- TNH
- Nino34
- X10W
- LONGITUDE
- X100E
- EAWR
- X70E
- v wind 850
- X140E
- X40W
- SolarFlux
- X160E
- EPNP
- TempAnomaly
- u wind 850
- WP
none
- POL
- TMIN
- X20E
- EA
- u wind 200
- u wind 600
- SCA
- DMI
- PNA
- v wind 200
- v wind 600
- NAO
−2
0
2
4
6
8
10
log(PRCP+1)
2012
●
Positive resid
negative resid
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Table 6.1: Ordered e-values when
dropping one variable at a time
in the Indian summer precipitation
data
Figure 6.1: Comparing full model rolling
predictions with reduced models: (Top to
bottom) Bias across years, Mean Squared
Error (MSE) across years, density plots for
2012, stationwise residuals for 2012
20
All variables selected by our procedure have documented effects on Indian monsoon
(Krishnamurthy and Kinter III, 2003; Moon et al., 2013). The single largest contributor is
the maximum temperature variable, whose relation to precipitation based on the ClausiusClapeyron relation is now classical knowledge in physics (Singleton and Toumi, 2012). It
seems that wind velocities high up in the atmosphere are not significant contributors, and
the fact that many covariates are selected in the process highlights the complexity of the
system.
We checked out-of-sample prediction performance of the estimated model using a
rolling validation scheme. For each of the 10 test years: 2003–2012, we select important
variables from the model built on past 25 years’ data (i.e. use data from 1978–2002 for
2003, 1979-2003 for 2004 and so on), build a model using them and compare predictions on
test year obtained from this model with those from the full model. Figure 6.1 summarizes
results obtained through this process. Across all testing years, reduced model predictions
have less bias as well as are more stable (top two panels). The better approximations of
truth by reduced models is also evident from the density plot for 2012 in the third panel,
and the residuals do not show any spatial patterns (bottom panel).
7
Conclusion
We present above an expansive framework and principle, where the definition of a statistical model is broad, and estimation procedures as well as resampling algorithms very
general. In such a scenario, we propose a scheme for best subset selection using the newly
defined e-value. An extremely fast algorithm obtains the true model selection with probability tending to one as the sample size grows by fitting and using only a single model.
Simulation results show that our algorithm performs better than traditional methods
in three illustrative examples, and a case study on Indian summer precipitation identifies several important physical drivers of monsoon precipitation. Theoretical consistency
results of multiple kinds are provided.
An immediate goal for future research is incorporating dimension asymptotics where
the parameter dimension is allowed to grow with the sample size, as well as extremely
high-dimensional parameters in our framework. Higher order correctness is a possibility in
our context since we use resampling methods. The sensitivity of the results to the choice
of the depth functions, resampling algorithm and the tuning parameter τn associated with
it, and the way D(θ̂mn , [θ̂∗n ]) is summarized to obtain the e-value deserve further study.
Our results suggest excellent asymptotic properties that are borne out in the simulations,
but finite-sample performance of the procedure needs further scrutiny.
To our knowledge, no other non-penalized method is available in the literature in
which computational cost of the covariate selection step scales linearly with dimension of
the feature space. This makes the e-values a perfect candidate for being an accurate and
scalable model selection method in big data. Because our algorithm easily lends itself to
parallel computing across different sets of bootstrap samples, it opens up the possibility
21
to be combined with bootstrap methods on subsamples of the full dataset, like the moonbootstrap, bag of little bootstraps (Kleiner et al., 2014), or subsampled double bootstrap
(Sengupta et al., 2016). For ensemble methods in the machine learning literature, e.g.
random forest and bagging, by extension of the same logic e-values present an exciting
opportunity to explore the formulation of a method for consistent feature ranking.
Acknowledgements
This research is partially supported by the National Science Foundation (NSF) under
grants #IIS-1029711 and #DMS-1622483 and by the National Aeronautics and Space
Administration (NASA). The first author also acknowledges the University of Minnesota
Interdisciplinary Doctoral Fellowship program.
A
Proofs of theoretical results
Proof of Theorem 4.1. Part 1 follows directly from assumption (B3).
Part 2. Assuming now that Mn is an adequate model, we use (B1) property of D:
h
D(θ̂mn , [θ̂∗n ]) = D θ̂mn − θ0n , θ̂0n − θ0n
i
(A.1)
,
and decompose the first argument
θ̂mn − θ0n = (θ̂mn − θ̂0n ) + (θ̂∗n − θ0n ).
(A.2)
where θmnj equals θ0nj in indices j ∈ Sn and Cnj elsewhere.
Now we have, for any Mn ,
θ̂mn = θmn + a−1
n Tmn
where Tmn is non-degenerate at the Sn indices. For the first summand of the right-hand
side in (A.2) we get
θ̂mn − θ̂∗n = θmn − θ0n + Rn
(A.3)
where EkRn2 k = O(a−2
n ). Since Mn is an adequate model, θmn − θ0n = o(1). Thus,
substituting the above right-hand side in (A.2) we get
h
D θ̂mn − θ0n , θ̂∗n − θ0n
i
h
− D θ̂∗n − θ0n , θ̂∗n − θ0n
i
≤ kRn kα
(A.4)
from Lipschitz continuity of D(·) given in (B2). The result now follows.
Part 3. Since the depth function E is invariant under location and scale transformations, we have
h
i
D(θ̂mn , [θ̂∗n ]) = D an (θ̂mn − θ0n ), an (θ̂∗n − θ0n )
22
.
(A.5)
Decomposing the first argument,
an (θ̂mn − θ0n ) = an (θ̂mn − θmn ) + an (θmn − θ0n ).
(A.6)
Since Mn is inadequate, given δ > 0 there exists a subsequence indexed by jn such that
|Cnjn − θ0jn | > δ for all n. The result follows by application of condition (B4).
Proof of Theorem 4.2. Several derivations here are similar to those of the proof of Theorem 4.1. Consequently we skip some details.
We have
a
an
n
θ̂r1 mn − θ̂∗n ,
θ̂r∗n − θ̂∗n
D θ̂r1 mn , [θ̂r∗n ] = D
τ
τn
n
a
an
an
n
.
=D
θ̂r mn − θ̂mn −
θ̂mn − θ̂∗n ,
θ̂r∗n − θ̂∗n
τn 1
τn
τn
Conditional on the data, (an /τn )(θ̂r1 mn − θ̂mn ) has the same weak limit as an (θ̂mn − θmn ),
and (A.3) and τn → ∞ combine to give
an
P
θ̂mn − θ̂∗n → 0
τn
as n → ∞. The results follow directly now.
Proof of Theorem 4.3. Since we are dealing with a finite sequence of nested models, it is
enough to prove that en (M1n ) > en (M2n ) for large enough n, when both M1n and M2n
are strictly adequate models and M1n ≺ M2n .
Suppose T0 = E(0p , Ip , g). Affine invariance implies invariant to rotational transformations, and since the evaluation functions we consider decrease along any ray from
the origin because of (B5), E(θ, T0 ) is a monotonocally decreasing function of kθk for any
θ ∈ Rp . Now consider the models M01n , M02n that have 0 in all indices outside S1n and S2n ,
respectively. Take some θ10 ∈ Θ01n , which is the parameter space corresponding to M01n ,
and replace its (zero) entries at indices j ∈ S2n \ S1n by some non-zero δ ∈ Rp−|S2n \S1n | .
Denote it by θ1δ . Then we shall have
T
T
θ1δ > θ10
θ10
θ1δ
⇒
⇒
D(θ10 , T0 ) > D(θ1δ , T0 )
Es1 D(θ10 , T0 ) > Es1 D(θ1δ , T0 )
where Es1 denotes the expectation taken over the marginal of the distributional argument
T0 at indices S1n . Notice now that by construction θ1δ ∈ Θ02n , the parameter space
corresponding to M02n , and since the above holds for all possible δ, we can take expectation
over indices S2n \ S1n in both sides to obtain Es1 D(θ10 , T0 ) > Es2 D(θ20 , T0 ), with θ20
denoting a general element in Θ20 .
Combining (A1) and (A2) we get an Vn−1/2 (θ̂∗n −θ0n )
T0 . Denote Tn = [an Vn−1/2 (θ̂∗n −
θ0n )], and choose a positive < (Es1 D(θ10 , T0 ) − Es2 D(θ20 , T0 ))/2. Then, for large enough
n we shall have
|D(θ10 , Tn ) − D(θ10 , T0 )| <
⇒
|Es1 D(θ10 , Tn ) − Es1 D(θ10 , T0 )| <
23
following condition (B4). Similarly we have |Es2 D(θ20 , Tn ) − Es2 D(θ20 , T0 )| < for the
same n for which the above holds. This implies Es1 D(θ10 , Tn ) > Es2 D(θ20 , Tn ).
Now apply the affine transformation t(θn ) = Vn1/2 θn /an + θ0n to both arguments of the
depth function above. This will keep the depths constant following affine invariance, i.e.
D(t(θ10 ), [θ̂∗n ]) = D(θ10 , Tn ) and D(t(θ20 ), [θ̂∗n ]) = D(θ20 , Tn ). Since this transformation
maps Θ01n to Θ1n , the parameter space corresonding to M1n , we get Es1 D(t(θ10 ), [θ̂∗n ]) >
Es2 D(t(θ20 ), [θ̂∗n ]), i.e. en (M1n ) > en (M2n ).
For any inadequate model Mjn , k < j ≤ K, suppose Njn is the integer such that
en1 (Mjn1 ) < en1 (M∗n1 ) for all n1 > Njn . Part 3 of Theorem 4.1 ensures that such
an integer exists for every inadequate model. Now define N = maxk<j≤K Njn . Thus
en1 (M∗n1 ) is larger than e-values of all inadequate models Mjn1 for k < j ≤ K.
Proof of Corollary 4.1. By construction, M0 is nested in all other adequate models in
M0 . Hence Theorem 4.3 implies en (M0 ) > en (Mad ) > en (Minad ) for any adequate model
Mad and inadequate model Minad in M0 and large enough n.
Proof of Corollary 4.2. Consider j ∈ S0 . Then θ0 ∈
/ M−j , hence M−j is inadequate. By
choice of n1 from Corollary 4.1, e-values of all inadequate models are less than that of
M∗ , hence en1 (M−j ) < en1 (M∗ ).
On the other hand, suppose there exists a j such that en1 (M−j ) ≤ en1 (M∗ ) but
j∈
/ S0 . Now j ∈
/ S0 means that M−j is an adequate model. Since M−j is nested within
M∗ for any j, and the full model is always adequate, we have en1 (M−j ) > en1 (M∗ ) by
Theorem 4.3: leading to a contradiction and thus completing the proof.
Proof of Corollary 4.3. Corollary 4.2 implies that
S0 = {j : en (M−j ) < en (M∗ )}
Now define S̄0 = {j : ern (M−j ) < ern (M∗ )}. Theorem 4.2 implies that
ern (M−j ) = en (M−j ) + Rjn
ern (M∗ ) = en (M∗ ) + R∗n
such that Er |R∗n |2 = oP (1) and Er |Rjn |2 = oP (1) for all j. Hence P1 (S̄0 = S0 ) → 1
as n → ∞, P1 being probability conditional on the data. Similarly one can prove that
the probability conditional on the bootstrap samples that S̄0 = Ŝ0 holds goes to 1 as
R, R1 → ∞, which completes the proof.
References
Akaike, H. (1970). Statistical predictor identification. Ann. Inst. Statist. Math., 22(1):203–
217.
24
Bertsimas, D., King, A., and Mazumder, R. (2016). Best Subset Selection via a Modern
Optimization Lens. Ann. Statist., 44(2):813–852.
Bondell, H. D., Krishna, A., and Ghosh, S. K. (2010). Joint variable selection for fixed
and random effects in linear mixed-effects models. Biometrics, 66(4):1069–1077.
Bose, A. and Chatterjee, S. (2003). Generalized bootstrap for estimators of minimizers
of convex functions. J. Stat. Plan. Inference, 117(2):225–239.
Candes, E. and Tao, T. (2007). The Dantzig selector: Statistical estimation when p is
much larger than n. Ann. Statist., 35(6):2313–2351.
Chatterjee, S. and Bose, A. (2005). Generalized bootstrap for estimating equations. Ann.
Statist., 33(1):414–436.
Dietz, L. and Chatterjee, S. (2014). Logit-normal mixed model for Indian monsoon precipitation. Nonlin. Proc. Geophys., 21:934–953.
Dietz, L. and Chatterjee, S. (2015). Extreme thresholds in Indian monsoon precipitation
using logit-Normal mixed models. In Machine Learning and Data Mining Approaches
to Climate Science, pages 239–246. Springer, New York, NY, USA.
Fan, J. and Lv, J. (2008). Sure Independence Screening for Ultrahigh Dimensional Feature
Space. J. R. Statist. Soc. B, 70:849–911.
Fan, Y. and Li, R. (2012). Variable selection in linear mixed effects models. Ann. Statist.,
40(4):2043–2068.
Fang, K. T., Kotz, S., and Ng, K. W. (1990). Symmetric multivariate and related distributions. Monographs on Statistics and Applied Probability, volume 36. Chapman and
Hall Ltd., London.
Ghosh, S., Luniya, V., and Gupta, A. (2009). Trend analysis of indian summer monsoon
rainfall at different spatial scales. Atmos. Sci. Lett., 10:285–290.
Goswami, B., Venugopal, V., Sengupta, D., Madhusoodanan, M. S., and Xavier, P. K.
(2006). Increasing trend of extreme rain events over india in a warming environment.
Science, 314:1442–1445.
Jiang, J., Rao, J. S., Gu, Z., and Nguyen, T. (2008). Fence methods for mixed model
selection. Ann. Statist., 36(4):1669–1692.
Kleiner, A., Talwalkar, A., Sarkar, P., and Jordan, M. I. (2014). A scalable bootstrap for
massive data. J. R. Statist. Soc. B, 76(4):765–814.
Knutti, R., Furrer, R., Tebaldi, C., Cermak, J., and Meehl, G. A. (2010). Challenges in
combining projections from multiple climate models. J. Clim., 23(10):2739–2758.
25
Krishnamurthy, V. and Kinter III, J. L. (2003). The Indian monsoon and its relation
to global climate variability. In Rodo, X. and Comin, F. A., editors, Global Climate:
Current Research and Uncertainties in the Climate System, pages 186–236. Springer.
Krishnamurty, C. K. B., Lall, U., and Kwon, H.-H. (2009). Changing frequencey and
intensity of rainfall extremes over india from 1951 to 2003. J. Clim., 22(18):4737–4746.
Lai, R. C. S., Hannig, J., and Lee, T. C. M. (2015). Generalized Fiducial Inference for
Ultrahigh-Dimensional Regression. J. Amer. Statist. Assoc., 110(510):760–772.
Liu, R. and Singh, K. (1993). A Quality Index Based on Data Depth and Multivariate
Rank Tests. J. Amer. Statist. Assoc., 88:252–260.
Meza, J. and Lahiri, P. (2005). A note on the Cp statistic under the nested error regression
model. Surv. Methodol., 31:105–109.
Michel, R. and Pfanzagl, J. (1971). The accuracy of the normal approximation for minimum contrast estimates. Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte
Gebiete, 18(1):73–84.
Moon, J.-Y., Wang, B., Ha, K.-J., and Lee, J.-Y. (2013). Teleconnections associated with
Northern Hemisphere summer monsoon intraseasonal oscillation. Clim. Dyn., 40(1112):2761–2774.
Mosler, K. (2013). Depth statistics. In Becker, C., Fried, R., and Kuhnt, S., editors,
Robustness and Complex Data Structures, pages 17–34. Springer Berlin Heidelberg.
Narisetty, N. N. and He, X. (2014). Bayesian variable selection with shrinking and diffusing
priors. Ann. Statist., 42:789–817.
Natarajan, B. K. (1995). Sparse Approximate Solutions to Linear Systems. Siam. J.
Comput., 24:227–234.
Nguyen, T. and Jiang, J. (2014). Restricted fence method for covariate selection in
longitudinal data analysis. Biostatistics, 13(2):303–314.
Peng, H. and Lu, Y. (2012). Model selection in linear mixed effect models. J. Multivar.
Anal., 109:109–129.
Pfanzagl, J. (1969). On the measurability and consistency of minimum contrast estimates.
Metrika, 14(1):249–272.
Ronchetti, E. (2000). Data Segmentation and Model Selection for Computer Vision:
A Statistical Approach, chapter Robust Regression Methods and Model Selection.
Springer. pp 31–40.
26
Roc̆kovà, V. and George, E. I. (2016). The Spike-and-Slab LASSO. J. Amer. Statist.
Assoc., 0:0–0.
Schwarz, G. (1978). Estimating the dimension of a model. Ann. Statist., 6(2):461–464.
Sengupta, S., Volgushev, S., and Shao, X. (2016). A Subsampled Double Bootstrap for
Massive Data. J. Amer. Statist. Assoc., 111(515):1222–1232.
Shao, J. (1996). Bootstrap model selection. J. Amer. Statist. Assoc., 91(434):655–665.
Singleton, A. and Toumi, R. (2012). Super-Clausius-Clapeyron scaling of rainfall in a
model squall line. Quat. J. R. Met. Soc., 139(671):334–339.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. J. R. Statist. Soc.
B, 58(267–288).
Trenberth, K. E. (2011). Changes in precipitation with climate change. Clim. Res.,
47(1-2):123–138.
Trenberth, K. E., Dai, A., Rasmussen, R. M., and Parsons, D. B. (2003). The changing
character of precipitation. Bull. Am. Meteorol. Soc., 84(9):1205–1217.
Tukey, J. W. (1975). Mathematics and picturing data. In James, R., editor, Proceedings
of the International Congress on Mathematics, volume 2, pages 523–531.
Vaida, F. and Blanchard, S. (2005). Conditional Akaike information for mixed effects
models. J. Amer. Statist. Assoc., 93:120–131.
Wang, B., Ding, Q., Fu, X., Kang, I.-S., Jin, K., Shukla, J., and Doblas-Reyes, F. (2005).
Fundamental challenge in simulation and prediction of summer monsoon rainfall. Geophys. Res. Lett., 32(15).
Wang, L., Kim, Y., and Li, R. (2013). Calibrating Nonconvex Penalized Regression in
Ultra-high Dimension. Ann. Statist., 41:2505–2536.
Zhang, C.-H. and Zhang, S. S. (2014). Confidence intervals for low dimensional parameters
in high dimensional linear models. J. R. Statist. Soc. B, 76(1):217–242.
Zhu, L.-P., Li, L., Li, R., and Zhu, L.-X. (2011). Model-Free Feature Screening for
Ultrahigh-Dimensional Data. J. Amer. Statist. Assoc., 106(496):1464–1475.
Zou, H. (2006). The Adaptive Lasso and Its Oracle Properties. J. Amer. Statist. Assoc.,
101:1418–1429.
Zou, H. and Li, R. (2008). One-step sparse estimates in nonconcave penalized likelihood
models. Ann. Statist., 36:1509–1533.
27
Zuo, Y. (2003). Projection-based depth functions and associated medians. Ann. Statist.,
31:1460–1490.
Zuo, Y. and Serfling, R. (2000). General notions of statistical depth function. Ann.
Statist., 28(2):461–482.
28
| 10math.ST
|
A Revisit of Hashing Algorithms for Approximate Nearest
Neighbor Search
Deng Cai
The State Key Lab of CAD&CG, College of Computer Science, Zhejiang University, China
Alibaba-Zhejiang University Joint Institute of Frontier Technologies
[email protected]
arXiv:1612.07545v4 [cs.CV] 18 Feb 2018
ABSTRACT
Approximate Nearest Neighbor Search (ANNS) is a fundamental
problem in many areas of machine learning and data mining. During the past decade, numerous hashing algorithms are proposed
to solve this problem. Every proposed algorithm claims outperform other state-of-the-art hashing algorithms. However, the evaluation of these hashing papers was not thorough enough, and
those claims should be re-examined. The ultimate goal of an ANNS
method is returning the most accurate answers (nearest neighbors)
in the shortest time. If implemented correctly, almost all the hashing methods will have their performance improved as the code
length increases. However, many existing hashing papers only report the performance with the code length shorter than 128. In
this paper, we carefully revisit the problem of search with the hash
index and make a thorough comparison of many popular hashing
algorithms. Surprisingly, the random-projection-based Locality Sensitive Hashing (LSH) ranked the second among twelve compared
popular hashing algorithms, while all the other eleven hashing
papers claim to outperform LSH. Despite the extreme simplicity
of random-projection-based LSH, our results show that the capability of this algorithm has been far underestimated. For the sake
of reproducibility, all the codes used in the paper are released on
GitHub, which can be used as a testing platform for a fair comparison between various hashing algorithms.
CCS CONCEPTS
• Information systems → Search engine indexing;
KEYWORDS
ANNS, hashing, hamming ranking
ACM Reference Format:
Deng Cai. 2018. A Revisit of Hashing Algorithms for Approximate Nearest
Neighbor Search. In Proceedings of arXiv. ACM, New York, NY, USA, 11 pages.
https://doi.org/10.1145/nnnnnnn.nnnnnnn
1
INTRODUCTION
Nearest neighbor search plays an important role in many applications of machine learning and data mining. Given a dataset with
N entries, the cost for finding the exact nearest neighbor is O(N ),
which is very time consuming when the data set is large. So people
turn to Approximate Nearest Neighbor (ANN) search in practice
[1, 20]. Hierarchical structure (tree) based methods, such as KD-tree
[3] ,Randomized KD-tree [32], K-means tree [6], are popular solutions for the ANN search problem. These methods perform very
well when the dimension of the data is relatively low. However, the
performance decreases dramatically as the dimension of the data
increases [32].
During the past decade, hashing based ANN search methods
[8, 23, 36] received considerable attention. These methods generate
binary codes for high dimensional data points (real vectors) and try
to preserve the similarity among the original real vectors. A hashing algorithm generating l-bits code can be regarded as contains l
hash functions. Each function partitions the original feature space
into two parts, the points in one part are coded as 1, and the points
in the other part are coded as 0. When an l-bits code is used, the
hashing algorithm (l hash functions) partitions the feature space
into 2l parts, which can be named as hash buckets. Thus, all the
data points fall into different hash buckets (associated with different binary codes). Ideally, if neighbor vectors fall into the same
bucket or the nearby buckets (measured by the hamming distance
of binary codes), the hashing based methods can efficiently retrieve
the nearest neighbors of a query point.
One of the most popular hashing algorithms is Locality Sensitive
Hashing (LSH) [8, 13]. LSH is a name for a set of hashing algorithms,
and we can specially design different LSH algorithms for the different type of data [8]. For real vectors, random-projection-based
LSH [4] might be the most simple and popular one. This algorithm
uses random projection to partition the feature space.
LSH (random-projection-based) is naturally data independent.
Many data-dependent hashing algorithms [7, 9–12, 14–16, 18, 19,
21–24, 26–28, 31, 33, 35–40] have been proposed during the past
decades and all these algorithms claimed to have superior performance over the baseline algorithm LSH. However, the evaluation
of all these papers are not thorough enough, and their conclusions
are questionable:
• Despite the performance increases as the code length increases, most of the hashing papers [7, 10, 11, 14–16, 19, 22–
24, 26–28, 33, 35–40] only reported the performance with
the code length shorter than 128. How about we further increase the code length? And what is the up-bound of the
code length?
• The most popular implementation of using a hash index for
search is the so-called "hash bucket search" approach 12345 .
1 https://github.com/kayzhu/LSHash
2 https://github.com/willard-yuan/hashing-baseline-for-image-retrieval
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
2018. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00
https://doi.org/10.1145/nnnnnnn.nnnnnnn
3 https://github.com/kevinlin311tw/Caffe-DeepBinaryCode
4 https://github.com/thuml/HashNet
5 https://github.com/ymcidence/DeepSketchHashing
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Is it the right way of using hash index for high accuracy
required ANNS problems?
• Most of the hashing papers use the accuracy - # of retrieved samples curves to compare various hashing algorithms. This metric certainly does not include the possible
extra cost of using longer codes. Can we use other metrics
which includes all the costs, so that we can compare the
hashing algorithms with various non-hashing based ANNS
methods?
These problems have already been raised in the literature. [18]
pointed out that many data-dependent hashing algorithms have
the improvements over LSH, but "improvements occur only for
relatively small hash code sizes up to 64 or 128 bits". The figure
4 in this paper [18] shows that when 1,024-bits codes were used,
LSH is the best-performed hashing algorithm. However, [18] did
not provide a full analysis.
There is also a public available c++ program named FALCONN
on the GitHub6 which implements LSH algorithm. Since we can
record the search time of the program, the FALCONN (LSH) can
be fairly compared with other non-hashing based ANNS methods.
However, the results are not very encouraging7 . There are even
researchers pointing out that "LSH is so hopelessly slow and/or
inaccurate"8 . Is this true that "LSH is so hopelessly slow and/or
inaccurate"?
All these problems motivate us to revisit the problem of applying
hashing algorithms for ANNS. This study is carried on two popular million scale ANNS datasets (SIFT1M and GIST1M)9 , and the
findings are very surprising:
• Twelve popular hashing algorithms are thoroughly compared on these two datasets. Despite the fact that all the
other eleven data-dependent hashing algorithms claimed the
superiority over random-projection-based LSH, LSH ranked
the second among all the twelve compared algorithms in our
experiments. There is only one algorithm, Spherical Hashing [12], which consistently outperforms LSH on both two
datasets.
• We also compare our implementation of random-projectionbased LSH (RPLSH) with other five popular open source
ANNS algorithms (FALCONN, flann10 , annoy11 , faiss12 and
KGrpah13 ). On SIFT1M, if a high recall is required (higher
than 98%), RPLSH performs the second best among the six
compared methods. On GIST1M, if a high recall is required
(higher than 96%), RPLSH performs best among the six compared methods.
Despite the extreme simplicity of the random-projection-based LSH,
our results show that the capability of this algorithm has been far
underestimated. For the sake of reproducibility, all the codes are
6 https://github.com/FALCONN-LIB/FALCONN
7 http://www.itu.dk/people/pagh/SSS/ann-benchmarks/
8 http://www.kgraph.org/
9 http://corpus-texmex.irisa.fr/
10 https://github.com/mariusmuja/flann
11 https://github.com/spotify/annoy
12 https://github.com/facebookresearch/faiss
13 https://github.com/aaalgo/kgraph
Deng Cai
Algorithm 1 Search with a hash index
Require: base set D, query vector q, the hash functions, hash index
(binary codes for all the points in the base set), the pool size L
and the number K of required nearest neighbors,
Ensure: K points from the base set which are close to q.
1: Encode the query q to a binary code b;
2: Locate L points which have shortest hamming distance to b;
3: return the closet K points to q among these L points.
Algorithm 2 Hash bucket search subroutine for locate L closest
binary codes
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
Let hamming radius r = 0;
Let candidate pool set C = ∅;
while |C | < L do
Get the buckets list B with hamming distance r to b;
for all bucket bb in B do
Put all the points in bb into C;
if |C | >= L then
Keep the first L points in C;
return;
end if
end for
r = r + 1;
end while
released on GitHub1415 which can be used as a testing platform for
a fair comparison between various hashing methods.
2
SEARCH WITH A HASH INDEX
A hashing method transforms the real vectors to binary codes. The
binary codes can then be used as a hash index for online search.
The general procedure of searching with a hash index is as follows
(Suppose the user submit a query q and ask for K neighbors of the
query):
(1) The search system encodes the query to a binary code b use
the hash functions.
(2) The search system finds L points which are closest to b in
terms of the hamming distance.
(3) The search system performs a scan within the L points, returns the top K points which are closet to q.
This procedure is summarized in Algorithm 1, and there is only one
parameter L which can be used to control the accuracy-efficiency
trade-off. Line 1 and 3 are very straightforward, and nothing needs
to be discussed. However, it can be tricky in implementing line 2.
There are two popular procedures to complete the task in line 2.
We named them hash bucket search approach (Algorithm 2) and
hamming ranking approach (Algorithm 3). If the code length is l,
the hash bucket search approach builds 2l buckets and store the
binary vectors in these buckets.
Based on this search procedure, we can see that the time of
searching with a hash index contains three parts [17]:
(1) Coding time: the line 1 in Algorithm 1.
14 https://github.com/ZJULearning/MatlabFunc/tree/master/ANNS/Hashing
15 https://github.com/ZJULearning/hashingSearch
A Revisit of Hashing Algorithms for Approximate Nearest Neighbor Search
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Table 1: Data sets
Algorithm 3 Hamming ranking subroutine for locate L closest
binary codes
1:
2:
Compute the hamming distances between b and all the binary
codes in the data set;
Sort the points based on the hamming distances in ascending
order and keep the first L points;
(2) Locating time: the lines 2 in Algorithm 1.
(3) Scanning time: the line 3 in Algorithm 1.
The hash bucket search and the hamming ranking approaches will
need different locating times, and we analyze the pros and cons
of these two approaches here and verify this analysis in the next
section.
• If the search time is considered:
– The hamming ranking approach needs to compute the
hamming distance of the query code b to all the codes
stored in the database. This portion of time16 is dependent
on the code length l and independent with the parameter
L. Then we need to find the smallest L values (distances).
Overall, the locating time of the hamming ranking approach grows linearly with respect to the code length l
and the parameter L.
– The case of the hash bucket search approach is more
complicated. Given a binary code b, locating the hashing bucket corresponding to b costs O(1) time (by using
std::vector or std::unordered_map). If there are enough
points (larger than L) in this bucket, the total time cost of
the hash bucket search is O(1), which seems much more
efficient than the hamming ranking approach. However,
this is only the ideal case. In reality, there are always
not enough points (even no point) in this bucket, and we
need to increase the hamming radius r . Given an l-bits
code b, considering those hashing buckets whose hamming distance to b is smaller or equal to r . It is easy to
Í
show the number of these buckets is ri=0 li , which increases almost exponentially with respect to r and l. As
the L increases, to locate L points successfully, r has to be
increased. Overall, the locating time of the hash bucket
search approach grows exponentially with respect to
the code length l and at least linearly with respect to the
parameter L.
– When the code length is short and the number of required
points L is small, the hash bucket search approach will be
much more efficient than the hamming ranking approach.
However, as the code length increases and L increases,
the hash bucket search approach soon becomes extremely
slow.
• If the index size is considered:
– If the hamming ranking approach is used, we can easily
find that there is no additional data structure needed for
Algorithm 3 and we only need to store the binary codes
for the data points in the database. Suppose the database
contains n data points and we use l bits code, we need n 8l
16 It
is important to note that the hamming distance computation of two binary codes
can be extremely fast with smart algorithms. Please see [34].
data set
SIFT1M
GIST1M
dimension
128
960
base number
1,000,000
1,000,000
query number
10,000
1,000
Algorithm 4 random projection based LSH
Require: base set D with n m-dimensional points, code length l
Ensure: m × l projection matrix A and n l-bits binary codes..
1: Generate an m × l random matrix A from Gaussian distribution
N (0, 1);
2: Project the data onto l dimensional space using this random
matrix.
3: Convert the n × l dimensional data matrix to n l-bits binary
codes by the binarize operation.
byte. This index size may be the smallest one among most
of the popular ANNS methods even when l = 1024.
– If the hash bucket search approach is used, we need the
data structures (either std::vector or std::unordered_map)
to store the hash buckets. The index size will be much large
than the hamming ranking approach, especially when the
code length is considerable.
In summary, the hash bucket search approach is suitable for the
case that the code is not very long and L is small (e.g., near-duplicate
search). If the code length is very long and L is pretty large (e.g.,
high accuracy required ANNS), the hamming ranking approach
should be used.
3
THE RIGHT WAY TO USE HASHING FOR
ANNS
The previous section gave a brief analysis of the pros and cons of
two search procedures using the hash index. In this section, we will
experimentally verify our analysis.
The experiments are performed on two popular ANNS datasets,
SIFT1M and GIST1M17 . The basic statistics of these two datasets
are shown in Table 1.
We use the random-projection-based LSH [4] for its simplicity.
Given a dataset with dimensionality m, one can generate an m ×
l random matrix from Gaussian distribution N (0, 1) where l is
the required code length. Then the data will be projected onto l
dimensional space using this random matrix. The l dimensional
real representation of each point can then be converted to l-bits
binary code by a simple binarize operation 18 . The formal algorithm
is illustrated in Algorithm (4).
Given a query point, the search algorithm is expected to return K
points. Then we need to examine how many points in this returned
set are among the real K nearest neighbors of the query. Suppose
the returned set of K points given a query is R and the real K nearest
neighbors set of the query is R ′ , the precision and recall can be
defined [29] as
|R ′ ∩ R|
|R ′ ∩ R|
, recall(R) =
.
(1)
precision(R) =
|R|
|R ′ |
17 http://corpus-texmex.irisa.fr/
18 i -th
bit is 1 if i -th feature is grater or equal to 0; i -th bit is 0 if i -th feature is smaller
than 0;
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Deng Cai
80
LSH(16)
LSH(20)
LSH(24)
LSH(28)
LSH(32)
40
20
0
5
10
# of located samples
LSH(24)
LSH(28)
LSH(32)
LSH(36)
LSH(40)
40
20
0.5
15
×10
Time (s)
Recall (%)
1
4
1.5
2
2.5
# of located samples
(a) SIFT1M
×10
200
300
100
200
100
0
0
3
16 bit 20 bit 24 bit 28 bit 32 bit
16 bit 20 bit 24 bit 28 bit 32 bit
Hamming ranking approach
Hash bucket search approach
(b) GIST1M
(a) Locate 50000 points on SIFT1M
80
300
300
locating time
scanning time
20
60
LSH(24)
LSH(28)
LSH(32)
LSH(36)
LSH(40)
40
20
100
200
300
400
500
100
150
Time (s)
(c) Hamming ranking on SIFT1M
200
100
0
50
Time (s)
Time (s)
LSH(16)
LSH(20)
LSH(24)
LSH(28)
LSH(32)
40
Recall (%)
60
locating time
scanning time
5
80
Recall (%)
locating time
scanning time
(d) Hamming ranking on GIST1M
locating time
scanning time
Time (s)
Recall (%)
60
60
Time (s)
300
80
200
100
0
24 bit 28 bit 32 bit 36 bit 40 bit
24 bit 28 bit 32 bit 36 bit 40 bit
Hamming ranking approach
Hash bucket search approach
(b) Locate 50000 points on GIST1M
80
60
LSH(16)
LSH(20)
LSH(24)
LSH(28)
LSH(32)
40
20
10 1
10 2
10 3
Time (s)
(e) Hash bucket search on SIFT1M
Recall (%)
Recall (%)
80
60
LSH(24)
LSH(28)
LSH(32)
LSH(36)
LSH(40)
Figure 2: Where the search time spend on? The percentages
of the locating time (the scanning time) in the total search
time varies as the code length increases when we are fixing
to locate 50,000 points on two datasets.
40
20
10 1
10 2
10 3
Time (s)
(f) Hash bucket search on GIST1M
Figure 1: The ANN search results of LSH on two datasets. The
y-axis denotes the average recall while the x-axis denotes
(a,b) the # of located samples (c,d) the search time using the
hamming ranking approach (e,f) the search time using the
hash bucket search approach. We can see that if the hamming ranking approach is used, longer codes means better
performance. While if the hash bucket search approach is
used, 24bits and 28bits are the optimal code lengths for LSH
on SIFT1M and GIST1M respectively.
Since the sizes of R ′ and R are the same, the recall and the precision
of R are actually the same. We fixed K = 100 throughout our
experiments.
3.1
The Search Time against # of Located
Samples
The ultimate goal of an ANNS method is returning the most accurate answers (nearest neighbors) in the shortest time. Thus, the
search time must be reported together with the accuracy measures.
Naturally, the accuracy - time curve is an excellent choice to evaluate the performance of various hashing algorithms. As analyzed
in Section 2, the search time should include coding time, locating
time and scanning time altogether.
Some hashing papers[16, 26–28] reported the training and testing time of various hashing algorithms. In an ANNS system, the
training time of a hashing algorithm is just the indexing time, while
the testing time of a hashing algorithm is merely the coding time.
For almost all the popular hashing algorithms, this coding time
can be neglected comparing with locating time and scanning time.
Please see Table 5 for details. Thus, in our remaining experiments,
the search time only includes locating time and scanning time.
Many hashing papers [7, 9–12, 15, 19, 23, 27, 28, 31, 33, 37, 38,
40] reported accuracy - # of located samples curves of various
hashing algorithms. If the hamming ranking approach is used, one
can easily find that various hashing algorithms almost have the
same locating time with the equal L (# of located samples) and the
same code length. The equal L also means the same scanning time.
Thus, # of located samples is a good indicator of the search time, and
the accuracy - # of located samples curve can be used to compare
various hashing algorithms correctly. Moreover, since the locating
time grows slowly as the code length grows (see Figure 3), accuracy
- # of located samples curve can also be used to measure the
performance of the same hashing algorithm with different code
length as long as these code lengths are close to each other.
However, if the hash bucket search approach is used, # of located
samples is no longer a good indicator of the search time. The locating time grows exponentially as the code length grows. Even with
the same code length, different hashing algorithms have different
code distributions which result in different locating times.
Take Figure 1 as an example. Figure 1 plots the performance of
LSH on both two datasets. The code length grows from 16 to 32 on
SIFT1M and 24 to 40 on GIST1M. Figure 1 (a,b) show the recall # of located samples curves, (c,d) show the recall - time curves
A Revisit of Hashing Algorithms for Approximate Nearest Neighbor Search
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Table 2: Hamming radius VS. # of queries successfully located 50000 points on SIFT1M (total # of queries 10,000)
hamming radius
0
1
2
3
4
5
6
7
8
9
10
# of buckets (16 bits)
# of queries (16 bits) (1.752s)
1
0
16
3637
120
4932
560
1230
1,820
169
4,368
30
8,008
1
11,440
1
12,870
0
11,440
0
8,008
0
# of buckets (32 bits)
# of queries (32 bits) (237.8s)
1
0
32
0
496
0
4960
4
35,960
1710
201,376
4282
906,192
2691
3,365,856
1021
10,518,300
241
28,048,800
44
64,512,240
7
Table 3: Hamming radius VS. # of queries successfully located 50000 points on GIST1M (total # of queries 1,000)
hamming radius
0
1
2
3
4
5
6
7
8
9
10
# of buckets (24 bits)
# of queries (24 bits) (0.234s)
1
70
24
467
276
277
2,024
115
10,626
44
42,504
18
134,596
7
346,104
2
735,471
0
1,307,504
0
1,961,256
0
# of buckets (40 bits)
# of queries (40 bits) (240.0s)
1
0
40
0
780
27
9,880
252
91,390
302
658,008
205
3,838,380
107
18,643,560
57
76,904,685
27
273,438,880
14
847,660,528
9
10 3
150
100
50
20
24
28
10 1
10 0
0
16
10 2
32
1500
1000
500
0
16
20
Code length
24
28
32
Code length
0
5
10
Located samples
(a) Locate 50000 points on SIFT1M
15
×10 4
(a) 32 bits LSH code on SIFT1M
hash bucket search
hamming ranking
1500
1000
500
0
0.5
1
1.5
2
Located samples
2.5
3
×10 5
(b) 40 bits LSH code on GIST1M
10 4
250
hash bucket search
hamming ranking
200
Locating time (s)
Locating time (s)
2000
hash bucket search
hamming ranking
Locating time (s)
200
2000
hash bucket search
hamming ranking
Locating time (s)
hash bucket search
hamming ranking
Locating time (s)
Locating time (s)
250
150
100
50
0
24
28
32
36
40
Figure 4: The locating time grows linearly as the number
of locating sample increases for the hamming ranking approach while the locating time grows super-linearly as the
number of locating sample increases for the hash bucket
search approach.
hash bucket search
hamming ranking
10 2
10 0
10 -2
24
28
Code length
32
36
40
Code length
(b) Locate 50000 points on GIST1M
Figure 3: The locating time grows linearly as the code length
increases for the hamming ranking approach while the locating time grows exponentially as the code length increases
for the hash bucket search approach.
using the hamming ranking approach and (e,f) show the recall time curves using the hash bucket search approach. We can see
that (a, c) and (b, d) have almost the same curves while (a, e) and
(b, f) are significantly different. If the hamming ranking approach
is used, longer codes (not very long) means better performance.
While if the hash bucket search approach is used, 24-bits and 28bits are the optimal code lengths for LSH on SIFT1M and GIST1M
respectively. The hash bucket search approach with LSH 40-bits
code (actually not very long) on GIST1M is extremely slow.
Figure 1 (a,b) is very natural and reasonable. Since each bit is
a partition of the feature space and same code (0 or 1) on this bit
means two points are on the same side of this partition. If two
points share t bits same code, which means these two points are
on the same sides of t partitions. Neighbors in the hamming space
with longer code are more likely to be the real neighbors in the
original feature space.
Figure 2 shows how the percentages of the locating time (the
scanning time) in the total search time varies as the code length
increases when we are fixing to locate 50,000 points on two datasets.
Since the number of locating points is fixed, the scanning time will
be fixed. The overall search time increasing is caused by the locating
time. The locating time of the hamming ranking approach grows
slowly as the code length grows while the locating time of the hash
bucket search approach grows very quickly.
Figure 3 shows that the locating time grows linearly as the code
length increases for the hamming ranking approach while the locating time grows exponentially as the code length increases for the
hash bucket search approach. This result agrees with our analysis
in Section 2.
Table 2 and 3 explain the reason. These tables recorded the number of queries (the total number is 10,000 on SIFT1M and 1,000 on
GIST1M) which successfully located 50,000 samples in different
hamming radius with different code length utilizing hash bucket
search approach. As the analysis in Section 2, as the code length
increases, the number of hash buckets corresponding to the same
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Deng Cai
100
100
(1 table)
(32bit)
(2 table)
(64bit)
(4 table)
(128bit)
(8 table)
(256bit)
(16 table)
(512bit)
90
90
Recall (%)
(1 table)
(32bit)
(2 table)
(64bit)
(4 table)
(128bit)
(8 table)
(256bit)
(16 table)
(512bit)
70
60
50
40
Recall (%)
80
80
70
60
50
40
30
20
10 2
10 3
20
50
100
200
Time (s)
Time (s)
(a) SIFT1M
(b) GIST1M
Figure 5: The recall-time curves of 1) (black lines) the hash bucket search approach with different number of tables 2) (red
lines) the hamming ranking approach with different code length on SIFT1M and GIST1M.
100
80
(1 table)
(2 table)
(4 table)
(8 table)
(16 table)
60
Recall (%)
80
Recall (%)
3.2
90
70
60
(1 table)
(2 table)
(4 table)
(8 table)
(16 table)
50
40
30
40
10
2
10
3
Time (s)
(a) Locate 50000 points on SIFT1M
20
50
100
Time (s)
(b) Locate 50000 points on GIST1M
Figure 6: The recall-time curves of the hash bucket search
approach with different number of tables on SIFT1M and
GIST1M.
hamming radius increases almost exponentially. Moreover, as the
code length increases, the base points are more diversified. And we
need to increase the hamming radius to locate enough number of
samples. These two reasons cause the locating time grows exponentially as the code length increases for the hash bucket search
approach. As a result, the hash bucket search approach has no
practical meaning if the code length is longer than 64.
When the code length is fixed, Figure 4 shows that the locating time of the hamming ranking approach grows linearly as the
number of locating sample increases while the locating time of the
hash bucket search approach grows super-linearly as the number of
locating sample increases. And the slope of the hamming ranking
approach is much smaller than that of the hash bucket search approach. It just because as the number of locating sample increases,
we need to increase the hamming radius to locate enough samples
successfully. Thus, we need to visit much more hash buckets.
The Long Codes against The Multiple
Tables
Our analysis in the previous section indicates that the hash bucket
search approach has no practical meaning if the code length is
longer than 64 since the locating time grows exponentially as the
code length increases. Can we design a method for the hash bucket
search approach to take the advantages of long codes?
The answer is yes. A straightforward way is using multiple hash
tables to represent long codes. Suppose we have 128-bits codes, if
a single table with 128-bits is used, the extremely long locating
time will make the hash bucket search approach meaningless. If
we use four tables, each table is 32-bits. To locate L points, if we
only have one hash table, we have to increase the hamming radius
r if the points in all the buckets within the hamming distance r are
not enough. This will increase the locating time a lot. If we have
multiple hash tables, instead of growing r , we can scan the buckets
within the hamming radius r in all the tables, which gives us a more
substantial chance to locate enough data points.
Figure 6 shows the recall-time curves of LSH with the different number of tables (each table is 32-bits) on the two datasets.
The results are very consistent: LSH gains the advantage by using
multiple tables on both datasets. This multiple tables trick is more
effective on SIFT1M than on GIST1M.
If we use four 32-bits tables, actually we used 128-bits codes. A
natural question will be which approach is more efficient? The hash
bucket search approach with four 32-bits tables? Or the hamming
ranking approach with 128-bits codes?
Figure 5 shows the recall-time curves of hash bucket search
approach with the different number of tables and that of hamming
ranking approach with different code length, both on SIFT1M and
GIST1M. If we aim at high recall region, the hamming ranking
approach is much more efficient. However, the hash bucket search
approach can return results (with low recall) in a very short of
time. If the hamming ranking approach is used, the computation of
the hamming distances of the query code and the database codes
A Revisit of Hashing Algorithms for Approximate Nearest Neighbor Search
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Table 4: Indexing Time (s)
Method SIFT1M GIST1M
Method SIFT1M
GIST1M
LSH
3.35
12.33
SH
83.44
3864.3
BRE
680.3
369.49
IsoH
1.68
91.99
USPLH
1099.2
10508.3
KLSH
94.09
194.89
ITQ
88.3
1372.4
AGH1
296.84
389.91
SpH
2583.5
4041.1
AGH2
611.7
703.25
DSH
51.92
122.84
CH
263.78
365.34
ITQ, SH and IsoH learn 128 bits code on SIFT1M and 960 bits code
on GIST1M.
is unavoidable. Thus it is not possible for the hamming ranking
approach to return the answer (even not very accurate) in a very
short of time.
As summarized in Section 2, the hash bucket search with multiple tables approach is more suitable for the case that L is small
(e.g., near-duplicate search). And the hamming ranking approach is
more suitable for the case L will be pretty large (e.g., high accuracy
required ANNS).
4
A COMPREHENSIVE COMPARISON
Our analysis in the previous sections shows that the evaluations
in most of the existing hashing papers are not thorough enough
and the conclusions in those papers need to be re-examined. In this
section, we aim at providing a comprehensive comparison between
twelve popular hashing algorithms on SIFT1M and GIST1M.
4.1
Compared Algorithms
Twelve popular hashing algorithms compared in the experiments
are listed as follows:
• LSH is a short name for random-projection-based Locality
Sensitive Hashing [4] as in Algorithm 4. It is frequently used
as a baseline method in various hashing papers.
• SH is a short name for Spectral Hashing [36]. SH is based on
quantizing the values of analytical eigenfunctions computed
along PCA directions of the data.
• KLSH is a short name for Kernelized Locality Sensitive Hashing [23]. KLSH generalizes the LSH method to the kernel
space.
• BRE is a short name for Binary Reconstructive Embeddings
[22].
• USPLH is a short name for Unsupervised Sequential Projection Learning Hashing [33].
• ITQ is a short name for ITerative Quantization [9]. ITQ
finds a rotation of zero-centered data so as to minimize the
quantization error of mapping this data to the vertices of a
zero-centered binary hypercube.
• AGH is a short name for Anchor Graph Hashing [28]. It
aims at performing spectral analysis [2] of the data which
shares the same goal with Self-taught Hashing [39]. The
advantage of AGH over Self-taught Hashing is the computational efficiency. AGH uses an anchor graph [25] to speed
up the spectral analysis. There are two versions of AGH: one
layer AGH and two layer AGH. We use AGH1 and AGH2
to denote them.
Table 5: Coding Time (s)
Method
LSH
BRE
USPLH
ITQ
SpH
DSH
SIFT1M
0.03
0.25
0.05
0.01
0.11
0.04
GIST1M
0.016
0.043
0.02
0.025
0.027
0.02
Method
SH
IsoH
KLSH
AGH1
AGH2
CH
SIFT1M
1.21
0.007
0.34
0.86
1.06
0.96
GIST1M
6.82
0.034
0.06
0.11
0.22
0.10
• SpH is a short name for Spherical Hashing [12]. SpH uses
a hyperspherebased hash function to map data points into
binary codes.
• IsoH is a short name for Isotropic Hashing [21]. IsoH learns
the projection functions with isotropic variances for PCA
projected data. The main motivation of IsoH is that PCA directions with different variance should not be equally treated
(one bit for one direction).
• CH is a short name for Compressed Hashing [24]. CH first
learns a landmark (anchor) based sparse representation [5]
then followed by LSH.
• DSH is a short name for Density Sensitive Hashing [16]. DSH
finds the projections which aware the density distribution
of the data.
Since our problem is high accuracy required ANNS, we use the
hamming ranking approach for all the hashing algorithms. For SH,
ITQ, and IsoH, the eigenvectors of the covariance matrix will be
used. Thus, these three algorithms can only learn 128-bits code on
SIFT1M and 960-bits code on GIST1M. All the other nine algorithms
learn 1,024-bits code on both datasets.
For KLSH, AGH and CH, we need to pick a certain number of
anchor (landmark) points. We use the same 1,500 anchor points for
all these algorithms, and the anchor points are generated by using
kmeans with five iterations. Also, the number of nearest anchors is
set to 50 for AGH and CH algorithms. Please see [28] for details.
All the hashing algorithms are implemented in Matlab19 , and we
use these hashing functions to learn the binary codes for both base
vectors and query vectors. The binary codes can then be fed into
Algorithm 1 for search. The Matlab functions are run on an i7-5930K
CPU with 128G memory, and the Algorithm 1 is implemented in
c++ and run on an i7-4790K CPU and 32G memory. For the sake of
reproducibility, both the Matlab codes and c++ codes are released
on GitHub2021 .
Besides all the hashing algorithms, we also report the performance of brute-force search.
• brute-force. The performance of brute-force search is reported to show the advantages of using hashing methods. To
get a different recall, we just perform brute-force search on
a different percentage of the query number. For example, the
brute-force search time on 90% queries of the original query
set stands for the brute-force search time of 90% recall.
19 Almost
all the core parts of the matlab code of the hashing algorithms are written
by the original authors of the papers.
20 https://github.com/ZJULearning/MatlabFunc/tree/master/ANNS/Hashing
21 https://github.com/ZJULearning/hashingSearch
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
Deng Cai
claim in all these hashing papers since they never report the
performance using longer codes.
• If the locating time is ignored, the performance of most of the
algorithms increases as the code length increases. However,
on SIFT1M, the recall of USPLH decreases as the code length
exceeds 256 and the recalls of AGH1 and AGH2 decrease
as the code length exceeds 512. On GIST1M, the recall of
USPLH almost fixed as the code length increases and the
recalls of AGH1, AGH2, SH and IsoH decrease as the code
length exceeds 512. This result suggests the five hashing
algorithms, USPLH, AGH1, AGH2, SH, and IsoH may not be
able to learn very long discriminative codes.
100
90
95
70
LSH
SH
BRE
USPLH
ITQ
SpH
60
50
40
32
IsoH
DSH
KLSH
AGH1
AGH2
CH
64
Recall(%)
Recall(%)
80
90
85
80
75
128
LSH
BRE
USPLH
256
SpH
DSH
KLSH
512
Code length
AGH1
AGH2
CH
1024
Code length
(a) Locate 10000 samples on SIFT1M
100
90
95
70
Recall(%)
Recall(%)
80
60
50
90
85
80
40
75
30
20
32
64
128
70
256
512
Code length
LSH
SH
BRE
USPLH
ITQ
SpH
IsoH
DSH
1024
Code length
KLSH
AGH1
AGH2
CH
LSH
SH
BRE
USPLH
ITQ
SpH
IsoH
DSH
KLSH
AGH1
AGH2
CH
(b) Locate 50000 samples on GIST1M
Figure 7: The recall - codelength curves of various hashing algorithms when the number of the located samples are
fixed
4.2
Results
Since we use Matlab functions to learn the binary codes and c++
algorithm to search with a hash index, the coding time and later
time (locating time and scanning time) cannot be added together,
and we reported them separately.
Table 4 and 5 report the indexing (hash learning) time and coding
(testing) time of various hashing algorithms. The indexing (hash
learning) stage is performed off-line thus the indexing time is not
very crucial. The coding time of all the hashing algorithms is short.
Therefore the coding time can be ignored compared with locating
time and scanning time (see Figure 8) when we consider the total
search time.
Figure 7 shows how the recall changes as the code length increases of various hashing algorithms when the number of the located samples are fixed at 10,000 on SIFT1M and 50,000 on GIST1M.
This figure illustrates several interesting points:
• When the code length is 32, all of the other algorithms outperform LSH on both two datasets. When the code length
is 64, nine algorithms outperform LSH on SIFT1M, and ten
algorithms outperform LSH on GIST1M. However, when the
code length exceeds 512, LSH performs the best on SIFT1M
and the 3rd best on GIST1M. This result is consistent with
the finding in [18] that "many data-dependent hashing algorithms have the improvements over LSH, but improvements
occur only for relatively small hash code sizes up to 64 or 128
bits". This result is also consistent with the "outperform LSH"
As we discussed in Section 2, if the hamming ranking approach
is used, the memory usage for hashing methods’ index is extremely
small. Meanwhile, if the recall-time curve is used, the additional
cost of locating stage for longer codes automatically be considered.
Thus, we are ready to pick the best hashing algorithm in the case
that the code length can reach to 1,024 bits.
Figure 8 shows the recall-time curves of various hashing algorithms with the optimal code length on SIFT1M and GIST1M. We are
very surprised to find there is only one algorithm, Spherical Hashing, can consistently outperform LSH on both two datasets. KLSH
is superior to LSH on GIST1M but performs very bad on SIFT1M.
Despite the extreme simplicity of random-projection-based LSH,
our results show that the capability of this algorithm has been far
underestimated.
5
COMPARISON WITH OTHER POPULAR
ANNS METHODS
Given the surprising performance of random-projection-based LSH,
we want to compare it with some public popular ANNS software.
To be fair, we convert our Matlab code for random-projection-based
LSH to c++ code and make a complete search software (by c++).
We named our algorithm RPLSH. The codes are also released on
GitHub22 .
5.1
Compared Algorithms
We pick five popular ANNS algorithms for comparison which cover
various types such as tree-based, hashing-based, quantization-based
and graph-based approaches. They are:
(1) FALCONN23 is a wel-known ANNS library based on multiprobe locality sensitive hashing. It implements the hash
bucket search with multiple tables approach.
(2) flann is a well-known ANNS library based on trees [30]. It
integrates Randomized KD-tree [32] and Kmeans tree [6]. In
our experiment, we use the randomized KD-tree and set the
trees number as 32 on SIFT1M and 64 on GIST1M. The code
on GitHub can be found at24 .
(3) annoy is based on a binary search forest index. We use their
code on GitHub for comparison25 .
22 https://github.com/ZJULearning/RPLSH
23 https://github.com/FALCONN-LIB/FALCONN
24 https://github.com/mariusmuja/flann
25 https://github.com/spotify/annoy
100
100
98
98
96
96
94
94
Recall (%)
Recall (%)
A Revisit of Hashing Algorithms for Approximate Nearest Neighbor Search
92
90
92
90
88
88
86
86
84
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
84
60
100
150
200
300
400
30
50
100
Time (s)
brute-force
LSH(256)
SH(128)
BRE(256)
USPLH(128)
ITQ(128)
SpH(256)
IsoH(128)
DSH(256)
KLSH(256)
200
300
Time (s)
AGH1(256)
AGH2(256)
CH(256)
brute-force
LSH(1024)
SH(512)
BRE(512)
USPLH(64)
(a) SIFT1M
ITQ(960)
SpH(1024)
IsoH(256)
DSH(1024)
KLSH(1024)
AGH1(256)
AGH2(256)
CH(512)
(b) GIST1M
Figure 8: The recall-time curves of various hashing algorithms with the optimal code length on SIFT1M and GIST1M. The
number in parenthesis after the algorithm name is the code length.
(4) faiss is recently released26 by Facebook. It contains well
implemented code for state-of-the-art product quantization
[14] based methods both on CPU and GPU. The CPU version
is used for a fair comparison.
(5) KGraph has open source code GitHub at27 , which is based
on an approximate kNN Graph.
• Both the RPLSH and the FALCONN are LSH based algorithms. However, RPLSH significantly outperforms FALCONN when the recall is higher than 90% on both two
datasets. This is because FALCONN uses the hash buckets search with multiple tables approach while RPLSH uses
the hamming ranking approach.
• flann and annoy are tree-based methods. Compare the performance of two hashing-based methods with that of two
tree-based methods, we can find that tree-based methods are
more suitable for low dimensional data.
• faiss is a product quantization (PQ) [14] based method, and
it uses binary codes to approximate the original features.
Thus it is impossible for faiss to achieve a very high recall.
The advantage of this approximation is that we do not need
the original vectors which is a huge saving of the memory
usage [14].
• KGraph is a graph based algorithm, and it is KGraph’s author who says "LSH is so hopelessly slow and/or inaccurate"28 . However, RPLSH is significantly better than KGraph
on GIST1M when at the high recall region (higher than 96%)
• The different performance of RPLSH on SIFT1M and GIST1M
suggest RPLSH is more suitable for high dimensional data.
The experiments are carried out on a machine with an i7-4790k
CPU and 32G memory. The performance on the high precision
region for an ANNS algorithm will be more crucial in real applications. Thus we sample one percentage points from each base set
as its corresponding validation set and tune all the algorithms on
the validation sets to get their best performing indices at the high
precision region. For RPLSH, we use 512-bits code on SIFT1M and
1,024-bits code on GIST1M. For all the search experiments, we only
evaluate the algorithms on a single thread.
5.2
Results
The recall-time curves of the six compared ANNS methods are
shown in Figure 9. A number of interesting conclusions can be
made:
• On SIFT1M, if a high recall is required (higher than 98%),
RPLSH performs the second best among all the six compared methods. On GIST1M, if a high recall is required
(higher than 96%), RPLSH performs the best among all the
six compared methods.
6
CONCLUSIONS
We carefully studied the problem of using hashing algorithms for
ANNS. With careful implementation, random-projection-based LSH
is superior to many other popular hashing algorithms, and this
algorithm is extremely simple. All the codes used in the paper are
26 https://github.com/facebookresearch/faiss
27 https://github.com/aaalgo/kgraph
28 http://www.kgraph.org/
Deng Cai
100
100
98
98
96
94
RPLSH
FALCONN
flann
annoy
faiss
KGraph
92
90
Recall (%)
Recall (%)
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
RPLSH
FALCONN
flann
annoy
faiss
KGraph
96
94
92
90
10 1
10 2
10 3
Time (s)
(a) SIFT1M
10 1
10 2
10 3
Time (s)
(b) GIST1M
Figure 9: The recall-time curves of six popular ANNS methods on SIFT1M and GIST1M.
released on GitHub, which can be used as a testing platform for a
fair comparison between various hashing methods.
REFERENCES
[1] Sunil Arya and David M. Mount. 1993. Approximate Nearest Neighbor Queries
in Fixed Dimensions. In Proceedings of the Fourth Annual ACM/SIGACT-SIAM
Symposium on Discrete Algorithms.
[2] M. Belkin and P. Niyogi. 2001. Laplacian Eigenmaps and Spectral Techniques for
Embedding and Clustering. In Advances in Neural Information Processing Systems
14.
[3] Jon Louis Bentley. 1975. Multidimensional Binary Search Trees Used for Associative Searching. Commun. ACM 18, 9 (1975), 509–517.
[4] Moses S. Charikar. 2002. Similarity Estimation Techniques from Rounding Algorithms. In Proceedings of the Thiry-fourth Annual ACM Symposium on Theory of
Computing. 380–388.
[5] Xinlei Chen and Deng Cai. 2011. Large Scale Spectral Clustering with LandmarkBased Representation. In Proceedings of the Twenty-Fifth AAAI Conference on
Artificial Intelligence.
[6] Keinosuke Fukunaga and Patrenahalli M. Narendra. 1975. A Branch and Bound
Algorithm for Computing k-Nearest Neighbors. IEEE Trans. Comput. 100, 7 (1975),
750–753.
[7] Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. 2013. Optimized Product
Quantization for Approximate Nearest Neighbor Search. In 2013 IEEE Conference
on Computer Vision and Pattern Recognition.
[8] Aristides Gionis, Piotr Indyk, and Rajeev Motwani. 1999. Similarity Search in
High Dimensions via Hashing. In Proceedings of the 25th International Conference
on Very Large Data Bases.
[9] Yunchao Gong and Svetlana Lazebnik. 2011. Iterative quantization: A procrustean
approach to learning binary codes. In The 24th IEEE Conference on Computer
Vision and Pattern Recognition.
[10] Junfeng He, Wei Liu, and Shih-Fu Chang. 2010. Scalable similarity search with
optimized kernel hashing. In Proceedings of the 16th ACM International Conference
on Knowledge Discovery and Data Mining.
[11] Kaiming He, Fang Wen, and Jian Sun. 2013. K-Means Hashing: An AffinityPreserving Quantization Method for Learning Binary Compact Codes. In 2013
IEEE Conference on Computer Vision and Pattern Recognition.
[12] Jae-Pil Heo, Youngwoon Lee, Junfeng He, Shih-Fu Chang, and Sung-Eui Yoon.
2012. Spherical hashing. In 2012 IEEE Conference on Computer Vision and Pattern
Recognition.
[13] Piotr Indyk and Rajeev Motwani. 1998. Approximate Nearest Neighbors: Towards
Removing the Curse of Dimensionality. In Proceedings of the Thirtieth Annual
ACM Symposium on the Theory of Computing.
[14] Hervé Jégou, Matthijs Douze, and Cordelia Schmid. 2011. Product Quantization
for Nearest Neighbor Search. IEEE Trans. Pattern Anal. Mach. Intell. 33, 1 (2011),
117–128.
[15] Zhongming Jin, Yao Hu, Yue Lin, Debing Zhang, Shiding Lin, Deng Cai, and
Xuelong Li. 2013. Complementary Projection Hashing. In IEEE International
Conference on Computer Vision.
[16] Zhongming Jin, Cheng Li, Yue Lin, and Deng Cai. 2014. Density Sensitive Hashing.
IEEE Trans. Cybernetics 44, 8 (2014), 1362–1371.
[17] Zhongming Jin, Debing Zhang, Yao Hu, Shiding Lin, Deng Cai, and Xiaofei. He.
2014. Fast and Accurate Hashing Via Iterative Nearest Neighbors Expansion.
IEEE transactions on cybernetics 44, 11 (2014), 2167–2177.
[18] Alexis Joly and Olivier Buisson. 2011. Random maximum margin hashing. In The
24th IEEE Conference on Computer Vision and Pattern Recognition.
[19] Yannis Kalantidis and Yannis S. Avrithis. 2014. Locally Optimized Product Quantization for Approximate Nearest Neighbor Search. In 2014 IEEE Conference on
Computer Vision and Pattern Recognition.
[20] Jon M. Kleinberg. 1997. Two Algorithms for Nearest-Neighbor Search in High
Dimensions. In Proceedings of the Twenty-Ninth Annual ACM Symposium on the
Theory of Computing.
[21] Weihao Kong and Wu-Jun Li. 2012. Isotropic Hashing. In Advances in Neural
Information Processing Systems 25.
[22] Brian Kulis and Trevor Darrell. 2009. Learning to Hash with Binary Reconstructive Embeddings. In Advances in Neural Information Processing Systems 22,
NIPS.
[23] Brian Kulis and Kristen Grauman. 2009. Kernelized locality-sensitive hashing for
scalable image search. In IEEE 12th International Conference on Computer Vision.
[24] Yue Lin, Rong Jin, Deng Cai, Shuicheng Yan, and Xuelong Li. 2013. Compressed
Hashing. In 2013 IEEE Conference on Computer Vision and Pattern Recognition.
[25] Wei Liu, Junfeng He, and Shih-Fu Chang. 2010. Large graph construction for scalable semi-supervised learning. In Proceedings of the 27th International Conference
on Machine Learning.
[26] Wei Liu, Cun Mu, Sanjiv Kumar, and Shih-Fu Chang. 2014. Discrete Graph
Hashing. In Advances in Neural Information Processing Systems 27.
[27] Wei Liu, Jun Wang, Rongrong Ji, Yu-Gang Jiang, and Shih-Fu Chang. 2012. Supervised hashing with kernels. In 2012 IEEE Conference on Computer Vision and
Pattern Recognition.
[28] Wei Liu, Jun Wang, Sanjiv Kumar, and Shih-Fu Chang. 2011. Hashing with
Graphs. In Proceedings of the 28th International Conference on Machine Learning.
[29] John Makhoul, Francis Kubala, Richard Schwartz, and Ralph Weischedel. 2000.
Performance Measures For Information Extraction. In Proceedings of DARPA
Broadcast News Workshop.
[30] Marius Muja and David G. Lowe. 2014. Scalable Nearest Neighbor Algorithms
for High Dimensional Data. IEEE Transactions on Pattern Analysis and Machine
Intelligence 36, 11 (2014), 2227–2240.
[31] Mohammad Norouzi and David J. Fleet. 2011. Minimal Loss Hashing for Compact
Binary Codes. In Proceedings of the 28th International Conference on Machine
Learning,ICML 2011.
[32] Chanop Silpa-Anan and Richard Hartley. 2008. Optimised KD-trees for fast image
descriptor matching. In Proceedings of the 2008 IEEE Conference on Computer Vision
and Pattern Recognition.
[33] Jun Wang, Sanjiv Kumar, and Shih-Fu Chang. 2010. Sequential Projection Learning for Hashing with Compact Codes. In Proceedings of the 27th International
Conference on Machine Learning.
[34] Henry S. Warren. 2012. Hacker’s Delight, Chapter 5. Addison-Wesley Professional.
A Revisit of Hashing Algorithms for Approximate Nearest Neighbor Search
[35] Yair Weiss, Rob Fergus, and Antonio Torralba. 2012. Multidimensional Spectral
Hashing. In 12th European Conference on Computer Vision.
[36] Yair Weiss, Antonio Torralba, and Robert Fergus. 2008. Spectral Hashing. In
Advances in Neural Information Processing Systems 21, NIPS.
[37] Bin Xu, Jiajun Bu, Yue Lin, Chun Chen, Xiaofei He, and Deng Cai. 2013. Harmonious Hashing. In Proceedings of the 23rd International Joint Conference on
Artificial Intelligence.
[38] Hao Xu, Jingdong Wang, Zhu Li, Gang Zeng, Shipeng Li, and Nenghai Yu. 2011.
Complementary hashing for approximate nearest neighbor search. In IEEE International Conference on Computer Vision.
[39] Dell Zhang, Jun Wang, Deng Cai, and Jinsong Lu. 2010. Self-taught hashing for
fast similarity search. In Proceeding of the 33rd International ACM SIGIR Conference
on Research and Development in Information Retrieval.
[40] Kang Zhao, Hongtao Lu, and Jincheng Mei. 2014. Locality Preserving Hashing.
In Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence.
arXiv, February 18, 2018, Hangzhou, Zhejiang, China
| 1cs.CV
|
OptLayer - Practical Constrained Optimization for
Deep Reinforcement Learning in the Real World
arXiv:1709.07643v2 [cs.RO] 23 Feb 2018
Tu-Hoa Pham1 , Giovanni De Magistris1 and Ryuki Tachibana1
Abstract— While deep reinforcement learning techniques
have recently produced considerable achievements on many
decision-making problems, their use in robotics has largely
been limited to simulated worlds or restricted motions, since
unconstrained trial-and-error interactions in the real world
can have undesirable consequences for the robot or its environment. To overcome such limitations, we propose a novel
reinforcement learning architecture, OptLayer, that takes as
inputs possibly unsafe actions predicted by a neural network
and outputs the closest actions that satisfy chosen constraints.
While learning control policies often requires carefully crafted
rewards and penalties while exploring the range of possible
actions, OptLayer ensures that only safe actions are actually
executed and unsafe predictions are penalized during training.
We demonstrate the effectiveness of our approach on robot
reaching tasks, both simulated and in the real world.
I. I NTRODUCTION
Over the recent years, the rise of deep neural network
architectures in multiple fields of science and engineering
has come together with considerable achievements for deep
reinforcement learning techniques, sometimes outperforming humans on high-dimensional problems such as Atari
games [1] and Go [2]. The robotics field has also greatly
benefitted from these advances, yielding impressive results
on tasks otherwise difficult to model explicitly, e.g., manipulation from visual inputs [3] and robust locomotion on
challenging terrains [4]. While visual understanding neural
network pipelines can often be trained in an end-to-end
fashion (e.g., inferring object poses directly from raw pixel
information), robot motion is typically initialized in the real
world from (partially) supervised learning, which supposes
the availability of ground-truth data or expert demonstrations. In contrast, the discovery of robot motion behaviours
from scratch is often limited to simulation due to time
(to learn basic task features), safety (e.g., collisions), and
other experimental constraints (e.g., resetting knocked-down
obstacles to a chosen state). However, transferring control
policies learned in simulation to reality remains a challenge,
in particular due to model uncertainties (kinodynamics) [5]
and variability between simulated and real observations [6].
Enabling deep reinforcement learning in the real world is
thus crucial for tasks that cannot be learned in simulation
and for which expert policies are not known.
In this paper, we propose a practical method to constrain
neural network predictions to lie within a domain defined by
safety constraints (see Fig. 1). Our work capitalizes on the
state of the art in multiple areas of robotic control, optimization and reinforcement learning (Section II). We summarize
1 IBM
Research AI, Tokyo, Japan.
(a) Unconstrained action predictions can be dangerous to run on a real robot.
(b) Explicit constraints can be incorporated to guarantee safe robot actions.
Fig. 1.
3D reaching with obstacle avoidance using reinforcement learning.
the necessary technical background and important challenges
for real-world reinforcement learning (Section III), in particular a strong dependency on the chosen reward structure.
• We augment a neural network architecture with a
constrained optimization layer, OptLayer, that enforces
arbitrary constraints on the predicted robot actions
(Section IV). OptLayer is fully differentiable, enabling
future end-to-end learning under safety constraints.
• We tackle the problem of safe reinforcement learning
using OptLayer and propose a reward strategy that
is simple to implement and readily compatible with
existing policy optimization techniques (Section V).
• Our approach makes exploration during training more
efficient while always satisfying safety constraints. We
demonstrate its effectiveness by learning 3D reaching
with collision avoidance on an industrial manipulator,
on tasks of increasing difficulty, both simulated and in
the real world (Section VI).
Finally, we discuss challenges we encountered, current limitations and future extensions of our work (Section VII).
II. R ELATED W ORK
Robot motion is traditionally achieved by solving a multiparameter optimization problem yielding a sequence of states
achieving a given task (e.g., track a given end effector
trajectory) while respecting constraints motivated by safety
and performance (e.g., joint limits). When constraints and
task objectives are respectively linear and quadratic in the
robot variables, these problems can be formulated and solved
as quadratic programs (QP) [7]. Still, the resulting control
policy can be unsatisfactory due to uncertainties on the robot
kinodynamic model and sensor feedback (e.g., during locomotion, errors on center of mass coordinates can invalidate
robot stability constraints), the chosen constraints being too
conservative (making the motion suboptimal), or the task
dynamics being non-trivial to model (e.g., finding the optimal
trajectory to follow).
Reinforcement learning (RL) techniques have demonstrated impressive results in this context. In [1], convolutional
neural networks (CNN) were trained to play Atari games
using only pixels and game score as inputs, achieving humanlevel performance on multiple games without need for expert demonstrations or feature engineering. A guided policy
search (GPS) method was proposed by [3] to predict motor
torques for a PR2 robot from image and joint state inputs
only, training the policy in a data-efficient manner using
supervised learning (SL) and local controllers trained by
RL. An overview and benchmark of modern RL algorithms
for continuous control (e.g., for robotic applications) was
presented in [8], along with valuable reference implementations released to the community. We thus chose to build
our work upon the top-performing benchmarked algorithm,
Trust-Region Policy Optimization (TRPO) [9], which we
detail further in Section III-A. TRPO was notably used
recently to discover dynamic locomotion behaviours such as
crouching and jumping on simulated humanoids [4].
Applied in the real world, RL techniques require special
attention to guarantee safe exploration during training, and
safe execution at inference time. A thorough survey of safe
RL was presented in [10], which classified techniques into
two categories: those incorporating safety terms in the policy
optimality criterion, and those modifying the exploration process using additional information. In particular, our work was
inspired by [11], which proposed a new trust region method,
Constrained Policy Optimization (CPO), to guarantee policy
performance improvement with near-constraint satisfaction at
each iteration, along with important results on convergence
speed and worst-case constraint violation. Still, real-world
robot experiments may require exact constraint satisfaction,
though possibly at the cost of less (time and data) efficient
training. Alternatively, constraints can be enforced on the
neural network predictions directly: a priori, e.g., by placing
limits on the range of possible actions before sampling [12];
or a posteriori, e.g., by ensuring that predictions do not send
the robot’s end effector outside of a chosen bounding sphere,
reprojecting it inside if needed [13]. Our approach extends
these ideas by reshaping the range of possible actions at each
time step to ensure that resulting predictions always satisfy
the exact safety constraints (not approximations thereof),
during both training and inference. Our work also benefitted
from [14], which proposed a neural network architecture
to solve QP optimization problems in the context of SL,
with applications to Sudoku puzzles. We extend this research
towards RL for real-world systems with physical constraints.
III. BACKGROUND AND M OTIVATION
In this section, we summarize the RL algorithm our work
builds upon, TRPO (Section III-A). We then consider the
motivating example of a 2D reaching task (Section III-B) and
illustrate the limitations of RL without explicit consideration
of constraints (Section III-C), namely that learning a safe
policy involves unsafe actions in the first place, with great
sensitivity to the associated reward and cost structure.
A. Reinforcement Learning Nomenclature
We consider an infinite-horizon discounted Markov
Decision Process (MDP) characterized by a tuple
(S, A, P, R, ρ0 , γ), with S the set of states (e.g., robot
joint configurations), A the set of actions (e.g., motor
commands), P : S × A × S → [0, 1] the transition
probability distribution to go from one state to another
by taking a specific action, R : S × A × S → R the
function of associated rewards, ρ0 : S → [0, 1] the initial
state probability distribution, γ ∈ [0, 1) a discount factor.
Denoting by π : S × A → [0, 1] a stochastic policy,
the objective is typically to maximize the discounted
∞
P
i
γ R(si , ai , si+1 ) , with
expected return η(π) = E
τ
i=0
τ = (s0 , a0 , . . . ) a state-action trajectory, s0 ∼ ρ0 ,
ai ∼ π(·|si ) and si+1 ∼ P (·|si , ai ). With the same
notations, we denote by Qπ the state-action value function,
Vπ the value function and Aπ the advantage function:
"∞
#
X
l
γ R(si+l , ai+l , si+l+1 ) ,
Qπ (si , ai ) =
E
si+1 ,ai+1 ,...
l=0
(1)
"
Vπ (si ) =
E
ai ,si+1 ,...
∞
X
#
γ l R(si+l , ai+l , si+l+1 ) , (2)
l=0
Aπ (s, a) = Qπ (s, a) − Vπ (s).
(3)
We consider in particular the case of a policy πθ parameterized by a vector θ (e.g., neural network parameters). In
TRPO, the policy πθk is iteratively refined by solving the
following optimization problem at each iteration k:
πθ (a|s)
Aπθk (s, a)
(4)
θk+1 = argmax
E
s∼ρθk ,a∼πθk πθk (a|s)
θ
such that
with ρθ (s) =
E [DKL (πθk (·|s)||πθ (·|s))] ≤ δKL ,
s∼ρθk
∞
P
i=0
(5)
γ i P (si = s) the discounted state-visitation
frequencies when actions are chosen according to πθ , DKL
the Kullback-Leibler divergence (measuring how much two
probability distributions differ from each other), and δKL
a step size controlling how much the policy is allowed to
change at each iteration. Eq. (4) is then solved numerically
by sampling state-action trajectories following πθk and averaging over samples. Our work builds on top of the OpenAI
Baselines reference implementation for TRPO1 , in which a
neural network N predicts state values jointly with actions.
In the rest of the paper, we refer to it as a procedure
U PDATE N ETWORK (N , (si , ai , ri , vi )i ) that iteratively optimizes N from sequences of state-action-reward-value tuples
(si , ai , ri , vi )i=0,...,N over finite horizon N .
end θelbow
effector
forearm
target
upper arm
θshoulder
obstacle
T YPE
rdist
rcoll
rprox
R EWARD
−dT
−20βcoll
if collision
else 0
2 if dT ≤ 3 cm
else 0
(a) 2-DoF robot, target and obstacle. (b) 2D-reaching reward structure.
Fig. 2. 2D reacher description and rewards, with dT the distance between
target and end-effector and βcoll ∈ {1, 5, 10, 50} a penalization coefficient.
B. Motivating Example
We consider a 2-DoF robot with parallel revolute joints.
The robot’s base link is rigidly linked to the world such that it
evolves in a plane (x, y) perpendicular to the gravity vector
−gz. Its two main links, upper arm and forearm, are of length
0.1 m. The task consists in reaching a target point with the
robot’s end effector while avoiding a spherical obstacle, both
on the plane (x, y). We denote by θ = (θshoulder , θelbow ) the
robot’s joint angles, with θshoulder not limited and θelbow ∈
[−π, +π], and by θ̇, θ̈ its joint velocities and accelerations,
respectively. We depict the 2D reacher environment, E2D , in
Fig. 2a. Each episode is initialized as follows:
1) The robot is reset to the initial state θ = θ̇ = θ̈ = 0.
2) The target’s coordinates are randomly sampled following a uniform distribution of ±0.27 cm along x and y
(possibly out of reach for the robot).
3) The obstacle is initialized following the same distribution. If it collides at initialization with the robot, its
position is re-sampled until it is no more the case.
We discretize time into N = 200 steps of duration ∆T =
0.01 s. The state vector si at step i, or time ti = i∆T , is:
• pfa the 3D position of the forearm’s base,
• θelbow the elbow joint angle,
• θ̇ the joint velocities,
• pee the 3D position of the end effector,
• pt −pee the target position pt relative to the end effector,
• po − pee the obstacle position po , also relative to pee .
The robot is controlled in position by providing, at step i, the
desired joint position at step i + 1. We define a zero-centered
action ai as the joint step ∆θi = θi+1 − θi to perform
between two consecutive time steps. Finally, executing ai in
the environment yields three rewards, rdist (reward on low
distance to target), rcoll (collision penalization), rprox (bonus
on proximity to target), detailed in Fig. 2b. We denote by ri
the total reward at step i. In order to assess the sensitivity
of RL to the reward structure, we parameterize the collision
reward rcoll with a coefficient βcoll ∈ {1, 5, 10, 50}.
C. Preliminary Results
We implement the environment E2D within the OpenAI
Gym and Roboschool framework [15], which provides a
unified interface to train and test neural network policies.
1 https://github.com/openai/baselines
We use the physics engine within Roboschool to monitor
external collisions, between any robot link and the obstacle,
and auto-collisions, between the end effector and the base.
Throughout this paper, we take as neural network N a simple
multilayer perceptron (MLP) with two hidden layers of size
32 each. At each step i, the environment produces a state
vector si , which is fed into N to produce an action and
value pair ai , vi . Executing ai in E2D results in a reward
ri and an updated state si+1 . We construct such sequences
(si , ai , ri , vi )i=0,...,N over episodes of up to N = 200
timesteps, interrupted in case of collision. On a 8-core
computer, we train N on four instances of E2D in parallel,
with the TRPO procedure U PDATE N ETWORK of Section IIIA. We report the evolution of the reward over 15000 episodes
of training, for four values of βcoll , along with the cumulated
number of collisions in Fig. 3. We observe the following:
• The reward evolution throughout training is greatly
dependent on the collision penalization weight.
• With βcoll = 1, the low penalty associated to collisions
rcoll = −20 results in collision avoidance never being
learnt. The optimal strategy consists in quickly going to
the target to get the proximity reward rprox and possibly
hitting the obstacle to start a new episode.
• Conversely, with βcoll = 50, the large penalty rcoll =
−1000 results in fewer collisions but discourages exploration. The optimal strategy then consists in staying
in the vicinity of the start pose and ignoring further
targets.
• βcoll = 5 and 10 (rcoll = −100 and −200, respectively)
are comparable in terms of rewards, but the latter results
in half the number of collisions. This confirms the
importance of carefully crafting the reward structure.
These results show that, without explicitly consideration
of safety constraints during RL, it is difficult to learn a
safe policy without violating such constraints many times.
Overall, while we believe it is ultimately possible to design a
reward structure that efficiently decreases collisions over time
(e.g., empirically or by inverse reinforcement learning [16]),
for the sake of RL in the real world, it is crucial to be able
to learn to satisfy constraints without ever violating them.
IV. C ONSTRAINED O PTIMIZATION L AYER
A. OptLayer Architecture
With x a vector of nx variables to optimize, we consider
quadratic programs (QP) of the form:
1 T
x Px + qT x
2
such that Gx ≤ h
min
x
and Ax = b.
(6)
(7)
(8)
Eq. (6) is a quadratic objective in x, with P a square
matrix and q a vector of respective size nx × nx and
nx ; Eq (7) a set of nin linear inequalities, with G and
h of respective size nin × nx and nin ; and Eq (8) a set
of neq linear equalities, with A and b of respective size
neq × nx and neq . Our goal is to ensure that the action
coll
Cumulated
collisions [10]
Reward [100]
4
=1
coll
4
=5
coll
4
= 10
2
2
2
2
0
0
0
0
2
2
2
2
4
4
4
4
6
6
6
6
0
10000
100
0
10000
20
50
0
10000
10
0
0
10000
Episode
0
0
Episode reward
10000
Episode
= 50
0
10000
0
10000
Episode
care is given to constraint matrices, which for robotic applications can be constant or affine in si , require auxiliary
state variables, or activate only under particular conditions.
Although only inequality constraints appear in the following
exposition, equality constraints are treated the same way.
1) Objective matrices: We minimize the (squared) L2
ei :
distance between optimization variables x and prediction a
1
1 T
1
ei ||2 = xT x − a
e a
eTi x + a
ei .
min ||x − a
x 2
2
2 i
0
0
10000
Episode
Reward moving average
Collision count
θ̇min ≤ θ̇ ≤ θ̇max ,
Fig. 3.
Reward and collision count for unconstrained 2D reaching task
with collision penalization coefficient βcoll ∈ {1, 5, 10, 50}, averaged over
three seeds. Moving average over 40 consecutive episode rewards.
v
Environment
s
Agent
Neural
network
World
Constraint
parameters
e
a
s
OptLayer
a∗
(a) Prediction-correction architecture.
OptLayer: min 12 xT Px + qT x such that Gx ≤ h and Ax = b
x
c
Compute constraint
violation cost
P, q, G, h, A, b
···
e
a
s
Build
constraint
Init.
matrices P,q,G optim.
h,A,b
var.
Optim.
iter. 1
∆x0
∆y0
∆λ0
Optim.
iter. n
∆µ0
+
···
∆xn−1
∆yn−1
∆λn−1
∆µn−1
+
x0 , y0 , λ0 , µ0 x1 , y1 , λ1 , µ1 xn−1 , yn−1 , λn−1 , µn−1 xn , yn , λn , µn
(b) OptLayer computes the safe action a∗ that is closest to neural network
e, while quantifying how much the latter violates safety constraints.
prediction a
Fig. 4.
OptLayer constrained optimization pipeline.
at predicted by a neural network N satisfies constraints
such as Eqs. (7) and (8). In the following, we denote by
ei the action predicted by the neural network N based on
a
the current state si . We propose a constrained optimization
layer, OptLayer, that computes the action a∗i that is closest
ei while also satisfying constraints depending on si . We
to a
depict the overall architecture in Fig. 4a. While multiple
solutions are already available for solving QPs alone (as a
forward pass), end-to-end neural network training requires
a backward pass, in particular computing the gradients of
the outputs with respect to the input parameters. We thus
construct objective and constraint matrices as differentiable
ei (Section IVexpressions of OptLayer’s inputs si and a
B). We then solve the QP with an interior point method
implemented as fully differentiable layers (Section IV-C).
B. Quadratic Program Initialization
Our first goal is to construct objective and constraint
ei and si . Special
matrices as differentiable expressions of a
(11)
Eq (11) thus defines a set of constraint matrices Gcst =
[Inx , −Inx ]T and hcst = [∆T θ̇max , −∆T θ̇min ]T that are
constant throughout the robot’s motion.
3) Constraint matrices affine in the state vector: Consider
now joint position limits θmin , θmax :
θmin ≤ θ ≤ θmax .
Select a∗
a∗ = xn
···
(10)
with θ̇min , θ̇max velocity limits specified by the manufacturer
(typically, θ̇min = −θ̇max ). At step i + 1, approximating joint
velocities θ̇i+1 with ∆θi /∆T yields the following inequalities on action hypothesis x = ∆θi , following Eq. (7):
[Inx ]x ≤ [∆T θ̇max ] and [−Inx ]x ≤ [−∆T θ̇min ].
c
e
a
(9)
eTi a
ei does not depend on x, it does not affect the
Since a
optimum. Eq. (9) can thus be written as Eq. (6), with P =
Inx the identity matrix of size nx × nx and q = −e
ai .
2) Constant constraint matrices: In general, the constraints that affect a robot can be static or depend on its state.
For the former, consider for example joint velocity limits:
5
10
0
coll
4
(12)
In this case, the range of possible actions depends on the
current joint positions. Since θi+1 = θi + x, we have:
[Inx ]x ≤ [θmax − θi ] and [−Inx ]x ≤ [θi − θmin ].
(13)
Since θi is a member of the state vector si , we can construct
a selection matrix Hθ (with only 0 and ±1) such that
[−θi , θi ]T = Hθ si . Eq. (13) defines a set of constraint
matrices that are constant or affine in si : Gaff = [Inx , −Inx ]T
and haff = Hθ si + [θmax , −θmin ]T .
4) Auxiliary state parameters: In some cases, constraint
matrices may not be formulated as a simple function of si .
Consider for example limits τmin , τmax on joint torques τ:
τmin ≤ τ ≤ τmax .
(14)
With q, q̇, q̈ the robot’s generalized coordinates (constant
base pose and joint angles θ), generalized velocities (zero
base velocities 06 and joint velocities θ̇) and generalized accelerations (zero base accelerations 06 and joint accelerations
θ̈), respectively, let H(q) be its mass matrix, C(q, q̇) the
bias vector comprising Coriolis, centrifugal and gravitational
forces, and Fbase the 6-element vector of the external forces
and torques exerted at the robot’s (fixed) base. In the absence
of other contact forces, the equations of motion are then:
Fbase
H(q)q̈ + C(q, q̇) =
.
(15)
τ
The first 6 rows of Eq. (15) allow the computation of Fbase .
We take Hτ (q) and Cτ (q, q̇) as the rows indexed 7 and
onwards in H(q) and C(q, q̇), respectively, such that:
06
τ = Hτ (q)q̈ + Cτ (q, q̇), with q̈ =
.
(16)
θ̈
At step i, estimating θ̈i by forward difference θ̈i = (θ̇i+1 −
θ̇i )/∆T and θ̇i+1 = (θi+1 − θi )/∆T = x/∆T yields:
θ̇i
x
−
.
2
∆T
∆T
Combining Eqs. (16) and (17) makes τi affine in x:
Hτ (qi ) 06
Hτ (qi ) 06
τi =
+
C
(q,
q̇)
−
.
τ
x
θ̇i
∆T 2
∆T
θ̈i =
(17)
(18)
Since Hτ (qi ) and Cτ (qi , q̇i ) can not be expressed as a trivial function (e.g., linear) of si in its current form, we adjust
the environment E to compute them internally and append
them to si . Explicitly, we compute Hτ (qi ) and Cτ (qi , q̇i )
from the robot’s joint states and physical parameters using
the KDL library [17] and store their elements as additional
elements of si . The joint torque constraints of Eq. (14) can
then be expressed following Eq. (7), with constraint matrices
Gaux and haux assembled from si (extended) directly.
5) Conditional constraint matrices: Finally, some constraints may be active only under some conditions. We
consider the collision avoidance constraints of [18]. We
denote by Ll a convex object associated to a robot link l, pl
the closest point between Ll and the environment, dl and nl
the corresponding distance and unit vector, oriented towards
pl . If dl is smaller than a threshold dM , named influence
distance, then the following constraint is enabled:
−ξ
dl − dm
≤ d˙l ,
dM − dm
(19)
with d˙l the distance velocity, dm a threshold security distance, and ξ a positive velocity damping coefficient. Intuitively, Eq. (19) ensures that the distance dl between the robot
and its environment cannot decrease faster than a chosen
rate while dm ≤ dl ≤ dM , and that dl strictly increases if
dl < dm . Eq. (19) can be equivalently written as a constraint
on q̇, with J(q, pl ) the Jacobian matrix of Ll at pl when
obstacles are static (otherwise it suffices to adjust d˙l with
their velocities):
dl − dm
−ξ
≤ nTl J(q, pl ) q̇.
(20)
dM − dm
Similarly to Section IV-B.4, the state vector si can be
extended with the elements of J(q, pl )T nl to form constraint
matrices Gcond and hcond . We also extend si with dl and
denote by Hdl the selection matrix such that Hdl si = dl .
With htest = [dM ] − Hdl si , conditional constraints and their
activation can be expressed following:
If {0 ≤ htest } , then enable {Gcond x ≤ hcond } .
(21)
Note that activation condition {0 ≤ htest } can only depend
on state si , the action x being optimized subsequently.
C. Quadratic Program Solving
1) Assembling constraint matrices for batch solving:
Using the notations of Section IV-B, we consider first a
set of individual constraint matrices that are always active:
Gcst , hcst , Gaff , haff , Gaux , haux . We concatenate them into
base constraint matrices Gbase = [Gcst , Gaff , Gaux ]T and
hbase = [hcst , haff , haux ]T to be considered at every step.
When solving only one QP at a time, conditional constraints
matrices Gcond , hcond can be appended to the complete constraint matrices G, h of Eq. (7) only when active, i.e.:
(
T
G = [Gbase , Gcond ]
If {0 ≤ htest } :
,
(22)
T
h = [hbase , hcond ]
(
G = Gbase
otherwise :
.
(23)
h = hbase
However, deep neural networks are commonly trained using mini-batches of multiple input-output pairs in parallel.
A technical consequence of parallel computing is that all
elements processed together must be the same size. With
Gsub (resp. hsub ) a submatrix of Gbase (resp. hbase ) of size
that of Gcond (resp. hcond ), we replace Eq. (23) with:
(
T
G = [Gbase , Gsub ]
.
(24)
If not {0 ≤ htest } :
T
h = [hbase , hsub ]
G and h being of constant size thus enables batch solving.
2) Iterative resolution: Provided objective and constraint
matrices P, q, G, h, A, b, we implement the interior point
method for solving QPs described in [7]. With x still
denoting the variables to optimize, λ slack variables for the
QP, µ and y dual variables respectively associated to the
inequality and equality constraints, finding the optimum x∗
amounts to solving a sequence of linear systems. First:
x
−q
P
GT
AT
G −Inin 0nin ,neq µ = h .
(25)
y
b
A 0neq ,nin 0neq ,neq
We then use the solution to Eq. (25) to set x0 = x, y0 = y,
and initialize λ0 and µ0 depending on the value of µ. We
thus obtain a starting point (x0 , λ0 , µ0 , y0 ). The QP is then
solved iteratively. For (xk , λk , µk , yk ) at iteration k, with
Λk = diag(λk ) and Mk = diag(µk ), update directions
are computed by solving linear systems of the form [7]:
∆xk
P
0nin ,nin
GT
AT
0nin ,nx
Mk
Λk
0nin ,neq
∆λk = dk , (26)
G
Inin
0nin ,nin 0nin ,neq ∆µk
A
0neq ,nin 0neq ,nin 0neq ,neq ∆yk
with dk a vector depending on the current iteration
(xk , λk , µk , yk ) and QP matrices P, q, G, h, A, b. [14]
showed that such problems could be efficiently solved on
the GPU and released a public implementation using the PyTorch framework. In our custom implementation, we solved
Eqs. (25) and (26) using linear system solvers provided
within the Tensorflow framework. We iteratively update
xk+1 = xk + ∆xk , λk+1 = λk + ∆λk , µk+1 = µk +
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
procedure B UILD T RAJ(E, N , O, constrained)
S, V, R ← (), ()
. States, values, rewards
e A∗ ← (), ()
A,
. Predicted, optimal actions
C ← ()
. Constraint violation costs
s = E.reset(); end = false
. Get initial state
while not end do
S.add(s)
. Store current state
e, v ← N (s)
a
. Predict action, value
e)
a∗ , c ← O(s, a
. Constrain prediction
e
A.add(e
a); V .add(v); A∗ .add(a∗ ); C.add(c)
if constrained then
s, r, end = E.do(a∗ ) . Respects constraints
else
s, r, end = E.do(e
a) . Can violate constraints
end if
R.add(r)
. Store reward
end while
e R, V, A∗ , C
return S, A,
end procedure
Fig. 5. Build a training trajectory for network N and optimization layer
O in environment E. If constrained = false, execute raw predictions.
∆µk , yk+1 = yk + ∆yk , for k = 0, . . . , kmax . While early
stopping criteria can be employed for single QP solving,
we take kmax constant for batch solving. Noting that the
problems we solve typically converge in about 5 iterations,
we empirically set kmax = 10. The final value for x is then
e
outputted as the action a∗ closest to the initial prediction a
that satisfies all safety constraints. We depict the complete
OptLayer pipeline in Fig. 4b.
V. C ONSTRAINED R EINFORCEMENT L EARNING
We propose a reward strategy that ensures that safety
constraints are never violated throughout training and is
readily compatible with existing RL methods, such as TRPO.
A. Trajectory Sampling
Recall from Section III-B that for unconstrained RL,
we collect sequences of state-action-reward-value tuples
ei , ri , vi )i=0,...,N by iteratively executing neural network
(si , a
e in the environment E. Having established that
predictions a
executing unconstrained actions can be dangerous in the real
world, we use OptLayer to produce corrected actions a∗ and
execute those instead in E. Additionally, given constraint
matrices A, b, G, h, it is possible to quantify how much
e violate those. We define equality
unconstrained predictions a
and inequality violation costs ceq and cin , respectively:
ceq = ||Ae
a − b||,
(27)
cin = || max(Ge
a − h, 0nin )||.
(28)
From Eq. (27), ceq increases any time Ae
a 6= b, element-wise.
Similarly, from Eq (28), only elements such that Ge
a > h
contribute to cin by taking the element-wise maximum with
zero. In practice, the rows of A, b (resp. G, h) can be
scaled with an arbitrary non-zero (resp. positive) coefficient
without changing the QP solution. We thus uniformize their
individual contributions in ceq (resp. cin ) by normalization
with the coefficients making each row of A (resp. G) of
unit norm. We denote by c = ceq + cin the total constraint
violation and compute it within OptLayer, together with a∗ .
We define a procedure B UILD T RAJ in Fig. 5, that takes
as inputs an environment E, a neural network N with
OptLayer O, a flag constrained determining whether constrained or unconstrained actions are executed, and outputs
states, raw predictions, rewards, values, corrected actions and
ei , ri , vi , ai , ci )i . By setting
constraint violation costs (si , a
constrained = false, unconstrained actions are executed.
We only do so in simulation. For real-world RL, we set
constrained = true and only execute corrected actions.
B. Policy Update Strategies
With the procedure U PDATE N ETWORK described in Section III-A, we update the network N with sequences of the
form (si , ai , ri , vi )i . We assess four ways to do so.
1) Unconstrained Predictions (UP): We run B UILD T RAJ
with constrained = false and update N with
ei , ri , vi )i ), as in SecU PDATE N ETWORK (N , (si , a
tion III-B. This is our baseline for unconstrained RL.
2) Constrained - learn Predictions (CP): We run
B UILD T RAJ with constrained = true and update
ei , ri , vi )i ). Note
N with U PDATE N ETWORK (N , (si , a
that different state sequences are produced compared to
UP, as corrected (not predicted) actions are executed.
3) Constrained - learn Corrections (CC): We run
B UILD T RAJ with constrained = true and update
N with U PDATE N ETWORK (N , (si , a∗i , ri , vi )i ).
4) Constrained - learn Predictions and Corrections (CPC):
We run B UILD T RAJ with constrained = true
and compute discounted rewards rei using constraint violation costs ci : rei = ri − ci . We
first update N with raw predictions and discounted
ei , rei , vi )i ). We
rewards, U PDATE N ETWORK (N , (si , a
update N a second time, using corrected actions
and non-discounted rewards from initial sampling:
U PDATE N ETWORK (N , (si , a∗i , ri , vi )i ).
Intuitively, the CP policy update strategy amounts to considering OptLayer separately from the network and as part of
the environment. In the CC strategy, we try to learn corrected
actions directly, that can possibly greatly differ from initial
neural network predictions. With CPC, we first associate
raw predictions with discounted rewards before associating
corrected actions with better rewards.
VI. E XPERIMENTS
We now illustrate the application of OptLayer in the
context of RL to learn 3D reaching tasks using a 6-DoF
industrial manipulator (Section VI-A), both in simulation
(Section VI-B) and in the real world (Section VI-C).
A. 3D Reaching with Obstacle Avoidance
We perform our experiments on a 6-DoF industrial manipulator (Universal Robots UR5), equipped with a forcetorque sensor (Robotiq FT-150) and a gripper (Robotiq 3Finger Adaptive Robot Gripper). Although we do not use
rprox
(b) 3D-reaching reward structure.
Fig. 6. 3D reacher environment, with dT the distance between target and
gripper. In simulation, obstacle (red) and target (green) are floating spheres.
In reality, we position them using handles with visual markers.
the latter to actually grasp objects in our experiments, we
keep it for the sake of realism, making it even more crucial
to preserve safety. We control the robot by directly sending
desired trajectories through the Robot Operating System
(ROS) middleware, in joint space, without making use of
any safety mechanism available. We define a 3D reacher
environment, E3D , where the goal is to send consecutive
joint commands to reach a target point in 3D while avoiding
collisions with the environment and the robot itself. At any
step i, joint states θi , θ̇i , θ̈i and actions ai = θi+1 − θi are
now of size 6. The state vector si contains joint positions and
velocities θi , θ̇i , end effector position taken at the center of
the finger ends pee,i , and target, obstacle 3D positions relative
to pee,i . Each episode is run following the same procedure
as the 2D case, except for one adjustment, motivated by
real-robot training. At the end of each episode, we reset the
robot by replaying the trajectory in reverse. This is important
since going back from end to initial pose by joint space
interpolation can lead to collisions in between. We set the
maximum number of time steps to N = 100, with duration
∆T = 0.1 s. We implement E3D to enable parallel training on
multiple physical robots or simulation instances (Gazebo).
B. Evaluating Policy Update Strategies
To evaluate the UP, CP, CC and CPC policy update strategies, we initialize safety constraint matrices as described in
Section IV. While we could formulate them manually based
on the specifications of our robot system (UR5 robot, forcetorque sensor, gripper), it is in fact possible to partially
automate their definition. We leverage the availability of
standardized robot description formats (e.g., URDF, MJCF)
often directly released by robot manufacturers. We thus
automatically generate constraint matrices for joint limits by
simply parsing kinematic and physical parameters from such
files. For collision constraints, we manually construct convex
hulls for every robot link to enable fast distance computation [19]. We then automate the formulation of collision
constraint matrices from a set of distances to monitor, of the
form {(gripper, forearm), (elbow, obstacle), . . . }. In total, we
enforce 36 joint constraints (position, velocity, torque) and 20
collision constraints (robot, floating obstacle, environment).
In simulation, we train neural networks on E3D using UP,
CP, CC and CPC. For each policy update strategy, we report
rewards and collisions throughout 7000 episodes in Fig. 7a.
CP
CC
CPC
7.5
7.5
7.5
7.5
5.0
5.0
5.0
5.0
2.5
2.5
2.5
2.5
0.0
0.0
0.0
0.0
2.5
2.5
2.5
2.5
5.0
5.0
5.0
5.0
7.5
7.5
7.5
7.5
0
Cumulated
collisions [10]
(a) UR5 robot, obstacle and target.
R EWARD
−10dT
−100
if collision
else 0
10 if dT ≤ 5 cm
else 0
Reward [100]
UP
T YPE
rdist
rcoll
5000
20
0
0
5000
5000
0.05
0.05
0.00
0.00
0.00
0.05
0
0
0.05
5000
Episode
Episode reward
0.05
0
5000
Episode
0
5000
0
5000
Episode
0.05
0
5000
Episode
Reward moving average
Collision count
(a) Reward and collision count. Moving average over 40 consecutive episodes.
UP
CP
CC
CPC
T OTAL T IME
10 h19 min
12 h28 min
12 h22 min
12 h30 min
S T /E P
88
100
100
100
T IME /S T
60 ms
64 ms
64 ms
64 ms
E P TO R
5350
4950
N/A
3250
T IME TO R
7 h53 min
8 h49 min
N/A
5 h48 min
(b) Training time over 7000 episodes E P, steps S T, average reward R = 700.
Fig. 7. Training results for 3D reaching with a 6-DoF robot, with policy
update strategies UP (unconstrained) and CP, CC, CPC (constrained).
We observe the following. First, unconstrained training with
UC results in the robot successfully learning to reach (top
left), but not in effectively learning collision avoidance (bottom left). In contrast, using OptLayer successfully ensures
that collisions never happen for all policy update strategies.
In particular, CP realizes rewards that are comparable to UP
while never putting the robot at risk. Recall that in CP, from
the perspective of the network, OptLayer can be seen as
part of the environment, since the network is updated using
e only. In constrast, updating the network
raw predictions a
using only corrected actions a∗ within CC does not enable
learning to reach. We believe this is due to a∗ being initially
e to be representative of the stochastic policy
too far from a
we optimize within TRPO, following Eq. (4). To do so
e using
successfully, CPC first updates the network with a
discounted rewards prior to using a∗ . Subsequently, CPC
appears to reach the performance of UP and CP with far
fewer episodes. For example, a 700 average reward is reached
in 5350 episodes by UP, 4950 by CP, 3250 by CPC. Detailed
computation times (Fig. 7b) show that, though OptLayer
induces some overhead (4 ms per step), the gain in terms of
episodes makes CPC faster than UP even in terms of time.
C. Real-World Experiments
Having trained control policies in simulation, we execute
them on a real robot. In these experiments, we position
obstacle and target with respect to the robot using physical
handles marked with 2D barcodes and tracked with an RGB
camera. First, we consider a normal instance of E3D with
Fig. 8. From left to right: static obstacle, moving target; moving obstacle,
moving target; static obstacle, static target learning on a real robot.
static target and obstacle. We execute the unconstrained policy learned with UP, NUP , to reach a target behind a balloon
(Fig. 1a). While reaching was successful, the balloon was
knocked down, illustrating that collision avoidance was not
effectively learned. We do not use NUP further. In contrast,
using the constrained policy learned with CPC, NCPC , results
in successfully reaching the target while avoiding the balloon
(Fig. 1b). As a qualitative example, we consider the case
where the target is moving, carried by a toy train around
a balloon (Fig. 8, left). Although NCPC was not explicitly
trained on this task, the target was successfully followed
without hitting the balloon. We then consider the case where
the obstacle is also moving. Through the collision avoidance
constraints of Eq. (19), we verify that bringing the obstacle
close to the robot forces the latter to move away from
it. These results illustrate that combining control policies
learned on specific tasks with OptLayer can help adapt
to changing conditions, while always maintaining safety.
Finally, although position control policies can be learned in
simulation for E3D , real-world training can be necessary, e.g.,
for tasks involving force control [20] and physical interaction
with the environment [13]. We thus learn a simplified version
of E3D , in which the real robot has to reach a static target
without touching a static obstacle. For ease of visualization,
we use a cake for the latter (Fig. 8, right), never hitting it
over 6 h of training. Our approach can thus be used to learn
a robot control policy from scratch without active human
supervision. We present our complete experimental results
in the supplementary material2 .
VII. D ISCUSSION AND F UTURE W ORK
Our work establishes that stochastic control policies can be
efficiently coupled with constrained optimization to enable
RL in the real world, where safety is crucial. Our method
is simple to implement and readily compatible with existing RL techniques. We illustrated its application on neural
network policies optimized with TRPO and demonstrated its
effectiveness on a 3D reaching task with collision avoidance.
Real-world robot experiments showed in particular that our
system could accomodate reasonable changes in environment
and task conditions, even without dedicated training.
Still, our work is subject to some limitations in its current
implementation. While we showed that our approach could
speed up training by better exploiting unsafe predictions,
2 https://www.youtube.com/watch?v=7liBbk3VjWQ
learning control policies from scratch may remain too timeconsuming even for simple tasks. We would thus like
to investigate novel strategies for action space exploration
tailored to robotics. Another limitation is that formalizing
safety constraints can be difficult, e.g., under dynamic model
uncertainties. Even when constraints are well defined, their
activation may be contingent on the accuracy of other perception systems (e.g., vision). Our work would thus benefit
from further advances in learning perception and control
under uncertainties [5], [6]. Next, we intend to tackle other
challenging robot learning and control tasks, such as bipedal
locomotion and multi-agent collaboration. In the long term,
we plan to extend our approach to problems besides robotics,
such as decision-making systems for finance or healthcare.
R EFERENCES
[1] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.
Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski,
et al., “Human-level control through deep reinforcement learning,”
Nature, 2015.
[2] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van
Den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam,
M. Lanctot, et al., “Mastering the game of go with deep neural
networks and tree search,” Nature, 2016.
[3] S. Levine, C. Finn, T. Darrell, and P. Abbeel, “End-to-end training of
deep visuomotor policies,” J. Mach. Learn. Res., 2016.
[4] N. Heess, S. Sriram, J. Lemmon, J. Merel, G. Wayne, Y. Tassa, T. Erez,
Z. Wang, A. Eslami, M. Riedmiller, et al., “Emergence of locomotion
behaviours in rich environments,” arXiv, 2017.
[5] P. Christiano, Z. Shah, I. Mordatch, J. Schneider, T. Blackwell,
J. Tobin, P. Abbeel, and W. Zaremba, “Transfer from simulation to real
world through learning deep inverse dynamics model,” arXiv, 2016.
[6] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel,
“Domain randomization for transferring deep neural networks from
simulation to the real world,” in Proc. IEEE-RSJ Int. Conf. Intell.
Robot. Syst., 2017.
[7] J. Mattingley and S. Boyd, “Cvxgen: A code generator for embedded
convex optimization,” Optimization and Engineering, 2012.
[8] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel,
“Benchmarking deep reinforcement learning for continuous control,”
in Proc. Int. Conf. Mach. Learn., 2016.
[9] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, “Trust
region policy optimization,” in Proc. Int. Conf. Mach. Learn., 2015.
[10] J. Garcıa and F. Fernández, “A comprehensive survey on safe reinforcement learning,” J. Mach. Learn. Res., 2015.
[11] J. Achiam, D. Held, A. Tamar, and P. Abbeel, “Constrained policy
optimization,” in Proc. Int. Conf. Mach. Learn., 2017.
[12] Y. F. Chen, M. Liu, M. Everett, and J. P. How, “Decentralized noncommunicating multiagent collision avoidance with deep reinforcement learning,” in Proc. IEEE Int. Conf. Robot. Autom., 2017.
[13] S. Gu, E. Holly, T. Lillicrap, and S. Levine, “Deep reinforcement learning for robotic manipulation with asynchronous off-policy updates,”
in Proc. IEEE Int. Conf. Robot. Autom., 2017.
[14] B. Amos and J. Z. Kolter, “Optnet: Differentiable optimization as a
layer in neural networks,” in Proc. Int. Conf. Mach. Learn., 2017.
[15] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman,
J. Tang, and W. Zaremba, “Openai gym,” arXiv, 2016.
[16] C. Finn, T. Yu, J. Fu, P. Abbeel, and S. Levine, “Generalizing skills
with semi-supervised reinforcement learning,” in Proc. Int. Conf.
Learn. Repr., 2017.
[17] R. Smits, H. Bruyninckx, and E. Aertbeliën, “Kdl: Kinematics and
dynamics library,” http://www.orocos.org/kdl, 2011.
[18] F. Kanehiro, F. Lamiraux, O. Kanoun, E. Yoshida, and J.-P. Laumond,
“A local collision avoidance method for non-strictly convex polyhedra,” Robot. Sci. Syst., 2008.
[19] J. Pan, S. Chitta, and D. Manocha, “Fcl: A general purpose library
for collision and proximity queries,” in Proc. IEEE Int. Conf. Robot.
Autom., 2012.
[20] T. Inoue, G. De Magistris, A. Munawar, T. Yokoya, and R. Tachibana,
“Deep reinforcement learning for high precision assembly tasks,” in
Proc. IEEE-RSJ Int. Conf. Intell. Robot. Syst., 2017.
| 2cs.AI
|
1
A Scalable Algorithm for Tracking an Unknown
Number of Targets Using Multiple Sensors
arXiv:1607.07647v1 [cs.DS] 26 Jul 2016
Florian Meyer, Member, IEEE, Paolo Braca, Member, IEEE, Peter Willett, Fellow, IEEE,
and Franz Hlawatsch, Fellow, IEEE
Abstract—We propose a method for tracking an unknown
number of targets based on measurements provided by multiple
sensors. Our method achieves low computational complexity
and excellent scalability by running belief propagation on a
suitably devised factor graph. A redundant formulation of data
association uncertainty and the use of “augmented target states”
including binary target indicators make it possible to exploit
statistical independencies for a drastic reduction of complexity.
An increase in the number of targets, sensors, or measurements
leads to additional variable nodes in the factor graph but not
to higher dimensions of the messages. As a consequence, the
complexity of our method scales only quadratically in the number
of targets, linearly in the number of sensors, and linearly in the
number of measurements per sensors. The performance of the
method compares well with that of previously proposed methods,
including methods with a less favorable scaling behavior. In
particular, our method can outperform multisensor versions of
the probability hypothesis density (PHD) filter, the cardinalized
PHD filter, and the multi-Bernoulli filter.
Index Terms—Multitarget tracking, data association, belief
propagation, message passing, factor graph, sensor network.
I. I NTRODUCTION
A. Multitarget Tracking Using Multiple Sensors
Multitarget tracking is important in many applications including surveillance, autonomous driving, biomedical analytics, robotics, and oceanography [1]–[4]. Multitarget tracking
aims at estimating the states—i.e., positions and possibly further parameters—of moving objects (targets) over time, based
on measurements provided by sensing devices such as radar,
sonar, or cameras [3]. Often information from multiple sensors
is required to obtain satisfactory reliability and accuracy. The
number of targets is usually unknown [4] and there is a data
association uncertainty, i.e., an unknown association between
measurements and targets [3].
Traditional methods for multitarget tracking model the
target states as a random vector, i.e., an ordered list of
random variables, and estimate them jointly with the random
association variables. Examples are the joint probabilistic
F. Meyer and P. Braca are with the NATO STO Centre for Maritime Research and Experimentation, La Spezia 19126, Italy (e-mail:
{florian.meyer, paolo.braca}@cmre.nato.int). P. Willett is with the Department of ECE, University of Connecticut, Storrs, CT 06269, USA (e-mail:
[email protected]). F. Hlawatsch is with the Institute of Telecommunications, TU Wien, 1040 Vienna, Austria (e-mail: [email protected]).
This work was supported by the NATO Supreme Allied Command Transformation under projects SAC000601 and SAC000608, by the Naval Postgraduate School via ONR contract N00244-16-1-0017, by the Austrian Science
Fund (FWF) under project P27370-N30, and by the National Sustainability
Program of the European Commission under project LO1401. Parts of this
paper were previously presented at Fusion 2015, Washington D.C., USA, July
2015 and at Fusion 2016, Heidelberg, Germany, July 2016.
data association (JPDA) filter [3] and the multiple hypothesis
density tracker (MHT) [5] and their extensions to multiple
sensors [6]–[8]. Most of these methods assume that the number
of targets is fixed and known, which is typically not true in
practice. Because of this assumption, most traditional methods
do not solve the track management problem, i.e., they are
unable to create or cancel a track when a target appears or
disappears, respectively. Track management extensions of the
single-sensor JPDA filter and single-sensor MHT include the
joint integrated probabilistic data association (JIPDA) filter [9],
the joint integrated track splitting (JITS) filter [10], and the
search-initialize-track filter [11].
A more recent class of multitarget tracking methods is based
on finite set statistics (FISST). These methods calculate an
approximation of the posterior multiobject probability density
function (pdf), which is a joint distribution of the unordered
target states. Typically, this quantity is then used to estimate a
(possibly unordered) set of target states, which is described as
a random finite set. Notable examples include the probability
hypothesis density (PHD) filter [4], [12], [13], the cardinalized
PHD (CPHD) filter [4], [14], [15], the Bernoulli filter [16],
and the multi-Bernoulli (MB) filter [4], [17]. These FISSTbased tracking methods avoid the data association problem and
implicitly perform track management. However, most existing
methods are restricted to a single sensor.
Even more recently, FISST-based multitarget tracking methods using labeled random finite sets have been proposed. These
filters track an unknown number of targets that are identified
by an (unobserved) label, and thus are able to estimate individual target tracks. In particular, the labeled multi-Bernoulli
(LMB) filter [18] and the generalized LMB filter [19] achieve
good estimation accuracy with a computational complexity
that is similar to that of the CPHD filter. Alternatively,
the track-oriented marginal Bernoulli/Poisson (TOMB/P) filter and the measurement-oriented marginal Bernoulli/Poisson
(MOMB/P) filter proposed in [20] can estimate individual
target tracks by integrating probabilistic data association into
FISST-based sequential estimation; they are not based on
labeled random finite sets.
In the case of low-observable targets, i.e., targets leading
to measurements with a low signal-to-noise ratio, reliable
detection and tracking using a single sensor may be impossible. Theoretical results [21] suggest that the probability of
detection can be strongly improved by increasing the number
of sensors. Unfortunately, the computational complexity of
optimum multisensor-multitarget tracking scales exponentially
in the number of sensors, number of targets, and number of
measurements per sensor [22]–[25]. Computationally feasible
2
multisensor-multitarget tracking methods include the iteratorcorrector (C)PHD or briefly IC-(C)PHD filter [26], the approximate product multisensor (C)PHD filter [27], and the
partition-based multisensor (C)PHD (MS-(C)PHD) filter [23],
[24]. These methods either use approximations of unknown fidelity and thus may not be able to fully realize the performance
gains promised by multiple sensors, or they still scale poorly
in relevant system parameters. A further disadvantage of the
IC-(C)PHD filter is the strong dependence of its performance
on the order in which the sensor measurements are processed
[23], [24], [26], [27]. We note that the methods in [18]–[20]
have been formulated only for a single sensor.
B. The Proposed Method and Other Message Passing Methods
Here, we propose a multisensor method for multitarget
tracking with excellent scalability in the number of targets,
number of sensors, and number of measurements per sensor.
Our method allows for an unknown, time-varying number
of targets (up to a specified maximally possible number of
targets), i.e., it implicitly performs track management. These
advantages are obtained by performing ordered estimation
using belief propagation (BP) message passing, based on the
sum-product algorithm [28]–[31]. Contrary to most FISSTbased methods, which calculate an approximation of the
joint posterior multiobject pdf, BP provides accurate approximations of the marginal posterior pdfs for the individual
targets. These are then used to perform Bayesian detection
and estimation of the target states.
The proposed BP method is derived by formulating a
detection-estimation problem involving all the target states,
existence variables, and association variables—which are modeled via random vectors—for all times, targets, and sensors.
We use a redundant formulation of data association uncertainty
in terms of both target-oriented and measurement-oriented
association vectors [32], [33], and “augmented target states”
that include binary target existence indicators. In contrast to
FISST-based techniques, the joint augmented target state is
ordered and has a fixed number of components.
By this new formulation of the multisensor-multitarget
detection-estimation problem, the statistical structure of the
problem can be described by a factor graph, and the problem
can be solved using loopy BP message passing. The advantage
of the BP approach is that it exploits conditional statistical
independencies for a drastic reduction of complexity [28]–
[31]. We use a “detailed” factor graph in which each target
state and each association variable is modeled as an individual
node. Because this factor graph involves only low-dimensional
variables, the resulting BP algorithm does not perform highdimensional operations. As a consequence, the complexity of
our method scales only quadratically in the number of targets,
linearly in the number of sensors, and linearly in the number
of measurements per sensors (assuming a fixed number of
message passing iterations). In addition, because our method
uses particle-based calculations of all messages and beliefs, it
is suited to general nonlinear, non-Gaussian measurement and
state evolution models.
Simulation results in a challenging scenario with intersecting targets demonstrate that our method exhibits excellent
scalability and, at the same time, its performance compares
well with that of previously proposed methods. This includes
methods with a less favorable scaling behavior, namely, cubic
in the number of measurements and in the number of targets.
In particular, our method can outperform the IC-PHD and ICCPHD filters [26], the IC-MB filter [17], and the MS-PHD
and MS-CPHD filters [23], [24]. Furthermore, its performance
does not depend on an assumed order of processing the
measurements of the different sensors.
To the best of our knowledge, previously proposed BP
methods for multisensor-multitarget tracking are limited to the
method presented in [34] and our previous method in [35]. In
[34], all target states and association variables at one time step
are modeled as a joint state. This results in a tree-structured
factor graph for which BP is exact but also in an unappealing
scalability in the number of targets. In contrast, our method is
based on a detailed (but loopy) factor graph that gives rise to
low-dimensional messages and, in turn, results in the attractive
scaling properties described above. Furthermore, both methods
[34], [35] assume that the number of targets is known, whereas
our present method is suited to an unknown number of targets.
BP-based methods have also been proposed for the problems
of data association alone or data association within a multitarget tracking scheme where the tracking itself is not done by
BP. In particular, BP has been used in [32] and [33] to calculate
approximate marginal association probabilities for a single
sensor; in [36] to calculate exact marginal association probabilities for a single sensor; in [37] to calculate approximate
association probabilities for multiple sensors with overlapping
regions of interest; and in [20] to calculate approximate
association probabilities for a single sensor. In contrast to
these methods, our method uses BP for the overall multisensormultitarget tracking problem, of which data association is only
a part.
C. Paper Organization
This paper is organized as follows. The system model and
the multisensor-multitarget tracking problem are described in
Section II, and a statistical formulation of the problem is
presented in Section III. In Section IV, we briefly review
the framework of factor graphs and BP message passing.
Section V develops the proposed multisensor-multitarget tracking method. A particle-based implementation is presented in
Section VI. Section VII proposes a scheme for choosing the
birth and survival parameters. In Section VIII, relations of the
proposed method to existing methods are discussed. Finally,
simulation results in a scenario with intersecting targets are reported in Section IX. We note that this paper advances over the
preliminary account of our method provided in our conference
publication [38] by adding a particle-based implementation, a
scheme for choosing birth and survival parameters, a detailed
discussion of relations to existing methods, additional performance results, and an experimental verification of scaling
properties.
II. S YSTEM M ODEL
AND
P ROBLEM S TATEMENT
In this section, we describe our system model and formulate
the multitarget detection-estimation problem to be solved.
3
A. Potential Targets and Sensor Measurements
We consider at most K targets with time-varying states. We
describe this situation by introducing potential targets (PTs)
k ∈ K , {1, . . . , K}. The existence of the PTs is modeled
by binary variables rn,k ∈ {0, 1}, i.e., PT k exists at time
n if and only if rn,k = 1. We also define the vector rn ,
T
rn,1 · · · rn,K . The state xn,k of PT k at time n consists
of the PT’s position and possibly further parameters. It will
be convenient to formally consider a PT state xn,k also if
rn,k = 0. We define the augmented state yn,k , [xTn,k rn,k ]T
T
T
T
and the joint augmented state yn , yn,1
· · · yn,K
.
There are S sensors s ∈ S , {1, . . . , S} that produce
“thresholded” measurements resulting from a detection process (as performed, e.g., by a radar or sonar device). Let
(s)
(s)
(s)
zn,m , m ∈ Mn , 1, . . . , Mn
denote the measurements
produced by sensor s at time n. We also define
T the stacked
(s)T
(s)T
(s)
and zn ,
measurement vectors zn , zn,1 · · · z
(s)
n
(1)T
n,M
(S)T T
(1)
(S) T
zn · · · zn
, and the vector mn , Mn · · · Mn
.
Because the measurements are thresholded, the multitarget
tracking problem is complicated by a data association un(s)
certainty: it is unknown which measurement zn,m originated
from which PT k, and it is also possible that a measurement
(s)
zn,m did not originate from any PT (false alarm, clutter)
or that a PT did not generate any measurement of sensor s
(missed detection) [3], [4]. We make the usual assumption
that at any time n, an existing target can generate at most one
measurement at sensor s, and a measurement at sensor s can
be generated by at most one existing target [3], [4]. The PTmeasurement associations at sensor s and time n can then be
described by the target-oriented association variables
(s)
m ∈ Mn , at time n, PT k generates measure
(s)
ment m at sensor s
an,k ,
at time n, PT k is not detected by
0,
sensor s.
(s)
(1)T
(s)
(s) T
We also define an , an,1 · · · an,K and an , an · · ·
(S)T T
an
.
Following [32] and [33], we also use an alternative description of the PT-measurement associations in terms of the
measurement-oriented association variables
k ∈ K, at time n, measurement m at
sensor s is generated by PT k
(s)
bn,m ,
(1)
0,
at
time n, measurement m at
sensor s is not generated by a PT.
(s)
T
(1)T
(s)
(s)
We also define bn , bn,1 · · · b
and bn , bn · · ·
(s)
n,Mn
(S)T T
(s)
bn
. The description in terms of bn is redundant in that
(s)
(s)
bn can be derived from an and vice versa.
B. Target Detection and State Estimation
The problem considered is detection of the PTs k ∈ K
(i.e., of the binary target existence variables rn,k ) and estimation of the target states xn,k from the past and present
measurements of all the sensors s ∈ S, i.e., from the total
T
measurement vector z , zT1 · · · zTn . In a Bayesian setting,
this essentially amounts to calculating the marginal posterior
existence probabilities p(rn,k = 1|z) and the marginal posterior
pdfs f (xn,k |rn,k = 1, z). Target detection is performed by
comparing p(rn,k = 1|z) to a threshold Pth , i.e., PT k is
considered to exist if p(rn,k = 1|z) > Pth [39, Ch. 2]. For
the detected targets k, an estimate of xn,k is then produced
by the minimum mean-square error (MMSE) estimator [39,
Ch. 4]
Z
x̂MMSE
,
xn,k f (xn,k |rn,k = 1, z)dxn,k .
(2)
n,k
This Bayesian two-stage detection-estimation procedure is
also employed by the JITS method [10], the JIPDA filter
[9], and certain FISST-based algorithms, e.g., [4], [20]. The
main problem to be solved now is to find a computationally
feasible recursive (sequential) calculation of p(rn,k = 1|z) and
f (xn,k |rn,k = 1, z).
III. S TATISTICAL F ORMULATION
Next, we present a statistical formulation of the system
model and the multitarget detection-estimation problem.
A. Target States
While in our model a PT state xn,k is formally defined
also if rn,k = 0, the states of nonexisting PTs are obviously
irrelevant. Accordingly, all pdfs and BP messages defined
for an augmented state, φ(yn,k ) = φ(xn,k , rn,k ), have the
property that for rn,k = 0,
φ(xn,k , 0) = φn,k fD (xn,k ),
(3)
where fD (xn,k ) is a “dummy pdf.” The form (3) must be
consistent with a message multiplication operation (such as
Equation (24) in Section IV), in the sense that the resulting
message product can still be expressed as in (3). This implies
that the dummy pdf fD (xn,k ) satisfies fD2 (xn,k ) = fD (xn,k )
for all values of xn,k . Because fD (xn,k ) must also integrate
to one, it follows that fD (xn,k ) is 1 on an arbitrary support of
area/volume 1 and 0 outside that support. Let
Z
φ(rn,k ) , φ(xn,k , rn,k )dxn,k .
(4)
We then have for rn,k = 0
Z
Z
φ(0) = φ(xn,k , 0)dxn,k = φn,k fD (xn,k )dxn,k = φn,k .
(5)
Furthermore, using (4) and (5) yields
X Z
φ(xn,k , rn,k )dxn,k
rn,k ∈{0,1}
=
Z
φ(xn,k , 0)dxn,k +
Z
φ(xn,k , 1)dxn,k
= φn,k + φ(1).
R
P
Hence, if
rn,k ∈{0,1} φ(xn,k , rn,k )dxn,k = 1, i.e., if
φ(xn,k , rn,k ) is a true pdf in the sense of being normalized,
then φn,k + φ(1) = 1. In that case, φn,k = φ(0) can be
4
interpreted as a probability of nonexistence of PT k, i.e., of
the event rn,k = 0, and φ(1) can be interpreted as a probability
of existence of PT k, i.e., of the event rn,k = 1.
The augmented target states yn,k = [xTn,k rn,k ]T are
assumed to evolve independently according to Markovian
dynamic models [3], [4], and at time n = 0, they are assumed
statistically independent across k with prior pdfs f (y0,k ) =
T
f (x0,k , r0,k ). Thus, the pdf of y , y0T · · · ynT factorizes as
f (y) =
K
Y
f (y0,k )
k=1
n
Y
f (yn′,k |yn′−1,k ).
(6)
assumption [3], [4] that given y, a, and m, the measurements
(s)
zn are conditionally independent across time n and sensor
index s, the global likelihood function factorizes as
n Y
S
Y
f (z|y, a, m) =
n′ =1 s=1
If PT k existed at time n − 1, i.e., rn−1,k = 1, then the
probability that it still exists at time n, i.e., rn,k = 1, is given
s
, and if it still exists at time
by the survival probability pn,k
n, its state xn,k is distributed according to the state transition
pdf f (xn,k |xn−1,k ). Thus, for rn−1,k = 1, we have
(
s
)fD (xn,k ), rn,k = 0
(1− pn,k
f (xn,k , rn,k |xn−1,k , 1) =
s
pn,k f (xn,k |xn−1,k ), rn,k = 1.
(8)
b
s
A possible strategy for choosing pn,k
, pn,k
, and f b (xn,k ) is
presented in Section VII. We note that our previous work in
[35], which assumed that the number of targets is known, is a
s
special case of this setup that uses survival probabilities pn,k
=
b
=0
1 (existing targets always survive), birth probabilities pn,k
R(no targets are born), and initial prior pdfs f (x0,k , r0,k ) with
f (x0,k , 1)dx0,k = 1 (at time n = 0, all targets exist with
probability 1).
B. Sensor Measurements
An existing target k is detected by sensor s (in the sense
(s)
that the target generates a measurement zn,m at sensor s) with
(s)
probability Pd (xn,k ), which may depend on the target state
xn,k . The number of false alarms at sensor s is modeled by a
Poisson probability mass function (pmf) with mean µ(s), and
the distribution of each false alarm
measurement at sensor s
(s)
is described by the pdf fFA zn,m [3], [4].
The dependence of the measurement vector z = zT1 · · ·
T
zTn on the numbers-of-measurements vector m , mT1 · · ·
T
T
mTn , the augmented state vector y = y0T · · · ynT , and the
T
T
association vector a , a1 · · · aTn is described by the global
likelihood function f (z|y, a, m). With the commonly used
(9)
(s)
Assuming in addition that the different measurements zn,m
(s)
at sensor s are conditionally independent given yn , an , and
(s)
Mn , we have the further factorization [3], [4]
(s)
Mn
n′ =1
Here, the single-target augmented state transition pdf
f (yn,k |yn−1,k ) = f (xn,k , rn,k |xn−1,k , rn−1,k ) is given as
follows. If PT k did not exist at time n − 1, i.e., rn−1,k = 0,
then the probability that it exists at time n, i.e., rn,k = 1,
b
is given by the birth probability pn,k
, and if it does exist at
time n, its state xn,k is distributed according to the birth pdf
f b (xn,k ). Thus, for rn−1,k = 0, we have
(
b
(1− pn,k
)fD (xn,k ), rn,k = 0
f (xn,k , rn,k |xn−1,k , 0) =
b
f b (xn,k ),
rn,k = 1.
pn,k
(7)
(s)
(s)
(s)
f zn′ yn′ , an′ , Mn′ .
f
z(s)
n
(s)
yn , a(s)
n , Mn
=
Y
fFA z(s)
n,m
m=1
!
(s)
f
z
x
n,k
(s)
Y
n,an,k
. (10)
×
(s)
z
f
(s)
FA
(s)
k∈D
an ,rn
(s)
n,an,k
targets detected at
Here, Dan ,rn denotes the set of existing
(s)
(s)
sensor s and time n, i.e., Dan ,rn , k ∈ K : rn,k = 1, an,k 6=
(s)
(s)
0 . If zn is observed and thus fixed, Mn is fixed as well
and (10) can be written as
f
z(s)
n
(s)
yn , a(s)
n , Mn
=C
z(s)
n
K
Y
k=1
(s)
,
g xn,k , rn,k , an,k ; z(s)
n
(11)
(s)
zn
where C
is a normalization factor that depends only on
(s)
(s)
(s)
zn and g xn,k , rn,k , an,k ; zn is defined as
(s)
f zn,m xn,k
(s)
(s)
, an,k = m ∈ Mn
(s)
(s)
=
g xn,k , 1, an,k ; z(s)
z
f
FA n,m
n
(s)
1,
a
=0
n,k
(s)
g xn,k , 0, an,k ; z(s)
= 1.
n
(12)
Inserting (11) into (9) yields
f (z|y, a, m) = C(z)
n Y
S Y
K
Y
n′ =1 s=1 k=1
(s)
(s)
g xn′,k , rn′,k , an′,k ; zn′ ,
(13)
where C(z) is a normalization factor that depends only on z.
C. Joint Prior Distribution of Association Variables and Numbers of Measurements
(s)
(s)
Under the assumption that given y, the an and the Mn
are conditionally independent across n and s [3], [4], the joint
prior pmf of the association vector a and the numbers-ofmeasurements vector m given y factorizes as
p(a, m|y) =
n Y
S
Y
n′ =1 s=1
(s)
(s)
p an′ , Mn′ yn′ .
(14)
(s)
Assuming a random permutation of the measurements zn,m ,
(s)
m ∈ Mn at sensor s, with each permutation equally likely,
it is shown in [3], [11] that
5
(s)
p a(s)
n , Mn yn
= ψ a(s)
n
×
e
(µ
(s)
(s)
(s) Mn −|Dan ,rn |
)
(s)
Mn !
Y
(s)
k′∈D
/ an ,rn
Y
(s)
Pd (xn,k )
(s)
k∈Dan ,rn
(s)
(s)
1 an,k′ − rn,k′ Pd (xn,k′ ) ,
(15)
ψ
,
(
(s)
1, otherwise,
(16)
=C
Mn(s)
K
Y
k=1
and 1(a) denotes the indicator function of the event a = 0
(i.e., 1(a)
= 1 if a = 0 and 0 otherwise). Note that the factor
(s)
ψ an enforces the exclusion assumptions stated in Section
II-A, i.e., that each existing target can generate at most one
measurement at sensor s and each measurement at sensor s
can be generated by at most one target. We can express (15)
as
(s)
(s)
ψ a(s)
p a(s)
n
n , Mn yn = C Mn
×
K
Y
(s)
(s)
h xn,k , rn,k , an,k ; Mn
k=1
, (17)
(s)
where C Mn is a normalization factor depending only on
(s)
(s)
(s)
Mn and h xn,k , rn,k , an,k ; Mn is defined as
(s)
(s)
h xn,k , 1, an,k ; Mn
(s)
(s)
(s)
Pd (xn,k ) ,
an,k ∈ Mn
µ(s)
=
(s)
(s)
1− Pd (xn,k ), an,k = 0
(s)
(s)
h xn,k , 0, an,k ; Mn(s) = 1 an,k .
(18)
(s)
Using the measurement-oriented association vectors bn
defined in (1) alongside with the target-oriented association
(s)
(s)
vectors an , the exclusion-enforcing function ψ an in (16)
can be formally replaced by the function [32], [33]
(s)
ψ
with
Fig. 1.
Factor graph representing
ψ1 (x1 , x2 )ψ2 (x2 )ψ3 (x2 , x3 ).
(s)
0, ∃k, k ′ ∈ K such that an,k = an,k′ 6= 0
(s)
a(s)
n , bn
=
K M
n
Y
Y
k=1 m=1
Ψ
(s)
an,k , b(s)
n,m ,
(s)
(s)
0, an,k = m, bn,m 6= k
(s)
(s)
(s)
or bn,m = k, an,k 6= m
Ψ an,k , b(s)
n,m ,
1, otherwise.
(19)
Using this redundant reformulation, and defining b , bT1 · · ·
T
bTn , the prior pmf p(a, m|y) in (14) can be formally rewritten
as
n Y
S
Y
(s)
(s)
(s)
(20)
p(a, b, m|y) =
p an′ , bn′ , Mn′ yn′ ,
n′ =1 s=1
with the single-sensor prior pmfs (cf. (17) and (19))
x3
ψ2
(s)
(s)
p a(s)
yn
n , bn , M n
ψ3
x2
!
where
a(s)
n
ψ1
x1
−µ(s)
h
f (x|z)
the factorization
(s)
Mn
Y
(s)
(s)
Ψ
xn,k , rn,k , an,k ; Mn
m=1
∝
(s)
an,k , b(s)
n,m .
Thus, Equation (20) can be expressed as
p(a, b, m|y) = C(m)
n Y
K
S Y
Y
(s)
n′ =1 s=1 k=1
(s)
Mn′
×
(s)
h xn′,k , rn′,k , an′,k ; Mn′
Y
(s)
(s)
Ψ an′,k , bn′,m ,
(21)
m=1
where C(m) is a normalization factor depending only on m.
The factorization in (21) constitutes an important basis for our
development of the proposed BP method in Section V.
IV. R EVIEW OF BP M ESSAGE PASSING
We briefly review factor graphs and the generic BP message
passing scheme, which constitute the main methodological
basis of the proposed multisensor-multitarget tracking method.
Consider the problem of estimating parameter vectors xk ,
k ∈ {1, . . . , K} from a measurement vector z. Bayesian
estimation of xk relies on the posterior pdf f (xk |z) [40].
This pdf is a marginal pdf of the joint posterior pdf f (x|z),
K
where x = [xk ]k=1 ; however, direct marginalization of f (x|z)
is usually infeasible. An efficient marginalization can be
achieved if the posterior pdf f (x|z) factorizes, i.e.,
f (x|z) ∝
Q
Y
q=1
ψq x(q) .
(22)
Here, each factor argument x(q) comprises certain parameter
vectors xk (each xk can appear in several x(q) ) and ∝
indicates equality up to a normalization factor. Note that for
compactness, our notation
does not indicate the dependence
of the factors ψq x(q) on z.
The factorization structure (22) can be represented
by a
T
factor graph [31]. As an example, for x = xT1 xT2 xT2 ,
the factor graph representing the factorization f (x|z) ∝
ψ1 (x1 , x2 )ψ2 (x2 )ψ3 (x2 , x3 ) is shown in Fig. 1. In a factor
graph, each parameter variable xk is represented by a variable
node and each factor ψq (·) by a factor node (depicted in Fig.
1 by a circle and a square, respectively). Variable node “xk ”
and factor node “ψq ” are adjacent, i.e., connected by an edge,
if the variable xk is an argument of the factor ψq (·).
Belief propagation (BP), also known as the sum-product
algorithm [28], is based on a factor graph and aims at
computing the marginal posterior pdfs f (xk |z) in an efficient
6
way. For each node, certain messages are calculated, each of
which is passed to one of the adjacent nodes. For each variable
node, the incoming and outgoing messages are functions of the
corresponding variable. More specifically, consider a variable
node “xk ” and an adjacent factor node “ψq ”, i.e., the variable
xk is part of the argument x(q) of ψq x(q) . Then, the message
passed from factor node “ψq ” to variable node “xk ” is given
by
Z
Y
ζψq →xk (xk ) = ψq x(q)
η xk′ →ψq (xk′ ) dxk̄ , (23)
k′6=k
Q
where k′6=k η xk′ →ψq (xk′ ) denotes the product of the messages passed to factor node “ψRq ” from all adjacent variable nodes except “xk ”, and . . . dxk̄ denotes integration with respect to all constituent vectors of x(q) except
xk . For example, the message passed from factor node
“ψ ” to variable node “x2 ” in Fig. 1 is ζψ1 →x2 (x2 ) =
T
R 1
ψ1 (x1 , x2 )η x1 →ψ1 (x1 )dx1 ; note that x(1) = xT1 xT2 .
The message η xk →ψq (xk ) passed from variable node “xk ”
to factor node “ψq ” is given by the product of the messages passed to variable node “xk ” from all adjacent factor
nodes except “ψq ”. For example, in Fig. 1, the message
passed from variable node “x2 ” to factor node “ψ1 ” is
η x2 →ψ1 (x2 ) = ζψ2 →x2 (x2 )ζψ3 →x2 (x2 ). Message passing is
started at variable nodes with only one edge (which pass a
constant message) and/or factor nodes with only one edge
(which pass the corresponding factor). Note that BP can also
be applied to factorizations involving discrete variables by
replacing integration with summation in (23).
Finally, for each variable node “xk ”, a belief f˜(xk ) is
calculated as the product of all incoming messages (passed
from all adjacent
factor nodes) followed by a normalization
R
such that f˜(xk )dxk = 1. For example, in Fig. 1,
f˜(x2 ) ∝ ζψ1 →x2 (x2 )ζψ2 →x2 (x2 )ζψ3 →x2 (x2 ).
and the marginal posterior pdf f (xn,k |rn,k = 1, z) underlying MMSE state estimation (see (2)) can be obtained from
f (xn,k , rn,k |z) according to
A. Joint Posterior pdf and Factor Graph
The marginal posterior pdf f (yn,k |z) = f (xn,k , rn,k |z) is a
marginal density of the joint posterior pdf f (y, a, b|z), which
involves all the augmented states, all the target-oriented and
measurement-oriented association variables, and all the measurements of all sensors, at all times up to the current time n.
In the following derivation of f (y, a, b|z), the measurements
(s)
z are observed and thus fixed, and consequently Mn and m
are fixed as well. Then, using Bayes’ rule and the fact that a
implies b, we obtain
f (y, a, b|z) = f (y, a, b, m|z)
∝ f (z|y, a, b, m)f (y, a, b, m)
= f (z|y, a, m) p(a, b, m|y)f (y).
Inserting (6) for f (y), (13) for f (z|y, a, m), and (21) for
p(a, b, m|y) then yields the final factorization
f (y, a, b|z) ∝
K
Y
f (y0,k )
k=1
×
S
Y
s=1
V. T HE P ROPOSED BP- BASED M ULTISENSOR M ULTITARGET T RACKING M ETHOD
The marginal posterior existence probability p(rn,k = 1|z)
underlying target detection as discussed in Section II-B can
be obtained from the marginal posterior pdf of the augmented
target state, f (yn,k |z) = f (xn,k , rn,k |z), according to
Z
p(rn,k = 1|z) = f (xn,k , rn,k = 1|z)dxn,k ,
(25)
(26)
An efficient approximate calculation of f (xn,k , rn,k |z) can
be obtained by performing BP message passing on a factor
graph that expresses the factorization of the joint posterior
pdf involving all relevant parameters. This factor graph will
be derived next.
(24)
If the factor graph is a tree, i.e., without loops, then the belief
f˜(xk ) is exactly equal to the marginal posterior pdf f (xk |z).
For factor graphs with loops, BP is applied in an iterative
manner, and the beliefs f˜(xk ) are only approximations of
the respective marginal posterior pdfs f (xk |z); these approximations have been observed to be very accurate in many
applications [28]–[30]. In this iterative “loopy BP” scheme,
there is no canonical order in which the messages should be
calculated, and different orders may lead to different beliefs.
The choice of an appropriate order of message calculation will
be an important aspect in our development of the proposed
method.
f (xn,k , rn,k = 1|z)
.
p(rn,k = 1|z)
f (xn,k |rn,k = 1, z) =
with
n
Y
f (yn′,k |yn′−1,k )
n′ =1
υ
(s)
(s)
yn′,k , an′,k ; zn′
(s)
Mn′
Y
(s)
(s)
Ψ an′,k , bn′,m ,
m=1
(27)
(s)
(s)
, g xn,k , rn,k , an,k ; z(s)
υ yn,k , an,k ; z(s)
n
n
(s)
× h xn,k , rn,k , an,k ; Mn(s) . (28)
This factorization can be represented graphically by a factor
graph as explained in Section IV. This factor graph is depicted
for one time step in Fig. 2; it provides the starting-point for
our development of the proposed BP method.
B. BP Method
As discussed in Section IV, approximations f˜(yn,k ) of
the marginal posterior pdfs f (yn,k |z) can be obtained in an
efficient way by running iterative BP message passing [28],
[30], [31] on the factor graph in Fig. 2. Since this factor graph
is loopy, we have to decide on a specific order of message
computation. We choose this order according to the following
rules: (i) Messages are not sent backward in time1 [41]. (ii)
1 This is equivalent to the approximative assumption that the target states
are conditionally independent given the past measurements, as is done in the
derivation of the JPDA filter [3].
7
=
s=S
f˜1
f1
γ1
α1
y1
α1
υ1
s=1
α1
γ1
υ1
Ψ1,1
β1
η1
ν1,1
ζ1,1
b1
a1
ζK,1
νM,1
Ψ1,M
f˜K
γK
fK
αK
yK
αK
γK
ΨK,1
υK
αK
υK
βK
ηK
ν1,K
ζ1,M
aK
νM,K
bM
ζK,M
ΨK,M
Z
(s)
α(xn,k , 1) dxn,k
υ xn,k , 1, an,k ; z(s)
n
(s)
+ 1 an,k αn,k . (31)
(s)
(s)
In the last expression, we used υ xn,k , 0, an,k ; zn =
(s)
1 an,k , which follows from (28) with (12) and (18).
2) Iterative data association (this part of the BP method
closely follows [32], [33], [42]). In iteration p ∈ {1, . . . ,
P }, the following calculations are performed for all
(s)
measurements m ∈ Mn :
X
Y (p−1) (s)
(p)
(s)
(s)
νm→k an,k =
ζk′ →m bn,m
Ψ an,k , b(s)
n,m
(s)
Fig. 2. Factor graph representing the factorization of the joint posterior pdf
f (y, a, b|z) in (27), depicted for one time step. For simplicity, the time
index n and sensor index s are omitted, and the following
short notations
(s)
(s)
are used: fk , f (yn,k |yn−1,k ), υk , υ yn,k , an,k ; zn , f˜k , f˜(yn,k ),
(s)
(s)
(s)
(s)
Ψk,m , Ψ an,k , bn,m , αk , α(yn,k ), βk , β an,k , ηk , η an,k ,
(p)
(s)
(p)
(s)
γk , γ (s) (yn,k ), νm,k , νm→k an,k , and ζk,m , ζk→m bn,m .
k′ ∈K\{k}
bn,m
and
(32)
X
(p)
(s)
(s)
ζk→m b(s)
β an,k Ψ an,k , b(s)
n,m =
n,m
(s)
an,k
×
Y
(s)
(p)
(s)
νm′ →k an,k . (33)
m′ ∈Mn \{m}
Iterative message passing is only performed at each time step
and at each sensor separately—i.e., in particular, for the loops
connecting different sensors we only perform a single message
passing iteration—and only for data association. With these
rules, the generic BP rules for calculating messages and beliefs
as summarized in Section IV yield the following BP message
passing operations at time n.
First, a prediction step is performed for all PTs k ∈ K, i.e.,
Z
X
α(xn,k , rn,k ) =
f (xn,k , rn,k |xn−1,k , rn−1,k )
rn−1,k ∈{0,1}
× f˜(xn−1,k , rn−1,k ) dxn−1,k .
Here, f˜(xn−1,k , rn−1,k ) was calculated at the previous time
n−1. Inserting (7) and (8) for f (xn,k , rn,k |xn−1,k , rn−1,k ), we
obtain the following expressions of α(xn,k , rn,k ): for rn,k = 1,
b
s
α(xn,k , 1) = pn,k
f b (xn,k ) f˜n−1,k + pn,k
Z
× f (xn,k |xn−1,k )f˜(xn−1,k , 1) dxn−1,k , (29)
P
P
P
Here,
is short for
and
(s)
(s)
(s)
bn,m
b
∈{0,...,K}
an,k
n,m
P
is short for a(s) ∈{0,...,M (s) } . The operations (32) and
n
n,k
(33) constitute an iteration loop, which is initialized (for
p = 0) by
X
(s)
(s)
(0)
(34)
β an,k Ψ an,k , b(s)
ζk→m b(s)
n,m .
n,m =
(s)
an,k
An efficient implementation of (32) and (33) is described
in [42] and [33]. After
the last iteration p = P , the
(P )
(s)
(s)
messages νm→k an,k , m ∈ Mn are multiplied, i.e.,
(s)
Mn
η
(s)
an,k
=
Y
m=1
(s)
(P )
νm→k an,k .
(35)
3) Measurement update:
X
(s)
(s)
η an,k
γ (s) (xn,k , 1) =
υ xn,k , 1, an,k ; z(s)
n
(s)
an,k
(s)
(s)
γn,k = η an,k = 0 .
(36)
Finally, beliefs f˜(yn,k ) = f˜(xn,k , rn,k ) approximating
the marginal posterior pdfs f (yn,k |z) = f (xn,k , rn,k |z) are
s
b
f˜(xn−1,k , 1) dxn−1,k . obtained as the products
f˜n−1,k + 1− pn,k
αn,k = 1− pn,k
(30)
S
Y
1
R
˜
f (xn,k , 1) =
γ (s) (xn,k , 1)
(37)
α(xn,k , 1)
We note Rthat f˜n−1,k =
f˜(xn−1,k , 0)dxn−1,k and
Cn,k
s=1
αn,k =
α(xn,k , 0)dxn,k (cf. (5)). Furthermore, since
S
˜(xn,k , rn,k ) is normalized, so is α(xn,k , rn,k ), i.e.,
Y
fP
(s)
R
˜n,k = 1 αn,k
f
γn,k ,
(38)
α(x
,
r
)dx
=
1.
Thus,
we
have
α
=
n,k
n,k
n,k
n,k
C
rn,k
∈{0,1}
n,k
R
s=1
1 − α(xn,k , 1)dxn,k .
After the prediction step, the following steps are performed with the normalization constant
Z
for all PTs k ∈ K and all sensors s ∈ S in parallel:
S
S
Y
Y
(s)
(s)
C
=
α(x
,
1)
γ
(x
,
1)dx
+
α
γn,k .
n,k
n,k
n,k
n,k
n,k
1) Measurement evaluation:
s=1
s=1
X Z
(39)
(s)
(s)
(s)
β an,k =
υ xn,k , rn,k , an,k ; zn
2
Note that fD (xn,k ) = fD (xn,k ) (cf. Section III-A) was used to
rn,k ∈{0,1}
× α(xn,k , rn,k ) dxn,k
obtain (38). Because the belief f˜(xn,k , 1) in (37) approximates
and for rn,k = 0,
Z
8
(j)
(j)
x̄n−1,k , w̄n−1,k
Prediction step maps
(j)
(j)
x̄n−1,k , w̄n−1,k
J
∼ f˜(yn−1,k ),
j=1
J
∼ f˜(yn−1,k )
j=1
k ∈ K (from the previous time step n−1)
to
(j)
α(j) J +I
∼
j=1
xn,k , wn,k
α(yn,k ), k ∈ K (see Section VI-A)
For all sensors s:
(s)
Particle-based measurement evaluation provides β̃ an,k , k ∈ K (see (42))
(s)
(s)
• Iterative data association calculates η̃ a
n,k from β̃ an,k , k ∈ K (see (32)–(35))
•
(j)
(j) J +I
Importance sampling calculates xn,k , wn,k j=1 ∼ f˜(yn,k ), k ∈ K from
(j)
α(j) J +I
(s)
xn,k , wn,k
∼ α(yn,k ) and η̃ an,k , s ∈ S (see Section VI-C)
j=1
The quantities pen,k and x̂n,k are determined
(j)
(j) J +I
from xn,k , wn,k j=1 ∼ f˜(yn,k ), k ∈ K
(see (45) and (46), respectively)
(j)
(j) J +I
Resampling maps xn,k , wn,k j=1 ∼ f˜(yn,k ) to
(j)
(j) J
x̄n,k , w̄n,k j=1 ∼ f˜(yn,k ), k ∈ K (see Section VI-D)
pen,k , x̂n,k , k ∈ K
(j)
(j) J
x̄n,k , w̄n,k j=1 ∼ f˜(yn,k ), k ∈ K
(for processing at the next time step n + 1)
Fig. 3. Flowchart of a particle-based implementation of the proposed BP method. The symbol ∼ expresses the fact that a set of particles and weights
represents a certain distribution.
f (xn,k , rn,k = 1|z), it can be substituted for f (xn,k , rn,k =
1|z) in (25) and (26), thus providing the basis for Bayesian
target detection and state estimation as discussed in Section
II-B. A particle-based implementation of the above BP method
that avoids the explicit evaluation of integrals and message
products will be presented in Section VI.
The “data association” iteration loop (32)–(35) involves
solely messages related to discrete random variables. Being
based on loopy BP, it does not perform an exact marginalization [28], [30], [31]. However, its high accuracy has been
demonstrated numerically [32], [33] (see also Section IX-B),
and its convergence has been proven [33], [42].
C. Scalability
The main advantage of the BP message passing method
described in Section V-B is its scalability. Assuming a fixed
number P of message passing iterations, the computational
complexity of calculating the (approximate) marginal posterior
pdfs of all the target states is only linear in the number of
sensors S (see (37), (38)). Moreover, the complexity of the
operations (31)–(36) performed for a given sensor s ∈ S
(s)
(s)
scales as O KMn , where Mn increases linearly with the
number of PTs K and with the number of false alarms. Thus,
the overall complexity of our algorithm scales linearly in the
number of sensors and quadratically in the number of PTs.
Note that in practical implementations, measurement gating
[3] can be used to further improve scalability.
Such favorable scaling is a consequence of the “detailed”
factorization (27). This factorization, in turn, is due to the
redundant formulation of the joint state-association estimation
task in terms of both target-related and measurement-related
association variables as described in Sections II-A and III-C.
Using this factorization, an increase in the number of PTs,
the number of sensors, or the number of measurements leads
to additional variable nodes in the factor graph (see Fig. 2)
but not to higher dimensions of the messages passed between
the nodes. The scalability of the proposed algorithm will be
further analyzed in Section IX-C.
VI. PARTICLE -BASED I MPLEMENTATION
For general nonlinear and non-Gaussian measurement and
state evolution models, the integrals in (2), (29), (30), and (31)
as well as the message products in (37) and (38) typically
cannot be evaluated in closed form and are computationally
infeasible. Therefore, we next present an approximate particlebased implementation of these operations. In this implementation, each belief f˜(xn,k , rn,k ) is represented by a set of
(j)
(j) J
particles and corresponding weights x̄n,k , w̄n,k j=1 . More
(j)
(j) J
specifically, f˜(xn,k , 1) is represented by x̄n,k , w̄n,k j=1 ,
and f˜(xn,k , 0) is given implicitly by the normalization
propR
erty of f˜(xn,k , rn,k ), i.e., f˜(xn,k , 0) = 1 − f˜(xn,k , 1)dxn,k .
Contrary to conventional particle filtering [43], [44], the parti(j)
cle weights w̄n,k , j ∈ {1, . . . , J} do not sum to one; instead,
e
pn,k
,
J
X
j=1
(j)
w̄n,k
≈
Z
f˜(xn,k , 1)dxn,k .
(40)
R
Note that since f˜(xn,k , 1)dxn,k approximates the posterior
probability of target existence p(rn,k = 1|z), it follows that the
e
sum of weights pn,k
is approximately equal to p(rn,k = 1|z).
The particle operations discussed in the remainder of this
section are performed for all PTs k ∈ K in parallel. The
resulting particle-based implementation of the overall BP
method is summarized in Fig. 3.
9
A. Prediction
n ≥ 1 and for each PT k∈ K, J particles and weights
For
J
(j)
(j)
e
x̄n−1,k , w̄n−1,k = pn−1,k
/J j=1 representing the previous belief f˜(xn−1,k , rn−1,k ) were calculated at the previous
time n − 1 as described further below. Weighted particles
J+I
(j)
α(j)
xn,k , wn,k j=1 representing the message α(xn,k , 1) in
(29) are now obtained as follows.2 First, for each parti(j)
(j)
cle x̄n−1,k , j ∈ {1, . . . , J}, one particle xn,k is drawn
(j)
from f xn,k x̄n−1,k . Next, I additional “birth particles”
(j) J+I
xn,k j=J+1 are drawn from the birth pdf f b (xn,k ). Finally,
α(j)
weights wn,k , j ∈ {1, . . . , J + I} are obtained as
α(j)
wn,k
=
(
(j)
s
pn,k
w̄n−1,k ,
j ∈ {1, . . . , J}
b
pn,k
(1 −
j ∈ {J +1, . . . , J +I}.
e
pn−1,k
)/I ,
(41)
importance sampling [43], [44]. To that end, we first rewrite
the belief f˜(xn,k , rn,k ) in (37), (38) by inserting (36), i.e.,
f˜(xn,k , 1) ∝ α(xn,k , 1)
S X
Y
s=1 a(s)
n,k
f˜n,k ∝ αn,k
S
Y
(s)
s=1
A(j)
α(j)
wn,k = wn,k
S X
Y
(j)
B. Measurement Evaluation
(s)
For each sensor s ∈ S, an approximation β̃ an,k of the
(s)
message β an,k in (31) can be calculated from the weighted
(j)
α(j) J+I
particles xn,k , wn,k j=1 representing α(xn,k , rn,k ) as
J+I
X
α(j)
(j)
(s)
(s)
wn,k
υ xn,k , 1, an,k ; z(s)
β̃ an,k =
n
j=1
+1
(s)
an,k
1−
J+I
X
j=1
!
α(j)
wn,k
(s)
(j)
(s)
η̃ an,k ,
υ xn,k , 1, an,k ; z(s)
n
j ∈ {1, . . . , J + I}.
Note that this expression is based on importance sampling
with proposal density α(xn,k , 1) (represented by the weighted
(j)
α(j) J+I
particles xn,k , wn,k j=1 ). Similarly, we calculate a single
nonnormalized weight corresponding to (44) as
! S
J+I
X α(j) Y
(s)
B
wn,k
η̃ an,k = 0 ,
wn,k = 1 −
s=1
j=1
in which 1 −
PJ+I
α(j)
wn,k approximates αn,k .
(j)
(j) J+I
Next, weighted particles
xn,k , wn,k j=1 representing
the belief f˜(xn,k , rn,k ) are obtained by using the particles
(j) J+I
xn,k j=1 representing α(xn,k , rn,k ) and calculating the corresponding weights as
j=1
A(j)
.
(42)
(j)
(s)
(s) α(j)
Here,
wn,k provides a Monte
j=1 υ xn,k , 1, an,k ; zn
R
(s)
(s)
Carlo approximation [44] of υ xn,k , 1, an,k ; zn α(xn,k , 1)
PJ+I α(j)
(s)
×dxn,k in (31), and 1 an,k 1 − j=1
wn,k provides an
PJ+I α(j)
(s)
approximation of 1 an,k αn,k in (31). Note that j=1
wn,k
can be interpreted as a “predicted existence probability,” and
PJ+I α(j)
thus 1 − j=1 wn,k can be interpreted as a “predicted
nonexistence probability,” which approximates αn,k .
PJ+I
C. Data Association, Measurement Update, Belief Calculation
(s)
The approximate messages β̃ an,k obtained in (42) are
used in the data association
loop, i.e., they are substituted for
(s)
the messages β an,k in (33) and (34). After convergence of
(s)
the data association loop, approximate messages η̃ an,k (ap(s)
proximating the messages η an,k in (35)) are then available
for all PTs k and all sensors s.
Next, the measurement update step (36) and the belief
calculation step (37), (38) are implemented by means of
2 Note
(44)
(s)
Here, we also replaced η an,k by its particle-based approxi(s)
mation η̃ an,k , even though we do not indicate this additional
approximation in our notation f˜(xn,k , rn,k ). We now calculate
nonnormalized weights corresponding to (43) as
s=1 a(s)
e
Here, pn−1,k
= j=1 w̄n−1,k (cf. (40)). Note that the pro(j)
posal distribution [43], [44] underlying (41) is f xn,k xn−1,k )
for j ∈ {1, . . . , J} and f b (xn,k ) for j ∈ {J + 1, . . . , J + I}.
A more general expression for particle-based prediction with
an arbitrary proposal distribution can be found in [16].
(43)
η̃ an,k = 0 .
n,k
PJ
(s)
(s)
η̃ an,k
υ xn,k , 1, an,k ; z(s)
n
that αn,k in (30) is again given implicitly by these weighted particles
since α(xn,k , rn,k ) is normalized.
(j)
wn,k
=
B +
wn,k
wn,k
PJ+I
j ′ =1
A(j ′ )
wn,k
.
PJ+I A(j)
B
Here, wn,k
+ j=1 wn,k is a particle-based approximation
of the normalization constant Cn,k in (39). We note that
e
pn,k
=
J+I
X
(j)
wn,k .
(45)
j=1
D. Target Detection, State Estimation, Resampling
(j)
(j) J+I
The weighted particles xn,k , wn,k j=1 can now be used
for target detection and estimation. First, for each PT k, an
e
approximation pn,k
of the existence probability p(rn,k = 1|z)
(j) J+I
is calculated from the particle weights wn,k j=1 as in (45).
e
PT k is then detected (i.e., considered to exist) if pn,k
is above
a threshold Pth (cf. Section II-B). For the detected targets k,
an approximation of the MMSE state estimate x̂MMSE
in (2)
n,k
is calculated according to
x̂n,k =
J+I
1 X
e
pn,k
(j)
(j)
wn,k xn,k .
(46)
j=1
Finally, as a preparation for the next time step n + 1, a resampling step [43], [44] is performed to reduce the number of
particles to J and to avoid degeneracy effects. The resampling
10
results in equally weighted particles
sponding weights are given by
j ∈ {1, . . . , J}.
VII. C HOICE
OF
B IRTH
(j)
w̄n,k
AND
=
J
; the correj=1
PI+J (j ′ )
1
w̄n,k = J j ′ =1 wn,k ,
(j)
x̄n,k
S URVIVAL PARAMETERS
We next present a scheme for choosing the birth pdfs
f b (xn,k ), birth probabilities pbn,k , and survival probabilities
psn,k (see Section III-A). This scheme is heuristic but results
in scalability with respect to the number of sensors S and, as
demonstrated in Section IX-B, in good detection and tracking
performance. It is based on the standard assumption that the
number of newly born targets obeys a Poisson distribution with
mean µb , and existing targets survive with a fixed, specified
probability ps [4].
We first distinguish between “reliable” and “unreliable” PTs
at time n − 1 by comparing the PT existence probabilities
e
pn−1,k
with a reliability threshold Rth : PT k is considered
e
reliable at time n−1 if pn−1,k
> Rth and unreliable otherwise.
r
u
Let Kn−1 and Kn−1 denote the sets of indices k of reliable and
r
unreliable PTs at time n −1, respectively. For PTs k ∈ Kn−1
,
b
we set the birth and survival probabilities at time n to pn,k = 0
and psn,k = ps , respectively. Since pbn,k = 0, no birth pdf is
needed at time n (cf. (7)).
u
u
|,
, we set psn,k = 0 and pbn,k = µb /|Kn−1
For PTs k ∈ Kn−1
and we construct the birth pdf f b (xn,k ) as follows. Consider
(s0 )
(s0 )
an arbitrary sensor s0 , and let Zn−1
, zn−1,m
(s0 )
m∈Mn−1
denote the set of measurements of that sensor at time n −1.
(s0 )
(Note that Zn−1
corresponds to the measurement vector
(s0 )
(s0 )
zn−1 = zn−1,m ]m∈M(s0 ) , with the difference that the
(s )
n−1
(s )
0
0
elements zn−1,m
are considered unordered in Zn−1
.) We
(s0 )
u
b
partition Zn−1 into disjoint subsets Zn−1,k
, k ∈ Kn−1
such
b
that the cardinalities of the Zn−1,k differ at most by 1, i.e.,
b
b
u
. Then, based on
|Zn−1,k
| − |Zn−1,l
| ≤ 1 for any k, l ∈ Kn−1
b
the kth measurement set Zn−1,k , we construct a corresponding
“adaptive birth pdf” as
Z
b
f b (xn,k ) , f (xn,k |xn−1,k )f b (xn−1,k ; Zn−1,k
)dxn−1,k ,
b
b
) is constructed using Zn−1,k
where the pdf f b (xn−1,k ; Zn−1,k
and prior knowledge (e.g., about the target velocity) as discussed in [45]. Particles representing f b (xn,k ) are obtained
b
) and performing
by drawing particles from f b (xn−1,k ; Zn−1,k
particle-based prediction [44].
VIII. R ELATION
TO
•
E XISTING M ETHODS
Several aspects of the proposed method are related to
existing methods, as discussed next.
• The hybrid model for data association using both targetoriented and measurement-oriented association variables
was previously proposed in [32] and [33]. In [32], BP is
used to estimate optical flow parameters. In [33], BP is
used for data association (without multitarget tracking).
• Our model for target existence was previously used by
the search-initialize-track filter in [11], which, however,
is not BP-based, considers only a single sensor, does not
•
•
employ the hybrid model for data association, and uses
a different track initialization scheme.
In the case of a single target and a single sensor, the
proposed method reduces to the particle-based implementation of the Bernoulli filter [16] (which is derived using
the FISST framework).
The TOMB/P filter [20], which is effectively a FISSTbased variant of the JIPDA filter that uses BP and the
hybrid data association model, differs from the proposed
method in the following respects: it is restricted to
a single sensor and to linear-Gaussian state evolution
and measurement models—see [46] for an extension to
nonlinear, non-Gaussian models—and the number of PTs
(tracks) varies over time.
If the parameters of the proposed method are chosen
such that all targets exist at all times (this special case
was mentioned in Section III-A, and was considered in
our previous work in [35]), then the method becomes
similar to the Monte Carlo JPDA filter [8] in that it uses
a similar particle-based processing scheme. However,
contrary to the Monte Carlo JPDA filter, the proposed
method performs data association by means of BP, is
based on the hybrid model for data association, and can
also be used when the number of targets is unknown.
IX. S IMULATION R ESULTS
Next, we report simulation results assessing the performance
of our method and comparing it with that of five previously
proposed methods for multisensor-multitarget tracking.
A. Simulation Setting
We simulated up to five actual targets whose states consist
of two-dimensional (2D) position and velocity, i.e., xn,k =
[x1,n,k x2,n,k ẋ1,n,k ẋ2,n,k ]T. The targets move in a region
of interest (ROI) given by [−3000, 3000] × [−3000, 3000]
according to the constant-velocity motion model, i.e., xn,k =
Axn−1,k + Wun,k , where A ∈ R4×4 and W ∈ R4×2 are
chosen as in [47] and un,k ∼ N (0, σu2 I2 ) with σu2 = 0.025
is an independent and identically distributed (iid) sequence of
2D Gaussian random vectors. The birth distribution fb (xn,k ),
b
the birth probabilities pn,k
, and the survival probabilities
s
pn,k were chosen as described in Section VII, using the
global birth probability p b = 0.01 and the global survival
probability p s = 0.999. The number of PTs was set to K = 8.
We considered a challenging scenario where the five target
trajectories intersect at the ROI center. The target trajectories
were generated by first assuming that the five targets start from
initial positions uniformly placed on a circle of radius 1000
and move with an initial speed of 10 toward the ROI center,
and then letting the targets start to exist at times n = 5, 10,
15, 20, and 25.
The sensors are located uniformly on a circle of radius
3000 and perform range and bearing measurements within
a measurement range of 6000. More specifically, within the
measurement range, the target-generated measurements are
given by
T
(s)
+ vn,m
,
x̃n,k − p(s) ϕ x̃n,k , p(s)
z(s)
n,m =
11
B. Performance Comparison
We compare the proposed BP method with particle implementations of the IC-PHD filter [4], [13], [26], the ICCPHD filter [4], [15], [26], the IC-MB filter [4], [17], and
the partition-based MS-PHD and MS-CPHD filters [23], [24].
The “IC-” filters are straightforward multisensor extensions
performing a single-sensor update step sequentially for each
sensor [4], [6], [26]. The partition-based MS-(C)PHD filters
approximate the exact multisensor (C)PHD filters; they can
outperform the “IC-” filters but have a higher computational
complexity. Since the trellis algorithm used for partition extraction in the original formulation of the MS-(C)PHD filter
[23], [24] is only suitable for a Gaussian mixture implementation of the filter, it was adapted to a particle-based implementation. We note that the exact multisensor (C)PHD filters are
not computationally feasible for the simulated scenario since
their complexity scales exponentially in the number of sensors
and in the number of measurements per sensor [22], [25], [48].
The performance of the various methods is measured by the
Euclidean distance based OSPA metric with a cutoff parameter
of 200 [49].
The (C)PHD-type filters use 24000 particles to represent
the PHD of the target states, and they perform kmeans++
clustering [50] for state estimation. The IC-MB filter uses
3000 particles to represent each Bernoulli component. The
maximum numbers of subsets and partitions used by the MS(C)PHD filter are 120 and 720, respectively, similarly to [23],
[24]. With the above-mentioned parameters, the runtime per
time scan for a MATLAB implementation without gating on a
single core of an Intel Xeon E5-2640 v3 CPU was measured
as 0.07s for the proposed method, 0.11s for the IC-PHD filter,
0.14s for the IC-CPHD filter, 0.27s for the IC-MB filter,
13.21s for the MS-PHD filter, and 13.82s for MS-CPHD filter.
The high runtimes of the MS-(C)PHD filters are due to the fact
that in our particle-based implementation, the trellis algorithm
used for partition extraction is computationally intensive. We
note that the efficient extraction of high-quality partitions in a
particle-based implementation of the MS-(C)PHD filter is still
an open problem.
200
IC-MB
IC-PHD
MS-PHD
IC-CPHD
MS-CPHD
BP (proposed)
MOSPA
160
120
80
40
0
25
50
75
100
125
150
n
Fig. 4. MOSPA error versus time n for S = 3 sensors, Pd = 0.8, and µ(s) = 2.
120
IC-MB
IC-PHD
MS-PHD
IC-CPHD
MS-CPHD
BP (proposed)
100
80
MOSPA
T
(s)
where x̃n,k , [x
1,n,k x2,n,k ] , p is the position of sensor
(s)
s, ϕ x̃n,k , p
is the angle (in degrees) of the vector x̃n,k
(s)
relative to the vector p(s), and vn,m ∼ N (0, Cv ) with Cv =
2
2
diag{10 , 0.5 } is an iid sequence of 2D Gaussian random
(s)
vectors. The false alarm pdf fFA zn,m is linearly increasing
on [0, 6000] and zero outside [0, 6000] with respect to the range
component, and uniform on [0◦ , 360◦) with respect to the
angle component. In Cartesian coordinates, this corresponds
to a uniform distribution on the sensor’s measurement area.
The mean number of false alarm measurements is µ(s) = 2 if
not noted otherwise.
Our implementation of the proposed method used J = 3000
particles and I = 3000 birth particles for each PT. We
performed P = 20 BP iterations for iterative data association.
The threshold for target detection was Pth = 0.5, and the
reliability threshold was Rth = 10−3. We simulated 150 time
steps n.
60
40
20
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Pd
Fig. 5. Time-averaged MOSPA error versus Pd for S = 3 sensors and µ(s) =
2.
Fig. 4 shows the mean OSPA (MOSPA) error—averaged
over 400 simulation runs—of all methods versus time n,
assuming S = 3 sensors with a detection probability of
(s)
Pd (xn,k ) = Pd = 0.8. The error exhibits peaks at times
n = 5, 10, 15, 20, and 25 because of target births. However,
very soon after a target birth, the proposed method as well
as the IC-CPHD, MS-PHD, and MS-CPHD filters are able to
reliably estimate the number of targets. The proposed method
is seen to outperform all the other methods. In particular, it
outperforms the IC-CPHD, MS-PHD, and MS-CPHD filters
mainly because particle implementations of (C)PHD filters
involve a potentially unreliable clustering step. This clustering
step is especially unreliable for targets that are close to each
other. This fact explains the higher MOSPA error of the ICCPHD, MS-PHD, and MS-CPHD filters around n = 100, i.e.,
around the time when the target trajectories intersect in the
ROI center. Finally, the MOSPA error of the IC-PHD and
IC-MB filters is seen to be significantly larger than that of
the other methods; this is caused by the inability of these
filters to reliably estimate the number of targets. We note
that for sensors with different probabilities of detection, the
performance loss of IC-(C)PHD filters relative to MS-(C)PHD
filters tends to be larger than in Fig. 4 [23], [24].
Fig. 5 shows the time-averaged MOSPA error—averaged
over time steps n ∈ {50, . . . , 150}—versus the detection
probability Pd for S = 3 sensors. For all methods, as expected,
the MOSPA error decreases with decreasing Pd . Fig. 6 shows
the time-averaged MOSPA error versus the number of sensors
12
0.2
IC-MB
IC-PHD
MS-PHD
IC-CPHD
MS-CPHD
BP (proposed)
100
80
MOSPA
runtime (s)
120
0.15
0.1
0.05
0
60
2
4
6
8
10
12
14
16
18
20
50
60
70
80
90
14
16
18
20
S
40
20
0
1
2
3
5
4
7
6
8
runtime (s)
0.8
S
0.6
0.4
0.2
0
Fig. 6. Time-averaged MOSPA error versus number of sensors S for Pd = 0.6
and µ(s) = 2.
1
10
20
30
40
µ(s)
IC-MB
IC-PHD
MS-PHD
IC-CPHD
MS-CPHD
BP (proposed)
120
100
MOSPA
runtime (s)
0.4
80
0.3
0.2
0.1
0
2
4
6
8
10
12
number of actual targets
60
Fig. 8. Average runtime per time step of the proposed method. Top: versus
S for five actual targets and µ(s) = 2; center: versus µ(s) for S = 3 and five
actual targets; bottom: versus the number of actual targets for S = 3, µ(s) = 2,
and the number of PTs set to K = 5, 7, . . . , 23.
40
20
0
1
10
30
20
40
50
µ(s)
Fig. 7. Time-averaged MOSPA error versus mean number of false alarms
µ(s) for Pd = 0.6 and S = 3.
S for Pd = 0.6. It can be seen that the MOSPA error of the
MS-PHD and MS-CPHD filters increases for S larger than 5;
this is because the chosen maximum numbers of subsets (120)
and partitions (720) are too small for that case. (We note that
choosing larger maximum numbers of subsets and partitions
leads to excessive simulation times.) Finally, Fig. 7 shows the
time-averaged MOSPA error versus the mean number of false
alarms µ(s) for S = 3 and Pd = 0.6. As expected, the MOSPA
error of all methods increases with growing µ(s). In addition,
Figs. 5–7 again show that the proposed method outperforms
the other methods. We note that the poor performance of the
IC-MB filter is due to the approximation used by that filter,
which is accurate only for a high Pd and a very low µ(s) [17].
C. Scalability
Finally, we investigate how the runtime of our method scales
in the number of sensors S and the number of actual targets,
and how it depends on the mean number of false alarms µ(s).
Fig. 8 shows the average runtime per time step n versus
S, µ(s), and the number of actual targets for a MATLAB
implementation on a single core of an Intel Xeon E5-2640
v3 CPU. The runtime was averaged over 150 time steps and
400 simulation runs. The probability of detection was set to
Pd = 0.6. These results confirm the linear scaling of the
runtime in S and in the mean number of measurements per
sensor (which grows linearly with µ(s) ). The scaling in the
number of actual targets is seen to be roughly quadratic.
Further investigation showed that the scaling in the number
of actual targets is linear if the number of PTs is held fixed,
and similarly, the scaling in the number of PTs is linear if
the number of actual targets is held fixed. The low absolute
complexity of the proposed method is evidenced by the fact
that for 20 actual targets, S = 3 sensors, and µ(s) = 2, the
computations per time step n require less than 0.4s.
X. C ONCLUSION
We developed and demonstrated the application of the
belief propagation (BP) scheme to the problem of tracking
an unknown number of targets using multiple sensors. The
proposed BP-based multitarget tracking method exhibits low
complexity and excellent scaling properties with respect to
all relevant systems parameters. This is achieved through
the use of “augmented target states” including binary target
indicators and the establishment of an appropriate statistical
model involving a redundant formulation of data association
uncertainty [33]. The complexity of our method scales only
quadratically in the number of targets, linearly in the number
of sensors, and linearly in the number of measurements per
sensors. Simulation results in a challenging scenario with
intersecting targets showed that the proposed method can outperform previously proposed methods, including methods with
a less favorable scaling behavior. In particular, we observed
significant improvements in OSPA performance relative to
various multisensor extensions of the PHD, CPHD, and multiBernoulli filters.
Promising directions for future research include extensions
of the proposed BP method that adapt to time-varying en-
13
vironmental conditions, e.g., to a time-varying probability of
detection [51], and distributed variants for use in decentralized
wireless sensor networks with communication constraints [52].
A direction of theoretical interest would be a FISST-based
derivation of multisensor-multitarget tracking algorithms using
BP for data association.
R EFERENCES
[1] W. Koch, Tracking and Sensor Data Fusion: Methodological Framework
and Selected Applications. Berlin, Germany: Springer, 2014.
[2] B.-N. Vo, M. Mallick, Y. Bar-Shalom, S. Coraluppi, R. Osborne,
R. Mahler, and B.-T. Vo, “Multitarget tracking,” in Wiley Encyclopedia
of Electrical and Electronics Engineering, M. Peterca, Ed. Hoboken,
NJ, USA: Wiley, 2015.
[3] Y. Bar-Shalom and X.-R. Li, Multitarget-Multisensor Tracking: Principles and Techniques. Storrs, CT, USA: Yaakov Bar-Shalom, 1995.
[4] R. Mahler, Statistical Multisource-Multitarget Information Fusion. Norwood, MA, USA: Artech House, 2007.
[5] D. B. Reid, “An algorithm for tracking multiple targets,” IEEE Trans.
Autom. Control, vol. 24, no. 6, pp. 843–854, Dec. 1979.
[6] L. Y. Pao and C. W. Frei, “A comparison of parallel and sequential
implementations of a multisensor multitarget tracking algorithm,” in
Proc. ACC-95, vol. 3, Seattle, WA, USA, Jun. 1995, pp. 1683–1687.
[7] S. Deb, M. Yeddanapudi, K. Pattipati, and Y. Bar-Shalom, “A generalized S-D assignment algorithm for multisensor-multitarget state
estimation,” IEEE Trans. Aerosp. Electron. Syst., vol. 33, no. 2, pp.
523–538, Apr. 1997.
[8] J. Vermaak, S. J. Godsill, and P. Perez, “Monte Carlo filtering for multi
target tracking and data association,” IEEE Trans. Aerosp. Electron.
Syst., vol. 41, no. 1, pp. 309–332, Jan. 2005.
[9] D. Musicki and R. Evans, “Joint integrated probabilistic data association:
JIPDA,” IEEE Trans. Aerosp. Electron. Syst., vol. 40, no. 3, pp. 1093–
1099, Jul. 2004.
[10] D. Musicki and R. J. Evans, “Multiscan multitarget tracking in clutter
with integrated track splitting filter,” IEEE Trans. Aerosp. Electron. Syst.,
vol. 45, no. 4, pp. 1432–1447, Oct. 2009.
[11] P. Horridge and S. Maskell, “Searching for, initiating and tracking
multiple targets using existence probabilities,” in Proc. FUSION-09,
Seattle, WA, USA, Jul. 2009, pp. 611–617.
[12] R. Mahler, “Multitarget Bayes filtering via first-order multitarget moments,” IEEE Trans. Aerosp. Electron. Syst., vol. 39, no. 4, pp. 1152–
1178, Oct. 2003.
[13] B.-N. Vo, S. Singh, and A. Doucet, “Sequential Monte Carlo methods
for multitarget filtering with random finite sets,” IEEE Trans. Aerosp.
Electron. Syst., vol. 41, no. 4, pp. 1224–1245, Oct. 2005.
[14] R. Mahler, “PHD filters of higher order in target number,” IEEE Trans.
Aerosp. Electron. Syst., vol. 43, no. 4, pp. 1523–1543, Oct. 2007.
[15] B.-T. Vo, B.-N. Vo, and A. Cantoni, “Analytic implementations of the
cardinalized probability hypothesis density filter,” IEEE Trans. Signal
Process., vol. 55, no. 7, pp. 3553–3567, Jul. 2007.
[16] B. Ristic, B.-T. Vo, B.-N. Vo, and A. Farina, “A tutorial on Bernoulli
filters: Theory, implementation and applications,” IEEE Trans. Signal
Process., vol. 61, no. 13, pp. 3406–3430, Jul. 2013.
[17] B.-T. Vo, B.-N. Vo, and A. Cantoni, “The cardinality balanced multitarget multi-Bernoulli filter and its implementations,” IEEE Trans. Signal
Process., vol. 57, no. 2, pp. 409–423, Feb. 2009.
[18] S. Reuter, B.-T. Vo, B.-N. Vo, and K. Dietmayer, “The labeled multiBernoulli filter,” IEEE Trans. Signal Process., vol. 62, no. 12, pp. 3246–
3260, Jun. 2014.
[19] B.-N. Vo, B.-T. Vo, and D. Phung, “Labeled random finite sets and the
Bayes multi-target tracking filter,” IEEE Trans. Signal Process., vol. 62,
no. 24, pp. 6554–6567, Dec. 2014.
[20] J. L. Williams, “Marginal multi-Bernoulli filters: RFS derivation of
MHT, JIPDA and association-based MeMBer,” IEEE Trans. Aerosp.
Electron. Syst., vol. 51, no. 3, pp. 1664–1687, Jul. 2015.
[21] P. Braca, S. Marano, V. Matta, and P. Willett, “Asymptotic efficiency
of the PHD in multitarget/multisensor estimation,” IEEE J. Sel. Topics
Signal Process., vol. 7, no. 3, pp. 553–564, Jun. 2013.
[22] R. Mahler, “The multisensor PHD filter: I. General solution via multitarget calculus,” in Proc. SPIE-09, Orlando, FL, USA, Apr. 2009.
[23] S. Nannuru, M. Coates, M. Rabbat, and S. Blouin, “General solution
and approximate implementation of the multisensor multitarget CPHD
filter,” in Proc. IEEE ICASSP-15, Brisbane, Australia, Apr. 2015, pp.
4055–4059.
[24] ——, “Multisensor CPHD filter,” 2016, available online:
http://arxiv.org/abs/1504.06342.
[25] E. Delande, E. Duflos, P. Vanheeghe, and D. Heurguier, “Multi-sensor
PHD: Construction and implementation by space partitioning,” in Proc.
IEEE ICASSP-11, Prague, Czech Republic, May 2011, pp. 3632–3635.
[26] S. Nagappa and D. Clark, “On the ordering of the sensors in the iteratedcorrector probability hypothesis density (PHD) filter,” in Proc. SPIE-11,
Orlando, FL, USA, Apr. 2011, pp. 26–28.
[27] R. Mahler, “Approximate multisensor CPHD and PHD filters,” in Proc.
FUSION-10, Edinburgh, UK, Jul. 2010, pp. 26–29.
[28] F. R. Kschischang, B. J. Frey, and H.-A. Loeliger, “Factor graphs and
the sum-product algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp.
498–519, Feb. 2001.
[29] H. Wymeersch, Iterative Receiver Design.
New York, NY, USA:
Cambridge University Press, 2007.
[30] M. J. Wainwright and M. I. Jordan, “Graphical models, exponential
families, and variational inference,” Found. Trends Mach. Learn., vol. 1,
Jan. 2008.
[31] H.-A. Loeliger, “An introduction to factor graphs,” IEEE Signal Process.
Mag., vol. 21, no. 1, pp. 28–41, Jan. 2004.
[32] M. Chertkov, L. Kroc, F. Krzakala, M. Vergassola, and L. Zdeborová,
“Inference in particle tracking experiments by passing messages between
images,” PNAS, vol. 107, no. 17, pp. 7663–7668, Apr. 2010.
[33] J. L. Williams and R. Lau, “Approximate evaluation of marginal association probabilities with belief propagation,” IEEE Trans. Aerosp.
Electron. Syst., vol. 50, no. 4, pp. 2942–2959, Oct. 2014.
[34] Z. Chen, L. Chen, M. Cetin, and A. S. Willsky, “An efficient message
passing algorithm for multi-target tracking.” in Proc. FUSION-09,
Seattle, WA, USA, Jul. 2009, pp. 826–833.
[35] F. Meyer, P. Braca, P. Willett, and F. Hlawatsch, “Scalable multitarget
tracking using multiple sensors: A belief propagation approach,” in Proc.
FUSION-15, Washington D.C., USA, Jul. 2015, pp. 1778–1785.
[36] P. Horridge and S. Maskell, “Real-time tracking of hundreds of targets
with efficient exact JPDAF implementation,” in Proc. FUSION-06,
Florence, Italy, Jul. 2006, pp. 1–8.
[37] L. Chen, M. J. Wainwright, M. Cetin, and A. S. Willsky, “Data
association based on optimization in graphical models with application
to sensor networks,” Math. Comp. Model., vol. 43, no. 9–10, pp. 1114–
1135, 2006.
[38] F. Meyer, P. Braca, P. Willett, and F. Hlawatsch, “Tracking an unknown
number of targets using multiple sensors: A belief propagation method,”
in Proc. FUSION-16, Heidelberg, Germany, Jul. 2016.
[39] H. V. Poor, An Introduction to Signal Detection and Estimation. New
York, NY: Springer, 1994.
[40] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Upper Saddle River, NJ, USA: Prentice-Hall, 1993.
[41] H. Wymeersch, J. Lien, and M. Z. Win, “Cooperative localization in
wireless networks,” Proc. IEEE, vol. 97, no. 2, pp. 427–450, Feb. 2009.
[42] P. O. Vontobel, “The Bethe permanent of a nonnegative matrix,” IEEE
Trans. Inf. Theory, vol. 59, no. 3, pp. 1866–1901, Mar. 2013.
[43] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A tutorial
on particle filters for online nonlinear/non-Gaussian Bayesian tracking,”
IEEE Trans. Signal Process., vol. 50, no. 2, pp. 174–188, Feb. 2002.
[44] A. Doucet, N. de Freitas, and N. Gordon, Sequential Monte Carlo
Methods in Practice. New York, NY, USA: Springer, 2001.
[45] B. Ristic and S. Arulampalam, “Bernoulli particle filter with observer
control for bearings-only tracking in clutter,” IEEE Trans. Aerosp.
Electron. Syst., vol. 48, no. 3, pp. 2405–2415, Jul. 2012.
[46] T. Kropfreiter, F. Meyer, and F. Hlawatsch, “Sequential Monte Carlo
implementation of the track-oriented marginal multi-Bernoulli/Poisson
filter,” in Proc. FUSION-16, Heidelberg, Germany, Jul. 2016.
[47] J. H. Kotecha and P. M. Djuric, “Gaussian particle filtering,” IEEE Trans.
Signal Process., vol. 51, no. 10, pp. 2592–2601, Oct 2003.
[48] R. Mahler, “The multisensor PHD filter: II. Erroneous solution via
Poisson magic,” in Proc. SPIE-09, Orlando, FL, USA, Apr. 2009.
[49] D. Schuhmacher, B.-T. Vo, and B.-N. Vo, “A consistent metric for
performance evaluation of multi-object filters,” IEEE Trans. Signal
Process., vol. 56, no. 8, pp. 3447–3457, Aug. 2008.
[50] G. Gan, C. Ma, and J. Wu, Data Clustering: Theory, Algorithms, and
Applications. Philadelphia, PA, USA: SIAM, 2007.
[51] B.-T. Vo, B.-N. Vo, R. Hoseinnezhad, and R. Mahler, “Robust multiBernoulli filtering,” IEEE J. Sel. Topics Signal Process., vol. 7, no. 3,
pp. 399–409, Jun. 2013.
[52] K. Sohraby, D. Minoli, and T. Znati, Wireless Sensor Networks: Technology, Protocols, and Applications. Hoboken, NJ, USA: Wiley, 2007.
| 8cs.DS
|
1
Joint Relay Selection and Power Allocation in
Large-Scale MIMO Systems with Untrusted Relays
and Passive Eavesdroppers
arXiv:1709.01341v1 [cs.IT] 5 Sep 2017
Ali Kuhestani, Student Member, IEEE, Abbas Mohammadi Senior M ember, IEEE ,
and Mohammadali Mohammadi, M ember, IEEE
Abstract—In this paper, a joint relay selection and power
allocation (JRP) scheme is proposed to enhance the physical layer
security of a cooperative network, where a multiple antennas
source communicates with a single-antenna destination in presence of untrusted relays and passive eavesdroppers (Eves). The
objective is to protect the data confidentially while concurrently
relying on the untrusted relays as potential Eves to improve
both the security and reliability of the network. To realize this
objective, we consider cooperative jamming performed by the
destination while JRP scheme is implemented. With the aim
of maximizing the instantaneous secrecy rate, we derive a new
closed-form solution for the optimal power allocation and propose
a simple relay selection criterion under two scenarios of noncolluding Eves (NCE) and colluding Eves (CE). For the proposed
scheme, a new closed-form expression is derived for the ergodic
secrecy rate (ESR) and the secrecy outage probability as security
metrics, and a new closed-form expression is presented for the
average symbol error rate (SER) as a reliability measure over
Rayleigh fading channels. We further explicitly characterize the
high signal-to-noise ratio slope and power offset of the ESR
to highlight the impacts of system parameters on the ESR. In
addition, we examine the diversity order of the proposed scheme
to reveal the achievable secrecy performance advantage. Finally,
the secrecy and reliability diversity-multiplexing tradeoff of the
optimized network are provided. Numerical results highlight
that the ESR performance of the proposed JRP scheme for
NCE and CE cases is increased with respect to the number of
untrustworthy relays.
Index Terms—Physical-layer security, untrusted relay, joint
relay selection and power allocation.
I. I NTRODUCTION
L
ARGE-scale multiple-input multiple-output (MIMO) system as a promising solution of the fifth-generation (5G)
wireless communication networks provides significant performance gains in terms of energy saving and spectral efficiency
[1], [2]. This new technology deploys simple coherent processing methods, e.g., maximum ratio transmission (MRT)
across arrays of hundreds of or even more antennas at base
station (BS) and supports tens of or more mobile users (MUs)
[1]–[3]. An attractive feature of large-scale MIMO systems is
that they offer a significant security improvement compared
to a conventional MIMO systems, as with large-scale multiple
This work was supported by the INSF and contract number 94023761.
A. Kuhestani and A. Mohammadi are with Department of Electrical
Engineering, Amirkabir University of Technology, Tehran, Iran (emails:
[email protected]; [email protected]).
M. Mohammadi is with the Faculty of Engineering, Shahrekord University,
Shahrekord, Iran (e-mail: [email protected]).
antennas (LSMA) at the BS, a narrow directional beam can be
radiated toward the desired terminal. Accordingly, the received
signal power at the desired terminal is several orders of magnitude higher than that at any non-coherent passive eavesdropper
(Eve) [3]1 . However, the security benefits of LSMA systems
are severely hampered in cooperative networks, where the
intermediate nodes may be potential Eves [6].
Owing to the broadcast nature of wireless communication,
the information transmission between legitimate users can be
simply captured by Eves. Accordingly, physical layer security
(PLS) as a promising approach to enhance the confidentiality
of wireless communications has attracted a lot of interest
[7], [8]. Physical layer secure transmission is provisioned by
intelligently exploiting the time varying properties of fading
channels, instead of relying on conventional cryptographic
techniques [7]. Among the proposed PLS solutions, cooperative relaying, cooperative jamming (CJ) and a mixed of these
two techniques have recently attracted a great deal of interest
[9]–[18]. Cooperative relaying can enhance the PLS through
implementing distributed beamforming or opportunistic relaying (OR) [9], [10]. In contrast to distributed beamforming
that suffers from high complexity due to the need of large
overhead, OR imposes low system overhead by choosing only
one relay which offers the best secrecy performance for the
network [10]. As a complementary approach, CJ schemes
can also be implemented by the network’s nodes to transmit
jamming signals towards the Eves in order to degrade the
received signal-to-noise ratio (SNR) at the Eves [11]–[18].
This goal can be achieved through applying the following
three methods: 1) Source-based jamming [11], [12] in which
the source transmits a mixed signal carrying the information
signal and the jamming signal, 2) Friendly jammer-based
jamming [12]–[15] in which a relay [12]–[14] or an external
friendly jammer (FJ) [15] contributes to provide confidential
communication, and 3) Destination-based CJ (DBCJ) [16]–
[24], in which the destination itself contributes to degrade the
received signal at an external Eve [16]–[18] or at a helper
intermediate node who may act as an Eve [6]. Among the
presented methods, the DBCJ policy can be implemented
simply compared to the first and second methods, where
the destination performs self-interference cancellation with
regard to its prior information of jamming signal. Contrary to
1 It is worth noting that in contrast to passive eavesdropping, an active
Eve may cause pilot contamination at the BS and pose more serious security
problem to the network [4], [5].
2
DBCJ, for the first and second methods, the destination must
be perfectly aware of the pre-defined jamming signal, while
it is unknown for the Eves. This a priori known jamming
signal, generated by using some pseudo-random codes or some
cryptographic signals, is hard to implement and transfer to the
destination confidentially. Transferring this known jamming
signal from the source to destination imposes more challenges
to the network, when an untrusted relay collaborates for data
transmission.
From a perspective of security, a trusted relay can friendly
assist to protect the confidential message from being eavesdropped by illegitimate nodes, while an untrusted relay may
intentionally overhear the information signal when relaying.
In some communication networks, an untrusted relay may
collaborate to provide a reliable communication [6]. This
scenario occurs in large-scale wireless systems such as heterogeneous networks, device-to-device (D2D) communications
and Internet-of-things (IoT) applications [5], where confidential messages are often retransmitted by multiple intermediate nodes. It is therefore necessary to answer this question
that whether exploiting the untrusted relay is still beneficial
compared with direct transmission (DT) and if so, what the
appropriate relaying strategy should be.
To achieve a positive secrecy rate in untrusted relaying
networks, the DBCJ policy was first introduced in [6]. Then
several recent works have investigated the performance of the
DBCJ policy in presence of a single [19], [20] or multiple
[21]–[24] untrusted relays. Specifically, the secrecy performance of DBCJ with optimal power allocation (OPA) in presence of an untrusted amplify-and-forward (AF) relay investigated in [19] and then comprehensively studied in [20]. Taking
into account non-colluding untrusted relaying, the authors in
[21]–[23] derived lower bound expressions for the ergodic
secrecy rate (ESR) performance in the absence [21], [22] or
presence [23] of source jamming. The researchers in [21]–[23]
indicate that increasing the number of untrustworthy relays
degrades the ESR in contrast to the case of trustworthy relays.
To be specific, the authors in [22] found that the diversity
order of OR is restricted to unity independent of the number
of untrusted relays. All the aforementioned works [21]–[24]
investigated the cooperative untrusted relaying networks in
absence of passive Eves and without considering achievable
secrecy rate optimization.
While the recent literature [19] presented a solid work for
OPA in presence of a single untrusted relay, the impact of
OPA on a more realistic cooperative network with multiple
untrusted relays and passive Eves has not been studied yet. In
this paper, motivated by the recent literature on LSMA-based
relaying systems [1], [2], [5], [19], [26], [27], we present a
comprehensive research to improve the PLS of cooperative
networks in presence of untrusted relays and passive Eves2 .
Along this line, we consider a new cooperative network consisting of a large-scale MIMO source, a destination, multiple
non-colluding untrusted AF relays [21]–[24], and multiple
non-colluding Eves (NCE) [10], [28] who hide their existence
in the network. In contrast to vast studies on the secrecy
2 Henceforth,
we interchangeably use the terms ‘passive Eve’ and ‘Eve’.
performance of untrusted relaying networks [19]–[24], we take
into account joint relay selection and power allocation (JRP)
in our network by considering the information leakage in the
second phase of transmission. We further investigate a worstcase scenario, where passive Eves may collaborate together
to maximize the total received SNR. Toward this end, we
consider the scenario that the maximal ratio combining (MRC)
is performed over all the received signals from the source
and the selected relay to perform more harmful attacks. This
event is referred to as colluding Eves (CE) [28], [29]. As
a benchmark, we next study the conventional DT dispensing
with the relays to compare with the proposed JRP transmission
policy. In contrast to [30], in which DT policy with transmit
antenna selection is analyzed for MIMO wiretap channels,
we utilize the simple MRT beamformer [2] in our proposed
DT policy to maximize the received SNR at the intended
receiver. Furthermore, different from [30] that considered one
multiple antennas Eve, we consider two cases of NCE and
CE. Therefore, our system model and the related analysis are
completely different compared with [30].
The main contributions of the paper are summarized as
follows:
1) We develop a JRP scheme to maximize the instantaneous
secrecy rate of the network for both NCE and CE
cases. Our findings highlight that the large-scale MIMO
approach as a powerful mathematical tool offers a new
simple relay selection criterion. The proposed criterion
only requires the channel state information (CSI) of the
relays-destination links.
2) Based on the proposed JRP scheme, new closed-form
expressions are derived for the probability of positive
secrecy rate and the ESR of NCE and CE cases over
Rayleigh fading channels. Furthermore, new compact
expressions are derived for the asymptotic ESR. Our
asymptotic results highlight that the probability of positive secrecy transmission tends to one for the optimized
JRP scheme as the number of relays K grows. Moreover,
we find that the ESR performance improves as K
increases. We further characterize the high SNR slope
and power offset of the ESR to highlight the impact of
system parameters on the ESR performance.
3) For the proposed JRP scheme, new closed-form expressions are derived for the secrecy outage probability
(SOP) of both NCE and CE cases over Rayleigh fading
channels. In order to shed insights into the system
performance, new simple expressions are derived for
the SOP in the high SNR regime. Our asymptotic
results highlight that the proposed JRP scheme enjoys
the diversity order of K. Next, the secrecy diversitymultiplexing trade-off (DMT) is examined to express the
trade-off between the error probability and the data rate
of the proposed JRP scheme.
4) For DT policy, a new tight lower bound is derived for
the ESR of NCE and CE cases. We also derive a new
closed-form expression for the SOP of those cases. Our
results show that the high SNR slope and the secrecy
diversity order of this transmission policy are zero.
3
First Phase
Second Phase
Relay
1
Eve 1
Eve
Relay
i
Destination
Eve L
Eve
Source
Relay
K
Fig. 1. Secure transmission in a cooperative network in presence of multiple
untrusted relays and multiple passive Eves. Under DBCJ policy, relay i is
selected to amplify the source’s signal. The solid and dashed lines denote the
first and second phases of transmission, respectively.
5) To illustrate the reliability of the network, we calculate
the average symbol error rate (SER) as the performance
measure. New exact and asymptotic closed-form expressions are derived for the SER performance of the
proposed JRP scheme over Rayleigh fading channels.
Our results highlight that the proposed scheme offers the
diversity order of K. We further determine the reliability
DMT of the proposed scheme.
The remainder of this paper is organized as follows. In
Section II we present the system model and preliminaries.
Next, in Section III, the relay selection criterion is introduced
and the related OPA is evaluated. Performance metrics are
evaluated in Section IV. Numerical results are presented in
Section V, followed by conclusions in Section VI.
Notation: We use bold lower case letters to denote vectors.
IN and 0N ×1 denote the Identity matrix and the zeros matrix,
respectively. k.k and (.)H denote the Euclidean norm and
conjugate transpose operator, respectively; Ex {·} stands for
the expectation over the random variable x; Pr(·) denotes the
probability; fX (·) and FX (·) denote the probability density
function (pdf) and cumulative distribution function (cdf) of
the random variable (r.v.) X, respectively; CN (µ, σ 2 ) denotes
a circularly symmetric complex Gaussian r.v. with mean µ
and variance σ 2 ; Ei(x), Q(x) and Ψ(x) are the exponential
integral [31, Section (8.21)], the Q-function [31, Section
(8.25)] and the psi function [31, Section (8.25)], respectively.
[·]+ = max{0, x} and max stands for the maximum value.
⌈x⌉ is the smallest integer that is larger than or equal to x.
II. S YSTEM M ODEL
We consider a cooperative LSMA-based network consisting
of one multiple antennas source equipped with an array of
Ns antennas, one single-antenna destination, K single-antenna
untrusted AF relays and L single-antenna passive Eves3 as
depicted in Fig. 1. The untrusted relays in our network are
so-called semi-trusted, i.e., they are trusted at the service
level while they are untrusted at the data level4 [6]. All the
passive Eves are randomly located around the source, the relay
nodes and the destination, and they hide their existence in the
network. We mention that, in this paper, the term “malicious
node” includes both the untrusted relays and passive Eves.
We further assume that each node operates in a half-duplex
mode. Based on time division duplexing (TDD) operation, the
source obtains downlink CSI through uplink training. Then
the acquired CSI is used to generate the low implementation
complexity MRT precoding matrix [26].
We assume that the untrusted relays extract the information
signal solely based on its own observation and they adopt
selection combining (SC) [23], while relying on the Eves’
behavior, the following two eavesdropping scenarios are considered in our work:
1) Non-colluding eavesdroppers case: In this case, each
passive Eve individually overhears the information signal without any collaboration with other Eves. For this
case, we assume SC is applied at the Eves. As will be
validated via numerical examples, the presented analysis
can still be utilized for the case of MRC at the malicious
nodes.
2) Colluding eavesdroppers case: In this case, all the
passive Eves can connect to a data center to share their
information, leading to extract more information. For
this case, we assume MRC technique is applied at the
Eves to enhance the intercept probability. This case can
be considered as a worst-case scenario from the security
viewpoint [28].
Some additional assumptions and definitions are as follows:
• Complex Gaussian channel vector from the source to the
malicious node l: hsl ∼ CN (0Ns ×1 , µsl INs ).
• Complex Gaussian channel vector from the source to the
destination: hsd ∼ CN (0Ns ×1 , µsd INs ).
• Complex Gaussian channel from the selected relay i to
the destination: hid ∼ CN (0, µid ).
• Complex Gaussian channel from the selected relay i to
the malicious node l: hil ∼ CN (0, µil ).
• All the channels satisfy the reciprocity theorem [21].
• The additive white Gaussian noise (AWGN) at each
receiver nm ,m ∈ {i, l, d}, is a zero-mean complex
Gaussian r.v. with variance N0 .
• The total transmit power of each phase is limited by
P . One practical reason is the impact of co-channel
interference between adjacent networks which should be
taken into account in network design.
P
• ρ = N is the transmit SNR of the system.
0
3 The presence of multiple Eves in a wireless network may be a realistic
scenario in which the malicious nodes act as harmful attackers to the
authorized terminals [5], [29].
4 The service level trust conveys this concept that the accurate CSI of the
communication links can be given to the source through relay’s cooperation,
and the relay retransmits the received information signal toward the destination. However, this collaboration is untrustworthy at the data level, i.e., the
relay may decipher the confidential information from their received signal.
4
γsi = ρkhsi k2 , γid = ρ|hid |2 , γil = ρ|hil |2 , γld =
hH
m
2
sm
ρ|hld |2 , γsd = ρkhsd k2 and γnull,l
= ρ| |hsm
| hsl | ,
m ∈ {i, d}.
• γ̄si = ρµsi , γ̄id = ρµid , γ̄sl = ρµsl , γ̄ld = ρµld and
γ̄sd = ρµsd .
In the following, we describe two transmission policies
adopted in this paper to establish the PLS in our considered
network; namely DBCJ policy and DT policy.
•
A. DBCJ Policy
In some communication networks, the direct channel gain
between the source and destination may be so weak when
the source and destination are located far apart or within
heavily shadowed areas. In this condition, cooperative relaying
offers a promising solution. This scenario has been extensively
implemented in most previous works [10], [12]–[14], [18],
where the source and relays pertain to a group, while the
destination and passive Eves are placed in another group.
Some networks including this scenario are the mobile ad-hoc
networks (MANETs) [9], the LTE cellular systems [9], [14]
and the IoT network [5], [25]. The aim of DBCJ policy is to
deteriorate the received signal at the malicious nodes and to
not allow them to decipher the information signal.
Now, we proceed to highlight the necessity of adopting jamming signal in our relay-aided network. Toward this end and
for the first step, we investigate the conventional AF relaying
scheme in which CJ is not utilized for data transmission. Since
the selected relay i itself is a curious node, the instantaneous
secrecy rate is given by [21]
h1
1
i+
γsi γid
Rsconv =
− log2 (1 + γsi ) .
log2 1 +
2
1 + γsi + γid
2
(1)
xy
According to the fact that 1+x+y
< min(x, y) ≤ x [21],
one can easily conclude that Rsconv = 0. This result indicates
that the achievable secrecy rate of the conventional untrusted
relaying scheme without employing CJ is zero. Therefore, a
jamming signal must be sent to degrade the received signal
at the untrusted relay. In this paper, due to the simplicity of
DBCJ policy compared with source-based jamming and FJbased jamming, we adopt the DBCJ policy to provide perfect
secure transmission. It is worth noting that the analysis in
this paper can be simply extended to the case of utilizing an
external FJ [15].
Under DBCJ policy, since the nodes operate in a half-duplex
mode, the direct link between the source and destination is
missed. The whole transmission is performed based on a
time division multiple-access (TDMA) protocol including the
broadcast phase (first phase) and the relaying phase (second
phase). During the first phase, while the source transmits the
information signal with power λP , the destination concurrently radiates the artificial noise with power (1 − λ)P , where
λ ∈ (0, 1) is the power allocation factor. It is worth noting
that this power allocation scheme provides insights for the
power allocation of the source and destination. This power
allocation approach has been exploited in several works for
both performance analysis and network optimization design
[10], [19]–[21], [23]. During the second phase, the selected
relay normalizes its received signal and forwards it with power
P . Finally, the destination decodes the source information
by subtracting the self-interference signal. Note that due to
the broadcast nature of wireless communication, the signal
transmitted in both the phases could be eavesdropped. In this
paper, contrary to most recent literature in multiple untrusted
relays [21]–[24] that ignored the information leakage in the
second phase of transmission, we consider a more realistic
scenario. In our considered scenario, the malicious nodes
intercept the transmissions of both the source and the selected
relay and then try to capture the information.
Denoting xs and xd as the information signal and the jamming signal, respectively, the received signal at the malicious
node l (l ∈ {1, ..., K + L}) in first phase can be expressed as
√
p
(1)
yl = λP wH hsl xs + (1 − λ)P hld xd + nl ,
(2)
hsi
represents the MRT beamformer at the
where w = kh
si k
source and i (i ∈ {1, ..., K}) indicates the index of the selected
relay Ri . In the second phase, the relay amplifies its received
signal by an amplification factor of
s
P
G=
,
(3)
2
λP khsi k + (1 − λ)P |hid |2 + N0
and broadcasts the message xi = Gyi . During the second
phase, all the malicious nodes hear the signal diffused by Ri .
Thus, the received signal at the malicious node l (l 6= i) can
be expressed as
(2)
yl
(1)
= Ghil yi
+ nl .
(4)
Furthermore, at the destination, after performing selfinterference cancellation, the corresponding signal is given by
√
(5)
yD = λP Gkhsi khid xs + Ghid nj + nd .
Based on (2), the received signal-to-interference-and-noiseratio (SINR) at the selected relay i and the SINR at the
malicious node l during the first phase of transmission are
respectively, given by
γi =
λγsi
(1 − λ)γid + 1
and
(1)
γl
=
i
λγnull,l
.
(1 − λ)γld + 1
(6)
By substituting (2) into (4), the received SINR at the malicious
node l during the second phase can be expressed as
(2)
γl
=
λγsi γil
.
λγsi + (1 + (1 − λ)γid )(1 + γil )
(7)
Moreover, by invoking (5), the SINR at the destination can be
obtained as
λγsi γid
γDi =
.
(8)
λγsi + (2 − λ)γid + 1
The instantaneous secrecy rate corresponding to the selected
relay i can be expressed as [6]
i+
1h
log2 (1 + γDi ) − log2 (1 + γE ) ,
(9)
Rs(i) =
2
where γE is the amount of information leaked to the malicious
5
nodes. In case of NCE, the information leakage is given by
o
n
(1)
(2)
N CE
.
(10)
=
max
γ
,
γ
,
γ
γE
i
l
l
i
Let i⋆ denote the index of the best selected relay. A relay
that maximizes the instantaneous secrecy rate is selected as
i⋆ = arg max Rs(i) .
1≤l≤K+L,l6=i
1≤i≤K
Moreover, the information leakage for CE case is given by
[28], [29]
(14)
The optimal relay selection criterion (14) requires the CSI of
both hops and all inter-relay channels. Therefore, it is rather
L
n
o
X
(1)
(2)
(1)
(2)
CE
max {γi , γl , γl },
γEi = max
(γl + γl ) , (11) complicated to be implemented in practice, especially when
1≤l≤K,l6=i
the number of source antennas and the number of relays are
|
{z
} |l=1
{z
}
large. To alleviate this issue, motivated by LSMA at the source,
Information leakage 1
Information leakage 2
we will propose a simple relay selection criterion for both NCE
where information leakages 1 and 2 describe the information and CE cases.
leakage for untrusted relays and passive Eves, respectively.
Remark 1: In practical cooperative networks, e.g., an ultra- A. Non-colluding Eavesdroppers
dense heterogeneous wireless network, both trusted and unFor NCE case, based on (10), we need to find the maximum
trusted relays exist. In such a network, to combat the po- received SINR at the malicious nodes. Due to deploying an
tential security attacks from untrusted relays, reliable relay LSMA at the source and by leveraging the Cauchy-Schwarz
authentication techniques are implemented, i.e., the network inequality, γ (1) in (6) can be upper bounded by
l
needs to discriminate between the trusted and untrusted relays
i
λN CE γnull,l
and then adopts the suitable secure transmission policy. The
(1)
< γi ,
(15)
γl =
authentication can be accomplished by traditional key-based
(1 − λN CE )γld + 1
cryptographic methods or new physical layer authentication where λ
N CE represents the power allocation factor related to
(2)
(PLA) methods [32], [33]. A novel technique to implement
NCE case. Furthermore, γl in (7) can be upper bounded by
the PLA is the multi-attribute multi-observation (MAMO)
λN CE γsi γil
(2)
technique [33] in which the joint verification of the received
γl <
(1 + (1 − λN CE )γid )(1 + γil )
signal strength indicator (RSSI) and the hardware imperfecλN CE γsi
tions of the transceivers are exploited to authenticate the
<
= γi .
(16)
relay nodes. PLA can also be implemented based on this
1 + (1 − λN CE )γid + 1
feature that the device-dependent hardware impairment inAccording to (6), (15) and (16), the information leakage in
phase/quadrature (I/Q) imbalance related to the reception and
N CE
(10) is simplified as γE
= γi . As such, the instantaneous
transmission of the relays is unique [32].
secrecy rate in (9) is simplified as
1 + γ i+
1h
Di
B. DT Policy
log2
.
(17)
Rs(i) =
2
1
+
γ
i
In DT policy, when the source transmits the information
△ 1+γDi
signal to the destination, all the relays and the Eves listen. Let φ(λ
N CE ) =
1+γi . Notably φ(0) ≥ 1. Therefore,
Accordingly, the received SNR at the malicious node l and
⋆
for OPA φ(λN CE ) ≥ 1, where λ⋆N CE is the OPA factor
the SNR at the destination are respectively, represented by
corresponding to NCE case. Therefore, the operator [·]+ in
(17) can be dropped. For NCE case, we have the following
(a) ρĥdl
(a)
d
γl = γnull,l
≈
and γD = ρkhsd k2 ≈ Ns γ sd , (12) key result.
Ns µsd
∆
2
where ĥdl = |hH
sd hsl | . In (12), (a) follows from the law of
large numbers [34] due to large Ns . Applying the LindebergLevy central limit theorem [34], the r.v. ĥdl can be approximated as CN (0, Ns µsd µsl ). It is worth pointing out that ĥdl is
very well approximated as a Gaussian r.v. even for small Ns
d
[35]. Therefore, γnull,l
can be considered as an exponential
d
r.v. with mean γ null,l = ρµsl .
For DT policy, the information leakage for NCE and CE
cases can be respectively, expressed as [28], [29]
N CE
γE
=
max {γl }
1≤l≤K+L
and
CE
γE
= max
1≤l≤K
n
γl ,
L
X
j=1
o
γj .
(13)
III. J OINT R ELAY S ELECTION AND P OWER A LLOCATION
In this section, we first introduce the optimal relay selection
scheme. Then motivated by the LSMA [1], [2], our proposed
JRP scheme is presented for both NCE and CE cases.
Proposition 1: For a large number of antennas at the source,
1+γ
the function φ(λN CE ) = 1+γDii is a quasiconcave function of
λN CE in the feasible set and the optimal solution is given by
√
2γid
.
(18)
λ⋆N CE =
γsi
Proof: By substituting the expressions (6) and (8) into
(17), forming the function φ(λN CE ), and then taking the first
derivative of it with respect to λN CE , we obtain
φ′ (λN CE ) =
γsi C1 λN CE 2 + C2 λN CE + C3
−
2
2 ,
si
si
− λN CE + 1
λN CE γγid
− λN CE + 2
γid λN CE γγid
(19)
si
si
where C1 = ( γγid
− 1) (γid + 1) γγid
+ 2 γid − 1 , C2 =
si
4 γid + 4 γγid
− 4 and C3 = −2 γid + 4. For a large number
6
of antennas at the source, γsi ≫ γid , the coefficients are simγ2
γsi
plified as C1 = (γid + 1) γsi
and C3 = −2γid + 4.
2 , C2 = 4 γ
id
id
′
By solving φ (λN CE ) = 0, the single feasible solution is
γ 3 (1+γ )
obtained as (18). Since φ′ (1) = − si 4γ 3 id < 0, we conclude
id
that φ(λN CE ) is a quasiconcave function in the feasible set.
Thanks to the LSMA at the source, by applying the law
of large numbers [34], we have γsi ≈ Ns γ si . Therefore, the
optimal solution (18) that requires the CSI of both hops can
be further simplified as
√
2|hid |2
⋆
λN CE ≈
.
(20)
Ns µsi
The proposed OPA factor in (20) requires only the CSI of the
selected relay-destination link and the statistical mean of the
source-selected relay link.
By substituting (20) into (6) and (8), we obtain
√
γid
N CE
N CE
√ . (21)
≈
and
γD
= γi ≈ 2
γE
i
i
1+ 2
Therefore, by using (21), the instantaneous secrecy rate in (9)
can be rewritten as
1 + γ CE
1
Di
√
.
(22)
Rs(i) = log2
2
1+ 2
The interesting result in (22) indicates that the instantaneous
secrecy rate only depends on the transmit SNR and the relaydestination link. As a consequence, the high complexity relay
criterion (14) can be approximated as
i⋆
CE
= arg max γD
i
1≤i≤K
= arg max |hid |2 .
1≤i≤K
(23)
The proposed relay selection criterion in (23) requires only the
relays-destination channel gains and hence, enjoys from low
complexity and energy consumption. Furthermore, the relay
selection scheme in (23) can be easily implemented using the
distributed timer technique in [36].
In practice, the proposed JRP scheme can be implemented
as follows: Before data transmission, the relays are scheduled
to transmit pilot symbols [26]. Using the pilots, the source
and the destination can estimate their channels. Then
according to (23), the destination computes the strongest
link between itself and the relays and hence, the relay index
i⋆ is selected. Afterward, the destination broadcasts i⋆ and
pilot symbols to estimate the destination-relay i⋆ link. Then
the relay i⋆ forwards a quantized version of the estimated
destination-relay i⋆ to the source (l bits). Finally, both the
source and destination tune their optimal transmit power
to start communication. Accordingly, the overall overhead
required for broadcasting is equal to l + ⌈log2 K⌉ bits.
source to transmit the information signal. Based on this,
λ⋆CE < λ⋆N CE ≪ 1, where λ⋆CE denotes the OPA factor for
(2)
CE case. Therefore, γl in (7) can be approximated as
(2)
γl ≈
(a)
λCE γsi γil
λCE γsi
= γi ,
≈
(1 + (1 − λCE )γid )(1 + γil ) (1 − λCE )γid + 1
(24)
where (a) follows from the high SNR assumption. The interesting result in (24) expresses that the amount of information
leakage to the malicious nodes in the second phase of transmission is approximately the same as the amount of information
leaked to the selected relay in the first phase. Based on this
new result and given this fact that, the received signal by
the malicious nodes in the second phase of transmission is
a degraded version of the emitted signal by the selected relay,
CE
γE
in (11) is changed to
i
CE
= γi +
γE
i
L
X
(1)
γl .
(25)
l=1
For the sake of tractability, according to the fact that λ∗CE ≪ 1,
(1)
γi and γl in (6) can be rewritten as
γi ≈
i
λCE γnull,l
λCE γsi
(1)
, γl ≈
. (26)
(1 − λCE )(γid + 1)
(1 − λCE )(γld + 1)
By substituting (26) into (25), we obtain
CE
=
γE
i
λCE ∆
,
1 − λCE
(27)
i
PL γnull,l
γsi
where ∆ = γid
l=1 γld +1 . Following the same steps
+1 +
as in Proposition 1, at the high SNR regime, the OPA factor
corresponding to CE case can be obtained as
r
2γid
λ∗CE =
.
(28)
γsi ∆
We can conclude from (28) that by increasing the number of
Eves, the most amount of the total power is assigned to the
destination to inject jamming signal.
Remark 2: In practice, often may not be feasible to achieve
the CSI of passive Eves. Based on this practical issue, we consider a scenario, where only the second order statistics related
to the Eves’ are available which is a common assumption in the
literature [10], [13], [14], [18]. Furthermore, for mathematical
simplicity, we assume that the relaying and eavesdropping
channels are independent and identically distributed, i.e., for
relaying channels with 1 ≤ 1 ≤ K, we consider µsl = µsr and
µld = µrd , and for eavesdropping channels with 1 ≤ l ≤ L,
we consider µsl = µse and µld = µed [9], [13], [14].
B. Colluding Eavesdroppers
For CE case, the Eves jointly try to decode the information
signal based on MRC processing and hence, the amount
of overheard information increases. To tackle this problem,
more power should be dedicated to the destination to confuse
the Eves compared to λ⋆N CE in (18) and less power to the
Here, we consider the case with large number of passive
Eves. As will be observed in numerical examples, the analysis
are valid even for moderate number of Eves. Based on the law
of large numbers and since the first and the second hops are
7
case E{ γsi1 ∗ }, we used Lemma 2.9 in [38]
n 1 o
1
=
.
E
∗
γsi
(Ns − 1)ρµsr
independent, we have
L
i
n
o n 1 o
X
γnull,l
i
≈ L E γnull,l
E
γld + 1
γld + 1
l=1
Lµse e ρµed
1 ∆
=−
Ei −
= θ̂,
µed
ρµed
1
(29)
where we used [31, Eq. (3.352.4)]. By substituting (29) into
λ⋆CE in (28) and then substituting into (27) and (8), we obtain
r
γid
γ
CE
CE
q id
2 + 2 θ̂ and γD
.
=
≈
γE
i
i
γsi
id
θ̂
1 + 2 + 2 γγsi
(30)
CE
CE
By substituting γE
and γD
in (30) into (9), the instantai
i
neous secrecy rate is obtained. Accordingly, we find that the
optimal relay selection scheme corresponding to the maximum
achievable secrecy rate needs the excessive implementational
overhead. The optimal relay selection based on both hops is
out of the scope of this paper. As such, we adopt the suboptimal simple relay selection scheme proposed for NCE case in
(23). We mention that since the second hop has a dominant
impact on quantifying the received SNR at the destination,
the relay selection criterion in (23) can be considered as a
suboptimal criterion to enhance the secrecy rate. Moreover,
in case of very LSMA at the source, the suboptimal relay
selection criterion in (23) will be the near-optimal one.
For the selected relay i∗ with the relay selection criterion in
(23) and with very LSMA at the source, we can use the wellknown approximation of minimum mean square error (MMSE)
estimator as follows [1], [37]
o n 1 o
n γ ∗ o (a) n
γi∗ d
i d
≈E
= E γi∗ d E
γsi∗
γsi∗
γsi∗
(b) µrd (Ψ(K + 1) + ̺) △
= η,
(31)
=
(Ns − 1)µsr
where ̺ = 0.577 [31] is Euler’s constant. In (31), (a) follows
from the fact that the two-hops are independent and (b) follows
from evaluating E{γi∗ d } and E{ γsi1 ∗ }. For the first case,
E{γi∗ d }, we have
o Z ∞
n
1 − Fγi∗ d (x) dx
E γi∗ d =
0
(32)
= ρµrd Ψ(K + 1) + ̺ ,
where the last equality follows from using the cdf of γi∗ d
which can be obtained based on order statistics [34]
Fγi⋆ d (γ) =
K h
Y
γ i
1 − exp (−
) .
γ id
i=1
=1+
K
2X
−1
(33)
X γ
(−1)|un | exp −
,
γ id
n=1
i∈u
(34)
n
where |un | denotes the cardinality of the n-th non-empty
subcollection of the K relays, and the last expression follows
from applying the binomial expansion theorem. For the second
Therefore, by substituting (31) into (30), we arrive at
q
γ∗
CE
CE
qi d
γE
≈
2 + 2η θ̂ and γD
=
.
i∗
i∗
1 + 2 + 2η θ̂
(35)
(36)
CE
We conclude from (36) that the constant value γE
increases
i∗
by increasing the number of Eves in the network. By equipping
the source with √very LSMA Ns → ∞, this information
leakage tends to 2 the same as NCE case.
Comparing (21) for NCE and (36) for CE, we can define
the new parameter C to integrate the performance analysis for
NCE and CE cases. Based on this definition, we have
p
γ∗
pi d
, (37)
γEi∗ ≈ 2(1 + C) and γDi∗ ≈
1 + 2(1 + C)
where C = 0 for NCE case and C = η θ̂ for CE case.
Remark 3: Both the network power optimization and
performance analysis presented in this paper, can be routinely
extended to the scenario that the untrusted relays and passive
Eves collaboratively decode the information signal based on
MRC technique.
Remark 4: According to the results in (15) and (24), the
received SNR at the malicious nodes in the second phase
of transmission is more than that one in the first phase of
transmission. Therefore, when SC is adopted at the malicious
nodes, and all of them cooperate to decode the information
signal, the network power optimization and performance
analysis are similar to the NCE case.
IV. P ERFORMANCE A NALYSIS
In this section, we derive new closed-form expressions for
the probability of positive secrecy rate, the ESR and the
SOP as the metrics of security and the SER as the reliability
measure. Both the proposed JRP scheme (which is based on
DBCJ policy) and the DT policy are studied.
A. Probability of Positive Secrecy Rate
In this section, we proceed to derive the probability of
positive secrecy rate for the proposed DBCJ and DT policies.
DBCJ Policy: By substituting (37) into (9) and using the
proposed relay selection criterion in (23), the probability of
positive secrecy rate can be expressed as
n ⋆
o
DBCJ
Ppos
= Pr Rs(i ) > 0
o
n
p
= Pr γi⋆ d > 2(1 + C) + 2(1 + C)
p
= 1 − Fγi⋆ d 2(1 + C) + 2(1 + C)
K h
2(1 + C) + p2(1 + C) i
Y
(a)
,
1− exp −
= 1−
γ id
i=1
(38)
8
where (a) follows from substituting (33). It can be concluded
from (38) that, the proposed JRP is not efficient when the
average transmit SNR of the second hop γ id (which is a
function of the transmit SNR ρ and the distance-dependent
channel gain µid ) is low. This observation is not surprising,
because when the DBCJ policy is adopted, due to the halfduplex operation of the nodes, the direct link between the
source and destination is vanished. As such, the destination
only relies on the second hop to receive the information signal
form the source. Consequently, the reliability of confidential
communication is degraded when the relays are far from the
destination or the transmit SNR is low. In the high SNR
regime, (38) is simplified as
K
K Y
p
1
DBCJ
.
Ppos
= 1 − 2(1 + C) + 2(1 + C)
γ
i=1 id
(39)
We deduce from (39) that the probability of positive secrecy
rate approaches one as the transmit SNR or the number
of relays increases. The reason is that as K grows, the
occurrence probability of a stronger channel between relays
DBCJ
and the destination increases and thus, Ppos
approaches one.
DT Policy: For this policy, we will study the NCE case and
CE case separately, below.
1) NCE Case: In this case, by combining (9), (10) and (12),
the probability of positive secrecy rate can be obtained as
n
DT,NCE
= Pr log2 1 + Ns γ sd
Ppos
o
d
>0
− log2 1 + max γnull,l
1≤l≤K+L
≈
K+L
Y
l=1
1−e
−
Ns µsd
µsl
l
,
(40)
d
where the last expression follows from the fact that γnull,l
can be approximated as an exponential r.v., as mentioned in
DT,NCE
Section II-B. This result reveals that Ppos
→ 0 as the
number of malicious nodes goes to infinity. Furthermore, one
DT,NCE
can obtain Ppos
→ 1 as Ns → ∞, which is an intuitive
observation, since by increasing Ns , the source with an LSMA
focuses the transmission energy toward the direction of the
selected relay. Hence, the strength of the received signals at
the malicious nodes will be low enough. Based on this, the
malicious nodes fail to extract the information. Furthermore, in
contrast to DBCJ policy that the confidential communication is
solely dependent on the second hop channel, expression (40)
illustrates that since the probability of positive secrecy rate
for DT policy is independent of the transmit SNR, the DT
can offer secure transmission even for low transmit SNRs by
deploying large number of source antennas Ns .
CE
2) CE Case: As mentioned in (13), in this case γE
=
PL
max1≤l≤K {γl , j=1 γj }. To obtain the probability of positive
CE
secrecy rate, we first find the cdf of γE
. Using order statistics
∆
R
[34], the cdf of γE = max1≤l≤K {γl } is given by [9], [34]
FγER (γ) = 1 +
K
2X
−1
X γ
.
(−1)|un | exp −
γ sl
n=1
l∈un
(41)
The following lemma from [39] helps to obtain the cdf of
PL
Eve ∆
γE
= j=1 γj in (13).
Lemma 1: Let {Xi }ni=1 , n > 1, be independent exponential
r.vs with distinct averages µi . Then the cdf of their sum W =
X1 + X2 + ... + Xn is given by
n
X
− x
Γ(µj ) 1 − e µj , x > 0.
(42)
FW (x) =
j=1
∆
where Γ(µj ) =
Q
n
1
i=1 µi
µj
1
k6=j,k=1 ( µk
Qn
− µ1 )
.
j
CE
Using (41) and leveraging Lemma 1, the cdf of γE
in (13)
can be expressed as
FγECE (γ) = FγER (γ)FγEEve (γ)
K
−1
L
2X
X γ X
− x
= 1+
(−1)|un | exp (−
Γ(γ̄sj )(1 − e γ̄sj ) .
)
γ sl
n=1
j=1
l∈un
(43)
Hence, the probability of positive secrecy rate for DT policy
under the presence of CE can be expressed as
DT,CE
Ppos
= FγECE (Ns γ sd ),
(44)
CE
where the cdf of γE
is in (43). From (44), we observe that
when Ns → ∞, the secure communication is established,
while for large number of malicious nodes, the secure transmission is harmed.
B. Ergodic Secrecy Rate
The ESR as a useful secrecy metric characterizes the
average of the achievable instantaneous rate difference
between the legitimate channel and the wiretap channel. In
the following, we derive new accurate closed-form expressions
for the ESR of both DBCJ based and DT policies.
DBCJ Policy: By substituting (37) into (9) and using the
relay selection criterion in (23), the ESR can be written as
Z ∞
1
γ
DBCJ
p
fγi⋆ d (γ)dγ
Rs
=
ln 1 +
2 ln 2 0
1 + 2(1 + C)
p
1
− log2 1 + 2(1 + C)
2
i
h
Z ∞
(γ)
1
−
F
γ
⋆
i
d
1
(a)
p
=
dγ
2 ln 2 0 γ + 1 + 2(1 + C)
p
1
(45)
− log2 1 + 2(1 + C) ,
2
where (a) follows from using integration by parts. To obtain
a closed-form solution for the ESR, by substituting (34) into
(45) and using [31, Eq. (3.352.4)], we get
√
2 −1
P
1+ 2(1+C)
1 X
|un |
i∈un
γ id
(−1) e
=
2 ln 2 n=1
X 1 + p2(1 + C) 1
p
− log2 (1 + 2(1 + C)).
× Ei −
γ id
2
i∈u
K
RsDBCJ
n
(46)
9
We remark that the ESR in (46) is explicitly characterized by
the average channel gains between the relays and destination,
and the transmit SNR of the network. As observed, increasing
the number of antennas at the source Ns has no impact on
the ESR performance. For a single-relay network without any
passive Eve, the ESR in (46) is simplified as
√
1 + √2 1
√
1+ 2
1
DBCJ
γ 1d
e
− log2 (1 + 2),
Rs
=−
Ei −
2 ln 2
γ 1d
2
(47)
which was derived in [19], [20]. As such, the works [19], [20]
can be considered as a special case of our work.
We can further calculate the ESR performance in the high
SNR regime when ρ → ∞ by applying the general asymptotic
form given by [30]
(48)
Rs∞ = S∞ log2 ρ − L∞ ,
impacts of µid , the number of relays K and the number of
passive Eves L on the ESR and shows that the power offset
is independent of Ns . As expected, we find that increasing
the number of passive Eves increases the ESR power offset
which corresponds to a decrease in the ESR.
DT Policy: In the following, we investigate the ESR performance of DT policy for both NCE and CE cases.
where S∞ is the high SNR slope in bits/s/Hz/ (3 dB) and L∞
is the high SNR power offset in 3 dB units. These parameters
are two key performance factors that explicitly examine the
N CE
1) NCE Case: In this case, by substituting γE
in (13)
ESR performance at the high SNR regime which are defined
and γD in (12) into (9), we have
respectively, as [30]
nh
RsDT,NCE = E log2 1 + Ns γ sd
Rs∞
Rs∞
S∞ = lim
. (49)
and L∞ = lim log2 ρ −
i+ o
ρ→∞ log2 ρ
ρ→∞
S∞
d
− log2 1 + max γnull,l
(54)
1≤l≤K+L
We mention that the high SNR slope is also recognized as
h
the maximum multiplexing gain or the number of degrees of
≥ log2 1 + Ns γ sd
freedom [30]. Based on (37), in the high SNR regime with
oi+ △
n
DT,NCE
d
,
= Rs,LB
− E log2 1 + max γnull,l
ρ → ∞, we have ln(1 + γDi ) ≈ ln(γDi ). As such, the ESR
1≤l≤K+L
of the proposed JRP scheme can be represented by
(55)
n
o
p
1
n
o
E ln(γi⋆ d ) − log2 1 + 2(1 + C) . (50)
Rs∞ =
2 ln 2
where the inequality follows from the fact E max{x, y} ≥
n
o
By taking the derivative of (34) with respect to γ to obtain
max E{x}, E{y} [21]. Following the same steps presented
the pdf of γi∗ d , and then applying [31, Eq. (4.331.1)], the
to derive the ESR of DBCJ policy, (55) is given by
asymptotic ESR can be obtained as
h
DT,NCE
K
R
=
log
1
+
N
γ
2
−1
s
h
i
2
sd
s,LB
X 1
1 X
Rs∞ =
(−1)|un | ln(
)+̺
K+L
2
−1
P
X 1 i+
2 ln 2 n=1
γ
1
1 X
i∈un id
,
(−1)|vn | e i∈vn γ sj Ei −
−
p
ln 2 n=1
γ sj
i∈v
(51)
− log2 1 + 2(1 + C) .
n
(56)
Using (49), the high SNR slope is given by
where |vn | denotes the cardinality of the n-th non-empty
1
S∞ = ,
(52) subcollection of the K + L nodes. We remark that at the
2
high SNR regime, the instantaneous secrecy rate (54) is further
P2K −1
where we used the fact that n=1 (−1)|un | = −1. Expres- simplified as
i+
h
sion (52) highlights that the number of source antennas Ns and
DT,N CE
. (57)
max ĥld
= 2 log2 Ns µsd − log2
the presence of collaborative eavesdropping have no impact on Rs
1≤l≤K+L
the ESR slope.
Expression (57) states that the instantaneous secrecy rate
Furthermore, the high SNR power offset is derived as
of DT is independent of the transmit SNR, i.e., a secrecy
K
2X
−1
rate ceiling appears when the transmit SNR increases. In
X 1
̺
)+
L∞ = −
(−1)|un | log2 (
other words, the high SNR slope of DT policy is zero. We
µ
ln 2
n=1
i∈un id
conclude that, unlike the DBCJ policy, the DT cannot achieve
p
+ 2 log2 1 + 2(1 + C) ,
(53) high secure transmission rates. Furthermore, as observed
from (57), the ESR decreases by increasing the number of
where a decrease in the power offset corresponds to an malicious nodes while the ESR increases as Ns grows.
increase in the ESR. Expression (53) characterizes the
10
CE
2) CE Case: In this case, by using γE
in (13), a tight
lower bound for the ESR of CE case can be obtained as
h
n
oi+
DT,CE
CE
Rs,LB
= log2 1 + Ns γ sd − E log2 1 + γE
L
X
1
1
Γ(γ̄sj )e γ̄sj Ei −
= log2 1 + Ns γ sd +
γ̄sj
j=1
h
−
+
K
L
2X
−1 X
(−1)|un | Γ(γ̄sj )e
P
1
l∈un γ sl
Ei(−
n=1 j=1
K
L
2X
−1X
1
(−1)|un| Γ(γ̄sj )e γ̄sj
+
1
)
γ̄sl
P
1
l∈un γ sl
n=1 j=1
X 1 i+
1
× Ei −
,
−
γ̄sj
γ sl
(58)
l∈un
CE
where we used the cdf of γE
in (43) and [31, Eq. (3.352.4)].
Similar to NCE case, a secrecy rate ceiling appears in the
high SNR regime which limits the secrecy performance of
DT policy.
C. Secrecy Outage Probability
The overall SOP denoted by Pout is defined as the
probability that a system with the instantaneous secrecy rate
Rs is unable
o the target transmission rate Rt , i.e.,
n to support
Pout = Pr Rs < Rt .
DBCJ Policy: By Substituting (37) into the SOP definition,
we obtain
o
n
DBCJ
et
Pout
= Pr γi⋆ d < R
K h
R
Y
bt i
et =
,
(59)
1 − exp −
= Fγi⋆ d R
γ id
i=1
p
p
et = 1 + 2(1 + C) 22Rt (1 + 2(1 + C)) − 1 .
where R
To derive (59), we used the cdf of γi∗ d in (34). Expression (59)
indicates that the SOP approaches zero as the transmit SNR
ρ → ∞. For a single-relay case and without passive Eve, the
overall SOP is simplified as
(1 + √2)(22Rt (1 + √2) − 1)
DBCJ
Pout = 1 − exp −
, (60)
γ 1d
which was derived in our previous work [20]. Therefore, this
new work extends the recent work [20].
Now, we look into the high SNR regime and investigate
the diversity order. In the high SNR regime, the closed-form
expression in (59) can be written as
DBCJ
et )K
≈ (R
Pout
K
Y
1
.
γ
i=1 id
(61)
By inspecting (61), we interestingly find that the diversity
order of the system equals to the number of untrusted relays
K. To justify this new result intuitively, we can say that in the
considered LSMA-based network, under applying the OPA,
the first hop channel becomes deterministic and only the
second hop contributes for signal transmission. Increasing the
number of untrusted relays, actually increases the probability
of having a stronger link in the second hop. Therefore, the
secrecy diversity order of the system increases which in turn
decreases the secrecy outage probability.
DT Policy: To study the SOP of DT policy, we have two
cases of NCE and CE.
N CE
1) NCE Case: By substituting γE
in (13) and γD in
(12) into (9), and then formulating the SOP, we obtain
o
n
1 + Ns γ sd
DT,NCE
<
R
Pout
= Pr log2
t
d
1 + max1≤l≤K+L γnull,l
n
o
1 + Ns γ sd
d
= 1− Pr
max γnull,l
<
−
1
1≤l≤K+L
2 Rt
K+L
i
Y h
1 1 + Ns γ sd
= 1−
1 − exp −
−
1)
.
(
γ sl
2 Rt
l=1
(62)
DT,NCE
Pout
The expression (62) indicates that
→ 1 as the
number of malicious nodes grows. One can easily conclude
from (62) that an error floor occurs at the high SNR regime.
This means that the secrecy diversity order of MRT-based DT
policy is zero. This result is the same as the results in [30],
where the authors considered transmit antenna selection at
the transmitter and receive generalized selection combining
DT,NCE
at the receiver. Moreover, we can conclude Pout
→ 0 as
Ns → ∞.
2) CE Case: In this case, we obtain
o
n
1 + N γ
s sd
DT,CE
<
R
Pout
= Pr log2
t
CE
1 + γE
1 + N γ
s sd
= 1 − FγECE
−
1
,
(63)
R
2 t
where FγECE (γ) is in (43). Expression (63) states that similar
to NCE case, the SOP of CE case tends to a constant in the
high SNR regime. Furthermore, we find that the SOP of CE
case approaches zero as Ns → ∞.
D. Average Symbol Error Rate
In this subsection, we evaluate the SER of the proposed
JRP scheme. The SER of the DT scheme is available in the
literature [39, Sec. 3] and is omitted for brevity.
The instantaneous SER
√ of coherent modulation is in the
form of Ps = αm Q( βm γD ), where γD represents the
received SNR, and αm and βm depend on the modulation
type. Specifically, for rectangular M -ary quadrature amplitude
3
modulation, αm = 4(1 − √1M ) and βm = M−1
. Moreover,
for M -ary phase-shift keying (M ≥ 4), αm = 2 and
π
βm = 2 sin2 M
[39].
Using the average SER of coherent modulation as [39]
Z ∞
t2
t2
αm
exp(− )dt,
(64)
Ps = √
FγD
βm
2
2π 0
11
the average SER of the proposed JRP system using (33) can
be expressed as
Z ∞
αm
t2
DBCJ
√
Ps
=
exp(− )
2
2π 0
p
K
h
Y
(1 + 2(1 + C))t2 i
dt
1 − exp −
×
βm γ id
i=1
K
2X
−1
αm
(−1)|un |
r
=
,
1+
√
2
P
1+ 2(1+C)
n=1
1 + 2 i∈un
βm γ
id
(65)
where the last expression follows
applying
q the binoR ∞ from
2
mial expansion and the fact 0 e−qx dx = 12 πq [31, Eq.
(3.318.3)]. For a single-relay system, the average SER is
s
βm γ 1d
αm
√
.
(66)
Ps =
1−
2
βm γ 1d + 2 + 2
In the high
√ SNR regime,
√ i.e., γ̄id → ∞, we have 1 −
(1+ 2C)t2
(1+ 2C)t2
exp(− βm γ
) ≈ βm γ
. Accordingly, the expression
id
id
(65) can be approximated as
K
p
αm 1 + 2(1 + C) Z ∞
t2
PsDBCJ ≈
t2k e− 2 dt. (67)
√
QK
K
2πβm i=1 γ id
0
q
R ∞ 2k −px2
π
Using the facts that 0 x e
dx = (2k−1)!!
p [31, Eq.
2(2p)k
(3.416.3)] with (2k−1)!! = (2k)!
2k k! , the high SNR approximation
of the overall average SER can be obtained as
p
αm (2K)!(1 + 2(1 + C))K
DBCJ
Ps
≈
.
(68)
QK
K 2K+1 K!
βm
i=1 γ id
By inspecting (68), we observe that the DBCJ policy achieves
a diversity order of K.
V. D IVERSITY-M ULTIPLEXING T RADEOFF
As is known, multi-antenna systems offer two different
types of benefits in a fading channel: diversity order and
multiplexing gain [39]. The DMT describes this trade-off by
presenting a trade-off between the error probability and the
data rate of a system [39]. Recently, the secrecy DMT of a
MIMO wiretap channel has been addressed in [40], where a
zero-forcing (ZF) transmit scheme is utilized. In the following,
we evaluate the DMT of the proposed JRP scheme from the
perspectives of security and reliability.
A. Secrecy Perspective
Let d and r be the diversity order and the multiplexing
△
2 Pout
gain of the presented system, defined as d = − lim loglog
ρ
ρ→∞
△
Rt
,
ρ→∞ log2 ρ
and r = lim
2
respectively. Substituting (61) into the
diversity order definition and using Rt = r log2 ρ, the DMT
can be obtained as
p
PK
2Rt
(1
+
γ
−K
log
2
2(1
+
C))−1
log
2
2 id
i=1
d = lim
ρ→∞
log2 ρ
= K(1 − 2r).
(69)
It is observed that when the multiplexing gain is not utilized
(r → 0), the diversity order equals the maximum value d = K,
which is consistent with our previous results. On the other
hand, a multiplexing gain of r = 0.5 is achieved as d → 0.
This is because, in the presented cooperative network, it takes
two time slots to complete the transmission of one traffic flow,
and hence, the maximum multiplexing gain of such network
is just 0.5. This can be dealt with utilizing a two-way network.
B. Reliability Perspective
log2 Ps
.
ρ→∞ log2 ρ
In this case, d = − lim
By substituting (68) into
the diversity order definition and using Rt = r log2 ρ, the
DMT is given by
d = K(1 − r).
(70)
One can conclude from (70) that in the absence of a multiplexing gain (r → 0), the diversity order is K. On the other
hand, a multiplexing gain of r = 1 is achieved as d → 0.
VI. N UMERICAL R ESULTS AND D ISCUSSION
In this section, numerical examples are presented to verify
the accuracy of the derived performance metrics (ESR, overall
SOP and average SER) for the proposed JRP transmission
scheme. Numerical curves for the exact JRP scheme, marked
with filled circles, are obtained using the optimum relay
selection criterion in (14) together with the OPA, which is
numerically evaluated for a finite number of source antennas
using the bisection method. We compare our proposed JRP
scheme for both the NCE and CE cases with other well-known
transmission schemes as listed below:
1) Equal power allocation (EPA) between source and destination (λN CE = λCE = 0.5) with random relay
selection, which is denoted by “EPRR”,
2) OPA between source and destination with random relay
selection, which is denoted by “OPRR”,
3) EPA between source and destination with the optimum
relay selection in (14), which is denoted by “EPRS”,
4) The exact JRP scheme for the scenario of adopting the
MRC technique at the malicious nodes for NCE case,
and
5) The DT policy, wherein both the untrusted relays and
passive Eves are considered as pure Eves.
To verify the accuracy of the derived LSMA-based expressions for the ESR of JRP, the SOP of JRP, the SER of JRP, the
ESR of DT for NCE, the ESR of DT for CE, the SOP of DT for
NCE, the SOP of DT for CE and the SER of JRP in (46), (59),
(65), (56), (58), (62), (63) and (65) respectively, we conduct
Monte-Carlo simulations, where results are shown in Figs. 2
12
Ergodic secrecy Rate (bits/s/Hz)
4
3.5
5
JRP for K=1 using SC, Eq. (46)
JRP for K=1 using MRC
JRP for K=5 using SC, Eq. (46)
JRP for K=5 using MRC
EPRR, K={1,5}
DT for K=1, Eq. (56)
DT for K=5, Eq. (56)
3
Proposed JRP, K=1
4.5
4
Ergodic Secrecy Rate (bits/s/Hz)
4.5
EPRR,
K={1,5}
2.5
2
Proposed JRP, K=5
1.5
DT, K=1
1
JRP for K=1, Eq. (46)
JRP for K=5, Eq. (46)
EPA for K=1
EPRR for K=5
DT for K=1, Eq. (58)
DT for K=5, Eq. (58)
3.5
Proposed JRP, K=1
3
2.5
EPRR,
K={1,5}
Proposed JRP, K=5
2
1.5
1
DT, K=5
0.5
0
0
DT, K=5
0.5
5
10
15
20
25
Transmit SNR [dB]
30
35
0
0
5
10
15
20
25
Transmit SNR [dB]
DT, K=1
30
35
Fig. 2. ESR versus the transmit SNR for the NCE case and for different
transmission schemes. We consider K = {1, 5} relays, the multi-passive Eve
L = 5 scenario, and Ns = 16 source antennas. The asymptotic curves for
the JRP scheme are shown with dashed lines using Eq. (51), while the filled
circles depict the Monte-Carlo simulations.
Fig. 3. ESR versus the transmit SNR for the CE case and for different
transmission schemes. We consider K = {1, 5} relays, the multi-passive Eve
L = 5 scenario, and Ns = 16 source antennas. The asymptotic curves for
the JRP scheme are shown with dashed lines using Eq. (51), while the filled
circles depict the Monte-Carlo simulations.
- 7. Furthermore, we plot the derived asymptotic expressions
for the ESR, SOP and SER of the proposed JRP scheme given
by (51), (61) and (68), respectively. As will be observed from
Figs. 2 - 7, our LSAM-based closed-form expressions well
matched with the simulation results, and the asymptotic curves
well approximate the exact curves in the high SNR regime. For
simplicity and without loss of generality, we assume that the
source, destination and relay(s) are located at the positions
(-1,0), (0,0), and (1,0) respectively, and the passive Eves are
placed near the relays to overhear the maximum information.
Unless otherwise stated, the values of network parameters
are: number of antennas at the source Ns = 16, number of
untrusted relays K = {1, 5}, number of passive Eves L = 5,
the target rate Rt = 1 bits/s/Hz and the distance-dependent
path loss factor α = 3.
Figure 2 shows the achievable ESR versus the transmit SNR
ρ for NCE case. Our observations are summarized as follows:
1) The ESR performance of the proposed JRP scheme is a
monotonically increasing function of the transmit SNR ρ
while the ESR of the DT scheme converges to a constant
value as presented in Section IV-B. The reason behind
this behavior of the JRP scheme is that the injected
jamming signal by the destination only degrades the
received information signal at the malicious nodes and
has no effect on the overall two-hop signal reception at
the destination. As such, the secrecy performance of the
DBCJ-based JRP scheme is superior to the performance
of the DT policy in the high SNR regime, while the
opposite behavior is observed at the low SNR regime.
The reason is that by equipping the source with an
LSMA and adopting an MRT beamformer the received
SNR at the destination for DT becomes considerable,
while the information leakage is negligible as computed
in (12). We also mention that, as discussed in Section IVA, the secrecy performance of the JRP is not satisfactory
when the average transmit SNR of the second hop is
low. Therefore, the proposed JRP suffers from w secrecy
performance loss in the low SNR regime.
2) The ESR performance of the JRP scheme increases as
K grows, which can be concluded form (22). This is
because by increasing K, the probability of emerging
a stronger channel between the relays and destination
grows and accordingly, the secrecy rate increases. Contrary to the JRP scheme, the ESR of the DT policy decreases by increasing K. The reason is that according to
(12), the received SNR at the destination is deterministic,
while based on (13), the amount of information leakage
increases as K grows.
3) The secrecy rate advantage of JRP scheme compared
with EPRR is obvious. For example, the SNR gap
between the proposed JRP scheme and EPRR is about
8 dB for K = 1 case; this gap is about 13 dB for
K = 5 case to achieve the target transmission rate of 2
bits/s/Hz.
4) The secrecy performance of the proposed JRP scheme
with SC at the malicious nodes is close to the secrecy
performance of the exact JRP with MRC technique.
In Fig. 3, we plot the ESR performance versus the transmit
SNR ρ for the CE case. As can be seen, the proposed JRP
scheme which is based on the near-optimal relay selection criterion in (23) is in perfect agreement with the exact numerical
results across the entire SNR range of interest. As can be seen
13
JRP for NCE with L={5,50} & CE with L=5
JRP for CE with L=50
OPRR for NCE with L={5,50} & CE with L=5
OPRR for CE with L=50
EPRS for NCE with L={5, 50} & CE with L=5
EPRS for CE with L=50
DT for NCE with L=5
DT for NCE with L=50 & CE with L=5
DT for CE with L=50
Ergodic Secrecy Rate (bits/s/Hz)
5
4
Proposed JRP
3
2
OPRR Policy
EPRS Policy
1
0
0
10
10
10
40
60
Number of Relays (K)
80
100
DT, K=1
−2
Proposed JRP, K=1
10
10
20
−1
EPRR, K={1,5}
10
DT Policy
0
DT, K=5
Secrecy Outage Probability
6
−3
−4
−5
0
Proposed JRP, K=5
JRP for K=1, Eq. (59)
JRP for K=5, Eq. (59)
EPRR for K={1,5}
DT for K=1, Eq. (62)
DT for K=5, Eq. (62)
10
20
30
Transmit SNR [dB]
40
Fig. 4. ESR versus the number of untrusted relays for the transmit SNR of
ρ = 20 dB. We consider the multi-Eve scenario with L = {5, 50}, where the
number of source antennas Ns = 256. The filled circles are obtained with
the Monte-Carlo simulations.
Fig. 5. Overall SOP versus the transmit SNR for the NCE case with K =
{1, 5} relays and L = 5 passive Eves. The number of source antennas is set
to Ns = 16 and the target rate to Rt = 1 bits/s/Hz. The asymptotic curves
are shown with dashed lines using Eq. (61), while the filled circles depict the
Monte-Carlo simulations.
from Fig. 3, the ESR performance of the proposed JRP scheme
increases by increasing K, while the opposite is observed for
DT policy. Furthermore, we can conclude from Fig. 3 that
the ESR of the proposed scheme significantly outperforms the
EPRR policy.
Figure 4 examines the impact of the number of untrusted
relays K on the ESR performance of the proposed JRP scheme
for the NCE and CE cases. We set the number of source
antennas to Ns = 256 and the transmit SNR ρ = 20 dB.
We also consider two cases of small L = 5 and large L = 50
number of passive Eves. The following observations can be
made from Fig. 4:
can be extracted by Eves, leading to ESR loss relative
to NCE case.
3) When OPRR is adopted in the network, the ESR performance for both NCE and CE cases is a constant function
of K. The reason is that when one relay is randomly
selected out of K available relay nodes and OPA is
applied, it is statically equivalent to the single-relay
scenario with OPA. For large number of Eves L = 50,
due to the collaboration between Eves, the ESR of CE
case is somewhat fewer than the NCE case.
4) The ESR performance of DT policy for NCE is a
decreasing function of K. The reason is that by increasing K, the probability of emerging a stronger wiretap
channel increases and therefore, the ESR decreases.
However, for CE with large number of Eves, the ESR is
a constant function with respect to K. The reason is that
both the amount of information leakage and the received
SNR at the destination are independent of the number
of untrusted relays. We note that according P
to (13) and
CE
the law of large numbers, we have γE
= L
j=1 γj ≈
Lρµse which is independent of K.
5) The ESR performance of EPRS for both NCE and
CE cases is an increasing function of K. This result
highlights the effectiveness of relay selection in LSMAbased security networks even without applying OPA.
1) As predicted by the analytical expressions and discussions in Section IV-B, the ESR performance of the
proposed JRP scheme for NCE and CE cases is a
monotonically increasing function of the number of
untrusted relays K. This new result highlights that,
unlike the results in [21], [22], the proposed LSMAbased scheme increases the secrecy rate of the network
by utilizing more untrusted relays.
2) The proposed JRP scheme for the NCE case with any
number of Eves offers a superior ESR performance
relative to the CE case with large number of Eves
L = 50. This event can be justified simply based on the
new results presented in Section III-B. The reason is that
by equipping the source with an LSMA, the amount of
information leaked to the malicious nodes in the first of
transmission is negligible compared to that one in the
second phase. Therefore, for CE with small number of
Eves L = 5, the ESR performance is the same as NCE
case, while with large number of Eves, some information
In Figs. 5 and 6, we compare the overall SOP versus the
transmit SNR ρ for NCE and CE cases, respectively. The
results in Figs. 5 and 6 indicate that the SOP performance advantage of our proposed secure transmission scheme compared
to the EPRR policy. As can be readily observed from these
figures, unlike DT policy that the SOP performance converges
14
10
0
10
DT, K=5
0
DT, K=1
10
EPRR, K={1,5}
10
10
10
10
Proposed JRP, K=1
−2
−3
−4
−5
0
10
Proposed JRP, K=5
10
JRP for K=1, Eq. (59)
JRP for K=5, Eq. (59)
EPRR for K=1
EPRR for K=5
DT for K=1, Eq. (63)
DT for K=5, Eq. (63)
10
20
Transmit SNR [dB]
Average Symbol Error Rate
Secrecy Outage Probability
10
−1
10
30
40
Fig. 6. Overall SOP versus the transmit SNR for the CE case with K =
{1, 5} relays and L = 5 passive Eves. The number of source antennas is
set to Ns = 16 and the target rate to Rt = 1 bits/s/Hz. The asymptotic
curves are shown with dashed lines using Eq. (61), while the filled circles
depict the Monte-Carlo simulations.
to a nonzero constant as ρ → ∞, the SOP of the proposed JRP
scheme for both NCE and CE cases approaches zero in the
asymptotic SNR regime. Furthermore, as mentioned in Section
IV-C, the proposed JRP scheme achieves the diversity order
equal to the number of untrusted relays, which can be obtained
simply from the asymptotic curves in Figs. 5 and 6. Finally,
we observe from Fig. 6 that, while the DT policy fails to
establish confidential communication for target secrecy rate of
Rt = 1 bits/s/Hz, the proposed JRP scheme enhances the
PLS remarkably. This highlights the efficiency of the proposed
JRP scheme.
To study the level of reliability of the proposed secure
transmission scheme, we plot the average SER metric in Fig.
7. We consider the multi-relay scenario with K = 5 and QPSK
modulation. For this network topology, the DT policy offers
a superior average SER relative to the proposed JRP scheme.
The reason is that the DT policy provisions a better SNR at the
destination compared with the proposed scheme. Furthermore,
as observed, the average SER of NCE case is lower than the
CE case. The reason is that according to the received SINR
CE
N CE
at the destination in (37), we have γD
≤ γD
. Moreover,
we can conclude from Fig. 7 that the proposed JRP scheme is
better compared to OPRR. For example, the SNR gap between
the optimized network and OPRR is approximately 13 dB for
NCE and 12 dB for CE, respectively, when Ps = 10−2 . This
is because the proposed JRP selects the relay with the largest
second hop channel to signal transmission. Consequently, the
achievable reliability of the proposed scheme is higher than the
random relay selection of the OPRR policy. Evidently, unlike
the OPRR, the proposed scheme attains diversity order d = 5.
10
OPRR Policy
−1
−2
−3
DT Policy
Proposed JRP
−4
−5
0
JRP for NCE, Eq. (65)
JRP for CE, Eq. (65)
OPRR for NCE
OPRR for CE
DT [39]
5
10
15
20
Transmit SNR [dB]
25
30
Fig. 7. Average SER versus the transmit SNR for the number of relays K = 5
and multi-Eve L = 5 scenario. The number of source antennas Ns = 16 and
QPSK modulation is used. The asymptotic curves are shown with dashed lines
using Eq. (68), while the filled circles depict the Monte-Carlo simulations.
VII. C ONCLUSION
In this paper, we addressed secure transmission in a two-hop
relaying network including a multiple antennas source, a single
antenna destination, single antenna untrusted relays and single
antenna passive Eves. We considered two practical scenarios
of NCE and CE. A novel JRP scheme has been proposed for
security enhancement of the two NCE and CE networks. For
the proposed JRP scheme, a new closed-form expression was
presented for the ESR and the SOP as security metrics, and a
new closed-form expression was derived for the average SER
as a reliability measure over Rayleigh fading channels. We
further evaluated the high signal-to-noise ratio slope and power
offset of the ESR to reveal the impacts of system parameters on
the achievable ESR. Our findings highlighted that the diversity
order of the proposed JRP scheme is equal to the number of
untrusted relays. Numerical results presented that the ESR of
the proposed JRP scheme for NCE and CE cases increases as
the number of untrustworthy relays grows.
It would be interesting to extend the results in this paper to
the case, where trusted relays are also exist in the network.
Thus, we would like to consider this case as future work.
Acknowledgments
The authors deeply thank Prof. Dimitris Toumpakaris and Prof.
Lajos Hanzo for their constructive comments to improve the
paper. The authors also would like to thank the editor and the
anonymous reviewers for the constructive comments.
R EFERENCES
[1] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta,
O. Edfors, and F. Tufvesson, “Scaling up MIMO: Opportunities and
challenges with very large arrays,” IEEE Sig. Proc. Mag., vol. 30, no.
1, pp. 40-46, Jan. 2013.
15
[2] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Commun. Mag., vol.
52, no. 2, pp. 186-195 , Feb. 2014.
[3] D. Kapetanovic, G. Zheng, and F. Rusek, “Physical layer security for
massive MIMO: An overview on passive eavesdropping and active
attacks,” IEEE Commun. Mag., vol. 53, no. 6, pp. 21-27, Jun. 2015.
[4] Y. Wu, R. Schober, D. W. K. Ng, C. Xiao, and G. Caire, “Secure
massive MIMO transmission with an active eavesdropper,” IEEE Trans.
Inf. Theory., vol. 62, no. 7, pp. 3880-3900, July 2016.
[5] D. B. Rawat, T. White, M. S. Parwez, C. Bajracharya, and M. Song,
“Evaluating secrecy outage of physical layer security in large-scale
MIMO wireless communications for cyber-physical systems,” IEEE
Internet of Things Journal, Apr. 2017.
[6] X. He and A. Yener, “Two-hop secure communication using an untrusted
relay: A case for cooperative jamming,” in Proc. IEEE Global Commun.
Conf. (Globecom), New Orleans, LA, Dec. 2008, pp. 15.
[7] A. Mukherjee, S. A. A. Fakoorian, J. Huang, and A. L. Swindlehurst,
“Principles of physical layer security in multiuser wireless networks: A
survey,” IEEE Commun. Surveys Tuts., vol. 16, no. 3, pp. 1550-1573,
Feb. 2014.
[8] X. Chen, D. W. K. Ng, W. Gerstacker, and H-H. Chen, “A survey on
multiple-antenna techniques for physical layer security,” IEEE Commun.
Surveys Tuts., vol. 19, no. 2, pp. 1027-1053, Jun. 2017.
[9] Y. Zou, X. Wang, W. Shen, and L. Hanzo, “Security versus reliability
analysis of opportunistic relaying,” IEEE Trans. Veh. Tech., vol. 63, no.
6, pp. 2653-2661, Jul. 2014.
[10] C. Wang, H.-M. Wang, and X.-G. Xia, “Hybrid opportunistic relaying
and jamming with power allocation for secure cooperative networks,”
IEEE Trans. Wireless Commun., vol. 14, no. 2, pp. 589-605, Feb. 2015.
[11] H.-M. Wang, C. Wang, and D. W. K. Ng, “Artificial noise assisted
secure transmission under training and feedback”, IEEE Trans. on Signal
Process., vol. 63, no. 23, pp. 6285-6298, Dec. 2015.
[12] D. Fang, N. Yang, M. Elkashlan, P. L. Yeoh, and J. Yuan, “Cooperative
jamming protocols in two-hop amplify-and-forward wiretap channels,”
in Proc. IEEE Int. Conf. Commun. (ICC), Budapest, Hungary, pp. 21882192, Nov. 2013.
[13] I. Krikidis, J. S. Thompson, and S. McLaughlin, “Relay selection
for secure cooperative networks with jamming,” IEEE Trans. Wireless
Commun., vol. 8, no. 10, pp. 5003-5011, Oct. 2009.
[14] L. Wang, Y. Cai, Y. Zou, W. Yang and L. Hanzo, “Joint relay and
jammer selection improves the physical layer security in the face of
CSI feedback delays,” IEEE Trans. Veh. Technol., vol. 65, no. 8, pp.
6259-6274, Aug. 2016.
[15] Z. Han, N. Marina, M. Debbah, and A. Hjorungnes, “Physical layer
security game: Interaction between source, eavesdropper and friendly
jammer,” EURASIP J. Wireless Commun. Netw., vol. 2009, pp. 4529071 452907-10, Mar. 2009.
[16] Y. Liu, J. Li and A. P. Petropulu, “Destination assisted cooperative
jamming for wireless physical-layer security,” IEEE Trans. Inf. Foren.
Sec., vol. 8, no. 4, pp. 682-694, Apr. 2013.
[17] T. T. Kim and H. V. Poor, “On the secure degrees of freedom of relaying
with half-duplex feedback,” IEEE Trans. Inf. Theory., vol. 57, no. 1, pp.
291-302, Jan. 2011.
[18] S. I. Kim, I. M. Kim and J. Heo, “Secure transmission for multiuser
relay networks,” IEEE Trans. Wireless Commun., vol. 14, no. 7, pp.
3724-3737, July 2015.
[19] L. Wang, M. Elkashlan, J. Huang, N. H. Tran, and T. Q. Duong, “Secure
transmission with optimal power allocation in untrusted relay networks,”
IEEE Wireless Commun. Lett., vol. 3, no. 3, pp. 289-292, Jun. 2014.
[20] A. Kuhestani and A. Mohammadi, “Destination-based cooperative jamming in untrusted amplify-and-forward relay networks: Resource allocation and performance study,” IET Commun., vol. 10, no. 1, pp. 17-23,
Feb. 2016.
[21] L. Sun, T. Zhang, Y. Li, and H. Niu, “Performance study of two-hop
amplify-and-forward systems with untrustworthy relay nodes,” IEEE
Trans. Veh. Technol., vol. 61, no. 8, pp. 3801-3807, Oct. 2012.
[22] J.-B. Kim, J. Lim, and J. Cioffi, “Capacity scaling and diversity order
for secure cooperative relaying with untrustworthy relays,” IEEE Trans.
Wireless Commun., vol. 14, no. 7, pp. 3866-3876, July 2015.
[23] L. Sun, P. Ren, Q. Du, Y. Wang, and Z. Gao, “Security-aware relaying
scheme for cooperative networks with untrusted relay nodes,” IEEE
Commun. Lett., vol. 19, no. 3, pp. 463-466, Sep. 2014.
[24] W. Wang, K. C. Teh, and K. H. Li, “Relay selection for secure successive
AF relaying networks with untrusted nodes,”IEEE Trans. Inf. Forensics
Security., vol. 11, no. 11, pp. 2466-2476, Nov. 2016.
[25] A. Mukherjee, “Physical-layer security in the Internet of Things: Sensing
and communication confidentiality under resource constraints,” Proc.
IEEE, vol. 103, no. 10, pp. 1747-1761, Oct. 2015.
[26] J. Chen, H. Chen, H. Zhang and F. Zhao, “Spectral-energy efficiency
tradeoff in relay-aided massive MIMO cellular networks with pilot
contamination.,” IEEE Access, vol. 4, no. , pp. 5234-5242, Sept. 2016.
[27] Z. Zhang, Z. Chen, M. Shen, and B. Xia,“Spectral and energy efficiency
of multipair two-way full-duplex relay systems with massive MIMO”
IEEE J. Sel. Areas in Commun., vol. 34, no. 4, pp. 848-863, Apr. 2016.
[28] X. Zhou, R. K. Ganti and J. G. Andrews, “Secure wireless network
connectivity with multi-antenna transmission,” IEEE Wireless Commun.
Lett., vol. 10, no. 2, pp. 425-430, Feb. 2011.
[29] P. L. Yeoh, N. Yang, and K. J. Kim, “Secrecy outage probability of
selective relaying wiretap channels with collaborative eavesdropping,”in
Proc. IEEE Global Commun. Conf. (Globecom), Dec. 2015, pp. 1-6.
[30] L. Wang, M. Elkashlan, J. Huang, R. Schober, and R. K. Mallik,
“Secure transmission with antenna selection in MIMO Nakagami-m
fading channels,” IEEE Trans. Wireless Commun., vol. 13, no. 11, pp.
6054-6067, Nov. 2014.
[31] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, 7th ed. New York: Academic, 2007.
[32] P. Hao, X. Wang, and A. Behnad, “Relay authentication by exploiting
I/Q imbalance in amplify-and-forward system,”in Proc. IEEE Global
Commun. Conf. (Globecom), Dec. 2014, pp. 613-18.
[33] X. Wang, P. Hao and L. Hanzo, “Physical-layer authentication for wireless security enhancement: Current challenges and future developments,”
IEEE Commun. Mag., vol. 54, no. 6, pp. 152-158, June 2016.
[34] A. Papoulis, Probability, Random Variables, and Stochastic Processes.
New York: McGraw-Hill, 1984.
[35] Z. Wang and G. B. Giannakis, “Outage mutual information of space-time
MIMO channels,” IEEE Trans. Inf. Theory, vol. 50, no. 4, pp. 657-662,
Apr. 2004.
[36] A. Bletsas, A. Khisti, D. P. Reed, and A. Lippman, “A simple cooperative
diversity method based on network path selection”, IEEE J. Sel. Areas
Commun., vol. 24, no. 3 pp. 659-672, Mar. 2006.
[37] S.M. Kay, Fundamentals of Statistical Signal Processing, Volume I:
Estimation Theory, Prentice Hall, 1998.
[38] A. M. Tulino and S. Verdu, “Random matrix theory and wireless communications,” Foundations and Trends in Communications and Information
Theory, vol. 1, no. 1, pp. 1-182, Jun. 2004.
[39] J. G. Proakis, Digital Communications. 5th ed. McGraw-Hill, New York,
2008.
[40] Z. Rezki and M. S. Alouini, “Secure diversity-multiplexing tradeoff of
zero-forcing transmit scheme at finite-SNR,” IEEE Trans. Commun., vol.
60, no. 4, pp. 1138-1147, Apr. 2012.
| 7cs.IT
|
Wasserstein and total variation distance
between marginals of Lévy processes
arXiv:1710.02715v1 [math.PR] 7 Oct 2017
Ester Mariucci∗
Markus Rei߆
07 October 2017
Abstract
We present upper bounds for the Wasserstein distance of order p between the marginals of Lévy processes, including Gaussian approximations for jumps of infinite activity. Using the convolution structure, we
further derive upper bounds for the total variation distance between the
marginals of Lévy processes. Connections to other metrics like Zolotarev
and Toscani-Fourier distances are established. The theory is illustrated
by concrete examples and an application to statistical lower bounds.
Keywords: Lévy Processes; Wasserstein Distance; Total Variation; ToscaniFourier distance; Statistical lower bound.
1
Introduction
Lévy processes form the prototype of continuous-time processes with a continuous diffusion and a jump part. In applications, there is a high interest to
disentangle these parts based on discrete observations. While Aït-Sahalia and
Jacod [1] among many others propose an asymptotically (as the observation
distances become smaller) consistent test on the presence of jumps for general
semimartingale models, Neumann and Reiß [17] argue that, already inside the
class of α-stable processes with α ∈ (0, 2], no uniformly consistent test exists.
The subtle, but important, difference is the uniformity over the class of processes. Mathematically, the difference is that on the Skorokhod path space
D([0, T ]) α-stable processes, α ∈ (0, 2), induce laws singular to that of Brownian motion (α = 2), while their respective marginals at tk = kT /n, k = 0, . . . , n
for n fixed, have equivalent laws, which even converge in total variation distance
as α → 2 to those of Brownian motion. It is our aim here to shed some light
on the geometry of the marginal laws of one-dimensional Lévy processes and
to quantify the distance of the marginal laws non-asymptotically as a function
of the respective Lévy characteristics (b, σ 2 , ν). The marginals form, of course,
infinitely divisible distributions, but we prefer here the process point of view
which is sometimes more intuitive.
Let us recall the fundamental result by Gnedenko and Kolmogorov [11].
∗ Humboldt-Universität zu Berlin, Germany. E-mail: [email protected]. This
research was supported by the Federal Ministry for Education and Research through the
Sponsorship provided by the Alexander von Humboldt Foundation.
† Humboldt-Universität zu Berlin, Germany. E-mail: [email protected]
1
Theorem 1 ([11]). Marginals of Lévy processes X n = (Xtn )t≥0 with characteristics (bn , σn2 , νn ) converge weakly to marginals of a Lévy process X = (Xt )t≥0
with characteristics (b, σ 2 , ν) if and only if
w
bn → b and σn2 δ0 + (x2 ∧ 1)νn (dx) −
→ σ 2 δ0 + (x2 ∧ 1)ν(dx),
w
where δ0 is the Dirac measure in 0 and −
→ denotes weak convergence of finite
measures.
As a particular example, consider the compound Poisson process with Lévy
ε
measure δ−ε2ε+δ
that has jumps of size ε and −ε both at intensity 2ε12 . Then as
2
ε ↓ 0, the marginals converge to those of a standard Brownian motion, which can
also be derived from Donsker’s Theorem. Below, we shall be able to quantify
this rate of convergence for general Lévy processes in terms of the (stronger)
p-Wasserstein distances Wp . The derived Gaussian approximation of the small
jump part relies on the fine analysis by Rio [19] of the approximation error
in Wasserstein distance for the central limit theorem. This is the subject of
Theorem 9, of which the following is a simplified statement:
Result. Let X S (ε) be a Lévy process with characteristics (0,R0, νε ) where νε is
ε
a Lévy measure with support in [−ε, ε]. Introducing σ̄ 2 (ε) = −ε x2 ν(dx), there
exists a constant C depending only on p such that:
√
Wp (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤ C min ( tσ̄(ε), ε) ≤ Cε.
A Gaussian approximation of the small jumps of Lévy processes has already
been employed, for example when simulating trajectories of Lévy processes with
infinite Lévy measure (see e.g. [3]).
The above result is actually an intermediate step for the more general Corollary 3, that bounds the p-Wasserstein distance Wp in ℓr (Rn ) as follows:
Result. Let X j , j = 1, 2, be two Lévy processes with characteristics (bj , σj2 , νj ),
j = 1, 2. Then for all ε ≥ 0, T > 0 and n ∈ N we have
1
1
n
2
2
n
Wp (XkT
−
X
)
,
(X
−
X
)
kT /n
(k−1)T /n k=1
/n
(k−1)T /n k=1
1
1
1
≤ T n r −1 |b1 (ε) − b2 (ε)| + T 1/2 n r − 2 |σ1 + σ̄1 (ε) − σ2 − σ̄2 (ε)|
+C
2
X
j=1
where bj (ε) := bj −
1
1
1
1
2,B
min (T 1/2 n r − 2 σ̄j (ε), n r ε) + n r Wp (XT1,B
/n (ε), XT /n (ε)),
R
R
2
|x|≤ε x ν(dx) and C is
2,B
Wp (XT1,B
/n (ε), XT /n (ε)), involving
ε<|x|≤1 xνj (dx),
σ̄j2 (ε) :=
depending only on p. The term
larger ε, can be bounded as in Theorem 10.
a constant
the jumps
Sometimes we can even obtain bounds on the total variation distance, which
for statistical purposes, especially testing, is particularly meaningful. The currently available bound in the literature is by Liese [15].
Theorem 2 ([15, Cor. 2.7]). For Lévy processes X 1 and X 2 with characteristics
(b1 , σ12 , ν1 ) and (b2 , σ22 , ν2 ), respectively, introduce the squared Hellinger distance
of the Lévy measures (put ν0 = ν1 + ν2 ):
r
2
Z r
dν1
dν2
2
H (ν1 , ν2 ) :=
(x) −
(x) ν0 (dx).
dν0
dν0
R
2
Then the total variation distance between the laws of Xt1 and Xt2 is bounded as:
kL (Xt1 ) − L (Xt2 )kT V
r
2
1
exp − tH 2 (ν1 , ν2 )
≤ 2 1 − 1 − H 2 N (b̃1 t, σ12 t), N (b̃2 t, σ22 t)
2
R1
R1
with b̃1 = b1 − −1 xν1 (dx), b̃2 = b2 − −1 xν2 (dx).
Note that the bound is very loose or even trivial in the case ν2 = 0 and
λ1 = ν1 (R) > 1/t because then tH 2 (ν1 , ν2 ) = tλ1 > 1. So, this bound does not
allow to deduce a total variation approximation of Brownian motion by jump
processes of infinite jump activity like α-stable processes with α ↑ 2. In fact, for
pure jump Lévy processes these bounds are analogous to the bounds by Mémin
and Shiryayev [16] in the path space D([0, T ]), where pure jump processes and
Brownian motion have singular laws (for other results on distances on D([0, T ])
see e.g. [5, 8, 13, 14]). Our main idea is to use the convolutional structure of
the laws to transfer bounds from Wasserstein to total variation distance. This
strategy is implemented for Lévy processes with a non-zero Gaussian component
(but without any restriction on the Lévy measures, which can be infinite, and
even with infinite variation) and yields Theorem 14:
Result. For Lévy processes X 1 and X 2 with characteristics (bj , σj2 , νj ) and
σj > 0, j = 1, 2, we have for all t > 0, ε ∈ [0, 1]:
q
p
√ p
t
b1 (ε) − b2 (ε) + 2 σ12 + σ̄12 (ε) − σ22 + σ̄22 (ε)
2π
p
p
kL (Xt1 ) − L (Xt2 )kT V ≤
σ12 + σ̄12 (ε) ∨ σ22 + σ̄22 (ε)
s
2
q
X
2
2 (ε), ε
tσ̄
min
2
+
j
πtσj2
2
j=1
+ t|λ1 (ε) − λ2 (ε)| + t(λ1 (ε) ∧ λ2 (ε))
ν1ε
νε
− 2
λ1 (ε) λ2 (ε)
with the above notation, νjε = νj (· \ (−ε, ε)) and λj (ε) = νjε (R).
,
TV
The results proven in this paper provide further insight in the geometry
of the space of discretely observed Lévy processes. At the same time, their
nonasymptotic character finds fruitful applications in nonparametric statistics,
when proving general lower bounds in a minimax sense. The technology is
shown at work in Section 5.2, making the original proof by Jacod and Reiß
[12] for volatility estimation under high activity jumps simpler and much more
transparent.
The paper is organized as follows. In Section 2 we review basic properties
of the Wasserstein distances and discuss their relationship with the Zolotarev
and Toscani-Fourier distances. Then we recall the main non-asymptotic bounds
for the Wasserstein distances in the CLT and introduce Lévy processes. Section
3 derives bounds between marginals of Lévy processes in Wasserstein distance.
The main focus is on the small jump part, which is treated in Theorem 9 and
for which the tightness of the bounds is discussed in detail, first for concrete
examples and then more generally using a lower bound via the Toscani-Fourier
distance. Main results are presented in Section 3.3. Section 4 introduces properties of the total variation distance and then shows how bounds in Wasserstein or
3
Toscani-Fourier distance transfer under convolution to total variation bounds,
see e.g. Proposition 4 and Proposition 7. For Gaussian convolutions the different bounds are first compared and then applied to the marginals of Lévy
processes. Section 5 is devoted to the application of the total variation bounds
for proving the minimax-optimality of integrated volatility estimators in the
presence of jumps proposed in [12].
2
2.1
Preliminaries
The Wasserstein distances
Let (X , d) be a Polish metric space. Given p ∈ [1, ∞), let Pp (X ) denote the
space of all Borel probability measures µ on X such that the moment bound
Eµ [d(X, x0 )p ] < ∞
holds for some (and hence all) x0 ∈ X .
Definition 1. Given p ≥ 1, for any two probability measures µ, ν ∈ Pp (X ), the
Wasserstein distance of order p between µ and ν is defined by
o
n
1
(1)
Wp (µ, ν) = inf [E[d(X ′ , Y ′ )p ]] p , L (X ′ ) = µ, L (Y ′ ) = ν ,
where the infimum is taken over all random variables X ′ and Y ′ having laws µ
and ν, respectively. We abbreviate Wp (X, Y ) = Wp (L (X), L (Y )) for random
variables X, Y with laws L (X), L (Y ) ∈ Pp (X ).
The following lemma introduces some properties of the Wasserstein distances
that we will use throughout the paper. For a proof, the reader is referred to
[24], Chapter 6.
Lemma 1. The Wasserstein distances have the following properties:
(1) For all p ≥ 1, Wp (·, ·) is a metric on Pp (X ).
(2) If 1 ≤ p ≤ q, then Pq (X ) ⊆ Pp (X ), and Wp (µ, ν) ≤ Wq (µ, ν) for every
µ, ν ∈ Pq (X ).
(3) Given a sequence (µn )n≥1 and a probability measure µ in Pp (X )
lim Wp (µn , µ) = 0
n→∞
if and only if (µn )n≥1 converges to µ weakly and for some (and hence all)
x0 ∈ X
Z
Z
d(x, x0 )p µ(dx).
d(x, x0 )p µn (dx) =
lim
n→∞
X
X
(4) The infimum in (1) is actually a minimum; i.e., there exists a pair (X ∗ , Y ∗ )
of jointly distributed X -valued random variables with L (X ∗ ) = µ and
L (Y ∗ ) = ν, such that
Wp (µ, ν)p = E[d(X ∗ , Y ∗ )p ].
4
Following the terminology used in [25], we can say that the Wasserstein
distances are ideal metrics since they possess the following two properties.
Lemma 2. Let X be a separable Banach space. For any three X -valued random
variables X, Y, Z, with Z independent of X and Y , the inequality
Wp (X + Z, Y + Z) ≤ Wp (X, Y )
holds. Furthermore, for any real constant c, we have
(2)
Wp (cX, cY ) = |c|Wp (X, Y ).
Proof. Lemma 1 guarantees the existence of two random variables X ∗ , Y ∗ , independent of Z, such that
Wp (X, Y ) = (E[d(X ∗ , Y ∗ )p ])
1/p
.
We have:
1
1
Wp (X + Z, Y + Z) ≤ (E[d(X ∗ + Z, Y ∗ + Z)p ]) p = (E[d(X ∗ , Y ∗ )p ]) p = Wp (X, Y ).
The equality (2) follows by homogeneity of the expectation.
An immediate corollary of Lemma 2 is the subadditivity of the metric Wp
under independence (or equivalently under convolution of laws).
Corollary 1. If X1 , . . . , Xn are independent random variables as well as Y1 , . . . , Yn ,
then
n
X
Wp (Xi , Yi ).
Wp (X1 + · · · + Xn , Y1 + · · · + Yn ) ≤
i=1
Proof. By induction, it suffices to prove the case n = 2. Let X̃2 be a random
variable equal in law to X2 and independent of Y1 and of X1 . By means of
Lemma 2 we have
Wp (Y1 + X̃2 , Y1 + Y2 ) ≤ Wp (X̃2 , Y2 ) = Wp (X2 , Y2 ),
Wp (X1 + X2 , Y1 + X̃2 ) ≤ Wp (X1 , Y1 ).
Hence, by triangle inequality Wp (X1 + X2 , Y1 + Y2 ) ≤ Wp (X1 , Y1 ) + Wp (X2 , Y2 )
follows.
A useful property of the Wasserstein distances is their good behaviour with
respect to products of measures.
Lemma
3 (Tensorisation). Let (X , d) be Rn endowed with the distance d(x, y) =
Pn
r 1/r
( i=1
, rN
≥ 1, for all x = (x1 , . . . , xn ), y = (y1 , . . . , yn ) and let
i − yi | )
N|x
n
µ = i=1 µi and ν = ni=1 νi be two product measures on Rn . Then,
p
Wp (µ, ν) ≤ max(n
p
r −1
, 1)
n
X
i=1
Wp (µi , νi )p .
In the special case where µ1 = · · · = µn and ν1 = · · · = νn , the following stricter
inequality holds:
p
Wp (µ, ν)p ≤ n r Wp (µ1 , ν1 )p .
5
Proof. Thanks to Point (4) in Lemma 1, we can always find two random vectors
X ∗,n = (X1∗ , . . . , Xn∗ ), Y ∗,n = (Y1∗ , . . . , Yn∗ ) with independent coordinates such
that µi = L (Xi∗ ), νi = L (Yi∗ ) and Wp (µi , νi ) = E[|Xi∗ −Yi∗ |p ]1/p . In particular,
we have:
p/r
X
n
.
(3)
|Xi∗ − Yi∗ |r
Wp (µ, ν)p ≤ E[d((Xi∗ )i , (Yi∗ )i )p ] = E
i=1
If p ≥ r, by means of the elementary inequality (z1 + · · · + zn )q ≤ nq−1 (z1q +
· · · + znq ), q ≥ 1, we deduce from (3) that
Wp (µ, ν)p ≤ np/r−1
n
X
i=1
E[|Xi∗ − Yi∗ |p ] = np/r−1
n
X
i=1
Wp (µi , νi )p .
Similarly, if p < r, the proof follows by the inequality (|z1 |+ · · · + |zn |)1/q ≤
|z1 |1/q + · · · + |zn |1/q , q ≥ 1.
In the case where µ1 = · · · = µn and ν1 = · · · = νn , one may choose
X1∗ = · · · = Xn∗ and Y1∗ = · · · = Yn∗ . The conclusion readily follows.
The distance W1 is commonly called the Kantorovich-Rubinstein distance
and it can be characterized in many different ways. Some useful properties of
the distance W1 are the following.
Proposition 1. [See [9]] Let X and Y be integrable real random variables.
Denote by µ and ν their laws and by F and G their cumulative distribution
functions, respectively. Then the following characterizations of the Wasserstein
distance of order 1 hold:
Z
1. W1 (X, Y ) = |F (x) − G(x)|dx,
R
2. W1 (X, Y ) =
Z
0
1
|F −1 (t) − G−1 (t)|dt,
3. W1 (X, Y ) = supkψkLip ≤1
R
R ψdµ −
R
ψdν
, the supremum being taken
R
over all ψ satisfying the Lipschitz condition |ψ(x) − ψ(y)|≤ |x − y|, for
all x, y ∈ R. This property is generally called Kantorovich-Rubinstein
formula.
2.2
Wasserstein, Zolotarev and Toscani-Fourier distances
Let µ, ν be two probability measures on R endowed with the distance d(x, y) =
|x − y|, x, y ∈ R. Writing p > 0 as p = m + α with m ∈ N0 and 0 < α ≤ 1,
denote by Fp the Hölder class of real-valued bounded functions f on R which
are m-times differentiable with
|f (m) (x) − f (m) (y)| ≤ |x − y|α .
Definition 2. The Zolotarev distance Zp between µ and ν is defined by
Z
Z
Zp (µ, ν) = sup
f dµ − f dν .
f ∈Fp
R
6
R
Remark 1. It is easy to see that the functional Zp is a metric. For p = 0
the metric Zp is defined by the relation Z0 = limp→0 Zp and F0 is the set of
Borel functions satisfying the condition |f (x) − f (y)|≤ Ix6=y . Thanks to the
characterisation of the total variation given in Property 6 below, it follows that
Z0 (µ, ν) = kµ − νkT V . Also, by means of the Kantorovich-Rubinstein formula,
recalled in Property 1, we have Z1 (µ, ν) = W1 (µ, ν).
The following result shows that the Wasserstein distance of order p is bounded
by the p-th root of the Zolotarev distance Zp . This fact, together with Theorem
4 below, will be a useful tool to control the Wasserstein distances between the
increments of compound Poisson processes.
Theorem 3 (See [19], Theorem 3.1). For any p ≥ 1 there exists a positive
constant cp such that for any pair (µ, ν) of laws on the real line with finite
absolute moments of order p
p
(Wp (µ, ν)) ≤ cp Zp (µ, ν).
Theorem 4. [See [25], Theorem 1.4.3] Let (Xi )i≥1 and (Yi )i≥1 be sequences
of independent random variables and N be an integer-valued random variable
independent of the random variables from both sequences. Then,
Zp
X
N
i=1
Xi ,
N
X
Yi
i=1
∞
X
≤
k=1
P(N ≥ k)Zp (Xk , Yk ).
Theorem 5 (See [25], Theorem 1.4.2.). Let X and Y be integrable real random
variables with laws µ and ν, respectively. Then the following characterization
of the Zolotarev distance holds: for any p ≥ 1
Z Z x
(x − u)p−1
Zp (X, Y ) =
(µ − ν)(du) dx,
Γ(p)
−∞
where Γ denotes the Gamma function.
Let P1 and P2 be two probability measures on the real line. We will denote
by ϕ1 (resp. ϕ2 ) the characteristic function of P1 (resp. P2 ), i.e.
Z
eiux P1 (dx).
ϕ1 (u) =
R
Also, denote by L b (R) (resp. L b (C)) the class of real-valued (resp. complexvalued) bounded functions on R with Lipschitz norm bounded by 1.
Definition 3. For s > 0, the Toscani-Fourier distance of order s, denoted by
Ts , is defined as:
Ts (P1 , P2 ) =
|ϕ1 (u) − ϕ2 (u)|
.
|u|s
u∈R\{0}
sup
The distance introduced in Definition 3 first appeared in [7], under the name
“Fourier-based metrics”, to study the trend to equilibrium for solutions of the
space-homogeneous Boltzmann equation for Maxwellian molecules. After that,
it has been used in several other works, and especially linked to the kinetic
theory, see [2] for an overview. In [24], T2 is called the “Toscani distance”.
7
Proposition 2. For all p ≥ 1
1
Wp (P1 , P2 ) ≥ √ T1 (P1 , P2 ).
2
Proof. Thanks to Lemma 1 and Property 1
Wp (P1 , P2 ) ≥ W1 (P1 , P2 ) =
1
sup
≥ √
2 ψ∈L b (C)
sup
ψ∈L b (R)
Z
R
Z
ψdP1 −
R
Z
ψdP1 −
Z
R
ψdP2
ψdP2 .
R
iux
For all u ∈ R \ {0}, let us consider the function Ψu (x) = e u and observe that
the Lipschitz norm of Ψu is 1. It immediately follows that
Z
Z
Z
Z
sup
ψdP1 − ψdP2 ≥ sup
Ψu dP1 − Ψu dP2 = T1 (P1 , P2 ).
ψ∈L b (C)
2.3
R
R
u∈R\{0}
R
R
Wasserstein distances in the central limit theorem
The class of Wasserstein metrics proves to be very useful in estimating the convergence rate in the central limit theorem. We recall some results. Let (Yi )i≥1
be a sequence of centred i.i.d. random variables
Pn with finite and positive vari1
ance σ 2 . We denote by µn the law of √nσ
2
i=1 Yi . For i.i.d. centred random
variables with finite absolute third moment, Esseen [4] proved the following
result.
Theorem 6. [See e.g. [18], Theorem 16] For any n ≥ 1,
1
E|Y1 |3
.
W1 (µn , N (0, 1)) ≤ √
2 n (Var[Y1 ])3/2
The constant
1
2
in this inequality cannot be improved.
A bound for the Wasserstein distances of order r ∈ (1, 2] is due to Rio [19]:
Theorem 7. [See [19], Theorem 4.1] For any n ≥ 1 and any r ∈ (1, 2], there
exists some positive constant C depending only on r such that
1/r
E[|Y1 |r+2 ]
Wr (µn , N (0, 1)) ≤ C √
r+2 .
n(Var[Y1 ]) 2r
For r > 2 and i.i.d. random variables with a finite absolute moment of order
r, we have the following:
Theorem 8. [See [21]] For any n ≥ 1 and r > 2, there exists some positive
constant C, depending only on r, such that
1/r
E[|Y1 |r ]
1
1
Wr (µn , N (0, 1)) ≤ C p
nr −2 .
Var[Y1 ]
8
If one only assumes finite absolute moment of order r, this rate cannot be
improved. In particular, under this assumption, the classical rate of convergence
√1 cannot be recovered for r > 2. For that reason, from now on, we will only
n
focus on the case r ∈ [1, 2].
2.4
Lévy processes
(b,σ,ν)
Let us denote by Pt
the marginal law at time t ≥ 0 of a Lévy process X
with characteristics (b, σ 2 , ν), i.e. (see Theorem 8.1 in [22])
Z
u2 σ 2
iux
iuXt
+ (e − 1 − iuxI|x|≤1 )ν(dx)
) = exp t iub −
E(e
2
R
Z
u2 σ 2
= exp t iub(ε) −
+ (eiux − 1 − iuxI|x|≤ε )ν(dx) ,
2
R
R
(b,σ,ν)
denotes
where b(ε) := b− ε<|x|≤1 xν(dx), for all ε ∈ (0, 1]. Equivalently, Pt
the infinitely divisible law with characteristics (bt, σ 2 t, νt). X can be characterised via the Lévy-Itô decomposition (see [22]), that is via a canonical representation with independent components X = X (1) + X (2) (ε) + X S (ε) + X B (ε):
For all ε ∈ (0, 1]
X
∆Xs I(η,ε] (|∆Xs |) − t(b(ε) − b(η))
Xt = σWt + b(ε)t + lim
η→0
+
X
0<s≤t
0<s≤t
∆Xs I(ε,+∞) (|∆Xs |),
where W is a standard Brownian motion, ∆Xs := Xs − limr↑s Xr is the jump at
time s of X, X S (ε) is a pure jump martingale containing only small jumps and
X B (ε) is a finite variation part containing jumps larger in absolute value than ε.
Thus X B (ε) is a compound Poisson process with intensity λε := ν(R \ (−ε, ε))
ν(dx)
and jump distribution Fε (dx) = ν(R\(−ε,ε))
I(ε,+∞) (|x|). In the following, someP
Nt
B
times we will write Xt (ε) as i=1 Yi where N is a Poisson process of intensity
λε independent of the sequence (Yi )i≥0 of i.i.d. random variables having distribution Fε . Also, for a given Lévy process X we define an auxiliary characteristic
σ̄ : R+ → R capturing the variance induced by small jumps:
Z
2
σ̄ (ε) :=
x2 ν(dx).
|x|≤ε
3
Wasserstein distances for Lévy processes
Let X j , j = 1, 2, be two Lévy processes with characteristics (bj , σj2 , νj ), j =
1, 2. As we will see later, thanks to Corollary 1 and the Lévy-Itô decomposition, in order to control Wp (Xt1 , Xt2 ) it is enough to separately control
the Wasserstein distances between two Gaussian random variables as well as
Wp (L (Xtj,S (ε)), N (0, tσ̄j2 (ε))) and Wp (Xt1,B (ε), Xt2,B (ε)). A bound for the
Wasserstein distances between Gaussian distributions is given by:
9
Lemma 4. [See [10], Prop. 7]
W2 (N (m1 , σ12 ), N (m2 , σ22 )) =
p
(m1 − m2 )2 + (σ1 − σ2 )2 .
Upper bounds for Wp (L (Xtj,S (ε)), N (0, tσ̄j2 (ε))) and Wp (Xt1,B (ε), Xt2,B (ε))
will be the subject of Sections 3.1 and 3.2, respectively.
3.1
Distances between marginals of small jump Lévy processes
Let X be a Lévy process with Lévy measure ν and denote by X S (ε) the Lévy
process associated with the small jumps of X, following the notation introduced
in Section 2.4.
Theorem 9. For any p ∈ [1, 2], there exists a positive constant C such that
R ε
√
|x|p+2 ν(dx) 1/p
S
2
Wp (L (Xt (ε)), N (0, tσ̄(ε) )) ≤ C min
tσ̄(ε), −ε 2
σ̄ (ε)
√
≤ C min
tσ̄(ε), ε .
(4)
√
In particular, for p = 1 the bound is min(2 tσ̄(ε), 12 ε).
Remark 2. The inequality
√
Wp (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤ W2 (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤ 2 tσ̄(ε)
is clear from the definition of W2 , noting that tσ̄ 2 (ε) is the second moment of
both arguments. The interest of Theorem 9 lies in the bound
Wp (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤ 2ε,
(5)
which after renormalisation yields
S
Xt (ε)
Cε
Wp L √
, N (0, 1) ≤ √
.
tσ̄(ε)
tσ̄(ε)
Thus, not surprisingly in view of the central limit theorem, the Gaussian approximation is better as t is large. Also whenever σ̄ 2 (ε) is much larger than ε2 ,
then a Gaussian approximation is valid, e.g. for α-stable processes with α > 0
and ε small, see Example 1 below.
Remark 3. The upper bound (4) gives in general the right order. Indeed, let
ε
us consider for ε > 0 the Lévy measure νε = δ−ε2ε+δ
and denote by Y (ε) the
2
corresponding (centred) pure jump Lévy process, i.e. Yt (ε) = ε(Nt1 (ε)−Nt2 (ε)) is
the rescaled difference of two independent Poisson processes of intensity λ = 2ε12
Rε
each. In particular, observe that σ̄ 2 (ε) = 1 and −ε |x|3 νε (dx) = ε.
Let us develop the case p = 1. Applying the scheme of proof proposed in
[19], see proof of Theorem 5.1, we show that there exists a constant K such that
√
W1 (L (Yt (ε)), N (0, t)) ≥ K min( t, ε).
To see that, we consider the cases where t ≤ ε2 and t > ε2 separately.
10
• t ≤ ε2 : From the definition of the Wasserstein distance of order 1 it follows
that
W1 (L (Yt (ε)), N (0, t)) ≥ E[|N (0, t)|]P(Nt1 (ε) + Nt2 (ε) = 0)
r
r
2t − t2
2 1√
ε
≥
e
=
t.
π
πe
• t ≥ ε2 : Again, by the definition of the Wasserstein distance of order 1, we
find that
i
h
√
W1 (L (Yt (ε)), N (0, t)) ≥ E min| tN − nε|
n∈Z
[
ε √
≥ P ( t/ε)N ∈
[n + 1/4, n + 3/4]
4
n∈Z
√
with N ∼ N (0, 1). Since in this case ( t/ε)N has variance at least one,
there exists a constant K such that
W1 (L (Yt (ε)), N (0, t)) ≥ Kε.
In the case p ∈ (1, 2] Wp is even larger than W1 . For the case p = 2 see also [6].
Example 1. Let us illustrate Theorem 9 for the class of α-stable Lévy processes
with a Lévy density proportional to |x|11+α , α ∈ [0, 2). For all ε ∈ (0, 1], let us
denote by X S (ε) the Lévy process describing the small jumps and by νI[−ε,ε]
its Lévy measure, i.e.
X S (ε) ∼ (0, 0, νI[−ε,ε] ),
ν(dx) =
Cα
dx,
|x|1+α
for some constant Cα . In particular, we have
Z ε
ε2−α
σ̄ 2 (ε) =
x2 ν(dx) = 2Cα
.
2−α
−ε
Therefore an application of Theorem 9 guarantees the existence of a constant
C, possibly depending on p and α, such that:
X S (ε)
√ α
t
, N (0, t) ≤ C min ( t, ε 2 ),
Wp L
σ̄(ε)
∀t > 0, ∀ε ∈ (0, 1], ∀p ∈ [1, 2].
(6)
Equation (6) validates the intuition that a Gaussian approximation of the small
jumps is the better the more active the small jumps are. Indeed, the approximation in (6) is better when α is larger.
Let us now prove Theorem 9. For that we need to recall the following lemma:
Lemma 5 (See [20], Lemma 6.). Let X be a Lévy process with Lévy measure ν.
R
=0
If a Borel function f : R → R satisfies |x|≥1 f (x)ν(dx) < ∞, limx→0 fx(x)
2
and f (x)(|x|2 ∧1)−1 is bounded, then
Z
1
lim E[f (Xt )] = f (x)ν(dx).
t→0 t
11
Proof of Theorem 9. Let us introduce n random variables defined by Yj =
√
S
S
n(Xtj/n
(ε) − Xt(j−1)/n
(ε)). The Yj ’s are i.i.d. centred random variables with
Pn
2
variance equal to tσ̄ (ε) and such that XtS (ε) = √1n j=1 Yj . An application
√
S
(ε)
of Theorems 7 and 6 (using the fact that Yj has the same law as nXt/n
and the homogeneity property of the Wasserstein distances stated in Lemma 2)
gives
W1 (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤
S
nE[|Xt/n
(ε)|3 ]
2tσ̄ 2 (ε)
.
Let us now argue that
lim sup
t→0
E[|XtS (ε)|3 ]
≤
t
Z
|x|<ε
|x|3 ν(dx).
Indeed, applying Lemma 5 to the family f (x) = fR (x) = |x|3 I[−R,R] (x) for
R > ε, we deduce that
Z
E[|XtS (ε)|3 I|XtS (ε)|≤R ]
lim
=
|x|3 ν(dx).
t→0
t
|x|<ε
R
Thus, using the fact that E[(XtS (ε))4 ] = t |x|<ε x4 ν(dx) + 3t2 σ̄ 4 (ε), we get
E[|XtS (ε)|3 ] ≤ E[|XtS (ε)|3 I|XtS (ε)|≤R ] + E[(XtS (ε)4 /R)I|XtS (ε)|>R ]
Z
1
t
x4 ν(dx) + 3t2 σ̄ 4 (ε) .
≤ E[|XtS (ε)|3 I|XtS (ε)|≤R ] +
R
|x|<ε
Therefore, for any R > ε,
E[|XtS (ε)|3 ]
lim sup
≤
t
t→0
Z
3
|x|<ε
|x| ν(dx) +
R
|x|<ε
x4 ν(dx)
R
Taking the limit as R → ∞, we conclude. It follows that
W1 (L (XtS (ε)), N (0, tσ̄ 2 (ε)))
≤ lim sup
n→∞
S
nE[|Xt/n
(ε)|3 ]
2tσ̄ 2 (ε)
≤
.
Rε
3
−ε |x| ν(dx)
.
2σ̄ 2 (ε)
Moreover, by definition of the Wasserstein distance of order 1 and denoting by
N a centered Gaussian random variable with variance tσ̄ 2 (ε), we have
p
W1 (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤ E[|XtS (ε)|] + E[|N |] ≤ 2 tσ̄ 2 (ε).
We deduce
W1 (L (XtS (ε)), N (0, tσ̄ 2 (ε)))
Rε
3
p
−ε |x| ν(dx)
2
.
≤ min 2 tσ̄ (ε),
2σ̄ 2 (ε)
Similarly, by means of Theorem 7, for p ∈ (1, 2]
Wp (L (XtS (ε)), N (0, tσ̄ 2 (ε)))
√
S p+2 1/p
C E[| nXt/n | ]
≤ lim sup √
n
tσ̄ 2 (ε)
n→∞
R ε
|x|p+2 ν(dx) 1/p
−ε
≤C
σ̄ 2 (ε)
12
and also
1/p
1/p
p
Wp (L (XtS (ε)), N (0, tσ̄ 2 (ε))) ≤ E[|XtS (ε)|p ]
+ E[|N |p ]
≤ 2 tσ̄ 2 (ε).
The upper bound (4) follows by the fact that
Rε
−ε
|x|p+2 νi (dx)
σ̄i2 (ε)
≤ εp .
Theorem 9 can be used to bound the Wasserstein distances between the
increments of the small jumps of two Lévy processes.
R
Corollary 2. For all ε ∈ (0, 1] let X j (ε) ∼ (− ε<|x|≤1 xνj (dx), σj2 , νj I[−ε,ε] ) be
two Lévy processes with σ̄j2 (ε) 6= 0 and σj ≥ 0, j = 1, 2. Then, for all p ∈ [1, 2],
there exists a constant C, only depending on p, such that
R ε
2
X
√
|x|p+2 νj (dx) 1/p
1
2
C min
Wp (Xt (ε), Xt (ε)) ≤
tσ̄j (ε), −ε 2
σ̄j (ε)
j=1
q
q
2
σ̄12 (ε) + σ12 − σ̄22 (ε) + σ22 .
+t
Proof. This is a consequence of Theorem 9 and Lemma 4.
3.2
Distances between random sums of random variables
Theorem 10. Let (Xi )i≥1 and (Yi )i≥1 be sequences of i.i.d. random variables
with Yi ∈ L2 and N , N ′ be two positive integer-valued random variables with N
(resp. N ′ ) independent of (Xi )i≥1 (resp. (Yi )i≥1 ). Then, for 1 ≤ p ≤ 2,
Wp
X
N
′
N
X
Xi ,
i=1
i=1
Yi
1/p
≤ min (cp E[N ]Zp (X1 , Y1 )) , E[N p ]1/p Wp (X1 , Y1 )
+ Wp (N, N ′ )E[|Y1 |p ]
1/p
with the constant cp from Theorem 3.
Proof. By the triangle inequality,
X
X
X
N′
N ′′
N ′′
N
N′
N
X
X
X
Yi ,
Yi ,
Yi + Wp
Xi ,
Yi ≤ Wp
Xi ,
Wp
i=1
i=1
i=1
i=1
i=1
(7)
i=1
where N ′′ is independent of (Yi )i≥1 and with the same law as N . Thanks to Theorems 3 and 4, the first summand in (7) is bounded by (cp E[N ]Zp (X1 , Y1 ))1/p .
Alternatively, this summand can be estimated via Jensen’s inequality joined
PÑ
PN
PN ′′
p
p
with the fact that Wp ( i=1 Xi , i=1 Yi ) ≤ E[| i=1 (Xi − Yi )| ], Ñ indepen′′
dent of (Xi , Yi )i≥1 and L (Ñ ) = L (N ) = L (N ), as follows:
X
Ñ
Ñ
p
X
p
p−1
|Xi − Yi | ≤ E[N p ]E[|X1 − Y1 |p ].
≤ E Ñ
(Xi − Yi )
E
i=1
i=1
Therefore,
Wp
X
N
i=1
′′
Xi ,
N
X
i=1
Yi
p
X
Ñ
p
(Xi − Yi ) , Ñ independent of (Xi , Yi )i≥1
≤ inf E
i=1
≤ E[N p ]Wp (X1 , Y1 )p .
13
To control the second summand, we proceed similarly
Wp
X
N ′′
i=1
′
Yi ,
N
X
i=1
Yi
p
X
N′
N ′′
p
X
′′
′
Yi , N , N independent of (Yi )i≥1
Yi −
≤ inf E
i=1
i=1
≤ E[|Y1 |p ]Wp (N ′′ , N ′ )p ,
which, by noting L (N ′′ ) = L (N ), concludes the proof.
In the preceding theorem one term is bounded alternatively by the Zolotarev
or the Wasserstein distance between X1 and Y1 . The difference is the factor in
front which is either the first or the pth moment of N . If N is likely to be large,
then better bounds can be obtained by profiting from the variance stabilisation
for centred sums. Since the larger jumps are not our main issue, this is not
pursued further.
In the Poisson case the moments and the Wasserstein distances can be easily
analysed.
Proposition 3. Let N and N ′ be two Poisson random variables of mean λ and
λ′ , respectively. Let us denote by m(p,ℓ) the moment of order p of a Poisson
random variable of mean ℓ, i.e.
p
i
X
p
1X
i−j i
i p
(−1)
jp.
ℓ
, where
:=
m(p,ℓ) :=
j
i!
i
i
j=0
i=1
Then the following upper bound holds for p ≥ 1 :
Wp (N, N ′ )p ≤ m(p,|λ−λ′ |) .
In particular,
W1 (N, N ′ ) ≤ |λ − λ′ |
Wp (N, N ′ )p ≤ |λ − λ′ |+|λ − λ′ |p ,
1 < p ≤ 2.
(8)
(9)
Proof. Without loss of generality, let us suppose λ′ ≥ λ and let N ′′ be a Poisson
random variable with mean λ − λ′ , independent of N . Thanks to Lemma 2 we
have
Wp (N, N ′ )p = Wp (N ′ + N ′′ , N ′ )p ≤ Wp (0, N ′′ )p ≤ E[(N ′′ )p ] = m(p,λ′ −λ) .
To deduce (8) and (9) we use the fact that m(1,ℓ) = ℓ, m(2,ℓ) = ℓ + ℓ2 and
E[(N ′′ )p ] ≤ E[N ′′ ]2−p E[(N ′′ )2 ]p−1 for p ∈ (1, 2] by Hölder’s inequality.
3.3
First main result
We will use the notation introduced in Section 2.4. In accordance with that,
for any given Lévy process X j with characteristics (bj , σj2 , νj ), X j,B (ε) will be
a compound Poisson process with Lévy measure νj (dx)I(ε,∞) (|x|), i.e.
j
Xtj,B (ε)
=
Nt
X
i=1
14
(j)
Yi
where N j is a Poisson process of intensity λj (ε) := νj (R\(−ε, ε)) independent of
(j)
the sequence of i.i.d. random variables (Yi )i≥0 having distribution Fεj (dx) =
I(ε,∞) (|x|)
λj (ε) νj (dx). Recall from Proposition 3 that m(p,ℓ) denotes the moment of
order p of a Poisson random variable of mean ℓ.
Theorem 11. Let X j , j = 1, 2, be two Lévy processes with characteristics
(bj , σj2 , νj ), j = 1, 2. For all p ∈ [1, 2], for all ε ∈ [0, 1] and all t ≥ 0, the
following estimate holds
1/2
2
2
Wp (Xt1 , Xt2 ) ≤ t2 (b1 (ε) − b2 (ε)) + t(σ1 + σ̄1 (ε) − σ2 − σ̄2 (ε))
+C
2
X
j=1
min
√
tσ̄j (ε), ε + Wp (Xt1,B (ε), Xt2,B (ε)),
for some constant C, only depending on p. Introducing Lt (ε) := t|λ1 (ε)−λ2 (ε)|,
we have
(1)
(2)
Wp (Xt1,B (ε), Xt2,B (ε)) ≤ ((tλ1 (ε))1/p + tλ1 (ε))Wp (Y1 , Y1 )
(2) p 1/p
+ (Lt (ε)1/p + Lt (ε))E[(Y1 ) ]
.
Proof of Theorem 11. By some abuse of notation let N (µ, σ 2 ) denote a random
variable with this distribution. Then, thanks to the Lévy-Itô decomposition, we
have
Xti = N (tbi (ε), tσi2 ) + Xti,S (ε) + Xti,B (ε)
with independent summands. Hence, by subadditivity we get
Wp (Xt1 , Xt2 ) ≤ Wp (N (tb1 (ε), tσ12 ) + Xt1,S (ε), N (tb2 (ε), tσ22 ) + Xt2,S (ε))
+ Wp (Xt1,B (ε), Xt2,B (ε)).
Observe that
Wp (N (tb1 (ε), tσ12 ) + Xt1,S (ε), N (tb2 (ε), tσ22 ) + Xt2,S (ε))
≤ Wp (N (tb1 (ε), tσ12 ) + Xt1,S (ε), N (tb1 (ε), t(σ12 + σ̄12 (ε))))
+ Wp (N (tb2 (ε), tσ22 ) + Xt2,S (ε), N (tb2 (ε), t(σ22 + σ̄22 (ε))))
+ Wp (N (tb1 (ε), t(σ12 + σ̄22 (ε))), N (tb2 (ε), t(σ22 + σ̄22 (ε))))
≤ Wp (Xt1,S (ε), N (0, tσ12 (ε))) + Wp (Xt2,S (ε), N (0, tσ̄22 (ε)))
+ Wp (N (tb1 (ε), t(σ12 + σ̄12 (ε))), N (tb2 (ε), t(σ22 + σ̄22 (ε)))),
where in the second inequality we used again Lemma 2. An application of
Theorem 9 together with Point (2) in Lemma 1 and Lemma 4 allows us to
bound
Wp (N (tb1 (ε), tσ12 ) + Xt1,S (ε), N (tb2 (ε), tσ22 ) + Xt2,S (ε))
by the quantity
2
√
1/2
X
2
2
tσ̄j (ε), ε ,
min
t2 (b1 (ε) − b2 (ε)) + t(σ1 + σ̄1 (ε) − σ2 − σ̄2 (ε))
+C
j=1
for some constant C only depending on p. Finally, Wp (Xt1,B (ε), Xt2,B (ε)) is
bounded by means of Theorem 10 and Proposition 3.
15
We now address the problem of how to compute the Wasserstein distance
between n given increments of two Lévy processes. To that end, fix a time span
1
1
2
T > 0, a sample size n ∈ N and consider the sample (XkT
/n −X(k−1)T /n , XkT /n −
2
n
X(k−1)T /n )k=1 . From Property 3 we know that we can measure the distance be2
2
n
1
1
n
tween the random vectors (XkT
/n − X(k−1)T /n )k=1 and (XkT /n − X(k−1)T /n )k=1
in terms of the Wasserstein distance between the marginals. This observation
combined with Theorem 11 allows us to obtain an upper bound for the Wasserstein distance of order p between the increments of these Lévy processes.
Corollary 3. Let X j , j = 1, 2, be two Lévy processes with characteristics
(bj , σj2 , νj ), j = 1, 2. Then, with respect to the ℓr -metric on RN given by
PN
1/r
d(x, y) := ( i=1 |xi − yi |r ) , r ≥ 1, for all p ∈ [1, 2], ε ≥ 0, T > 0, n ∈ N we
have
n
2
n
2
1
1
Wp (XkT
/n − X(k−1)T /n )k=1 , (XkT /n − X(k−1)T /n )k=1
1
1
1
≤ T n r −1 |b1 (ε) − b2 (ε)| + T 1/2 n r − 2 |σ1 + σ̄1 (ε) − σ2 − σ̄2 (ε)|
+C
2
X
j=1
1
1
1
1
2,B
min (T 1/2 n r − 2 σ̄j (ε), n r ε) + n r Wp (XT1,B
/n (ε), XT /n (ε)),
2,B
where C is a constant depending only on p. The term Wp (XT1,B
/n (ε), XT /n (ε))
can be bounded as in Theorem 10 with t = T /n.
In the Euclidean case r = 2 we see that in the bound for the Wasserstein
distance the drift part disappears as n → ∞ (T fixed), while the Gaussian
part remains invariant and the Gaussian approximation of small jumps gives
an error of order min(σ̄j (ε), n1/2 ε). The bound on the larger jumps scales as
n1/2 (T /n + (T /n)1/2 ) (for p = 1 even as T /n1/2 ) so that the entire bound on
the Wasserstein distance remains bounded as n → ∞.
3.4
Lower bounds
Applying the general lower bound established in Proposition 2 to Lévy processes,
we get the following result:
Corollary 4. Let W be a Brownian motion and X ε be a pure jump Lévy process
with jumps of absolute value less than ε ∈ (0, 1]. This means that X ε has Lévy
triplet (0, 0, νε ), supp(νε ) ⊂ [−ε, ε] and characteristic function
Z ε
ε
iuXtε
ϕt (u) = E[e
] = exp t
(eiux − 1 − iux)νε (dx) .
Let σ̄ 2 (ε) =
R
−ε
x2 νε (dx). Then for p ≥ 1
R
R
exp t (eiux − 1 − iux)νε (dx) − exp t
√
Wp (Xtε , σ̄(ε)Wt ) ≥ sup
2|u|
u∈R
By the bound given in Proposition 2 we usually do not lose in approximation
order as the following lower bound examples demonstrate.
16
(iux)2
2 νε (dx)
.
Let us start with the general case that at ε = 1 we have a standardised pure
jump process
X 1 withRE[Xt1 ] = 0, Var[Xt1 ] = t as for Brownian motion, which
R
means xν1 (dx) = 0, x2 ν1 (dx) = 1. Then rescaling as in Donsker’s Theorem
we consider Xtε := εXε1−2 t such that νε (B) = ε−2 ν1 (ε−1 B) for Borel sets B,
σ̄ 2 (ε) = 1 and
Z 1
ϕεt (u) = exp tε−2
(eiεux − 1)ν1 (dx) .
−1
R 3
Let us further assume that q3 := x ν1 (dx) 6= 0. Then, taking into account the
first two moments, a Taylor expansion yields
Z 1
itεu3
tu2
−2
tε
−
q3 + O(tε2 u4 ).
(eiεux − 1)ν1 (dx) = −
2
3!
−1
For t > ε2 we thus obtain at u0 = t−1/2
2
q
√
−1/2
|ϕε (u0 ) − e−tu0 /2 |
3
/3!+O(ε2 /t)
= t1/2 e−1/2 e−iq3 εt
−1 = ε √ +O(ε/ t) .
u0
e3!
Hence, by Corollary 4 there are constants M > 0 and c > 0 such that for all
t ≥ M ε2
Wp (Xtε , Wt ) ≥ cε.
For t ≤ M ε2 we obtain at u0 = (8M λ1 /t)1/2 with λ1 = ν1 (R)
2
R
−2
|ϕεt (u0 ) − e−tu0 /2 |
t1/2
(cos(εu0 x)−1)ν1 (dx)
≥
etε
− e−4Mλ1
1/2
u0
(8M λ1 )
e−2Mλ1 − e−4Mλ1 1/2
t .
≥
(8M λ1 )1/2
We conclude
√
∀t > 0, ε ∈ (0, 1] : Wp (Xtε , Wt ) ≥ K min( t, ε)
for some positive
R constant K, depending on ν1 , but independent of t and ε,
whenever q3 = x3 ν1 (dx) 6= 0.
Even for symmetric Lévy measures,
√ not inducing skewness of the distribution, we can attain the order min( t, ε). If we consider ν1 = 12 (δ−1 + δ1 ),
ε
, we arrive at the same conclusion by computing the distance T1
νε = δ−ε2ε+δ
2
between Yt (ε) and N (0, t) as in Remark 3:
2
− exp − tu2
exp − t 1−cos(uε)
ε2
.
T1 (L (Yt (ε)), N (0, t)) = sup
u
u∈R
If t ≥ ε2 , we choose u =
2π
ε
If t < ε2 , the choice u =
3
√
t
and get
2π 2 t
1 − e−2π2
ε
1 − exp − 2
ε.
≥
T1 (L (Yt (ε)), N (0, t)) ≥
2π
ε
2π
gives
−2
−9
√ e −e 2
T1 (L (Yt (ε)), N (0, t)) ≥ t
.
3
√
We conclude that also in this case W1 (L (Yt (ε)), N (0, t)) ≥ K min( t, ε)
holds for some positive constant K, independent of t and ε.
17
4
Total variation bounds via convolution
4.1
Notation and some useful properties
Let (X , F ) be a measurable space and let µ and ν be two probability measures
on (X , F ).
Definition 4. The total variation distance between µ and ν is defined as
kµ − νkT V = sup |µ(A) − ν(A)|.
A∈F
Lemma 6. The total variation distance has the following properties.
R
1. kµ − νkT V = 12 supkΨk∞ ≤1 | X Ψ(x)(µ − ν)(dx)|.
2. kµ − νkT V = inf (P(X 6= Y ) : L (X) = µ, L (Y ) = ν).
Remark 4. Let X be a discrete set, equipped with the Hamming metric
d(x, y) = Ix6=y . In this case, thanks to Property 2. above, for any probability measures µ and ν on X we have
W1 (µ, ν) = kµ − νkT V .
The total variation distance does not always bound the Wasserstein distance,
because the latter is also influenced by large distances. However, thanks to the
following classical result, one can get some control on Wp given a bound on the
total variation distance.
Theorem 12. [See [24], Theorem 6.13] Let µ and ν be two probability measures
on a Polish space (X , d). Let p ∈ [1, ∞) and x0 ∈ X . Then
1
Wp (µ, ν) ≤ 2 p′
Z
p1
d(x0 , x)p |µ − ν|(dx) ,
1
1
= 1.
+
p p′
In particular, if p = 1 and the diameter of X is bounded by D, then
W1 (µ, ν) ≤ 2Dkµ − νkT V .
In Proposition 4 we will show an inequality that can be thought of as an
inverse of the one above. Namely, the total variation distance between two
measures convolved with a common measure can be bounded by a multiple of
the Wasserstein distance of order 1.
4.2
Wasserstein distance of order 1 and total variation
distance
Recall that a real function g is of bounded variation if its total variation norm
is finite, i.e.
nX
P −1
|g(xi+1 ) − g(xi )|< ∞,
kgkBV = sup
P ∈P i=0
where the supremum is taken over the set P = {P = (x0 , . . . , xnP ) : x0 < x1 <
· · · < xnP } of all finite ordered subsets of R. We will denote by BV (R) the
space of functions of bounded variation.
We now state a lemma that will be useful in the following.
18
Lemma 7. Let g be a real function of bounded variation and φ ∈ L∞ (R) such
that
Z
φ(y) g(t − y) − lim g(x) dy
hφ (t) =
x→−∞
R
is well defined. Then,
(10)
sup khφ kLip ≤ kgkBV .
kφk∞ ≤1
Proof. The proof is an easy consequence of the following classical results on
Lebesgue-Stieltjes measures:
1. For every right-continuous function g: R → R of bounded variation there
exists a unique signed measure µ such that
(11)
µ(] − ∞, x]) = g(x) − lim g(y).
y→−∞
2. Let φ ∈ L∞ (R) and let g ∈ BV (R) be a right-continuous function.
Let µ
R
be the finite signed measure associated to g as in (11).Then φ(t−y)µ(dy)
is well defined, measurable in t ∈ R and bounded in absolute value by
kφk∞ kgkBV .
More precisely,
let µ be the finite signed measure associated to g. It is enough to
R
prove that φ(t − y)µ(dy) is the weak
R derivative of hφ since then, using Point 2.
above, we deduce that khφ kLip = k φ(· − y)µ(dy)k∞ ≤ kφk∞ kgkBV and hence
(10). The claim above follows by Fubini’s Theorem: for all T > 0
Z TZ
Z Z
φ(t − y)µ(dy)dt =
I[0,T ] (u + y)φ(u)du µ(dy)
0
Z
= φ(u)(g(T − u) − g(−u))du
Z
T
= φ(u) g(t − u) − lim g(x) du
.
x→−∞
R
Hence, φ(t−y)µ(dy) is the weak derivative of
as desired.
R
t=0
φ(u)(g(t−u)−limx→−∞ g(x))du
Proposition 4. Let µ and ν be two measures on (R, B(R)) and G be an absolutely continuous measure with respect to the Lebesgue measure admitting a
density g of bounded variation. Then the total variation distance between the
convolution measures µ ∗ G and ν ∗ G is bounded by
kµ ∗ G − ν ∗ GkT V ≤
kgkBV
W1 (µ, ν).
2
Proof.
kµ ∗ G − ν ∗ GkT V =
1
sup
2 kφk∞ ≤1
1
sup
=
2 kφk∞ ≤1
1
sup
=
2 kφk∞ ≤1
Z
R
φ(x)(µ ∗ G − ν ∗ G)(dx)
Z Z
R
R
R
R
Z Z
19
φ(x)g(x − t)(µ − ν)(dt) dx
φ(x)g(x − t)dx (µ − ν)(dt) ,
the supremum
being taken over compactly supported functions φ. Denote by
R
hφ (t) = R φ(x)g(x − t)dx. From the last equality it follows that
Z
1
sup
sup
kµ ∗ G − ν ∗ GkT V ≤
ψ(t)(µ − ν)(dt) ,
2 kφk∞ ≤1 kψkLip ≤khφ kLip R
hence, applying Lemma 7 and Proposition 1, we deduce that
Z
kgkBV
kgkBV
sup
W1 (µ, ν).
kµ ∗ G − ν ∗ GkT V ≤
ψ(t)(µ − ν)(dt) =
2 kψkLip ≤1 R
2
The upper bound established in Proposition 4 is sharp. To see that, let us
consider the following example.
Example 2. Let µ = δ0 , ν = δε and G = N (0, 1) for some ε > 0. Denoting
by ϕ the density of a random variable N ∼ N (0, 1) and by Φ its cumulative
distribution function, we have
Z
ε
ε
ε
1
kν ∗ G − µ ∗ GkT V =
−Φ −
= 2Φ
−1
|ϕ(x) − ϕ(x − ε)|dx = Φ
2 R
2
2
2
ε
= √ + O(ε2 ).
2π
q
At the same time it is easy to see that W1 (µ, ν) = ε and kgkBV = π2 . Therefore, the upper bound established in Proposition 4
ε
1
kν ∗ G − µ ∗ GkT V ≤ √ W1 (µ, ν) = √
2π
2π
is exactly the correct estimate up to the first order.
4.3
Total variation distance and Toscani-Fourier distances
R
For any Lebesgue density f introduce its Fourier transform F f (u) = eiux f (x)dx.
A first elementary result linking the total variation distance between convolution
measures to Toscani-Fourier metrics is the following.
Proposition 5. Let µ, ν and G be probability measures and suppose that its
characteristic functions ϕµ , ϕν , ϕG are differentiable. Assume that G has
a Lebesgue density g with mth weak derivative g (m) . Then, for all k, j, r ∈
{1, . . . , m}, we have
kµ ∗ G − ν ∗ GkT V
√
√
|ϕ′µ (u) − ϕ′ν (u)| (j)
kg
k
≤ C Tk (µ, ν)kg (k) k2 + 2Tr (µ, ν)k(xg(x))(r) k2 + 2 sup
2
|u|j
u∈R
for some numerical constant C > 0.
Proof. First of all, remark that if any one among the kg (•) k2 , k(xg(x))(•) k2 ,
|ϕ′ (u)−ϕ′ (u)|
T• (µ, ν), or supu∈R µ |u|• ν
appearing above is infinite, then there is nothing to prove. Therefore, from now on, we will assume that they are all finite.
20
Since G admits a density g with respect to Lebesgue measure, µ ∗ G and ν ∗ G
have densities g ∗ µ and g ∗ ν.
Using the Cauchy-Schwarz inequality we have
Z
p
1
1
√
kµ ∗ G − ν ∗ GkT V =
1 + x2 |g ∗ µ(x) − g ∗ ν(x)|dx
2
2
1+x
≤ C(kg ∗ µ − g ∗ νk2 +kx(g ∗ µ − g ∗ ν)k2 ),
for some numerical constant C > 0. For all k > 0 an application of the Plancherel
identity yields
Z
1
|ϕµ (u) − ϕν (u)|2
1
2
2
kg ∗ µ − g ∗ νk2 =
|ϕG (u)|2 uk du.
kϕG (ϕµ − ϕν )k2 =
2π
2π
uk
Hence,
kg ∗ µ − g ∗ νk2 ≤
In the same way we also have
r
1
|ϕµ (u) − ϕν (u)| k/2
ku ϕG k2 .
sup
2π u∈R
|u|k/2
kx(g ∗ µ(x) − g ∗ ν(x))k22 ≤
1 ′
1
kϕ (ϕµ − ϕν )k22 + kϕG (ϕ′µ − ϕ′ν )k22
π G
π
and we conclude as before that for all r, j > 0
kx(g ∗ µ(x) − g ∗ ν(x))k2
r
r
|ϕ′µ (u) − ϕ′ν (u)| j/2
1
1
|ϕµ (u) − ϕν (u)| r/2 ′
ku ϕG k2 +
ku ϕG k2 .
sup
sup
≤
r/2
π u∈R
π u∈R
|u|
|u|j/2
It remains to apply the inverse Fourier transform.
Using a different set of hypotheses, one can also establish the following relation between the total variation distance and the Toscani-Fourier distance.
Proposition 6. Let µ, ν and G be real probability measures absolutely continuous with respect to the Lebesgue measure. Let fµ , fν and g denote their
densities and Fµ and Fν denote the cumulative distribution functions of µ and
ν. Suppose that F g ∈ L1 and that Fµ − Fν ∈ L1 . Further suppose that the
graphs of fµ ∗ g and fν ∗ g intersect in at most N points. Then,
Z
N
T1 (µ, ν) |F g(u)|du.
kµ ∗ G − ν ∗ GkT V ≤
2π
Proof. As in the proof of Proposition 4, let us introduce the function
Z
hφ (t) :=
φ(x)g(t − x)dx
R
and recall that
kµ ∗ G − ν ∗ GkT V =
1
sup
2 kφk∞ ≤1
21
Z
R
hφ (t)(µ − ν)(dt) .
Using an integration by part and Plancherel identity, we get
Z
Z
hφ (t)(µ − ν)(dt) =
h′φ (t)(Fµ (t) − Fν (t))dt
R
R
Z
1
=
F h′φ (u)F (Fµ − Fν )(u)du
2π
Z
ϕµ (u) − ϕν (u)
1
F h′φ (u)
du
=
2π
−iu
Z
1
≤ T1 (µ, ν)
|F h′φ (u)|du.
2π
(12)
Also observe that
sup
kφk∞ ≤1
Z
hφ (t)(µ − ν)(dt) =
R
where
φ̃(x) =
(
−1
1
if
if
Z
R
hφ̃ (t)(µ − ν)(dt)
fµ ∗ g(x) < fν ∗ g(x),
fµ ∗ g(x) ≥ fν ∗ g(x).
Let us denote by −∞ = x0 < x1 < · · · < xN < xN +1 = +∞ the points of
intersections between the graphs of fµ ∗ g and fν ∗ g. In particular φ̃(u) =
P
i
± N
i=0 (−1) I[xi ,xi+1 ) (u) with the sign depending on the sign of fµ ∗ g − fν ∗ g
on (−∞, x1 ). Thus,
h′φ̃ (u)
= ±2
N
X
= ±2
N
X
j=1
(−1)j g(u − xj ).
In particular, we get that
F h′φ̃ (u)
(−1)j F g(u)eiuxj ,
j=1
hence |F h′φ̃ (u)|≤ 2N |F g(u)|. This fact, together with (12), concludes the proof.
Let us observe that another way to link the total variation distance between
convolution measures to the Toscani-Fourier distance is offered by Theorem 2.21
in [2] joint with Proposition 4. More precisely, Theorem 2.21 in [2] states that,
under appropriate hypotheses on µ and ν,
1/3
18M
T2 (µ, ν)1/6 ,
W1 (µ, ν) ≤
π
with M = max {E[X 2 ], E[Y 2 ]}, X ∼ µ and Y ∼ ν. Therefore, from Proposition
4, it follows that
1/3
9M
kµ ∗ G − ν ∗ GkT V ≤ kgkBV
T2 (µ, ν)1/6 ,
4π
where g denotes the density of G. Using some ideas from the proof of Theorem
2.21 in [2] we will be able to prove the following general result.
22
Proposition 7. Let µ, ν ∈ Pj (R), j ≥ 1, and G be a measure, absolutely
continuous with respect to the Lebesgue measure. Suppose that the density g of
G is j-times weakly differentiable with jth derivative g (j) ∈ L2 . Then,
1/(2j+1)
kµ ∗ G − ν ∗ GkT V ≤ Cj
2j/(2j+1)
kg (j) k2
Tj (µ, ν)2j/(2j+1) ,
where Cj = max (E[|X + Z|j ], E[|Y + Z|j ]) with X ∼ µ, Y ∼ ν, Z ∼ G and Z
independent of X and Y .
Proof. Using the same notation as in Proposition 6, we have for all R > 0
Z
1
kµ ∗ G − ν ∗ GkT V =
|g ∗ µ(x) − g ∗ ν(x)|dx
2
Z R
Z
1
1
|g ∗ µ(x) − g ∗ ν(x)|dx + j
|x|j |g ∗ µ(x) − g ∗ ν(x)|dx
≤
2
R |x|>R
−R
Z R
Cj
1
|g ∗ µ(x) − g ∗ ν(x)|dx + j .
≤
2
R
−R
By Cauchy-Schwarz inequality,
Z R
√
|g ∗ µ(x) − g ∗ ν(x)|dx ≤ 2Rkg ∗ µ − g ∗ νk2
−R
C
2/(2j+1)
j
holds. Taking R = ( √2kg∗µ−g∗νk
)
2
kµ ∗ G − ν ∗ GkT V ≤
we get
1 1/(2j+1) √
( 2kg ∗ µ − g ∗ νk2 )2j/(2j+1) .
C
2 j
Using Plancherel identity and the properties of the Fourier transform, we deduce
that
1
1
kF (g ∗ µ) − F (g ∗ ν)k22 =
kF g(ϕµ − ϕν )k22
2π
2π
Z
|ϕµ (u) − ϕν (u)|2
1
du
|F g(u)|2 u2j
=
2π
u2j
Z
1
≤
Tj (µ, ν)2 |F g(u)|2 u2j du
2π
1 2
T (µ, ν)kF g (j) k22 = Tj2 (µ, ν)kg (j) k22 .
=
2π j
kg ∗ µ − g ∗ νk22 =
It follows that
1/(2j+1)
kµ ∗ G − ν ∗ GkT V ≤ Cj
(Tj (µ, ν)kg (j) k2 )2j/(2j+1) .
Remark 5. To better understand the upper bounds presented above, let us
specialise to the case G = N (0, σ 2 ). In order to compare the results presented
in Propositions 5–7 let us start by observing that the following equalities hold.
R
x2
u2 σ2
1
• If g(x) = √2πσ
e− 2σ2 , then F g(u) = e− 2 . Therefore, |F g(u)|du =
√
2π
σ .
23
that kg ′ k22 =
q
2
σ2 π .
2
2
x
x
2
√−x e− 2σ2 and g ′′ (x) = √ 1
e− 2σ2 (−1 + σx2 ). It follows
2πσ3
2πσ3
R∞
2
√1
, kg ′′ k22 = πσ1 5 0 e−y (y 2 −1)2 dy = 4√1πσ5 and kgkBV =
4 πσ3
• Also, g ′ (x) =
We are now able to compare the previous results for independent random variables Z ∼ N (0, σ 2 ), X ∼ µ, Y ∼ ν.
Proposition 5 for T1 : With a numerical constant C > 0, independent of the
laws of X, Y, Z,
1
1
kL (X + Z) − L (Y + Z)kT V ≤ C T1 (X, Y ) √ +
σ
σ3
|ϕ′X (u) − ϕ′Y (u)|
1
.
+ √ sup
|u|
σ 3 u∈R
Proposition 6: Let N be the number of intersections between the graphs of the
densities of X + Z and Y + Z. Then,
N T1 (X, Y )
.
kL (X + Z) − L (Y + Z)kT V ≤ √
σ
2π
Proposition 7 for T1 :
kL (X + Z) − L (Y + Z)kT V
max (E[|X + Z|], E[|Y + Z|]) 1/3 (T (X, Y ))2/3
1
√
≤
.
σ
4 π
Proposition 7 for T2 :
kL (X + Z) − L (Y + Z)kT V
1/5
(T2 (X, Y ))4/5
max (E[(X + Z)2 ], E[(Y + Z)2 ])
.
≤
16π
σ2
Proposition 4 + Theorem 2.21 in [2]:
kL (X + Z) − L (Y + Z)kT V ≤
9 max (E[X 2 ], E[Y 2 ])
√
2π 5
1/3
(T2 (X, Y ))1/6
.
σ
We see that Proposition 7 gives a much tighter bound than Proposition 4 +
Theorem 2.21 in [2] when T2 (X, Y ) is small.
4.4
Main total variation results
As it was the case in Section 3, in order to obtain an upper bound for the total
variation distance between the marginals Xt1 and Xt2 of two Lévy processes it
is enough to separetely control the total variation distance between Gaussian
distributions, between the small jumps and the corresponding Gaussian component and finally between the big jumps. The latter can be controlled by means
of the following result.
24
Theorem 13. Let (Xi )i≥1 and (Yi )i≥1 be sequences of i.i.d. random variables
a.s. different from zero and N , N ′ be two Poisson random variables with N
(resp. N ′ ) independent of (Xi )i≥1 (resp. (Yi )i≥1 ). Denote by λ (resp. λ′ ) the
mean of N (resp. N ′ ). Then,
L
N
X
i=1
Xi − L
N′
X
Yi
i=1
′
TV
≤ (λ ∧ λ′ )kL (X1 ) − L (Y1 )kT V +1 − e−|λ−λ | .
Proof. Without loss of generality, let us suppose that λ ≥ λ′ and write λ =
α + λ′ , α ≥ 0. By triangle inequality,
L
N
X
i=1
N′
X
Yi
Xi − L
i=1
TV
≤ L
N
X
i=1
+ L
N ′′
X
Xi
Xi − L
i=1
N ′′
X
i=1
TV
N′
X
Yi
Xi − L
i=1
TV
,
(13)
where N ′′ is a random variable independent of (Xi )i≥1 and with the same law as
N ′ . The first addendum in (13) can be bounded as follows. Let P be a Poisson
random variable independent of N ′′ and (Xi )i≥1 with mean α. Then,
L
N
X
i=1
N ′′
X
Xi
Xi − L
TV
i=1
= L
′′
+P
NX
i=1
≤ δ0 − L
N′
X
Xi
Xi − L
P
X
i=1
Xi
i=1
TV
TV
where the last bound follows by subadditivity of the total variation distance.
By definition, it is easy to see that
δ0 − L
P
X
i=1
Xi
TV
=P
P
X
i=1
Xi 6= 0 ≤ 1 − e−α .
In order to bound the second addendum in (13) we condition on N ′ and use
again the subadditivity of the total variation joined with the fact that L (N ′ ) =
L (N ′′ ):
L
N ′′
X
i=1
Xi − L
N′
X
i=1
Yi
TV
=
X
L
i=1
n≥0
≤
X
n≥0
n
X
Xi − L
n
X
i=1
Yi
TV
P(N ′ = n)
nkL (X1 ) − L (Y1 )kT V P(N ′ = n)
= λ′ kL (X1 ) − L (Y1 )kT V .
The treatment of the small jumps is the subject of the following result:
25
Proposition 8. Let X be a pure jump Lévy process with Lévy measure ν. Introduce νε = νI|x|≤ε . Then, for all Σ > 0 and ε ∈ (0, 1], we have
(0,Σ,νε )
Pt
−
(0,
Pt
√
Σ2 +σ̄2 (ε),0)
TV
r
2
(0,0,νε )
(0,σ̄(ε),0)
P
,
P
W
1
t
t
πtΣ2
r
p
2
ε
≤
.
min 2 tσ̄ 2 (ε),
2
πtΣ
2
≤
Proof. This follows by applying first Proposition 4 and then Theorem 9.
As a consequence of the above estimates on the Wasserstein distances, we
obtain a bound for the total variation distance of the marginals of Lévy processes
with non-zero Gaussian components.
Theorem 14. With the same notation used in Theorem 11 and Section 2.4,
for all t > 0, ε ∈ (0, 1] and for all σi > 0, i = 1, 2, we have:
(b1 ,σ1 ,ν1 )
Pt
(b ,σ2 ,ν2 )
− Pt 2
q
TV
p
√ p
b1 (ε) − b2 (ε) + 2 σ12 + σ̄12 (ε) − σ22 + σ̄22 (ε)
p
p
≤
σ12 + σ̄12 (ε) ∨ σ22 + σ̄22 (ε)
s
2
q
X
2
ε
min 2 tσ̄i2 (ε),
+
2
πtσi
2
i=1
t
2π
ν1ε
νε
− 2 kT V ,
λ1 (ε) λ2 (ε)
+ t|λ1 (ε) − λ2 (ε)| + t(λ1 (ε) ∧ λ2 (ε))k
with νjε = νj (· ∩ (R \ (−ε, ε))) and λj (ε) = νjε (R).
Proof. By subadditivity of the total variation distance and by the triangle inequality,
(b ,σ ,ν )
Pt 1 1 1
−
(b ,σ ,ν )
Pt 2 2 2
TV
≤
2
X
(0,σ ,ν (ε))
Pt i i
i=1
(b1 (ε),
+ Pt
√
−
(0,
Pt
σ12 +σ̄12 (ε),0)
√
σi2 +σ̄i2 (ε),0)
TV
(b2 (ε),
− Pt
√
σ22 +σ̄22 (ε),0)
TV
+ kL (Xt1,B (ε)) − L (Xt2,B (ε))kT V .
The proof follows from Proposition 8, the classical bound
√
√1 |µ1 − µ2 |+ 2|σ1 − σ2 |
.
kN (µ1 , σ12 ) − N (µ2 , σ22 )kT V ≤ 2π
σ1 ∨ σ2
and Theorem 13.
Another useful result follows directly from Proposition 7 with j = 1 and
allows to bound the total variation distance for Lévy processes with positive
Gaussian part by the Toscani-Fourier distance for the same Lévy processes, but
with a smaller Gaussian part.
26
Theorem 15. Let X i ∼ (bi , σi2 , νi ), with σi > 0, i = 1, 2, be two Lévy processes.
e i ∼ (bi , σ 2 − Σ2 , νi ),
For any Σ ∈ (0, σ1 ∧ σ2 ) consider the Lévy processes X
i
i = 1, 2. Then,
kL (Xt1 ) − L (Xt2 )kT V ≤
5
5.1
e 1, X
e 2 ))2/3
max (E[|Xt1 |], E[|Xt2 |])1/3 (T1 (X
t
t
√
.
(16π)1/6 Σ t
A statistical application
Lower bounds in the minimax sense
One of the main goals in statistics is to estimate a quantity of interest from
the data. There are different criteria that can be used to judge the quality
of an estimator. In nonparametric statistics it is common to use a minimax
approach. Let us recall the classical setting. From the data (X1 , . . . , Xn ) one
wants to recover a quantity of interest θ (e.g. θ is the density of the observations,
or the regression function, or the Léyy density, or the diffusion coefficient, etc.).
In practice θ is unknown (but supposed to belong to a certain parameter space
Θ) and one needs to estimate it via an estimator (a measurable function of
the data) θ̂n = θ̂n (X1 , . . . , Xn ). To measure the accuracy of the estimator one
computes the minimax risk
R∗n := inf sup E[d2 (θ, Tn )],
Tn θ∈Θ
where the infimum is taken over all possible estimators Tn of θ and d is a
semi-distance on Θ. Furthermore, one says that a positive sequence (ψn )n≥1 is
an optimal rate of convergence of estimators on (Θ, d) if there exist constants
C < ∞ and c > 0 such that
lim sup ψn−2 R∗n ≤ C
(upper bound)
(14)
lim inf ψn−2 R∗n ≥ c,
(lower bound).
(15)
n→∞
and
n→∞
The goal is then to construct an estimator θn∗ such that
sup E[d2 (θn∗ , θ)] ≤ C ′ ψn2
θ∈Θ
where (ψn )n≥1 is the optimal rate of convergence and C ′ < ∞ is a constant.
The usual way to proceed is to build an estimator θ̂n of θ and start the
investigation about its performance firstly via an upper bound like (14). This
is important since the first thing to check is that the considered estimator is at
least consistent, that is automatically implied if supθ∈Θ E[d2 (θ, θ̂n )] → 0. After
that, a natural question is whether one could construct a better (in terms of rate
of convergence in the class (Θ, d)) estimator. In order to ensure that it is not
possible to obtain a better estimator than the one already constructed one has
to prove a lower bound, that is it is needed to prove that the rate of convergence
of any other possible estimator of θ will not be faster than the rate obtained in
the upper bound. This is in general a difficult task and we refer to Chapter 2
in [23] for general techniques to prove lower bounds. Without recalling all the
27
steps needed to prove a lower bound following [23], let us stress here that one of
the fundamental ingredients is to have a fine upper bound for the total variation
distance or other measure distances. To that end the estimates in Section 4.4
can be of general interest to prove lower bounds in the minimax sense.
One situation when this general procedure applies is the following, where we
show how to simplify the arguments used in [12] in order to prove the desired
lower bound for an estimator of the integrated volatility.
5.2
How to simplify the proof of the lower bound in [12]
In [12], the authors consider a one-dimensional Itô-semimartingale X with characteristics (B, C, ν):
Bt =
Z
t
bs ds,
Ct =
0
Z
t
cs ds,
ν(dt, dx) = dtFt (dx).
0
r
They assume that X belongs to the class SA
of all Itô-semimartingales that
satisfy
Z
|bt |+ct + (|x|r ∧1)Ft (dx) ≤ A ∀t ∈ [0, 1].
Their goal is to estimate the integrated volatility C at time 1, C(X)1 , from
high-frequency observations X i , i = 0, . . . , n. They have an upper bound for
n
an estimator of C(X)1 and they want to prove that the rate of convergence
attained by that estimator is optimal. To that aim they need to prove that any
uniform rate ψn for estimating C(X)1 satisfies
ψn ≥ (n log n)−
2−r
2
if r > 1.
(16)
Following [23], their strategy consists in finding two Lévy processes X i ∼
(bi , σi2 , Fi ), i = 1, 2, such that
1. σ12 − σ22 = an := (n log n)−
R
2. (|x|r ∧1)Fi (dx) ≤ K,
2−r
2
, r ∈ (0, 2),
1
2
3. kL ((Xi/n
)1≤i≤n ) − L ((Xi/n
)1≤i≤n )kT V → 0 as n → ∞.
The construction in [12] of the Lévy processes as well as the proof of the convergence in total variation stated in Point 3. above is very involved. Let us now
see how to use Theorem 15 to prove (16) more easily. To that aim consider two
sequences of Lévy processes X 1,n ∼ (0, 1 + an , Fn1 ) and X 2,n ∼ (0, 1, Fn2 ) with
Lévy measures Fn1 and Fn2 satisfying the following conditions:
R
• R (|x|r ∧1)Fni (dx) ≤ K, i = 1, 2.
R
• Define Ψi,n := R (eiux − 1 − iuxI|x|≤1 )Fni (dx), i = 1, 2. Then Ψ1,n and
Ψ2,n are real positive functions such that
Ψ2,n (u) =
an
+ Ψ1,n (u),
2
28
∀|u|< un := 2
p
n log n.
(17)
It is not difficult to see that Lévy measures Fn1 and Fn2 satisfying such conditions
always exist. In particular, it follows from (17) that the X i,n , i = 1, 2, have the
same characteristic function for all |u|< un , i.e.:
i
u2
h
1,n
Ψ1,n (u)
,
(1 + an ) −
E eiuX1/n = exp −
2n
n
i
h
2
2,n
u
Ψ2,n (u)
E eiuX1/n = exp −
.
−
2n
n
1,n
2,n
In order to apply Theorem 15 let us observe that X1/n
(resp. X1/n
) is equal
1
e 1,n
in law to the convolution between a Gaussian distribution N (0, ) and X
1/n
8n
e 2,n ∼ (0, 7 , F 2 )). We obtain
e 2,n ), where X
e 1,n ∼ (0, 7 + an , F 1 ) (resp. X
(resp. X
n
n
1/n
8
8
1,n
2,n
kL (X1/n
) − L (X1/n
)kT V ≤
32 1/6
π
(Cn n3/4 T1 (µn , νn ))
2/3
,
1,n
2,n
where Cn2 = max (E[|X1/n
|], E[|X1/n
|]). We are therefore left to compute T1 (µn , νn )
1,n
2,n
and show that nkL (X1/n
) − L (X1/n
)kT V → 0.
T1 (µn , νn ) = sup
exp ( −
u2 7
2n ( 8
+ an ) −
Ψ1,n (u)
)
n
u
u∈R
h
7u
exp ( − 16n
) exp ( −
2
= sup
|u|>un
≤
exp ( −
un
7u2n
16n )
=
2
u an
2n
−
Ψ1,n (u)
)
n
− exp ( −
− exp ( −
u
7u2
16n
−
Ψ2,n (u)
)
n
Ψ2,n (u)
)
n
i
log n
)
exp ( − 7×4n
n−9/4
16n
√
= √
.
2 n log n
2 log n
Hence,
kµn ∗ Gn − νn ∗ Gn kT V ≤
2/3
2/3
32 1/6 Cn n−1
n−9/4
=
Cn n3/4 √
.
π
(log n)1/3
log n
Therefore,
q
1,n
2,n
nkL (X1/n
) − L (X1/n
)kT V
32 1/12 C 1/3
n
√
→ 0,
≤
π
log n
1,n
2,n
kL ((Xi/n
)1≤i≤n ) − L ((Xi/n
)1≤i≤n )kT V ≤
as desired.
References
[1] Y. Aït-Sahalia and J. Jacod: Testing for jumps in a discretely observed
process. The Annals of Statistics, 37, (2009), 184–222.
[2] J. A. Carrillo and G. Toscani: Contractive probability metrics and asymptotic behavior of dissipative kinetic equations. Riv. Mat. Univ. Parma, 6,
(2007), 75–198.
29
[3] R. Cont and P. Tankov: Financial modelling with jump processes. Chapman
& Hall/CRC Financial Mathematics Series, Chapman & Hall/CRC, Boca
Raton, FL, 2004.
[4] C. G. Esseen: On mean central limit theorems. Kungl. Tekn. Högsk. Handl.
Stockholm, 121, (1958), 31.
[5] P. Étoré and E. Mariucci: L1-distance for additive processes with timehomogeneous Lévy measures. Electronic Communications in Probability,
19, (2014), 1–10.
[6] N. Fournier: Simulation and approximation of Lévy-driven stochastic differential equations. ESAIM Probab. Stat., 15, (2011), 233–248.
[7] G. Gabetta, G. Toscani, and B. Wennberg: Metrics for probability distributions and the trend to equilibrium for solutions of the boltzmann equation.
Journal of statistical physics, 81, (1995), 901–934.
[8] J. Gairing, M. Högele, T. Kosenkova, and A. Kulik: Coupling distances between Lévy measures and applications to noise sensitivity of SDE. Stochastics and Dynamics, 15, (2015).
[9] A. L. Gibbs and F. E. Su: On choosing and bounding probability metrics.
International statistical review, 70, (2002), 419–435.
[10] C. R. Givens and R. M. Shortt: A class of Wasserstein metrics for probability distributions. Michigan Math. J., 31, (1984), 231–240.
[11] B. V. Gnedenko and A. N. Kolmogorov: Limit distributions for sums of
independent random variables. English edition, Addison-Wesley, 1954.
[12] J. Jacod and M. Reiß: A remark on the rates of convergence for integrated
volatility estimation in the presence of jumps. The Annals of Statistics, 42,
(2014), 1131–1144.
[13] J. Jacod and A. N. Shiryaev: Limit theorems for stochastic processes,
Springer-Verlag Berlin, volume 288, 1987.
[14] Y. A. Kutoyants: Statistical Inference for Spatial Poisson Processes, Lecture
Notes in Statistics 134, Springer, New York, volume 134, 2012.
[15] F. Liese: Estimates of Hellinger integrals of infinitely divisible distributions.
Kybernetika, 23, (1987), 227–238.
[16] J. Mémin and A. N. Shiryayev: Distance de Hellinger-Kakutani des lois correspondant à deux processus à accroissements indépendants. Z. Wahrsch.
Verw. Gebiete, 70, (1985), 67–89.
[17] M. H. Neumann and M. Reiß: Nonparametric estimation for Lévy processes
from low-frequency observations. Bernoulli, 15, (2009), 223-248.
[18] V. V. Petrov: Sums of independent random variables. Springer-Verlag,
Berlin Heidelberg New York, 1975.
[19] E. Rio: Upper bounds for minimal distances in the central limit theorem.
Ann. Inst. Henri Poincaré Probab. Stat., 45, (2009), 802–817.
30
[20] L. Rüschendorf and J. Woerner: Expansion of transition distributions of
Lévy processes in small time. Bernoulli, 8, (2002), 81–96.
[21] A. I. Sakhanenko: Estimates in an invariance principle. Limit theorems of
probability theory, Trudy Inst. Mat., 5, 1985, 27–44.
[22] K. I. Sato: Lévy processes and infinitely divisible distributions. Cambridge Studies in Advanced Mathematics, Cambridge University Press,
Cambridge, volume 68, 1999.
[23] A. B. Tsybakov: Introduction to nonparametric estimation. Springer Series
in Statistics, Springer, New York, 2009.
[24] C. Villani: Optimal transport: Old and New. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences],
Springer-Verlag, Berlin, volume 338, 2009.
[25] V. M. Zolotarev: Modern theory of summation of random variables. Modern
Probability and Statistics, VSP, Utrecht, 1997.
31
| 10math.ST
|
Device-to-Device Aided Multicasting
Thomas Varela Santana∗† , Richard Combes† and Mari Kobayashi†
∗ Orange
arXiv:1801.05224v1 [cs.IT] 16 Jan 2018
† Laboratoire
Labs Networks, Châtillon, France, Email: [email protected]
des Signaux et Systèmes (L2S), CentraleSupélec-CNRS-Université Paris-Sud, Gif-sur-Yvette, France,
Email: {richard.combes, mari.kobayashi}@centralesupelec.fr
Abstract—We consider a device-to-device (D2D) aided multicast channel, where a transmitter wishes to convey a common
message to many receivers and these receivers cooperate with
each other. We propose a simple computationally efficient scheme
requiring only statistical channel knowledge at transmitter. Our
analysis in general topologies reveals that, when the number of
receivers K grows to infinity, the proposed scheme guarantees
a multicast rate of 21 log2 (1 + β ln K) with high probability for
any β < β ⋆ where β ⋆ depends on the network topology. This
scheme undergoes a phase transition at threshold β ⋆ ln K where
transmissions are successful/unsuccessful with high probability
when the SNR is above/below this threshold. We also analyze the
outage rate of the proposed scheme in the same setting.
Index Terms—Device-to-Device, Multicasting.
I. I NTRODUCTION
We consider the multicast channel, where a single transmitter sends a common message to many receivers in the
presence of fading [1]–[3]. Since the common message must
be decoded by all receivers, the multicast capacity of the
fading broadcast channel is limited by the worst user. For
the case of the i.i.d. Rayleigh fading channel, it is known
that the multicast capacity vanishes inversely to the number
K of users as K grows [4]. Despite its vanishing rate, the
multicast channel is relevant to two scenarios in wireless
crowded networks. The first scenario is the wireless edge
caching. It has been shown that the traffic during the peak
hours can be significantly reduced by caching popular contents
during off-peak hours, and delivering these popular contents
using multicasting (see e.g. [5], [6] and references therein). A
novel user selection scheme requiring only statistical channel
knowledge has been proposed recently in [7] for the case of
i.i.d. Rayleigh fading channel. The second scenario is so-called
enhanced Multimedia Broadcast Multicast Service (eMBMS)
aided by Proximity Service (ProSe), standardized by the 3rd
Generation Partnership Project (3GPP). These services reduce
the network load by multicasting common data to public
safety devices [8] which can use Device-to-Device (D2D)
communication [9]. In this work, motivated by these two
scenarios, we study D2D-aided multicasting without Channel
State Information at the Transmitter (CSIT) to overcome the
vanishing effect of the multicast rate. We consider a general
network topology where the transmitter has only statistical
channel knowledge. In this setup, we address the fundamental
question: can D2D without CSIT increase the achievable
multicasting rate ?
We propose a two-stage transmission scheme assuming
no CSIT: in the first stage the transmitter sends a common
message, while in the second stage, the subset of users
whom have successfully decoded retransmit this information
simultaneously by using the same codeword. We study this
two-stage scheme for two metrics of interest:
• The average multicast rate, which is the expected number
of successfully decoded bits per channel use achieved by
a user chosen uniformly at random;
• The outage rate, which is the maximum rate at which
all the users can decode the message with an error
probability of ǫ.
We show that by carefully choosing the transmission rate, we
can answer positively to our main question. More specifically,
our contributions is two-fold and summarized below:
1) We identify conditions on the network topology for which
the average multicast rate grows with the number of users
K and the outage rate does not vanish with K.
2) We provide tractable, asymptotic expressions for both the
multicast rate and the outage rate in the regime of a large
number of users.
A similar two-phase scheme has been studied in [10]. However
this work is different from ours in its assumption and concept.
First, the channel statistic of all users is assumed to be
symmetric in [10]. In fact, this is a special case of our
model corresponding to a single class C = 1 (see subsection
II-D). Second, the metric of the work [10] is on the error
probability decay by exploiting channel hardening via spacetime code, while we aim at achieving a scalable multicast rate
by exploiting multiuser diversity.
The rest of the paper is organized as follows. Section II
presents the system model. Sections III and IV studies the
multicast and outage rate metrics respectively. Numerical
experiments are provided in Section V. Section VI concludes
the paper. We denote by ln the natural logarithm and log2 the
binary logarithm.
II. M ODEL
A. Channel Model
We consider a multicast channel, where a station (indexed
by 0) wants to convey the same message to K users indexed
by i = 1, . . . , K. Time is slotted, and in each time slot, yi the
received signal by i = 1, . . . , K is given by:
yi =
K
X
j=0
xj hj,i + ni ,
with (xi )i=0,...,K the signal transmitted by i = 0, . . . , K, h =
(hi,j )i,j=0,...,K are the channel coefficients and ni ∼ N (0, 1)
is Additive White Gaussian Noise (AWGN). The channel coefficients h are assumed to be independent, Gaussian complex
random variables with mean 0 and variance γi,j = E(|hi,j |2 ).
The transmitted signals have unit power E(|xi |2 ) ≤ 1. This
is without loss of generality,
otherwise simply replace xi by
p
√ xi 2 and hi,j by E(|xi |2 )hi,j .
E(|xi | )
Matrix Γ = (γi,j )i,j=0,...,K represents the channel statistics,
and captures the topology of the network. Channel coefficients
during successive time slots are assumed to be independent.
No channel state information is available at the transmitter:
the transmission strategy should only depend on the statistics
of the channel (i.e. Γ), but not on the channel coefficients h.
B. Proposed scheme
Throughout the article we will study the following transmission scheme, which utilizes two time slots. Let h, h′ , the
channel coefficients during the two successive time slots.
• During the first time slot, the station broadcasts a message
at rate log2 (1 + s), where s is a parameter of the scheme
and can be chosen depending on the channel statistics Γ.
• At the end of the first time slot, users attempt to decode,
and user i decodes successfully if and only if the received
Signal to Noise Ratio (SNR) is greater than s, i.e.
|h0,i |2 ≥ s (1 )
• All users that have successfully decoded the message
in the first slot retransmit this message in the second
slot. As before, user j decodes successfully at the end
ofPthe second time slot if the SNR is above s, i.e.
K
| j=1 Zj (s)h′j,i |2 ≥ s, where Z1 (s), . . . , ZK (s) are
binary variables with Zj (s) = 1 if j has decoded
successfully during the first time slot and 0 otherwise.
C. Performance Measures
We say that user i decodes if and only if he successfully
decodes either in the first or second slot. Denote by
i (s) the
PP
K
1
probability that i = 1, . . . , K decodes, P̄ (s) = K
i=1 Pi (s)
the probability that a user chosen uniformly at random
amongst i = 1, . . . , K decodes, and P+ (s) probability that
all users decode. We study two performance measures for our
problems: the multicast rate:
Rm =
1
max{log2 (1 + s)P̄ (s)},
2 s≥0
which is the expected number of bits received by a user chosen
uniformly at random per time slot, and the outage rate:
1
log2 (1 + s) with s solution to P+ (s) = 1 − ǫ.
2
which is the largest rate such that all users decode with
probability at least 1 − ǫ, with ǫ some fixed reliability level.
Both performance measures are interesting for different scenarios: Rm seems more appropriate when considering say
video streaming where the goal is that on average most users
receive enough information, whereas Ro seems more suited to
applications such as broadcasting of safety information, where
it is important that all users obtain the message.
D. The Block Model
To study large systems, we introduce a block model for
Γ. Users 1, . . . , K are partitioned in classes 1, . . . , C, where
ci ∈ {1, . . . , C} indicates the class of user i, and c0 = 0 by
convention. There are Kαc users of
c, with αc > 0 the
Pclass
C
proportion of users of class c and c=1 αc = 1. Matrix Γ is
a block matrix, so that the mean channel gains between two
users depends solely on their class. Namely, γi,j = gci ,cj for
all i, j = 0, . . . , K. Define G = maxc,c′ =0,...,C gc,c′ the largest
entry of Γ, so that maxi,j=0,...,K γi,j = G. This model can
represent any network topology, as the number of classes may
be arbitrary. We introduce the following natural assumption.
Assumption 1: Any class of users can be reached in two
transmissions, so that for all c = 1, . . . , C there exists c′ such
that g0,c′ gc′ ,c > 0.
E. Baseline
We will compare the performance of this scheme to the
most natural baseline which is the same scheme where only
the first time slot is used. The performance of the baseline is
recalled below.
Proposition 1: The performance of the baseline scheme is:
Rm = max{log2 (1 + s)
s≥0
C
X
αc e
c=1
−gs
0,c
}
=
K→∞
C
1
1 1 X αc −1
.
= O
ln
Ro = log2 1 +
K→∞
K
1−ǫ
g
K
c=1 0,c
Proof. In the baseline scheme, user i successfully decodes if
and only if |h0,i |2 ≥ s. Since |h0,i |2 follows an exponential
distribution with mean γ0,i , we have:
Pi (s) = P(|h0,i |2 ≥ s) = e
− γs
0,i
,
and averaging:
P̄ (s) =
C
K
X
s
1 X − γ0,i
− s
=
αc e g0,c .
e
K i=1
c=1
which is the announced result.
Once again i decodes successfully if and only if |h0,i |2 ≥ s,
|h0,i |2 follows an exponential distribution with mean γ0,i and
h0,1 , . . . , h0,K are independent so that:
Ro =
1 This is valid if we assume that time slots are long enough so that one may
use an efficient code for the AWGN channel.
O(1),
P+ (s) = P(|h0,i |2 ≥ s , i = 1, . . . , K) =
K
X
1
= exp −s
γ
i=1 0,i
(
)
K
Y
i=1
P(|h0,i |2 ≥ s)
C
X
αc
= exp −sK
g
c=1 0,c
(
)
.
The outage rate is log2 (1 + s) with s solution to the equation
P+ (s) = 1 − ǫ, so replacing yields the announced result.
III. M ULTICAST R ATE
•
In this section we study the multicast rate of the proposed
scheme, and show that it undergoes a phase transition in the
regime of a large number of users K → ∞, so that the
probability of success P̄ (s) becomes constant-by-parts and
may be computed explicitly as a function of g and α.
A. Success probability
•
We first prove Proposition 2, a formula for the probability
of success Pi (s), which will serve as the backbone of our
analysis. This result shows that the success probability Pi (s)
can be controlled by examining the fluctuations of Xi (s), a
sum of independent Bernoulli random variables.
Proposition 2: For any i = 1, . . . , K and s ≥ 0 we have:
s
− s
,
Pi (s) = 1 − (1 − e γ0,i )E 1 − exp −
Xi (s)
where Z1 (s), . . . , ZK (s) aresindependent random variables in
PK
−
{0, 1} with E(Zi (s)) = e γ0,i and Xi (s) = j=1 Zj (s)γj,i .
′
(h′i,j )i,j=0,...,K
Proof. Consider h = (hi,j )i,j=0,...,K and h =
the channel coefficients during the first and second time slot
respectively. By assumption h is independent from h′ . Denote
by Z(s) = (Z1 (s), . . . , ZK (s)) the outcome of the first time
slot, where Zi (s) = 1 if user i decodes correctly at the first
time slot. We may write Zi (s) = 1{|h0,i |2 ≥ s}. Since
|h0,i |2 has exponential
distribution with mean γ0,i , it follows
− γs
0,i
. Furthermore, since h0,1 , . . . , h0,K are
that E(Zi (s)) = e
independent, so are Z1 (s), . . . , ZK (s).
Consider user i. Conditionally to the value of Z(s), user i
does not decode successfully in the first phase if and only if
Zi (s) = 0. If Zi (s) = 0, he does not P
decode successfully
in the second phase if and only if | j6=i h′j,i Zj (s)|2 =
P
PK
′
2
| K
hj,i Zj (s)|2 has expoj=1 hj,i Zj (s)| ≤ s, where |
PK j=1
nential distribution with mean j=1 γj,i Zj (s) = Xi (s), since
Z(s) is independent of h′ . Hence
P(i does not decode|Z(s)) = (1 − Zi (s))(1 − e
Taking expectations over Z(s) yields the result
− X s(s)
i
).
.
B. Asymptotic behavior
We now analyze how the multicast rate scales in the regime
of a large number of users. Define:
βc =
•
max {g0,c′ 1{gc′ ,c > 0}},
c′ =1,...,C
the largest value of g0,c′ , the mean channel gain from the
station to class c′ , where class c′ can communicate with
class c in the second phase i.e. gc′ ,c > 0. Further define
the minimum value β ⋆ = minc=1,...,C βc . Theorem 1 shows
that in the limit of a large number of users, for any user i,
the success probability Pi undergoes a phase transition at the
value βci ln K. Namely transmissions are always successful
above this threshold, and always unsuccessful below. This has
3 consequences:
Our scheme transmits at rate 12 log2 (1 + β ln K) with an
arbitrarily high probability of success for any β < β ⋆ .
As K → ∞, the multicast rate of our scheme scales
as O(ln ln K) while the baseline yields a multicast rate
of O(1). So considering two slots instead of one has a
dramatic impact on performance. Considering more than
two time slots does not seem to improve this scaling.
To obtain an order-optimal rate, one can set s = β ln K
for any β < β ⋆ , so that optimizing over s is not needed.
Theorem 1: (i) For any β > 0 and i = 1, . . . , K we have:
(
1 if β < βci
Pi (β ln K) →
K→∞
0 otherwise.
(ii) We have:
C
X
P̄ (β ln K) →
K→∞
c=1
αc 1{β < βc },
Proof. From Proposition 2, we have:
Pi (β ln K) = 1 − (1 − K
We have K
β
0,ci
−g
β
0,ci
−g
)E(1 − e
− X β(βlnlnKK)
i
).
→ 0 so that:
K→∞
lim Pi (β ln K) = lim E(e
K→∞
− X β(βlnlnKK)
i
K→∞
).
The β > βci case. Assume that β > βci , and we control the
expectation of Xi (β ln K). Since E(Zj (s)) = e
− γβ
0,j
s
0,cj
−g
:
β
1−
C
X
αc gc,ci K g0,c
=
β ln K
c=1
K
X (β ln K) X
γj,i K
i
=
E
β ln K
β ln K
j=1
β
1−
C
X
αc gc,ci K βci
≤
β ln K
c=1
→ 0
K→∞
since for all c either gc,ci = 0 or g0,c ≤ βci , and β > βci .
Therefore, Xiβ(βlnlnKK) converges to 0 in L1 so that it converges
1
to 0 in distribution as well. Since x 7→ e− x is both continuous
and bounded we get:
lim Pi (β ln K) = lim E(e
K→∞
− X β(βlnlnKK)
i
K→∞
) = 0.
The β < βci case. Now consider β < βci . We control the
moments of Xj (s):
E(Xi (s)) =
K
X
γj,i e
−γs
0,j
,
j=1
and since Z1 (s), . . . , ZK (s) are independent:
var(Xi (s)) =
K
X
2
γj,i
e
−γs
j=1
≤G
K
X
j=1
γj,i e
0,j
(1 − e
−γs
0,j
−γs
0,j
)
= GE(Xi (s)).
1
Apply Chebychev’s inequality:
Proof. Define f (x) = 1−e− x . Throughout the proof consider
Xi (β ln K)
E(Xi (s))
E(Xi (s)) i fixed, and define V = β ln K .
P Xi (s) ≤
≤ P |Xi (s) − E(Xi (s))| ≥
From Proposition 2:
2
2
− β
4G
4var(Xi (s))
1 − Pi (β ln K) = (1 − K γ0,i )E(f (V )) ∼ E(f (V )).
≤
≤
K→∞
E(Xi (s))2
E(Xi (s))
− γβ
since K 0,i → 0. Define:
using the previous bound. Hence
K→∞
PC
E(Xi (s))
4G
1− g β
0,c
P Xi (s) ≥
≥1−
.
(1)
c=1 αc gc,ci K
2
E(Xi (s))
→ ∞.
vK = E(V ) =
K→∞
β ln K
1
Since x 7→ e− x is increasing:
since β < β ⋆ . To complete the proof, it suffices to show that
− X β(βlnlnKK)
E(f (V ))
)
E(Xi (β ln K)) − E(X2β(βlnlnKK))
i
e
≥ P Xi (β ln K) ≥
.
2
E(e
i
f (V ) = f (V )1{V ≤ (1 − δ)vK } + f (V )1{V ≥ (1 − δ)vK }
β
E
Xi (β ln K)
β ln K
≤ 1{V ≤ (1 − δ)vK } + f ((1 − δ)vK ).
1−
C
X
αc gc,ci K g0,c
=
β ln K
c=1
1− ββc
αĉ gĉ,ci K
β ln K
≥
i
So taking expectations and using Chernoff’s inequality, which
is recalled as proposition 6 in appendix:
→ ∞.
E(f (V )) ≤ P(V ≤ (1 − δ)vK ) + f ((1 − δ)vK )
K→∞
≤ e−
Replacing in (1) we deduce
E(Xi (β ln K))
→ 1,
P Xi (β ln K) ≥
K→∞
2
and e
− E(X2β(βlnlnKK))
i
vK e−
δ2 v
3
K
lim Pi (β ln K) = lim E(e
i
K→∞
→ ∞ so that f (vK )
K→∞
∼ 1
K→∞ vK
and
→ 0 we get:
1
E(f (V ))
≤
.
f
(v
)
1
−
δ
K
K→∞
lim sup
) = 1.
Lower bound: Similarly:
which completes the proof of statement
(i). Statement (ii)
1 PK
follows from the fact that P̄ (s) = K
i=1 Pi (s).
Corollary 2: For any β < β ⋆ we have:
Rm ≥
+ f ((1 − δ)vK ).
K→∞
K→∞
K→∞
δ 2 vK
3
Now using the facts vK
→ 1 so that:
− X β(βlnlnKK)
f (vK ).
Now consider δ ∈ (0, 1) fixed and let us bound E(f (V )).
Upper bound: Since x 7→ f (x) is decreasing:
Consider ĉ such that gĉ,ci > 0 and g0,ĉ = βci . Then:
∼
K→∞
f (V ) = f (V )1{V ≤ (1 + δ)vK } + f (V )1{V ≥ (1 + δ)vK }
≥ 1{V ≤ (1 + δ)vK }f ((1 + δ)vK )
So taking expectations and using Chernoff’s inequality, which
is recalled as proposition 6 in appendix:
1 − o(1)
log2 (1 + β ln K) , K → ∞.
2
f (V ) ≥ P(V ≤ (1 + δ)vK )f ((1 + δ)vK )
≥ (1 − e−
C. Non asymptotic behavior
We may state Theorem 3, a further result which gives a
tractable, accurate approximation for 1 − Pj (β ln K) in the
regime where Pj (β ln K) → 1 i.e. whenever β < βci . Two
K→∞
main facts are worth mentioning:
• This approximation is very accurate even for modest
size systems (say K ≥ 50) as shown by our numerical
experiments (see section V).
• Due to it’s accuracy, it allows to find the optimal value
of s given g and α in a tractable manner, so that finite
size systems can be dealt with efficiently.
Theorem 3: Consider β < βc , then:
1 − Pi (β ln K) ∼ 1 − exp P
K→∞
C
c=1
Using the facts vK
e
δ2 v
− 3K
δ 2 vK
3
→
K→∞
)f ((1 + δ)vK ).
∞ so that f (vK )
1
∼
K→∞ vK
and
→ 0 we get:
K→∞
lim inf
K→∞
1
E(f (V ))
≥
.
f (vK )
1+δ
Putting it together: The above holds for any δ ∈ (0, 1). So
we have proven that δ arbitrairly small:
E(f (V ))
1
E(f (V ))
1
≤ lim inf
≤ lim sup
≤
.
K→∞ f (vK )
1+δ
f
(v
)
1
−
δ
K
K→∞
Hence we have proven
β ln K
αc gc,ci K
1− g β
0,c
E(f (V ))
.
which concludes the proof.
∼
K→∞
f (vK ).
IV. O UTAGE R ATE
We now turn to the outage rate, and we show that in the
regime of a large number of users, the outage rate may be
computed explicitly as a function of g and α. We further show
that, while the outage rate of the baseline scheme vanishes
when K → ∞ our scheme guarantees a constant outage rate.
A. Success probability
As in the multicast case, we express the outage rate as a
function of X1 (s), . . . , XK (s) in Proposition 3.
Proposition 3: For any s ≥ 0 we have:
)#
"
(
K
X
1 − Zi (s)
,
P+ (s) = E exp −s
Xi (s)
i=1
where Z1 (s), . . . , ZK (s) ares independent random variables in
−γ
0,i for i = 1, . . . , K and X (s) =
{0,
i
PK1} and E(Zi (s)) = e
j=1 Zj (s)γj,i .
Proof. We use the same notation as in the proof of Proposition 2. User iPsuccessfully decodes if and only if either
K
Zi (s) = 1 or | j=1 Zj (s)h′j,i |2 ≥ s. So i decodes if and
only if:
K
X
2
Zj (s)h′j,i
≥ s(1 − Zi (s)).
j=1
j=1
2
≥ s(1 − Zi (s)) , ∀i|Z(s)
2
j=1
(
= exp −s
K
X
1 − Zi (s)
i=1
Theorem 4: For any s ≥ 0 we have:
(
)
− s
C
X
αc (1 − e g0,c )
P+ (s) → exp −s
.
PC
−g s ′
K→∞
0,c
c=1
c′ =1 αc′ gc′ ,c e
Proof. We first prove the following fact:
K
X
1 − Zi (s)
Xi (s)
i=1
Xi (s)
≥ s(1 − Zi (s))|Z(s)
)
s
P
→
K→∞
K
X
1 − Zi (s)
PK
i=1
B. Asymptotic behavior
We now prove our main result concerning outage rate stated
in Theorem 4. The main
element is to show that, for
P proof
1−Zi (s)
any s random variable K
i=1 Xi (s) concentrates around its
expectation when the number of users K grows large. The
following facts should be noted:
• In the regime of a large number of users, the outage
rate converges to a non-zero finite value. Recall that
the baseline scheme only guarantees a vanishing outage
rate. Hence our scheme is very efficient in countering
the variability of the channel coefficients and induces a
perfect channel hardening.
E(Xi (s))
.
≤
Xi (s)
s
−
K
X
1 − Zi (s)
i=1
Xi (s)
i=1
E(Xi (s))
s
−
K
X
Zi (s) − e γ0,i
1 − Zi (s)
−
+
.
E(Xi (s))
E(Xi (s))
i=1
We now prove that both terms go to 0 in probability.
First term We have:
Xi (s)
K
−
X
1
1 − Zi (s)
1
−
≤
E(Xi (s))
X
(s)
E(X
i
i (s))
i=1
=
K
X
|Xi (s) − E(Xi (s))|
Xi (s)E(Xi (s))
i=1
Furthermore
E(Xi (s)) = K
with
m(s) =
′ 2
j=1 Zj (s)hj,i |
′
since conditional to Z(s), |
has exponential
distribution with mean Xi (s), and h has independent entries.
Averaging over Z(s) yields the result.
i=1
−
K
X
1 − e γ0,i
C
X
αc gc,ci e
c=1
.
−
K
X
1 − e γ0,i
We bound the error as follows:
i=1
K
K
X
Y
P
=
Zj (s)h′j,i
i=1
•
K
X
1 − Zi (s)
Now:
P(all decode|Z(s))
K
X
Zj (s)h′j,i
=P
The outage rate can be computed explicitly as a function
of g and α (see corollary 5) by a zero-finding method
such as bisection or Newton-Raphson.
In fact, this asymptotic result provides a very good
approximation even for systems of modest size (say
K ≥ 10), as shown by numerical experiments (see
section V).
•
min
max
c′ =1,...,C c=1,...,C
−gs
0,c
.
≥ m(s)K.
n
o
− s
αc gc,c′ e g0,c > 0.
where m(s) > 0 from assumption 1.
For i = 1, . . . , K, Xi (s) is a sum of K random variables
bounded by G, so that Hoeffding’s inequality yields:
√
1
P(|Xi (s) − E(Xi (s))| ≥ G K ln K) ≤ 2 .
K
Using a union bound over i = 1, . . . , K:
√
P( max |Xi (s) − E(Xi (s))| ≥ G K ln K)
i=1,...,K
≤
K
X
√
1
P(|Xi (s) − E(Xi (s))| ≥ G K ln K) ≤
K
i=1
→ 0.
K→∞
Therefore the event
√
A = { max |Xi (s) − E(Xi (s))| ≤ G K ln K}
i=1,...,K
Second term Compute the second moment of the second
term:
#
" K
s
− s
− s
K
2
X Z (s) − e− γ0,i
X
e γ0,i (1 − e γ0,i )
i
=
E
E(Xi (s))
E(Xi (s))2
i=1
i=1
s
−
K
X
i=1
1
1
≤
E(Xi (s))2
m(s)2 K
→ 0,
K→∞
since E(Zi (s)) = e γ0,i and Z1 (s), . . . , ZK (s) are independent. So the second term goes to 0 in probability, since L2
convergence implies convergence in probability.
Putting it together We have proven that:
K
X
1 − Zi (s)
i=1
Xi (s)
s
P
→
K→∞
=
C
X
c=1
−
K
X
1 − e γ0,i
E(Xi (s))
i=1
αc (1 − e
PC
c′ =1
−gs
0,c
αc′ gc′ ,c e
)
−g
s
.
0,c′
Since convergence in probability implies convergence in distribution and x 7→ e−x is continuous and bounded on [0, ∞)
we obtain the result:
)
(
− s
C
X
αc (1 − e g0,c )
.
P+ (s) → exp −s
PC
−g s ′
K→∞
0,c
c=1
c′ =1 αc′ gc′ ,c e
Corollary 5: When K → ∞, the outage rate converges
1
log
2 (1 + s) where s is the unique solution to:
2
s
C
X
c=1
αc (1 − e
PC
c′ =1
−gs
0,c
αc′ gc′ ,c e
)
−g
s
0,c′
= ln
0.8
0.6
0.4
0
200
1
.
1−ǫ
and for ǫ ≈ 0 we have s ≈ 0 and a Taylor development gives:
v
u
u
1
ln 1−ǫ
u
s≈u
−1
t PC α PC
c
′ gc′ ,c
α
′
c
c =1
c=1 g0,c
400
600
800
0.6
0.5
baseline
simulation
approx
0.4
0.3
0.2
0.1
0
1000
200
400
600
number of users (K)
number of users (K)
(a) Scenario (a)
(b) Scenario (b)
800
1000
0.8
0.6
0.4
0.2
baseline
simulation
approx
0
200
V. N UMERICAL E XPERIMENTS
In this section we present numerical experiments and show
that our theoretical analysis provides accurate predictions of
the system’s behavior. For each figure, three curves are presented: ”baseline” is the performance of the baseline scheme,
”simulation” is the exact performance of the proposed scheme
obtained by simulation, and ”approx” is the analytical approximation of the proposed scheme’s performance. For the
multicast rate the approximation is given by Theorem 3 and
baseline
simulation
approx
0.2
multicast rate (bits / channel use)
≤
multicast rate (bits / channel use)
→ 0.
K→∞
for the outage rate, the approximation is given by Theorem 4.
We consider three scenarios:
(a) A single class C = 1, with α = (1) , g0,1 = 46 dBm
and g1,1 = 23 dBm.
(b) Two classes C = 2, with α = (0.5, 0.5), g0,1 = 46
dBm,g0,2 = 0, g1,1 = g2,2 = 23 dBm and g1,2 = g2,1 = 13
dBm.
(c) A cell of radius 250 m, with K users drawn uniformly
at random in this area. The mean channel gains are taken
as γi,j = ρi − 128 − 36.4 log10 (di,j ) dBm where di,j is the
distance between i and j, expressed in kilometers, and ρi = 46
dBm if i = 0 and ρi = 23 dBm otherwise. For this scenario,
the results are averaged over 100 random realizations.
Scenario (a) is the homogeneous case where all users are
close to the station and to each other, scenario (b) is a case
where users of class 1 are close to the station, and users of
class 2 are far away from the station, so that in order to receive
data they require users of class 1 to act as a relay. Finally
scenario (c) represents the case where users arrive uniformly
in a cell area, which seems like an acceptable model for a real
system in high load (i.e. when there are many active users).
For all scenarios ǫ = 10−2 . For the multicast rate, figure 1
presents the multicast rate Rm as a function of the number of
users K (when s is chosen optimally), while figure 2 presents
the multicast rate as a function of s, for various values of K.
For outage rate, figure 3 presents the outage rate as a function
of the number of users K, while figure 4 presents the outage
probability P+ (s) as a function of s for system size K = 100.
multicast rate (bits / channel use)
occurs with high probability
√ when K → ∞. Consider K large
enough so that m(s)K − K ln K ≥ 0. If A occurs:
√
K
X
K K ln K
|Xi (s) − E(Xi (s))|
√
≤
Xi (s)E(Xi (s))
m(s)K(m(s)K − K ln K)
i=1
400
600
800
1000
number of users (K)
(c) Scenario (c)
Fig. 1: Multicast rate Rm versus number of users K
These numerical experiments point out three facts:
•
The proposed scheme is vastly superior to the baseline
scheme as soon as there are more than a few users, as
predicted by our asymptotic analysis.
0.3
0.2
0.1
0
0
1
2
3
4
0.6
10 0
0.5
0.4
0.3
0.2
10 -2
10 -4
0
0
5
1
2
3
4
5
SNR (linear)
10 -2
10 -1
10 -2
error probability
0.4
0.2
(b) Scenario (b), K = 102
10 -2
10 -4
baseline
simulation
approx
0
2
3
4
5
10 0
10 0
0.6
1
10 -1
SNR (linear)
(a) Scenario (a), K = 102
baseline
simulation
approx
0.8
10 0
SNR (linear)
(b) Scenario (c), K = 102
(a) Scenario (c), K = 10
0
baseline
simulation
approx
10 -1
baseline
simulation
approx
0.1
SNR (linear)
multicast rate (bits / channel use)
10 0
baseline
simulation
approx
error probability
baseline
simulation
approx
error probability
multicast rate (bits / channel use)
multicast rate (bits / channel use)
0.4
6
10 -2
SNR (linear)
10 -1
10 0
SNR (linear)
(c) Scenario (c), K = 103
baseline
simulation
approx
0.04
0.02
0
20
40
60
80
1
2
outage rate (bits / channel use)
0.06
log2 (1 + s)P̄ (s) vs SNR s
1
VI. C ONCLUSION
0.8
0.6
baseline
simulation
approx
0.4
0.2
0
20
100
40
60
80
100
number of users (K)
(a) Scenario (a)
(b) Scenario (b)
A PPENDIX
0.04
0.03
baseline
simulation
approx
0.02
0.01
0
40
60
80
100
number of users (K)
(c) Scenario (c)
Fig. 3: Outage rate Ro versus number of users K
•
We have studied a new scheme for broadcasting a message to users with receiver-to-receiver communication, without
channel state information at the transmitter. We have shown
that, in the large system limit, the performance of this scheme
can be completely characterized. For finite systems, we also
have provided tractable approximations for performance measures which are very accurate even for systems of modest
sizes.
0.05
20
•
Fig. 4: Outage rate, probability of error 1 − P+ (s) vs SNR s
× 10-3
number of users (K)
outage rate (bits / channel use)
outage rate (bits / channel use)
Fig. 2: Multicast, effective rate
(c) Scenario (c), K = 102
The proposed approximations/asymptotic expressions derived above predict the performance of the proposed
scheme very accurately, even for systems of modest size,
say K ≥ 40 for the multicast rate and K ≥ 15 for the
outage rate.
As a consequence, setting the parameter s can be done
in a simple and tractable manner to obtain good practical
performance for systems of modest size.
We recall below a few basic concentration inequalities.
Proofs may be found, for instance, in [11].
Proposition 4: [Chebychev’s inequality] Consider Y a random variable. Then for all δ > 0 we have:
P(|Y − E(Y )| ≥ δ) ≤
var(Y )
.
δ2
Proposition 5: [Hoeffding’s inequality] Consider
Y1 , . . . , YK independent random variables in [0, 1]. Then for
all δ > 0 we have:
!
K
2
1 X
P
(Yi − E(Yi )) ≥ δ ≤ e−2Kδ .
K i=1
Proposition 6: [Chernoff’s inequality] Consider Y1 , . . . , YK
independent
random variables in [0, 1]. Define yK =
PK
E(Y
).
Then
for all δ ∈ (0, 1) we have:
i
i=1
!
K
X
yK δ2
Yi ≥ (1 + δ)yK ≤ e− 3 .
P
i=1
and
P
K
X
i=1
Yi ≤ (1 − δ)yK
!
≤ e−
yK δ2
3
.
R EFERENCES
[1] M. J. Lopez, “Multiplexing, scheduling, and multicasting strategies for
antenna arrays in wireless networks,” MIT, Tech. Rep., 2004.
[2] P. K. Gopala and H. El Gamal, “On the throughput-delay tradeoff in
cellular multicast,” in IEEE WNCMC, 2005.
[3] N. D. Sidiropoulos, T. N. Davidson, and Z.-Q. Luo, “Transmit beamforming for physical-layer multicasting,” IEEE Trans. on Signal Processing, vol. 54, no. 6, pp. 2239–2251, 2006.
[4] N. Jindal and Z.-Q. Luo, “Capacity limits of multiple antenna multicast,”
in Proc. of ISIT, 2006.
[5] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,”
IEEE Trans. on Information Theory, vol. 60, no. 5, pp. 2856–2867,
2014.
[6] G. Paschos, E. Bastug, I. Land, G. Caire, and M. Debbah, “Wireless
caching: Technical misconceptions and business barriers,” IEEE Communications Magazine, vol. 54, no. 8, pp. 16–22, 2016.
[7] K.-H. Ngo, S. Yang, and M. Kobayashi, “Scalable content delivery
with coded caching in multi-antenna fading channels,” IEEE Trans. on
Wireless Communications, vol. 17, no. 1, pp. 548–562, 2018.
[8] Nokia, “LTE networks for public safety services, white paper, 2016.”
[9] 3GPP, “3rd generation partnership project: Technical specification group
radio access network: Study on lte device to device proximity services;
radio aspects (release 12),” 3GPP, March 2014.
[10] A. Khisti, U. Erez, and G. W. Wornell, “Fundamental limits and scaling
behavior of cooperative multicasting in wireless networks,” IEEE Trans.
on Information Theory, vol. 52, no. 6, pp. 2762–2770, 2006.
[11] S. Boucheron, G. Lugosi, and P. Massart, Concentration inequalities. A
nonasymptotic theory of independence. Oxford University Press, 2013.
multicast rate (bits / channel use)
0.2
baseline
simulation
approx
0.15
0.1
0.05
0
0
2
4
SNR (linear)
6
multicast rate (bits / channel use)
0.5
baseline
simulation
approx
0.4
0.3
0.2
0.1
0
0
1
2
3
SNR (linear)
4
5
6
error probability
10 0
baseline
simulation
approx
10 -1
10 -2
10 -1
SNR (linear)
10 0
error probability
10 0
10 -2
baseline
simulation
approx
10 -4
10 -2
10 -1
SNR (linear)
10 0
error probability
10 0
10 -2
10 -4
baseline
simulation
approx
10 -2
10 -1
SNR (linear)
10 0
multicast rate (bits / channel use)
0.4
baseline
simulation
approx
0.3
0.2
0.1
0
0
1
2
3
SNR (linear)
4
5
6
multicast rate (bits / channel use)
0.8
baseline
simulation
approx
0.6
0.4
0.2
0
0
1
2
3
SNR (linear)
4
5
6
multicast rate (bits / channel use)
baseline
simulation
approx
0.6
0.5
0.4
0.3
0.2
0.1
0
0
2
4
SNR (linear)
6
8
multicast rate (bits / channel use)
1
baseline
simulation
approx
0.8
0.6
0.4
0.2
0
0
2
4
SNR (linear)
6
8
| 7cs.IT
|
New order bounds in differential elimination algorithms✩
Richard Gustavson
arXiv:1602.00246v2 [math.AC] 30 Nov 2016
CUNY Graduate Center, Ph.D. Program in Mathematics, 365 Fifth Avenue, New York, NY 10016, USA
Alexey Ovchinnikov
CUNY Queens College, Department of Mathematics, 65-30 Kissena Boulevard, Queens, NY 11367, USA
Gleb Pogudin
Johannes Kepler University, Institute for Algebra, Science Park II, 3rd Floor, 4040 Linz, Austria
Abstract
We present a new upper bound for the orders of derivatives in the Rosenfeld-Gröbner algorithm.
This algorithm computes a regular decomposition of a radical differential ideal in the ring of
differential polynomials over a differential field of characteristic zero with an arbitrary number
of commuting derivations. This decomposition can then be used to test for membership in the
given radical differential ideal. In particular, this algorithm allows us to determine whether a
system of polynomial PDEs is consistent.
Previously, the only known order upper bound was given by Golubitsky, Kondratieva,
Moreno Maza, and Ovchinnikov for the case of a single derivation. We achieve our bound by
associating to the algorithm antichain sequences whose lengths can be bounded using the results
of León Sánchez and Ovchinnikov.
Keywords: Polynomial differential equations; differential elimination algorithms;
computational complexity
1. Introduction
The Rosenfeld-Gröbner algorithm is a fundamental algorithm in the algebraic theory of differential equations. This algorithm, which first appeared in (Boulier et al., 1995, 2009), takes as
its input a finite set F of differential polynomials and outputs a representation of the radical differential ideal generated by F as a finite intersection of regular differential ideals. The algorithm
has many applications; for example, it can be used to test membership in a radical differential
ideal, and, in conjunction with the differential Nullstellensatz, can be used to test the consistency
✩ This work was partially supported by the NSF grants CCF-095259, CCF-1563942, DMS-1606334, by the NSA
grant #H98230-15-1-0245, by CUNY CIRG #2248, by PSC-CUNY grant #69827-00 47, by the Austrian Science Fund
FWF grant Y464-N18.
Email addresses: [email protected] (Richard Gustavson), [email protected]
(Alexey Ovchinnikov), [email protected] (Gleb Pogudin)
Preprint submitted to Elsevier
December 1, 2016
of a system of polynomial differential equations. See (Golubitsky et al., 2008) for a history of
the development of the Rosenfeld-Gröbner algorithm and similar decomposition algorithms.
The Rosenfeld-Gröbner algorithm has been implemented in Maple as a part of the
DifferentialAlgebra package. In order to determine the complexity of the algorithm, we
need to (among other things) find an upper bound on the orders of derivatives that appear in
all intermediate steps and in the output of the algorithm. The first step in answering this question was completed in (Golubitsky et al., 2008), in which an upper bound in the case of a single
derivation and any ranking on the set of derivatives was found. If there are n unknown functions
and the order of the original system is h, the authors showed that an upper bound on the orders
of the output of the Rosenfeld-Gröbner algorithm is h(n − 1)!.
In this paper, we extend this result by finding an upper bound for the orders of derivatives
that appear in the intermediate steps and in the output of the Rosenfeld-Gröbner algorithm in the
case of an arbitrary number of commuting derivations and a weighted ranking on the derivatives.
We first compute an upper bound for the weights of the derivatives involved for an arbitrary
weighted ranking; by choosing a specific weight, we obtain an upper bound for the orders of the
derivatives. For this, we construct special antichain sequences in the set Zm
>0 ×{1, . . . , n} equipped
with a specific partial order. We then use (León Sánchez and Ovchinnikov, 2016) to estimate the
lengths of our sequences. A general analysis of lengths of antichain sequences began in (Pierce,
2014) and continued in (Freitag and León Sánchez, 2016).
We show that an upper bound for the weights of derivatives in the intermediate steps and in
the output of the Rosenfeld-Gröbner algorithm is given by h fL+1 , where h is the weight of our input system of differential equations, { f0 , f1 , f2 , . . . } is the Fibonacci sequence {0, 1, 1, 2, 3, 5, . . .},
and L is the maximal possible length of a certain antichain sequence (that depends solely on h,
the number m of derivations, and the number n of unknown functions). For m = 2, we refine this
upper bound in a new way by showing that the weights of the derivatives in question are bounded
above by a sequence defined similarly to the Fibonacci sequence but with a slower growth rate.
By choosing a specific weight, we are able to produce an upper bound for the orders of the
derivatives in the intermediate steps and in the output of the Rosenfeld-Gröbner algorithm. Note
that this bound is different from the upper bounds for the effective differential Nullstellensatz
(D’Alfonso et al., 2014; Gustavson et al., 2016a), which are higher and also depend on the degree
of the given system of differential equations. Our result is an improvement of (Gustavson et al.,
2016b) because it allows us to compute sharper order upper bounds with respect to specific
derivations than the previous upper bound did, and because of the refinement in the case m = 2.
For example, if n = 2 and h = 3, 4, 5, the new bound is 3, 8, 33 times better, respectively.
The paper is organized as follows. In Section 2, we present the background material from differential algebra that is necessary to understand the Rosenfeld-Gröbner algorithm. In Section 3,
we describe this algorithm as it is presented in (Hubert, 2003), as well as two necessary auxiliary
algorithms. In Section 4, we prove our main result on the upper bound. In Section 5, we calculate
the upper bound for specific values using the results of (León Sánchez and Ovchinnikov, 2016).
In Section 6, we give an example showing that the lower bound for the orders of derivatives in
the Rosenfeld-Gröbner algorithm is at least double-exponential in the number of derivations.
2. Background on differential algebra
In this section, we present background material from differential algebra that is pertinent to
the Rosenfeld-Gröbner algorithm. For a more in-depth discussion, we refer the reader to (Hubert,
2003; Kolchin, 1973).
2
Definition 1. A differential ring is a commutative ring R with a collection of m commuting
derivations ∆ = {∂1 , . . . , ∂m } on R.
Definition 2. An ideal I of a differential ring is a differential ideal if δa ∈ I for all a ∈ I, δ ∈ ∆.
√
For a set A ⊆ R, let (A), (A), [A], and {A} denote the smallest ideal, radical ideal,√differential
ideal, and radical differential ideal containing A, respectively. If Q ⊆ R, then {A} = [A].
Remark 3. In this paper, as usual, we also use the braces {a1 , a2 , . . .} to denote the set containing
the elements a1 , a2 , . . .. Even though this notation conflicts with the above notation for radical
differential ideals (used here for historical reasons), it will be clear from the context which of the
two objects we mean in each particular situation.
In this paper, k is a differential field of characteristic zero with m commuting derivations.
The set of derivative operators is denoted by
o
n
Θ := ∂i11 · · · ∂imm : i j ∈ Z>0 , 1 6 j 6 m .
For Y = {y1 , . . . , yn } a set of n differential indeterminates, the set of derivatives of Y is
ΘY := {θy : θ ∈ Θ, y ∈ Y}.
Then the ring of differential polynomials over k is defined to be
k{Y} = k{y1 , . . . , yn } := k[θy : θy ∈ ΘY].
We can naturally extend the derivations ∂1 , . . . , ∂m to the ring k{Y} by defining
i +1
∂ j ∂i11 · · · ∂imm yk := ∂i11 · · · ∂ jj · · · ∂imm yk .
For any θ = ∂i11 · · · ∂imm ∈ Θ, we define the order of θ to be
ord(θ) := i1 + · · · + im .
For any derivative u = θy ∈ ΘY, we define
ord(u) := ord(θ).
For a differential polynomial f ∈ k{Y} \ k, we define the order of f to be the maximum order of
all derivatives that appear in f . For any finite set A ⊆ k{Y} \ k, we set
H(A) := max{ord( f ) : f ∈ A}.
(1)
For any θ = ∂i11 · · · ∂imm and positive integers c1 , . . . , cm ∈ Z>0 , we define the weight of θ to be
w(θ) = w ∂i11 · · · ∂imm := c1 i1 + · · · + cm im .
Note that if all of the ci = 1, then w(θ) = ord(θ) for all θ ∈ Θ. For a derivative u = θy ∈ ΘY,
we define the weight of u to be w(u) := w(θ). For any differential polynomial f ∈ k{Y} \ k, we
define the weight of f , w( f ), to be the maximum weight of all derivatives that appear in f . For
any finite set A ⊆ k{Y} \ k, we set
W(A) := max{w( f ) : f ∈ A}.
3
Definition 4. A ranking on the set ΘY is a total order < satisfying the following two additional
properties: for all u, v ∈ ΘY and all θ ∈ Θ, θ , id,
u < θu
and u < v =⇒ θu < θv.
A ranking < is called an orderly ranking if for all u, v ∈ ΘY,
ord(u) < ord(v) =⇒ u < v.
Given a weight w, a ranking < on ΘY is called a weighted ranking if for all u, v ∈ ΘY,
w(u) < w(v) =⇒ u < v.
Remark 5. Note that if w ∂i11 · · · ∂imm = i1 + · · · + im (that is, w(θ) = ord(θ)), then a weighted
ranking < on ΘY is in fact an orderly ranking.
From now on, we fix a weighted ranking < on ΘY.
Definition 6. Let f ∈ k{Y} \ k.
• The derivative u ∈ ΘY of highest rank appearing in f is called the leader of f , denoted
lead( f ).
• If we write f as a univariate polynomial in lead( f ), the leading coefficient is called the
initial of f , denoted init( f ).
• If we apply any derivative δ ∈ ∆ to f , the leader of δ f is δ(lead( f )), and the initial of δ f is
called the separant of f , denoted sep( f ).
Given a set A ⊆ k{Y} \ k, we will denote the set of leaders of A by L(A), the set of initials of
A by IA , and the set of separants of A by S A ; we then let HA = IA ∪ S A be the set of initials and
separants of A.
For a derivative u ∈ ΘY, we let (ΘY)<u (respectively, (ΘY)6u ) be the collection of all derivatives v ∈ ΘY with v < u (respectively, v 6 u). For any derivative u ∈ ΘY, we let A<u (respectively,
A6u ) be the elements of A with leader < u (respectively, 6 u), that is,
A<u := A ∩ k[(ΘY)<u ] and
A6u := A ∩ k[(ΘY)6u ].
We can similarly define (ΘA)<u and (ΘA)6u , where
ΘA := {θ f : θ ∈ Θ, f ∈ A}.
Given f ∈ k{Y} \ k such that deglead( f ) ( f ) = d, we define the rank of f to be
rank( f ) := lead( f )d .
The weighted ranking < on ΘY determines a pre-order (that is, a relation satisfying all of the
properties of an order, except for the property that a 6 b and b 6 a imply that a = b) on k{Y} \ k:
Definition 7. Given f1 , f2 ∈ k{Y} \ k, we say that
rank( f1 ) < rank( f2 )
if lead( f1 ) < lead( f2 ) or if lead( f1 ) = lead( f2 ) and deglead( f1 ) ( f1 ) < deglead( f2 ) ( f2 ).
4
Definition 8. A differential polynomial f is partially reduced with respect to another differential
polynomial g if no proper derivative of lead(g) appears in f , and f is reduced with respect to g
if, in addition,
deglead(g) ( f ) < deglead(g) (g).
A differential polynomial is then (partially) reduced with respect to a set A ⊆ k{Y} \ k if it is
(partially) reduced with respect to every element of A.
Definition 9. For a set A ⊆ k{Y} \ k, we say that A is:
• autoreduced if every element of A is reduced with respect to every other element.
• weak d-triangular if L(A) is autoreduced.
• d-triangular if A is weak d-triangular and every element of A is partially reduced with
respect to every other element.
Note that every autoreduced set is d-triangular. Every weak d-triangular set (and thus every
d-triangular and autoreduced set) is finite (Hubert, 2003, Proposition 3.9). Since the set of leaders
of a weak d-triangular set A is autoreduced, distinct elements of A must have distinct leaders. If
u ∈ ΘY is the leader of some element of a weak d-triangular set A, we let Au denote this element.
Definition 10. We define a pre-order on the collection of all weak d-triangular sets, which
we also call rank, as follows. Given two weak d-triangular sets A = {A1 , . . . , Ar } and B =
{B1 , . . . , B s}, in each case arranged in increasing rank, we say that rank(A) < rank(B) if either:
• there exists a k 6 min(r, s) such that rank(Ai ) = rank(Bi) for all 1 6 i < k and rank(Ak ) <
rank(Bk ), or
• r > s and rank(Ai ) = rank(Bi) for all 1 6 i 6 s.
We also say that rank(A) = rank(B) if r = s and rank(Ai ) = rank(Bi ) for all 1 6 i 6 r.
We can restrict this ranking to the collection of all d-triangular sets or the collection of all
autoreduced sets.
Definition 11. A characteristic set of a differential ideal I is an autoreduced set C ⊆ I of minimal
rank among all autoreduced subsets of I.
Given a finite set S ⊆ k{Y}, let S ∞ denote the multiplicative set containing 1 and generated
by S . For an ideal I ⊆ k{Y}, we define the colon ideal to be
I : S ∞ := {a ∈ k{Y} : ∃s ∈ S ∞ with sa ∈ I}.
If I is a differential ideal, then I : S ∞ is also a differential ideal (Kolchin, 1973, Section I.2).
Definition 12. For a differential polynomial f ∈ k{Y} and a weak d-triangular set A ⊆ k{Y},
a differential partial remainder f1 and a differential remainder f2 of f with respect to A are
∞
differential polynomials such that there exist s ∈ S ∞
A , h ∈ HA such that s f ≡ f1 mod [A] and
h f ≡ f2 mod [A], with f1 partially reduced with respect to A and f2 reduced with respect to A.
5
We denote a differential partial remainder of f with respect to A by pd-red( f, A) and a differential remainder of f with respect to A by d-red( f, A). There are algorithms to compute
pd-red( f, A) and d-red( f, A) for any f and A (Hubert, 2003, Algorithms 3.12 and 3.13). These
algorithms have the property that
rank(pd-red( f, A)), rank(d-red( f, A)) 6 rank( f );
since we have a weighted ranking, this implies that
w(pd-red( f, A)), w(d-red( f, A)) 6 w( f ).
Definition 13. Two derivatives u, v ∈ ΘY are said to have a common derivative if there exist
φ, ψ ∈ Θ such that φu = ψv. Note this is the case precisely when u = θ1 y and v = θ2 y for some
y ∈ Y and θ1 , θ2 ∈ Θ.
Definition 14. If u = ∂i11 · · · ∂imm y and v = ∂1j1 · · · ∂mjm y for some y ∈ Y, we define the least common
derivative of u and v, denoted lcd(u, v), to be
1 , j1 )
m , jm )
lcd(u, v) = ∂max(i
· · · ∂max(i
y.
m
1
Definition 15. For f, g ∈ k{Y} \ k, we define the ∆-polynomial of f and g, denoted ∆( f, g), as
follows. If lead( f ) and lead(g) have no common derivatives, set ∆( f, g) = 0. Otherwise, let
φ, ψ ∈ Θ be such that
lcd(lead( f ), lead(g)) = φ(lead( f )) = ψ(lead(g)),
and define
∆( f, g) := sep(g)φ( f ) − sep( f )ψ(g).
Definition 16. A pair (A, H) is called a regular differential system if:
• A is a d-triangular set
• H is a set of differential polynomials that are all partially reduced with respect to A
• S A ⊆ H∞
• for all f, g ∈ A, ∆( f, g) ∈ ((ΘA)<u) : H ∞ , where u = lcd(lead( f ), lead(g)).
Definition 17. Any ideal of the form [A] : H ∞ , where (A, H) is a regular differential system, is
called a regular differential ideal.
Every regular differential ideal is a radical differential ideal (Hubert, 2003, Theorem 4.12).
Definition 18. Given a radical differential ideal I ⊆ k{Y}, a regular decomposition of I is a finite
collection of regular differential systems {(A1 , H1 ), . . . , (Ar , Hr )} such that
I=
r
\
[Ai ] : Hi∞ .
i=1
Due to the Rosenfeld-Gröbner algorithm, every radical differential ideal in k{Y} has a regular
decomposition.
6
Definition 19. A d-triangular set C is called a differential regular chain if it is a characteristic
set of [C] : HC∞ ; in this case, we call [C] : HC∞ a characterizable differential ideal.
Definition 20. A characteristic decomposition of a radical differential ideal I ⊆ k{Y} is a representation of I as an intersection of characterizable differential ideals.
As we will recall in Section 3, every radical differential ideal also has a characteristic decomposition.
3. Rosenfeld-Gröbner algorithm
Below we reproduce the Rosenfeld-Gröbner algorithm from (Hubert, 2003, Section 6). This
algorithm relies on two others, called auto-partial-reduce and update, which we also include.
We include these two auxiliary algorithms because, in Section 4, we will study their effect on the
growth of the weights of derivatives in Rosenfeld-Gröbner.
Rosenfeld-Gröbner takes as its input two finite subsets F, K ∈ k{Y} and outputs a finite set
A of regular differential systems such that
\
{F} : K ∞ =
[A] : H ∞ ,
(2)
(A,H)∈A
where A = ∅ if 1 ∈ {F} : K ∞ .
If we have a decomposition of {F} : K ∞ as in (2), we can compute, using only algebraic
operations, a decomposition of the form
\
(3)
{F} : K ∞ =
[C] : HC∞ ,
C∈C
where C is finite and each C ∈ C is a differential regular chain (Hubert, 2003, Algorithms 7.1
S
and 7.2). This means that an upper bound on (A,H)∈A W(A ∪ H) from (2) will also be an upper
S
bound on C∈C W(C) from (3).
Rosenfeld-Gröbner has many immediate applications. For example, if K = {1}, then
{F} : K ∞ = {F}, so in this case, Rosenfeld-Gröbner computes a regular decomposition of {F},
which then also gives us a characteristic decomposition of {F} by the discussion in the previous
paragraph.
The weak differential Nullstellensatz says that a system of polynomial differential equations
F = 0 is consistent (that is, has a solution in some differential field extension of k) if and only if
1 < [F] (Kolchin, 1973, Section IV.2). Thus, since Rosenfeld-Gröbner(F, K) = ∅ if and only if
1 ∈ {F} : K ∞ , we see that F = 0 is consistent if and only if Rosenfeld-Gröbner(F, {1}) , ∅.
More generally, Rosenfeld-Gröbner and its extension for computing a characteristic decomposition of a radical differential ideal allow us to test for membership in a radical differential
ideal, as follows. Suppose we have computed a characteristic decomposition
\
{F} =
[C] : HC∞ .
C∈C
Now, a differential polynomial f ∈ k{Y} is contained in {F} if and only if f ∈ [C] : HC∞ for all
C ∈ C; this latter case is true if and only if d-red( f, C) = 0, which can be tested using (Hubert,
2003, Algorithm 3.13).
Rosenfeld-Gröbner, auto-partial-reduce, and update rely on the following tuples of differential polynomials:
7
Definition 21. A Rosenfeld-Gröbner quadruple (or RG-quadruple) is a 4-tuple (G, D, A, H) of
finite subsets of k{Y} such that:
• A is a weak d-triangular set, HA ⊆ H, D is a set of ∆-polynomials, and
• for all f, g ∈ A, either ∆( f, g) = 0 or ∆( f, g) ∈ D or
∆( f, g) ∈ (Θ(A ∪ G)<u ) : Hu∞ ,
where u = lcd(lead( f ), lead(g)) and Hu = HA<u ∪ (H \ HA ) ∩ k[(ΘY)<u ].
Algorithm: Rosenfeld-Gröbner, (Hubert, 2003, Algorithm 6.11)
Data: F, K finite subsets of k{Y}
Result: A set A of regular differential systems such that:
• A is empty if it has been detected that 1 ∈ {F} : K ∞
T
• {F} : K ∞ =
[A] : H ∞ otherwise
(A,H)∈A
S := {(F, ∅, ∅, K)};
A := ∅;
while S , ∅ do
(G, D, A, H) := an element of S;
S̄ = S \ (G, D, A, H);
if G ∪ D = ∅ then
A := A ∪ auto-partial-reduce(A, H);
else
p := an element of G ∪ D;
Ḡ, D̄ := G \ {p}, D \ {p};
p̄ := d-red(p, A);
if p̄ = 0 then
S̄ := S̄ ∪ {(Ḡ, D̄, A, H)};
else
if p̄ < k then
p̄i := p̄ − init( p̄) rank( p̄) p̄ s := deglead( p̄) ( p̄) p̄ − lead( p̄) sep( p̄);
S̄ := S̄ ∪ {update(Ḡ, D̄, A, H, p̄), (G ∪ { p̄ s , sep( p̄)}, D̄, A, H ∪
{init( p̄)}), (Ḡ ∪ { p̄i , init( p̄)}, D̄, A, H)};
end
end
end
S := S̄;
end
return A;
Remark 22. The RG-quadruple that is output by update satisfies additional properties that we
do not list, as they are not important for our analysis. For more information, we refer the reader
to (Hubert, 2003, Algorithm 6.10)
8
Algorithm: auto-partial-reduce, (Hubert, 2003, Algorithm 6.8)
Data: Two finite subsets A, H of k{Y} such that (∅, ∅, A, H) is an RG-quadruple
Result:
• The empty set if it is detected that 1 ∈ [A] : H ∞
• Otherwise, a set with a single regular differential system (B, K) with L(A) = L(B),
HB ⊆ K, and [A] : H ∞ = [B] : K ∞
B := ∅;
for u ∈ L(A) increasingly do
b := pd-red(Au , B);
if rank(b) = rank(Au ) then
B := B ∪ {b};
else
return (∅);
end
end
K := HB ∪ {pd-red(p, B) : p ∈ H \ HA };
if 0 ∈ K then
return (∅);
else
return {(B, K)};
end
Algorithm: update (Hubert, 2003, Algorithm 6.10)
Data:
• A 4-tuple (G, D, A, H) of finite subsets of k{Y}
• A differential polynomial p reduced with respect to A such that (G ∪ {p}, D, A, H) is an
RG-quadruple
Result: A new RG-quadruple (Ḡ, D̄, Ā, H̄)
u := lead(p);
G A := {a ∈ A | lead(a) ∈ Θu};
Ā := A \ G A ;
Ḡ := G ∪ G A ;
D̄ := D ∪ {∆(p, a) | a ∈ Ā} \ {0};
H̄ := H ∪ {sep(p), init(p)};
return (Ḡ, D̄, Ā ∪ {p}, H̄);
9
4. Order upper bound
Given finite subsets F, K ⊆ k{Y}, let h = W(F ∪ K). Our goal is to find an upper bound for
[
W
(A ∪ H) ,
(A,H)∈A
where A = Rosenfeld-Gröbner(F, K), in terms of h, m (the number of derivations), and n (the
number of differential
indeterminates).
By then choosing a specific weight, we can find an upper
S
bound for H (A,H)∈A (A ∪ H) in terms of m, n, and H(F ∪ K).
We approach this problem as follows. Every (A, H) ∈ A is formed by applying auto-partialreduce to a 4-tuple (∅, ∅, A′, H ′ ) ∈ S. Thus, it suffices:
• to bound how auto-partial-reduce increases the weight of a collection of differential polynomials (it turns out to not increase the weight), and
• to bound W(G ∪ D ∪ A ∪ H) for all (G, D, A, H) added to S throughout the course of
Rosenfeld-Gröbner.
We accomplish the latter by determining when the weight of a tuple (G, D, A, H) added to S is
larger than the weights of the previous elements of S and bounding W(G ∪ D ∪ A ∪ H) in this
instance, and then bounding the number of times we can add such elements to S.
N
There is a sequence {(Gi , Di , Ai , Hi )}i=0
corresponding to each regular differential system
(A, H) in the output of Rosenfeld-Gröbner, where N = N(A,H) , such that (Gi+1 , Di+1 , Ai+1 , Hi+1 )
is obtained from (Gi , Di , Ai , Hi ) during the while loop, (G0 , D0 , A0 , H0 ) = (F, ∅, ∅, K), and
(A, H) = auto-partial-reduce(AN , HN ).
We begin with an auxiliary result, which is an analogue of the first property from
(Golubitsky et al., 2009, Section 5.1).
Lemma 23. For every f ∈ Ai and i < j, there exists g ∈ A j such that lead( f ) ∈ Θ lead(g). In
particular, if p is reduced with respect to A j , then p is reduced with respect to Ai for all i < j.
Proof. It is sufficient to consider the case j = i + 1. If (Gi+1 , Di+1 , Ai+1 , Hi+1 ) was obtained from
(Gi , Di , Ai , Hi ) without applying update, then Ai = Ai+1 . Otherwise, either f ∈ Ai \ G Ai (we use
the notation from update), or f ∈ G Ai . In the former case, f ∈ Ai+1 as well, so we can set g = f .
In the latter case, lead( f ) ∈ Θ lead(p), so we can set g = p.
We define a partial order 4 on the set of derivatives ΘY as follows. For u, v ∈ ΘY, we say
that u 4 v if there exists θ ∈ Θ such that θu = v. Note that this implies that u and v are both
derivatives of the same y ∈ Y.
Definition 24. An antichain sequence in ΘY is a sequence of elements S = {s1 , s2 , . . . } ⊆ ΘY
that are pairwise incomparable in this partial order.
N
Given a sequence {(Gi , Di , Ai , Hi )}i=0
as above (where N = N(A,H) for some regular differential
system (A, H) in the output of Rosenfeld-Gröbner), we will construct an antichain sequence
S = {s1 , s2 , . . .} ⊆ ΘY inductively going along the sequence {(Gi , Di , Ai , Hi )}. Suppose S j−1 =
{s1 , . . . , s j−1 } has been constructed after considering (G0 , D0 , A0 , H0 ), . . . , (Gi−1 , Di−1 , Ai−1 , Hi−1 ),
where S 0 = ∅. A 4-tuple (Gi , Di , Ai , Hi ) can be obtained from the tuple (Gi−1 , Di−1 , Ai−1 , Hi−1 )
in two ways:
10
(1) We did not perform update. In this case, we do not append a new element to S .
(2) We performed update with respect to a differential polynomial p̄. If there exists sk ∈ S j−1
such that lead( p̄) 6 sk , we do not append a new element to S j−1 . Otherwise, let s j =
lead( p̄) and define S j = {s1 , . . . , s j }. In the latter case, we set k j = i. We also set k0 = 0.
Theorem 25. The sequence {s j } is an antichain sequence in ΘY and, for all j > 1,
w(s j ) 6 h f j ,
where { f j } is the Fibonacci sequence.
For m = 2, we provide a refined version of Theorem 25. Let { f (n, h)k } be the sequence:
f (n, h)0 = 0, f (n, h)1 = f (n, h)2 = h
(4)
f (n, h)k = f (n, h)k−1 + f (n, h)k−2
for k 6 n + 1
f (n, h)k = f (n, h)k−1 + f (n, h)k−2 − 1 for k > n + 1.
Proposition 26. For m = 2 the sequence {s j } satisfies, for all j > 1,
w(s j ) 6 f (n, h) j.
We will prove Proposition 26 while proving Theorem 25, highlighting the case m = 2.
Proof. Let i < j. Assume that s j < si . Then, p is not reduced with respect to Aki , which
contradicts Lemma 23. On the other hand, the case s j 4 si is impossible by the construction of
the sequence, so {s j } is an antichain sequence.
Let L denote the length of the sequence {s j }. We denote the maximal j ∈ Z>0 such that k j 6 i
by anti-ki . For all i > 0, let us set j = anti-ki and prove by induction on i that
!
i
S
(1) W
(Gt ∪ Dt ∪ Ht ) 6 h f j+1 ;
t=0
(2) W
i
S
t=0
!
At 6 h f j ;
(3) For all distinct elements of
i
S
At , the weights of the least common derivatives of their
t=0
leaders do not exceed h f j+1 .
If m = 2, let F0 = 0, F1 = F2 = h. We will show that there exists a sequence {Fr } such that
• for all r > 1, w(sr ) 6 Fr and
• Fr = Fr−1 + Fr−2 − 1 for all r > 3 except at most n − 1 of them, for which Fr = Fr−1 + Fr−2 .
In the latter case, we will say that r is a jump index. Note that 2 is not a jump index by the
definition, although F2 = F1 + F0 .
For each such sequence, the induction hypothesis will be the following:
!
!
i
i
S
S
(1) W
(Gt ∪ Dt ∪ Ht ) 6 F j+1 for j < L and W
(Gt ∪ Dt ∪ Ht ) 6 F L+1 + 1 for j = L;
t=0
t=0
11
(2) W
i
S
t=0
!
At 6 F j ;
(3) For all distinct elements of
i
S
At , the weights of the least common derivatives of their
t=0
leaders do not exceed F j+1 for j < L and F L+1 + 1 for j = L;
(4) If, in either of (1) or (3), the equality holds in the case j = L, then, for every q, 1 6 q 6 n,
a
b
the sequence {sr } contains ∂1q yq and ∂2q yq for some aq and bq .
In the base case i = 0 = k0 , we have
W(G0 ∪ D0 ∪ H0 ) = h = h f1 (F1 in the case m = 2)
and
W(A0 ) = W(∅) = 0 = h f0 (F0 in the case m = 2).
There are two distinct cases for i + 1:
(1) Case i + 1 < k j+1 (so anti-ki+1 = j). Then, (Gi+1 , Di+1 , Ai+1 , Hi+1 ) was obtained from
(Gi , Di , Ai , Hi ) in one of the following ways:
(a) We did not perform update. In this case, Ai+1 = Ai and
W(Gi+1 ∪ Di+1 ∪ Hi+1 ) 6 W(Gi ∪ Di ∪ Hi ).
(b) We performed update with respect to a differential polynomial p such that lead( f ) ∈
S
Θ lead(p) for some f ∈ it=0 At . In this case,
i
[
W(Ai+1 ) 6 W At .
t=0
Then, for all g ∈ At (t 6 i),
w(∆(p, g)) 6 w(lcd(lead(g), lead( f ))),
which is bounded by h f j+1 (by F j+1 or F L+1 + 1 in the case m = 2) due to the
third inductive hypothesis. Since Di+1 \ Di consists of some of these polynomials,
Gi+1 \ Gi ⊆ Ai , and Hi+1 \ Hi = {sep(p), init(p)}, then
W(Gi+1 ∪ Di+1 ∪ Hi+1 ) 6 W(Gi ∪ Di ∪ Hi ).
(2) Case i + 1 = k j+1 (so now anti-ki+1 = j + 1). We performed update with respect to a
differential polynomial p, which is a result of reduction of some p̃ ∈ Gi ∪ Di with respect
to Ai . Then
W(Ai+1 ) 6 max(W(Ai ), w(p)) 6 h f j+1 .
Moreover, for every g ∈
i
S
At ,
t=0
w(lcd(lead(g), lead(p))) 6 h f j + h f j+1 = h f j+2 .
12
(5)
Since Di+1 \ Di consists of some of these polynomials, Gi+1 \ Gi ⊆ Ai , and Hi+1 \ Hi =
{sep(p), init(p)}, we have
W(Gi+1 ∪Di+1 ∪ Hi+1 ) 6 max(W(Gi ∪ Di ∪ Hi ), h f j+2 ) = h f j+2 .
In the case m = 2, instead of (5), we obtain
w(lcd(lead(g), lead(p))) 6 w(lead(p)) + w(lead(g))
(6)
If (6) is strict, we have
w(lcd(lead(g), lead(p))) 6 w(lead(p)) + w(lead(g)) − 1 6 F j + F j+1 − 1 = F j+2 ,
and j + 2 is not a jump index. Otherwise, (6) turns out to be an equality. In this case, the
only possibility is lead(p) = ∂a1 yr and lead(g) = ∂b2 yr (or vice versa) for some r. Note that,
for every r, such a situation occurs at most once. Consider the following two cases:
a
b
(a) For every q, 1 6 q 6 n, the sequence s1 , . . . , s j+1 already contains ∂1q yq and ∂2q yq
for some aq and bq . In this case, s1 , . . . , s j+1 already form an antichain sequence that
cannot be extended further, so j + 1 = L. We set F L+1 = F L + F L−1 − 1, so we can
bound the right-hand side of (6) from above by F L+1 + 1.
(b) Otherwise, we just set F j+2 = F j+1 + F j , so j + 2 is a jump index, and we still have
less than n of them.
Since w(s j ) 6 W(Ak j ) 6 h f j , this completes the proof of Theorem 25.
In order to complete the proof of Proposition 26, it is sufficient to show that, for every such
sequence {F j }, for all j, f (n, h) j > F j . Let i1 , . . . , in−1 denote the jump indices of {F j }. Note
that { f (n, h) j } is uniquely defined as a sequence of the same type as {F j } with jump indices
3, . . . , n + 1. It is sufficient to prove that, after decreasing any jump index of {F j } by one, we
obtain a sequence which is not smaller than {F j }. Then, since we will obtain { f (n, h) j } after some
number of such operations and the jump indices of { f (n, h) j } cannot be further decreased, we
will have that { f (n, h) j } is the largest such sequence. The claim is true since, before decreasing
i j , the sequence was of the form
. . . , Fi j −2 ,
Fi j −1 = Fi j −3 + Fi j −2 − 1,
Fi j = Fi j −1 + Fi j −2 = Fi j −3 + 2Fi j −2 − 1, . . .
but, after decreasing i j by one, it will be of the form
. . . , Fi j −2 ,
Fi j −1 = Fi j −3 + Fi j −2 ,
Fi j = Fi j −1 + Fi j −2 − 1 = Fi j −3 + 2Fi j −2 − 1, . . .
Since the rest of terms obey the same recurrence for both sequences, the latter is not smaller than
the former.
Let n = {1, . . . , n}. Define the degree of an element ((i1 , . . . , im ), k) ∈ Zm
>0 ×n to be i1 +· · ·+im .
i
im
1
Given a weight w ∂1 · · · ∂m = c1 i1 + · · · + cm im on Θ, define a map from the set of derivatives
ΘY to the set Zm
>0 × n by
∂i11 · · · ∂imm yk 7→ ((c1 i1 , . . . , cm im ), k).
13
Note the degree of the image of θy in Z>0 × n is equal to the weight of θy in ΘY.
Under this map, the partial order 4 on ΘY determines a partial order 4 on Zm
>0 × n by saying
((i1 , . . . , im ), k) 4 (( j1 , . . . , jm ), l) ⇐⇒ k = l and ir 6 jr for all r, 1 6 r 6 m.
Thus, every antichain sequence of ΘY determines an antichain sequence of Zm
>0 × n. Every
2014,
Lemma
4.4).
antichain sequence of Zm
×
n
(and
thus
of
ΘY)
is
finite
(Pierce,
>0
Given an increasing function f : Z>0 → Z>0 , we say that f bounds the degree growth of an
antichain sequence S = {s1 , . . . , sk } ⊆ Zm
>0 × n if deg(si ) 6 f (i) for all 1 6 i 6 k. By (Pierce,
2014, Lemma 4.9), there is an upper bound on the length of an antichain sequence of Zm
>0 × n
with degree growth bounded by f , and this bound depends only on m, n, and f . Let Lnf,m be the
maximal length of an antichain sequence of Zm
>0 × n with degree growth bounded by f .
Theorem 27. Let F, K ⊆ k{Y} be finite subsets with h = W(F ∪ K), L = Lnf,m , and A =
Rosenfeld-Gröbner(F, K), where f (i) = h fi with { fi } the Fibonacci sequence. Then
[
(A ∪ H) 6 h fL+1 .
W
(A,H)∈A
Proof. Since w(pd-red(p, B)) 6 w(p) for any p ∈ k{Y} and weak d-triangular set B, we have
W(B ∪ K) 6 W(A ∪ H), where {(B, K)} = auto-partial-reduce(A, H). Hence, it suffices to
bound W(G ∪ D ∪ A ∪ H) whenever the tuple (G, D, A, H) is added to S in Rosenfeld-Gröbner.
By Theorem 25 and the correspondence between antichain sequences of ΘY and Zm
>0 × n, we
obtain an antichain sequence of Zm
×
n
of
degree
growth
bounded
by
f
(i),
so
the
length
of this
>0
sequence (and thus the sequence from Theorem 25) is at most L.
In the proof of Theorem 25, it is shown that for all i 6 N, for j := anti-ki , we have
i
[
W (Gt ∪ Dt ∪ At ∪ Ht ) 6 h f j+1 .
t=1
Since the largest possible j is the length of the antichain sequence (and this j is equal to anti-kN ),
for every (Gi , Di , Ai , Hi ), we have
W(Gi ∪ Di ∪ Ai ∪ Hi ) 6 h fL+1 .
Since every (G, D, A, H) added to S is equal to (Gi , Di , Ai , Hi ) for some i, this ends the proof.
Corollary 28. Let m = 2, F, K ⊆ k{Y} be finite subsets with h = W(F ∪ K), L = Lnf,m , and
A = Rosenfeld-Gröbner(F, K), where f (i) = f (n, h)i with { f (n, h)i } given by (4). Then
[
W
(A ∪ H) 6 f (n, h)L+1 .
(A,H)∈A
Proof. Replacing h fi with f (n, h)i everywhere in the proof of Theorem 27, we obtain an argument that is valid in all cases except for the case in which, for every q, 1 6 q 6 n, the
a
b
antichain sequence {s j } contains ∂1q yq and ∂2q yq for some aq and bq . In this case, we still have
W(Ai ) 6 f (n, h)L for all i. We will prove that W(Hi ) 6 f (n, h)L+1 for all i. For i < kL , this
inequality follows from the proof of Theorem 25. For i > kL , every h added to Hi is reduced with
14
respect to Ai (see Rosenfeld-Gröbner). The definition of k j implies that the set of leaders of Ak j
contains s j . While performing update for Ai , every leader s of Ai either survives or is replaced
with s̃ such that s is a derivative of s̃. Hence, for all i > k j , the set of leaders of Ai contains either
s j or s̃ such that s j is a derivative of s̃. Thus, since h is reduced with respect to Ai for i > kL , for
every variable ∂a1 ∂b2 yq occurring in h, we have a < aq and b < bq . Thus,
b −1
a −1
w(h) 6 max w ∂1q yq + w ∂2q yq 6 f (n, h)L + f (n, h)L−1 − 2 < f (n, h)L+1 .
16q6n
We can use Theorem 27 and Corollary 28 to bound the orders of the output RosenfeldGröbner. Let F, K ⊆ k{Y} be two finite subsets, and define a weight w on Θ such that
W(F ∪ K) = H(F ∪ K).
(7)
This can always be done by letting w(θ) = ord(θ) for all derivatives θ, but there are sometimes
other weights that lead to equation (7) being satisfied.
Example 29. We provide examples of differential polynomials f that arise as part of systems of
PDEs for which it is possible to construct a nontrivial weight w such that w( f ) = ord( f ). We
note that we are not applying Rosenfeld-Gröbner to these examples; we simply present them to
demonstrate that there are nontrivial weights satisfying equation (7).
(1) Consider the heat equation
ut − α · (u xx + uyy ) = 0,
f (u) := ∂t u − α · (∂2x u + ∂2y u) ∈ k{u},
where u(x, y, t) is the unknown, α is a positive constant, and k{u} has derivations
{∂ x , ∂y , ∂t }. If we define a weight w on Θ by
w ∂ix ∂yj ∂kt = i + j + 2k,
then w( f ) = 2 = ord( f ).
(2) Consider the K-dV equation
φt + φ xxx + 6φφ x = 0,
f (φ) := ∂t φ + ∂3x φ + 6φ∂ x φ ∈ k{φ},
where φ(x, t) is the unknown and k{φ} has derivations {∂ x , ∂t }. Define a weight w on Θ by
j
w ∂ix ∂t = i + 3 j,
so that w( f ) = 3 = ord( f ).
Using Theorem 27, Corollary 28, and (7), we obtain the following order bound for the output
of Rosenfeld-Gröbner:
Corollary 30. Let F, K ⊆ k{Y} be finite subsets with h = H(F ∪ K), L = Lnf,m , A =
Rosenfeld-Gröbner(F, K), where f (i) = f (n, h)i with { f (n, h)i } the
sequence
given by (4) if m = 2
and f (i) = h fi with { fi } the Fibonacci sequence if m > 2. Let w ∂i11 · · · ∂imm = c1 i1 + · · · + cm im be
a weight defined on Θ such that W(F ∪ K) = H(F ∪ K). Then, for all g ∈ A,
f (n,h)
L+1
if m = 2
ci
ord(g, ∂i) 6
h
f
L+1
if m > 2.
ci
15
5. Specific values
In order to apply the results of the previous section, we need to be able to effectively compute
Lnf,m . (Pierce, 2014) only proved the existence of this number, without an analysis of how to
construct it. (Freitag and León Sánchez, 2016) constructed an upper bound for m = 1, 2. The
first analysis for the case of arbitrary m appears in (León Sánchez and Ovchinnikov, 2016).
Let f : Z>0 → Z>0 be an increasing function. Let us define a function Ψ f,m : Z>0 × Zm
>0 →
Z>0 by the following relations:
Ψ f,m (i, (0, . . . , 0, um)) = i,
Ψ f,m (i − 1, (u1 , . . . , ur , 0, . . . , 0, um))
= Ψ f,m (i, (u1, . . . , ur − 1, f (i) − f (i − 1) + um + 1, 0, . . . , 0)), r < m − 1, ur > 0,
Ψ
(i
f,m − 1, (u1 , . . . , um ))
= Ψ f,m (i, (u1, . . . , um−1 − 1, f (i) − f (i − 1) + um + 1)),
um−1 > 0.
Proposition 31 ((León Sánchez and Ovchinnikov, 2016, Corollary 3.10)). The maximal length
of an antichain sequence in Zm
>0 with degree growth bounded by f does not exceed
Ψ f,m (1, ( f (1), 0, . . . , 0)).
Let us also define the sequence ψ0 , ψ1 , . . . by the relations ψ0 = 0 and
ψi+1 = Ψ fi ,m (1, ( fi (1), 0, . . . , 0)) + ψi ,
fi (x) := f (x + ψi ).
Proposition 32 ((León Sánchez and Ovchinnikov, 2016, Corollary 3.14)). The maximal length
of an antichain sequence in Zm
>0 × n with degree growth bounded by f does not exceed ψn .
Now, let us apply this technique to the functions f1 (i) = f (n, h)i and f2 (i) = h fi . Then,
by Theorem 27 and Corollary 28, an upper bound on the weights of the output of RosenfeldGröbner will be f1 (Lnf1 ,m + 1) if m = 2 and f2 (Lnf2 ,m + 1) if m > 2. In general, we do not have
a formula for Lnf1 ,m and Lnf2 ,m for arbitrary h, m, n that improves the one given in Proposition 32;
however, we can compute Lnf1 ,m and Lnf2 ,m for some specific values of h, m, n.
If W(F ∪ K) = H(F ∪ K) = h, we can use Corollary 30 to produce perhaps sharper
bounds for the order of the elements of Rosenfeld-Gröbner(F, K) with respect to particular derivations. In the examples that follow,
we calculate upper bounds for ord(g, ∂1) for
i
1
g ∈ Rosenfeld-Gröbner(F, K), where w ∂1 · · · ∂imm = c1 i1 + · · · + cm im in the case in which
c1 = 2 and the case in which c1 = 3. We note that in the tables that follow, “N/A” appears
whenever we cannot have the given initial order h with given ci as part of the weight function.
(1) Assume that n = 1 and m = 2. Then the maximal length of an antichain sequence does
not exceed h + 1. In this case, the weights of the resulting polynomials are bounded by
f (1, h)h+2 , which results in the following table:
h
f (1, h)h+2
ord(g, ∂1), c1 = 2
ord(g, ∂1), c1 = 3
1
1
N/A
N/A
2
4
2
N/A
3
11
5
3
16
4
25
12
8
5
55
26
17
6
106
53
35
7
205
102
68
8
386
193
128
9
713
356
237
10
1297
648
432
(2) Assume that m = 2 and n is arbitrary. Then the maximal length of an antichain sequence
does not exceed bn , where bn satisfies b1 = h + 1 and bn+1 = f (n, h)bn+1 + bn + 1, which
results in the following table:
n
2
2
2
2
3
h
1
2
3
4
1
bn
5
9
18
34
11
f (n, h)bn+1
4
77
9,960
31, 206, 974
90
ord(g, ∂1 ), c1 = 2
N/A
38
4,980
15, 603, 487
N/A
ord(g, ∂1 ), c1 = 3
N/A
N/A
3,320
10, 402, 324
N/A
(3) Assume that m = 3 and n = 1. We can construct the maximal length antichain sequence of
Z3>0 using the methods of (León Sánchez and Ovchinnikov, 2016) and the function f (i) =
h fi , resulting in the following sequence:
(h, 0, 0), (h − 1, 1, 0), (h − 1, 0, h + 1), (h − 2, 2h + 2, 0), . . . ,
(h − 2, 0, h f2h+6 − (h − 2)), . . . , (h − i, h fci−1 +1 − (h − i), 0), . . . ,
(h − i, 0, h fci − (h − i)), . . . , (0, h fch−1+1 , 0), . . . , (0, 0, h fch ),
where the sequence ci is given by c0 = 1 and for 1 6 i 6 h,
ci = ci−1 + 1 + h fci−1 +1 − (h − i).
As a result, we see that the maximal length of an antichain sequence is equal to ch .
Below is a table of some maximal lengths Lnf,m and weights f (Lnf,m + 1), where f (i) = h fi ,
for m = 3, 4, and 5:
m
3
3
4
5
n
1
1
1
1
h
1
2
1
1
length
3
10
5
20
weight
3
178
8
10,946
6. Order lower bound
This section gives a lower bound for the orders of the output of Rosenfeld-Gröbner, coming
from the lower bound for degrees of elements of a Gröbner basis from (Yap, 1991). To be
specific, we show that for m, h sufficiently large, there is a set of r differential polynomials
F ⊆ k{y} of order at most h, where k is equipped with m derivations, r ∼ m/2, and k is constant
with respect to all of the derivations, such that if A = Rosenfeld-Gröbner(F, {1}), then
[
r
H
(A ∪ H) > h2 .
(8)
(A,H)∈A
The arguments presented here are standard, and we include them for completeness. We first note
the following standard fact about differential ideals generated by linear differential polynomials.
17
Proposition 33. Suppose F, K ⊆ k{Y} are composed of linear differential polynomials. Then
the output of Rosenfeld-Gröbner(F, K) is either empty or consists of a single regular differential
system (A, H) with A and H both composed of linear differential polynomials.
Suppose now we apply Rosenfeld-Gröbner to (F, {1}), where F consists of linear differential
polynomials, in order to obtain a regular decomposition of {F}. Since every element of F is
linear, [F] is a prime differential ideal, so by Proposition 33, we have
[F] = {F} = [A] : H ∞
for some regular differential system (A, H), with A and H both composed of linear differential
polynomials. Since every element of A is linear, after performing scalar multiplications and
addition, A can be transformed to an autoreduced set Ā without affecting the leaders and orders
of elements of A. Since (A, H) is a regular differential system, Ā is a characteristic set of [F]. So,
it suffices to find a lower bound on the orders of elements of linear characteristic sets in k{Y}.
There is a well-studied one-to-one correspondence between polynomials in k[x1 , . . . , xm ] and
homogeneous linear differential polynomials in k{y} with m derivations and k a field of constants:
X
X
(9)
ci1 ,...,im ∂i11 · · · ∂imm y.
ci1 ,...,im xi11 · · · ximm ↔
Any orderly ranking on Θy then determines a graded monomial order on k[x1 , . . . , xm ].
Given a polynomial f ∈ k[x1 , . . . , xm ], let f˜ ∈ k{y} be its corresponding differential polynomial under (9). By the discussion above, if we have a collection of polynomials f1 , . . . , fr ∈
k[x1 , . . . , xm ], we can construct a characteristic set C = {C1 , . . . , C s } of [ f˜1 , . . . , f˜r ] ⊆ k{y} consisting of homogeneous linear differential polynomials, and so each Ci ∈ k{y} is in fact equal to
g̃i for some gi ∈ k[x1 , . . . , xm ].
Proposition 34 (cf. (Wu, 2005, page 352),(Gerdt, 2005)). With the notation above, {g1 , . . . , g s } ⊆
k[x1 , . . . , xm ] is a Gröbner basis of the ideal I = ( f1 , . . . , fr ).
By Proposition 34, we can thus find a lower bound for the orders of the output of RosenfeldGröbner via a lower bound for the degrees of elements of a Gröbner basis, as we do in the
following example.
Example 35. This example demonstrates the lower bound (8) for the orders of the output of
Rosenfeld-Gröbner. In (Yap, 1991, Section 8), for m, h sufficiently large, a collection of m
algebraic polynomials f1 , . . . , fr of degree at most h in m algebraic indeterminates, with r ∼ m/2,
is constructed such that any Gröbner basis of ( f1 , . . . , fr ) with respect to a graded monomial
r
order has an element of degree at least h2 .
As a result of the previous discussion, we have a collection of differential polynomials
F = f˜1 , . . . , f˜r ∈ k{y} of order h with m derivations such that any linear characteristic set
r
of [ f˜1 , . . . , f˜r ] will contain a differential polynomial of order at least h2 . Since in this case
{(A, H)} = Rosenfeld-Gröbner(F, {1}) can be transformed into a linear characteristic set without affecting the orders of the elements, this means that
r
H(A ∪ H) > h2 .
18
References
Boulier, F., Lazard, D., Ollivier, F., Petitot, M., 1995. Representation for the radical of a finitely generated differential
ideal. In: ISSAC’95: Proceedings of the 1995 international symposium on Symbolic and algebraic computation. ACM
Press, New York, NY, USA, pp. 158–166.
URL http://dx.doi.org/10.1145/220346.220367
Boulier, F., Lazard, D., Ollivier, F., Petitot, M., 2009. Computing representations for radicals of finitely generated differential ideals. Applicable Algebra in Engineering, Communication and Computing 20 (1), 73–121.
URL http://dx.doi.org/10.1007/s00200-009-0091-7
D’Alfonso, L., Jeronimo, G., Solernó, P., 2014. Effective differential Nullstellensatz for ordinary DAE systems with
constant coefficients. Journal of Complexity 30 (5), 588–603.
URL http://dx.doi.org/10.1016/j.jco.2014.01.001
Freitag, J., León Sánchez, O., 2016. Effective uniform bounds in partial differential fields. Advances in Mathematics 288,
308–336.
URL http://dx.doi.org/10.1016/j.aim.2015.10.013
Gerdt, V., 2005. Involutive algorithms for computing Gröbner bases. In: Computational commutative and noncommutative algebraic geometry. Vol. 196 of NATO Science Series III: Computer and Systems Sciences. IOS, Amsterdam.
URL http://arxiv.org/abs/math/0501111
Golubitsky, O., Kondratieva, M., Moreno Maza, M., Ovchinnikov, A., 2008. A bound for Rosenfeld-Gröbner algorithm.
Journal of Symbolic Computation 43 (8), 582–610.
URL http://dx.doi.org/10.1016/j.jsc.2007.12.002
Golubitsky, O., Kondratieva, M., Ovchinnikov, A., Szanto, A., 2009. A bound for orders in differential Nullstellensatz.
Journal of Algebra 322 (11), 3852–3877.
URL http://dx.doi.org/10.1016/j.jalgebra.2009.05.032
Gustavson, R., Kondratieva, M., Ovchinnikov, A., 2016a. New effective differential Nullstellensatz. Advances in Mathematics 290, 1138–1158.
URL http://dx.doi.org/10.1016/j.aim.2015.12.021
Gustavson, R., Ovchinnikov, A., Pogudin, G., 2016b. Bounds for orders of derivatives in differential elimination algorithms. In: ISSAC ’16: Proceedings of the 2016 ACM International Symposium on Symbolic and Algebraic Computation. ACM Press, New York, NY, USA, pp. 239–246.
URL http://dx.doi.org/10.1145/2930889.2930922
Hubert, E., 2003. Notes on triangular sets and triangulation-decomposition algorithms II: Differential systems. In:
Langer, U., Winkler, F. (Eds.), Symbolic and Numerical Scientific Computations. Vol. 2630 of Lecture Notes in
Computer Science. Springer, pp. 40–87.
URL http://dx.doi.org/10.1007/3-540-45084-X_2
Kolchin, E. R., 1973. Differential algebra and algebraic groups. Academic Press, New York, pure and Applied Mathematics, Vol. 54.
León Sánchez, O., Ovchinnikov, A., 2016. On bounds for the effective differential Nullstellensatz. Journal of Algebra
449, 1–21.
URL http://dx.doi.org/10.1016/j.jalgebra.2015.10.009
Pierce, D., 2014. Fields with several commuting derivations. Journal of Symbolic Logic 79 (1), 1–19.
URL http://dx.doi.org/10.1017/jsl.2013.19
Wu, W., 2005. On the construction of Groebner basis of a polynomial ideal based on Riquier-Janet theory. In: Wang, D.,
Zheng, Z. (Eds.), Differential Equations with Symbolic Computation. Trends in Mathematics. Birkhäuser Verlag, pp.
351–368.
URL http://dx.doi.org/10.1007/3-7643-7429-2_20
Yap, C., 1991. A new lower bound construction for commutative Thue systems with applications. Journal of Symbolic
Computation 12, 1–27.
URL http://dx.doi.org/10.1016/S0747-7171(08)80138-1
19
| 0math.AC
|
1
Figure
Figure11
Joint Person Re-identification and Camera Network
Topology Inference in Multiple Cameras
arXiv:1710.00983v1 [cs.CV] 3 Oct 2017
Yeong-Jun Cho, Su-A Kim*, Jae-Han Park*, Kyuewang Lee, Student Member, IEEE
and Kuk-Jin Yoon, Member, IEEE
Abstract—Person re-identification is the task of recognizing
or identifying a person across multiple views in multi-camera
networks. Although there has been much progress in person reidentification, person re-identification in large-scale multi-camera
networks still remains a challenging task because of the large
spatio-temporal uncertainty and high complexity due to a large
number of cameras and people. To handle these difficulties,
additional information such as camera network topology should
be provided, which is also difficult to automatically estimate,
unfortunately. In this study, we propose a unified framework
which jointly solves both person re-identification and camera network topology inference problems with minimal prior
knowledge about the environments. The proposed framework
takes general multi-camera network environments into account
and can be applied to online person re-identification in largescale multi-camera networks. In addition, to effectively show
the superiority of the proposed framework, we provide a new
person re-identification dataset with full annotations, named
SLP, captured in the multi-camera network consisting of nine
non-overlapping cameras. Experimental results using our person
re-identification and public datasets show that the proposed
methods are promising for both person re-identification and
camera topology inference tasks.
Index Terms—Person re-identification, Camera network topology inference, Non-overlapping camera network, Person reidentification dataset
I. I NTRODUCTION
Nowadays, a large number of surveillance cameras installed
at public places (e.g. offices, stations, airports, and streets)
are used to prevent untoward incidents, monitor localities,
or to track down evidences. In this context, multi-camera
networks have been studied to monitor larger areas or specific
individuals. Person re-identification, studied extensively over
the past decade, is the task of automatically recognizing
and identifying a person using multiple views obtained from
multi-camera networks. Nevertheless, the re-identification in
large-scale multi-camera networks still remains a challenging
task because of the large spatio-temporal ambiguity and high
complexity due to the large numbers of cameras and people.
It becomes even more challenging when camera views do
This work was done when all authors were in Gwangju Institute of
Science and Technology. * These authors contributed equally to this work.
Yeong-Jun Cho, Jae-Han Park and Kuk-Jin Yoon are with the School of
Electrical Engineering and Computer Science, Gwangju Institute of Science
and Technology, South Korea. Currently, Su-A Kim is with Intel Visual
Computing Institute Saarland Informatics Campus, Germany and Kyuewang
Lee is with ASRI, Department of Electrical and Computer Engineering,
Seoul National University. Kuk-Jin Yoon is a corresponding author. E-mail:
[email protected]
Cam
CamBB
Cam
CamAA
??
Large
Largenumber
numberofofcameras
cameras
Cam
CamCC
??
??
Large
Largenumber
number
of
ofobjects
objects
Cam
CamNN
(a) Spatio-temporal uncertainties
between cameras
(b) High complexity due to large
numbers of cameras and people
Fig. 1. Challenges of large-scale person re-identification
not overlap each other. As shown in Fig. 1 (a), spatiotemporal uncertainty due to the unknown geometrical relationship between cameras in a multi-camera network makes
re-identification difficult. Unless some prior knowledge about
the camera network is given, re-identification should be done
by thoroughly matching the person of interest with all the other
people who appear in the camera network within the given
time interval. This exhaustive search method is slow in general
and generates unsatisfactory results in many cases because
it is hard to find a correct match among a large number of
candidates, as there might be many people having appearances
similar to the person of interest.
Most of the previous studies [1]–[3] conducted an exhaustive search to re-identify people across multiple cameras,
relying solely on the appearance of people. These methods
work quite well when the numbers of cameras and people
are small, but cannot effectively handle the aforementioned
challenges in the large-scale problem illustrated in Fig. 1
(b). To resolve the complexity problem and to improve the
accuracy of the large-scale person re-identification, the number
of matching candidates should be constrained and reduced by
inferring and exploiting the spatio-temporal relation between
cameras, referred to as the camera network topology.
In recent years, several camera network topology inference
methods [4], [5] have been proposed for large-scale person reidentification. These methods infer the topology of a camera
network based on the simple occurrence correlation between
the people’s entry and exit events. However, no appearancebased validation is performed for topology inference. Hence,
the inferred topology can be inaccurate in crowded scenes.
Actually, it is a chicken-and-egg situation: the camera network topology can be more accurately inferred from the reidentification results while person re-identification can be im-
2
proved by utilizing the camera network topology information.
The main idea of this study is that the camera network topology inference and person re-identification can be achieved
jointly while complementing each other. Based on this idea, we
propose a unified framework, which automatically solves both
the problems together. To the best of our knowledge, this is the
first attempt to solve both the problems jointly. In the proposed
framework, we first infer the initial camera network topology
using only highly reliable re-identification results obtained
by the proposed multi-shot matching method. This initial
topology is used to improve the person re-identification results,
and the improved re-identification results are then used to
refine the camera network topology. This procedure is repeated
until the estimated camera network topology converges. Once
we estimate the reliable camera network topology in the
initialization stage, we utilize it for the online person reidentification and update the camera network topology with
time.
To sum up, we propose a multi-shot person reidentification method that exploits a time-efficient random
forest. (Sec. III-A). We also propose a fast and accurate camera
network topology inference method in Sec. III-B. In Sec. III-C,
we describe a proposed online person re-identification method.
The methods are effectively integrated in our framework.
It is important to note that our proposed framework runs
automatically with minimal prior knowledge about the environments. Besides the proposed methods, we also provide a
new Synchronized Large-scale Person re-identification dataset
named SLP (Sec. IV) captured in the multi-camera network
consisting of nine non-overlapping cameras. To validate our
unified framework, we extensively evaluate the performance
and compare it with state-of-the art methods.
Our previous study is [6]. Differences between the previous
study and this study are as follows. In this study, we propose
and validate online person re-identification after the camera
network topology initialization stage (Sec. III-C). During the
online person re-identification, the camera network topology
updates itself with time. We describe complete details of
the proposed methods and provide more experimental results
to explicitly validate our proposed methods. In Sec. VI, we
compare our proposed method with state-of-the-art methods in
terms of both quantitative and qualitative performances, and
show the results using another public dataset.
II. P REVIOUS W ORKS
Person re-identification methods can be categorized into
non-contextual and contextual methods as summarized in [7].
In general, non-contextual methods rely only on appearances
of people and measure visual similarities between people to
establish correspondences, while contextual methods exploit
additional contexts such as human pose, camera parameters,
geometry, and camera network topology.
A. Non-contextual Methods
In order to identify people across non-overlapping views,
most studies rely on appearances of people by utilizing
appearance-based matching methods with feature learning
or metric learning techniques. For feature learning, many
studies [1], [8], [9] have tried to design visual descriptors
to describe the appearance of people suitably. Regarding the
metric learning, several methods such as KISSME [10] and
LMNN-R [11] have been proposed and applied to the reidentification problem [2], [12], [13]. Some studies [10], [13]
extensively evaluated and compared several metric learning
methods and showed the effectiveness of the metric learning
for re-identification. In order to utilize more plentiful appearance information, various multi-shot matching methods exploiting multiple appearances were proposed in [1], [14], [15].
In addition, person re-identification methods based on deep
learning were proposed in [3], [16], which simultaneously
learn both features and distance metrics.
Although many non-contextual methods have improved the
performance of person re-identification, the challenges such as
spatio-temporal uncertainty between non-overlapping cameras
and high computational complexity still remain.
B. Contextual Methods
Several studies [17]–[19] using human pose priors have
been proposed recently. These methods calibrate cameras to
obtain camera parameters and estimate human poses. Person
re-identification with estimated human poses can then be
efficiently carried out. The methods mitigate ambiguities by
seeking various poses of people.
To resolve the spatio-temporal ambiguities, many studies
have tried to employ camera network topology and camera
geometry. Several studies [20]–[22] assume that the camera network topology information is available and show the
effectiveness of the topological information. However, the
topology is not available in the real-world scenario. Thus,
many studies have tried to infer the camera network topology
in an unsupervised manner. Makris et al. [4] proposed a
topology inference method that simply observes the entry
and exit events of the targets and measures the correlations
between the events to establish the camera network topology.
This method was applied and extended in [23]–[25]. Similarly,
Loy et al. [5], [26] also proposed topology inference methods
to understand multi-camera activity by measuring correlation
or common information across simple activity patterns.
The aforementioned topology inference methods [4], [5],
[24]–[26], so-called event-based approaches are practical since
they do not require any appearance matching steps such as reidentification or inter-camera tracking for topology inference.
However, the topology inferred by the event-based approach
can be inaccurate since the topology may be inferred from
false event correlations. Actually, false event correlation may
easily occur when people pass through blind regions irregularly. On the other hand, a few studies [27], [28] inferred
the topology using person re-identification results. These
methods tend to be more robust to noise than the eventbased approaches. However, they do not refine or update
the initially inferred topology with time while performing reidentification. Thus, these methods may not fully exploit the
topology information.
3
label 𝑁 𝑐𝐵
label 𝑗
…
…
1
…
…
2
𝑙
𝑐
𝐾𝑗 𝐵
1st
Tth
tree
tree
𝑐
𝑦|𝐯𝑗,𝑙𝐵
𝑐
𝑝𝑐𝐴 𝑦|𝐯𝑗 𝐵
𝑐
𝑝Τ𝐴
𝑐
𝑦|𝐯𝑗,𝑙𝐵
Ensemble each tree result
& multi-shot test result
𝑦
Zone-to-Zone (Sec III-B2)
topology inference
Iterative update
of person re-id and camera topology
Enhance person re-id accuracy
using topology information
CAM C
(Sec III-A)
Re-identifying
time
iterations
𝑝(∆𝑡)
𝜇last Random forest
time
2nd Random forest
…
Iterate until the topology converges
T
ID #32
ID #32
T
𝑝(∆𝑡)
… T
…
T
Updated topology
Initial topology
time
*Edge thickness: connectivity confidence
2T
Refine the camera topology
using updated re-id results
Blind area
(Sec. III-C)
Re-identify person using
online updated camera topology
searching reduced
matching candidates
0
CAM B
Online Person Re-identification
(Sec. III-B3)
CAM C
CAM B
CAM-to-CAM topology
Valid link
Zone
CAM A
Max probability
𝑦𝑖∗
(Sec III-B1)
Blind area
…
𝑐
𝑝1 𝐴
CAM-to-CAM
topology inference
Valid link
CAM
CAM A
Test index 𝑗: 𝐯𝑗,𝑙𝑐𝐵
CAM 𝑐𝐴
Offline Camera Network Topology Initialization
Zone to zone topology
Re-id based on Random Forest
CAM 𝑐𝐵
Update camera topology
using re-identified person
Fig. 2. The proposed unified framework for person re-identification and camera network topology inference.
III. P ROPOSED U NIFIED F RAMEWORK
2nd RF
Figure 2 illustrates the proposed unified framework for person re-identification and camera network topology inference.
In the proposed framework, we first train random forestbased person classifiers (Sec. III-A) for efficient person reidentification. Subsequently, we jointly estimate and refine
the camera network topology and person re-identification results (Sec. III-B) using the trained random forests. Finally, we
perform an online person re-identification using the inferred
camera topology (Sec. III-C).
Most of the previous studies mainly focused on enhancing
the re-identification performance. However, when handling a
large number of people, time complexity is also very important
for building a practical re-identification framework. To this
end, we utilize a random forest algorithm [29] for multi-shot
person re-identification and incorporate it into our framework.
We denote the k-th appearance of person i in the camera cA
cA
as vi,k
. A set of the appearances of people in the camera is
expressed as,
n
o
cA
D cA =
vi,k
, yi |1 ≤ i ≤ N cA , 1 ≤ k ≤ KicA ,
(1)
where yi is the label of person i (normally yi is set to i). N cA
is the number of people visible in camera cA and KicA is the
number of appearances of person i in camera cA . We then
train a random forest classifier using the appearance set DcA .
After the random forest classifier is trained, we can estimate
the probability distribution of classificationPby aggregating
T
decision tree outputs as: pcA (y|v) = T1 τ =1 pcτA (y|v),
cA
where pτ , T denote a trained decision tree and the number
of decision trees, respectively. Testing a person j (i.e. probe)
in another camera cB to the trained random forest classifier is
defined as,
p
cB
y|vj,l
T
1 X cA cB
p
y|vj,l .
=
T τ =1 τ
T
1st RF
time
2T
3rd RF
last RF
Fig. 3. Training a series of random forest classifiers based on the overlapping
time window T .
To obtain a multi-shot person re-identification result, we test
multiple appearances of each person and average the multiple
results as
A. Random Forest-based Person Re-identification
cA
0
4th RF
(2)
K
cB
j
1 X cA
cB
pcA (y|vjcB ) = cB
p (y|vj,l
),
Kj
(3)
l=1
where KjcB is the number of appearances
of the probe. From
the probability distribution pcA y|vjcB , we choose a final
matched label as:
yi∗ =
argmax
yi
∈{1,2,...,N cA }
pcA (yi |vjcB ).
(4)
As the result of the multiple appearance matching test, we
obtain a corresponding pair (vycA∗ , vjcB ) between the camera
i
cA and cB . Finally, we calculate a similarity score for the
matching pair by selecting the smallest matching score as per
the following equation [1],
− mink,l
S vycA∗ , vjcB = e
i
c
c
B
vyA
∗ ,k −vj,l
i
2
.
(5)
The similarity score lies on [0, 1] and it is used in Sec. III-B
for inferring the topology.
The tree structure of the random forest method makes
the multi-shot test very fast: its computational complexity is
O(N K log N )1 . On the other hand, the conventional exhaustive re-identification approaches comparing every person pair
1 When analyzing the computational complexity, we assume that the number
of objects in cA and cB are the same (i, j = {1, 2, .., N }) and each object
contains the same number of image patches (k, l = {1, 2, .., K}). Details
are described in APPENDIX.
4
0.0771638
0.56574
0.2
0.2
𝑐𝑜𝑛𝑓 𝑝 ∆𝑡
Fitted curve
0.1
𝑐𝑜𝑛𝑓 𝑝 ∆𝑡
= 0.07
50
150
Fitted curve
0.1
0.05
0.05
0
-50
0.15
y
y
0.15
= 0.56
0
50
100
150
200
250
time
300
0
-50
0
100
200
250
time
300
x
x
(a) CAM1 – CAM2
(b) CAM1 – CAM4
Fig. 4. Examples of estimated transition distributions with connectivity confidences
involve higher computational cost: O(N 2 K 2 ). Over and above
the superiority in terms of the computational cost, our method
also delivers a high person re-identification accuracy as shown
in Sec. VI-A1.
B. Camera Network Topology Initialization
Camera network topology represents spatio-temporal relations and connections between the cameras in the network. It
involves the inter-camera transition distributions between two
cameras, which denote the transition distributions of objects
across cameras according to time, and represent the strength
of connectivity between the cameras. In general, the topology
is represented as a graph G = (V, E), where the vertices V
denote the cameras and the edges E denote the inter-camera
transition distributions, as shown in Fig. 8 (b).
1) CAM-to-CAM topology inference: To build the CAMto-CAM topology, we first estimate the transition distributions
between the cameras. In this study, we estimate the transition
distributions based on the person re-identification results.
Before performing person re-identification, we select key
appearances among the multiple appearances of each person
based on the method in [30]. By selecting key appearances
uncorrelated with each other, we can efficiently reduce the
matching complexity and avoid over-fitting problems.
After the key appearance selection, we split a whole group
of people into several sub-groups according to their timestamps and train a series of random forest classifiers with
overlapping time window T as shown in Fig. 3. Next, we
search correspondences of people who disappeared in a camera
using the trained random forest classifiers of other cameras.
Initially, we have no transition distributions between cameras
to utilize. Hence, we consider every pair of cameras in the
camera network and search for correspondences within a wide
time interval. When a person disappears at time t in a certain
camera, we search the correspondence of the person from the
other cameras within time range [t − T, t + T ]. If there are
multiple random forest classifiers overlapping with the time
range, we test all random forest classifiers and select the most
reliable one among them as a correspondence. Thanks to our
overlapping search scheme, we can alleviate the risk of failure
in correspondence search.
When these initial correspondences are given, we only use
highly reliable correspondences to infer transition distributions
between cameras. We regard a correspondence as a reliable
one with a high similarity score when S(vycA∗ , vjcB ) > θsim .
i
The inference procedure of transition distribution is as follows:
(1) Calculate the time difference between correspondences and
make a histogram of the time difference. (2) Normalize the
histogram using the total number of reliable correspondences.
We denote the transition distribution as p (∆t). Figure 4 shows
two distributions: Fig. 4 (a) is obtained from a pair of cameras
having a strong connection, and Fig. 4 (b) from a pair of
cameras having a weak or no connection.
Connectivity Check
Based on the estimated transition
distributions, we automatically identify whether a pair of
cameras is connected or not. We assume that the transition
distribution follows a normal distribution if there is a topological connection. Based on this assumption, we fit a Gaussian
model N (µ, σ 2 ) to the distribution p (∆t) and measure the
connectivity of a pair of cameras based on the following
observations:
• Variance of p (∆t): In general, most of the people reappear
after a certain transition time µ; therefore, the variance in
the transition distribution (σ 2 ) is not extremely large and the
distribution shows a clear peak.
• Fitting error: Even though the distribution is obtained from
a pair of cameras with a weak connection, the variance in the
distribution can be small and the distribution can have a clear
peak due to noise. In order to measure the connectivity robust
to noise, we consider the model fitting error E (p (∆t)) ∈ [0, 1]
calculated by R-squared statistics.
Based on the above observations, we newly define a connectivity confidence between a pair of cameras as
conf (p (∆t)) = e−σ · (1 − E (p (∆t))) .
(6)
The connectivity confidence lies on [0, 1]. We regard a pair
of cameras as a valid link when conf (p (∆t)) > θconf .
Compared to a previous method [4], which only takes the
variance of a distribution into account, our method is more
robust to noise of distributions. Using the defined confidences,
we check every pair of cameras and reject invalid links as
shown in Fig. 8. We can see that many camera pairs in the
camera network have weak connections. Therefore, we can
greatly reduce computation time and save resources. Only the
valid pairs of cameras proceed to the next step.
Through this topology inference procedure, CAM-to-CAM
topology is represented as
Gcam = (Vcam , Ecam ) ,
Vcam ∈ {ci |1 ≤ i ≤ Ncam } ,
Ecam ∈ pij (∆t)|1 ≤ i ≤ Ncam , 1 ≤ j ≤ Ncam , i 6= j ,
(7)
where Ncam is the number of cameras in the camera network,
ci is a ith camera, and pij (∆t) is a transition distribution
between ci and cj .
acc
0.235
0.5474
0.6631
5
0.897959
0.4
𝑐𝑜𝑛𝑓 𝑝 ∆𝑡
1
= 0.63
0.4
1
𝑐𝑜𝑛𝑓 𝑝 ∆𝑡
= 0.86
0.3
y
y
0.2
0.2
0.2
0.1
0.1
0.1
0
-20
0
20
40
60
time
80
= 0.93
0.3
y
0.3
𝑐𝑜𝑛𝑓 𝑝 ∆𝑡
0.4
0
-20
0
x
(a) Iteration 1
20
40
60
time
80
0
-20
x
0
20
40
60
time
80
x
(c) Iteration
5
(b) Iteration 2
Fig. 5. Iterative update of the topology and re-identification results. Re-id accuracy is improved from 23.5% to 66.31%.
2) Zone-to-Zone topology inference: In this step, we estimate transition distributions between zones in cameras and
build a Zone-to-Zone topology. For each camera, a set of entry
and exit zones is automatically learned by [31]. Note that we
only consider exit-to-entry zone pairs when two zones belong
to different cameras. Other pairs of zones such as exit-to-exit,
entry-to-entry, and entry-to-exit are not considered.
A person who disappeared at an exit zone at time t is likely
to appear at an entry zone in a different camera within a certain
time interval T . Therefore, we search the correspondence
of the disappeared person from the entry zones in different
cameras within the time range [t, t + T ]. Similarly, as in
Sec. III-B1, we train a series of random forest classifiers for
each entry zone and measure connectivity confidences of all
possible pairs of zones using only reliable correspondences.
Through this step, many invalid pairs of zones between the
cameras can be ignored, and we can formulate Zone-to-Zone
topology as:
Gzone = (Vzone , Ezone ) ,
Vzone ∈ ci(k) |1 ≤ i ≤ Ncam , 1 ≤ k ≤ Zi ,
Ezone ∈{pij
(kl) (∆t)|1 ≤ i ≤ Ncam , 1 ≤ j ≤ Ncam ,
(8)
Algorithm 1: Camera network topology initialization
Data: Image sequences and object identities in each view
during initialization stage
Result: Re-identification result (people correspondences),
Camera network topology
for every camera pair do
Gcam = CAM-to-CAM topology inference;
if conf pij (∆t) > θconf then
Gzone = Zone-to-Zone topology inference;
if conf pij
(kl) (∆t) > θconf then
while pij
do
(kl) (∆t) converging
ij
ij
1
T ij
T(kl) =
ij
(kl)U − T(kl)L
1−E p(kl) (∆t)
Gzone = Topology update;
end
end
end
end
i 6= j, 1 ≤ k ≤ Zi , 1 ≤ l ≤ Zj }
where Ncam is the number of cameras in the camera network,
Zi is the number of zones of ith camera. ci(k) is a kth zone of
the ith camera. pij
(kl) (∆t) is a transition distribution between
ci(k) and cj(l) . In the next section, we iteratively update the
valid links between zones and build a camera topology map
of the camera network.
3) Iterative update of person re-identification
and camera network topology: After conducting the Zone-toZone topology inference, we obtain an initial topology map
between every pair of zones in the camera network. However,
the initial topology map can be inaccurate, because it is inferred from noisy initial re-identification results. Although we
consider only highly reliable re-identification results to infer
the initial topology, the re-identification results are erroneous
because we cannot exploit the topology information initially
for re-identification.
As mentioned above, camera network topology information and re-identification results can complement each other.
The inferred topology of the camera network can enhance
the person re-identification performance, and the person reidentification can assist the topology inference. Therefore, we
update person re-identification results and the camera network
topology in an iterative manner. Assuming that we have a
transition distribution p (∆t) ∼ N (µ, σ 2 ) between two zones.
Each zone belongs to different cameras. Then, the iterative
update procedure is as follows:
•
Step 1. Update the time window T . The initial time
window T was set quite wide, but now we can narrow it
down based on the inferred topology. First, we find the
lower and upper time bounds (TL , TU ) of the transition
distribution p (∆t) with a constant R as,
p (TL ≤ ∆t ≤ TU ) =
R
.
100
(9)
We set R as 95 following 3-sigma rule in order to cover
a majority of the distribution (95%) and ignore some
outliers (5%). Then, using the time bounds obtained, the
time window T is updated as,
T =
1
(TU − TL ) ,
1 − E (p (∆t))
(10)
where E (p (∆t)) is a Gaussian model fitting error rate.
When the fitting error is high, the time window T
becomes large. Thanks to our update strategy, we can
avoid the over-fitting of the topology during the update
steps.
6
Algorithm 2: Online person re-identification and
camera network topology update
Data: Image sequences, Object identities in each view,
Camera network topology
Result: Re-identification result (people correspondences),
Camera network topology
TABLE I
L IST OF PUBLIC PERSON RE - IDENTIFICATION DATASETS . W E REFERRED
TO THE LIST OF PERSON RE - IDENTIFICATION DATASETS IN [32].
dataset
VIPeR [33]
QMUL iLIDS [34]
GRID [5]
CAVIAR4ReID [35]
3DPeS [36]
PRID2011 [37]
V47 [38]
WARD [39]
Softbio [40]
CUHK 2 [41]
CUHK 3 [42]
for every zone
pair do
if conf pij
(kl) (∆t) > θconf then
Find correspondences;
p0ij
(kl) (∆t) = Transition distribution update;
P
0ij
> 0.1 then
(∆t)
(∆t)
−
p
if ∆t pij
(kl)
(kl)
% Fit Gaussian model to p0ij
(kl) (∆t)
pij
(kl) (∆t)
=
p0ij
(kl) (∆t)
year # IDs # cam
NLPR MCT [43]
2
∼ N (µ, σ );
end
end
end
Step 2. Re-train a series of random forests of an entry
zone with the updated time window T as shown in the
Fig. 3.
• Step 3. Find correspondences of the disappeared people
at an exit zone. Based on the topology, a person who
disappeared at time t at the exit zone is expected to appear
around the time (t + µ) at the entry zone of the other
camera. Using the topological information, we search
the correspondence of the person from a trained random
forest having the center of time slot close to (t + µ).
• Step 4. Update the transition distribution using reliable correspondences with a high similarity score
S(vycA∗ , vjcB ) > θsim .
i
This procedure (Step 1 – Step 4) is repeated until the
transition distribution converges. We perform the procedure
for all Zone-to-Zone pairs in the camera network topology
Gzone . Consequently, the camera topology map in the camera
network is refined. If the topology does not change anymore
or shows subtle changes during several iterations, we stop the
procedure. The above procedure improves the performance
of re-identification as well as the accuracy of the topology
inference, as shown in Fig. 5. The overall algorithm of the
proposed camera network topology initialization process is
summarized in Algorithm. 1.
RAiD
iLIDS-VID
Market1501
PRW
MARS
DukeMTMC4ReID [48],
Airport
[44]
[14]
[45]
[46]
[47]
[49]
[32]
SLP
2007 632
2009 119
2009 250
2011
72
2011 192
2011 934
2011
47
2012
70
2012 150
2013 1816
2014 1360
235
2014
255
2014
43
2014 300
2015 1501
2016 932
2016 1261
2017 1852
2017 9651
2017 2632
2
2
8
2
8
2
2
4
8
2
2
3
3
4
2
6
6
6
8
6
9
multi back
traj. sync
-shot ground
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
•
C. Online Person Re-identification
and Camera Network Topology Update
Once we estimate the reliable camera network topology in
the initialization stage, we utilize it for the online person reidentification and online camera network topology update. The
proposed online process is as follows:
• Step 1. Pick out matching candidates of a person based on
the initialized camera network topology. When a person
disappeared at a time t, a set of people appearing at a
different camera within t + TL ≤ t + µ ≤ t + TU are the
matching candidates. The lower and upper time bounds
•
•
•
(TL , TU ) are evaluated using the method described in
Sec. III-B3.
Step 2. Find a correspondence among the matching
candidates. For efficiency, we only train random forest
classifiers where the number of matching candidates
is large unless we find the correspondence based on
exhaustive search.
Step 3. Update the transition distribution. We use the
matching result with a high similarity score (> θsim ) to
update the transition distribution, adding the new matching result to p(∆t) and normalizing the distribution. The
updated transition distribution is represented as p0 (∆t).
Step 4. Fit a new Gaussian model to the updated transition distribution p0 (∆t). The cost of model fitting is
quite high. Hence, we do not always fit the Gaussian
model, but only when the difference between the initial
distribution p (∆t) and the updated distribution p0 (∆t) is
sufficiently large.
P To this end, we measure the distribution
difference as ∆t (|p (∆t) − p0 (∆t)|) and fit the model
when the value is larger than 0.1 (i.e., 10 % variation).
Unless the value is larger than the threshold, p0 (∆t)
merely cumulates the update.
We perform the aforementioned procedure for all Zone-toZone pairs in Gzone . The overall algorithm of the proposed
online process is summarized in Algorithm. 2.
The online topology update step is necessary and useful
for the following reasons: First, the pose of a camera can
change due to vibration, heavy wind, and so on. Even a slight
change in pose can cause a considerable viewpoint change, and
it makes the initialized topology inaccurate. Second, people
transition time between cameras can change temporarily due
to many factors such as sidewalk conditions, weather, and
unusual movement of crowd (e.g. parade, march). The online
topology update can effectively handle the above issues.
7
TABLE II
D ETAILS OF OUR NEW DATASET: SLP.
Index
# ID
# frames
# annotated box
Resolution
Duration
Building B
CAM 1
CAM 2
CAM 3
CAM 4
CAM 5
CAM 6
CAM 7
CAM 8
CAM 9
Total
256
661
1,175
243
817
324
516
711
641
2,632
19,545
65,518
104,639
41,824
78,917
79,974
93,978
53,621
42,347
580,363
47,870
205,003
310,262
65,732
307,156
160,367
78,259
176,406
117,087
1,468,142
720×480 1920×1080 1920×1080 1440×1080 1440×1080 1440×1080 1920×1080 1920×1080 1920×1080
–
2h 13m
2h 12m
2h 22m
2h
2h 21m
2h
2h 38m
2h 29m
2h 28m
–
2 3 C4
1 4
C6
3
4
2
1 6 5
2
1
3
Building A C2 1 2
3
5 4
C1
1
C1
51 2
4
3
C5
C3
C7 Building C
3 4
C8
2 1
C9
C2
C3
C4
C5
C6
C7
C8
C9
2 3
1
Sidewalk
2
1
2
(a) Layout of a camera network
(b) Example frames of nine cameras (CAM1 – CAM9)
CAM
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
256
227
206
0
23
28
19
1
1
227
661
571
2
82
84
43
1
1
206
571
1175
98
568
90
168
57
56
0
2
98
243
155
0
37
16
16
23
82
568
155
817
61
109
55
54
28
84
90
0
61
324
0
0
0
19
43
168
37
109
0
516
281
269
1
1
57
16
55
0
281
711
633
1
1
56
16
54
0
269
633
641
(c) Numbers of true matching pairs
Fig. 6. A new synchronized large-scale person re-identification dataset: SLP.
Thanks to the accurate topology initialization (Sec. III-B)
and the online procedure (Sec. III-C), the camera network
topology can be maintained automatically with time. When
a new camera is added to an existing camera network during
the online stage, the camera network topology initialization
operates in parallel and the topology is updated efficiently.
IV. N EW P ERSON R E - IDENTIFICATION DATASET: SLP
To validate the performance of person re-identification
methods, numerous datasets have been published. Table. I
shows a list of the public person re-identification datasets
with their various characteristics, which have been constructed
for specific scenarios. For example, VIPeR [33] was constructed by two cameras and contains 632 people. Each camera
provides one single image for one person. On the contrary,
Softbio [40] includes 150 people captured from eight
cameras. In this dataset, each camera provides multiple images
for a person (multi-shot). However, despite the availability of
numerous published datasets, none of them reflect practical
large-scale surveillance scenarios, in which (1) multiple observations (multi-shot) and moving paths (traj.) for each person
from each camera are available, and (2) both the number of
people (# IDs) and the cameras (# cam) is large.
Most of the public datasets include a small number of people (# IDs < 200) [34]–[36], [38]–[40], [44] or cameras (# cam
< 5) [14], [33]–[35], [37], [38], [41]–[44] as shown in Table. I.
Some datasets provide just a single shot of each person [5],
[33]. Moreover, many datasets do not provide the required
annotation with information about people (traj.) such as their
movement paths [32], [34]–[36], [38], [41], [42], [45], [46] but
only give cropped multi-shot appearances. Furthermore, there
are only two datasets that provide camera synchronization
information or time stamps of all frames (sync) [43], [48].
Recently, M. Gou et al. [48] provided an re-identification
dataset named DukeMTMC4ReID. The dataset covers real
large-scale surveillance scenarios (e.g., synchronized cameras,
large number of cameras and people). However, it does not
give the trajectory of each person but only gives cropped multishot images. If the dataset provides additional information
such as background and traj., it will be very helpful for other
researchers and can be fully exploited.
In this study, we provide a new Synchronized Large-scale
Person re-identification dataset called SLP constructed for
practical large-scale surveillance scenarios. The main characteristics of our dataset are as follows:
• The number of people (# IDs): The total number
•
•
•
•
of people in the dataset is 2,632. Among them, 1,737
people appear in at least two cameras. The details of the
matching pairs are shown in Fig. 6 (c).
The number of cameras (# cam): The dataset was
captured from nine non-overlapping cameras. The layout
of the camera network and the example frames from nine
cameras are shown in Fig. 6 (a-b).
Multiple appearances availability (multi-shot): SLP
provides multiple appearances of each person. In this version, we do not provide the original image appearances
but provide pre-computed feature descriptors due to legal
restrictions. However, we expect that we will provide the
original image appearances of people in the future.
Scene background availability (background): SLP provides the scene background. Although it only contains
the background of the camera scene, it contains scene
structures such as buildings, ground plane, orthogonal
lines etc., which are quite useful for understanding the
scene.
People trajectory availability (traj.): The ground truth
tracking information of every person is available. It includes positions (x,y locations) and sizes (height, width)
of people in entire image sequences. To annotate and
generate people’s trajectories for all video sequences, we
2
TP
0
FP
(a)
8
-2
used an annotation tool ‘VATIC’ [50]. The total number
0.5
0.6 is 1,468,142.
0.65
0.7
0.75
of0.55
annotated boxes
• Camera synchronization (sync): Each camera is synchronized with a global timestamp. It provides all times2
tamps of video frames.
• Camera calibration: Each camera is calibrated and
camera parameters are given.
0
• Details of our dataset are summarized in Table II.
As mentioned above, our new dataset enables us to test
-2
person re-identification methods under practical large-scale
0.5surveillance
0.55 scenarios.
0.6 Compared
0.65 to previous
0.7 datasets,
0.75SLP
not only provides competitive number of person IDs and cameras but also additional information such as full annotations,
camera synchronization and camera parameters. In addition,
SLP is also applicable to other computer vision tasks such as
multi-object tracking, multi-camera activity recognition, and
so on. It is available on online: https://sites.google.com/view/
yjcho/project-pages/re-id topology.
0.5
0.6
0.7
PRID2011 [19]
TP
FP
0.5
0.6
0.7
V. S ETTINGS AND M ETHODOLOGY
A. Experimental Settings
Since we mainly focus on person re-identification and
camera topology inference problems, we assume that person
detection and tracking results are given. Most person reidentification researches follow this assumption and setting.
In order to evaluate each proposed method, we divided our
dataset into two subsets according to time. The first subset
contains 1-hour data starting from the global start time (AM
11:20). It is used in an initialization stage. The latter subset
including the remaining data is utilized in an online test stage.
In this study, every person image is resized to 128×48
pixels. We used the LOMO feature [51] to describe the
appearances of people. It first divides the image into six
horizontal stripes. Then it extracts an HSV color histogram
and builds descriptors based on Scale Invariant Local Ternary
Pattern (SILTP) from each stripe. The final descriptor from
the 128×48 image has 26,960 dimensions.
It is worthy to note that the proposed framework does
not utilize pre-matched results between cameras. In general,
many conventional person re-identification methods exploit
the pre-matched pairs for learning distance metrics such as
LMNN [12], ITML [2], KISSME [10] and Mahal [13]. The
metric learning methods are effective but not appropriate for
practical re-identification systems since the matched results
are not given at first. On the other hand, the proposed
framework runs fully automatically, providing the person
detection/tracking results of each camera.
B. Evaluation Methodology
In general, many previous works plot a Cumulative Match
Curve (CMC) [33] representing true match being found within
the first n ranks for evaluating person re-identification performances. In practice, however, all ranks (2,3,...,n) except rank1 failed to find the correct correspondences. For this reason,
rank-1 value is the most important one among all ranks and
several studies mainly measure the rank-1 accuracy. In this
occurrence
2
0.8
0.85
Similarity
TP
0
FP
0.9
3DPeS
-2
0.5
0.5
0.55
0.6
0.6
0.65
0.7
0.7
0.75
0.8
0.85
Similarity
2
TP
10
10
0
PRID2011
-2
0.55
0.6
0.6
0.65
0.7
0.7
0.75
Similarity
0.8
0.85
occurrence
15
5
5
00
00
00
00
0.9
Fig. 2
10
10
0.9
5
FP
0.5
0.5
15
5
0.2
0.4 0.6 0.8 conf
0.2 0.4 0.6 0.8
1
0.2
0.2 0.4
0.4 0.6
0.6 0.8
0.8 conf
1
(a)
(b)
Fig. 7. Observations for selecting reliable parameters
Similarity
0.8
0.85
0.9
study, we measure the rank-1 accuracy by
TP
Tgt ,
where T P is
the number of true matching results and Tgt is the total number
of ground truth pairs in the camera network. The evaluation
metric rank-1 accuracy reflects the overall re-identification
performance of the camera network.
To evaluate the accuracy of the camera network topology, we adopt two evaluation metrics: transition time
error and topology distance. When an inferred transition distribution and a ground truth
are given as
2
(p(∆t)∼N µ, σ 2 , pgt (∆t)∼N µgt , σgt
), the transition
time error is simply measured as |µ − µgt |. We defined the
topology distance as Bhattacharyya the distance that measures the difference
two probability
R between
distributions as
p
dB (p, pgt )= − ln
p(∆t)pgt (∆t) d∆t . If there are multiple links in the camera network, we measure the evaluation
metrics for all links and average them to get the final transition
time error and topology distance.
C. Parameters Selection
We used fixed parameters for all steps in the Sec. III
and applied the same parameters to all datasets SLP, NLPR
MCT [43]. We found that the performance of our framework
does not depend on each parameter. The parameters are fixed
as follows:
• θsim : To determine a proper θsim , we tested several other
datasets (3DPeS [36], PRID [37]) and analyzed the results.
Figure 7 (a) shows re-identification similarity scores of each
dataset. Marks ◦, ∗ denote true and false positives. To estimate
the accurate topology, we need to exclude false positives.
Based on the test results, we strictly set θsim as 0.7 for
selecting only true positives.
• θconf : Figure 7 (b) shows a histogram of calculated connectivity confidences between nine cameras (36 pairs) in the
CAM-to-CAM topology inference step. Weak connections
are easily classified, since they have very low confidences
(0 − 0.25). Based on this, we set θconf as 0.4 and fixed the
value for topology inference steps and all datasets.
• Initial T : We tested various T such as 300, 450, 600, which
are sufficiently wide to find correspondences under unknown
camera networks. We noticed that the results of topology
inference are not significantly affected by T . We set the initial
T as 600.
* The proposed framework has only one parameter (T ) to be
adjusted depending on the environments. However, T does not
critically affect the performance of the proposed system.
9
TABLE III
I NFERRED VALID Z ONE - TO -Z ONE LINKS AND GROUND - TRUTHS .
Exit
CAM1-ZONE1
CAM2-ZONE2
CAM3-ZONE2
CAM3-ZONE3
CAM4-ZONE4
CAM7-ZONE1
CAM8-ZONE1
Entry
CAM2-ZONE5
CAM3-ZONE1
CAM5-ZONE6
CAM7-ZONE3
CAM5-ZONE2
CAM8-ZONE2
CAM9-ZONE2
PE-LARSCALE
600초
CAM
µ
34.4
36.7
-0.42
4.8
30.2
28.2
11.6
µgt
34.7
36.3
-0.57
4.3
30.1
28.4
11.7
σ
6.25
8.03
3.49
4.8
13.4
21.3
4.82
PE-LARSCALE
σgt
6.04
5.79
3.23
3.5
12.5
6.36
4.24
Exit
CAM2-ZONE5
CAM3-ZONE1
CAM5-ZONE6
CAM7-ZONE3
CAM5-ZONE2
CAM8-ZONE2
CAM9-ZONE2
Entry
CAM1-ZONE1
CAM2-ZONE2
CAM3-ZONE2
CAM3-ZONE3
CAM4-ZONE4
CAM7-ZONE1
CAM8-ZONE1
µ
40.4
37.6
0.70
3.75
39.5
31.9
10.5
µgt
40.4
37.0
1.59
4.68
28.6
30.0
10.5
σ
7.62
10.3
3.43
2.16
3.82
2.41
4.03
σgt
5.93
8.90
2.32
3.04
14.8
4.02
4.08
For all the camera pairs, we illustrate a color map of
estimated
CAM-to-CAM connectivity confidence in Fig. 8(a).
2
3
Each
row
and
column indicates the index of the camera. When
4
5
5
the confidence value is greater than θconf , we regard the
5
2
8
5
2
8
6
corresponding camera pair as a valid link. As a result, the valid
7
8
camera links are drawn as Fig. 8(b). Each vertex indicates
3
6
6
1
6
1
9
9
9
the index of the camera and valid links are represented by
(a)7Connectivity confidences
(b) Ours
(c) DNPR [28]
edges. Unfortunately, CAM6 failed to be linked to CAM5. That
Ours
DNPR
is
MCT because the size of the person image patches is very small
Fig. 8. Results of CAM-to-CAM connectivity check. (a) Connectivity
due to the long distance from the camera. Therefore, it is
confidence
map.
(b),
(c)
Valid
camera
links
of
two
methods
(solid
line:
true
8
link, dotted line: missing).
hard
to distinguish the appearances of people. In addition,
2
3
CAM6
is quite isolated from other cameras as we can see
9
125
0.8
125
0.8
500
0.8
500
0.8
500
0.8
500
0.8
500
0.8
500
0.8
125
0.8
125
125
0.8
0.8
500
0.80.8
125
0.8
500
0.8
125
0.8
0.8
125
125
0.8
0.8
500
500
0.8 from0.8the
125
0.8
0.8
125
0.8
125
125
0.8
125
matching pairs of CAM6 in the Fig. 6(c). Compared to
500
500
0.8
sec
sec
sec
sec
Accumulated
time
Accumulated
timeComputational
Accumulated
time
Accumulated
time
Accumulated
time
Accumulated
time
Accumulated
time
Accumulated
time
Accumulated
time
Accumulated
time
Accumulated
Accumulated
time
Accumulated
Accumulated
time
Computational
time
Computational
time
time
Computational
time
Computational
time
Computational
time
Computational
time
Computational
time
Computational
time
Computational
time
1 0.80.8
500
500
450
450
450
450
450
450
Computational
time
450
450
Computational
time
450
450
450
Accuracy
Accuracy
0.7
0.7
0.7
0.7
0.7
Accuracy
Accuracy
0.7
Accuracy
Accuracy
Accuracy
Accuracy
0.7
Accuracy
Accuracy
0.7
0.7
0.70.7
0.7
0.7
0.7
0.70.7
0.7
0.7
0.7 450
0.7 DNPR
Accuracy
0.7
0.7
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
Accuracy
0.7
Accuracy
Accuracy
0.7
[28],
which also infers the CAM-to-CAM topology, our
0.7
0.7
Accuracy
Accuracy
Computational
time
Computational
time
450
450
100
100
400
400
400
400
400
100
400
400
100
100
100
400
100
400
100
100
400
100
100
100
100
100
400
400
Accuracy
Accuracy
0.70.7
Accuracy
Accuracy
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
method
provides
more accurate results. Our proposed method
0.6
0.6
0.6
400
0.6
400
350
350
Accumulated
time
350
350
Accumulated
time
350
350
350
0.6
350
350
0.6
350
350
350
0.60.6
350
350
7575
300
300
300
300
300
300
75
300
300
75
7575
75
300
75
75
300
75
75
7575
75
300
300
0.5
0.5
an accuracy of 87.5%, finding 7 links out of 8. The
0.5
0.5
0.5
0.5
0.5
0.5
0.50.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5 provides
0.50.5
0.5
0.5
0.50.5
300
300
250
250
250
250
250
250
250
250
250
250
250
250
0.50.5
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4 DNPR
0.4
0.4
0.4
0.4
0.4
0.40.4
0.4
0.4200
0.4
0.4
0.4
0.4
0.40.4 [28] method provides an accuracy of 62.5%, finding 5
0.4
0.4
5050
250
250
200
200
50
200
50
5050
50
50
50
200
200
200
200
200
200
200
50
200
50
5050
50
0.40.4
200
200
150
150
links
out of 8.
150
150
150
150
150
150
150
150
150
150
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.30.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.3
0.30.3
1
1
2
3
4
5
6
7
8
9
4
conf
1
1
1
4
0.9 0.9
3
7
4
3
7
0.8
0.8 0.8
0.7 0.7
0.6
0.6 0.6
0.5 0.5
0.4
0.4 0.4
0.3 0.3
0.2
0.1 0.1
100100
75 75
50 50
2525
150
150
100
100
100
100
100
100
100
100
100
100
100
100
0.2
0.2
0.20.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
100
100
50
50
5050505050
50
50
50
0.80.8
Accuracy of person re-identification
Accuracy of person re-identification
400
400
erson
son
person
rson
125125
Accuracy of person re-identification
Accuracy of person re-identification
25
25
2525
25
25
25
25
25
2525
25
00
Accuracy of
of person
person re-identification
re-identification
Accuracy
AccuracyComputational
of person
person re-identification
Accuracy
time
Accuracy of
of person re-identification
re-identification
Computational
time
AccuracyComputational
of person re-identification
time
time
Accuracy
of
person
re-identification
re-identification
AccuracyComputational
of
person
re-identification
Computational
time
Computational
time
Computational
time
Accuracy
of
re-identification
Accuracy
of person
person re-identification
re-identification
Computational time
Accumulated computational time
Accumulated
Accumulated computational
computational time
time
Accumulated computational
time
computational time
time
Accumulated
computational
Accumulated computational
time
computational
time
Accumulated
computational
time
Accumulated computational time
0
Accumulated computational time
Accumulated computational time
00
2
Accumulated computational time
Accumulated computational time
1
0.2 0.2
0.70.7
0.60.6
0.50.5
0.40.4
0.30.3
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.20.2
0.3
0.3
0.20.2
0 0
0.1
0.1
00.1
0.1
00.1
0.1
0.1
00050
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.10.1
0.1
0.1
0 0 00050
0.10.1
00 000 0000000
0.1
0.1
0.1
0.1
CAM
ZONE
Iter3
Iter4
Iter5
CAM
ZONEIter1
Iter1Iter2
Iter2
Iter3
Iter4
Iter5
Ours
Exhaustive
Ours
Exhaustive
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
Ours
Ours
Base0.20.2
CAM
CAM
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
CAM
ZONE
Iter1
Iter2
Iter3
Iter4
Iter5
Ours
Base
Ours
Base
Ours
Base Base
Ours
Base
Ours
Base
Ours
Base
Ours
Base
Ours
Base
Ours
Base
Ours
Base
0 0
0.10.1
Ours
Base
Ours
Base
0 0
0.10.1
25 25
(a)
(b)
CAMCAM
ZONEZONE
Iteration
Iteration
Iteration
Iteration
Iteration
CAMCAM
ZONEZONE
Iteration
1 1Iteration
2 2Iteration
3 3Iteration
4 4Iteration
5 5
Fig. 9. Results of the person re-identification through the camera topology
initialization.
VI. E XPERIMENTAL R ESULTS
A. Evaluations using Our Dataset: SLP
1) Camera Network Topology Initialization: As explained
in Sec. III, we first inferred the CAM-to-CAM topology between all cameras. We then found the Zone-to-Zone topology
and iteratively updated the topology for the topology initialization. During the topology initialization process, we evaluate
both tasks: person re-identification and topology inference.
CAM-to-CAM topology inference result
In order to check the connectivity between two cameras,
person re-identification can be carried out bi-directionally.
However, for the sake of efficiency, we only used unidirectional re-identification results to check the connectivities
between the cameras. To this end, we assigned a camera
containing more people as a training camera (gallery) and a
camera containing fewer people as a testing camera (query).
That is because the problem of reidentifying a small set of
people from a large set is easier than the opposite case. In
addition, we can reduce the computational cost for checking
the connectivity by half.
0.20.2
Camera
topology initialization results
0.10.1
Figure 9 (a) represents the accuracy of person reidentification in each of the proposed initialization steps such
as CAM-to-CAM, Zone-to-Zone, and iterative update steps
(Sec. III-B1–III-B3). The accuracy of person re-identification
is 28.54% at the beginning, but it is consistently improved
by using inferred and refined camera topology information.
As a result, our method reaches 62.55% accuracy at the last
step of the topology initialization. In addition, it took only
112.46 seconds to conduct both person re-identification and
camera topology inference tasks with a large number of people
in the nine cameras (using Intel i7 CPU in MATLAB). Figure 10 shows the connectivity confidences between the linked
camera pairs. The connections become stronger during the
iterations. In the final iteration, most connections demonstrate
a high connectivity confidence (conf (p (∆t)) > 0.7). On the
other hand, a connection from Exit (CAM4-Zone4) to Entry
(CAM5-Zone2) is not a strong connection, since a small
number of people moved from CAM4 to CAM5 during the
topology initialization period (AM 11:20 – PM 12:20).
Figure 9 (b) shows the comparison of our proposed approach with a conventional approach called an Exhaustive
approach, which fully compares the multiple appearances
of people and exhaustively searches the correspondences of
people between the entry/exit zones without using camera
topology information. For the baseline of the exhaustive
method, we used a LOMO feature [51] descriptor. It shows
a low performance (52.06% person re-id accuracy) compared
10
C6
3
4
2
1 6 5
2
1
3
C2
C1
1
5
2
0.9
0.8
0.8
4
2
1 6 5
2
1
3
0.6
0.6
0.5
0.4
0.4
0.3
0.2
0.2
0.1
2 3
1
C2
C1
2
1
1
1
5
2
C9
2 3
1
3
4
4
2
1 6 5
2
1
3
C8
3 4
2 1
0
1
C2
C1
2
1
1
2
51 2
4
3
3
C5
C3
C7
0
C9
C6
2 3 C4
1 4
51 2
4
3
3
0.7
C8
3 4
2 1
3
4
1
C5
C3
C7
C6
2 3 C4
1 4
1
51 2
4
3
Confidence
2 3 C4
1 4
5
C5
C3
C7
3 4
2 1
2
3
4
C8
C9
2 3
1
2
1
1
2
2
(a) Initial
(b) Iteration 1
(c) Iteration 5
Fig. 10. The iterative update of connectivity confidences. The connectivity of each link has strengthened during the iterations. (best viewed in color).
TABLE IV
P ERFORMANCE COMPARISON WITH OTHER METHODS IN OFFLINE INITIALIZATION STAGE
Makris’s [4]
49.4
28.52
0.07
0.06
DNPR [28]
40.9
–
Cai’s [27]
47.0
4.84
LOMO [51] – Exhaustive
52.1
11.76
0.25
1
1
1
0.2
0.8
0.8
0.8
0.15
0.6
0.6
0.6
Ours
62.5
1.21
0.05
0.04
0.03
0.4
0.1
0.4
0.4
0.02
0
0
0
20
20
40
40
60
60
80
80
0
100
100
0.08
0
0
20
20
40
40
60
60
80
80
0
100
100
0.3
0.04
0.2
0.02
0.1
0
0
20
20
40
40
60
60
80
80
0
100
100
0.2
0.2
0
0
20
20
40
40
60
60
80
80
0
100
100
0
0
20
20
40
40
60
60
x
80
80
100
0
100
1
1
1
0.8
0.8
0.8
0.6
0.4
0.06
0
0.2
0.05
0.01
CAM8 – ZONE 1
↓
CAM9 – ZONE 2
Chen’s [25]
50.5
16.29
y
CAM3 – ZONE 3
↓
CAM7 – ZONE 3
Nui’s [23]
50.0
23.94
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0
0
0
20
20
40
60
80
100
0
20
40
60
80
100
0
20
40
60
80
100
0
40
60
80
100
y
nal version - connection
Performances \ Methods
Rank-1 accuracy (%)
Transition time error (sec)
0
0
20
20
40
40
60
60
80
80
100
100
0.2
0
0
20
20
40
40
60
60
80
80
100
100
0.2
0
0
20
20
40
40
60
60
x
80
80
100
0
100
2.5
1
1
2
0.8
0.8
1.5
0.6
0.6
0
20
40
60
80
100
0.8
0.15
0.6
0.1
y
CAM8 – ZONE 2
↓
CAM7 – ZONE 1
0.4
0.05
al version – no connection
0
0.2
0
0
20
20
40
40
60
60
80
80
100
0
100
0
0
(a) Makris et al. [4]
80
80
0
100
100
0
0
(b) Nui et al. [23]
20
20
40
40
60
60
80
80
0
100
100
0.4
0.2
0
0
20
20
(c) Chen et al. [25]
40
40
60
60
x
80
80
100
0
100
0
(d) Ours
1
1
0.08
0.8
0.8
0.06
0.6
0.6
0.06
0.06
0.05
0.05
0.04
0.04
0.03
0.03
0.04
0.4
0.02
0.02
0.02
0.2
20
40
60
80
100
(e) Ground truth
0.1
0.07
0
0.01
0
0
20
20
40
40
60
60
80
80
100
0
100
0.07
0.07
0.06
0.06
0
0
20
20
40
40
60
60
80
80
100
100
0
0
0
20
20
40
40
60
60
80
80
0
100
100
0.4
0.2
0
0
20
20
40
40
60
60
x
80
80
100
0
100
0.1
1
1
0.08
0.8
0.8
0.05
0.04
0.04
0.06
0.6
0.6
0.03
0.03
0.04
0.4
0.4
0.02
0.02
0.01
0.01
0.02
0.2
0
20
20
40
60
80
100
0
20
40
60
80
100
0
20
40
60
80
100
0
40
60
80
100
y
0.05
0
0
0
20
20
40
40
60
60
80
80
100
100
0
0.07
0.07
0.06
0.06
0.05
0.05
0.04
0.04
0.03
0.03
0.02
0.02
0.01
0
0
0
20
20
40
40
60
60
80
80
100
100
0
0
0
20
20
40
40
60
60
80
80
0
100
100
0.2
0
0
20
20
4040
x
6060
8080
100
100
0
0.1
1
1
0.08
0.8
0.8
0.06
0.6
0.6
0
20
40
60
80
100
y
CAM7 – ZONE 4
↓
CAM8 – ZONE 3
60
60
0.2
0.07
0.01
CAM4 – ZONE 1
↓
CAM5 – ZONE 2
40
40
0.4
y
CAM3 – ZONE 1
↓
CAM7 – ZONE 2
20
20
1
0.5
0.04
0.4
0.02
0.2
0.01
0
0
20
20
40
40
60
60
(f) Makris et al. [4]
80
80
100
100
0
0
0
20
20
40
40
60
60
80
80
(g) Nui et al. [23]
100
100
0
0
0
20
20
40
40
60
60
80
80
(h) Chen et al. [25]
100
100
0
0.4
0.2
0
0
20
20
40
40
60
60
x
(i) Ours
80
80
100
100
0
0
20
40
60
80
100
(j) Ground truth
Fig. 11. Comparison of inferred transition distributions and ground truth. First three rows are valid links (a-e) and last three rows are invalid links (f-j).
to the proposed method. Moreover, it takes a much longer time
(337.27 seconds) than our proposed method.
shown in Fig. 11 (a-c,f-h). On the other hand, our results are
very similar to the ground truth (Fig. 11 (d,e,i,j)).
A list of valid Zone-to-Zone links inferred by the proposed
methods is summarized in Table. III. Note that the pair
of Exit (CAM3-ZONE2) – Entry (CAM5-ZONE6) has the
negative value of the transition time because these zones are
overlapped.The overall results are close to the ground-truth
2
N µgt , σgt
. We expect that the inferred transition distributions of valid links show Gaussian-like distributions and
those of invalid links show uniform or arbitrary distributions.
However, the previous methods [4], [23], [25] showed unclear
and noisy distributions for both valid and invalid links as
In Table. IV, we compared several methods in terms of
re-identification and topology accuracies. To ensure a fair
comparison of the previous methods with our method, we
applied the same baseline (e.g. feature descriptor, parameters,
experimental settings) to the previous methods except topology inference methods. The inferred topologies of previous
methods [4], [23], [25] are used for person re-identification
tasks by following the method in [5]. Note that we cannot
estimate the transition time error of DNPR [28] since the
method only infers CAM-to-CAM connections. As we can
11
see, our method shows superior performances for both reidentification and topology inference than other methods.
2) Online Person Re-identification Results: Based on the
inferred camera topology in the initialization stage, we conducted online person re-identification and compared it with
two different approaches. The first approach estimates the
camera topology based on Exhaustive search in the offline
initialization stage and uses the inferred topology for the online
re-identification test. Note that this approach exploits the
inferred topology but still fully compares multiple appearances
of people to find correspondences in the online test stage.
The second approach based on True matching estimates the
camera topology using ground truth re-identification pairs in
the offline initialization stage and performs the online reidentification test in the same way as our method.
As shown in Table V, the performance of our online reidentification test is comparable to that of True matching.
Our method also outperforms Exhaustive search in terms of
both re-identification accuracy and topology distance. Thanks
to the online topology update in the online re-identification
test stage, the topology distance of our method decreased
during the online test stage (0.076 → 0.023). As we can
see in the table, our method with the online topology update
(i.e., ours) performs better than our method without online
topology update (i.e., Ours– no update) in terms of both reidentification accuracy and topology distance in the online test
stage. It demonstrates that our online topology update and
re-identification methods are effective and complement each
other.
B. Evaluations using Public Dataset: NLPR MCT
In order to validate the proposed method using public
datasets, we tested the proposed method using the NLPR
MCT [43] dataset. The dataset contains four subsets with
different non-overlapping multi-camera networks and different
numbers of people. Among them, we used two datasets (i.e.,
DATA1, DATA2) containing larger numbers of people than
others (the number of people – DATA1: 235, DATA2: 255,
DATA3: 14, DATA4: 49). DATA1 and DATA2 have the same
layout and camera field of views as shown in Fig. 12 (a, b-d).
Duration of each dataset is 20 minutes. Since the datasets are
not large-scale datasets (the number of people is small), we
used whole videos (20min) for our proposed camera network
topology initialization method and did not perform the online
person re-identification.
CAM-to-CAM topology inference
Based on our proposed method, we inferred the CAM-toCAM topologies of both datasets (DATA1, 2). We found strong
connections between CAM1 and CAM2 in both datasets. Unfortunately, we could not find the connections from CAM2 to
CAM3 in both datasets, although there is a uni-directional link
from Exit (CAM2-Zone2) to Entry (CAM3-Zone2) because
of the severe illumination difference (indoor and outdoor) and
different radiometric parameters between the two cameras.
In this case, person re-identification methods based on
matching should handle these photometric issues, which is
1
2
C1
Sidewalk
1
2
C2
Building
C3
2
3
Lawn
1
(a) A layout of the dataset
(b) CAM1
(c) CAM2
(d) CAM3
Fig. 12. A public multi-camera person re-identification dataset: NLPR
MCT [43]
beyond the scope of this study. In this case, even the exhaustive
search approaches failed to find the connection between CAM2
and CAM3 because of the same reasons. We expect that by
adopting some color correction methods such as [25] before
re-identification or exploiting illumination invariant feature
descriptors for re-identification, we can find the connection
between CAM2 and CAM3. We will try this in future studies.
Camera topology initialization results
After the CAM-to-CAM topology inference, we infer the
topology of entry and exit zones. We found two valid links
between CAM1 and CAM2 in both the datasets. The first
link is Exit (CAM1-Zone2) – Entry (CAM2-Zone1) and the
second link is Exit (CAM2-Zone1) – Entry (CAM1-Zone2).
Through this step, invalid pairs of cameras and zones are
detected and ignored. In the next step, we iteratively update the
valid links between zones and build a final camera topology
map of the camera network.
After evaluating the Zone-to-Zone topology inference, we
iteratively update the valid links between zones. In the
case of NLPR MCT dataset [43], valid Zone-to-Zone links
converged within the second iteration. Table VI shows the
person re-identification accuracy according to each proposed
topology initialization step. Our method shows a superior reidentification performances: 93.2% and 91.5%, for the camera
pair CAM1-CAM2.
We compared our method to several previous topology
inference methods [4], [23], [25]. The previous methods
showed unclear and noisy transition distributions as shown in
Fig. 13 (a-c). Even though the transition distribution inferred
by Chen’s method [25] shows a peak of the distribution, the
overall distribution is too noisy to utilize. On the other hand,
our result is very close to the ground truth and shows a very
clear distribution as shown in Fig. 13 (d,e).
VII. D ISCUSSIONS
The proposed unified framework can infer camera network
topology while it performs person re-identification. In this
study, we assume that the transition distribution between the
two links follows a single-mode Gaussian model since the
12
TABLE V
P ERFORMANCE COMPARISON IN ONLINE TEST STAGE .
Offline initialization stage
Rank-1 accuracy
Topology dist
52.1%
5.620
D2
c1z2
–c2z1
c2z1
D2D2
c1z2
– –c2z1
c1z2
method1
method1
method1
0.076
100%
0
0.1
0.04
0.040.04
0.2 0.2 0.2
0.05
0.050.05
0 0
0
0.1 0.1 0.1
2020
60 608080
80 80100
100100
20 40
2020
40
8080
100
40404040 6060
6060
100
100
0 00 00 20
0
0 0
(a) Makris et al. [4]
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
2020
4040
x
6060
8080
100
100
6060
8080
100
100
0.1
0
0
2020
80 80
100 100 0
20 40
2020
40
60 60
80 80
100
40404040 6060
100
60608080
100
0 0 0 20
100
0 0 0
0 0 0
(b) Nui et al. [23]
c2z1
D2D2
c2z1
– –c1z2
D2
c2z1
–c1z2
c1z2
4040
y
0.3 0.3 0.3
20
60 60 8080
80 80100
100100
20 40
2020
40
8080
100
4040
6060
100
20
4040 6060
100
0 0000 20
2020
0.7
0.4 0.4 0.4
0.15
0.150.15
0.02
0.020.02
0
method3
method3
method3
0.1 0.10.1
0
0.2
0.5 0.5 0.5
0.06
0.060.06
0
0.4
0
0.2 0.20.2
0.08
0.080.08
0.5
0.3
62.5%
method2
method2
method2
0.1 0.10.1
0.6
y
LOMO [51] – Exhaustive
Ours – no update
Ours
True matching
Online test stage
Rank-1 accuracy
Topology dist
65.6%
0.046
72.3%
0.093
74.8%
0.023
0
75.6%
0.011
2020
(c) Chen et al. [25]
4040
x
6060
8080
100
100
0
00
(d) Ours
(e) Ground truth
0.7
Fig. 13. Comparison of inferred transition distributions and ground truth (DATA2 - Exit: CAM1-ZONE2 , Entry: CAM2-ZONE1).
0.150.15
0.15
0.080.08
0.08
0.060.06
0.06
0.1 0.1 0.1
0.6
0.080.080.08
0.5
0.4
0.1 0.10.1
0.060.060.06
0.3
0.040.04
0.04
0.040.040.04
TABLE VI
0.020.02
0.02
0
0.2
Computational time
0.050.05
0.05
0.1
0.020.020.02
R ANK -1 ACCURACIES AT DIFFERENT STAGES OF CAMERA TOPOLOGY
100
100
100
0 0 0 202020 404040 606060 808080 100
0 0 0 202020 404040 606060 808080 100
0 0 0 202020 404040 606060 808080 100
100
100
100
INITIALIZATION
.
0 0
0 0 0
20 20 20
40 40 40
60 60 60
80 80 80
100100100
DATA1 (CAM1 - CAM2)
DATA2 (CAM1 - CAM2)
0
0 0
0 0 0
20 20 20
40 40 40
60 60 60
80 80 80
100100100
0 0 0
0 0 0
20 20 20
40 40 40
CAM-to-CAM Zone-to-Zone Iteration 1
72.6%
76.7%
93.2%
69.8%
70.8%
91.5%
model works well in most cases. In general, most previous
study [4], [23] also follow the same assumption. Although
it works well in most cases, it may fail in more complex
situations, e.g., multiple paths between two zones. We will
consider this issue in further studies.
VIII. C ONCLUSIONS
In this study, we proposed a unified framework to automatically solve both person re-identification and camera
network topology inference problems. In addition, to validate
the performance of person re-identification in the practical
large-scale surveillance scenarios, we provided a new person re-identification dataset called SLP. We qualitatively and
quantitatively evaluated and compared the performance of the
proposed framework with state-of-the-art methods using public
and our own datasets. The results show that the proposed
framework is promising for both person re-identification and
camera topology inference and superior to other frameworks
in terms of both speed and accuracy.
A PPENDIX
As we mentioned in the paper, the traditional exhaustive reidentification approaches examine every person pair between
two different cameras (cA , cB ). Assuming that the numbers
of objects in cA and cB are the same as N and each
object contains the same number of appearances as K, it is
obvious that the exhaustive approaches need (N 2 K 2 ) time for
appearance comparisons for re-identification. Using the BigO notation, the computational complexity
of the exhaustive
approach is defined as O N 2 K 2 .
On the other hand, in the proposed method, the tree structure
of the random forest method makes the multi-shot person reidentification very fast. In return, the random forest containing
60 60 60
80 80 80
2
Total time
0
100100 100
00
2020
4040
6060
8080
(sec)
100
100
1.5
1
0.5
0
100
200
300
400
The number of person
Fig. 14. Computation time of the proposed method based on the number of
people in each camera We fixed the number of appearances of each person
used for multi-shot matching (K = 30). The computation time includes
random forest training and testing time.
multiple decision trees needs training time. If we denote the
number of decision trees and the random forest training time
as M and T respectively, we need (M N K log N ) time for
the appearance comparisons and T for forest training. Here,
note that the number of decision trees is a constant value. We
fixed it as 10 in our experiments. Therefore, we can ignore the
number of the decision trees M in the Big-O notation since
M N, K. In addition, the random forest training time in
the training stage can be also be ignored because the training
time T linearly increases when the number of cameras in a
camera network increases whereas the test time exponentially
increases when the number of cameras increases. Actually, the
computational complexity of the test stage is proportional to
the number of possible camera pairs: L C2 , where L is the
number of cameras in a camera network. For that reason, the
training time T can be ignored when the camera network
is large. As a result, the complexity of proposed method is
approximated as O (N K log N ).
Fig. 14 shows the computation time according to the number
of people in each camera in the two-camera re-identification
case. As we can see, the computation time of the proposed
method increases linearly, not exponentially, as the number of
people increases.
13
R EFERENCES
[1] M. Farenzena, L. Bazzani, A. Perina, V. Murino, and M. Cristani, “Person re-identification by symmetry-driven accumulation of local features,”
in IEEE Conference on Computer Vision and Pattern Recognition, 2010.
[2] J. V. Davis, B. Kulis, P. Jain, S. Sra, and I. S. Dhillon, “Informationtheoretic metric learning,” in IEEE International Conference on Machine
Learning, 2007.
[3] E. Ahmed, M. Jones, and T. K. Marks, “An improved deep learning
architecture for person re-identification,” Differences, 2015.
[4] D. Makris, T. Ellis, and J. Black, “Bridging the gaps between cameras,”
in IEEE Conference on Computer Vision and Pattern Recognition, 2004.
[5] C. C. Loy, T. Xiang, and S. Gong, “Time-delayed correlation analysis for
multi-camera activity understanding,” International Journal of Computer
Vision, 2010.
[6] Y.-J. Cho, J.-H. Park, S.-A. Kim, K. Lee, and K.-J. Yoon, “Unified framework for automated person re-identification and camera
network topology inference in camera networks,” arXiv preprint
arXiv:1704.07085, 2017.
[7] A. Bedagkar-Gala and S. K. Shah, “A survey of approaches and trends
in person re-identification,” Image and Vision Computing, 2014.
[8] C. Liu, S. Gong, C. C. Loy, and X. Lin, “Person re-identi-fication: What
features are important?” in European Conference on Computer Vision,
2012.
[9] R. Zhao, W. Ouyang, and X. Wang, “Learning mid-level filters for person
re-identification,” in IEEE Conference on Computer Vision and Pattern
Recognition, 2014.
[10] M. Koestinger, M. Hirzer, P. Wohlhart, P. M. Roth, and H. Bischof,
“Large scale metric learning from equivalence constraints,” in IEEE
Conference on Computer Vision and Pattern Recognition, 2012.
[11] M. Dikmen, E. Akbas, T. S. Huang, and N. Ahuja, “Pedestrian recognition with a learned metric,” in Asian Conference on Computer Vision,
2011.
[12] K. Q. Weinberger, J. Blitzer, and L. K. Saul, “Distance metric learning
for large margin nearest neighbor classification,” in Advances in Neural
Information Processing Systems, 2005.
[13] P. M. Roth, M. Hirzer, M. Köstinger, C. Beleznai, and H. Bischof,
“Mahalanobis distance learning for person re-identification,” in Person
Re-Identification, 2014.
[14] T. Wang, S. Gong, X. Zhu, and S. Wang, “Person re-identification by
video ranking,” in European Conference on Computer Vision, 2014.
[15] Y. Li, Z. Wu, and R. J. Radke, “Multi-shot re-identification with
random-projection-based random forests,” in IEEE Winter Conference
on Applications of Computer Vision, 2015.
[16] D. Yi, Z. Lei, S. Liao, and S. Z. Li, “Deep metric learning for
person re-identification,” in IEEE International Conference on Pattern
Recognition, 2014.
[17] S. Bak, F. Martins, and F. Bremond, “Person re-identification by pose
priors,” in IS&T/SPIE Electronic Imaging, 2015.
[18] Z. Wu, Y. Li, and R. J. Radke, “Viewpoint invariant human reidentification in camera networks using pose priors and subjectdiscriminative features,” 2015.
[19] Y.-J. Cho and K.-J. Yoon, “Improving person re-identification via poseaware multi-shot matching,” in IEEE Conference on Computer Vision
and Pattern Recognition, 2016.
[20] O. Javed, Z. Rasheed, K. Shafique, and M. Shah, “Tracking across
multiple cameras with disjoint views,” IEEE International Conference
on Computer Vision, 2003.
[21] Y. Cai and G. Medioni, “Exploring context information for inter-camera
multiple target tracking,” in IEEE Winter Conference on Applications of
Computer Vision, 2014.
[22] A. Rahimi, B. Dunagan, and T. Darrell, “Simultaneous calibration
and tracking with a network of non-overlapping sensors,” in IEEE
Conference on Computer Vision and Pattern Recognition, 2004.
[23] C. Niu and E. Grimson, “Recovering non-overlapping network topology
using far-field vehicle tracking data,” in IEEE International Conference
on Pattern Recognition, 2006.
[24] C. Stauffer, “Learning to track objects through unobserved regions,”
2005.
[25] X. Chen, K. Huang, and T. Tan, “Object tracking across non-overlapping
views by learning inter-camera transfer models,” Pattern Recognition,
2014.
[26] C. C. Loy, T. Xiang, and S. Gong, “Incremental activity modeling in
multiple disjoint cameras,” 2012.
[27] Y. Cai, K. Huang, T. Tan, and M. Pietikainen, “Recovering the topology
of multiple cameras by finding continuous paths in a trellis,” in IEEE
International Conference on Pattern Recognition, 2010.
[28] N. Martinel, G. L. Foresti, and C. Micheloni, “Person reidentification
in a distributed camera network framework,” IEEE transactions on
cybernetics, 2016.
[29] L. Breiman, “Random forests,” Machine learning, 2001.
[30] S. H. Han, K. J. Yoon, and I. S. Kweon, “A new technique for shot
detection and key frames selection in histogram space,” 2000.
[31] D. Makris and T. Ellis, “Automatic learning of an activity-based semantic
scene model.” in IEEE Advanced Video and Signal Based Surveillance,
2003.
[32] S. Karanam, M. Gou, Z. Wu, A. Rates-Borras, O. Camps, and R. J.
Radke, “A comprehensive evaluation and benchmark for person reidentification: Features, metrics, and datasets,” arXiv:1605.09653, 2016.
[33] D. Gray, S. Brennan, and H. Tao, “Evaluating appearance models
for recognition, reacquisition, and tracking,” in IEEE International
Workshop on Performance Evaluation for Tracking and Surveillance,
2007.
[34] W.-S. Zheng, S. Gong, and T. Xiang, “Associating groups of people,”
in British Machine Vision Conference, 2009.
[35] D. S. Cheng, M. Cristani, M. Stoppa, L. Bazzani, and V. Murino,
“Custom pictorial structures for re-identification,” in British Machine
Vision Conference, 2011.
[36] D. Baltieri, R. Vezzani, and R. Cucchiara, “3dpes: 3d people dataset
for surveillance and forensics,” in Proceedings of the 2011 joint ACM
workshop on Human gesture and behavior understanding, 2011.
[37] M. Hirzer, C. Beleznai, P. M. Roth, and H. Bischof, “Person reidentification by descriptive and discriminative classification,” in IEEE
Scandinavian conference on Image analysis, 2011.
[38] S. Wang, M. Lewandowski, J. Annesley, and J. Orwell, “Reidentification of pedestrians with variable occlusion and scale,” in IEEE
International Conference on Computer Vision Workshops, 2011.
[39] N. Martinel and C. Micheloni, “Re-identify people in wide area camera
network,” in IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2012.
[40] A. Bialkowski, S. Denman, S. Sridharan, C. Fookes, and P. Lucey, “A
database for person re-identification in multi-camera surveillance networks,” in IEEE Digital Image Computing: Techniques and Applications,
2012.
[41] W. Li and X. Wang, “Locally aligned feature transforms across views,”
in IEEE Conference on Computer Vision and Pattern Recognition, 2013.
[42] W. Li, R. Zhao, T. Xiao, and X. Wang, “Deepreid: Deep filter pairing
neural network for person re-identification,” in IEEE Conference on
Computer Vision and Pattern Recognition, 2014.
[43] W. Chen, L. Cao, X. Chen, and K. Huang, “An equalised global
graphical model-based approach for multi-camera object tracking,”
arXiv:1502.03532, 2015.
[44] A. Das, A. Chakraborty, and A. K. Roy-Chowdhury, “Consistent reidentification in a camera network,” in European Conference on Computer Vision, 2014.
[45] L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian, “Scalable
person re-identification: A benchmark,” in IEEE International Conference on Computer Vision, 2015.
[46] L. Zheng, H. Zhang, S. Sun, M. Chandraker, and Q. Tian, “Person reidentification in the wild,” arXiv preprint arXiv:1604.02531, 2016.
[47] L. Zheng, Z. Bie, Y. Sun, J. Wang, C. Su, S. Wang, and Q. Tian, “Mars:
A video benchmark for large-scale person re-identification,” in European
Conference on Computer Vision, 2016.
[48] M. Gou, S. Karanam, W. Liu, O. Camps, and R. J. Radke,
“Dukemtmc4reid: A large-scale multi-camera person re-identification
dataset,” in IEEE Conference on Computer Vision and Pattern Recognition Workshops, July 2017.
[49] E. Ristani, F. Solera, R. Zou, R. Cucchiara, and C. Tomasi, “Performance
measures and a data set for multi-target, multi-camera tracking,” in
European Conference on Computer Vision Workshops on Benchmarking
Multi-Target Tracking, 2016.
[50] C. Vondrick, D. Patterson, and D. Ramanan, “Efficiently scaling up
crowdsourced video annotation,” International Journal of Computer
Vision.
[51] S. Liao, Y. Hu, X. Zhu, and S. Z. Li, “Person re-identification by
local maximal occurrence representation and metric learning,” in IEEE
Conference on Computer Vision and Pattern Recognition, 2015.
14
PLACE
PHOTO
HERE
Yeong-Jun Cho received the B.S. degree in information and communication engineering from Korea
Aerospace University and the M.S. degree in information and communications from Gwangju Institute of Science and Technology (GIST), in 2012
and 2014, respectively. He is currently pursuing
the ph.D. degree as a member of the Computer
Vision Laboratory in GIST. His research interests
include computer vision topics such as person reidentification, multi-object tracking, object detection
and medical image analysis.
estimation, and
Su-A Kim received the B.S. degree from the Department of Electronic and IT Media Engineering
from Seoul National University of Science and
Technology in 2013 and the M.S. degree in School
of Information and Communications from Gwangju
Institute of Science and Technology (GIST) in 2015.
She is currently a scientific researcher at Intel Visual
Computing Institute and pursuing the Ph.D. degree
in Saarland University. Her research interests include
key research topics in computer vision such as person re-identification, 3D object detection and pose
4D light field segmentation.
PLACE
PHOTO
HERE
Jae-Han Park received the B.S. degree in electrical
engineering and computer science from Gwangju
Institute of Science and Technology (GIST) in 2015.
He is currently pursuing M.S degree as a member of
Computer Vision Laboratory in GIST. His research
interests include person re-identification, video summarization and age estimation using deep learning.
PLACE
PHOTO
HERE
PLACE
PHOTO
HERE
Kyuewang Lee received the B.S. degree in Electrical Engineering and Computer Science from
Gwangju Institute of Science and Technology
(GIST) in 2016. He is currently pursuing the integrated M.S./Ph.D. degree as a member of the
Perception and Intelligence Laboratory in Seoul National University. His research interests include key
research topics in computer vision, such as visual
object tracking, multi-object tracking, and person reidentification.
Kuk-Jin Yoon received the B.S., M.S., and Ph.D.
degrees in Electrical Engineering and Computer
Science from Korea Advanced Institute of Science
and Technology (KAIST) in 1998, 2000, 2006,
PLACE
respectively. He was a post-doctoral fellow in the
PHOTO
PERCEPTION team in INRIAGrenoble, France, for
HERE
two years from 2006 and 2008 and joined the School
of Information and Communications in Gwangju
Institute of Science and Technology (GIST), Korea,
as an assistant professor in 2008. He is currently an
associate professor and a director of the Computer
Vision Laboratory in GIST.
| 1cs.CV
|
Ref: AAAI Conference on Artificial Intelligence, pages 2839–2845, Quebec City, Canada, July 2014.
A Generalized Genetic Algorithm-Based Solver
for Very Large Jigsaw Puzzles of Complex Types
Dror Sholomon and Eli (Omid) David∗ and Nathan S. Netanyahu†
arXiv:1711.06768v1 [cs.CV] 17 Nov 2017
Department of Computer Science
Bar-Ilan University
Ramat-Gan 52900, Israel
[email protected], [email protected], [email protected]
Abstract
In this paper we introduce new types of square-piece
jigsaw puzzles, where in addition to the unknown location and orientation of each piece, a piece might also
need to be flipped. These puzzles, which are associated
with a number of real world problems, are considerably
harder, from a computational standpoint. Specifically,
we present a novel generalized genetic algorithm (GA)based solver that can handle puzzle pieces of unknown
location and orientation (Type 2 puzzles) and (twosided) puzzle pieces of unknown location, orientation,
and face (Type 4 puzzles). To the best of our knowledge,
our solver provides a new state-of-the-art, solving previously attempted puzzles faster and far more accurately,
handling puzzle sizes that have never been attempted
before, and assembling the newly introduced two-sided
puzzles automatically and effectively. This paper also
presents, among other results, the most extensive set of
experimental results, compiled as of yet, on Type 2 puzzles.
Introduction
Jigsaw puzzles are a popular form of entertainment, first
produced around 1760 by John Spilsbury, a Londonian engraver and mapmaker. Given n different non-overlapping
pieces of an image, the player has to reconstruct the original image, taking advantage of both the shape and chromatic information of each piece. Despite the popularity and
vast distribution of jigsaw puzzles, their assembly is not trivial, from a computational standpoint, as this problem was
proven to be NP-hard (Altman 1989; Demaine and Demaine
2007). Nevertheless, a computerized jigsaw solver may have
applications in many real-world problems, arising in various domains such as archeology (Koller and Levoy 2006;
Brown et al. 2008), biology (Marande and Burger 2007),
chemistry (Wang 2000), literature (Morton and Levison
1968), speech descrambling (Zhao et al. 2007), art restoration (Fornasier and Toniolo 2005), image editing (Cho et
al. 2008), and the recovery of shredded documents or photographs (Justino, Oliveira, and Freitas 2006; Marques and
∗
www.elidavid.com
Nathan Netanyahu is also with the Center for Automation Research, University of Maryland, College Park, MD 20742 (e-mail:
[email protected]).
†
Freitas 2009; Cao, Liu, and Yan 2010; Deever and Gallagher
2012). Besides, as noted in (Goldberg, Malon, and Bern
2004), pursuing this topic may be justified solely due to its
intriguing nature.
The first to tackle the jigsaw problem, computationally,
were Freeman and Garder (1964). Their solver handles up to
9-piece puzzles, using only piece shape information. Kosiba
et al. (1994) were the first to facilitate the use of image content by their solver. Subsequent research turned to colorbased square-piece puzzles, instead of the earlier shapebased variants. Cho et al. (2010) presented a probabilistic
puzzle solver that can handle up to 432 pieces, given some
a priori knowledge of the puzzle (e.g., anchor pieces). Their
results were further improved by (Yang, Adluru, and Latecki 2011), who presented a so-called particle filter-based
solver and by (Pomeranz, Shemesh, and Ben-Shahar 2011),
who made a major contribution to the field by introducing
a fully automated jigsaw puzzle solver that handles puzzles
of up to 3,000 (square) pieces, without any prior knowledge
about the image. The latter solver treats puzzles with unknown piece location but assumes a known orientation. Gallagher (2012) was the first to handle puzzles with unknown
piece location and orientation (Type 2 puzzles), where each
piece might need to be translated and rotated (by 0, 90, 180,
or 270 degrees). This solver was tested on 432- and 1,064piece puzzles and a single 9,600-piece image. More recently,
a solver based on genetic algorithms (GA) (Holland 1975),
which can handle up to 22,755-piece puzzles, was presented
by(Sholomon, David, and Netanyahu 2013). Although capable of solving considerably larger puzzles, their solver is
restricted to known piece orientations (i.e., Type 1 puzzles).
In its most basic form, every puzzle solver requires an
evaluation function for the compatibility of adjacent pieces
and a tiling strategy for placing the pieces as accurately as
possible. Recent tiling strategies tend to be greedy and thus
are subject to the familiar disadvantages of greedy methods, i.e., they are more likely to be affected by local optima.
While (Sholomon, David, and Netanyahu 2013) successfully
employed a GA-based solver for Type 1 puzzles (unknown
piece location only), the question remains whether GAbased solvers could be applied to more challenging types,
namely Type 2 puzzles (unknown piece location and orientation) and Type 4 puzzles (two-sided pieces of unknown
location and orientation).
In this paper, we continue in the same vein of employing
genetic algorithms as a strategy for piece placement. We describe a detailed GA scheme for a solver capable of handling
both Type 1 and Type 2 puzzles more than twice as large as
puzzle sizes that have been attempted before, and present
extensive empirical results which demonstrate the efficiency
of the presented method in terms of speed and accuracy. We
further advance the state-of-the-art with respect to the jigsaw puzzle problem by extending our solver to handle also
two-sided puzzles (Type 4 puzzles), i.e., puzzles where the
correct face of each piece is also unknown (adding another
degree of freedom to each piece). Thus, the solver has to
find the correct location for each piece, its correct orientation (out of four possible angles), and its correct face (out
of two possible ones). This type of puzzle, which is representative of various real-world applications, is considerably
more complex. We present an extensive set of empirical results for all available datasets, establishing the effectiveness
of our solver in handling different images of different sizes.
Puzzle Types
The first discussion on different puzzle types appears in
(Gallagher 2012). In all types, n different non-overlapping
square pieces of an image are given and there exists a unique
tiling (arrangement) which is considered correct. Type 1 is
the most common variant handled; it refers to puzzles with
only piece location unknown. In Type 2 puzzles, a piece orientation is also unknown, allowing each piece to be rotated
by 0, 90, 180 or 270 degrees. As noted by (Gallagher 2012),
this puzzle type increases the complexity in several ways.
First, a pair of pieces can fit together in any of 16 possible configurations, multiplying the number of possible solutions by 4n in comparison to Type 1 puzzles. Second, an
algorithmic solver must consider both translations and rotations. Third, the puzzle reconstruction should be considered
in both landscape and portrait orientations. Type 3 puzzles,
consisting of pieces with known location and unknown orientation, are listed only for the sake of completeness.
We define here Type 4 puzzles, i.e., two-sided puzzles of
two images, where each piece face belongs to one of the
images. The solver has to determine the correct location
of each piece, its correct orientation, and its relevant face
(with respect to each of the two images). This problem version is motivated by real-world applications, e.g., a shredded document might have been printed on both sides before being shredded. (This is all the more applicable, in view
of current global environmental trends, encouraging doublesided printing.) The computational complexity in this case
increases in several ways. First, each pair of pieces might
now fit together in any of 64 possible configurations, multiplying the number of possible solutions by 8n , relatively to
Type 1 puzzles. Second, the solver has to consider now the
possibility of flipping a piece, in addition to its translation
and rotation. Third, the solver must always consider the two
images (being formed on the fly) when placing new pieces
and assessing the results, e.g., whether to assemble each image separately or assemble them simultaneously.
For completeness, one can define additional puzzle types
with all possibilities of (un)known piece location, orientation, and face.
Type
1
2
3
4
5
6
7
Unknown
Location
3
3
3
3
Unknown
Orientation
3
3
3
3
Unknown
Face
3
3
3
3
Table 1: Puzzle types according to different problem specifications.
Genetic Algorithms
In this section we provide a quick overview of the GA
methodology. A GA is a search procedure within a problem’s solution domain. Since examining all possible solutions of a specific problem is virtually infeasible, GAs offer
an optimization heuristic inspired by biological natural selection.
In GA terms, a solution to the problem (e.g., a suggested
arrangement of the puzzle’s pieces) is represented as an individual “organism” (i.e., chromosome) of a large population. Essentially, a GA attempts to reach an optimal solution
by mimicking the processes of natural selection and evolution; the “fittest” individuals of each generation reproduce,
creating offspring chromosomes. If defined correctly, the
crossover operator responsible for offspring creation should
allow for ”good” qualities to pass on from parents to children, in an attempt to create better offspring (i.e., solutions).
In each iteration (i.e., one generation of the algorithm), the
entire population is replaced by the many offspring created
by the crossover operation. (The total population size remains fixed throughout all the generations.)
In order to imitate natural selection, a chromosome’s reproduction rate, i.e., the number of times it is selected to
reproduce (and hence the number of its offspring), is set
directly proportionate to its fitness. The fitness, which is a
score obtained by a fitness function, represents the quality
of a given solution. The crossover should, thus, take place
mainly between higher-rated solutions.
The successful performance of a GA depends mainly
on the appropriate choice of chromosome representation,
crossover operator, and fitness function. The chromosome
representation and crossover operator should yield an enhanced solution by merging two “promising” chromosomes
(i.e., chromosomes representing promising partial solutions)
that are passed on to the next generation. Figure 1 provides
a pseudo-code of a common GA framework.
Puzzle Solving
In this section we present our generalized GA-based solver,
which is designed to handle more difficult puzzle types. The
population = create_random_population(POPULATION_SIZE);
for (i = 0; i < NUM_OF_GENERATIONS; ++i) {
new_population = NULL;
evaluate_population(&population);
for (j = 0; j < POPULATION_SIZE; ++j) {
parent1 = select_parent(population);
parent2 = select_parent(population);
child = crossover(parent1, parent2);
add_child_to_population(&new_population, child);
}
population = new_population;
}
solution = get_best(population);
Figure 1: Pseudocode of GA framework
GA starts from a fixed-size population of randomly generated solutions. In each iteration, the entire population is evaluated using the fitness function described below, and a new
population is (re)produced by employing the crossover operator to the selected chromosome pairs. We use the common
selection method of roulette wheel selection, where each
chromosome is selected to reproduce, with probability directly proportional to its fitness score.
We define each chromosome to be a complete solution to
the problem, i.e., a suggested tiling of all pieces, including
their orientation and face (if applicable). We now have to
supply an appropriate fitness function and a crossover operator.
The Fitness Function
The fitness function determines the quality of each chromosome (i.e., each solution), and hence the expected number of
its children. In every generation, all chromosomes are evaluated for the purpose of selection.
For fitness evaluation, we use the dissimilarity measure,
which was investigated thoroughly in previous comparative studies (Cho, Avidan, and Freeman 2010; Pomeranz,
Shemesh, and Ben-Shahar 2011) and found to be very effective. The dissimilarity measure relies on the premise that
adjacent jigsaw pieces in the original image tend to share
similar colors along their abutting edges, i.e., the sum (over
all neighboring pixels) of squared color differences (over all
three color bands) should be minimal. Let pieces pi , pj be
represented in normalized L*a*b* space by corresponding
W × W × 3 matrices, where W is the height/width of each
piece (in pixels). Assuming, for example, that pj is to the
right of pi , the piece dissimilarity in this case is given by:
v
uW 3
uX X
D(pi , pj ) = t
(pi (k, W, b) − pj (k, 1, b))2 . (1)
k=1 b=1
Obviously, to maximize the compatibility of two pieces,
their dissimilarity should be minimized.
For Type 2 puzzles we set the fitness function of a given
chromosome to be the sum of pairwise dissimilarities over
all adjacent edges. Given that the puzzle consists of (N ×M )
tiles, we represent each chromosome by an (N ×M ) matrix,
where a matrix entry xi,j (i = 1..N, j = 1..M ) corresponds
to a single puzzle piece and its orientation, we define its fitness as:
N M
−1
X
X
(D(xi,j , xi,j+1 )) +
i=1 j=1
N
−1 X
M
X
(D(xi,j , xi+1,j )) (2)
i=1 j=1
where the D term in each case is the dissimilarity of the two
pieces in question about their joint edge, taking into account
their actual rotations, as stored in the chromosome.
For Type 4 puzzles, we need to consider actually two abutting edges (i.e., an adjacent edge for each piece face), for every neighboring piece in a piece pair. We sum the pairwise
dissimilarities over all adjacent edges, computing effectively
a fitness score for each image of the two-sided puzzle. Thus,
the GA must balance the dissimilarity minimization on both
sides of the puzzle to obtain, hopefully, the correct reconstruction of both images.
Type 2 Crossover
Given a chromosome representation, the crossover operator receives two tile configurations (i.e., two parent chromosomes) and produces a new arrangement of the pieces (i.e.,
a new offspring chromosome). In general, the crossover operator should encourage “good” qualities to pass on from
the parents to their child. In particular, we would like an effective operator to meet the following three requirements.
First, it must create a valid child solution, where each piece
appears exactly once (i.e., no missing or duplicate pieces).
Second, the operator must support position independence of
puzzle segments assembled correctly. Namely, if a parent assembles correctly part of the image (i.e., a segment), modulo some spatial offset, the operator should accommodate
the required spatial translation and rotation of the entire segment in the resulting child. For example, Figure 2 features a
parent chromosome containing a person’s head correctly assembled, albeit spatially misplaced. The idea is to retain the
correctly assembled head segment but allow for its translation and rotation. Third, a proper heuristic should be applied
to detect correctly assembled segments (as the person’s head
in Figure 2).
Various works (Gallagher 2012; Sholomon, David, and
Netanyahu 2014) use a weighted graph representation for
the problem, where each node corresponds to a jigsaw piece
and each (graph) edge corresponds to a joint edge of two adjacent puzzle pieces. We denote the edges of piece pi (in a
clockwise manner) as pi .a, pi .b, pi .c and pi .d. For example,
the graph edge pi .b − pj .d denotes the geometric configuration where edge b of piece pi is adjacent to edge d of piece
pj . The weight of each graph edge is the dissimilarity of the
two-piece configuration (given by Eq. 1). For Type 2 puzzles, there are 16 possible edges between every two (graph)
vertices. This representation lends itself easily to an effective
crossover of correctly assembled segments in the parents.
The geometric relation pi .b − pj .d is invariant to both the
absolute spatial location and orientation of the pieces (e.g.,
translating and/or rotating both pieces by 90 degrees will not
affect this relation).
Arriving at a piece arrangement can be viewed analogously to constructing a spanning tree of n − 1 graph edges.
(a) Parent1
(b) Parent2
(c) 17 Pieces
(d) 91 Pieces
(e) 147 Pieces
(f) 229 Pieces
(g) 347 Pieces
(h) Child
Figure 2: Illustration of Type 2 crossover operation: (a) Parent1, (b) Parent2, (c)–(g) evolution of kernel growth until (h) a
complete child is obtained. Note how the operator manages to locate and compose the skier’s body from all differently located
and oriented parts assembled in both parents.
Although each jigsaw piece is represented exactly once in
this spanning tree, the correct dimensions and geometrical
feasibility of the corresponding image might not be satisfied. We propose a crossover procedure, analogous to Prim’s
algorithm (Prim 1957), for finding a minimal spanning tree
(MST). This procedure meets the required constraints while
attempting to construct a tree representing a better child solution.
Prim’s algorithm starts from a single vertex and grows the
sub-tree, one edge (and vertex) at a time, selecting at each
step a minimum-weight edge connecting a sub-tree vertex
and an external one. Similarly, we start constructing the puzzle from one piece and grow it by adding a single jigsaw
piece (a graph edge) at a time. At every iteration we review
all edges emanating from the partially grown piece kernel. In
addition to Prim’s basic constraint (of avoiding edge cycles),
the following requirements should be met upon adding a new
piece (associated with a joint edge of an adjacent piece).
First, the known image dimensions must not be breached.
Second, each piece edge must be accounted for only once for
all selected edges (thus avoiding infeasible geometrical configurations due to piece overlap, e.g., two competing pieces
for the same edge of an adjacent third piece). Thus, we ensure that each piece appears exactly once and that the resulting image is geometrically feasible. Finally, the constructed
tree is translated to a chromosome by recording the resulting
locations and orientations of all the pieces. Also recorded are
geometric configurations which are not part of the MST but
are implied by it.
Unlike Prim, instead of selecting literally the minimumweight edge in each iteration, we employ a heuristic based
on intrinsic knowledge of the parents and edge weights. The
procedure first selects an edge appearing in both parents;
e.g., if p3 is in the current kernel, both parents contain the
edge p3 .b − p6 .d, and all of the above constraints are met,
this edge will be selected. Second, if no common edge can be
added to the kernel, the procedure seeks a best-buddy edge
(described below) residing in either parent; e.g., if p3 is in
the current kernel, one parent contains the edge p3 .b − p6 .d,
which is a best-buddy edge, and all other constraints are
met, this edge will be picked. The notion of best-buddy
pieces was first introduced by (Pomeranz, Shemesh, and
Ben-Shahar 2011); two pieces are said to be best-buddies if
each piece considers the other as its most compatible piece
according to the compatibility measure defined. Extending
this notion to best-buddy edges is straightforward. Finally,
if none of the above holds, the procedure follows Prim’s
approach and chooses a minimal-weight edge, for which
all of the required constraints are satisfied. Note that only
edges emanating from the growing piece kernel are considered, in each iteration. Hence, each edge (jigsaw piece) addition introduces additional possible edges (i.e., edges shared
amongst the parents or best-buddy edges) that should be
considered next.
A subtle issue that requires careful consideration is that
of image dimensions. Since a piece orientation is unknown,
it cannot be determined, in advance, whether an N × M or
an M × N frame should be used. Choosing either frame,
in advance, could discard all chromosomes trying to assemble correctly rotated instances of the intended image by 90
or 270 degrees. We overcome this problem by maintaining
initially a flexible frame. After each piece assignment, we
check the farthest boundaries. Assuming M < N , once a
length of M + 1 is reached along one of the dimensions,
the frame must grow up to N along this same dimension.
This allows the images to be assembled in any direction,
with no boundary violations. Thus, the crossover operator
is invariant to the orientation chosen when creating a child.
Experimental results show, indeed, that the GA assembles
the images correctly in different orientations in each generation.
The above described procedure achieves all its predefined goals. The resulting chromosome (image) is valid, each
piece appears exactly once and both image dimensions and
geometric feasibility are maintained. Furthermore, segments
assembled correctly (through the addition at each step of
shared piece edges between parents or best-buddy edges)
are copied to the child solution. Since only piece adjacencies are copied, the segments might appear in different spatial positions and orientations in the two parents (see Figure
2), achieving position- and orientation-independence.
Type 4 Crossover
In this puzzle type each piece is two sided and might need
to be flipped. Segments assembled correctly might appear in
different absolute locations, different orientations, and opposite sides of the puzzle. For example, the same segment
might be located on opposite sides of two different chromosomes while two segments belonging to the same image
(e.g., a person’s head and their body) might be located on
opposite sides of the same chromosome.
Similarly to the discussion on Type 2 puzzles, eight edges
are now associated with each piece, four edges on each piece
face. We mark the piece edges as a, b, c, d and their opposite counterparts as a0 , b0 , c0 , d0 , respectively (e.g., if b is
the right piece edge, b0 is the left edge of the flipped face).
All graph edges will be labeled using the augmented edge
scheme. There are now 64 possible edges between every two
vertices; each piece edge should be considered during every
iteration of the crossover procedure (e.g., the procedure considers adding a piece pi , possibly rotated and/or flipped, to
the growing kernel). To maintain geometrical validity, we
add the following constraint. The flipping side of a piece
edge at the joint boundary of a two-piece configuration may
not be selected for another two-piece configuration, e.g., if
edge pi .b − pj .a0 is selected, then all MST edges incident on
pi .b0 or pi .a are prohibited.
Choosing an edge such as pi .b − pj .a0 effectively implies flipping one of the pieces. Hence, the operator easily
merges correctly assembled segments residing in different
image sides (e.g., copying the first segment, flipping one
piece from the second segment, and copying and flipping
the remaining pieces accordingly).
Interestingly, as a result of the above mentioned procedure, the crossover operator actually assembles both images
concurrently (this approach is different from how a human
would solve such a puzzle, and is considerably superior).
Thus, we achieved a concurrent puzzle solver, that exploits
“easier” segments in each image for achieving a better score.
The solver is completely invariant to the orientation and side
of the assembled images. Indeed, our experiments reveal that
these parameters change constantly in the best chromosome
of each generation.
Experimental Results
In all experiments we used the same GA parameters. Each
generation consists of 1,000 chromosomes, selection is due
to the roulette-wheel selection method (as previously mentioned), and the number of generations (for each run) is
30. (Based on our experience, the latter number seems to
achieve a good balance between accuracy and efficiency.)
Following previous works (Cho, Avidan, and Freeman
2010; Pomeranz, Shemesh, and Ben-Shahar 2011; Gallagher
# Pieces
432
540
805
5,015
10,375
22,755
Direct
94.58%
89.57%
87.76%
93.24%
96.18%
77.43%
Neighbor
94.86%
91.98%
92.07%
93.66%
97.05%
91.07%
Perfect
10
8
6
8
4
1
Run-time
8 sec
11.9 sec
22.1 sec
9.74 min
35.12 min
3.48 hr
Table 2: Accuracy results (under direct and neighbor comparison) on Type 2 puzzles, running the generalized GA five
times on each image of every 20-image set; average of best
result (per image) is shown for each image set.
#
of Pieces
432
540
805
5,015
10,375
22,755
Avg.
Best
94.86%
91.98%
92.07%
93.66%
97.05%
91.07%
Avg.
Worst
93.79%
90.60%
90.76%
93.19%
96.75%
90.47%
Avg.
Avg.
94.44%
91.33%
91.45%
93.42%
96.91%
90.80%
Avg.
Std. Dev.
0.39%
0.49%
0.48%
0.17%
0.11%
0.22%
Table 3: Accuracy results (under neighbor comparison) on
Type 2 puzzles, running generalized GA five times on each
image of every 20-image set; average of best, worst, and average score (and average standard deviation) per image is
given for each image set.
2012; Sholomon, David, and Netanyahu 2013), we evaluated
our scores using the direct comparison and neighbor comparison measures. Direct comparison returns the fraction of
pieces in the assembled puzzle that are in their correct absolute position. Neighbor comparison is the fraction of pairwise piece adjacencies that are correct. We also report the
number of images reconstructed perfectly in every set.
For Type 2 puzzles we tested our solver on all previously established benchmarks (Pomeranz, Shemesh, and
Ben-Shahar 2011; Sholomon, David, and Netanyahu 2013)
using standard tile dimensions of 28 × 28 pixels. These
benchmark datasets contain 20 images of 432-, 540-, 805-,
5,015-, 10,375-, and 22,755-piece puzzles. The largest Type
2 puzzle that has been attempted before is a single 9,600piece puzzle, i.e., we have tackled 20 puzzles more than
twice as large as this size. Due to the stochastic nature of
GAs, we ran the solver five times on each image in every
set, and recorded the best, worst, and average result over
these five runs. The average best results for each image set
(with respect to the direct comparison and neighbor comparison measures) is reported in Table 2. Comparing the results
achieved on the 432-piece dataset to the best results reported
in (Gallagher 2012, Table 6) reveals their algorithm yields a
90.4% accuracy, i.e., we obtained a significant improvement
of over 4%.
For completeness, we also report set averages (according
to the neighbor comparison) of the average and worst results
obtained for each image (over its five runs). Table 3 contains
these additional results for all image sets experimented with.
Despite the random nature of GAs, the results are consistent
(i.e., the small standard deviation suggests that a single run
could have sufficed).
(a) Puzzle - Front
(b) Generation 1 - Front
(c) Generation 2 - Front
(d) Final - Front
(e) Puzzle - Back
(f) Generation 1 - Back
(g) Generation 2 - Back
(h) Final - Back
Figure 3: Solution process of 432-piece Type 4 puzzle: (a) Given puzzle, and best chromosome obtained by the GA in the (b)
first, (c) second, and (d) last generation; (e)–(f) opposite sides of exact same chromosomes. Final chromosome’s accuracy is
100%. Largest Type 4 puzzle solved contains 10,375 pieces.
For Type 4 puzzles we did not experiment with the entire
benchmark, as each puzzle requires two images so running
the entire 400 possible pairs associated with each image set
would have been very tedious, if not infeasible. Instead, we
composed three two-sided puzzles (each containing two images) from the 5,015- and 10,375-piece puzzle sets. The first
(two-sided) puzzle from each set contained two perfectly
solved images as Type 2 puzzles, the second contained one
perfectly solved image and a different image, and the third
contained two images that were not solved perfectly as Type
2 puzzles. All puzzles with at least one perfectly solved side
were again solved perfectly. This result is quite remarkable,
as it attests to the GA’s power to assemble concurrently the
two images by focusing, presumably, on the “easier” one.
Even the more challenging puzzles were solved with far
greater accuracy than when solved separately, again due to
the GA solver’s ability to effectively incorporate the information to the opposite face of a given piece.
Finally, to further challenge our generalized GA solver, in
terms of handling real-world applications, we also attempted
a two-sided, 5,015-piece puzzle whose one side is an image
scene and its other side is a scanned document. Trying to
solve the document as a Type 1 puzzle, we obtained horrendous accuracy, probably due to the ineffectiveness of the
compatibility measure used, as all pieces are mostly white.
This experiment is the closest attempt known to us to automatically solve a large jigsaw shredded document. A scenario in which such a document contains an image on its
other side is not very frequent but is definitely possible.
Given such a puzzle, the generalized GA solver reaches
99.86% accuracy, reassembling almost completely the correct image(s). This was accomplished despite the possible
interference of the white piece faces.
All test results regarding Type 4 puzzles can be viewed in
Table 4. The tests were performed on a modern PC. Average
run-time of the solver on the larger, 10,375-piece puzzles
was 1.35 hours, a considerable improvement compared to
# Pieces
5,015
5,015
5,015
5,015
10,375
10,375
10,375
Images
09, 10
19, 07
05, 14
03, doc
01, 04
15, 11
15, 19
Direct
100%
100%
84.77%
99.92%
100%
100%
99.35%
Neigh.
100%
100%
85.73%
99.86%
100%
100%
99.20%
Run-time
20.08 min
17.88 min
28.57 sec
27.88 min
76.02 min
81.65 min
85 min
Table 4: Results for Type 4 GA on selected puzzles, under
direct and neighbor comparisons, including total run-time.
the 23.5 hours reported by (Gallagher 2012) for a slightly
smaller and far less complex, 9,600-piece Type 2 puzzle.
Conclusion
In this paper, we introduced a new puzzle type, the two-sided
puzzle, where the location, orientation, and face of each
piece is unknown. This type of puzzle is likely to stir great
interest, as it is supposedly the most complex type known;
more importantly, it is highly representative of real-world
applications, such as the reconstruction of shredded documents.
In addition, we presented the first GA-based solver capable of solving both Type 2 and Type 4 puzzles. Our generalized GA solver outperforms, to a significant extent, current
state-of-the-art solvers, as it is capable of solving Type 2
puzzles far more accurately and efficiently. Specifically, our
solver can handle Type 2 puzzles of up to 22,755 pieces, i.e.,
more than twice as large as the size of Type 2 puzzles that
have been reported. In addition, our solver is the only Type
4 solver, as of yet, and it manages, among other tasks, to reconstruct perfectly two-sided puzzles of up to 10,375 pieces.
We believe our contributions can be further generalized to
other, more difficult types of puzzles, in a similar manner.
References
[Altman 1989] Altman, T. 1989. Solving the jigsaw puzzle
problem in linear time. Applied Artificial Intelligence an
International Journal 3(4):453–462.
[Brown et al. 2008] Brown, B.; Toler-Franklin, C.; Nehab,
D.; Burns, M.; Dobkin, D.; Vlachopoulos, A.; Doumas, C.;
Rusinkiewicz, S.; and Weyrich, T. 2008. A system for highvolume acquisition and matching of fresco fragments: Reassembling Theran wall paintings. ACM Transactions on
Graphics 27(3):84.
[Cao, Liu, and Yan 2010] Cao, S.; Liu, H.; and Yan, S. 2010.
Automated assembly of shredded pieces from multiple photos. In IEEE International Conference on Multimedia and
Expo, 358–363.
[Cho, Avidan, and Freeman 2010] Cho, T.; Avidan, S.; and
Freeman, W. 2010. A probabilistic image jigsaw puzzle
solver. In IEEE Conference on Computer Vision and Pattern
Recognition, 183–190.
[Cho et al. 2008] Cho, T.; Butman, M.; Avidan, S.; and Freeman, W. 2008. The patch transform and its applications to
image editing. In IEEE Conference on Computer Vision and
Pattern Recognition, 1–8.
[Deever and Gallagher 2012] Deever, A., and Gallagher, A.
2012. Semi-automatic assembly of real cross-cut shredded
documents. In International Conference on Image Processing, 233–236.
[Demaine and Demaine 2007] Demaine, E., and Demaine,
M. 2007. Jigsaw puzzles, edge matching, and polyomino
packing: Connections and complexity. Graphs and Combinatorics 23:195–208.
[Fornasier and Toniolo 2005] Fornasier, M., and Toniolo, D.
2005. Fast, robust and efficient 2D pattern recognition
for re-assembling fragmented images. Pattern Recognition
38(11):2074–2087.
[Freeman and Garder 1964] Freeman, H., and Garder, L.
1964. Apictorial jigsaw puzzles: The computer solution of a
problem in pattern recognition. IEEE Transactions on Electronic Computers EC-13(2):118–127.
[Gallagher 2012] Gallagher, A. 2012. Jigsaw puzzles with
pieces of unknown orientation. In IEEE Conference on
Computer Vision and Pattern Recognition, 382–389.
[Goldberg, Malon, and Bern 2004] Goldberg, D.; Malon, C.;
and Bern, M. 2004. A global approach to automatic solution
of jigsaw puzzles. Computational Geometry: Theory and
Applications 28(2-3):165–174.
[Holland 1975] Holland, J. H. 1975. Adaptation in natural
and artificial systems. Ann Arbor, MI: University of Michigan Press.
[Justino, Oliveira, and Freitas 2006] Justino, E.; Oliveira, L.;
and Freitas, C. 2006. Reconstructing shredded documents
through feature matching. Forensic Science International
160(2):140–147.
[Koller and Levoy 2006] Koller, D., and Levoy, M. 2006.
Computer-aided reconstruction and new matches in the
forma urbis romae. Bullettino Della Commissione Archeologica Comunale di Roma 103–125.
[Kosiba et al. 1994] Kosiba, D. A.; Devaux, P. M.; Balasubramanian, S.; Gandhi, T. L.; and Kasturi, K. 1994. An automatic jigsaw puzzle solver. In International Conference on
Pattern Recognition, volume 1, 616–618. IEEE.
[Marande and Burger 2007] Marande, W., and Burger, G.
2007. Mitochondrial DNA as a genomic jigsaw puzzle. Science 318(5849):415–415.
[Marques and Freitas 2009] Marques, M., and Freitas, C.
2009. Reconstructing strip-shredded documents using color
as feature matching. In ACM Symposium on Applied Computing, 893–894.
[Morton and Levison 1968] Morton, A. Q., and Levison, M.
1968. The computer in literary studies. In IFIP Congress,
1072–1081.
[Pomeranz, Shemesh, and Ben-Shahar 2011] Pomeranz, D.;
Shemesh, M.; and Ben-Shahar, O. 2011. A fully automated
greedy square jigsaw puzzle solver. In IEEE Conference on
Computer Vision and Pattern Recognition, 9–16.
[Prim 1957] Prim, R. C. 1957. Shortest connection networks and some generalizations. Bell system technical journal 36(6):1389–1401.
[Sholomon, David, and Netanyahu 2013] Sholomon,
D.;
David, O. E.; and Netanyahu, N. S. 2013. A genetic
algorithm-based solver for very large jigsaw puzzles.
In IEEE Conference on Computer Vision and Pattern
Recognition, 1767–1774.
[Sholomon, David, and Netanyahu 2014] Sholomon,
D.;
David, O. E.; and Netanyahu, N. S. 2014. Genetic
algorithm-based solver for very large multiple jigsaw
puzzles of unknown dimensions and piece orientation.
In Proceeding of the 16th Genetic and Evolutionary
Computation Conference. To appear.
[Wang 2000] Wang, C.-S. E. 2000. Determining molecular
conformation from distance or density data. Ph.D. Dissertation, Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science.
[Yang, Adluru, and Latecki 2011] Yang, X.; Adluru, N.; and
Latecki, L. J. 2011. Particle filter with state permutations
for solving image jigsaw puzzles. In IEEE Conference on
Computer Vision and Pattern Recognition, 2873–2880.
[Zhao et al. 2007] Zhao, Y.; Su, M.; Chou, Z.; and Lee, J.
2007. A puzzle solver and its application in speech descrambling. In WSEAS International Conference on Computer Engineering and Applications, 171–176.
| 9cs.NE
|
Approximate Recovery in Changepoint Problems, from `2
Estimation Error Rates
arXiv:1606.06746v2 [stat.ME] 2 Dec 2016
Kevin Lin1
1
James Sharpnack2
Alessandro Rinaldo1
1
Ryan J. Tibshirani
Carnegie Mellon University, 2 University of California at Davis
Abstract
In the 1-dimensional multiple changepoint detection problem, we prove that any procedure
with a fast enough `2 error rate, in terms of its estimation of the underlying piecewise constant
mean vector, automatically has an (approximate) changepoint screening property—specifically,
each true jump in the underlying mean vector has an estimated jump nearby. We also show,
again assuming only knowledge of the `2 error rate, that a simple post-processing step can be
used to eliminate spurious estimated changepoints, and thus delivers an (approximate) changepoint recovery property—specifically, in addition to the screening property described above, we
are assured that each estimated jump has a true jump nearby. As a special case, we focus on the
application of these results to the 1-dimensional fused lasso, i.e., 1-dimensional total variation
denoising, and compare the implications with existing results from the literature. We also study
extensions to related problems, such as changepoint detection over graphs.
Keywords: changepoint detection, fused lasso, total variation denoising, approximate recovery
1
Introduction
Consider the 1-dimensional multiple changepoint model
yi = θ0,i + i ,
i = 1, . . . , n,
(1)
where i , i = 1, . . . , n are i.i.d. errors, and θ0,i , i = 1, . . . , n is a piecewise constant mean sequence,
having a set of changepoints
S0 = i ∈ {1, . . . , n − 1} : θ0,i 6= θ0,i+1 .
This is a well-studied problem, and there is a large body of literature on estimation of the piecewise
constant mean vector θ0 ∈ Rn in this model, as well as detection of its changepoints θ0 . Though
estimation (of θ0 ) and detection (of its changepoints) are clearly related pursuits, they are different
enough that most works on the changepoint problem are focused on one or the other. For example,
1-dimensional total variation denoising (Rudin et al. 1992), i.e., the 1-dimensional fused lasso (Tibshirani et al. 2005), has been primarily studied from the perspective of its estimation properties.
Meanwhile, segmentation methods like binary segmentation (Vostrikova 1981, Venkatraman 1992)
and wild binary segmentation (Fryzlewicz 2014) have been mostly studied for their detection properties. In this paper, we assert that the estimation and detection problems are very closely linked,
in the following sense: any estimator with `2 estimation error guarantees automatically has certain
approximate changepoint detection guarantees, and not surprisingly, a faster `2 estimation error rate
here translates into a stronger statement about approximate detection. We use this general link to
establish new approximate changepoint recovery results for the 1d fused lasso, an estimator that is
given central focus in our work.
1
1.1
Background and related work
Given a data vector y ∈ Rn from a model as in (1), the 1-dimensional fused lasso (1d fused lasso,
or simply fused lasso) estimate is defined by
n
n−1
X
1X
θb = argmin
(yi − θi )2 + λ
|θi − θi+1 |,
2 i=1
θ∈Rn
i=1
(2)
where λ ≥ 0 serves as a tuning parameter. This was proposed by Tibshirani et al. (2005)1 , though
the same idea had been proposed in signal processing much earlier, under the name total variation
(TV) denoising, by Rudin et al. (1992). There has been plenty of statistical theory developed for the
fused lasso, e.g., Mammen & van de Geer (1997), Davies & Kovac (2001), Rinaldo (2009), Harchaoui
& Lévy-Leduc (2010), Qian & Jia (2012), Rojas & Wahlberg (2014), Dalalyan et al. (2014). In
particular, Mammen & van de Geer (1997) and Dalalyan et al. (2014) derived `2 error rates for the
fused lasso, under different settings (different assumptions on θ0 ). We review these in Section 2.1, and
compare the latter to a new related result that we establish in Section 3. Harchaoui & Lévy-Leduc
(2010) and Rojas & Wahlberg (2014) derived approximate changepoint recovery properties for the
fused lasso. We review these in Section 2.2, and compare them to our own results on approximate
recovery in Section 4.
The literature on general approaches in multiple changepoint detection is enormous, and we do
not give an extensive overview, but we do summarize some relevant work of Donoho & Johnstone
(1994, 1998), Fryzlewicz (2007), Boysen et al. (2009), Fryzlewicz (2014), Frick et al. (2014), Fryzlewicz (2016) in Sections 2.1 and 2.2, and revisit these results in more detail in Section 4 when we
compare them to our new approximate changepoint recovery results. Extensions of the fused lasso
such as trend filtering and the graph-based fused lasso have been analyzed by, e.g, Sharpnack et al.
(2012), Tibshirani (2014), Wang et al. (2016), Hutter & Rigollet (2016), which are discussed later
in Section 6.
1.2
Notation
For a vector θ ∈ Rn , we write S(θ) for the set of its changepoint indices, i.e.,
S(θ) = i ∈ {1, . . . , n − 1} : θi 6= θi+1 .
b for the changepoints of the mean θ0 in (1), and the fused
We abbreviate S0 = S(θ0 ) and Sb = S(θ)
lasso estimate θb in (2), respectively. Throughout, we will use the words “changepoint” and “jump”
interchangeably. We will also make use of the following quantities defined in terms of θ0 . The size of
S0 is denoted s0 = |S0 |. For convenience, we write S0 = {t1 , . . . , ts0 }, where 1 ≤ t1 < . . . < ts0 < n,
and by convention, t0 = 0, ts0 +1 = n. The smallest distance between jumps in θ0 is denoted by
Wn =
min
i=0,1...,s0
(ti+1 − ti ),
(3)
and the smallest distance between consecutive levels of θ0 by
Hn = min |θ0,i+1 − θ0,i |.
i∈S0
(4)
Our notation here makes the dependence of Wn , Hn on n explicit (of course stemming from the fact
that the mean vector θ0 itself changes with n, though for simplicity we suppress this notationally.)
1 In Tibshirani et al. (2005), the authors actually used an additional ` penalty on θ itself, to induce componentwise
1
sparsity; here we do not consider this extension, and simply refer to the estimator in (2) as the fused lasso.
2
For a matrix D ∈ Rm×n , we write DS to extract rows of D indexed by a subset S ⊆ {1, . . . , m},
and D−S as shorthand for D−S , where −S = {1, . . . , m}\S. Unless otherwise specified, the notation
D ∈ R(n−1)×n will be used to denote the difference operator
−1
1
0 ... 0
0 −1
1 ... 0
D= .
(5)
.
.
.. ...
..
0
0 . . . −1 1
√
n
For a vector x ∈ R , we define its scaled `2 norm kxkn = kxk2 / n, and its discrete total variation
TV(x) =
n−1
X
|xi − xi+1 | = kDxk1 .
(6)
i=1
For two discrete sets A, B, we define the metrics
d(A|B) = max min |a − b| and dH (A, B) = max d(A|B), d(B|A)}.
b∈B a∈A
(7)
The former metric can seen as a one-sided screening distance from B to A, measuring the furthest
distance of an element in B to its closest element in A. The latter metric is traditionally known as
the Hausdorff distance between A and B. Note that if A is empty, then we have d(A|B) = 0, and if
B is empty, then d(A|B) = ∞; this makes dH (A, B) = ∞ if either A or B are empty.
For deterministic sequences an , bn we write an = O(bn ) to denote that an /bn is bounded for large
enough n, an = Ω(bn ) to denote that bn /an is bounded for large enough n, and an = Θ(bn ) to denote
that both an = O(bn ) and an = Ω(bn ). We also write an = o(bn ) to denote that an /bn → 0, and
an = ω(bn ) to denote that bn /an → 0. Finally, we write An = OP (Bn ) for random sequences An , Bn
to denote that An /Bn is bounded in probability, and An = oP (Bn ) to denote that An /Bn → 0 in
probability.
1.3
Summary of results
A summary of our contributions is as follows.
• New `2 error analysis for the fused lasso, under strong sparsity. In Section 3, we give
a new `2 estimation error analysis for the fused lasso, in the case s0 = O(1), which we refer to
as the “strong sparsity” case. Theorem 4 provides the bound
log n log log n
2
b
kθ − θ0 kn = OP
,
n
for the fused lasso estimate θb in (2). This is sharper than the previously established error rate
of log2 n/n, from Dalalyan et al. (2014), for the fused lasso under strong sparsity, and quite
close to the “oracle” rate of log n/n under strong sparsity, as we discuss in Remark 10. Our
theorem also applies beyond the case of a constant sparsity level s0 , and gives an explicit error
bound in terms of s0 . We believe that the proof of Theorem 4 is interesting in its own right,
as it leverages a new quantity that we call a lower interpolant to approximate the fused lasso
estimate in a certain sense using 2s0 + 2 piecewise monotonic segments, which allows for finer
control of the sub-Gaussian complexity.
• Bound on the screening distance, based on `2 error. In Section 4.1, we derive a bound
e of any estimator θ,
e given
on the sreening distance from S0 to the detected changepoints S(θ)
2
e
a bound on its `2 error rate kθ − θ0 kn = OP (Rn ). Specifically, in Theorem 8, we show that
nRn
e
d S(θ) | S0 = OP
.
Hn2
3
To emphasize, this bound on the screening distance is agnostic about the details of the estie provided that its `2 error rate Rn is known. As two principal applications, we plug
mator θ,
in the known error rate Rn for the fused lasso under two different settings—weak and strong
sparsity—to derive new screening results on the fused lasso in Corollaries 9 and 10. Perhaps
surprisingly (since these screening bounds are not based on fine-grained analysis of the fused
lasso, but on achieved `2 rates alone), these results provide interesting conclusions in each of
their own settings, as we discuss in Remarks 14 and 15.
• Bound on the Hausdorff distance, based on `2 error and a post-processing step. In
Section 4.2, we give a bound on the Hausdorff distance between S0 and the detected changee of any estimator θ,
e given a bound kθe − θ0 k2 = OP (Rn ) and a simple filtering-based
points S(θ)
n
e that occur far away from elements of S0 .
technique to remove spurious changepoints in S(θ)
e of changepoints satisfies
In particular, Theorem 11 states that the filtered set SF (θ)
e S0 ≤ nRn νn → 1 as n → ∞,
P dH SF (θ),
Hn2
where νn is any diverging sequence (i.e., diverging as slowly as desirable). As two applications,
we consider post-processing the changepoints from the fused lasso estimator, in the weak and
strong sparsity settings, in Corollaries 14 and 15. We compare these to existing approximate
changepoint recovery results in the literature in Remarks 19 and 20; the summary is that
under strong sparsity, our result on the post-processed fused lasso is comparable with the best
known recovery results, but under weak sparsity, our result is worse than the guarantees given
in Frick et al. (2014), Fryzlewicz (2014, 2016) for other changepoint estimators. It should be
reiterated that, unlike other results in the literature which are based on detailed analyses of
specific changepoint estimators, our results are generic and based only on `2 error properties,
making them widely applicable. Therefore, a lack in sharpness in some cases, such as the weak
sparsity case, is perhaps not unexpected.
• Practical guidelines for post-processing. In Section 4.3, we present a modification of the
aforementioned post-processing rule, which guarantees that the filtered set has at most 3e
s+2
e In Section 5, we describe a data-driven procedure to determine an
elements, where se = |S(θ)|.
appropriate threshold level for the filter, and we also conduct detailed empirical investigations
of our proposals.
• Extension to piecewise linear segmentation, and graph changepoint detection. In
Section 6, we give extensions of our screening results to two related settings: piecewise linear
segmentation and changepoint detection over graphs. For piecewise linear segmentation, the
main screening result is in Theorem 19, and its specialization to the trend filtering estimator
is in Corollary 21; for graph changepoint detection, the main result is in Theorem 23, and its
specialization to the 2d fused lasso estimator is in Corollary 24.
2
Preliminary review of existing theory
We review existing statistical theory for the fused lasso, first on `2 estimation error, and then on
(approximate) changepoint recovery.
2.1
Review: `2 estimation error
We begin by describing two major results on the quantity
kθb − θ0 k2n ,
4
the squared `2 estimation error between the fused lasso estimate θb in (2) and the mean θ0 in (1). In
somewhat of an abuse of notation, we will simply refer to the above quantity as the `2 estimation
error, or estimation error for short.
The first result, from Mammen & van de Geer (1997), studies what may be called the “weak
sparsity” case, in which the total variation of θ0 is controlled. Before stating this, we recall that a
random variable Z is said to have a mean zero sub-Gaussian distribution provided that
E(Z) = 0 and P(|Z| > t) ≤ M exp − t2 /(2σ 2 ) for t ≥ 0,
(8)
for some constants M, σ > 0.
Theorem 1 (Fused lasso error rate, weak sparsity setting, Theorem 10 of Mammen &
van de Geer 1997). Assume the data model in (1), with errors i , i = 1, . . . , n i.i.d. from a subGaussian distribution as in (8). Also assume that TV(θ0 ) ≤ Cn , for a nondecreasing sequence Cn .
−1/3
Then for a choice of tuning parameter λ = Θ(n1/3 Cn ), the fused lasso estimate θb in (2) satisfies
kθb − θ0 k2n = OP (n−2/3 Cn2/3 ).
Remark 1 (Consistency, optimality). This shows that the fused lasso estimator is consistent
when Cn = o(n). When Cn = O(1), its estimation error rate is n−2/3 , which is in fact the minimax
optimal rate as θ0 varies over the class of signals with bounded total variation, i.e., θ0 ∈ {θ ∈ Rn :
TV(θ) ≤ C} for a constant C > 0 (Donoho & Johnstone 1998). For explanations of the above
theorem and this minimax result, in notation that is more consistent with that of the current paper,
see Tibshirani (2014).
The second result, from Dalalyan et al. (2014), studies what may be called the “strong sparsity”
case, in which the number of changepoints s0 in θ0 is controlled.
Theorem 2 (Fused lasso error rate, strong sparsity setting, Proposition 4 of Dalalyan
2
et al. 2014). Assume the data model in (1),
pwith errors i , i = 1, . . . , n drawn i.i.d. from N (0, σ ).
Then for a choice of tuning parameter λ = 2n log(n/δ), the fused lasso estimate θb in (2) satisfies
s0 log(n/δ)
n
kθb − θ0 k2n ≤ c
log n +
,
n
Wn
with probability at least 1 − 2δ, for all δ > 0 and all n ≥ N , where c, N > 0 are constants, and recall
Wn is the minimum distance between jumps in θ0 , as in (3).
Remark 2 (The roles of s0 , Wn ). When the number of jumps s0 in θ0 grows quickly enough with
n, the error rate in Theorem 2 will become worse than that in Theorem 1. Given s0 jumps, in the
best case, the minimum gap Wn between jumps scales as Wn = Θ(n/s0 ), which delivers a rate of
s0 log2 n/n + s20 log n/n in Theorem 2. When s0 scales faster than n1/6 (log n)−1/2 , we can see that
this is slower than the n−2/3 rate delivered by Theorem 1 (assuming Cn = O(1)).
Of course, Theorem 2 is most useful when s0 = O(1). When this is true, and additionally Wn =
Θ(n), we see that the theorem implies that the fused lasso has `2 error kθb − θ0 k2n = OP (log2 n/n).
This is a very fast rate, nearly equal to the “parametric rate” of 1/n associated with estimating a
finite-dimensional parameter.
Remark 3 (Alternative fused lasso error rate, strong sparsity setting). When s0 = O(1),
Proposition 2 in Harchaoui & Lévy-Leduc (2010) proves that the fused lasso p
has estimation error
kθb − θ0 k2n = O(log n/n) with probability approaching 1, under a choice λ = Θ( log n/n3 ). But the
b which we
authors must also assume that the number of changepoints in the fused lasso estimate θ,
b
might denote as sb = |S|, is bounded with probability tending to 1. This seems to be an unrealistic
assumption, given the required scaling for λ. Theoretically, we remark that such a small choice of
5
p
λ, on the order of log n/n3 , does not √
match the much larger choices dictated
p by Theorems 2 and
4, both on the order of approximately n. Empirically, when λ scales as log n/n3 , we find that
the number of estimated changepoints in θb often grows very large, even when θ0 has few jumps and
the signal-to-noise ratio is quite high.
Remark 4 (Comparable error rates of other estimators). Various other estimators obtain
comparable estimation error rates toPthose descibed above for the fused lasso. The
Potts estimator,
Pn−1
n−1
defined by replacing the `1 penalty i=1 |θi − θi+1 | in (2) with the `0 penalty i=1 1{θi 6= θi+1 },
and denoted say by θbPotts , has been shown to satisfy kθbPotts − θ0 k2n = O((log n/n)2/3 ) a.s. when
TV(θ0 ) = O(1), and kθbPotts − θ0 k2n = O(log n/n) a.s. when s0 = O(1), by Boysen et al. (2009).
Wavelet denoising (under weak conditions on the wavelet basis), denoted by θbwav , has been shown
to satisfy Ekθbwav − θ0 k2n = O(n−2/3 ) when TV(θ0 ) = O(1), by Donoho & Johnstone (1998), and
Ekθbwav − θ0 k2n = O(log2 n/n) when s0 = O(1), by Donoho & Johnstone (1994). Combining unbalanced Haar (UH) wavelets with a basis selection method, Fryzlewicz (2007) gave an estimator θbUH
with EkθbUH − θ0 k2n = O(log2 n/n2/3 ) when TV(θ0 ) = O(1), and EkθbUH − θ0 k2n = O(log2 n/n) when
s0 = O(1). Though they are not written in this form, the results in Fryzlewicz (2016) imply that his
“tail-greedy” unbalanced Haar (TGUH) estimator, θbTGUH , satisfies kθbTGUH − θ0 k2n = O(log2 n/n)
with probability tending to 1, when s0 = O(1).
2.2
Review: changepoint recovery
Next, we review the relevant results on the quantities
b S0 ).
d(Sb | S0
or dH (S,
b in the fused lasso
The former is the screening distance from S0 to the set of changepoints Sb = S(θ)
b
b
estimate θ in (2); the latter is the Hausdorff distance between S0 and S; recall, both metrics were
defined in (7). We use the term “approximate screening” to mean that d(Sb | S0 ) is controlled, and
b S0 ) is controlled, though often times we will drop the
“approximate recovery” to mean that dH (S,
word “approximate” from either term, for brevity. Below we summarize two results from Harchaoui
& Lévy-Leduc (2010).
Theorem 3 (Fused lasso approximate screening and recovery results, strong sparsity
setting, Propositions 3 and 4 of Harchaoui & Lévy-Leduc 2010). Assume the data model in
(1), where the errors i , i = 1, . . . , n are i.i.d. from a sub-Gaussian distribution as in (8). Assume
also that (i) s0 = O(1), (ii) P(b
s ≥ s0 ) → 1, and that rn is a sequence satisfying (iii) rn ≤ Wn , (iv)
rn = ω(max{log n/Hn2 , λ/Hn }), where, recall, Wn is the minimum distance between changepoints in
θ0 , as defined in (3), and Hn is the minimum gap betwen levels of θ0 , as defined in (4). Then the
fused lasso estimate θb in (2) with tuning parameter λ satisfies
P d(Sb | S0 ) ≤ rn → 1 as n → ∞.
Under assumptions (i), (ii’) P(b
s = s0 ) → 1, (iii), and (iv’) rn = ω(max{log n/Hn2 , log (n5 /λ2 )/Hn2 }),
we instead have
b S0 ) ≤ rn → 1 as n → ∞.
P dH (S,
Remark 5 (Stringency of conditions). Assumption (ii’) in the above result, needed for the
bound on the Hausdorff distance, states that the number of estimated changepoints sb in θb equals
the number of changepoints s0 in θ0 with probability tending to 1, which is of course a very strong
assumption. Assumption (ii), needed for the bound on the one-sided screening distance, states that
sb ≥ s0 , which is itself fairly strong, though believable if λ is chosen to be small enough.
The tuning parameter λ in fact plays an important role in the achieved rates rn in Theorem 3.
In order to satisfy condition (iv) on rn , a choice of λ = log n/Hn gives the tightest possible scaling
6
for rn . Then we require rn to grow faster than log n, e.g., at the rate log2 n. This choice is basically
the same as that discussed in Harchaoui & Lévy-Leduc (2010); with this choice, their results show
that the screening distance achieved by the fused lasso estimator is at most log2 n, with probability
tending to 1. However, the choice of λ here is worrisome—it is considerably smaller than the choices
known to achieve reasonable error rates, specifically
in the strong sparsity setting, with s0 = O(1),
√
where we expect λ to scale at something like a n rate (see Theorems 2 and 4). With such a small
b
choice of λ = log n/Hn , there would likely be a very large number of estimated changepoints in θ,
rendering a quantity like the screening distance uninteresting.2 The same critique could made be
about condition (iv’), needed for the bound on the Hausdorff distance. In particular, condition (ii’)
seems unrealistic unless λ is chosen to be much larger.
√
For the reasons just described, we will consider a larger scaling of λ = Θ( n), when comparing
Theorem 3 to our new results on changepoint screening and recovery in Sections 4.1 and 4.2.
Remark 6 (Other fused lasso recovery results). Several other results have appeared in the
literature regarding changepoint recovery for the fused lasso. Rinaldo (2009) studied exact recovery
of changepoints (in which the achieved Hausdorff distance would be zero). There is an error in the
proof of his Theorem 2.3, which invalidates the result.3 Qian & Jia (2012) studied a modification
of the fused lasso defined by transforming the fused lasso problem (2) into a lasso problem with
particular design matrix X, and then applying a step that “preconditions” y and X. The authors
concluded that exact recovery is possible with probability
√ at tending to 1, as long as the minimum
on
signal gap and tuning parameter satisfy Hn ≥ λ = ω( log n). This is a very strong requirement
√
the scaling of the signal gap Hn ; Sharpnack et al. (2012) showed that, when Hn ≥ ω( log n), even
simple pairwise thresholding (i.e., thresholding based on the observed absolute differences |yi −yi+1 |,
i = 1, . . . , n − 1) achieves exact recovery. Most recently, Rojas & Wahlberg (2014) established an
impossibility result for the fused lasso estimator when θ0 exhibits a “staircase” pattern, which means
that DS0 θ0 has two consecutive positive or negative values; specifically, these authors proved that for
b S0 )/n remains bounded away from zero with nonzero
such a staircase pattern, the quantity dH (S,
asymptotic probability. For non-staircase patterns in θ0 , the authors also showed, under certain
b S0 )/n converges to zero in probability.
assumptions, that dH (S,
Remark 7 (Comparable recovery properties of other estimators). It is worth describing
relevant changepoint recovery properties of various methods in the literature. Boysen et al. (2009)
showed that the Potts estimator, denoted by θbPotts , satisfies dH (S(θbPotts ), S0 ) = O(log n) a.s., when
s0 = O(1). Frick et al. (2014) proposed a simultaneous multiscale changepoint estimator (SMUCE),
using an `0 -penalized optimization problem (like the Potts estimator), and under weak assumptions
on Wn , Hn , proved that their estimator θbSMUCE satisfies dH (S(θbSMUCE ), S0 ) = O(log n/Hn2 ) with
probability tending to 1. There is quite a large body of literature on binary segmentation (BS). To
the best of our knowledge, the sharpest analysis for BS is in Fryzlewicz (2014), who also proposed
and analyzed a “wild” (i.e., randomized) variant of the method (WBS). Denoting these two estimators by θbBS and θbWBS , Fryzlewicz (2014) established that dH (S(θbBS ), S0 ) = O(n2 log n/(Wn2 Hn2 ))
and dH (S(θbBS ), S0 ) = O(log n/Hn2 ), both with probability tending to 1, and both under certain
restrictions on Wn , Hn , these restrictions being stronger for BS than for WBS. Very recently, Fryzlewicz (2016) proved that his tail-greedy unbiased Haar estimator, denoted by θbTGUH , satisfies
dH (S(θbTGUH ), S0 ) = O(log2 n) with probability tending to 1, under weak conditions on Wn , Hn . All
of these results will be revisited in greater detail in Remarks 19 and 20.
Lastly, it should be noted that many of the methods described here also come with a guarantee
(under possibly additional conditions) that they correctly identify the number of changepoints s0 in
θ0 , with probability tending to 1. We refer the reader to the references above, for details.
2 Of course, with λ = 0, the screening distance achieved by the fused lasso is trivially zero. Hence, when studying
screening distance, it is implicitly understood that some other aspect of θb must be kept in balance. In our work, we
study screening distances while maintaining that θb must exhibit good estimation performance, as measured by its `2
error rate.
3 See the correction note posted at http://www.stat.cmu.edu/ arinaldo/Fused_Correction.pdf.
~
7
3
Error analysis under strong sparsity
In this section, we derive a new `2 estimation error bound for the fused lasso in the strong sparsity
case, improving on the result of Dalalyan et al. (2014) stated in Theorem 2. Our proof is based on
the concept of a lower interpolant, which as far as we can tell is a new idea that may be of interest
in its own right. We first state our error bound.
Theorem 4 (Fused lasso error rate, strong sparsity setting). Assume the data model in (1),
with errors i , i = 1, . . . , n i.i.d. from a sub-Gaussian distribution as in (8). Then for a choice of
tuning parameter λ = (nWn )1/4 , the the fused lasso estimate θb in (2) satisfies
!
r
n
s
0
(log s0 + log log n) log n +
,
kθb − θ0 k2n ≤ γ 2 c
n
Wn
with probability at least 1 − exp(−Cγ), for all γ > 1 and n ≥ N , where c, C, N > 0 are constants.
Remark 8 (The roles of s0 , Wn ). When s0 grows quickly enough with n, the error rate provided
by the above theorem will become worse than the weak sparsity rate in Theorem 1. Given s0 evenly
3/2
spaced jumps, so that Wn = Θ(n/s0 ), the rate in Theorem 4 is (log s0 + log log n)s0 log n/n + s0 /n;
2/9
−2/3
when s0 grows faster than n , this is slower than the n
rate in Theorem 1 (assuming Cn =
O(1)). Theorem 4 gives the fastest rate when s0 = O(1), Wn = Θ(n), this being (log n log log n)/n),
an improvement over the rate in Theorem 2. This comparison, and the comparison to other results
in the literature, will be drawn out in more detail in the last remark of this section.
Remark 9 (Expectation bound). An expectation bound follows more of less directly from the
high probability bound in Theorem 4. Define the random vairable
kθb − θ0 k22
M=
p
,
c2 s0 ((log s0 + log log n) log n + n/Wn )
√
which we know has the tail bound P(U > u) ≤ exp(−C u) for u > 1, and observe that
Z ∞
Z ∞
√
E(U ) =
P(U > u) du ≤ 1 +
exp(−C u) du.
0
1
The right-hand side is a finite constant, and this gives the result
s0
Ekθb − θ0 k2n ≤ c
(log s0 + log log n) log n +
n
r
!
n
,
Wn
where the constant c > 0 is adjusted to be larger, as needed.
Here is an overview of the proof of Theorem 4. The details are deferred until Appendix A, and
the proofs of the lemmas stated below are given in Appendix B. We consider a decomposition
b
θb − θ0 = P0 (θb − θ0 ) + P1 θ,
where P0 is the projection matrix onto the piecewise constant structure inherent to the mean θ0 ,
and P1 = I − P1 . To give more detail, recall that we write S0 = {t1 , . . . , ts0 } for the changepoints
in θ0 , ordered as in t1 < . . . < ts0 , and we write t0 = 0 and ts0 +1 = n for convenience. Furthermore,
define Bj = {tj + 1, . . . , tj+1 }, and write 1Bj ∈ Rn for the indicator of block Bj , for j = 0, . . . , s0 .
With this notational setup, we may now define P0 as the projection onto the (s0 + 1)-dimensional
linear subspace R = span{1B0 , . . . , 1Bs0 }. It is common practice (e.g., see van de Geer (2000)) to
8
bound the estimation error by bounding the empirical process term > (θb − θ0 ), where ∈ Rn is the
vector of errors in the data model (1). Using the decomposition above, this becomes
> (θb − θ0 ) = > δb + > x
b,
b The parameter δb lies in an (s0 + 1)-dimensional space,
where we define δb = P0 (θb − θ0 ) and x
b = P1 θ.
>b
which makes bounding δ relatively easy. Bounding the term > x
b requires a much more intricate
argument, which is spelled out in the following lemmas. Lemma 5 is a deterministic result ensuring
the existence of what we call the lower interpolant zb to the vector x
b. This interpolant approximates
x
b using roughly 2s0 +2 monotonic pieces, and its empirical process term > zb can be finely controlled,
as shown in Lemma 6. The residual from the interpolant approximation, denoted w
b=x
b − zb, has
an empirical process term > w
b that is more crudely controlled, in Lemma 7. Put together, as in
> x
b = > zb + > w,
b gives the final control on > x
b.
Before stating Lemma 5, we define the class of vectors containing the lower interpolant. Given
any collection of changepoints t1 < . . . < ts0 (and t0 = 0, ts0 +1 = n), let M be the set of “piecewise
monotonic” vectors z ∈ Rn , with the following properties, for each i = 0, . . . , s0 :
(i) there exists a point t0i such that ti + 1 ≤ t0i ≤ ti+1 , and the absolute value |zj | is nonincreasing
over the segment j ∈ {ti + 1, . . . , t0i }, and nondecreasing over the segment j ∈ {t0i , . . . , ti+1 };
(ii) the signs remain constant on the monotone pieces,
sign(zti ) · sign(zj ) ≥ 0,
sign(zti+1 ) · sign(zj ) ≥ 0,
j = ti + 1, . . . , t0i ,
j = t0i + 1, . . . , ti+1 .
Now we state our lemma that characterizes the lower interpolant.
Lemma 5. Given changepoints t0 < . . . < ts0 +1 , and any x ∈ Rn , there exists a vector z ∈ M (not
necessarily unique), such that the following statements hold:
kD−S0 xk1 = kD−S0 zk1 + kD−S0 (x − z)k1 ,
√
4 s0
kzk2 ,
kDS0 xk1 = kDS0 zk1 ≤ kD−S0 zk1 + √
Wn
kzk2 ≤ kxk2 and kx − zk2 ≤ kxk2 ,
(9)
(10)
(11)
where D ∈ R(n−1)×n is the difference matrix in (5). We call a vector z with these properties a lower
interpolant to x.
Loosely speaking, the lower interpolant zb can be visualized by taking a string that lies initially
on top of x
b, is nailed down at the changepoints t0 , . . . ts0 +1 , and then pulled taut while maintaining
that it is not greater (elementwise) than x
b, in magnitude. Here “pulling taut” means that kDb
z k1 is
made small. Figure 1 provides illustrations of the interpolant zb to x
b for a few examples.
Note that zb consists of 2s0 + 2 monotonic pieces. This special structure leads to a sharp concentration inequality. The next lemma is the primary contributor to the fast rate given in Theorem 4.
Lemma 6. Given changepoints t1 < . . . < ts0 , there exists constants cI , CI , NI > 0 such that when
∈ Rn has i.i.d. sub-Gaussian components satisfying (8),
!
p
|> z|
> γcI (log s0 + log log n)s0 log n ≤ 2 exp − CI γ 2 c2I (log s0 + log log n) ,
P sup
z∈M kzk2
for any γ > 1, and n ≥ NI .
9
●
8
10
●
●●
●●●
●●
●
● ●● ●●
●
●●●● ● ●●
●
●
●
●
●●
● ● ●
●
●
●
●
●
●
● ● ● ●● ●
●
●●● ●●
● ●
●●
●● ●●
●● ●●●●
●●
● ●●
●
●●
● ●●
●
● ●●●
●
●●
●●
●●●
●
● ●●● ●●● ● ●●
●
●
●●
●●●
● ● ● ●●
● ●●
●
●● ●
●
●●●
●
●
●
●● ●●●
●●●●
●●
●
●
●
●
●
●
●
●
●●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
● ●
●●
●●
●
●
●●
●●●
●
●●
●
●
●
●
●●●● ● ● ●
●● ●
●●
●●
●●
●
●
●●●●●
●
●●
●●●
●
●
●
●
●
●●
●●
●
●●●
●●
●
●
●●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
● ●●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●●
●● ● ●
●●●● ●
●●
●●
● ●●
●●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●●● ● ●●●
●● ● ●●
●
● ●●●●●● ● ●●●
●●●
● ●● ● ●
●●
●●●
●● ●
●
●
●
●●●
●
●
●
●●● ●●● ●
●●
●●
●
●●
●
●
●● ● ●●●
●
●
●
●●
●
●
●
●●
●
●
●
● ●●
●
●
●●●
●●●●●
●
●●●
●
●
●● ●●
●
●
●
●
●●●●●●●
●
●●
●
●
●●●
●
●●
●●●
●
●
● ●
●●
●
●●
●
●
●● ●●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
● ●
●
● ● ●● ●●●
●
●
●●
●
●● ●
● ●
●●
●●
●●●
●● ●
●●●
●
●● ●● ●●
●● ● ●●
●● ●●
●
●
●●
●●●●●
●●
●● ●● ●
● ●
●●
●
● ●●
●●● ●●
●●●●
●
●
●●●●
●
●●
●
●
●
●● ●
●
●●
●● ●●● ●●
●
●
●●●
●
●
●
●
●
●● ●
●● ● ●
●●
●
●
●
● ● ●● ●
●
●
●
●
●
0
200
400
600
6
4
2
0
−2
800
0
200
400
600
800
600
800
Index
−1.5
−3
−2
−0.5 0.0
−1
0.5
0
1.0
Index
●
●●
●●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●●●●●
●
●●
●●●
●
●
●
●
●
●
●●
●
●●
●●
●●●
●
● ●●
●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
● ●●
● ●
●● ●
●
●
●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●
●●
●●●●●●
●●
●●●●
●
●● ●
●●●
●●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●●●
●●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●●
●
●
●● ●
●
●
●
●
●●
●
●●
●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●●●
●●●
● ●●
●●●●
●
●
●
●
●
●●●
●● ●
●
● ●●●
●●
●●
●●●
●
●
●
●● ● ● ●●●● ●
● ●
●
● ● ●● ●
●● ● ●
●
●
●
●
●
●●
●
● ●●
●
●
●
●
●●●● ●
●
●
●
●
●
● ●
●
●●
●
●● ● ●
●●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●●
●
●●
●●
●●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
●●●
●
●●
●
●●●
●
●●●●
●
●
●
●●
●●
●●
●●
●●
●
●●
●●
●
●●
●●
●●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●●
●●
● ●● ●
● ●●
●●
●●●●●
●
●●●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●● ●●
●
●●●
●
●
●●●
●●●
●
●●
●
●
●
●
●
●
●●● ●●
●
●
●
● ●●●
● ●●●
●●●
●●
●
●
●
●●
●
●
●● ●
●●●
●●
●●
●●●●
●
●
●●
●● ●● ● ●
●
●
1
−5
0
5
●
0
200
400
600
800
0
200
400
Figure 1: The lower interpolants for two examples (in the left and right columns), each with n = 800 points.
In the top row, the data y (in gray) and underlying signal θ0 (red) are plotted across the locations 1, . . . , n.
Also shown is the fused lasso estimate θb (blue). In the bottom row, the error vector x
b = P1 θb is plotted (blue)
as well as the interpolant (black), and the dotted vertical lines (red) denote the changepoints t1 , . . . ts0 of θ0 .
Finally, the following lemma controls the residuals, w
b=x
b − zb.
Lemma 7. Given changepoints t1 < . . . < ts0 , there exists constants cR , CR > 0 such that when
∈ Rn has i.i.d. sub-Gaussian components satisfying (8),
√
|> w|
1/4
P sup p
> γcR (ns0 )
≤ 2 exp(−CR γ 2 c2R s0 ),
⊥
kD−S0 wk1 kwk2
w∈R
for any γ > 1, where R⊥ is the orthogonal complement of R = span{1B0 , . . . , 1Bs0 }.
We conclude this section with a remark comparing Theorem 4 to related results in the literature.
Remark 10 (Comparison to Theorem 2). We compare Theorem 4 to the strong sparsity result
in Dalalyan et al. (2014), as stated p
in Theorem 2. For any s0 , Wn , the former rate is sharper than
the latter, since log s0 ≤ log n and n/Wn ≤ n/Wn . Moreover, when s0 = O(1) and Wn = Θ(n),
the rates are (log n log log n)/n versus log2 n/n, in Theorems 4 and 2 respectively. Furthermore,
in this √
setting, we note that the scaling of the tuning parameter investigated by both theorems is
λ = Θ( n).
As discussed in Remarks 3 and 4, essentially all rates from comparable estimators in the strong
sparsity case scale as log2 n/n, with the exception being the Potts estimator, which has a rate of
log n/n. Therefore the improvement from log2 n/n to (log n log log n)/n offered by Theorem 4 could
certainly be viewed as nontrivial. An error rate faster than log n/n in the strong sparsity case seems
likely unattainable by any method, as Donoho & Johnstone (1994) showed that an oracle wavelet
estimator (that is allowed the optimal choice of wavelet threshold for each problem instance) still
has an expected estimation error on the order of log n/n.
10
4
Approximate changepoint screening and recovery
We develop results on approximate screening of changepoints by the fused lasso, and approximate
recovery of changepoints after a post-processing step has been applied to the fused lasso estimate.
A distinctive feature of our results is that their proofs based on only the `2 estimation error rates
achieved by the fused lasso. In fact, in their most general form, our results imply certain changepoint
screening and recovery properties for any estimation method that has a known `2 error rate, which
clearly has implications well beyond the fused lasso.
4.1
Results on approximate changepoint screening
We present a theorem that takes a general estimator θe of θ0 , with a known `2 estimation error rate,
e
and infers a bound on the screening distance between changepoints of θ0 and those of θ.
Theorem 8 (Generic screening result). Let θ0 ∈ Rn be a piecewise constant vector, and θe ∈ Rn
be an estimator that satisfies the error bound kθe − θ0 k2n = OP (Rn ). Assume that nRn /Hn2 = o(Wn ),
where, recall, Hn is the minimum gap between adjacent levels of θ0 , defined in (4), and Wn is the
minimum distance between adjacent changepoints of θ0 , defined in (3). Then
nRn
e
,
d S(θ) | S0 = OP
Hn2
e is the set of changepoints in θ,
e S0 = S(θ0 ) is the set of changepoints in θ0 , and d( · | · )
where S(θ)
is the one-sided screening distance, as defined in (7).
Proof. The proof is derived from the `2 rate. Fix any > 0, C > 0. By assumption, we know that
there is an integer N1 > 0 such that
C
2
e
P kθ − θ0 kn > Rn ≤ ,
4
for all n ≥ N1 . We also know that there is an integer N2 > 0 such that CnRn /Hn2 ≤ Wn for all n ≥
e | S0 ) > rn .
N2 . Let N = max{N1 , N2 }, take n ≥ N , and let rn = bCnRn /Hn2 c. Suppose that d(S(θ)
e
Then, by definition, there is a changepoint ti ∈ S0 such that no changepoints of θ are within rn of
ti , which means that θej is constant over j ∈ {ti − rn + 1, . . . , ti + rn }. Denote
z = θeti −rn +1 = . . . = θeti = θeti +1 = . . . = θeti +rn .
We then form the lower bound
t +r
iXn
2
2 rn
2
1
rn
rn Hn2
C
θej − θ0,j =
z − θ0,ti +
z − θ0,ti +1 ≥
> Rn ,
n j=t −r +1
n
n
2n
4
i
(12)
n
where the first inequality holds because (x − a)2 + (x − b)2 ≥ (a − b)2 /2 for all x (the quadratic in
x here is minimized at x = (a + b)/2), and the second because rn = bCnRn /Hn2 c. Therefore, we see
e | S0 ) > rn implies
that d(S(θ)
t +r
kθe − θ0 k2n ≥
iXn
2
1
C
θej − θ0,j > Rn ,
n j=t −r +1
4
i
n
which implies
e | S0 > rn ≤ P kθe − θ0 k2 > C Rn ≤ ,
P d S(θ)
n
4
for all n ≥ N , completing the proof.
11
Remark 11 (Conditions on Wn , Hn ). The condition that nRn /Hn2 = o(Wn ) in Theorem 8 is not
strong. Consider the simple case in which Hn = Ω(1) and Wn = Θ(n). This condition reduces to
Rn = o(1), requiring only that the estimator θe in question be consistent. The theorem then gives
e | S0 ) = OP (nRn ) on the screening distance obtained by θ.
e
the bound d(S(θ)
Remark 12 (Generic setting: no particular assumptions on data model, or estimator).
Importantly, Theorem 8 assumes no data model whatsoever, and treats θe as a generic estimator of
θ0 . Of course, through the statement kθe − θ0 k2n = OP (Rn ), one sees that θe is random, constructed
from data that depends on θ0 , but no specific data model is required, nor are any specific properties
of θe (other than its `2 error rate). This flexibility allows for the result to be applied in any problem
setting in which one has control of the estimation error of a piecewise constant parameter θ0 . Apart
from the applications of Theorem 8 to the fused lasso estimator, where we consider data from a
standard model as in (1), with θ0 being the mean, and i.i.d. sub-Gaussian errors (see Corollaries 9
and 10 below), we could instead suppose that data is distributed according to, e.g., a Poisson model
with natural parameter θ0 ,
yi ∼ Pois(eθ0,i ), i = 1, . . . , n.
(13)
If we knew of an estimate θe for θ0 such that the estimation error kθe − θ0 k2n was analyzable, then we
e In
could use Theorem 8 to infer a bound on the screening distance between changepoints of θ0 and θ.
this paper, we do not describe particular applications of Theorem 8 beyond the sub-Gaussian error
model in (1), since we are not aware of estimation error guarantees outside of this model. However,
establishing `2 estimation error rates for models like (13) (which may be used to describe say copy
number data in genetics), and interpreting the resulting changepoint approximation guarantees
would be an interesting topic for future work. (For model (13), and other likelihood-based models
with a piecewise constant parameter θ0 , we suspect that the fused lasso provides a basis for a good
estimator: simply replace the squared error loss in (2) by the negative log likelihood.)
We present two different corollaries of Theorem 8 for the fused lasso. The first is given by using
Theorem 1 and the associated rate Rn = n−2/3 Cn in the weak sparsity case, and the second is given
by using Theorem 4 and the associated rate Rn = (log n log log n)/n in the strong sparsity case. The
proofs are immediate and are hence omitted.
Corollary 9 (Fused lasso screening result, weak sparsity setting). Assume the conditions
1/3 √
of Theorem 1, so that TV(θ0 ) ≤ Cn for a sequence Cn . Also assume that Hn = ω(n1/6 Cn / Wn ).
−1/3
Let θb be the fused lasso estimate in (2), with λ = Θ(n1/3 Cn ). Then
1/3 2/3
n Cn
b
d S | S0 = OP
.
Hn2
Corollary 10 (Fused lasso screening result, strong sparsity setting). Assume
the conditions
p
(log
n
log log n)/n).
of Theorem 2, so that s0 = O(1) and Wn = Θ(n). Also
assume
that
H
=
ω(
n
√
Let θb be the fused lasso estimate in (2), with λ = Θ( n). Then
log n log log n
d Sb | S0 = OP
.
Hn2
2
Remark 13 (Conditions on Wn√, Hn ). We have rewritten the condition
p that nRn /Hn = o(Wn ) in
1/6 1/3
Theorem 8 as Hn = ω(n Cn / Wn ) in Corollary 9, and Hn = ω( (log n log log n)/n) in Corollary 10 (note that in the latter, we are assuming that Wn = Θ(n)).
Remark 14 (Screening under weak sparsity). Corollary 9 handles a difficult setting in which
the number of changepoints s0 in θ0 can grow quickly with n, and yet it still provides a reasonable
b | S0 ) provided that Cn is not too large (i.e., TV(θ0 ) is not
bound on the screening distance d(S(θ)
growing too quickly), or Hn is large enough (i.e., the minimum signal gap in θ0 is large enough). As
12
an example, suppose that s0 = Θ(n1/6 ), and the changepoints in θ0 are evenly spread out, so that
1/3
Wn = Θ(n5/6 ). Then Corollary 9 implies, provided that Hn = ω(n−1/4 Cn ),
1/3 2/3
n Cn
= oP (n5/6 ),
d Sb | S0 = OP
Hn2
so for each true changepoint, there is at least one estimated changepoint that is much closer to it
than all of the other true changepoints (each of which is at least a distance Wn = Θ(n5/6 ) away).
1/3
From the condition Hn = ω(n−1/4 Cn ), and the fact that we must always have Cn ≥ s0 Hn (recall
1/6
s0 = Θ(n )), we can be more explicit here about the allowable ranges for Hn , Cn : combining the
last two relationships gives Cn = ω(n−1/8 ), and then Hn = ω(n−7/24 ). Hence, the minimum signal
gap requirement here is very reasonable, allowing
Hn to shrink to 0, just not too quickly (this is far
√
from a trivial regime, e.g., with Hn = ω( log n), when simple thresholding of pairwise differences
achieves perfect recovery, as shown in Sharpnack et al. (2012)).
Remark 15 (Comparison to Theorem 3). Corollary 10 provides a similar conclusion to that in
Harchaoui & Lévy-Leduc (2010), restated in Theorem 3: in a strong sparsity setting, the fused lasso
has a well-controlled screening distance, only slightly larger than log n/Hn2 . However, we note that
Corollary 10 guarantees this screening bound under a natural choice for the tuning parameter λ,
known to provide good `2 estimation performance (see Theorem 4), whereas Theorem 3 implicitly
requires λ to be very small, which seems unnatural (see Remark 5).
p
Remark 16 (Changepoint detection limit). The restriction
√ that Hn = ω( (log n log log n)/n)
in Corollary 10 is very close to the optimal limit of Hn = ω(1/ n) for changepoint detection: Duembgen & Walther (2008) showed that in Gaussian changepoint model with a single elevated region,
and Wn = √
Θ(n), there is no test for detecting a changepoint that has asymptotic power 1 unless
Hn = ω(1/ n). See also Chan & Walther (2013).
4.2
Post-processing for approximate changepoint recovery
e in such a way that aims
We study a procedure for post-processing the estimated changepoints in θ,
to eliminate changepoints of θe that lie far away from changepoints of θ0 . Our procedure is based on
convolving θe with a filter that resembles the mother Haar wavelet. Consider
e =
Fi (θ)
i+b
1 Xn e
1
θj −
bn j=i+1
bn
i
X
θej ,
for i = bn , . . . , n − bn ,
(14)
j=i−bn +1
for an integral bandwidth bn > 0. Our result in this subsection asserts that, by evaluating the filter
e at all locations i = bn , . . . , n − bn , and retaining only locations at which the filter value is large
Fi (θ)
(in magnitude), we can approximately recovery the changepoints of θ0 , in the Hausdorff metric.
Theorem 11 (Generic recovery result). Let θ0 ∈ Rn be a piecewise constant vector, and θe ∈ Rn
be an estimator that satisfies the error bound kθe − θ0 k2n = OP (Rn ). Consider the following procedure:
we evaluate the filter in (14) with bandwidth bn at all locations i = bn , . . . , n − bn , and we keep only
the locations whose filter value is greater than or equal to a threshold level τn , in magnitude. Denote
the resulting “filtered” set by
n
o
e = i ∈ {bn , . . . , n − bn } : |Fi (θ)|
e ≥ τn .
SF (θ)
(15)
If the bandwidth and threshold values satisfy bn = ω(nRn /Hn2 ), 2bn ≤ Wn , and τn /Hn → ρ ∈ (0, 1)
as n → ∞, then we have
e S0 ≤ bn → 1 as n → ∞,
P dH SF (θ),
where dH ( ·, · ) is the Hausdorff distance, as defined in (7).
13
Proof. The proof is not complicated conceptually, but requires some careful bookkeeping. Also, we
make use of a few key lemmas whose details will be given later. Fix > 0 and C > 0. Let N1 > 0
be an integer such that for all n ≥ N1 ,
P kθe − θ0 k2n > CRn ≤ .
2
Set = min{ρ, 1 − ρ}/2. As bn = ω(nRn /Hn2 ), there is an integer N2 > 0 such that for all n ≥ N2 ,
2CnRn
≤ (0.99Hn )2 .
bn
As τn /Hn → ρ ∈ (0, 1), there is an integer N3 > 0 such that for all n ≥ N3 ,
(ρ − )Hn ≤ τn ≤ (ρ + )Hn .
Set N = max{N1 , N2 , N3 }, and take n ≥ N . Note that ≤ ρ − and ρ + ≤ 1 − by construction,
and thus by the last two displays,
r
r
2CnRn
2CnRn
< τn < Hn −
.
(16)
bn
bn
Now observe
e S0 > bn ≤ P d SF (θ)
e | S0 > bn + P d S0 | SF (θ)
e > bn .
P dH SF (θ),
(17)
We focus on bounding each term on the right-hand side above separately. For the first term on the
e ≥ τn for all ti ∈ S0 , then d(SF (θ)
e | S0 ) ≤ bn . By the
right-hand side in (17), observe that if Fti (θ)
contrapositive,
e | S0 > bn ≤ P |Ft (θ)|
e < τn for some ti ∈ S0
P d SF (θ)
i
r
2CnRn
e
for some ti ∈ S0 ,
(18)
≤ P |Fti (θ)| < Hn −
bn
where in the second line we used the upper bound on τn in (16). Suppose that kθe − θ0 k2n ≤ CRn ;
e can be made with this error bound in place.
then, for ti ∈ S0 , Lemma 13 tells us how small |Fti (θ)|
Specifically, define
a = (−1/bn , . . . , −1/bn , 1/bn , . . . , 1/bn )
{z
} |
{z
}
|
bn times
and also r =
and c = (θ0,ti −bn +1 , . . . , θ0,ti +bn ),
bn times
√
CnRn . Then Lemma 13 implies the following: if kθe − θ0 k2n ≤ CRn , then
r
r
2CnRn
2CnRn
>
e
|Fti (θ)| ≥ |a c| − rkak2 ≥ |θ0,ti +1 − θ0,ti | −
≥ Hn −
.
bn
bn
Therefore, continuing on from (18),
r
2CnRn
e
e
P d SF (θ) | S0 > bn ≤ P |Fti (θ)| < Hn −
for some ti ∈ S0
bn
≤ P kθe − θ0 k2n > CRn
≤ .
2
14
It suffices to consider the second term in (17), and show that this is also bounded by /2. Note
that
e
e
P d S0 | SF (θ) > bn ≤ P |Fi (θ)| ≥ τn at some i such that θ0,i−bn +1 = . . . = θ0,i+bn
r
e > 2CnRn at some i such that θ0,i−b +1 = . . . = θ0,i+b .
≤ P |Fi (θ)|
n
n
bn
(19)
In the second inequality we used the lower bound on τn in (16). Similar to the previous argument,
suppose that kθe − θ0 k2n ≤ CRn ; for any location i in consideration in (19), Lemma 12 tells us how
e can be made with this error bound in place. Defining
large |Fi (θ)|
a = (−1/bn , . . . , −1/bn , 1/bn , . . . , 1/bn )
{z
} |
{z
}
|
bn times
and r =
√
and c = (θ0,i−bn +1 , . . . , θ0,i+bn ),
bn times
CnRn , as before, the lemma says the following: if kθe − θ0 k2n ≤ CRn , then
r
2CnRn
>
e
|Fi (θ)| ≤ |a c| + rkak2 =
.
bn
Hence, continuing on from (19),
r
2CnRn
e
e
P d S0 | SF (θ) > bn ≤ P |Fi (θ)| >
at some i such that θ0,i−bn +1 = . . . = θ0,i+bn
bn
≤ P kθe − θ0 k2n > CRn
≤ ,
2
completing the proof.
Remark 17 (Comparison to Theorem 8). Though they are stated differently, the rates in Theorems 8 and 11 for approximate changepoint screening and recovery, respectively, are comparable.
To see this, note that the conclusion in the latter implies
e | S0 ≤ dn → 1 as n → ∞,
P d S(θ)
for any sequence dn = ω(nRn /Hn2 ), which is in line with Theorem 11. (The original conclusion that
e | S0 ) = OP (nRn /H 2 ) is a somewhat stronger statement, though the difference is not major.)
d(S(θ)
n
Remark 18 (Generic setting: no particular assumptions on data model, or estimator).
To emphasize a similar point to that in Remark 12, Theorem 11 does not use a specific data model,
and considers any estimator θe for which we have `2 error control, kθe − θ0 k2n = OP (Rn ). This makes
it a very flexible and broadly applicable result. When the data comes from a model as in (1), where
θ0 is the mean and we have i.i.d. sub-Gaussian errors, we can apply Theorem 11 to the fused lasso,
given our knowledge of its `2 error rate (see Corollaries 14 and 15 below). It could also be applied,
under the same data model, to many other estimators whose `2 error rates are known (such as the
Potts estimator, and unbalanced Haar wavelets). Moreover, it could be useful under different data
models, like the Poisson model in (13), as it would provide approximate recovery guarantees for any
method with a fast enough `2 estimation error rate. (Note that the post-processing step using the
filter (14) does not itself require assumptions about the data.) We do not consider such extensions
in the current paper, but they suggest interesting directions for future work.
15
The proof of Theorem 11 relied on two lemmas, that we state below. Their proofs are based on
simple arguments in convex analysis and deferred until Appendix C.
Lemma 12. Given a, c ∈ Rm , r ≥ 0, the optimal value of the (nonconvex) optimization problem
max |a> x| subject to kx − ck2 ≤ r
(20)
x∈Rm
is |a> c| + rkak2 .
Lemma 13. Given a, c ∈ Rm , r ≥ 0 such that |a> c| − rkak2 ≥ 0, the optimal value of the (convex)
optimization problem
minn |a> x| subject to kx − ck2 ≤ r
(21)
x∈R
>
is |a c| − rkak2 .
We finish this subsection with two corollaries of Theorem 11 for the fused lasso estimator, in the
weak and strong sparsity cases. The proofs are immediate and are thus omitted.
Corollary 14 (Fused lasso recovery result, weak sparsity setting). Assume the conditions
of Theorem 1, so that TV(θ0 ) ≤ Cn for a sequence Cn . Let θb be the fused lasso estimate in (2), with
−1/3
b as described in Theorem 11, to proλ = Θ(n1/3 Cn ), and consider applying the filter in (14) to θ,
b If the bandwidth and threshold satisfy bn = bn1/3 Cn2/3 ν 2 /H 2 c ≤ Wn /2
duce a filtered set SbF = SF (θ).
n
n
for a sequence νn → ∞, and τn /Hn → ρ ∈ (0, 1), then
2/3
n1/3 Cn νn2
b
P dH (SF , S0 ) ≤
→1
Hn2
as n → ∞.
Corollary 15 (Fused lasso recovery result, strong sparsity setting). Assume the conditions
b
of Theorem
√ 2, so that s0 = O(1) and Wn = Θ(n). Let θb denote the fused lasso estimate in (2), with
λ = Θ( n), and consider applying the filter in (14) to θ,
as in Theorem 11, to produce a filtered set
b If the bandwidth and threshold values satisfy bn = b(log n log log n)ν 2 /H 2 c ≤ Wn /2 for
SbF = SF (θ).
n
n
a sequence νn → ∞, and τn /Hn → ρ ∈ (0, 1), then
(log n log log n)νn2
P dH (SbF , S0 ) ≤
→ 1 as n → ∞.
Hn2
Remark 19 (Recovery under weak sparsity, comparison to BS). Corollary 14 considers a
challenging setting in which the number of changepoints s0 in θ0 could be growing quickly with n,
and the only control that we have is TV(θ0 ) ≤ Cn . We draw a comparison here to known results on
binary segmentation (BS). Corollary 14 on the (filtered) fused lasso and Theorem 3.1 in Fryzlewicz
(2014) on the BS estimator θbBS , each under appropriate conditions on Wn , Hn , state that
2/3
n1/3 Cn log n
dH (SbF , S0 ) ≤
Hn2
versus
n log n
dH S(θbBS ), S0 ≤ c
Hn2
respectively,
(22)
where c > 0 is a constant, and
probability approaching 1. The result on SbF
√both bounds hold with1/3
2/3
is obtained by choosing νn = log n and then bn = bn Cn log n/Hn2 c in Corollary 14. Examining
(22), we see that, when Cn scales more slowly than n, Theorem 11 provides the stronger result: the
2/3
term n1/3 Cn will be smaller than n, and thus the bound on dH (SbF , S0 ) will be sharper than that
BS
b
on dH (S(θ ), S0 ).
But we must also examine the specific restrictions that each result in (22) places on s0 , Wn , Hn .
Consider the simplification Wn = Θ(n/s0 ), corresponding to a case in which the changepoints in θ0
2/3
are spaced evenly apart. For Corollary 14, starting with the condition n1/3 Cn log n/Hn2 ≤ Wn /2,
plugging in the relationship Cn ≥ s0 Hn , and rearranging to derive a lower bound on the minimum
16
5/4
signal gap, gives Hn = Ω(s0 n−1/2 log3/4 n). If s0 = Θ(n2/5 ), then we see that the minimum signal
gap requirement becomes Hn = Ω(log3/4 n), which is growing with n and is thus too stringent to be
interesting (recall, as discussed previously, that Sharpnack et al. √
(2012) showed simple thresholding
of pairwise differences achieves perfect recovery when Hn = ω( log n)). Hence, to accommodate
signals for which Hn remains constant or even shrinks with n, we must restrict the number of jumps
in θ0 according to s0 = O(n2/5−δ ), for any fixed δ > 0. Meanwhile, inspection of Assumption 3.2 in
Fryzlewicz (2014) reveals that his Theorem 3.1 requires s0 = O(n1/4−δ ), for any δ > 0, in order to
handle signals such that Hn remains constant or shrinks with n. In short, the (effectively) allowable
range for s0 is larger for Theorem 11 than for Theorem 3.1 in Fryzlewicz (2014). Even when we look
within their common range, Theorem 11 places weaker conditions on Hn . As an example, consider
s0 = Θ(n1/6 ) and Wn = Θ(n5/6 ). The fused lasso result in (22) requires Hn = Ω(n−7/24 log4/3 n),
and the BS result in (22) requires Hn = Ω(n−1/6+δ ), for any δ > 0. Finally, to reiterate, the fused
lasso result in (22) gives a better Hausdorff recovery bound when Cn is small compared to n; at the
extreme end, this is better by a full factor of n2/3 , when Cn = O(1).
While the post-processed fused lasso looks favorable compared to BS, based on its approximate
changepoint recovery properties in the weak sparsity setting, we must be clear that the analyses
for other methods—wild binary segmentation (WBS), the simultaneous multiscale changepoint estimator (SMUCE), and tail-greedy unbiased Haar (TGUH) wavelets—are still much stronger in this
setting. Such methods have Hausdorff recovery bounds that are only possible for the post-processed
fused lasso (at least, using our current analysis technique) when we assume strong sparsity. This is
discussed next.
Remark 20 (Recovery under strong sparsity, comparison to other methods). When s =
O(1) and Wn = Θ(n), Corollary 10 shows that the post-processed fused lasso estimator delivers a
Hausdorff bound of
log2 n
,
(23)
dH (SbF , S0 ) ≤
Hn2
on the set Sbp
F of filtered changepoints, with probability approaching 1. This is obtained by choosing
(say) νn = log n/ log log n and bn = blog2 n/Hn2 c√≤ Wn /2 in the corollary. The effective restriction
√
on the minimum signal gap is thus Hn = Ω(log n/ n), which is quite reasonable, as Hn = ω(1/ n)
is needed for any method to detect a changepoint with probability tending to 1 (recall Remark 16).
Several other methods—the Potts estimator (Boysen et al. 2009), binary segmentation (BS) and wild
binary segmentation (WBS) (Fryzlewicz 2014), the simultaneous multiscale changepoint estimator
(SMUCE) (Frick et al. 2014), and tail-greedy unbiased Haar wavelets (TGUH) (Fryzlewicz 2016)—
all admit Hausdorff recovery bounds that essentially match (23), under similarly weak restrictions on
Hn . But, it should be noted that the latter three methods—WBS, SMUCE, and TGUH—continue
to enjoy these same sharp Hausdorff bounds outside of the strong sparsity setting, namely, their
analyses do not require that s0 = O(1) and Wn = Θ(n), and instead just place weak restrictions on
the allowed combinations of Wn , Hn (e.g., the analysis of WBS in Fryzlewicz (2014) only requires
Wn Hn2 ≥ log n). These analyses (and those for all previously described estimators) are more refined
than that given in Corollary 15: they are based on specific properties of the estimator in question.
The corollary, on the other hand, follows from Theorem 8, which uses a completely generic analysis
that only assumes knowledge of the `2 error rate.
4.3
Post-processing on a reduced set
Recall that the strategy studied in Theorem 11 was to apply the Haar filter in (14) at each location
i = bn , . . . , n − bn and then check for large absolute values. Computationally, this not expensive—it
only requires O(n) operations—but there is an undesirable feature of this strategy with respect to
e it is possible
practical usage. Writing the original number of estimated changepoints as se = |S(θ)|,
e
in practice for the size of the filtered set SF (θ) in (15) to be much larger than se, if the bandwidth
17
and threshold parameters are not set appropriately. Indeed, as the filter is being applied at n − 2bn
locations, it is possible for the filtered set to have precisely this many elements.
Here we propose a modified strategy that runs the filter on (at most) 3e
s + 2 changepoints, and
then as usual, keeps only changepoints whose absolute filter values are large. This modified strategy
has essentially same the theoretical guarantee of approximate changepoint recovery as the original
“exhaustive” strategy from Section 4.2, but enjoys the practical advantage that, no matter how the
bandwidth and threshold parameters are chosen, the final set of detected changepoints is bounded
in size by 3 times the number of changepoints in θe (plus 2, to be precise). Before stating the main
result of this subsection, we introduce a “candidate” set for locations for changepoints,
n
o
e = i ∈ {bn , . . . , n − bn } : i ∈ S(θ),
e or i + bn ∈ S(θ),
e or i − bn ∈ S(θ)
e ∪ {bn , n − bn }. (24)
IC (θ)
These are estimated changepoints, locations that are at a distance bn from estimated changepoints,
or boundary points.
Theorem 16 (Generic recovery result, reduced post-processing). Assume the conditions of
Theorem 11, but consider a modified strategy in which we only evaluate the filter in (14) at locations
e in (24), and define a “reduced” set of filtered points based on the locations
in the candidate set IC (θ)
whose filter value is at least τn ,
n
o
e = i ∈ IR (θ)
e : |Fi (θ)|
e ≥ τn .
SR (θ)
(25)
Then, subject to the same conditions on bn , τn as in Theorem 11, we have
e S0 ≤ 2bn → 1 as n → ∞.
P dH SR (θ),
Proof. We will show that
n
o n
o
e S0 ≤ bn ⊆ dH SR (θ),
e S0 ≤ 2bn ,
dH SF (θ),
(26)
Since the left-hand side occurs with probability tending to 1, by Theorem 11, so will the right-hand
side. To show the desired containment, recall that, by the definition of Hausdorff distance,
n
o n
o n
o
e S0 ≤ bn = d S0 | SF (θ)
e ≤ bn ∩ d SF (θ)
e | S0 ≤ bn .
dH SF (θ),
(27)
Inspecting the first term on the right-hand side of (27), we observe
n
o n
o n
o
e ≤ bn ⊆ d S0 | SF (θ)
e ≤ 2bn ⊆ d S0 | SR (θ)
e ≤ 2bn ,
d S0 | SF (θ)
(28)
e ⊆ SF (θ).
e Inspecting the second term on the right-hand
where the last containment holds as SR (θ)
e
side of (27), we use Lemma 17 which states that for each j ∈ {bn , . . . , n − bn }, there exists i ∈ IC (θ)
e ≥ |Fj (θ)|.
e Using this, we see
such that |i − j| ≤ bn and |Fi (θ)|
n
o n
o
e | S0 ≤ bn = for all ` ∈ S0 , there exists j ∈ SF (θ)
e such that |` − j| ≤ bn
d SF (θ)
n
o
e such that |` − i| ≤ 2bn
⊆ for all ` ∈ S0 , there exists i ∈ IC (θ)
n
o
e | S0 ≤ 2bn .
= d SR (θ)
(29)
Above, we have used Lemma 17 for the containment in the second line. Combining (27), (28), and
(29), we have established (26), as desired.
The proof of Theorem 16 relied on the following lemma. Its proof can be found in Appendix D.
e be the candidate set defined in (24). For every location j ∈ {bn , . . . , n − bn }
Lemma 17. Let IC (θ)
e > 0, there exists a location i ∈ IC (θ)
e such that |i − j| ≤ bn and |Fi (θ)|
e ≥ |Fj (θ)|.
e
where |Fj (θ)|
18
5
Implementation considerations and experiments
We develop a data-driven procedure to determine the threshold level τn of the filter in (14), used to
e from an estimate θ,
e as described in (15) in Theoderive a post-processed set of changepoints SF (θ)
rem 11. We also present a number of simulation results to support and complement the theoretical
developments in this paper.
A data-driven procedure for choosing τn . Let A(·) denote a fitting algorithm that, applied
to data y, outputs an estimate θe of θ0 (e.g., A(y) could be the minimizer in (2), so that its output is
the fused lasso estimate). In Algorithm 1, we present a heuristic but intuitive method for choosing
e Aside from
the threshold level τn , based on (entrywise) permutations of the residual vector y − θ.
the choice of fitting algorithm A(·), we must specify a number of permutations B to be explored, a
bandwidth bn for the filter in (14), and a quantile level q ∈ (0, 1). The intuition behind Algorithm 1
is to set τn large enough to suppress “false positive” changepoints 100 · q% of the time (according
to the permutations). This is revisited later, in the discussion of the simulation results.
Some example settings: we may choose A(·) to be the fused lasso estimator, where the tuning
parameter λ is selected to minimize 5-fold cross-validation (CV) error, B = 100, and q = 0.95. The
choice of bandwidth bn is more subtle, and unfortunately, there is no specific answer that works for
all problems.4 But, the theory in the last section provides some general guidance: e.g., for problems
in which we believe there are a small number of changepoints (i.e., s0 = O(1)) of reasonably large
magnitude (i.e., Hn = Ω(1)), Theorem 11 instructs us to choose a bandwidth that grows faster than
log n log log n, so, choosing bn to scale as log2 n would suffice. We will use this scaling, as well as the
above suggestions for A(·), B, and q in all coming experiments, unless otherwise specified.
Algorithm 1 Permutation-based approach for choosing τn
0. Input a fitting algorithm A(·), number of permutations B, bandwidth bn , and quantile level
q ∈ (0, 1).
e denote the changepoints, and r = y − θe the residuals.
1. Compute θe = A(y). Let Se = S(θ)
2. For each b = 1, . . . , B, repeat the following steps:
(a) Let r(b) be a randomly-chosen permutation of r, and define auxiliary data y (b) = θe + r(b) .
(b) Rerun the fitting algorithm on the auxiliary data to yield θe(b) = A(y (b) ).
(c) Apply the filter in (14) to θe(b) (with the specified bandwidth bn ), and record the largest
e Formally,
magnitude τb(b) of the filter values at locations greater than bn away from S.
τb(b) =
max
i∈{bn ,...,n−bn }:
e
d(S|{i})>b
n
Fi (θe(b) ) .
3. Output τbn , the level q quantile of the collection τb(b) , b = 1, . . . , B.
e or the
After running Algorithm 1 to compute τbn , the idea is to proceed with the full filter SF (θ)
e
e
reduced filter SR (θ), applied at the level τn = τbn , to the estimate θ computed on the original data y
at hand. In the experiments that follow, we use the reduced filter, though similar conclusions would
hold with the full filter.
4 We note that in some situations, problem-specific intuition can yield a reasonable choice of bandwidth b . Also,
n
it should be possible to extend Algorithm 1 to choose both τn and bn , but we do not pursue this, for simplicity.
19
0
200
400
600
10
0
5
●
●
●●
●● ●●
●●●
● ●
● ●●●●●
●●
●
●
●● ●
●
● ●●● ●●
●
●
●
●
●
●● ●
●
●
●●
●●
●●
●●
●
●
●●●
●●●
●
●●
● ●●●●
●
●
●●
●●
●
●
● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●● ●
●● ● ●
●●
●
●
●
●●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
● ●
●
●●
●
●
●
●
●
●
● ●●●
●
●
●
●
●●●●●
●●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●● ●
●
●
●●
●●
●
●
●●
●
●
●●
●●
●
●● ●
●
●
●
●
●
●
●
●
●
●●●●
●
●●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●●●●●
●
●
●
●
●● ●
●
●
●
●●
●
●
●
●
●
●●
●
●
●● ●●
●
●●●●
●
●
●●
●●
●
●●
●
●
●●
●●
●
●●
●●
●
●
●
●
●
●
●
●●●
●●●●
●
●●
●
●
●
●
●●
●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●
● ●●●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
● ●●●●●
●
●●
●
●●
●
●
●●
●
●
●●
●
● ● ●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●● ●●
●
●
●
●
●
●
●
●●
●
●●●
● ●● ●●
●●
●
●●
●●●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●● ●
● ●● ●
●
●
●
●
●
● ●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●● ●
●
●
●●
●● ●
●
●
●●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●●●
●● ●
● ●●
●
−5
−5
0
5
10
Simulation setup. In our experiments, we use the following simulation setup. For a given n, the
mean parameter θ0 ∈ Rn is defined to have s0 = 5 equally-sized segments, with levels 0, 2, 4, 1, 4,
from left to right. Data y ∈ Rn is generated around θ0 using i.i.d. N (0, 4) noise. Lastly, the sample
size n is varied between 100 and 10, 000, equally-spaced on a log scale. Figure 2 shows example data
sets with n = 774 and n = 10, 000.
800
●●● ●
●
● ●●
●●
● ●
●●●
●
●●
●
●
●
●●●●
●
●●
●●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●● ●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●●
●
●
● ●
●●●
●
●●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●●●●
●
●●
●●
●
●
●
●●● ●
●
0
2000
4000
6000
8000
10000
Figure 2: An example from our simulation setup for n = 774 (left) and n = 10, 000 (right), where in each
panel, the mean θ0 is plotted in red, and the data points in gray.
Evaluation of the filter. We demonstrate that the filter in (14), with bn = b0.25 log2 nc, can be
effective at reducing the Hausdorff distance between estimated and true changepoint sets. We first
illustrate the use of the filter in a single data example with n = 774, in Figure 3. As we can see, the
fused lasso originally places a spurious jump around location 250, but this jump is eliminated when
we apply the filter, provided that we set the threshold to be (say) τn = 0.5.
Figure 4 now reports the results from applying the filter in problems of sizes between n = 100
and n = 10, 000, using 50 trials for each n. We consider three different sets of changepoint estimates:
b the original changepoints from fused lasso estimate θb tuned with 5-fold CV tuning; SR (θ),
b
Sb = S(θ),
b with τn chosen by
the changepoints after applying the reduced filter as described in Theorem 16 to θ,
b an oracle set of changepoints given by trying a wide range of τn values and
Algorithm 1; and SO (θ),
choosing the value that minimizes the Hausdorff distance after filtering (this assumes knowledge of
S0 , and is infeasible in practice). These are labeled as “original”, “data-driven”, and “oracle” in the
figure, respectively. As we can see from the left and middle panels, the Hausdorff distance achieved
by the original changepoint set grows nearly linearly with n, but after applying the reduced filter,
the Hausdorff distance becomes very small, provided that n is larger than 1000 or so. Empirically,
the Hausdorff distance associated with the filtered set appears to grow very slowly with n, nearly
constant (slower than the the log n log log n rate guaranteed by Corollary 15). The right panel shows
that our data-driven choices of τn are not substantially different from those made by the oracle.
Screening distances, false positives. Figure 5 examines the outcomes from varying the filter
b The results
threshold τn in between 0 and 2, and then applying the reduced filter to produce SR (θ).
are aggregated over 500 trials when n = 774 (i.e., 500 data instances drawn from the simulation
b | S0 ) and “precision distance” d(S0 | SR (θ))
b are plotted
setup), and the screening distance d(SR (θ)
with τn . The former increases with τn , and the latter decreases; recall, the Hausdorff distance is the
maximum of the two. We see that threshold levels from 0.5 to 1 yield a small Hausdorff distance.
The left panel of Figure 6 shows the same results, but with the screening distance on the x-axis,
and the precision distance on the y-axis. The red dot marks the screening distance and precision
distance achieved by the data-driven rule from Algorithm 1, using B = 150 permutations. This lies
basically at the “elbow” of the curve, just as we would desire. The middle panel of the figure plots
the proportion of false positive detections (out of the 500 repetitions total) on the x-axis, versus the
20
10
●
●
●●
●●
●
●
●
●
● ● ● ●●●
●
●● ●●● ● ●●●
●
●
●
●
●
●●● ●●●
●
●●
●●● ●●
●●● ●●●● ●
● ●● ●●
●●
●
●● ● ●
● ●●
●●
● ●
● ●●
●
●
●
●●
●
●●●● ●●●
●●●
●
●
●
●●
●
●
●●●
●
●●
● ● ●●
● ●● ●● ●
●●
●●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●● ●●●
●
●
●
●
●● ●
●●
●
●
●
●
● ●● ●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●
●● ● ●
●●
●
●●●●●●●
●● ●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●● ●●
●●
●
●
●
● ●
●
●
●
●●
●
● ●●
●●●
●
●●
●
●
●●●●
●
●
●●●
●
●
●●
●●
●
●● ●
●
●● ●
●●
● ●
●
●●●
●
●●●
●
●
●
● ●● ●● ● ●●●
●●
●●
●●● ● ●●●●●●●●●
●
● ● ●●
●
● ●
●
●● ●
●
●
●●
●●
●● ●
●● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
● ●●● ● ●
●
●●
●
●
●
●
●
●
● ● ●●●
● ●
● ● ●● ●●
●●
● ●●
● ●
●●●● ●
●
●●
●●
●
●
●● ●●● ●●●●
●●
●●
●●
●●
●
●
●
●●
●● ●
●
●●●●●● ●
●
●
● ●●●
●●
●●●●●
●
●
●
●● ●
●
●●●● ●●
●
●
● ●
●●
●
●●
●
●●
●●●●
●
● ●●
●
●●
● ●●●
●●
● ●● ●●
●●
●
●● ● ●● ●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●●
●● ●●●●
●
●●
●
●●
●●
●●
● ●●
●●
●●
●● ●
●● ●●
● ●●
●●● ● ●●●
●● ●
●●
●
●●
●●
●● ●●●
●
● ●● ●●
● ●
●
●
●●●●
●
●●
●●●
●
●● ● ●● ●
●
●
●●
●
●
●
●
●
●
●
● ●●●
● ●●
●
●
●
●
●●
●
●●
●●
●
●●●● ●●● ● ●
●●
●
●● ●
●
●
●
● ●● ●
●
●
●
●
−5
0
Data values
5
●
200
400
2.0
0
600
800
Index
●
●
●●
1.5
Absolute filter values
●
●●
●●
1.0
●
●
●
●●
●
●●
● ●
●
● ●
●
0.5
●
●
●
●
●
●
0.0
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●● ●
●
● ●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
●
●
● ●
●
● ●
●
●
●
●
● ●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
200
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
400
●
●
●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
600
800
Figure 3: In the top plot, an example with n = 774 is shown from our simulation setup, where the data y is
0
●
●
2000 4000 6000 8000
1.6
1.4
1.2
1.0
Threshold
0
Data−driven
Oracle
●
●
●
●
●●
●
●●
●
● ●
●
●
●
●
●
●
●
0.6
●
●
●
●
●
●
●●● ●
●
0.8
300
200
Hausdorff distance
●
Data−driven
Oracle
●
●
2000 4000 6000 8000
n
n
●
0.4
●
●
●
●
●
●
●
●
●●● ●
●
100
●
0
1000
●
●
500
0
Hausdorff distance
1500
Original
Data−driven
Oracle
400
drawn in gray, the mean θ0 in red, and the fused lasso estimate θb in blue. In the bottom plot, the filter values
b i = 1, . . . , n are drawn in blue, and the threshold τn is drawn as a horizontal green line. Changepoints
Fi (θ),
before and after filtering are marked by short black lines along the bottom and top x-axes, respectively.
0
2000 4000 6000 8000
n
Figure 4: In the left panel, the Hausdorff distances between original changepoints, filtered changepoints with
a data-driven threshold, and filtered changepoints with an oracle threshold, are plotted (in black, blue, and
red, respectively). The results are aggregated across 50 trial runs for each sample size n; the solid dots display
the median values, and the vertical segments display the interquartile ranges (25th to 75th percentiles). The
middle panel zooms in on the Hausdorff distances for the data-driven and oracle filtering procedures, and the
right panel displays the choices of τn for these procedures.
21
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
150
0
50
100
Distance
200
250
300
Precision distance
Screening distance
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
0.5
1.0
1.5
2.0
Filter threshold
b | S0 ) (in black) and the precision distance d(S0 | SR (θ))
b (in red)
Figure 5: The screening distance d(SR (θ)
are shown as functions of the threshold τn used for the filtered set. These were aggregated over 500 trials,
in which n = 774; the dots display the median values, and the vertical segments are drawn from the 25th to
75th percentiles. We can see a substantial jump in the screening distance once a bit after τn = 1 and again
b suppressing all
after τn = 1.5, where the median value is close to one of the quartiles. This is due to SR (θ)
of the estimated changepoints near a particular true changepoint, at these critical values of τn .
proportion of true positive detections on the y-axis. Here, note, we define a false positive detection
to be the event that any estimated changepoint is more than bn away from all true changepoints, or
b > bn , and a true positive detection to be the event that all true
simply, the event that d(S0 | SR (θ))
b | S0 ) ≤ bn . Therefore,
changepoints have estimated changepoints at most bn away, or simply, d(SR (θ)
to be perfectly concrete, the x-axis and y-axis are displaying a certain type of false positive and true
positive rates (FPR and TPR), defined as
FPR =
b > bn
# trials in which d(S0 | SR (θ))
# of trials
and
TPR =
b | S0 ) ≤ bn
# trials in which d(SR (θ)
.
# of trials
The red dot again marks the FPR and TPR achieved by the data-driven rule in Algorithm 1 for
choosing the threshold, about 0.26 and 0.7, respectively. We might expect here, having set q = 0.95
in Algorithm 1, to see a FPR close to 0.05 (because the choice of threshold in Algorithm 1 precisely
controls the FPR at 0.05 over the permutations encountered in the procedure). However, this is not
the case on in our simulation, and the actual FPR is higher. This phenomenon is not specific to the
quantile choice of q = 0.95, as shown in the right panel of Figure 6. For a varying quantile level q in
between 0 and 1, we ran Algorithm 1, used the corresponding threshold for our filter, and measured
the FPR achieved by the filtered changepoint set. As we can see, the actual FPR is generally higher
than 1 − q.
Fused lasso fast `2 error rate, under strong sparsity. We finish by examining the (squared)
`2 error kθb − θ0 k2n as it scales with n, when the fused lasso estimate θb in (2) is tuned appropriately.
22
●●
●
●
●
●
●●●●●●●
●● ●●●
● ● ●
0.8
●
● ●●
●
●
●
●
●
●
●
●
●
●●
0
●
50
100
●
●
●
●
●
●
200
300
0.2
0.4
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.6
FDR
0.6
0.4
0.2
TPR
●
●
0
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●●
1.0
1.0
0.8
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
80
60
40
20
Precision distance
120
●
●●●●●●●●
●●●●●●●●
●●●●●●●
●●●●●
●●●●
●●●
●●●●
●●●●
●
●
●●
●●
●●
●●●
●●
●●●●
●●
●
●●●
●
●●
●●
●●
●
●
●●
●●
●●
●●
●●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
0.0
0.2
0.4
Screening distance
0.6
0.8
1.0
0.0
0.2
0.4
FPR
0.6
0.8
1.0
1−q
b and the screening distance d(SR (θ)
b | S0 ),
Figure 6: The left panel plots the precision distance d(S0 | SR (θ))
as the threshold τn is varied from 0 to 2. Shown are the median distances over 500 trials for a problem with
n = 774. The red dot marks the precision and screening distances achieved by the data-driven threshold level
chosen by Algorithm 1 with q = 0.95. The middle panel shows the same, but with true positive rate (TPR)
against false positive rate (FPR). The right panel shows the achieved FPR against 1−q, as the input quantile
level q is varied in Algorithm 1.
150
●
n * MSE
0.4
30
MSE
●
0
●
0.2
●
●
●
50
●
●
●
●
●
●
●
●
●
●
●
●
●
0.0
●
2000 4000 6000 8000
0
●
●
●
2000 4000 6000 8000
n
n
●
0
●
0
10
●
●
●
●
20
λ
40
0.6
●
100
50
0.8
60
For different sample sizes ranging from n = 100 to n = 10, 000, we generated 50 example data
sets from the same setup described previously, and on each data set, computed the fused lasso
estimate θb with 5-fold CV to select the tuning parameter λ. Figure 7 reports the median value of λ,
and the median achieved `2 error rate kθb − θ0 k2n , over the 50 trials, as functions of n. The results
support the theoretical conclusion in Theorem 4, as the achieved `2 error rate scales at about the
rate (log
√ n log log n)/n. Also, since s0 = O(1), the results support the theoretical result that λ scales
with n.
5
6
7
8
9
log(n)
Figure 7: The left panel shows the value of λ chosen to minimize 5-fold CV error over the fused lasso path,
aggregated over repetitions in our simulation setup, as the sample size n varies. This scales approximately as
√
n, which is drawn as a red curve (with a best-fitting constant). The middle panel shows the corresponding
(squared) `2 estimation error kθb − θ0 k2n , again aggregated over repetitions, as n varies. The scaling is about
(log n log log n)/n (red curve). The right panel plots the median achieved values of nkθb − θ0 k2n against log n;
this looks close to linear (red line), which provides empirical support to the claim that the fused lasso error
rate is indeed (log n log log n)/n instead of log2 n/n (as this would have appeared as a quadratic trend in the
right panel). In each panel, the vertical bars denote ±1 standard deviations.
23
6
Extensions
We study screening properties that are implied by `2 error properties in two related problems: first,
piecewise linear segmentation, and then, segmentation on graphs.
6.1
Piecewise linear segmentation
We now consider data from a model as in (1) but where θ0,i , i = 1, . . . , n is a piecewise linear (rather
than a piecewise constant) sequence. The main estimator of interest is linear trend filtering (Steidl
et al. 2006, Kim et al. 2009, Tibshirani 2014), which can be seen as an extension of the fused lasso
that penalizes second-order (rather than first-order) differences:
n
n−2
X
1X
(yi − θi )2 + λ
|θi − 2θi+1 + θi+2 |,
θb = argmin
2 i=1
θ∈Rn
i=1
(30)
for a tuning parameter λ ≥ 0. Several other estimators are available in the piecewise linear segmentation problem, but given its ties to the fused lasso (and our focus on the fused lasso thus far), we
focus on linear trend filtering in particular.
In terms of detection, we are now interested in the locations of nonzero second-order differences,
i.e., the “knots”, which mark the changes in slope across the entries of a parameter θ ∈ Rn :
θi−1 + θi+1
.
S2 (θ) = i ∈ {2, . . . , n − 1} : θi 6=
2
b We again write S0,2 = {t1 , . . . , ts }, where
We use the abbreviations S0,2 = S2 (θ0 ) and Sb2 = S2 (θ).
0
2 ≤ t1 < . . . < ts0 < n and s0 = |S0,2 |, and for convenience t0 = 0, ts0 +1 = n. We also carry forward
analogous definitions for Wn , Hn :
Wn =
min
i=0,1...,s0
(ti+1 − ti )
and Hn = min |θ0,i−1 − 2θ0,i + θ0,i+1 |.
i∈S0,2
(31)
Lastly, we define the discrete second-order total variation operator, acting on a vector x ∈ Rn , by
TV2 (x) =
n−1
X
|xi−1 − 2xi + xi+1 |.
i=2
The following describes the `2 estimation error of linear trend filtering, under weak sparsity.
Theorem 18 (Trend filtering error rate, weak sparsity setting, Theorem 10 of Mammen
& van de Geer 1997). Assume the data model in (1), with errors i , i = 1, . . . , n i.i.d. from a
sub-Gaussian distribution as in (8). Also assume that TV2 (θ0 ) ≤ Cn , for a sequence Cn . Then for
−3/5
λ = Θ(n1/5 Cn ), the linear trend filtering estimate θb in (30) satisfies
kθb − θ0 k2n = OP (n−4/5 Cn2/5 ).
Remark 21 (Consistency, optimality). The lemma shows that linear trend filtering is consistent
when Cn = o(n2 ). When Cn = O(1), its error rate is n−4/5 , which is in fact minimax optimal as
θ0 varies over the class of signals having bounded second-order total variation, i.e., θ0 ∈ {θ ∈ Rn :
TV2 (θ) ≤ C} for a constant C > 0 (Donoho & Johnstone 1998). As in the fused lasso case, we refer
the reader to Tibshirani (2014) for explanations of the above theorem and this minimax result, in
notation that is more consistent with that of the current paper.
24
Remark 22 (Strong sparsity, higher polynomial degrees). Results for linear trend filtering in
the strong sparsity setting, i.e., one in which s0 is assumed to be bounded (so that we are estimating
a piecewise linear function with few knots) are not currently available, to the best of our knowledge.
However, we suspect that the achieved error rate here will be close to the “parametric” 1/n rate, as
in Theorems 2 and 4, on the fused lasso. It is also worth noting that the extension of trend filtering
to fit piecewise polynomials of higher degrees (i.e., higher than 1, as in the current piecewise linear
case) is covered in Tibshirani (2014), where `2 estimation error rates (under weak sparsity) are also
derived. For simplicity, we do not consider the general piecewise polynomial setting in our study of
approximate screening, below, though such an extension should be possible.
Now we give our generic approximate screening result, analogous to that in Theorem 8.
Theorem 19 (Generic screening result, piecewise linear segmentation). Let θ0 ∈ Rn be a
piecewise linear vector, and θe ∈ Rn be an estimator satisfying the error bound kθe − θ0 k2n = OP (Rn ).
1/3 −2/3
Assume that n1/3 Rn Hn
= o(Wn ), where, recall, Wn , Hn are as defined in (31). Then
1/3 1/3
n Rn
e
.
d S2 (θ) | S0,2 = OP
2/3
Hn
Proof. The proof follows that of Theorem 8 closely, but differs in the lower bound asserted in (12).
As before, given any > 0, C > 0, as know that for some integer N1 > 0 and all n ≥ N1 ,
C3
2
e
P kθ − θ0 kn >
Rn ≤ .
4
1/3
−2/3
≤ Wn . Let
We also know that for some integer N2 > 0 and n ≥ n2 , it holds that Cn1/3 Rn Hn
1/3 −2/3
e | S0,2 ) > rn .
N = max{N1 , N2 }, take n ≥ N , and let rn = bCn1/3 Rn Hn c. Suppose that d(S2 (θ)
Then there is a knot ti ∈ S0,2 such that there are no knots in θe within rn of ti , which means that θej
displays a linear trend over the entire segment j ∈ {ti − rn , . . . , ti + rn }. Hence
t +r
1 iXn
n j=t −r
i
θej − θ0,j
2
≥
13rn3 ((θ0,ti +1 − θ0,ti ) − (θ0,ti − θ0,ti −1 ))2
13rn3 Hn2
C3
≥
>
Rn .
24n
24n
4
(32)
n
Here, the first inequality holds due to Lemma 20, the second holds by definition of Hn , and the third
e | S0,2 ) > rn implies the estimation error exceeds (C 3 /4)Rn ,
by definition of rn . We see that d(S2 (θ)
an event that we know occurs with probability at most , completing the proof.
The proof of Theorem 19 relied on the next lemma, to construct the key lower bound (32). The
lemma characterizes how well a piecewise linear function can be approximated by a linear one, and
is proved in Appendix E.
Lemma 20. Let f (x) be a piecewise linear function, defined over x = −r, . . . , r, by
(
a1 x for x ≥ 0
f (x) =
.
a2 x for x < 0
Let e
ax + eb be the optimal linear function for estimating f (x), according to squared error loss, i.e.,
(e
a, eb) = argmin
a,b∈R
Then
r
X
x=−r
r
X
2
f (x) − ax − b .
x=−r
2
13r3
f (x) − e
ax − eb ≥ (a2 − a1 )2
.
24
25
By combining Theorems 18 and 19, we have the following approximate screening result for linear
trend filtering. The proof is omitted.
Corollary 21 (Trend filtering screening result, weak sparsity setting). Assume the condi1/5
−3/2
tions in Theorem 18, thus T V2 (θ0 ) ≤ Cn for a sequence Cn . Also assume Hn = ω(n1/10 Cn Wn
).
−2/3
Let θb denote the linear trend filtering estimate in (30) with λ = Θ(n1/5 Cn ). Then
1/15 2/15
n
Cn
.
d Sb2 | S0,2 = OP
2/3
Hn
Remark 23 (Knot screening under weak sparsity). To give an example of a challenging
case
√
n)
knots,
that can be accommodated by Corollary
21,
consider
a
setting
in
which
θ
has
s
=
Θ(
0
0
√
1/5
evenly spread apart, so that Wn = Θ( n). Then, provided Hn = ω(n−13/20 Cn ), Corollary 21 says
1/15 2/15
√
n
Cn
b
= oP ( n),
d S2 | S0,2 = OP
2/3
Hn
so that each true knot has a detected knot that is much closer to it than all other true knots. Note
that Cn ≥ s0 Hn , and combining this with the requirement on Hn reveals the implicit requirement
Cn = ω(n−3/16 ), which in turn implies that Hn = ω(n−11/16 ). This seems to be a weak requirement
on the minimum nonzero change in slopes that is present in θ0 .
6.2
Changepoint detection on a graph
We depart from the 1-dimensional setting considered throughout the paper thus far, and study the
model (1) in a case where the mean parameter has components θ0,i , i = 1, . . . , n that correspond to
nodes V = {1, . . . , n} of a graph G, with edges E = {e1 , . . . , em }. Note that, for each ` = 1, . . . , m,
we may write e` = (i, j) for some nodes i, j (and all edges are to be considered undirected, so that
(i, j) and (j, i) are equivalent). Moreover, the mean θ0 is assumed to behave in a piecewise constant
fashion over the graph, which means that there are clusters of nodes over which θ0 admits constant
values, or, equivalently, θ0,i = θ0,j for many edges (i, j) ∈ E. For estimation of θ0 , we focus on
the graph fused lasso or graph-based total variation denoising (Tibshirani et al. 2005, Hoefling 2010,
Tibshirani & Taylor 2011, Sharpnack et al. 2012), defined by
n
X
1X
(yi − θi )2 + λ
|θi − θj |,
θb = argmin
2 i=1
θ∈Rn
(33)
(i,j)∈E
for a tuning parameter λ ≥ 0. When G is a 1d chain graph (i.e., E = {(1, 2), (2, 3), . . . , (n − 1, n)}),
the estimator in (33) reduces to the “usual” 1d fused lasso estimator in (2).
In the current graph-based setting, the “changepoints” of interest are actually edges for which
the corresponding nodes display differing values, under a vector θ ∈ Rn :
SG (θ) = (i, j) ∈ E : θi 6= θj .
b For an edge (i, j) ∈ E, let Pij denote the
We use the abbreviations S0,G = SG (θ0 ) and SbG = SG (θ).
set of paths in G centered around (i, j), and embedded within two constant clusters of nodes, i.e.,
Pij =
n
(i` , i`−1 ) . . . , (i1 , i), (i, j), (j, j1 ), . . . , (j`−1 , j` )
for any ` = 1, . . . , n :
o
θ0,i = θ0,i1 = . . . = θ0,i` and θ0,j = θ0,j1 = . . . = θ0,j` .
26
We now define Wn , Hn over the graph G, in an analogous fashion to our notions in the 1d setting,
Wn =
min
max
(i,j)∈S0,G p∈Pij
|p| − 1
2
and Hn =
min
(i,j)∈S0,G
|θ0,i − θ0,j |,
(34)
where we write |p| for the number of edges that form a path p. Note that, as defined, 2Wn + 1 is the
minimax length of any path centered around a changepoint in S0 ; in other words, by construction,
for each changepoint (i, j) ∈ S0 , there exists a path of at least Wn edges embedded entirely within a
cluster on either side of (i, j). When Wn is small, this is indicative of one of the constant clusters of
nodes in θ0 being small in size. For x ∈ Rn , we define its graph-based discrete total variation to be
X
TVG (x) =
|xi − xj |.
(i,j)∈E
Finally, we must precisely define our screening distance metric in the graph-based setting. For any
two edges e1 = (i1 , j1 ), e2 = (i2 , j2 ) ∈ E, let dG (e1 , e2 ) denote the length of the shortest path that
starts either i1 or j1 , and ends at either i2 or j2 . For sets A, B ∈ E, we define the screening distance
dG (A|B) =
max
min
e1 =(i1 ,j1 )∈B e2 =(i2 ,j2 )∈A
dG (e1 , e2 ).
Hence, if dG (A|B) = k, then for any edge in B, there is a path of at most k edges starting from this
edge, and ending at an edge in A.
Between Wang et al. (2016) and Hutter & Rigollet (2016), various estimation error rates are
available for the graph fused lasso. These results take on different forms, depending on the assumptions placed on θ0 and on the graph G. Below we recite a result from Hutter & Rigollet (2016) in
the case that G is a 2d grid graph. Here θb in (33) is called the 2d fused lasso estimate.
Theorem 22 (2d fused lasso error rate, weak and strong sparsity settings, Corollary 5
of Hutter & Rigollet 2016). Assume the data model in (1), with errors i , i = 1, . . . , n i.i.d.
2
from N
√
√(0, σ ). Assume that G is a 2d grid graph, with n nodes (hence the 2d grid is of dimension
n × n). Write s0 = |S0,G |, and TVG (θ0 ) ≤ Cn , for some nondecreasing sequence Cn . Then for
λ = Θ(log n), the 2d fused lasso estimate θb in (33) satisfies
log2 n
2
b
.
kθ − θ0 kn = OP min{s0 , Cn }
n
Remark 24 (Consistency, optimality). According to the theorem, the 2d fused lasso estimator
is consistent when either s0 = o(n/ log2 n) or Cn = o(n/ log2 n). Theorem 22 covers both the weak
and strong sparsity cases (since it allows us to draw conclusions involving either s0 or Cn ). In the
case of weak sparsity, the Cn log2 n/n rate achieved by the 2d fused lasso was recently shown to be
essentially minimax optimal (differing only by log factors), over the class of signals having bounded
total variation over the 2d grid G, i.e., θ0 ∈ {θ ∈ Rn : TVG (θ) ≤ Cn }, by Sadhanala et al. (2016).
Remark 25 (Other graphs). Basically the same result as in Theorem 22 holds for 3d and higherdimensional grids (except with one fewer log factor) (Hutter & Rigollet 2016). Estimation error
rates for various types of random graphs, the complete graph, and star graphs are derived in Wang
et al. (2016), Hutter & Rigollet (2016). For simplicity, we do not consider any of these cases when
we give an application of our generic graph screening result, below; however, given the availability
of `2 rates, we remark that results over different graph models (over than the 2d grid) are certainly
possible, and are just a matter of plugging in the proper rates in the proper settings.
Here is our generic graph screening result, analogous to that in Theorem 8 for the 1d chain.
27
Theorem 23 (Generic screening result, changepoint detection on a graph). Let θ0 ∈ Rn
be piecewise constant over a graph G, and θe ∈ Rn be an estimator that satisfies kθe − θ0 k2n = OP (Rn ).
Assume that nRn /Hn2 = o(Wn ), where, recall Wn , Hn are as defined in (34). Then
e | S0,G = OP nRn .
dG SG (θ)
Hn2
Proof. The proof is again very similar to the proof of Theorem 8. Fix > 0, C > 0, and let N1 > 0
be an integer such that, for n ≥ N1 ,
C
P kθe − θ0 k2n > Rn ≤ .
4
Let N2 > 0 be an integer such that, for n ≥ N2 , we have CnRn /Hn2 ≤ Wn . Let N = max{N1 , N2 },
e | S0,G ) > rn . By definition, there
take n ≥ N , and define rn = bCnRn /Hn2 c. Suppose that dG (SG (θ)
e
exists a changepoint (i, j) ∈ S0 such that no changepoints in θ are within rn of (i, j), which in our
distance metric, means that θek is constant over all nodes k that are rn away from i or j. Construct
an arbitrary path p centered around edge (i, j) with 2rn + 1 edges
n
o
p = (irn , irn −1 ), . . . , (i1 , i), (i, j), (j, j1 ) . . . , (jrn −1 , jrn ) ,
where θ0,i = θ0,i1 = . . . = θ0,irn and θ0,j = θ0,j1 = . . . = θ0,jrn . (This is possible because rn ≤ Wn .)
Denote
z = θeirn = . . . = θei = θej = . . . = θejrn .
Also let I(p) = {irn −1 , . . . , i1 , i, j, j1 , . . . , jrn −1 } denote the internal nodes of the path p. Then
1 X
n
k∈I(p)
θek − θ0,k
2
=
2 rn
2
rn
rn Hn2
C
z − θ0,i +
z − θ0,j ≥
> Rn ,
n
n
2n
4
where the first inequality holds because, as argued before, (x − a)2 + (x − b)2 ≥ (a − b)2 /2 for all x,
and the second by definition of rn . Invoking the assumed `2 error rate for θe completes the proof.
Combining Theorems 22 and 23 gives the next and final result, whose proof is omitted.
Corollary 24 (2d fused lasso screening result, weak and strong sparsity settings). Assume
the conditions in Theorem
22, so that
√ G is a 2d grid, and s0 = |S0,G |, TVG (θ0 ) ≤ Cn . Also assume
√ √
that Hn = ω(min{ s0 , Cn } log n/ Wn ). Let θb denote the 2d fused lasso estimate in (33), with the
choice of tuning parameter λ = Θ(log n). Then
log2 n
d SbG | S0,G = OP min{s0 , Cn }
.
Hn2
Remark 26 (Screening over a 2d grid). Consider, as a concrete example, a case in which
θ0 is
√
√
piecewise constant with just 2 pieces or clusters, over the 2d grid G (of dimension, recall, n × n).
Then s0 = |S0,G | reflects the length of the boundary separating the 2 pieces. In the typical case (in
which the
√ 2 pieces are of roughly equal size, and both have volume proportional to n), this scales as
s0 = Θ( n). Moreover, in the typical
√ case, the length Wn of the longest path on either side of this
boundary also scales as Wn = Θ( n). Here, Cn ≥ s0 Hn , which is larger than s0 unless Hn is quite
small (Hn ≤ 1). Thus when Hn is large (Hn > 1), we can use Corollary 24 to conclude that
d SbG | S0,G = OP
28
√
n log2 n
.
Hn2
√
Roughly speaking, this says if we were to place a “tube” of radius Bn = OP ( n log2 n/Hn2 ) around
the boundary edges S0,G , containing edges that are at distance of at most Bn from S0,G , then each
b
changepoint in S0,G has a corresponding
√ changepoint in SG lying inside this tube. Seeing
√ detected
as the entire grid itself is of dimension n × n, this statement is not really interesting unless Hn
is fairly large, say Hn = Θ(n1/8 log n). Then Bn = OP (n1/4 ), giving a reasonably tight tube around
the boundary S0,G .
An illustration of the true changepoints versus those detected by the 2d fused lasso, in a simple
simulated 2d image example, is given in Figure 8. See the figure caption for details.
Mean
Data
Estimate
Figure 8: An example on a 2d grid graph, of dimension 100 × 100 (i.e., n = 10, 000). The left panel displays
the mean θ0 ; the middle panel the data y, whose entries were generated by adding i.i.d. N (0, 1) noise to θ0 ;
and the right panel displays the 2d fused lasso estimate, whose tuning parameter λ was chosen to minimize
the `2 estimation error in retrospect. The mean only has two constant pieces, taking values 0 and 1, denoted
by dark gray and light gray colors, respectively, in the left panel. A consistent color scale is used throughout
the three plots. The true changepoint set S0,G is drawn in red, and the estimated changepoint set SbG in blue.
In this example, dG (SbG | S0,G ) = 1.
7
Discussion
We have derived a new `2 error bound for the fused lasso in a strong sparsity setting, which, to the
best of our knowledge, yields the sharpest available rate in this setting. We have also undertaken a
detailed study of the manner in which `2 error bounds for generic estimators θe can be used to prove
e and after simple post-processing, changepoint recovery results
changepoint screening results for θ,
e
for θ. As a prime example, we have derived new changepoint screening and recovery results for the
fused lasso estimator, in various settings, based solely on its `2 error guarantees, in these settings.
To reiterate, our general technique for analyzing changepoint screening and recovery properties is
not specific to the fused lasso, and is potentially much more broadly applicable, as it only assumes
knowledge of the `2 error rate of the estimator θe in question. This could be applied even outside of
the typical Gaussian data model.
We have also presented extensions to the piecewise linear segmentation and graph changepoint
detection problems, as well as detailed simulations. The code to run all our simulations is located
at https://github.com/linnylin92/fused_lasso, and relies on the R package genlasso.
29
References
Boysen, L., Kempe, A., Liebscher, V., Munk, A. & Wittich, O. (2009), ‘Consistencies and rates of
convergence of jump-penalized least squares estimators’, The Annals of Statistics 37(1), 157–183.
Chan, H. P. & Walther, G. (2013), ‘Detection with the scan and the average likelihood ratio’,
Statistica Sinica 23(1), 409–428.
Dalalyan, A. S., Hebiri, M. & Lederer, J. (2014), ‘On the prediction performance of the lasso’, arXiv
preprint arXiv:1402.1700 .
Davies, L. & Kovac, A. (2001), ‘Local extremes, runs, strings and multiresolution’, Annals of Statistics 21(1), 1–65.
Donoho, D. L. & Johnstone, I. M. (1994), ‘Ideal spatial adaptation by wavelet shrinkage’, Biometrika
81(3), 425–455.
Donoho, D. L. & Johnstone, I. M. (1998), ‘Minimax estimation via wavelet shrinkage’, Annals of
Statistics 26(8), 879–921.
Duembgen, L. & Walther, G. (2008), ‘Multiscale inference about a density’, The Annals of Statistics
36(4), 1758–1785.
Frick, K., Munk, A. & Sieling, H. (2014), ‘Multiscale change point inference’, Journal of the Royal
Statistical Society: Series B (Statistical Methodology) 76(3), 495–580.
Fryzlewicz, P. (2007), ‘Unbalanced Haar technique for nonparametric function estimation’, Journal
of the American Statistical Association 102(480), 1318–1327.
Fryzlewicz, P. (2014), ‘Wild binary segmentation for multiple change-point detection’, The Annals
of Statistics 42(6), 2243–2281.
Fryzlewicz, P. (2016), ‘Tail-greedy bottom-up data decompositions and fast multiple change-point
detection’. Available at http://stats.lse.ac.uk/fryzlewicz/tguh/tguh.pdf.
Harchaoui, Z. & Lévy-Leduc, C. (2010), ‘Multiple change-point estimation with a total variation
penalty’, Journal of the American Statistical Association 105(492).
Hoefling, H. (2010), ‘A path algorithm for the fused lasso signal approximator’, Journal of Computational and Graphical Statistics 19(4), 984–1006.
Hutter, J.-C. & Rigollet, P. (2016), ‘Optimal rates for total variation denoising’, arXiv preprint
arXiv:1603.09388 .
Kim, S.-J., Koh, K., Boyd, S. & Gorinevsky, D. (2009), ‘`1 trend filtering’, SIAM Review 51(2), 339–
360.
Mammen, E. & van de Geer, S. (1997), ‘Locally adaptive regression splines’, The Annals of Statistics
25(1), 387–413.
Qian, J. & Jia, J. (2012), ‘On pattern recovery of the fused lasso’, arXiv preprint arXiv:1211.5194 .
Rinaldo, A. (2009), ‘Properties and refinements of the fused lasso’, Annals of Statistics 37(5), 2922–
2952.
Rojas, C. R. & Wahlberg, B. (2014), ‘On change point detection using the fused lasso method’,
arXiv preprint arXiv:1401.5408 .
30
Rudin, L., Osher, S. & Faterni, E. (1992), ‘Nonlinear total variation based noise removal algorithms’,
Physica D: Nonlinear Phenomena 60(1–4), 259–268.
Sadhanala, V., Wang, Y.-X. & Tibshirani, R. J. (2016), ‘Total variation classes beyond 1d: Minimax
rates, and the limitations of linear smoothers’, arXiv preprint arXiv:1605.08400 .
Sharpnack, J., Rinaldo, A. & Singh, A. (2012), Sparsistency of the edge lasso over graphs, in ‘Proceedings of the 15th International Conference on Artificial Intelligence and Statistics’, pp. 1028–
1036.
Steidl, G., Didas, S. & Neumann, J. (2006), ‘Splines in higher order TV regularization’, International
Journal of Computer Vision 70(3), 214–255.
Tibshirani, R. J. (2014), ‘Adaptive piecewise polynomial estimation via trend filtering’, The Annals
of Statistics 42(1), 285–323.
Tibshirani, R. J. & Taylor, J. (2011), ‘The solution path of the generalized lasso’, Annals of Statistics
39(3), 1335–1371.
Tibshirani, R., Saunders, M., Rosset, S., Zhu, J. & Knight, K. (2005), ‘Sparsity and smoothness
via the fused lasso’, Journal of the Royal Statistical Society: Series B (Statistical Methodology)
67(1), 91–108.
van de Geer, S. (1990), ‘Estimating a regression function’, Annals of Statistics 18(2), 907–924.
van de Geer, S. (2000), Empirical Processes in M-Estimation, Cambridge University Press.
Venkatraman, E. S. (1992), Consistency results in multiple change-point problems, PhD thesis,
Department of Statistics, Stanford University.
Vostrikova, L. (1981), ‘Detecting ‘disorder’ in multidimensional random processes’, Soviet Mathematics Doklady 24, 55–59.
Wang, Y.-X., Sharpnack, J., Smola, A. & Tibshirani, R. J. (2016), ‘Trend filtering on graphs’,
Journal of Machine Learning Research . To appear.
31
A
Proof of Theorem 4
Here and henceforth, we write N (r, S, k · k) to denote the covering number of a set S in a norm k · k,
i.e., the smallest number of k · k-balls of radius r needed to cover S. We call log N (r,
√ S, k · k) the log
covering or entropy number. Recall that we define the scaled norm k · kn = k · k2 / n.
In the proof of Theorem 4, we will rely on the following result from van de Geer (1990) (which
is derived closely from Dudley’s chaining for sub-Gaussian processes).
Theorem 25 (Theorem 3.3 of van de Geer 1990). Assume that = (1 , . . . , n ) ∈ Rn has
i.i.d. components drawn from a sub-Gaussian distribution, as in (8). Consider a set X ⊆ Rn with
kxkn ≤ 1 for all x ∈ X , and let K(·) be a continuous function upper bounding the k · kn entropy of
X , i.e., K(r) ≥ log N (r, X , k · kn ). Then there are constants C1 , C2 , C3 , C4 > 0 (depending only on
M, σ, the parameters of the underlying sub-Gaussian distribution in (8)), such that for all t > C1 ,
with
Z t0 p
K(r) dr,
t > C2
0
2
where t0 = inf{r : K(r) ≤ C3 t }, we have
|> x|
P sup √ > t ≤ 2 exp(−C4 t2 ).
n
x∈X
Now we give the proof of Theorem 4.
b 2 and the other in
Proof of Theorem 4. The proof is given in two parts, one in which we bound kδk
2
2
b
b
b
which we bound kb
xk2 . Recall that δ = P0 (θ − θ0 ) and x
b = P1 θ. Each part begins with a different
“basic inequality”, established by comparing the fused lasso objective at different points. First, we
define the following events,
p
|> z|
Ω0 = sup
≤ γcI (log s0 + log log n)s0 log n ,
(35)
z∈M kzk2
|> w|
≤ γcR (ns0 )1/4 ,
(36)
Ω1 = sup
1/2
1/2
w∈R⊥ kD−S0 wk1 kwk2
√
|> δ|
Ω2 = sup
≤ γcS s0 ,
(37)
δ∈R kδk2
where γ > 1 is parameter free to vary in our analysis, cI , cR > 0 are the constants in Lemmas 6, 7,
and cS > 0 is a constant to be determined below. Focusing on the third event, we will lower bound
its probability by applying Theorem 25 to X = R ∩ {δ : kδkn ≤ 1}. Note that
log N (r, R ∩ {δ : kδkn ≤ 1}, k · kn ) ≤ (s0 + 1) log(3/r),
as R is (s0 + 1)-dimensional, and it is well-known that in Rd , the number of balls of radius r that
are needed to cover the unit ball is at most (3/r)d . The quantity t0 in Theorem 25 may be taken to
be t0 = inf{r : (s0 + 1) log(3/r) ≤ C3 C12 } = 3 exp(−C3 C12 /(s0 + 1)). The restrictions on t are hence
t > C1 , as well as
Z t0 p
t > C2
(s0 + 1) log(3/r) dr.
0
But, writing erf(·) for the error function,
Z
C2
0
t0
r
r
p
√
1 1
1
(s0 + 1) log(3/r) dr = ( s0 + 1) · 3C2 r log − erf
log
r
2
r
32
t0 /3
0
√
≤ C2 s0 ,
where the constant C2 > 0 is adjusted to be larger, as needed. Let us define cS = max{C1 , C2 } and
√
CS = C4 . Then we have by Theorem 25, for t = γcS s0 and any γ > 1,
√
√
|> δ|
|> δ|
2 2
√
1 − 2 exp(−CS γ cS s0 ) ≤ P sup
≤ γcS s0 = P sup
≤ γcS s0 = P(Ω2 ). (38)
nkδkn
δ∈R kδk2
δ∈R
b Comparing the objective in (2) at θb = P0 θb + P1 θb = θ0 + δb + x
Controlling δ.
b and at θ0 + x
b, we
have
b 1 ≤ kb
kδb + x
b − k22 + λkDθk
x − k22 + λkD(θ0 + x
b)k1 ,
and by rearranging terms we obtain our basic inequality,
b 2 ≤ 2δb> + λ kD(θ0 + x
b
kδk
b
)k
−
kD
θk
1
1 ,
2
(39)
which follows from the fact that δb> x
b = 0 (as they lie in orthogonal subspaces). Furthermore, since
θb = θ0 + δb + x
b,
b 1 ≤ kDδk
b 1 = kDS δk
b 1,
kD(θ0 + x
b)k1 − kDθk
0
where we used the triangle inequality, and the fact that D−S0 δb = 0. So from our basic inequality in
(39), we have that
b 2 ≤ 2δb> + λkDS δk
b 1,
kδk
2
0
b 2 , we get
and dividing by kδk
b 2≤2
kδk
b 1
kDS0 δk
|δb> |
+λ
.
b 2
b 2
kδk
kδk
Now observe that
b 1=
kDS0 δk
s0
X
i=1
|δbti+1 −δbti | ≤ 2
sX
0 +1
i=1
v
v
u
u s0 +1
r
sX
0 +1
u
u X ti − ti−1
s0 b
t
2
b
b
|δti | ≤ 2 (s0 + 1)
δti ≤ 4ts0
δbt2i = 4
kδk2 .
Wn
Wn
i=1
i=1
The second inequality used Cauchy-Schwartz, and the last equality used that δb is piecewise constant
on the blocks B0 , . . . , Bs0 , as δb ∈ R = span{1B0 , . . . , 1Bs0 }. Hence, on the event Ω2 , we have
r
√
s0
b
.
(40)
kδk2 ≤ 2γcS s0 + 4λ
Wn
Controlling x
b. We can establish our next basic inequality by comparing the objective in (2) at θb
b
and θ0 + δ,
b 1 ≤ kδb − k2 + λkD(θ0 + δ)k
b 1,
kb
x + δb − k22 + λkDθk
2
or, rearranged,
b 1 − kDS θk
b 1 − kD−S x
kb
xk22 ≤ 2> x
b + λ kDS0 (θ0 + δ)k
bk1
0
0
≤ 2> x
b + λ kDS0 x
bk1 − kD−S0 x
bk1 ,
(41)
where the first line used x
b> δb = 0 and D−S0 θ0 = D−S0 δb = 0, and the second used θb = θ0 + δb + x
b and
the triangle inequality.
33
Decompose x
b = zb + w,
b where zb ∈ M is the lower interpolant to x
b, as defined in Lemma 5, and
w
b=x
b − zb is the remainder. Combining the basic inequality in (41) with (9) and (10) from Lemma 5,
kb
xk22 ≤ 2> zb + 2> w
b + λ kDS0 zbk1 − kD−S0 zbk1 − kD−S0 wk
b 1
r
s0
kb
z k2 + 2> w
≤ 2> zb + 4λ
b − λkD−S0 wk
b 1.
(42)
Wn
On the event Ω0 in (35)
> zb ≤ γcI
p
(log s0 + log log n)s0 log nkb
z k2 .
Further, on the event Ω1 in (36), since P1 w
b ∈ R⊥ , kD−S0 P1 wk
b 1 = kD−S0 wk
b 1 , and kP1 wk
b 2 ≤ kwk
b 2,
1/2
1/2
> P1 w
b ≤ γcR (ns0 )1/4 kD−S0 wk
b 1 kwk
b 2 ,
Also, on the event Ω2 in (37), since P0 w
b ∈ R,
√
> P0 w
b ≤ γcS s0 kwk
b 2.
Hence, on the event Ω0 ∩ Ω1 ∩ Ω2 , combining the last three displays with (42),
r
p
√
s0
kb
xk22 ≤ 2 γcI (log s0 + log log n)s0 log n + 2λ
b 2+
kb
z k2 + 2γcS s0 kwk
Wn
1/2
1/2
2γcR (ns0 )1/4 kD−S0 wk
b 1 kwk
b 2
1/2
− λkD−S0 wk
b 1 . (43)
1/2
Consider the first case in which 2γcR (ns0 )1/4 kD−S0 wk
b 1 kwk
b 2 ≥ λkD−S0 wk
b 1 . Then
2
γcR √
kD−S0 wk
b 1≤4
ns0 kwk
b 2,
λ
and from (43), on the event Ω0 ∩ Ω1 ∩ Ω2 ,
kb
xk2 ≤ 2γcI
p
r
(log s0 + log log n)s0 log n + 4λ
√
4γ 2 c2R ns0
√
s0
.
+ 2γcS s0 +
Wn
λ
1/2
1/2
where in the above we used (11). In the case that 2γcR (ns0 )1/4 kD−S0 wk
b 1 kwk
b 2
we have from (43), on the event Ω0 ∩ Ω1 ∩ Ω2 ,
r
p
√
s0
+ 2γcS s0 .
kb
xk2 ≤ 2γcI (log s0 + log log n)s0 log n + 4λ
Wn
(44)
< λkD−S0 wk
b 1,
Therefore, the bound (44) always holds on the event Ω0 ∩ Ω1 ∩ Ω2 .
b 2 , combining (44) and (40), we see that
Putting it all together. As kθb − θ0 k2 ≤ kb
xk2 + kδk
r
√
p
4γ 2 c2R ns0
√
s0
kθb − θ0 k2 ≤ 4γcS s0 + 8λ
+ 2γcI (log s0 + log log n)s0 log n +
,
Wn
λ
on the event Ω0 ∩ Ω1 ∩ Ω2 . We see that there exists a constant c > 0, such that for large enough n,
and any γ > 1,
!
λ2
n
2
4
b
kθ − θ0 k2 ≤ γ cs0 (log s0 + log log n) log n +
+ 2 ,
(45)
Wn
λ
34
on the event Ω0 ∩ Ω1 ∩ Ω2 . Furthermore, using the union bound along with Lemmas 6, 7, and (38),
we find that
P (Ω0 ∩ Ω1 ∩ Ω2 )c ≤ 2 exp − CI γ 2 c2I (log s0 + log log n) +
√
2 exp(−CR γ 2 c2R s0 ) + 2 exp(−CS γ 2 c2S s0 ) ≤ exp(−Cγ 2 ),
for an appropriately defined constant C > 0. Optimizing the bound in (45) to choose the tuning
parameter λ yields λ = (nWn )1/4 . Plugging this in gives the final result.
B
Proofs of Lemmas 5, 6, 7
Proof of Lemma 5. We give an explicit construction of a lower interpolant z ∈ M to x, given the
changepoints 0 = t0 < . . . < ts0 +1 = n. We will use the notation a+ = max{0, a}. For i = 0, . . . , s0 ,
define z (i+) ∈ Rti+1 −ti by setting gi+ = sign(xti ) and
n
o
(i+)
zj
= gi+ · min (gi+ xti +1 )+ , . . . , (gi+ xti +j )+ , j = 1, . . . , ti+1 − ti .
Similarly, define z (i−) ∈ Rti+1 −ti by setting gi− = sign(xti+1 −1 ) and
n
o
(i−)
zj
= gi− · min (gi− xti +j )+ , . . . , (gi− xti+1 )+ , j = 1, . . . , ti+1 − ti .
(i+)
(i+)
(i−)
t
−t
i+1
i
is a nonincreasing sequence, and
Note that z1 = xti +1 and zti+1 −ti = xti+1 ; also, {|zj |}j=1
(i−) ti+1 −ti
{|zj |}j=1
is nondecreasing. Furthermore,
(i−)
(i+)
(i+)
(i−)
≥ 0,
j = 1, . . . , ti+1 − ti .
sign z1
· sign zj
≥ 0 and sign zti+1 −ti · sign zj
Lastly, notice that there exists a point j 0 ∈ 1, . . . , ti+1 − ti − 1 (not necessarily unique) such that
(i+)
= zj 0 +1 = zj
(i−)
= zj 0
min
zk
min
zk
k∈{1,...,ti+1 −ti }
k∈{1,...,ti+1 −ti }
(i+)
(i−)
(i+)
,
j = j 0 + 1, . . . , ti+1 − ti ,
(46)
(i−)
,
j = 1, . . . , j 0 .
(47)
= zj
(i+)
(i−)
We construct by zti +j = zj
for j = 1, . . . , j 0 , and zti +j = zj
for j = j 0 + 1, . . . , ti+1 − ti . Letting
0
0
ti = ti + j and repeating this process for i = 0, . . . , s0 , we have constructed z ∈ M.
We now verify the claimed properties for the constructed lower interpolant z. For i = 0, . . . , s0 ,
and any j = 1, . . . , ti+1 − ti , we have
(i+)
sign(zj
(i+)
|zj |
) · sign(xti +j ) ≥ 0,
≤ |xti +j |,
Further, for any j = 1, . . . , ti+1 − ti − 1,
sign (Dz (i+) )j · sign (Dx)ti +j ≥ 0,
(Dz (i+) )j ≤ (Dx)ti +j .
(48)
(49)
(50)
(51)
To see why (50) holds, note that sign(Dz (i+) )j ∈ {−1, 0}, (Dz (i+) )j < 0 imply (D(gi+ x)+ )ti +j < 0.
To see why (51) holds, if (Dz (i+) )j 6= 0, then we know that
n
o
(i+)
(i+)
|zj+1 − zj | ≤ min (gi+ xti +j+1 )+ , (gi+ xti +j )+ − (gi+ xti +j )+ ≤ |xti +j+1 − xti +j |,
35
where we used the observation that | min{a, b} − b| ≥ | min{a, b, c} − min{b, c}|.
It can be shown by nearly equivalent steps that z (i−) , and z both satisfy properties analogous
to (48)–(51). Using (48) and (49) on z gives (11). Using (50) and (51) on z gives (9) (note that if
sign(a) = sign(b) and |a| > |b|, then |a| = |b| + |a − b|). Because zti +1 = xti +1 and zti+1 = xti+1 for
all i = 0, . . . , s0 , we have the equality in (10) (since Dti z = zti +1 − zti = xti +1 − xti = Dti x).
Finally, for each i = 0, . . . , s0 , define t00i = t0i if |zt0i | ≥ |zt0i +1 | and t00i = t0i + 1 otherwise. Observe
that by (46) and (47), it holds that |zt00i | = minj=1,...,ti+1 −ti |zti +j |. The inequality in (10) is finally
established by the following chain of inequalities:
kDS0 zk1 =
=
s0
X
s0
X
|zti +1 − zti | ≤
i=1
s0
X
|zti +1 | + |zti |
i=1
|zti +1 | − |zt00i | + |zti | − |zt00i−1 | + |zt00i−1 | + |zt00i |
i=1
≤ kD−S0 zk1 + 2
s0
X
r
|zt00i | ≤ kD−S0 zk1 + 4
i=0
s0
kzk2 ,
Wn
where in the second inequality, we used |a| − |c| ≤ |a − c| ≤ |a − b| + |b − c|, and in the last inequality,
we used the above property of zt00i and
v
v
u s0
u s0
r
X
u
u X ti+1 − ti
√
s0
|zt00i | ≤ 2 s0 t
|zt00i |2 ≤ 2ts0
zt200 ≤ 2
kzk2 .
i
W
W
n
n
i=0
i=0
i=0
s0
X
This completes the proof.
Proof of Lemma 6. We consider ∈ Rn , an i.i.d. sub-Gaussian vector as referred to in the statement
of the lemma, and arbitrary z ∈ M. In this proof, we will also consider E(t) and Z(t), real-valued
functions over [0, n], constructed so that E(t) = dte for all t (i.e., E(t) is a step function), Z(t) = zt
for t = 1, . . . , n, and Z(t) is smooth and monotone over (ti , t0i ] and (t0i , ti+1 ] for i = 0, . . . , s0 . These
functions will also
R t satisfy the boundary conditions E(0) = 1 and Z(0) = z1 .
Let F (t) = 0 E(u) du. As is random, E(t) and F (t) are also random. It can be shown that
there exists constants cI , CI > 0 such that for any γ > 1,
P
p
|F (t) − F (ti )|
p
≤ γcI log s0 + log log n, for t ∈ (ti , ti+1 ], i = 0, . . . , s0
|t − ti |
!
≥ 1 − 2 exp − CI γ 2 c2I (log s0 + log log n) . (52)
So as not to distract from the main flow of ideas, we now proceed to prove Lemma 6, and we later
provide a proof of (52). Let Ω3 denote the event in consideration on the left-hand side of (52). By
integration by parts,
Z
t0i
E(t)Z(t) dt =
Z(t0i )(F (t0i )
Z
− F (ti )) −
ti
where Z 0 (t) =
Z
d
dt Z(t).
t0i
Z 0 (t)(F (t) − F (ti )) dt
ti
Thus, on the event Ω3 ,
t0i
E(t)Z(t) dt ≤ γcI
p
log s0 + log log n
|Z(t0i )|
ti
p
t0i − ti +
Z
t0i
ti
36
√
Z (t) t − ti dt
0
!
,
(53)
since Z 0 does not change sign within the intervals (ti , t0i ], (t0i , ti+1 ] (as z ∈ M). For n large enough,
we can upper bound the last term in (53) as follows
t0i
Z
√
Z 0 (t) t − ti dt =
Z
ti +n−1
√
Z 0 (t) t − ti dt +
√
Z 0 (t) t − ti dt .
(54)
ti +n−1
ti
ti
t0i
Z
Using integration by parts and the triangle inequality, on the second term in (54),
t0i
Z t0i
p
√
Z(ti + n−1 )
Z(t)
1
0
0
√
√
Z (t) t − ti dt = |Z(ti )| ti − ti +
+
dt .
2
n
t − ti
−1
−1
ti +n
ti +n
Z
0
(55)
By Cauchy-Schwartz on the last term in (55),
Z
t0i
ti +n−1
Z
Z(t)
√
dt ≤
t − ti
!1/2
t0i
t0i
Z
2
Z(t) dt
ti +n−1
Z
ti +n−1
!1/2
1
dt
t − ti
!1/2
t0i
p
Z(t)2 dt
≤
2 log n.
(56)
ti +n−1
Now examining the first term in (54),
ti +n−1
Z
√
Z 0 (t) t − ti dt ≤ n−1/2
Z
ti
ti +n−1
Z 0 (t) dt =
ti
|Z(ti + n−1 ) − Z(ti )|
√
.
n
But because we only require Z to be a piecewise monotonic and smooth interpolant then we are at
liberty to make Z(ti + n−1 ) = Z(ti ), forcing this term to be 0. In order to bound Z(t0i ), notice that
because |Z(t)| is non-increasing over the interval (ti , t0i ] we have that
Z(t0i )2 |t0i
Z
t0i
Z(t)2 dt.
− ti | ≤
(57)
ti
Combining (53)–(57), we have that on the event Ω3 ,
Z
t0i
r
E(t)Z(t) dt ≤ αn
2+
ti
log n
2
! Z
!1/2
t0i
2
Z(t) dt
ti
|Z(ti )|
+ αn √ .
n
(58)
√
where we have abbreviated αn = γcI log s0 + log log n. Through nearly identical steps we can show
that on the event Ω3 ,
Z
r
ti+1
E(t)Z(t) dt ≤ αn
2+
t0i
log n
2
! Z
!1/2
ti+1
|Z(ti+1 )|
√
.
n
(59)
1/2
kzk1
Z(t)2 dt
+ 2αn √ ,
n
(60)
2
Z(t) dt
+ αn
t0i
Therefore
Z
n
E(t)Z(t) dt ≤
0
s0
X
i=0
√
Z
t0i
Z
!
ti+1
E(t)Z(t) dt +
E(t)Z(t) dt
t0i
ti
r
≤ αn 2s0 + 2 2 +
37
log n
2
! Z
0
n
where in the second line we applied (58), (59), and the Cauchy-Schwartz
inequality. Because we can
Rn
2
1/2
choose
Z(t)
to
be
arbitrarily
close
to
z
over
all
t,
the
integral
(
Z(t)
is approaching kzk2
dte
0
Rn
√ dt)
and 0 E(t)Z(t) dt is approaching > z. Furthermore, because kzk1 ≤ nkzk2 , the first term in (60)
dominates. Hence on the event Ω3 , we have established yet
p
|> z| ≤ γcI (log s0 + log log n)s0 log nkzk2 ,
where the constant cI is adjusted to be larger, as needed. Noting that the event Ω3 does not depend
on z, the result follows.
0
Vi , where for each i = 0, . . . , s0 ,
Proof of claim (52). We will construct a covering for V = ∪si=0
r
r
n
n
Vi =
1A : A = {ti , . . . , t}, t = ti +1, . . . , n ∪
1A : A = {t, . . . , ti }, t = 1, . . . , ti −1 .
|A|
|A|
p
Note that our scaling is such
pthat, for any a = n/|A|1A , where A ⊆ {1, . . . , n}, we have kakn = 1.
Further, for any other b = n/|B|1B , where B ⊆ {1, . . . , n}, we have
|A ∩ B|
|A ∩ B|
|A \ B| |B \ A|
p
ka − bk2n = p
+
=2 1− p
+
,
(61)
|A|
|B|
( |A| − |B|)2
|A||B|
We first
√ construct a covering for each set Vi , i = 0, . . . , s0 , restricting our attention to a radius
0 < r < 2. Let α = d(1 − r2 /2)−2 e, and consider the set
Ci =
r
n
j
1A : A = ti , . . . , min{ti + α , n} , j = 1, . . . , dlog n/ log αe
|A|
r
n
j
∪
1A : A = max{ti − α , 1}, . . . , ti , j = 1, . . . , dlog n/ log αe .
|A|
Here, the set Ci has at most 2dlog n/ log αe ≤ 4 log n/ log α elements, and by (61), balls of radius r
around elements in Ci cover the set Vi . This establishes that
N (r, Vi , k · kn ) ≤
−2 log n
.
log(1 − r2 /2)
(62)
√
0
Vi can be obtained by just taking a
For a radius 0 < r < 2, the covering number for V = ∪si=0
union of the covers in (62) over i = 0, . . . , s0 , giving
s0
X
− log n
N (r, V, k · kn ) ≤
N (r, Vi , k · kn ) ≤ 2(s0 + 1)
.
log(1 − r2 /2)
i=0
(63)
√
√
Using (61) once more, the diameter of the set V is 2, hence if r ≥ 1/ 2, then we need only 1 ball
to cover V. Combining this fact with (63), we obtain
√
− log n
2(s + 1)
if 0 < r < 1/ 2
0
2
log(1 − r /2)
N (r, V, k · kn ) ≤
.
(64)
√
1
if r ≥ 1/ 2
Now let us apply Theorem
√ 25, with X = V. First, we remark that the quantity t0 in Theorem 25
may be taken to be t0 = 1/ 2. The bounds on t in the theorem are t > C1 , as well as
Z 1/√2 s
− log n
dr.
t > C2
log 2(s0 + 1)
log(1 − r2 /2)
0
38
Next, we know that the right-hand side above is upper bounded by
Z
C2
√
1/ 2
"
q
s
log 2(s0 + 1) log n +
0
s
= C2
log
−1
log(1 − r2 /2)
#
dr
v
Z 1/2 u
u
√
log 2(s0 + 1) log n
tlog
+ C2 2
2
0
!
1
log
1
1−x2
dx.
One can verify that the the integral in the second term above converges to a finite
√ constant (upper
bounded by 1 in fact). Thus the entire expression above is upper bounded by C2 log s0 + log log n,
where the constant C2 > 0 is adjusted to
√ be larger, as needed. Therefore, letting cI = max{C1 , C2 },
we may restrict our attention to t > cI log s0 + log log n in Theorem 25, and letting CI = C4 , the
conclusion reads, for t = γcI and γ > 1,
p
> a
P sup √ > γcI log s0 + log log n ≤ 2 exp − CI γ 2 c2I (log s0 + log log n) .
n
a∈V
p
Recalling the form of a = n/|A|1A ∈ V, the above may be rephrased as
Pt
p
j=t j
> γcI log s0 + log log n, for t = 1, . . . , n, i = 0, . . . , s0
P p i
|t − ti |
≤ 2 exp − CI γ 2 c2I (log s0 + log log n) . (65)
Finally, consider the following event
)
(
p
|F (t) − F (ti )|
p
≤ γcI log s0 + log log n, for t = 1, . . . , n, i = 0, . . . , s0 .
Ω4 =
|t − ti |
Rt
Pt
Recalling that E(t) = dte for all t ∈ [0, 1], we have F (t) = 0 E(u) du = j=0 j for t = 1, . . . , n. In
2 2
(65), we have thus shown P(Ω4 ) ≥ 1 − 2 exp(−CI γp
cI (log s0 + log log n)). Note that |F (t) − F (ti )|
is piecewise linear
with
knots
at
t
=
1,
.
.
.
,
n
and
|t − ti | is concave in between these knots, so if
p
√
|F (t) − F (ti )|/ |t − ti | ≤ γcI log s0 + log log n for t = 1, . . . , n, then the same bound must hold
over all t ∈ [0, n]. This shows that Ω4 ⊇ Ω3 , where Ω3 is the event in question in the left-hand side
of (52); in other words, we have verified (52).
For the proof of Lemma 7, we will need the following result from van de Geer (1990).
Lemma 26 (Lemma 3.5 of van de Geer 1990). Assume the conditions in Theorem 25, and
additionally, assume that for some ζ ∈ (0, 1) and K > 0,
K(r) ≤ Kr−2ζ ,
where, recall, K(r) is a continuous function upper bounding the entropy number log N (r, X , k · kn ).
Then there exists constants C0 , C1 (depending only on M, σ in (8)) such that for any t ≥ C0 ,
√
|> x|
P sup √
>
t
K
≤ exp(−C1 t2 K).
1−ζ
x∈X
nkxkn
Proof of Lemma 7. Recall that for i = 0, . . . , s0 , we define Bi = {ti + 1, . . . , ti+1 }. For i = 0, . . . , s0 ,
√
also define ni = |Bi |, the scaled norm k · kni = k · k2 / ni , and
n
o
Xi = w(i) ∈ Rni : (1(i) )> w(i) = 0, kD(i) w(i) k1 ≤ 1, kw(i) kni ≤ 1 .
39
Here, we write 1(i) ∈ Rni for the vector of all 1s, and D(i) ∈ R(ni −1)×n for the difference operator,
as in (5) but of smaller dimension. The set Xi is the discrete total variation space in Rni , where all
elements are centered and have scaled norm at most 1. From well-known results on entropy bounds
for total variation spaces (e.g., from Lemma 11 and Corollary 12 of Wang et al. (2016)), we have
C
,
r
for a universal constant C > 0. Hence we may apply Lemma 26 with X = Xi and ζ = 1/2: for the
random variable
(
)
(i)
|>
Bi w |
(i)
: w ∈ Xi ,
Mi = sup √
1/2
ni kw(i) kni
we may take t = γC0 in the lemma, for any γ > 1, and conclude that
√
P Mi > γC0 C ≤ exp(−C1 γ 2 C02 C).
log N (r, Xi , k · kni ) ≤
Notice that we may rewrite Mi as
(
Mi = sup
and therefore
P
(i)
|>
Bi w |
1/4
1/2
:w
1/2
ni kD(i) w(i) k1 kw(i) k2
(i)
|>
Bi w |
sup
1
w(i) ∈Rni , ( (i) )> w(i) =0
1/2
1/2
kD(i) w(i) k1 kw(i) k2
(i)
∈ R , (1 ) w
(i) >
ni
√ 1/4
> γC0 Cni
Using the union bound,
)
(i)
=0 ,
≤ exp(−C1 γ 2 C02 C).
(i)
|>
Bi w |
sup
P
(i) (i) 1/2
(i) 1/2
w(i) ∈Rni , (1(i) )> w(i) =0 kD w k1 kw k2
i=0,...,s0
√ 1/4
> γC0 Cni ≤ (s0 + 1) exp(−C1 γ 2 C02 C).
√
Define the constants cR = max{C0 C, 1} and CR = max{C1 /2, 1}. Then this ensures that we have
√
2CR γ 2 c2R s0 ≥ log(s0 + 1) for any γ > 1 and any s0 , thus
P
w
(i)
|>
Bi w |
sup
(i)
, (1 ) w
i=0,...,s0
ni
∈R
(i) >
(i)
√
> γcR (ni s0 )1/4 ≤ exp(−CR γ 2 c2R s0 ).
(i) (i) 1/2
(i) 1/2
=0 kD w k1 kw k2
The proof is completed by noting the following: if w ∈ R⊥ , then (1(i) )> wBi = 0 for all i = 0, . . . , s0 ,
and so on the event in consideration in the last display,
>
| w| ≤
s0
X
|>
Bi wBi |
≤
1/4
γcR s0
i=0
≤
1/4
γcR s0
s0
X
1/4
1/2
i=0
s0
X
!1/2
(i)
kD wBi k1
i=0
=
1/2
ni kD(i) wBi k1 kwBi k2
1/4
1/2
γcR s0 kD−S0 wk1
s0
X
!1/2
1/2
ni kwBi k2
i=0
s0
X
!1/2
1/2
ni kwBi k2
i=0
s0
X
!1/4
≤
1/4
1/2
γcR s0 kD−S0 wk1
=
i=0
1/4
1/2
1/2
γcR s0 kD−S0 wk1 kwk2 n1/4 ,
40
kwBi k22
s0
X
i=0
!1/4
ni
by two successive uses of Cauchy-Schwartz.
C
Proofs of Lemmas 12, 13
Both proofs follow from standard techniques in convex analysis.
Proof of Lemma 12. We first consider the convex optimization problem
min a> x subject to kx − ck2 ≤ r,
(66)
x∈Rm
whose Lagrangian may be written as, for a dual variable λ ≥ 0,
L(x, λ) = a> x + λ(kx − ck22 − r2 ).
The stationarity condition is a + λ(x − c) = 0, thus x = c − a/λ. By primal feasibility, kx − ck2 ≤ r,
we see that we can take λ = kak2 /r, which gives a solution x = c − ra/kak2 . The optimal value in
(67) is therefore a> x = a> c − rkak2 . By the same logic, the optimal value of the convex problem
max a> x subject to kx − ck2 ≤ r
(67)
x∈Rm
is a> c + rkak2 . Now we can read off the optimal value of (20) from those of (66), (67): its optimal
value is
max − a> c − rkak22 , a> c + rkak2 = |a> c| + rkak22 ,
completing the proof.
Proof of Lemma 13. The proof is nearly immediate from the proof of Lemma 12, above. Notice that
the optimal value of (21) is lower bounded by that of (66), which we already know is a> c − rkak22 .
But when the latter is nonnegative, this is also the optimal value of (21). Repeating the argument
with −a in place of a gives the result as stated in the lemma.
D
Proof of Lemma 17
To facilitate the proof, we define the concept of a local maximum among the absolute filter values:
e is be greater than or equal to the
a location i is a local maximum if its absolute filter value |Fi (θ)|
absolute values at neighboring locations, and strictly greater than at least one of these values (where
the boundary points are treated as having just one neighboring location). Specifically, a location i
must satisfy one of the following conditions
e < |Fi (θ)|,
e |Fi+1 (θ)|
e ≤ |Fi (θ)|,
e
|Fi−1 (θ)|
e ≤ |Fi (θ)|,
e |Fi+1 (θ)|
e < |Fi (θ)|,
e
|Fi−1 (θ)|
e < |Fi (θ)|
e
|Fi+1 (θ)|
if i ∈ {bn + 1, . . . , n − bn − 1},
(68)
if i ∈ {bn + 1, . . . , n − bn − 1},
(69)
if i = bn ,
(70)
if i = n − bn .
(71)
e < |Fi (θ)|
e
|Fi−1 (θ)|
e denote the set of local maximums derived from the filter with bandwidth bn , i.e., the set
Let L(θ)
of locations i satisfying one of the four conditions (68)–(71).
e ⊆ IC (θ).
e Fix i ∈ L(θ).
e The boundary cases, i = bn or i = n − bn , are
We first show that L(θ)
e
handled directly by the definition of IC (θ). Hence, we may assume that i ∈ {bn + 1, . . . , n − bn − 1},
and without a loss of generality,
e > |Fi−1 (θ)|
e
e ≥ |Fi+1 (θ)|,
e
|Fi (θ)|
and |Fi (θ)|
41
e > 0. This means that
as well as Fi (θ)
e > |Fi−1 (θ)|
e
e ≥ |Fi+1 (θ)|,
e
Fi (θ)
and Fi (θ)
which of course implies
e ≥ Fi+1 (θ).
e
and Fi (θ)
e > Fi−1 (θ)
e
Fi (θ)
Applying the definition of the filter in (14) gives
i+b
Xn
j=i+1
i+b
Xn
j=i+1
i
X
θej −
−
θej
i+b
n −1
X
j=i
j=i−bn +1
i
X
θej −
θej −
−
θej
i+b
n +1
X
j=i+2
j=i−bn +1
θej −
i−1
X
θej
>0
j=i−bn
i+1
X
θej
≥ 0,
j=i−bn +2
or, after simplification,
θei+bn − 2θei + θei−bn > 0
and
− θei+bn +1 + 2θei+1 − θei−bn +1 ≥ 0.
Adding the above two equations together, we get
− θei+bn +1 − θei+bn + 2 θei+1 − θei − θei−bn +1 − θei−bn > 0,
which implies at least one of the three bracketed pairs of terms must be nonzero, i.e., a changepoint
e ⊆ IC (θ).
e
must occur at one of the locations i, i + bn , or i − bn . The proves that L(θ)
e
Now we show the intended statement. Let j ∈ {bn , . . . , n − bn }, and i ∈ L(θ) be in the direction
e where j ≤ i, without a loss of generality (for the case i < j,
of ascent from j with respect to F (θ),
replace ` + bn below by ` − bn ). That is, the location i is a local maximum where
e ≤ |Fj+1 (θ)|
e ≤ . . . ≤ |Fi−1 (θ)|
e ≤ |Fi (θ)|.
e
|Fj (θ)|
(72)
If |i − j| ≤ bn , then we have the desired result, due to (72). If |i − j| > bn , then there must be at
e such that |` − j| ≤ bn . (To see this, note that if θej−b +1 = . . . = θej+b ,
least one location ` ∈ S(θ)
n
n
e = 0.) Thus, at least one of `, ` + bn lies in between j and i, and then again (72) implies
then Fj (θ)
the result, completing the proof.
E
Proof of Lemma 20
Our optimization problem may be rewritten as
(e
a, eb) = argmin b2 +
a,b∈R
r
X
(ax + b − a1 x)2 + (ax − b − a2 x)2 .
x=1
Taking a derivative of the criterion with respect to b and setting this equal to 0 gives
0=b+
r
X
(ax + b − a1 x − ax + b + a2 x),
x=1
i.e., we see that the optimal value is
Pr
eb = (a1 − a2 )
x
r(r + 1)
= (a1 − a2 )
.
2r + 1
2(2r + 1)
x=1
42
Taking a derivative of the criterion with respect to a and setting this equal to 0 gives
0=b+
r
X
(ax + b − a1 x + ax − b − a2 x),
x=1
i.e., we see that the optimal value is
a1 + a2
.
2
Plugging in e
a, eb into the criterion, and abbreviating cr = r(r + 1)/(2(2r + 1)), we can compute the
optimal criterion value:
e
a=
(a1 −
a2 )2 c2r
+
r
X
a2 − a1
x=1
2
2
x + (a2 − a1 )cr
+
a1 − a2
x − (a1 − a2 )cr
2
2
r
r
X
X
1
x2 + 2r(a1 − a2 )2 c2r + 2(a1 − a2 )2 cr
= (a1 − a2 )2 c2r + (a1 − a2 )2
x
2
x=1
x=1
2
2
r(r + 1)(2r + 1) r3 (r + 1)2
r2 (r + 1)2
2 r (r + 1)
= (a1 − a2 )
+
+
+
4(2r + 1)2
12
2(2r + 1)2
2(2r + 1)
2
3
2
r
r(r
+
1)(2r
+
1)
r
r
(r
+
1)
+
+
+
≥ (a1 − a2 )2
16
12
8
4
1
1
1
≥ (a1 − a2 )2 r3
+ +
6 8 4
13r3
= (a1 − a2 )2
.
24
43
| 10math.ST
|
arXiv:1405.3792v1 [cs.PL] 15 May 2014
Under consideration for publication in Theory and Practice of Logic Programming
1
Minimum Model Semantics for Extensional
Higher-order Logic Programming with Negation∗
Angelos Charalambidis
Department of Informatics & Telecommunications, University of Athens, Greece
(e-mail: [email protected])
Zoltán Ésik
Department of Computer Science, University of Szeged, Hungary
(e-mail: [email protected])
Panos Rondogiannis
Department of Informatics & Telecommunications, University of Athens, Greece
(e-mail: [email protected])
submitted 1 January 2003; revised 1 January 2003; accepted 1 January 2003
Abstract
Extensional higher-order logic programming has been introduced as a generalization of classical
logic programming. An important characteristic of this paradigm is that it preserves all the
well-known properties of traditional logic programming. In this paper we consider the semantics
of negation in the context of the new paradigm. Using some recent results from non-monotonic
fixed-point theory, we demonstrate that every higher-order logic program with negation has a
unique minimum infinite-valued model. In this way we obtain the first purely model-theoretic
semantics for negation in extensional higher-order logic programming. Using our approach, we
resolve an old paradox that was introduced by W. W. Wadge in order to demonstrate the
semantic difficulties of higher-order logic programming.
1 Introduction
Extensional higher-order logic programming has been proposed (Wadge 1991; Charalambidis et al. 2010; Charalambidis et al. 2013) as a generalization of classical logic
programming. The key idea behind this paradigm is that all predicates defined in a
program denote sets and therefore one can use standard extensional set theory in order to
understand their meaning and to reason about them. For example, consider the following
simple extensional higher-order program (Charalambidis et al. 2013) stating that a band
(musical ensemble) is a group that has at least a singer and a guitarist:
band(B):-singer(S),B(S),guitarist(G),B(G).
∗ This research is being supported by the Greek General Secretariat for Research and Technology, the
National Development Agency of Hungary, and the European Commission (European Regional Development Fund) under a Greek-Hungarian intergovernmental programme of Scientific and Technological
collaboration. Project title: “Extensions and Applications of Fixed Point Theory for Non-Monotonic
Formalisms”. It is also supported by grant no. ANN 110883 from the National Foundation of Hungary
for Scientific Research.
2
A. Charalambidis, Z. Ésik and P. Rondogiannis
Suppose that we also have a database of musicians:
singer(sally).
singer(steve).
guitarist(george).
guitarist(grace).
We can then ask the query ?-band(B). Since predicates denote sets, an extensional
higher-order language will return answers such as B = {sally, george} ∪ L, having the
meaning that every set that contains at least sally and george is a potential band.
A consequence of the set-theoretic nature of extensional higher-order logic programming is the fact that its semantics and its proof theory smoothly extend the corresponding
ones for traditional (ie., first-order) logic programming. In particular, every program has
a unique minimum Herbrand model which is the greatest lower bound of all Herbrand
models of the program and the least fixed-point of an immediate consequence operator
associated with the program; moreover, there exists an SLD resolution proof-procedure
which is sound and complete with respect to the minimum model semantics.
One basic property of all the higher-order predicates that can be defined in the language of (Charalambidis et al. 2013) is that they are monotonic. Intuitively, the monotonicity property states that if a predicate is true of a relation R then it is also true of
every superset of R. In the above example, it is clear that if band is true of a relation
B then it is also true of any band that is a superset of B. However, there are many natural higher-order predicates that are non-monotonic. Consider for example a predicate
single singer band which (apparently) defines a band that has a unique singer:
single singer band(B):-band(B),not two singers(B).
two singers(B):-B(S1),B(S2),singer(S1),singer(S2),not(S1=S2).
The predicate single singer band is obviously non-monotonic since it is satisfied by
the set {sally, george} but not by the set {sally, steve, george}. In other words, the
semantics of (Charalambidis et al. 2013) is not applicable to this extended higher-order
language. We are therefore facing the same problem that researchers faced more than
twenty years ago when they attempted to provide a sensible semantics to classical logic
programs with negation; the only difference is that the problem now reappears in a much
more general context, namely in the context of higher-order logic programming.
The solution we adopt is relatively simple to state (but non-trivial to materialize): it
suffices to generalize the well-founded construction (van Gelder et al. 1991; Przymusinski
1989) to higher-order programs. For this purpose, we have found convenient to use a relatively recent logical characterization of the well-founded semantics through an infinitevalued logic (Rondogiannis and Wadge 2005) and also the recent abstract fixed-point
theory for non-monotonic functions developed in (Ésik and Rondogiannis 2013; Ésik and
Rondogiannis 2014). This brings us to the two main contributions of the present paper:
• We provide the first model-theoretic semantics for extensional higher-order logic
programming with negation. In this way we initiate the study of a non-monotonic
formalism that is much broader than classical logic programming with negation.
• We provide further evidence that extensional higher-order logic programming is
a natural generalization of classical logic programming, by showing that all the
well-known properties of the latter also hold for the new paradigm.
Minimum Model Semantics for Extensional Higher-order LP with Negation
3
In the next section we provide an introduction to the proposed semantics for higherorder logic programming and the remaining sections provide the formal development of
this semantics. The proofs of all the results have been moved to corresponding appendices.
2 An Intuitive Overview of the Proposed Semantics
The starting point for the semantics proposed in this paper is the infinite-valued semantics for ordinary logic programs with negation, as introduced in (Rondogiannis and Wadge
2005). In this section we give an intuitive introduction to the infinite-valued approach
and discuss how it can be extended to the higher-order case.
The infinite-valued approach was introduced in order to provide a minimum model
semantics to logic programs with negation. As we are going to see shortly, it is compatible
with the well-founded semantics but it is purely model-theoretic1 . The main idea of this
approach can be explained with a simple example. Consider the program:
p ←
r ← ∼p
s ← ∼q
Under the well-founded semantics both p and s receive the value True. However, p is in
some sense “truer” than s. Namely, p is true because there is a rule which says so, whereas
s is true only because we are never obliged to make q true. In a sense, s is true only by
default. This gave the idea of adding a “default” truth value T1 just below the “real”
truth T0 , and (by symmetry) a weaker false value F1 just above (“not as false as”) the
real false F0 . We can then understand negation-as-failure as combining ordinary negation
with a weakening. Thus ∼ F0 = T1 and ∼ T0 = F1 . Since negations can effectively be
iterated, the infinite-valued approach requires a whole sequence . . . , T3 , T2 , T1 of weaker
and weaker truth values below T0 but above the neutral value 0; and a mirror image
sequence F1 , F2 , F3 , . . . above F0 and below 0. In fact, to capture the well-founded model
in full generality, we need a Tα and a Fα for every countable ordinal α. In other words,
the underlying truth domain of the infinite-valued approach is:
F0 < F1 < · · · < Fω < · · · < Fα < · · · < 0 < · · · < Tα < · · ·< Tω < · · · < T1 < T0
As shown in (Rondogiannis and Wadge 2005), every logic program P with negation has a
unique minimum infinite-valued model MP . Notice that MP is minimum with respect to a
relation ⊑ which compares interpretations in a stage-by-stage manner (see (Rondogiannis
and Wadge 2005) for details). As it is proven in (Rondogiannis and Wadge 2005), if we collapse all the Tα and Fα to True and False respectively, we get the well-founded model. For
the example program above, the minimum model is {(p, T0 ), (q, F0 ), (r, F1 ), (s, T1 )}. This
collapses to {(p, True), (q, False), (r, False), (s, True)}, which is the well-founded model
of the program.
As shown in (Rondogiannis and Wadge 2005), one can compute the minimum infinitevalued model as the least fixed point of an operator TP . It can easily be seen that TP is
not monotonic with respect to the ordering relation ⊑ and therefore one can not obtain
1
In the same way that the equilibrium logic approach of (Pearce 1996) gives a purely logical reconstruction of the stable model semantics.
4
A. Charalambidis, Z. Ésik and P. Rondogiannis
the least fixed point using the classical Knaster-Tarski theorem. However, TP possesses
some form of partial monotonicity. More specifically, as it is shown in (Rondogiannis and
Wadge 2005; Ésik and Rondogiannis 2014), TP is α-monotonic for all countable ordinals
α, a property that guarantees the existence of the least fixed point. Loosely speaking,
the property of TP being α-monotonic means that the operator is monotonic when we
restrict attention to interpretations that are equal for all levels of truth values that are
less than α. In other words, TP is monotonic in stages (but not overall monotonic).
The TP operator is a higher-order function since it takes as argument an interpretation
and returns an interpretation as the result. This observation leads us to the main concept
that helps us extend the infinite-valued semantics to the higher-order case. The key idea
is to demonstrate that the denotation of every expression of predicate type in our higherorder language, is α-monotonic for all ordinals α (see Lemma 5). This property ensures
that the immediate consequence operator of every program is also α-monotonic for all α
(see Lemma 7), and therefore it has a least fixed-point which is a model of the program.
Actually, this same model can also be obtained as the greatest lower bound of all the
Herbrand models of the program (see Theorem 2, the model intersection theorem). In
other words, the semantics of extensional higher-order logic programming with negation
preserves all the familiar properties of classical logic programming and can therefore be
considered as a natural generalization of the latter.
3 Non-Monotonic Fixed Point Theory
The main results of the paper will be obtained using some recent results from nonmonotonic fixed point theory (Ésik and Rondogiannis 2013; Ésik and Rondogiannis 2014).
The key objective of this area of research is to obtain novel fixed point results regarding
functions that are not necessarily monotonic. In particular, the results obtained in (Ésik
and Rondogiannis 2013; Ésik and Rondogiannis 2014) generalize the classical results
of monotonic fixed-point theory (namely Kleene’s theorem and also the Knaster-Tarski
theorem). In this section we provide the necessary material from (Ésik and Rondogiannis
2013; Ésik and Rondogiannis 2014) that will be needed in the next sections.
Suppose that (L, ≤) is a complete lattice in which the least upper bound operation is
W
denoted by
and the least element is denoted by ⊥. Let κ > 0 be a fixed ordinal. We
assume that for each ordinal α < κ, there exists a preordering ⊑α on L. We write x =α y
iff x ⊑α y and y ⊑α x. We define x ❁α y iff x ⊑α y but x =α y does not hold. Moreover,
we write x ❁ y iff x ❁α y for some α < κ. Finally, we define x ⊑ y iff x ❁ y or x = y.
Let x ∈ L and α < κ. We define (x]α = {y : ∀β < α x =β y}.
A key property that will be used throughout the paper is that if the above preordering
relations satisfy certain simple axioms, then the structure (L, ⊑) is a complete lattice;
moreover, every function f : L → L that satisfies some restricted form of monotonicity,
has a least fixed point. These ideas are formalized by the following definitions and results.
Definition 1
Let (L, ≤) be a complete lattice equipped with preorderings ⊑α for all α < κ. Then, L
will be called a basic model if and only if it satisfies the following axioms:
1. For all x, y ∈ L and all α < β < κ, if x ⊑β y then x =α y.
Minimum Model Semantics for Extensional Higher-order LP with Negation
5
2. For all x, y ∈ L, if x =α y for all α < κ then x = y.
F
3. Let x ∈ L and α < κ. Let X ⊆ (x]α . Then, there exists y (denoted by α X)
such that X ⊑α y 2 and for all z ∈ (x]α such that X ⊑α z, it holds y ⊑α z
and y ≤ z.
W
W
4. If xj , yj ∈ L and xj ⊑α yj for all j ∈ J then {xj : j ∈ J} ⊑α {yj : j ∈ J}.
Lemma 1
Let L be a basic model. Then, (L, ⊑) is a complete lattice.
Definition 2
Let A, B be basic models and let α < κ. A function f : A → B is called α-monotonic if
for all x, y ∈ A if x ⊑α y then f (x) ⊑α f (y).
It should be noted that even if a function f is α-monotonic for all α < κ, then it
need not be necessarily monotonic with respect to the relation ⊑ (for a counterexample,
see (Rondogiannis and Wadge 2005, Example 5.7, pages 453–454)). Therefore, the standard tools of classical fixed point theory (such as the Knaster-Tarski theorem), do not
suffice in order to find the least fixed point of f with respect to the relation ⊑.
m
Let us denote by [A → B] the set of functions from A to B that are α-monotonic for
all α < κ.
Theorem 1
m
Let L be a basic model and assume that f ∈ [L → L]. Then, f has a ⊑-least pre-fixed
point, which is also the ⊑-least fixed point of f .
The above theorem will be our main tool for establishing the fact that the immediate
consequence operator of any extensional higher order logic program, always has a least
fixed point, which is a model of the program.
4 The Syntax of the Higher-Order Language H
In this section we introduce the higher-order language H, which extends classical firstorder logic programming to a higher-order setting. The language H is based on a simple
type system that supports two base types: o, the boolean domain, and ι, the domain
of individuals (data objects). The composite types are partitioned into three classes:
functional (assigned to individual constants, individual variables and function symbols),
predicate (assigned to predicate constants and variables) and argument (assigned to
parameters of predicates).
Definition 3
A type can either be functional, predicate, argument, denoted by σ, π and ρ respectively
and defined as:
σ := ι | ι → σ
π := o | ρ → π
ρ := ι | π
We will use τ to denote an arbitrary type (either functional, predicate or argument one).
2
We write X ⊑α y iff forall x ∈ X it holds x ⊑α y.
6
A. Charalambidis, Z. Ésik and P. Rondogiannis
As usual, the binary operator → is right-associative. A functional type that is different
than ι will often be written in the form ιn → ι, n ≥ 1 (which stands for ι → ι → · · · → ι
(n + 1)-times). Moreover, it can be easily seen that every predicate type π can be written
uniquely in the form ρ1 → · · · → ρn → o, n ≥ 0 (for n = 0 we assume that π = o). We
can now proceed to the definition of H, starting from its alphabet and continuing with
expressions and program clauses:
Definition 4
The alphabet of the higher-order language H consists of the following:
1. Predicate variables of every predicate type π (denoted by capital letters such
as P, Q, R, . . .).
2. Predicate constants of every predicate type π (denoted by lowercase letters
such as p, q, r, . . .).
3. Individual variables of type ι (denoted by capital letters such as X, Y, Z, . . .).
4. Individual constants of type ι (denoted by lowercase letters such as a, b, c, . . .).
5. Function symbols of every functional type σ 6= ι (denoted by lowercase letters
such as f, g, h, . . .).
6. The following logical constant symbols: the constants false and true of type o;
the equality constant ≈ of type ι → ι → o; the generalized disjunction and
V
W
conjunction constants π and π of type π → π → π, for every predicate
type π; the generalized inverse implication constants ←π , of type π → π → o,
for every predicate type π; the existential quantifier ∃ρ , of type (ρ → o) → o,
for every argument type ρ; the negation constant ∼ of type o → o.
7. The abstractor λ and the parentheses “(” and “)”.
The set consisting of the predicate variables and the individual variables of H will be
called the set of argument variables of H. Argument variables will be usually denoted by
V and its subscripted versions.
Definition 5
The set of expressions of the higher-order language H is defined as follows:
1. Every predicate variable (respectively, predicate constant) of type π is an
expression of type π; every individual variable (respectively, individual constant) of type ι is an expression of type ι; the propositional constants false
and true are expressions of type o.
2. If f is an n-ary function symbol and E1 , . . . , En are expressions of type ι, then
(f E1 · · · En ) is an expression of type ι.
3. If E1 is an expression of type ρ → π and E2 is an expression of type ρ, then
(E1 E2 ) is an expression of type π.
4. If V is an argument variable of type ρ and E is an expression of type π, then
(λV.E) is an expression of type ρ → π.
W
V
5. If E1 , E2 are expressions of type π, then (E1 π E2 ) and (E1 π E2 ) are expressions of type π.
6. If E is an expression of type o, then (∼E) is an expression of type o.
7. If E1 , E2 are expressions of type ι, then (E1 ≈ E2 ) is an expression of type o.
Minimum Model Semantics for Extensional Higher-order LP with Negation
7
8. If E is an expression of type o and V is a variable of type ρ then (∃ρ V E) is
an expression of type o.
To denote that an expression E has type τ we will write E : τ . The notions of free and
bound variables of an expression are defined as usual. An expression is called closed if it
does not contain any free variables.
Definition 6
A program clause is a clause p ←π E where p is a predicate constant of type π and E is
a closed expression of type π. A program is a finite set of program clauses.
Example 1
The subset predicate can be defined in H as follows:
subset ←π→π→o λP.λQ. ∼∃X((P X)∧ ∼(Q X))
The subset predicate is defined by a λ-expression (which obviates the need to have the
formal parameters of the predicate in the left-hand side of the definition). Moreover, in
the right-hand side we have an explicit existential quantifier for the variable X (in Prolog,
if a variable appears in the body of a clause but not in the head, then it is implicitly
existentially quantified).
5 The Semantics of the Higher-Order Language H
In this section we specify the semantics of H. We start with the semantics of types and
proceed to the semantics of expressions.
The meaning of the boolean type o is equal to a partially ordered set (V, ≤) of truth
values. The number of truth values of V will be specified with respect to an ordinal κ > 0.
All the results of the paper hold for every initial selection of κ. The set (V, ≤) is therefore
F0 < F1 < · · · < Fα < · · · < 0 < · · · < Tα < · · · < T1 < T0
where α < κ.
Definition 7
The order of a truth value is defined as follows: order(Tα ) = α, order(Fα ) = α and
order(0) = +∞.
We can now define the meaning of all the types of our language as well as the corresponding relations ≤ and ⊑α . This is performed in the following definitions:
Definition 8
We define the relation ⊑α on the set V for each α < κ as follows:
1. x ⊑α x if order(x) < α;
2. Fα ⊑α x and x ⊑α Tα if order(x) ≥ α;
3. x ⊑α y if order(x), order(y) > α.
Notice that x =α y iff either x = y or order(x) > α and order(y) > α.
8
A. Charalambidis, Z. Ésik and P. Rondogiannis
Definition 9
Let D be a nonempty set. Then:
•
•
•
•
[[ι]]D = D, and ≤ι is the trivial partial order such that d ≤ι d, for all d ∈ D;
[[ιn → ι]]D = Dn → D. A partial order in this case will not be needed;
[[o]]D = V , and ≤o is the partial order of V ;
[[ι → π]]D = D → [[π]]D , and ≤ι→π is the partial order defined as follows: for all
f, g ∈ [[ι → π]]D , f ≤ι→π g iff f (d) ≤π g(d) for all d ∈ D;
m
• [[π1 → π2 ]]D = [[[π1 ]]D → [[π2 ]]D ], and ≤π1 →π2 is the partial order defined as follows:
for all f, g ∈ [[π1 → π2 ]]D , f ≤π1 →π2 g iff f (d) ≤π2 g(d) for all d ∈ [[π1 ]]D .
The subscripts in the above partial orders will often be omitted when they are obvious
from context.
Definition 10
Let D be a nonempty set and α < κ. Then:
• The relation ⊑α on [[o]]D is the relation ⊑α on V .
• The relation ⊑α on [[ρ → π]]D is defined as follows: f ⊑α g iff f (d) ⊑α g(d) for all
d ∈ [[ρ]]D . Moreover, f ❁α g iff f ⊑α g and f (d) ❁α g(d) for some d ∈ [[ρ]]D .
The following lemma expresses the fact that all the predicate types correspond to
semantic domains that are both complete lattices and basic models:
Lemma 2
Let D be a nonempty set and π be a predicate type. Then, ([[π]]D , ≤π ) is a complete
lattice and a basic model.
We now proceed to formally define the semantics of H:
Definition 11
An intepretation I of H consists of:
1.
2.
3.
4.
a nonempty set D called the domain of I;
an assignment to each individual constant symbol c, of an element I(c) ∈ D;
an assignment to each predicate constant p : π of an element I(p) ∈ [[π]]D ;
an assignment to each function symbol f : ιn → ι of a function I(f) ∈ Dn → D.
Definition 12
Let D be a nonempty set. A state s of H over D is a function that assigns to each
argument variable V of type ρ of H, of an element s(V) ∈ [[ρ]]D .
Definition 13
Let I be an interpretation of H, let D be the domain of I, and let s be a state over D.
Then, the semantics of expressions of H with respect to I and s, is defined as follows:
1.
2.
3.
4.
5.
[[false]]s (I) = F0
[[true]]s (I) = T0
[[c]]s (I) = I(c), for every individual constant c
[[p]]s (I) = I(p), for every predicate constant p
[[V]]s (I) = s(V), for every argument variable V
Minimum Model Semantics for Extensional Higher-order LP with Negation
9
6. [[(f E1 · · · En )]]s (I) = I(f) [[E1 ]]s (I) · · · [[En ]]s (I), for every n-ary function symbol f
7. [[(E1 E2 )]]s (I) = [[E1 ]]s (I)([[E2 ]]s (I))
8. [[(λV.E)]]s (I) = λd.[[E]]s[V/d] (I), where d ranges over [[type(V)]]D
W
W
W
9. [[(E1 π E2 )]]s (I) = π {[[E1 ]]s (I), [[E2 ]]s (I)}, where π is the least upper bound
function on [[π]]D
V
V
V
10. [[(E1 π E2 )]]s (I) = π {[[E1 ]]s (I), [[E2 ]]s (I)}, where π is the greatest lower
bound function
on [[π]]D
Tα+1 if [[E]]s (I) = Fα
11. [[(∼E)]]s (I) = Fα+1 if [[E]]s (I) = Tα
0
if [[E]]s (I) = 0
T0 , if [[E1 ]]s (I) = [[E2 ]]s (I)
12. [[(E1 ≈ E2 )]]s (I) =
F0 , otherwise
W
13. [[(∃V E)]]s (I) = d∈[[type(V)]] [[E]]s[V/d] (I)
D
For closed expressions E we will often write [[E]](I) instead of [[E]]s (I) (since, in this
case, the meaning of E is independent of s).
Lemma 3
Let E : ρ be an expression and let D be a nonempty set. Moreover, let s be a state over
D and let I be an interpretation over D. Then, [[E]]s (I) ∈ [[ρ]]D .
Definition 14
Let P be a program and let M be an interpretation over a nonempty set D. Then M will
be called a model of P iff for all clauses p ←π E of P, it holds [[E]](M ) ≤π M (p), where
M (p) ∈ [[π]]D .
6 Minimum Herbrand Model Semantics for H
In this section we demonstrate that every program of H has a unique minimum Herbrand
model which is the greatest lower bound of all the Herbrand models of the program, and
also the least fixed point of the immediate consequence operator of the program. We
start with the relevant definitions.
Definition 15
Let P be a program. The Herbrand universe UP of P is the set of all terms that can be
formed out of the individual constants3 and the function symbols of P.
Definition 16
A Herbrand interpretation I of a program P is an interpretation such that:
1. the domain of I is the Herbrand universe UP of P;
2. for every individual constant c of P, I(c) = c;
3. for every predicate constant p : π of P, I(p) ∈ [[π]]UP ;
3
As usual, if P has no constants, we assume the existence of an arbitrary one.
10
A. Charalambidis, Z. Ésik and P. Rondogiannis
4. for every n-ary function symbol f of P and for all t1 , . . . , tn ∈ UP , I(f) t1 · · · tn =
f t1 · · · tn .
A Herbrand state of a program P is a state whose underlying domain is UP . We denote
the set of Herbrand interpretations of a program P by IP .
Definition 17
A Herbrand model of a program P is a Herbrand interpretation that is a model of P.
Definition 18
Let P be a program. We define the following partial order on IP : for all I, J ∈ IP , I ≤IP J
iff for every π and for every predicate constant p : π of P, I(p) ≤π J(p).
Definition 19
Let P be a program. We define the following preorder on IP for all α < κ: for all I, J ∈ IP ,
I ⊑α J iff for every π and for every predicate constant p : π of P, I(p) ⊑α J(p).
The following two lemmas play a main role in establishing the two central theorems.
Lemma 4
Let P be a program. Then, IP is a complete lattice and a basic model.
Lemma 5 (α-Monotonicity of Semantics)
Let P be a program and let E : π be an expression. Let I, J be Herbrand interpretations
and s be a Herbrand state of P. For all α < κ, if I ⊑α J then [[E]]s (I) ⊑α [[E]]s (J).
Since by Lemma 4 the set IP is a basic model (and thus by Lemma 1 is a complete
d
lattice with respect to ⊑), every M ⊆ IP has a greatest lower bound M with respect to
⊑. We have the following theorem which generalizes the familiar model intersection theorem for definite first-order logic programs (Lloyd 1987), the model intersection theorem
for normal first-order logic programs (Rondogiannis and Wadge 2005, Theorem 8.6) and
the model intersection theorem for definite higher-order logic programs (Charalambidis
et al. 2013, Theorem 6.8).
Theorem 2 (Model Intersection Theorem)
d
Let P be a program and M be a nonempty set of Herbrand models of P. Then, M is
also a Herbrand model of P.
Definition 20
Let P be a program. The mapping TP : IP → IP is defined for every p : π and for every
W
I ∈ IP as TP (I)(p) = {[[E]](I) : (p ←π E) ∈ P}. The mapping TP will be called the
immediate consequence operator for P.
The following two lemmas are crucial in establishing the least fixed point theorem.
Lemma 6
Let P be a program. For every predicate constant p : π in P and I ∈ IP , TP (I)(p) ∈ [[π]]UP .
Lemma 7
Let P be a program. Then, TP is α-monotonic for all α < κ.
Minimum Model Semantics for Extensional Higher-order LP with Negation
11
Theorem 3 (Least Fixed Point Theorem)
Let P be a program and let M be the set of all its Herbrand models. Then, TP has a
d
least fixed point MP . Moreover, MP = M.
The construction of the least fixed point in the above theorem is similar to the one
given for (potentially infinite) propositional programs in (Rondogiannis and Wadge 2005,
Section 6). Due to space limitations, we provide a short outline of this procedure. In order
to calculate the least fixed point, we start with an interpretation, say I0 , which for every
predicate constant p of type ρ1 → · · · ρn → o, and for all d1 ∈ [[ρ1 ]]UP , . . . , dn ∈ [[ρn ]]UP ,
I0 (p) d1 · · · dn = F0 . We start iterating TP on this interpretation until we get to a point
where the additional iterations do not affect the F0 and T0 values. At this point, we
reset all the remaining values (regarding predicate constants and arguments that have
not stabilized) to F1 , getting an interpretation I1 . We start iterating TP on I1 , until we
get to a point where the additional iterations do not affect the F1 and T1 values. We
repeat this process for higher ordinals. In particular, when we get to a limit ordinal, say
α, we reset all the values that have not stabilized to a truth value of order less than α, to
Fα . The whole process is repeated for κ times. If the value of certain predicate constants
applied to certain arguments has not stabilized after the κ iterations, we assign to them
the intermediate value 0. The resulting interpretation is the least fixed point MP .
7 Resolving a Semantic Paradox of Higher-Order Logic Programming
One deficiency of extensional higher-order logic programming is the inability to define
rules (or facts) that have predicate constants in their heads. The reason of this restriction
is a semantic one and will be explained shortly. However, not all programs that use
predicate constants in the heads of clauses are problematic. For example, the program
computer scientist(john).
good profession(computer scientist).
has a clear declarative reading: the denotation of the computer scientist predicate is
the relation {john}, while the denotation of good profession is the relation {{john}}.
In (Wadge 1991), W. W. Wadge argued that allowing rules to have predicate constants
in their heads, creates tricky semantic problems to. Wadge gave a simple example (duplicated below) that revealed these problems; the example has since been used in other
studies of higher-order logic programming (such as for example in (Bezem 2001)). We
present the example in almost identical phrasing as it initially appeared.
Example 2
Consider the program:
p(a).
q(a).
phi(p).
q(b):-phi(q).
One candidate for minimum Herbrand model is the one in which p and q are true only
of a, and phi is true only of p. However, this means that p and q have the same extension, and so themselves are equal. But since p and q are equal, and phi holds for p, it
must also hold for q. The fourth rule forces us to add q(b), so that the model becomes
12
A. Charalambidis, Z. Ésik and P. Rondogiannis
{p(a), phi(p), q(a), q(b)} (in ad hoc notation). But this is problematic because p and
q are no longer equal and q(b) has lost its justification.
Problems such as the above led Wadge to disallow such clauses from the syntax of
the language proposed in (Wadge 1991). Similarly, the higher-order language introduced
in (Charalambidis et al. 2013) also disallows this kind of clauses.
However, under the semantics presented in this paper, we can now assign a proper
meaning to programs such as the above. Actually, higher order facts such as phi(p).
above, can be seen as syntactic sugar in our fragment. A fact of this form simply states
that phi is true of a relation if this relation is equal to p. This can simply be written as:
phi(P):-equal(P,p).
where equal is a higher-order equality relation that can easily be axiomatized in H using
the subset predicate (see Example 1):
equal ← λP.λQ.(subset P Q) ∧ (subset Q P).
One can compute the minimum model of the resulting program using the techniques
presented in this paper. The paradox of Example 2 is no longer valid since in the minimum
infinite-valued model the atom q(b) has value 0. Intuitively, this means that it is not
possible to decide whether q(b) should be true or false.
The above discussion leads to an easy way of handling rules with predicate constants
in their heads. The predicate constants are replaced with predicate variables and higherorder equality atoms are added in the bodies of clauses. Then, appropriate clauses defining
the equal predicates for all necessary types, are added to the program. The infinite valued
semantics of the resulting program is taken as the meaning of the initial program.
8 Future Work
We have presented the first, to our knowledge, formal semantics for negation in extensional higher-order logic programming. The results we have obtained generalize the
semantics of classical logic programming to the higher order setting. We believe that
the most interesting direction for future work is the investigation of implementation
techniques for (fragments of) H, based on the semantics introduced in this paper. One
possible option would be to examine the implementation of a higher order extension of
Datalog with negation. We are currently examining these possibilities.
References
Bezem, M. 2001. An improved extensionality criterion for higher-order logic programs. In
Proceedings of the 15th International Workshop on Computer Science Logic (CSL). SpringerVerlag, London, UK, 203–216.
Charalambidis, A., Handjopoulos, K., Rondogiannis, P., and Wadge, W. W. 2010. Extensional higher-order logic programming. In JELIA, T. Janhunen and I. Niemelä, Eds.
Lecture Notes in Computer Science, vol. 6341. Springer, 91–103.
Charalambidis, A., Handjopoulos, K., Rondogiannis, P., and Wadge, W. W. 2013. Extensional higher-order logic programming. ACM Transactions on Computational Logic 14, 3,
21:1–21:40.
Minimum Model Semantics for Extensional Higher-order LP with Negation
13
Ésik, Z. and Rondogiannis, P. 2013. A fixed point theorem for non-monotonic functions. In
Proceedings of 13th Panhellenic Logic Symposium, Athens, Greece.
Ésik, Z. and Rondogiannis, P. 2014. A fixed point theorem for non-monotonic functions.
CoRR abs/1402.0299.
Lloyd, J. W. 1987. Foundations of Logic Programming. Springer Verlag.
Pearce, D. 1996. A new logical characterisation of stable models and answer sets. In NMELP,
J. Dix, L. M. Pereira, and T. C. Przymusinski, Eds. Lecture Notes in Computer Science, vol.
1216. Springer, 57–70.
Przymusinski, T. C. 1989. Every logic program has a natural stratification and an iterated
least fixed point model. In PODS, A. Silberschatz, Ed. ACM Press, 11–21.
Rondogiannis, P. and Wadge, W. W. 2005. Minimum model semantics for logic programs
with negation-as-failure. ACM Transactions on Computational Logic 6, 2, 441–467.
van Gelder, A., Ross, K. A., and Schlipf, J. S. 1991. The well-founded semantics for
general logic programs. J. ACM 38, 3, 620–650.
Wadge, W. W. 1991. Higher-order Horn logic programming. In ISLP. 289–303.
Appendix A Proof of Lemma 2
We will make use of certain facts established in (Ésik and Rondogiannis 2014).
F
Suppose that L is a basic model. For each x ∈ L and α < κ, we define x|α = α {x}. It
was shown in (Ésik and Rondogiannis 2014) that x =α x|α and x|α =α x|β , x|α ≤ x|β for
W
all α < β < κ. Moreover, x = α<κ x|α . Also, for all x, y ∈ L and α < κ, it holds x =α y
iff x|α =α y|α iff x|α = y|α , and x ⊑α y iff x|α ⊑α y|α . And if x ⊑α y, then x|α ≤ y|α . It
is also not difficult to prove that for all x ∈ L and α, β < κ, (x|α )|β = x|min{α,β} . More
F
F
generally, whenever X ⊆ (z]α and β ≤ α < κ, it holds ( α X)|β = β X. And if α < β,
F
F
then ( α X)|β = α X. Finally, we will make use of the following two results from (Ésik
and Rondogiannis 2014):
Proposition 8
Let A, B be basic models and let α < κ. If fj : A → B is an α-monotonic function for
W
W
each j ∈ J, then so is f = j∈J fj defined by f (x) = j∈J fj (x).
Lemma 9
Let Z be an arbitrary set and L be a basic model. Then, Z → L is a basic model with
the pointwise definition of the order of relations ≤ and ⊑α for all α < κ.
Suppose that A, B are basic models. By Lemma 9 the set A → B is also a model,
where the relations ≤ and ⊑α , α < κ, are defined in a pointwise way (see (Ésik and
Rondogiannis 2014, Subsection 5.3) for details). It follows that for any set F of functions
W
A → B, F can be computed pointwisely. Also, when F ⊆ (f ]α for some f : A → B,
F
α F for α < κ can be computed pointwisely.
We want to show that whenever f : A → B, β < κ and F ⊆ (f ]β is a set of functions
F
m
m
such that F ⊆ [A → B], then β F ∈ [A → B]. We will make use of a lemma.
Lemma 10
Let L be a basic model. For all x, y ∈ L and α, β < κ with α 6= β, x|β ⊑α y|β iff either
β < α and x|β = y|β (or equivalently, x =β y), or β > α and x|α ⊑α y|α .
14
A. Charalambidis, Z. Ésik and P. Rondogiannis
Proof
Let x|β ⊑α y|β . If β < α then x|β = (x|β )|β = (y|β )|β = y|β . If β > α then x|α =
(x|β )|α ⊑α (y|β )|α = y|α .
Suppose now that β < α and x|β = y|β . Then (x|β )|α = x|β = y|β = (y|β )|α and thus
x|β =α y|β . Finally, let β > α and x|α ⊑α y|α . Then (x|β )|α = x|α ⊑α y|α = (y|β )|α and
thus x|β ⊑α y|β .
Remark 1
Under the above assumptions, if β < α, then x|β ⊑α y|β iff x|β =α y|β iff x|β = y|β .
Corollary 1
F
F
F
F
For all X, Y ⊆ L and α 6= β, β X ⊑α β Y iff β < α and β X = β Y , or β > α and
F
F
αY.
α X ⊑α
Proof
F F
F
F
F
Let x = β X and y = β Y . Then x = β X = β { β X} = x|β and y = y|β .
Let β < α. Then x ⊑α y iff x = y. Let β > α. Then x ⊑α y iff x|α ⊑α y|α . But
F
F F
x|α = α { β X} = α X and similarly for Y .
Lemma 11
Let A and B be basic models. Suppose that f : A → B and F ⊆ (f ]β (where β < κ) is a
F
m
set of functions in [A → B]. Then β F is also α-monotonic for all α < κ.
Proof
F
F
Suppose that α, β < κ and x ⊑α y in A. Then ( β F )(x) = β {f (x) : f ∈ F } and
F
F
( β F )(y) = β {f (y) : f ∈ F }. We have that f (x) ⊑α f (y) for all f ∈ F . Thus, if
F
F
α = β, then clearly ( β F )(x) ⊑α ( β F )(y).
F
F
Suppose that β < α. Then β {f (x) : f ∈ F } = β {f (y) : f ∈ F } since f (x) =β f (y)
F
F
for all f ∈ F . Thus, by Corollary 1, ( β F )(x) ⊑α ( β F )(y).
F
F
Suppose that β > α. Then ( β F )(x) ⊑α ( β F )(y) follows by Corollary 1 from
F
F
α {f (x) : f ∈ F } ⊑α
α {f (y) : f ∈ F }.
m
We equip [A → B] with the order relations ≤ and ⊑α inherited from A → B. We have
the following lemma:
Lemma 12
m
If A and B are basic models, then so is [A → B] with the pointwise definition of the
order of relations ≤ and ⊑α for all α < κ.
Proof
It is proved in (Ésik and Rondogiannis 2014) that the set of functions A → B is a basic
model with the pointwise definition of the relations ≤ and ⊑α , so that for all f, g : A → B
and α < κ, f ≤ g iff f (x) ≤ g(x) for all x ∈ A and f ⊑α g iff f (x) ⊑α g(x) for all x ∈ A. It
W
F
follows that for any F ⊆ B A and α < κ, F and α F can also be computed pointwise:
F
W
W
F
( F )(x) = {f (x) : x ∈ A} and ( α F )(x) = α {f (x) : f ∈ F }. By Proposition 8 and
W
Lemma 11, for all F ⊆ B A , if F is a set of functions α-monotonic for all α, then F and
F
m
β F are also α-monotonic for all α. Since the relations ≤ and ⊑α , α < κ on [A → B]
are the restrictions of the corresponding relations on B A , in view of Proposition 8 and
m
m
Lemma 11, [A → B] also satisfies the axioms in Definition 1, so that [A → B] is a basic
model.
Minimum Model Semantics for Extensional Higher-order LP with Negation
15
The following lemma is shown in (Ésik and Rondogiannis 2014, Subsection 5.2) and
will be used in the proof of the basis case of the next lemma:
Lemma 13
(V, ≤) is a complete lattice and a basic model.
Lemma 2
Let D be a nonempty set and π be a predicate type. Then, ([[π]]D , ≤π ) is a complete
lattice and a basic model.
Proof
Let π be a predicate type. We prove that [[π]]D is a basic model by induction on the
structure of π. When π = o, [[π]]D = V , a basic model. Suppose that π is of the sort
ι → π ′ . Then [[π]]D = D → [[π ′ ]]D , which is a basic model, since [[π ′ ]]D is a model by the
induction hypothesis. Finally, let π be of the sort π1 → π2 . By the induction hypothesis,
m
[[πi ]]D is a model for i = 1, 2. Thus, by Lemma 12, [[π]]D = [[[π1 ]]D → [[π2 ]]D ] is also a
basic model.
Remark 2
Let C denote the category of all basic models and α-monotonic functions. The above
results show that C is cartesian closed, since for all basic models A, B, the evaluation
function eval : (A × B) × A → B is α-monotonic (in both arguments) for all α < κ.
m
Indeed, suppose that f, g ∈ [A → B] and x, y ∈ A with f ⊑α g and x ⊑α y. Then
eval(f, x) = f (x) ⊑α g(x) = eval(g, x) by the pointwise definition of f ⊑α g. Also,
eval(f, x) = f (x) ⊑α f (y) = eval(f, y) since f is α-monotonic.
m
m
m
Since C is cartesian closed, for all f ∈ [B×A → C] there is a unique Λf ∈ [B → [A → C]]
in with f (y, x) = eval(Λf (y), x) for all x ∈ A and y ∈ B.
Appendix B Proofs of Lemmas 3, 4 and 5
Lemma 3
Let E : ρ be an expression and let D be a nonempty set. Moreover, let s be a state over
D and let I be an interpretation over D. Then, [[E]]s (I) ∈ [[ρ]]D .
Proof
If ρ = ι then the claim is clear. Let E be of a predicate type π. We prove simultaneously
the following auxiliary statement. Let α < κ, V : π, x, y ∈ [[π ′ ]]D . If x ⊑α y then
[[E]]s[V/x] (I) ⊑α [[E]]s[V/y] (I). The proof is by structural induction on E. We will cover
only the nontrivial cases.
Case (E1 E2 ): The main statement follows directly by the induction hypothesis of E1
and E2 . There are two cases. Suppose that E1 : π1 → π and E2 : π1 . Then [[E1 ]]s (I) ∈
m
[[π1 → π]]D = [[[π1 ]]D → [[π]]D ] and [[E2 ]]s (I) ∈ [[π1 ]]D by the induction hypothesis. Thus,
[[E1 ]]s (I) ([[E]]s (I)) ∈ [[π]]D . Suppose now that E1 : ι → π and E2 : ι. Then [[E1 ]]s (I) ∈
[[ι → π]]D = D → [[π]]D by the induction hypothesis and [[E2 ]]s (I) ∈ [[ι]]D = D. It follows
again that [[E1 ]]s (I) ([[E]]s (I)) ∈ [[π]]D .
Auxiliary statement: Let x, y ∈ [[π ′ ]]D and assume x ⊑α y. We have by definition
16
A. Charalambidis, Z. Ésik and P. Rondogiannis
[[(E1 E2 )]]s[V/x] (I) = [[E1 ]]s[V/x] (I) ([[E2 ]]s[V/x] (I)), and similarly for [[(E1 E2 )]]s[V/y] (I).
We have E1 : π1 → π and E2 : π1 or E1 : ι → π and E2 : ι. In the first case,
by induction hypothesis [[E1 ]]s[V/x] (I) ∈ [[π1 → π]]D , and thus is α-monotonic. Also,
[[E1 ]]s[V/x] (I) ⊑α [[E1 ]]s[V/y] (I) and [[E2 ]]s[V/x] (I) ⊑α [[E2 ]]s[V/y] (I) by the induction hypothesis. It follows that
[[E1 ]]s[V/x] (I) ([[E2 ]]s[V/x] (I)) ⊑α [[E1 ]]s[V/x] (I) ([[E2 ]]s[V/y] (I)) ⊑α [[E1 ]]s[V/y] (I) ([[E2 ]]s[V/y] (I)).
The second case is similar. We have [[E1 ]]s[V/x] (I) ⊑α [[E1 ]]s[V/y] (I) by the induction hypothesis, moreover, [[E2 ]]s[V/x] (I) = [[E2 ]]s[V/y] (I). Therefore, [[E1 ]]s[V/x] (I) ([[E2 ]]s[V/x] (I)) ⊑α
[[E1 ]]s[V/y] (I) ([[E2 ]]s[V/y] (I)).
Case (λV.E): Assume V : ρ1 and E : π2 . We will show that [[λV.E]]s (I) ∈ [[ρ1 → π2 ]]D . If
ρ1 = ι then the result follows easily from the induction hypothesis of the first statement.
Assume ρ1 = π1 . We show that [[λV.E]]s (I) ∈ [[π1 → π2 ]]D , that is, λd.[[E]]s[V/d] (I) is
α-monotonic for all α < κ. That follows directly by the induction hypothesis of the
auxiliary statement.
Auxiliary statement: It suffices to show that [[(λU.E)]]s[V/x] (I) ⊑α [[(λU.E)]]s[V/y] (I) and
equivalently for every d, [[E]]s[V/x][U/d] (I) ⊑α [[E]]s[V/y][U/d] (I) which follows from induction
hypothesis.
Lemma 4
Let P be a program. Then, IP is a complete lattice and a basic model.
Proof
From Lemma 2 we have that for all predicate types π, [[π]]UP is a complete lattice and a
basic model. It follows, by Lemma 9, that for all predicate types π, Pπ → [[π]]UP is also a
complete lattice and a model, where Pπ is the set of predicate constants of type π. Then,
Q
IP is π Pπ → [[π]]UP which is also a basic model (proved in (Ésik and Rondogiannis
2014)).
Lemma 5 (α-Monotonicity of Semantics)
Let P be a program and let E : π be an expression. Let I, J be Herbrand interpretations
and s be a Herbrand state of P. For all α < κ, if I ⊑α J then [[E]]s (I) ⊑α [[E]]s (J).
Proof
The proof is by structural induction on E.
Induction Base: The cases V, false, true are straightforward since their meanings do not
depend on I. Let I ⊑α J. If E is a predicate constant p then we have I(p) ⊑α J(p).
Induction Step: Assume that the statement holds for expressions E1 and E2 and let
I ⊑α J.
Case (E1 E2 ): It holds [[(E1 E2 )]]s (I) = [[E1 ]]s (I)([[E2 ]]s (I)). By induction hypothesis we
have [[E1 ]]s (I) ⊑α [[E1 ]]s (J) and therefore [[E1 ]]s (I)([[E2 ]]s (I)) ⊑α [[E1 ]]s (J)([[E2 ]]s (I)). We
perform a case analysis on the type of E2 . If E2 is of type ι and since I, J are Herbrand
interpretations, it is clear that [[E2 ]]s (I) = [[E2 ]]s (J) and therefore [[E1 ]]s (I)([[E2 ]]s (I)) ⊑α
[[E1 ]]s (J)([[E2 ]]s (J)). By definition of application we get [[(E1 E2 )]]s (I) ⊑α [[(E1 E2 )]]s (J).
If E2 is of type π then by induction hypothesis we have [[E2 ]]s (I) ⊑α [[E2 ]]s (J) and since
Minimum Model Semantics for Extensional Higher-order LP with Negation
17
[[E1 ]]s (J) is α-monotonic we get that [[E1 ]]s (J)([[E2 ]]s (I)) ⊑α [[E1 ]]s (J)([[E2 ]]s (J)). By transitivity of ⊑α and by the definition of application we conclude that [[(E1 E2 )]]s (I) ⊑α
[[(E1 E2 )]]s (J).
Case (λV.E1 ): It holds by definition that [[(λV.E1 )]]s (I) = λd.[[E1 ]]s[V/d] (I). It suffices
to show that λd.[[E1 ]]s[V/d] (I) ⊑α λd.[[E1 ]]s[V/d] (J) and equivalently that for every d,
[[E1 ]]s[V/d] (I) ⊑α [[E1 ]]s[V/d] (J) which holds by induction hypothesis.
W
W
W
Case (E1 π E2 ): It holds [[(E1 π E2 )]]s (I) = {[[E1 ]]s (I), [[E2 ]]s (I)}. It suffices to show
W
W
that {[[E1 ]]s (I), [[E2 ]]s (I)} ⊑α {[[E1 ]]s (J), [[E2 ]]s (J)} which holds by induction hypothesis and Axiom 4.
V
V
V
Case (E1 π E2 ): It holds [[(E1 π E2 )]]s (I) = {[[E1 ]]s (I), [[E2 ]]s (I)}. Let π = ρ1 → · · · →
V
ρn → o, it suffices to show for all di ∈ [[ρi ]]UP , {[[E1 ]]s (I) d1 · · · dn , [[E2 ]]s (I) d1 · · · dn } ⊑α
V
{[[E1 ]]s (J) d1 · · · dn , [[E2 ]]s (J) d1 · · · dn }. We define xi = [[Ei ]]s (I) d1 · · · dn and yi =
V
[[Ei ]]s (J) d1 · · · dn for i ∈ {1, 2}. We perform a case analysis on v = {x1 , x2 }. If v < Fα
V
V
V
V
or v > Tα then {x1 , x2 } = {y1 , y2 } and thus {x1 , x2 } ⊑α {y1 , y2 }. If v = Fα then
V
V
V
V
Fα ≤ {y1 , y2 } ≤ Tα and therefore {x1 , x2 } ⊑α {y1 , y2 }. If v = Tα then {y1 , y2 } =
V
V
V
Tα and thus {x1 , x2 } ⊑α {y1 , y2 }. If Fα < v < Tα then Fα < {y1 , y2 } ≤ Tα and
V
V
therefore {x1 , x2 } ⊑α {y1 , y2 }.
Case (∼ E1 ): Assume order([[E1 ]]s (I)) = α. Then, by induction hypothesis [[E1 ]]s (I) ⊑α
[[E1 ]]s (J) and thus order([[E1 ]]s (J)) ≥ α. It follows that order([[(∼ E1 )]]s (I)) > α and
order([[(∼ E1 )]]s (J)) > α and therefore [[(∼ E1 )]]s (I) ⊑α [[(∼ E1 )]]s (J).
W
Case (∃V.E1 ): Assume V is of type ρ. It holds [[(∃V.E1 )]]s (I) = d∈[[ρ]] [[E1 ]]s[V/d] (I). It
UP
W
W
suffices to show d∈[[ρ]] [[E1 ]]s[V/d] (I) ⊑α d∈[[ρ]] [[E1 ]]s[V/d] (J) which holds by inducUP
tion hypothesis and Axiom 4.
UP
Appendix C Proof of Theorem 2
We start by providing some necessary background material from (Ésik and Rondogiannis
d
2014) on how the
operation on a set of interpretations is actually defined.
d
d
Let x ∈ V . For every X ⊆ (x]α we define α X as follows: if X = ∅ then α X = Tα ,
otherwise
(V
V
l
X order( X) ≤ α
X=
α
Tα+1 otherwise
Let P be a program, I ∈ IP be a Herbrand interpretation of P and X ⊆ (I]α . For all
predicate constants p in P of type ρ1 → · · · → ρn → o and di ∈ [[ρi ]]UP and for all
d
d
d
i = {1, . . . , n}, it holds α X as ( α X)(p) d1 · · · dn = α {I(p) d1 · · · dn : I ∈ X}.
Let X be a nonempty set of Herbrand interpretations. By Lemma 4 we have that IP is
a complete lattice with respect to ≤ and a basic model. Moreover, by Lemma 1 it follows
that IP is also a complete lattice with respect to ⊑. Thus, there exist the least upper
bound and greatest lower bound of X for both ≤ and ⊑. We denote the greatest lower
V
d
d
bound of X as X and X with respect to relations ≤ and ⊑ respectively. Then, X
can be constructed in an symmetric way to the least upper bound construction described
in (Ésik and Rondogiannis 2014). More specifically, for each ordinal α < κ we define the
d
sets Xα , Yα ⊆ X and xα ∈ IP , which are then used in order to obtain X.
18
A. Charalambidis, Z. Ésik and P. Rondogiannis
d
Let Y0 = X and x0 = 0 Y0 . For every α, with 0 < α < κ we define Xα = {x ∈
d
T
X : ∀β ≤ α x =α xα }, Yα = β<α Xβ ; moreover, xα = α Yα if Yα is nonempty and
V
xα = β<α xβ if Yα is empty.
V
Finally, we define x∞ = α<κ xα . In analogy to the proof of (Ésik and Rondogiannis
d
2014) for the least upper bound it can be shown that x∞ = X with respect to the
relation ⊑. Moreover, it is easy to prove that by construction it holds xα =α xβ and
xβ ≥ xα for all β < α.
Lemma 14
Let P be a program, α < κ and Mα be a Herbrand model of P. Let M ⊆ (Mα ]α be a
d
nonempty set of Herbrand models of P. Then, α M is also a Herbrand model of P.
Proof
d
Assume α M is not a model. Then, there exists a clause p ← E in P and di ∈
d
d
[[ρi ]]D such that [[E]]( α M) d1 · · · dn > ( α M)(p) d1 · · · dn . Since for every N ∈ M
d
d
we have α M ⊑α N , using Lemma 5 we conclude [[E]]( α M) ⊑α [[E]](N ). Let x =
d
d
α {N (p) d1 · · · dn : N ∈ M}.VBy definition, x = ( α M)(p) d1 · · · dn .
If order(x) = α then x = {N (p) d1 · · · dn : N ∈ M}. If x = Tα then for all N ∈ M
d
we have N (p) d1 · · · dn = Tα . Moreover, [[E]]( α M) d1 · · · dn > Tα and by α-monotonicity
we have [[E]](N ) d1 · · · dn > Tα for all N ∈ M. Then, N (p) d1 · · · dn < [[E]](N ) d1 · · · dn
and therefore N is not a model (contradiction). If x = Fα then there exists N ∈ M such
that N (p) d1 · · · dn = Fα and since N is a model we have [[E]](N ) d1 · · · dn ≤ Fα . But
d
d
then, it follows [[E]]( α M) d1 · · · dn ≤ Fα and [[E]]( α M) d1 · · · dn ≤ x (contradiction).
d
If order(x) < α then x = Mα (p) d1 · · · dn . If x = Tβ then [[E]]( α M) d1 · · · dn > Tβ and
[[E]](Mα ) d1 · · · dn > Tβ . Then, we have Mα (p) d1 · · · dn < [[E]](Mα ) and thus Mα is not a
model of P (contradiction). If x = Fβ then [[E]](Mα ) d1 · · · dn ≤ Fβ and by α-monotonicity
d
d
[[E]]( α M) d1 · · · dn ≤ Fβ . Therefore, [[E]]( α M) d1 · · · dn ≤ x (contradiction).
If order(x) > α then x = Tα+1 and there exists model N ∈ M such that N (p) d1 · · · dn <
d
Tα . Moreover, we have [[E]]( α M) d1 · · · dn ≥ Tα and by α-monotonicity we conclude
[[E]](N ) d1 · · · dn ≥ Tα . But then, [[E]](N ) d1 · · · dn > N (p) d1 · · · dn and therefore N is not
a model of P (contradiction).
In the following, we will make use of the following lemma that has been shown in (Ésik
and Rondogiannis 2014, Lemma 3.18):
Lemma 15
If α ≤ κ is an ordinal and (xβ )β<α is a sequence of elements of L such that xβ =β xγ
V
W
and xβ ≤ xγ (xβ ≥ xγ ) whenever β < γ < α, and if x = β<α xβ (x = β<α xβ ), then
xβ =β x holds for all β < α.
Lemma 16
Let (Mα )α<κ be a sequence of Herbrand models of P such that Mα =α Mβ and Mβ ≤ Mα
V
for all α < β < κ. Then, α<κ Mα is also a Herbrand model of P.
Minimum Model Semantics for Extensional Higher-order LP with Negation
19
Proof
V
Let M∞ = α<κ Mα and assume M∞ is not a model of P. Then, there is a clause
p ← E and di ∈ [[ρi ]]D such that [[E]](M∞ ) d1 · · · dn > M∞ (p) d1 · · · dn . We define xα =
Mα (p) d1 · · · dn , x∞ = M∞ (p) d1 · · · dn , yα = [[E]](Mα ) d1 · · · dn and y∞ = [[E]](M∞ ) d1 · · · dn
for all α < κ. It follows from Lemma 15 that M∞ =α Mα and thus x∞ =α xα for all
α < κ. Moreover, using α-monotonicity we also have [[E]](M∞ ) =α [[E]](Mα ) and thus
y∞ =α yα for all α < κ. We distinguish cases based on the value of x∞ .
Assume x∞ = Tδ for some δ < κ. It follows by assumption that y∞ > Tδ . Then, since
x∞ =δ xδ it follows xδ = Tδ . Moreover, since y∞ =δ yδ and order(y∞ ) < δ it follows
yδ = y∞ > Tδ . But then, yδ > xδ (contradiction since Mδ is a model by assumption).
Assume x∞ = Fδ for some δ < κ. Then, since x∞ =δ xδ it follows xδ = Fδ . Then,
since Mδ is a model it follows yδ ≤ xδ and thus yδ ≤ Fδ . But then, since y∞ =δ yδ it
follows yδ = y∞ ≤ Fδ . Therefore, y∞ ≤ x∞ that is a contradiction to our assumption
that y∞ > x∞ .
Assume x∞ = 0. Then, y∞ > x∞ = 0. Let y∞ = Tβ for some β < κ. Then, since
yβ =β y∞ it follows yβ = Tβ . Since Mβ is a model of P it holds Tβ = yβ ≤ xβ , that is
xβ = Tγ for some γ ≤ β. Moreover, since x∞ =β xβ it follows that x∞ = Tγ that is a
contradiction to our assumption that x∞ = 0.
Theorem 2 (Model Intersection Theorem)
d
Let P be a program and M be a nonempty set of Herbrand models of P. Then, M is
also a Herbrand model of P.
Proof
d
We use the construction for M described in the beginning of this appendix. More
d
specifically, we define sets Mα , Yα ⊆ M and Mα ∈ IP . Let Y0 = M and M0 = 0 Y0 .
T
For every α > 0, let Mα = {M ∈ M : ∀β ≤ α M =α Mα } and Yα = β<α Mβ ;
V
d
moreover, Mα = α Yα if Yα is nonempty and Mα = β<α Mβ if Yα is empty. Then,
d
V
M = α<κ Mα . It is easy to see that Mα =α Mβ and Mβ ⊇ Mα for all β < α.
We distinguish two cases. First, consider the case when Yα is nonempty for all α < κ.
d
Then, Mα = α Yα and by Lemma 14 it follows that Mα is a model of P. Moreover, by
V
Lemma 16 we get that M∞ = α<κ Mα is also a model of P.
Consider now the case that there exists a least ordinal δ < κ such that Yδ is empty.
V
It holds (see (Ésik and Rondogiannis 2014)) that M∞ = α<δ Mδ . Suppose M∞ is
not a model of P. Then, there is a clause p ← E, a Herbrand state s and di ∈ [[ρi ]]D
such that [[E]](M∞ ) d1 · · · dn > M∞ (p) d1 · · · dn . We define xα = Mα (p) d1 · · · dn , x∞ =
M∞ (p) d1 · · · dn , yα = [[E]](Mα ) d1 · · · dn , and y∞ = [[E]](M∞ ) d1 · · · dn for all β ≤ α. We
distinguish cases based on the value of x∞ .
Assume x∞ = Tβ for some β < δ. It follows by assumption that y∞ > x∞ = Tβ .
Then, by Lemma 15 it holds that M∞ =β Mβ and we get x∞ =β xβ and therefore
xβ = Tβ . Moreover, by α-monotonicity we get [[E]](M∞ ) d1 · · · dn =β [[E]](Mβ ) d1 · · · dn
and it follows that y∞ =β yβ . Moreover, since y∞ > Tβ it follows yβ = y∞ > Tβ
d
and yβ > xβ . Since Yβ is not empty by assumption we have that Mβ = β Yβ and by
Lemma 14 we get that Mβ is a model of P (contradiction since yβ > xβ ).
Assume x∞ = Fβ for some β < δ. Then, by Lemma 15 it holds M∞ =β Mβ and
therefore x∞ =β xβ . It follows xβ = Fβ . Moreover, since Yβ is nonempty by assumption
20
A. Charalambidis, Z. Ésik and P. Rondogiannis
d
and by Lemma 14 it follows that Mβ = β Yβ is a model of P and thus yβ ≤ xβ = Fβ . By
α-monotonicity we get [[E]](M∞ ) =β [[E]](Mβ ) and therefore y∞ =β yβ ≤ Fβ . It follows
y∞ ≤ Fβ = x∞ (contradiction to the initial assumption y∞ > x∞ ).
Assume x∞ = Tδ . By assumption we have y∞ > x∞ = Tδ . Then, let y∞ = Tγ
for some γ < δ. By Lemma 15 it holds M∞ =γ Mγ and by α-monotonicity it follows
[[E]](M∞ ) =γ [[E]](Mγ ) and thus y∞ =γ yγ . It follows that yγ = Tγ . Moreover, since
d
γ < δ we know by assumption that Yγ is nonempty and therefore Mγ = Yγ and by
Lemma 14 Mγ is a model of P. It follows Tγ = yγ ≤ xγ , that is, xγ = Tβ for some
β ≤ γ < δ. Moreover, since x∞ =γ xγ it follows x∞ = Tβ that is a contradiction (since
by assumption x∞ = Tδ ).
Assume x∞ = Fδ . This case is not possible. Recall that Yα is not empty for all α < δ
d
d
d
and thus Mα = Yα . By the definition of α we observe that either α Yα ≤ Fα or
V
d
α<δ Mα it is not possible to have x∞ = Fδ .
α Yα ≥ Tα+1 . Then, since M∞ =
Assume x∞ = 0. This case does not arise. Again, Yα is not empty for all α < δ and
d
d
thus Mα = α Yα . Moreover, by definition of α , xα 6= 0 for all α < δ. Moreover, since
V
M∞ = α<δ Mα and since δ < κ it follows that the limit can be at most Tδ .
Appendix D Proofs of Lemmas 6, 7 and Theorem 3
Lemma 6
Let P be a program. For every predicate constant p : π in P and I ∈ IP , TP (I)(p) ∈ [[π]]UP .
Proof
It follows from the fact that [[π]]UP is a complete lattice (Lemma 2).
Lemma 7
Let P be a program. Then, TP is α-monotonic for all α < κ.
Proof
Follows directly from Lemma 5 and Proposition 8.
Lemma 17
Let P be a program. Then, M ∈ IP is a model of P if and only if TP (M ) ≤IP M .
Proof
An interpretation I ∈ IP is a model of P iff [[E]](I) ≤π I(p) for all clauses p ←π E in P iff
W
(p←E)∈P [[E]](I) ≤IP I(p) iff TP (I) ≤IP I.
Proposition 18
Let D be a nonempty set, π be a predicate type and x, y ∈ [[π]]D . If x ≤π y and x =β y
for all β < α then x ⊑α y.
Minimum Model Semantics for Extensional Higher-order LP with Negation
21
Proof
The proof is by structural induction on π.
Induction Basis: If x =β y for all β < α then either x = y or order(x), order(y) ≥ α. If
x = y then x ⊑α y. Suppose x 6= y. If order(x), order(y) > α then x =α y. If x = Fα
then clearly x ⊑α y. If x = Tα then Tα ≤ y and therefore y = Tα . The case analysis for
y is similar.
Induction Step: Assume that the statement holds for π. Let f, g ∈ [[ρ → π]]D and α < κ.
For all x ∈ [[ρ]]D and β < α, f (x) ≤ g(x) and f (x) =β g(x). It follows that f (x) ⊑α g(x).
Therefore, f ⊑α g.
Proposition 19
Let P be a program and I, J be Herbrand interpretations of P . If I ≤IP J and I =β J
for all β < α then I ⊑α J.
Proof
Let I, J ∈ IP and α < κ. For all predicate constants p and β < α, I(p) ≤ J(p) and
I(p) =β J(p). It follows by Proposition 18 that I(p) ⊑α J(p) and therefore, I ⊑α J.
Lemma 20
Let P be a program. If M is a model of P then TP (M ) ⊑ M .
Proof
It follows from Lemma 17 that if M is a Herbrand model of P then TP (M ) ≤IP M . If
TP (M ) = M then the statement is immediate. Suppose TP (M ) <IP M and let α denote
the least ordinal such that TP (M ) =α M does not hold. Then, TP (M ) =β M for all
β < α. Since TP (M ) <IP M , by Proposition 19 it follows that TP (M ) ⊑α M . Since
TP (M ) =α M does not hold, it follows that TP (M ) ❁α M . Therefore TP (M ) ⊑ M .
Theorem 3 (Least Fixed Point Theorem)
Let P be a program and let M be the set of all its Herbrand models. Then, TP has a
d
least fixed point MP . Moreover, MP = M.
Proof
It follows from Lemma 7 and Theorem 1 that TP has a least pre-fixed point with respect
to ⊑ that is also a least fixed point. Let MP be that least fixed point of TP , i.e., TP (MP ) =
MP . It is clear from Lemma 17 that MP is a model of P, i.e., MP ∈ M. Then, it follows
d
d
M ⊑ MP . Moreover, from Theorem 2 it is implied that M is a model and thus
d
from Lemma 20, M is a pre-fixed point of TP with respect to ⊑. Since MP is the least
d
d
pre-fixed point of P, MP ⊑ M and thus MP = M.
| 6cs.PL
|
Ref: ACM Genetic and Evolutionary Computation Conference (GECCO), pp. 189—190, Berlin, Germany, July 2017.
A Two-Phase Genetic Algorithm for Image Registration
Sarit Chicotay
Bar-Ilan University
Ramat-Gan, Israel
[email protected]
Eli (Omid) David1
Bar-Ilan University
Ramat-Gan, Israel
[email protected]
ABSTRACT
Image Registration (IR) is the process of aligning two (or more)
images of the same scene taken at different times, different
viewpoints and/or by different sensors. It is an important, crucial
step in various image analysis tasks where multiple data sources are
integrated/fused, in order to extract high-level information.
Registration methods usually assume a relevant transformation
model for a given problem domain. The goal is to search for the
"optimal" instance of the transformation model assumed with
respect to a similarity measure in question.
In this paper we present a novel genetic algorithm (GA)-based
approach for IR. Since GA performs effective search in various
optimization problems, it could prove useful also for IR. Indeed,
various GAs have been proposed for IR. However, most of them
assume certain constraints, which simplify the transformation
model, restrict the search space or make additional preprocessing
requirements. In contrast, we present a generalized GA-based
solution for an almost fully affine transformation model, which
achieves competitive results without such limitations using a twophase method and a multi-objective optimization (MOO) approach.
We present good results for multiple dataset and demonstrate the
robustness of our method in the presence of noisy data.
CCS CONCEPTS
• Computing methodologies~Search methodologies
Computing methodologies~Computer vision
•
KEYWORDS
Computer Vision, Genetic Algorithms, Image Registration, MultiObjective Optimization, Normalized Cross Correlation
1 INTRODUCTION
Image registration (IR) is an important, significant component
in many practical problem domains. Due to the enormous diversity
of IR applications and methodologies, automatic IR remains a
challenge to this day. A broad range of registration techniques has
been developed for various types of datasets and problem domains
[1], where typically, domain-specific knowledge is taken into
account and certain a priori assumptions are made to simplify the
model in question.
1
www.elidavid.com
Nathan Netanyahu is also with the Center for Automation Research,
University of Maryland, College Park, MD 20742.
2
Nathan S. Netanyahu2
Bar-Ilan University
Ramat-Gan, Israel
[email protected]
An affine transformation is one of the most commonly used
models. Since the search space is too large for a feasible exhaustive
search through the entire parameter space, the major challenge is to
avoid getting stuck at a local optimum when there are multiple
extrema in the similarity metric search space.
In order to overcome this problem, we present in this paper a
novel two-phase genetic algorithm (GA)-based approach for IR.
We devise a GA-based framework coupled with image processing
techniques to search efficiently for an optimal transformation with
respect to a given similarity measure. Due to our two-phase strategy
and a unique simultaneous optimization of two similarity measures
based on a multi-objective optimization (MOO) strategy, we obtain
good results over a relatively large search space assuming an almost
fully affine transformation model.
2
TWO-PHASE GA-BASED IR
This section describes briefly our two-phase GA-based
approach to optimize the devised similarity measures by utilizing
common IR tools. For a detailed presentation of this work see [2].
Our IR scheme searches for a transformation that generates a
maximal match in the overlap between the reference image and the
transformed sensed image, thus, the GA chromosome is defined by
six genes reflecting the effects represented by an affine
transformation; translation along the -and- axis, rotation, scale
factor, and shear along the -and- axis.
Two similarity measures are used in this work: (1) Euclidean
distance measure, which is applied to geometric feature points
extracted from both images, and (2) normalized cross correlation
(NCC) [3], which is an intensity-based measure.
The Euclidean distance measure computes the similarity
between two feature point sets,
and , extracted from the
reference and sensed image. We first tested our scheme using
manually selected features and showed that without assuming
correspondences our algorithm gives good registration results. We
then applied the measure to wavelet features obtained in a fullyautomatic mode from Simoncelli's steerable filters [4] based on a
wavelet transform. The Euclidean distance measure is calculated
for the two extracted feature sets as follows:
First, the feature points extracted from the sensed image are
warped according to the transformation assumed. For each warped
point ⃗ we determine its corresponding point ⃗ among the
unassigned reference feature points
⊆ , by finding its nearest
neighbor with respect to Euclidean distance ( ⃗, ⃗) , i.e.,
( )=
⃗
min ( , )
⃗
Finally, the similarity value is the value of the median Euclidean
distance among the correspondences found, i.e.,
( , )=
⃗
⃗,
( ⃗)
The second measure used is the normalized cross correlation
(NCC), which has been commonly applied to evaluate the degree
of similarity between two images. For image and the warped
image it is defined as:
∑ , ( ( , ) − ̅ )( ( , ) − )
( , )=
∑ , ( ( , ) − ̅) ∑ , ( ( , ) − )
where ̅ and are the average gray levels of images and .
Having performed several tests using each of these measures,
independently, as the fitness function, we noticed that the GA fails
to obtain consistent results with a single measure. In an attempt to
obtain a more robust IR scheme, we combined, therefore, the two
measures as part of our two-phase strategy.
2.1 Phase 1: Coarse Estimation
The goal of the first phase is to obtain an initial coarse estimate.
This is achieved using the Euclidean distance measure which is
expected to yield consistent candidate solutions that are "relatively"
close to the optimal solution.
The first phase completes when there is no "significant" update
for a predefined number of changes or when converging to some
predefined minimal distance measure.
2.2 Phase 2: Multi-Objective Optimization
The second phase starts with the population at the end of the
first phase. The Euclidean distance measure is combined with the
NCC measure which makes use of the full image data.
Ideally, we would like to optimize simultaneously the two
objective functions, however, in practice, they may not be
optimized simultaneously. Thus, we use a multi-objective
optimization approach that gives a partial ordering of solutions
based on Pareto dominance.
The second phase completes when there is no "significant"
update. We select among the pareto-optimal set the individual with
the best NCC value as the suggested solution.
(affected by contrast relationships). Table 1 and Figure 1 present
several results in a fully-automatic mode.
We compared also our results on multiple datasets to other
methods assuming a simpler transformation model and performed
additional tests on real images from INRIA database [5] that
underwent affine transformations. See [2] for full details.
4 CONCLUSIONS
In this paper we presented a novel two-phase GA-based image
registration algorithm, whose main advantage over existing
evolutionary IR techniques is that it provides a robust and
automatic solution for a (quasi) fully affine transformation which
is one of the most commonly used models for image registration.
We used the Euclidean distance measure and the NCC measure as
part of a two-phase strategy, supported by a novel MOO design,
which is used to optimize the two similarity measures
simultaneously.
We have tested extensively the proposed scheme, and
demonstrated its robustness to noisy data and consistency on
multiple datasets over successive runs.
Further research should be done to achieve a robust, fullyautomatic registration in more challenging scenarios.
REFERENCES
[1] J. Santamaría, O. Cordón, and S. Damas. A comparative study of state-ofthe art evolutionary image registration methods for 3D modeling, Computer
Vision and Image Understanding, vol. 115(9), pp. 1340–1354, 2011.
[2] http://u.cs.biu.ac.il/~nathan/GAImgReg/Two-Phase_GA_for_IR.pdf
[3] Y.R. Rao, N. Prathapani, and E. Nagabhooshanam. Application of
normalized cross correlation to image registation, International Journal of
Research in Engineering and Technology, vol. 3(5), pp. 12-16, 2014.
[4] A. Karasaridis and E.P. Simoncelli. A filter design technique for steerable
pyramid image transforms. Proceedings of the IEEE Inernational Conference
on Acoustics, Speech and Signal Processing, vol. 4(7-10), pp. 2387–2390,
1996.
[5] https://lear.inrialpes.fr/people/mikolajczyk/Database, INRIA, Mikolajczyk
evaluation database.
Table 1: Fully-automatic registration results of the images in
Figure 1 (RMSE in pixels).
Image
Boat
House
Avg. RMSE
1.37
1.34
σ RMSE
0.2
0.24
3 EMPIRICAL RESULTS
We tested our algorithm on a few dozens of synthetic and real
image datasets, including various satellite and outdoor scenes, in
both a semi-automatic and a fully-automatic mode.
The correctness of the final transformation is evaluated by the
root mean square error (RMSE) for manually selected points. We
consider RMSE value < 1.5 pixels as a good registration.
The semi-automatic mode yields good results in all of the cases
considered. The tests in a fully-automatic mode achieved
successful registration in RMSE terms in about 75% of the test
cases. Some of the failed cases can be recovered, though, if
additional measures/constraints are applied to the transformation's
parameters, e.g., using mutual information (MI) instead of NCC
2
Figure 1: (a), (d) Reference, (b), (e) sensed, and (c), (f)
registered images of "boat" and "house" pairs from [5].
| 1cs.CV
|
THE SYMMETRIC SIGNATURE
arXiv:1602.07184v2 [math.AC] 5 Sep 2016
HOLGER BRENNER AND ALESSIO CAMINATA
Abstract. We define two related invariants for a d-dimensional local ring (R, m, k) called syzygy and
differential symmetric signature by looking at the maximal free splitting of reflexive symmetric powers
of two modules: the top dimensional syzygy module SyzdR (k) of the residue field and the module of
Kähler differentials ΩR/k of R over k. We compute these invariants for two-dimensional ADE singularities
obtaining 1/|G|, where |G| is the order of the acting group, and for cones over elliptic curves obtaining 0 for
the differential symmetric signature. These values coincide with the F-signature of such rings in positive
characteristic.
Introduction
Let (R, m, k) be a Noetherian local ring of prime characteristic p and dimension d which is F-finite,
and with algebraically closed residue field k. For every e ∈ N, let q = pe , and let eR be the R-module
which is equal to R as abelian group and has scalar multiplication twisted via the e-th Frobenius
homomorphism, that is r ◦ s := rq s for r ∈ R and s ∈ eR. We decompose it as
e
R = Raq ⊕ Mq ,
where the module Mq has no free direct summands. The number aq is also called the free rank of eR
and denoted by frkR (eR), while if R is a domain, one has that qd = rankR (eR), the usual rank of an
R-module. Huneke and Leuschke [HL02] defined the F-signature of R as the limit
aq
s(R) := lim d .
e→+∞ q
They proved that the limit defining the F-signature exists assuming that R is Gorenstein and in
other cases. Watanabe and Yoshida [WY04] introduced the minimal relative Hilbert-Kunz multiplicity.
Then, Yao [Yao06] proved it actually coincides with the F-signature. Finally, Tucker [Tuc12] proved
that the F-signature exists for every reduced F-finite Noetherian local ring.
The F-signature is a real number between 0 and 1 and provides delicate information about the
singularities of R. Two principal results in this direction are the fact that s(R) = 1 if and only if the
ring is regular [WY00], and that s(R) > 0 if and only if R is strongly F-regular [AL03].
The main motivation for this paper is given by the following question.
Question. Does there exist an invariant analogous to the F-signature which is independent of the
characteristic and can be used to study singularities also in characteristic zero?
A possible natural attempt to define a characteristic zero version of the F-signature is the following.
If R is a reduced Z-algebra such that Spec R → Spec Z is dominant, then for every prime number p
we consider its reduction mod p, Rp := R ⊗Z (Z/pZ), and compute the F-signature s(Rp ). One may
ask whether the limit
lim s(Rp )
p→+∞
exists, and use this limit to define a characteristic 0 version of the F-signature.
Unfortunately, it is difficult to find an appropriate meaning and compute the previous limit, or
even determine whether it exists at all. If p1 and p2 are distinct prime numbers, the two Frobenius
homomorphisms are difficult to compare. If R is a domain then the rank of eRp1 and the rank of eRp2
are different, the first one being pe1 .
Mathematics Subject Classification (2010): 13A35, 13A50.
Keywords: F-signature, ADE singularities, free resolutions, Kähler differentials.
1
2
HOLGER BRENNER AND ALESSIO CAMINATA
In this paper we define a new invariant called symmetric signature which works in any characteristic.
As inspiration for our construction, we considered two important properties of the modules eR which
we would like to keep. The first fact is a result of Kunz [Kun69] which states that eR is a free module
(for all or for some e ∈ N) if and only if R is a regular ring. Next, if the ring R is Cohen-Macaulay
then eR is a maximal Cohen-Macaulay (MCM for short) module.
Let (R, m, k) be a Noetherian local k-algebra of dimension d. We look for an R-module with the
property that it is free if and only if R is regular. We have at least two modules with this property:
the top-dimensional syzygy module of the residue field SyzdR (k), and the module of (Kähler) differentials ΩR/k
of R over k or its dual.
If R is Cohen-Macaulay, then the module SyzdR (k) is MCM by the depth lemma, on the other hand
ΩR/k is not even reflexive in general. For this reason we prefer to work with its reflexive hull, Ω∗∗
,
R/k
which is called module of Zariski differentials of R over k. The sign (−)∗ stands for the functor HomR (−, R).
We want to study an asymptotic behaviour, so we construct two classes of modules (one for SyzdR (k),
and one for Ω∗∗
) indexed by a natural number q. To do this we introduce the functor of q-th reflexive
R/k
∗∗
q
symmetric powers SymR (−) , that is ordinary q-th symmetric powers of R-modules composed with
the reflexive hull.
The choice of studying symmetric powers is motivated by an unpublished work of Brenner and
Fischbacher-Weitz [BF09]. Symmetric powers of cotangent bundles on projective varieties have
been studied also by other authors such as Bogomolov and De Oliveira[BDO08], and Sakai [Sak78].
Instead, the reflexive hull is motivated by the fact that we want to stay inside the category of reflexive
modules, which coincides with the category of MCM R-modules if R is a normal two-dimensional
domain.
More precisely, we procede as follows. We consider SyzdR (k) (for Ω∗∗
the situation is analogous)
R/k
and we apply to it the q-th reflexive symmetric powers functor for a natural number q. We obtain a
reflexive R-module
∗∗
q
Sq := SymR SyzdR (k)
We decompose the module Sq as
Sq = Raq ⊕ Mq ,
with the module Mq containing no free direct summands, so that aq = frkR Sq the free rank of Sq . For
ease of notation we fix also bq := rankR Sq , and we introduce the following (Definition 1.11).
Definition. The real number
PN
q=0 aq
sσ (R) := lim PN
N→+∞
q=0 bq
is called (syzygy) symmetric signature of R, provided the limit exists.
Replacing the module SyzdR (k) with Ω∗∗
in the previous construction we define also the differential
R/k
symmetric signature of R, which we denote by sdσ (R).
In order to understand the behaviour of these new invariants in comparison with the F-signature
we concentrate on two important classes of examples: two-dimensional Kleinian singularities, and
coordinate rings of plane elliptic curves.
Two-dimensional Kleinian or ADE singularities are the rings of invariants R = SG of a power series
ring in two variables S = k~u, v under the linear action of certain finite subgroups G of SL(2, k), called
Klein groups. The field k is algebraically closed and such that its characteristic does not divide the
order of the acting group G. Watanabe and Yoshida [WY04] proved that if the characteristic of k is a
1
. We prove that the same holds for the symmetric
prime number then the F-signature of R is s(R) = |G|
and the differential symmetric signature in any characteristic (Corollary 3.18), that is
sσ (R) = sdσ (R) =
1
.
|G|
Our main tool to prove this result is the so-called Auslander correspondence (Theorem 2.4). This
states that for a Kleinian singularitiy R = SG there is a functorial one to one correspondence between
THE SYMMETRIC SIGNATURE
3
irreducible k-representations of G and indecomposable maximal Cohen-Macaulay R-modules. The
module corresponding to a k-representation V is given by A(V) := (S ⊗k V)G .
We prove that this construction commutes with q-th reflexive symmetric powers (Theorem 2.12),
that is we have
∗∗
q
q
A Symk (V) SymR (A(V)) .
q
q
Here, SymR (−) denotes the q-th symmetric power of an R-module, and Symk (V) is the q-th symmetric
power of the representation V.
This result is the key to translate the problem of computing the symmetric signature into a problem
in representation theory of finite groups. In fact, it turns out that the free rank of some MCM Rmodule M = A(V) is equal to the multiplicity of the trivial representation into V, and the latter can
be more easily computed using tools from representation theory, such as characters.
The symmetric signature of two-dimensional cyclic quotient singularities k~u, vG , with G cyclic,
is also 1/|G| and is computed by Katthän and the second author in the paper [CK16].
Let R be the coordinate ring of a plane elliptic curve over an algebraically closed field k. Then R is
not strongly F-regular, therefore its F-signature is s(R) = 0. We prove that the differential symmetric
signature of R is sdσ (R) = 0 (Theorem 4.14), provided that char k is not 2 or 3. For the syzygy symmetric
signature we give an upper bound, that is sσ (R) ≤ 21 (Corollary 4.18), provided that the limit exists.
The methods we use in this situation are geometric. We use the correspondence between graded
MCM R-modules and vector bundles over the smooth projective curve Y = Proj R, and we translate
the problem of computing sσ (R) and sdσ (R) into an analogous problem in the category VB(Y) of
vector bundles over Y. The main advantage of this approach is that over an elliptic curve Y the
indecomposable vector bundles have been classified and described by Atiyah [Ati57].
The structure of this paper is the following.
In Section 1 we give a short review of the F-signature and its properties. Then, we define the
symmetric and differential symmetric signatures, together with some other variants: a generalized
version for modules (Definition 1.16), and one for graded rings (Definition 1.17). Some easy properties
and consequences of the symmetric signature are also stated in this section.
In Section 2 we review the theory of the Auslander correspondence and we prove that the Auslander
functor commutes with reflexive symmetric powers (Theorem 2.12).
Section 3 is dedicated to the computation of the symmetric signatures of two-dimensional Kleinian
singularities, and Section 4 to the computation of the symmetric signature for coordinate rings of
plane elliptic curves.
1. F-signature and Symmetric signature
1.1. F-signature. We recall some general facts about rings of prime characteristic, and the definition
and basic properties of the F-signature. Let R be a commutative ring containing a field of prime
characteristic p. With the letter e we always denote a natural number, and with q = pe a power of the
characteristic. The Frobenius homomorphism is the ring homomorphism F : R → R, F(r) = rp , we often
consider also its iterates Fe : R → R, Fe (r) = rq . For any finitely generated R-module M, we denote by
e
M the R-module M, whose multiplicative structure is pulled back via Fe . The scalar multiplication
on eM is given by r · m := rq m, for any r ∈ R, m ∈ eM.
We say that R is F-finite if 1R is a finitely generated R-module, and we say that R is F-split if the
Frobenius map F : R → 1R splits. If (R, m, k) is a complete Noetherian local ring, then R is F-finite if
and only if [k1/p : k] is finite. We will always assume that our rings are F-finite, and that k is perfect,
that is [k1/p : k] = 1. The latter is not a strong restriction, but it is done simply to avoid that numbers
like [k1/p : k] appear in the definitions and in the results.
Remark 1.1. If R is F-finite and Cohen-Macaulay, then eR is a maximal Cohen-Macaulay module for
every e ∈ N. Actually, if x1 , . . . , xn is a maximal R-regular sequence, then it is also an eR-regular
sequence for every e ∈ N.
4
HOLGER BRENNER AND ALESSIO CAMINATA
Let (R, m, k) be a d-dimensional Noetherian reduced local ring of prime characteristic p, which is
F-finite and such that k is perfect. We recall the definition of free rank of an R-module M
frkR (M) := max{n : ∃ a split surjection ϕ : M ։ Rn }.
Since the cancellation property holds for finitely generated modules over a local ring (cf. [LW12,
Corollary 1.16]), frkR (M) is the unique integer such that we have a decomposition
M RfrkR (M) ⊕ N,
and the module N contains no free R-direct summands.
Definition 1.2 (Huneke-Leuschke, [HL02]). The F-signature of R, denoted by s(R) is the limit
frkR (eR)
.
e→+∞
qd
s(R) := lim
Tucker [Tuc12] proved that the F-signature exists for every reduced F-finite Noetherian local ring.
Remark 1.3. If R is a domain of dimension d, then rankR (eR) = qd . So we can write the limit defining
the F-signature as
frkR (eR)
s(R) = lim
.
e→+∞ rankR (eR)
The F-signature is always a real number in the interval [0, 1]. Using a result of Watanabe and
Yoshida [WY00], Huneke and Leuschke [HL02] proved that if R is Cohen-Macaulay then the extreme
value 1 is obtained if and only if the ring is regular. Then, Yao [Yao06] removed the Cohen-Macaulay
assumption.
Theorem 1.4. Let (R, m, k) be a reduced F-finite local ring of prime characteristic such that k is infinite and
perfect. Then s(R) = 1 if and only if R is regular.
Also the value s(R) = 0 has a special meaning in terms of the singularities of the ring, it is equivalent
to the ring being not strongly F-regular, an important notion in tight closure theory.
Theorem 1.5 (Aberbach-Leuschke, [AL03]). Let (R, m, k) be a reduced excellent F-finite local ring of prime
characteristic such that k is perfect. Then s(R) > 0 if and only if R is strongly F-regular.
Remark 1.6. If the ring R has dimension 0 or 1, then it is normal if and only if it is regular if and only
if it is strongly F-regular. So by Theorem 1.4 and Theorem 1.5 we have only two possibilities for the
F-signature of R. We have s(R) = 1 if R is regular, and s(R) = 0 otherwise.
Theorem 1.4 and Theorem 1.5 justify the statement that F-signature measures singularities. Roughly
speaking, the closer the F-signature to 1 is, the nicer the singularity. An important example is the
F-signature of quotient singularities computed by Watanabe and Yoshida.
Theorem 1.7 (Watanabe-Yoshida, [WY04]). Let R = k~x1 , . . . xn G be a quotient singularity over a field k of
prime characteristic p. Assume that the acting group G ⊆ GL(n, k) is a small finite group such that (p, |G|) = 1.
Then the F-signature of R is
1
.
s(R) =
|G|
We conclude with the notion of generalized F-signature, introduced by Hashimoto and Nakajima
[HN15]. Let C be a full subcategory of mod(R), the category of finitely generated R-modules, such
that C has the Krull-Remak-Schmidt property (KRS property for short) and eR ∈ C for every e ∈ N.
For example, if R is complete we may choose C = mod(R). Let M be an indecomposable object in C,
for each e ∈ N and q = pe we can consider the multiplicity aq of M inside eR. In other words, we can
write eR = Maq ⊕ Nq , with the module Nq containing no copies of M as direct summands.
Definition 1.8 (Hashimoto-Nakajima, [HN15]). The generalized F-signature of R with respect to M is
aq
s(R, M) := lim d ,
e→+∞ q
provided the limit exists.
THE SYMMETRIC SIGNATURE
5
Remark 1.9. If the ring R has finite F-representation type, then the generalized F-signature exists.
This is a consequence of [SVB96, Proposition 3.3.1] and of [Yao05, Theorem 3.11]. Moreover, Seibert
[Sei97] proved that in this situation the F-signature and the Hilbert-Kunz multiplicity of R are rational
numbers.
Theorem 1.10 (Hashimoto-Nakajima, [HN15]). Let R = k~x1 , . . . xn G be a quotient singularity over an
algebraically closed field k of prime characteristic p. Assume that the acting group G ⊆ GL(n, k) is a small finite
group such that that (p, |G|) = 1. Let Vt be an irreducible k-representation of G and let Mt = A(Vt ) = (S⊗k Vt )G
be the corresponding R-module via Auslander functor. Then the generalized F-signature of R with respect to
Mt is
rankR Mt
s(R, Mt ) =
.
|G|
The definition of the Auslander functor A and the relation between k-representations of G and
reflexive modules over the invariant ring R will be reviewed in Section 2.
1.2. Symmetric signature. For this section, let (R, m, k) be a local Noetherian k-domain of dimension
d (of any characteristic) and let q be a natural number. We consider the top-dimensional syzygy
module SyzdR (k) of the residue field, coming from a minimal free resolution of k, and the module of
(Kähler) differentials of R over k, denoted by ΩR/k .
It is an easy consequence of the Auslander-Buchsbaum-Serre theorem, and of Auslander-Buchsbaum
formula that SyzdR (k) is R-free if and only if R is regular. Similarly, under some mild conditions we
have that ΩR/k is R-free if and only if R is regular (see e.g. [Har77, Chapter II, Theorem 8.8]). Moreover
from the depth lemma follows that if R is Cohen-Macaulay then SyzdR (k) is maximal Cohen-Macaulay,
and in particular reflexive. On the other hand ΩR/k is not even reflexive in general (see e.g. [Her78a]),
so it is convenient to work with its reflexive hull Ω∗∗
, which is called module of Zariski (or regular)
R/k
differentials of R over k.
Let q be a natural number, we define the following two classes of R-modules.
∗∗
q
Sq : = SymR SyzdR (k)
,
∗∗
∗∗
q
q
Cq : = SymR Ω∗∗
=
Sym
.
Ω
R/k
R/k
R
Definition 1.11. The real numbers
PN
q=0
sσ (R) := lim PN
N→+∞
and
q=0
PN
frkR Sq
rankR Sq
q=0
sdσ (R) := lim PN
N→+∞
q=0
,
frkR Cq
rankR Cq
are called (syzygy) symmetric signature of R, and differential symmetric signature of R respectively,
provided the limits exist.
In the rest of the paper we will study both the symmetric signature and the differential symmetric
signature. For the clarity of exposition we will focus sometimes on one definition and we will say
what applies also to the other.
We don’t know whether the limit defining the symmetric signature always exists. To obtain an
invariant which is for sure well defined and exists, one may replace the limits in Definition 1.11 with
limits inferior.
Moreover one may ask why should we consider the limits of Definition 1.11, insted of the simpler
limit
frkR Sq
.
(1)
lim
q→+∞ rankR Sq
The main reason is that the limit (1) does not exist even in simple cases, as we see in Example 3.19.
However, when the simpler limit (1) exists, then also the symmetric signature exists and they coincide.
6
HOLGER BRENNER AND ALESSIO CAMINATA
Lemma 1.12. LetP(an )n∈N , (bn )n∈N be sequences of real numbers such that bn > 0 for all n ∈ N. Assume that
the infinite series n∈N bn diverges. If the limit limn→+∞ abnn exists, then also the limit
Pn
ak
lim Pnk=0
n→+∞
k=0 bk
exists and the two limits coincide.
Remark 1.13. The modules Sq and Cq are reflexive for every q ≥ 0. If R is Cohen-Macaulay of
dimension ≤ 2, then Sq and Cq are also MCM.
Example 1.14. If R is a regular ring of dimension d, then SyzdR (k) and Ω∗∗
are free modules. Then Sq
R/k
q
q
q
q
q
and C are also free, therefore frkR S = rankR S and frkR C = rankR C for every q. It follows that
the symmetric signatures are sσ (R) = 1 and sdσ (R) = 1. In particular, this happens if R is of dimension
0, since it is forced to be a field, being a domain.
Remark 1.15. If (R, m, k) has dimension 1, then the first syzygy of the residue field Syz1R (k) is just the
maximal ideal m. Since m is an ideal it has rank 1, therefore also all its symmetric
powershave rank 1,
∗∗
q
and the same holds for the reflexive hull. In other words, we have rankR SymR (Syz1R (k) = 1 for all
q ∈ N. Therefore, for each q we have only two possibilities: either Sq R and frkR Sq = rankR Sq = 1
or Sq is not free and frkR Sq = 0.
Now assume in addition that the category Ref(R) of finitely generated reflexive modules has the
KRS property. We fix an indecomposable object M in Ref(R), then for every q ∈ N we have a unique
decomposition
Sq = Maq ⊕ Nq ,
where the module Nq contains no copy of M as direct summand. The natural number aq is called the
multiplicity of M in Sq .
Definition 1.16. The number
sσ (R, M) := lim PN
N→+∞
PN
q=0
q=0 aq
rankR Sq
is called generalized symmetric signature of R with respect to M, provided the limit exists.
We can define the symmetric signature also for N-graded rings.
Let R be a standard graded Noetherian k-domain of dimension d. We work in the category modZ (R)
of finitely generated graded R-modules. In this category we consider the graded module SyzdR (k)
coming from a minimal graded free resolution of k as R-module, and the differential module ΩR/k ,
which is also naturally graded. Their q-th reflexive symmetric powers Sq and Cq are again graded
modules. We also introduce a graded version of the free rank. For any finitely generated graded
R-module M, we define the graded free rank of M as
gr
frkR (M) := max{n : ∃ a homogeneous of degree 0 split surjection ϕ : M ։ F,
with F free graded R-module of rank n}.
The main reason to introduce this definition is that we want work in the Krull-Schmidt category
modZ (R) of finitely generated graded R-modules, whose maps are R-linear homomorphisms of
degree 0. Thus, the previous definition is more natural in this setting.
Definition 1.17. The graded (syzygy) symmetric signature of R is defined by the limit
PN
gr q
q=0 frkR S
,
sσ (R) := lim PN
q
N→+∞
q=0 rankR S
provided it exists.
Replacing the module Sq with Cq in the definition above, we define also the graded version of the
differential symmetric signature.
THE SYMMETRIC SIGNATURE
7
Remark 1.18. Notice that in the definition of sσ (R) for graded rings we consider the graded free rank
instead of the free rank as in the local setting. Therefore some differences may occur. For example,
one should remind that all direct summands of the form R(a) for some integer a contribute to the
graded free rank part of the module.
2. Auslander correspondence and reflexive symmetric powers
We fix the setting for this section. Let S = k~u, v be a power series ring in two variables over
an algebraically closed field k. Let G ⊆ GL(2, k) be a finite group which is small, i.e. it contains
no pseudo-reflections, and such that the characteristic of k and the order |G| of G are coprime. The
group G acts on S via linear changes of variables, and we denote by R := SG the invariant ring under
this action. The invariant ring R is called quotient singularity and has the following properties: it
is a Noetherian local domain of dimension 2, normal, Cohen-Macaulay, complete, and an isolated
singularity (see e.g. [BD08, Theorem 4.1]). Moreover, Watanabe [Wat74a, Wat74b] proved that R is
Gorenstein if and only if G ⊆ SL(2, k). In this case R is called special quotient singularity. If G is a cyclic
group R is called cyclic quotient singularity. The Kleinian singularities, or ADE singularities are special
quotient singularities, which actually coincides with them if the base field has characteristic zero.
In Section 3 we will compute the symmetric signature of Kleinian singularities. In this section
we will review the theory of Auslander correspondence between linear k-representations of G and
MCM R-modules, and we will prove that the Auslander functor commutes with reflexive symmetric
powers. For a more detailed survey of these facts the reader may consult the books of Yoshino [Yos90],
and of Leuschke and Wiegand [LW12], or the second author’s Ph.D. thesis [Cam16].
2.1. The Auslander correspondence. We denote by k[G] the group ring of G over k, that is the kvector space with basis {e g : g ∈ G}, and with multiplication given on the basis elements by group
multiplication e g eh = e gh and extended linearly to arbitrary elements. The category mod(k[G]) of
finitely generated left modules over the group ring can be identified with the category of linear
k-representations of G. For this reason we will use indifferently the terminology of representation
theory or the terminology of module theory over k[G].
The skew
M group ring of G and S is denoted by S ∗ G and it is a free S-module on the elements of G
Sg, with a multiplicative structure given by (sg)(th) := sg(t) · gh for all s, t ∈ S, g, h ∈ G.
S∗G=
g∈G
The group ring and the skew group ring are in general non-commutative, so by k[G]-module and
S ∗ G-module we will always mean left module. We point out that an S ∗ G-module is just an S-module
with a compatible action of G: g(sm) = g(s)g(m) for all g ∈ G, s ∈ S and m ∈ M. The ring S is clearly
an S ∗ G-module, but not every S-module has a natural S ∗ G-module structure.
There is a natural functor F between the category mod(k[G]) and the category Proj(S ∗ G) of finitely
generated projective S ∗ G-modules, namely F (V) := S ⊗k V for every object V in mod(k[G]). The
functor F has a right adjoint given by F ′ (P) := P ⊗S k. The functors F and F ′ are an adjoint pair
[Yos90, Lemma 10.1], but in general they are not an equivalence of categories.
We remark that an S ∗ G-module is S ∗ G-projective if and only if it is S-projective, hence S-free,
since S is local.
We define two other functors G : Proj(S ∗ G) → Ref(R), G(M) = MG , and G′ : Ref(R) → Proj(S ∗ G),
G (N) = (S ⊗R N)∗∗ , where (−)∗ := HomS (−, S), and Ref(R) denotes the category of finitely generated
reflexive R-modules. Since R has dimension 2 the category Ref(R) coincides with the category
MCM(R) of maximal Cohen-Macaulay R-modules, and also with the category AddR (S) of R-direct
summands of S by a result of Herzog [Her78b], that is Ref(R) = MCM(R) = AddR (S).
′
Auslander [Aus86b] proved that the functors G and G′ are an equivalence of categories. We will
give a geometric proof of this fact, based on the following ring version of the Speiser’s Lemma of
Galois theory due to Auslander and Goldman [AG60], and Chase, Harrison, and Rosenberg [CHR65].
Lemma 2.1 (Speiser’s Lemma). Let A be a commutative domain, G a finite group of ring automorphisms of
A and denote by B = AG the invariant ring. Then the following facts are equivalent.
1) A is a separable B-algebra.
8
HOLGER BRENNER AND ALESSIO CAMINATA
2) For every g , idG in G, and every maximal ideal p in A, there exists a ∈ A such that a − g(a) < p.
3) The map ϕ : A ⊗B MG → M given by ϕ(a ⊗ m) = am is an isomorphism for every A ∗ G-module M.
Lemma 2.2. Let S = k~u, v, let G ⊆ GL(2, k) be a finite small group such that (char k, |G|) = 1, and let
R = SG . Let M be a projective S ∗ G-module and consider the canonical map
δ : S ⊗R MG → M,
given by δ(s ⊗ m) = sm. Then the following facts hold.
∼
e on the punctured spectrum U′ = Spec S\{m},
1) δ induces an isomorphism of coherent sheaves S ⊗R MG → M
where m = (u, v).
2) δ induces an isomorphism of S ∗ G-modules
(S ⊗R MG )∗∗ M.
Proof. The ring S is a Cohen-Macaulay isolated singularity, so coherent sheaves associated to MCM
S-modules are locally free on the punctured spectrum. Moreover we recall that projective S-modules
are MCM in this case. Therefore it follows that property 1) implies property 2).
p
We prove 1). Let f1 , . . . , fµ be elements in R which generate m up to radical, that is ( f1 , . . . , fµ ) = m
as ideals in S. These
elements exist, since the ring extension R ֒→ S is finite. Thus, we have an open
S
covering U′ = i D( fi ).
We claim that for every f = fi the induced map
δ f : S f ⊗R f MGf → M f
is an isomorphism. We prove the claim, then the Lemma will follow from it. In fact, because we
have a global homomorphism, we get a sheaf homomorphism defined on U′ which is locally an
isomorphism, so it is forced to be an isomorphism on U′ .
To prove the claim, let f be one of the fi ’s. We check that condition 2) of Lemma 2.1 is true for
A = S f and B = R f . Let p be a maximal ideal of A. If a − g(a) ∈ p for every a ∈ A and every g ∈ G, then
we have a ∈ p if and only if g(a) ∈ p, which is equivalent to say that p is a fix point for the action of G
on A. On the other hand, since G is small, its action on U′ is fixpoint-free, so we get a contradiction.
Therefore δ f is an isomorphism by Lemma 2.1 above.
Theorem 2.3. The functors G : Proj(S ∗ G) → Ref(R) and G′ : Ref(R) → Proj(S ∗ G) give an equivalence of
categories
Proj(S ∗ G) Ref(R).
Proof. See [Yos90, Proposition 10.9].
We compose the functors F and G and we obtain a functor A : mod(k[G]) → Ref(R), A(V) :=
(S ⊗k V)G called Auslander functor. The Auslander functor has a right adjoint given by A′ (N) :=
(S ⊗R N)∗∗ ⊗S k, for every reflexive R-module N. These functors give a one-one correspondence
between k-representations of G and maximal Cohen-Macaulay modules over R called Auslander
correspondence. We collect the properties of the functor A and of the Auslander correspondence in
the following theorem and corollary.
Theorem 2.4 (Auslander correspondence). The functors A : mod(k[G]) → Ref(R) and A′ : Ref(R) →
mod(k[G]) have the following properties.
1)
2)
3)
4)
5)
6)
A(V) A(W) if and only if V W.
A(V) is indecomposable in Ref(R) if and only V is an irreducible representation.
A(Homk (V, W)) HomR (A(V), A(W)) for every V, W ∈ mod(k[G]).
A(V ⊗k W) (A(V) ⊗R A(W))∗∗ for every V, W ∈ mod(k[G]).
If V0 is the trivial representation then A(V0 ) = R.
rankR A(V) = dimk V for every k[G]-module V.
THE SYMMETRIC SIGNATURE
9
Corollary 2.5. Let V1 , . . . , Vr be a complete set of non-isomorphic irreducible k-representations of G, and
fix Ni = A(Vi ) = (S ⊗k Vi )G . Then N1 , . . . , Nr is a complete set of non-isomorphic indecomposable MCM
R-modules. Moreover let V be a k-representation which decomposes as
V = V1n1 ⊕ · · · ⊕ Vrnr ,
for some natural numbers ni . Then the MCM R-module N := A(V) = (S ⊗k V)G decomposes as
N = N1n1 ⊕ · · · ⊕ Nrnr .
Example 2.6. Let Cn =< g > be the cyclic group of order n, with irreducible representations (Vt , ρt )
over an algebraically closed field k given by ρt (g) = ξt , for a fixed primitive n-th root of unity ξ ∈ k.
Assume that char k does not divide |G|. We can embed G into GL(2, k) via the representation V1 ⊕ Va ,
where a is a natural number such that (a, n) = 1, otherwise the representation is not faithful. In other
words, we consider the cyclic group generated by
!
ξ 0
.
0 ξa
This group acts linearly on S = k~u, v and the invariant subring R is generated by monomials ui v j
such that i + aj 0 mod n.
For each irreducible representation Vt, we have an indecomposable MCM R-module Mt = (S⊗k Vt )G .
A straightforward computation shows that this is given by
Mt = R ui v j : i + aj −t mod n .
2.2. Auslander functor and reflexive symmetric powers. We want to prove that the Auslander
functor commutes with q-th reflexive symmetric powers Symq (−)∗∗, that is
q
q
A(Symk (V)) (SymR (A(V)))∗∗
for every k[G]-module V. Since the Auslander functor is the composition of the functors F and G we
will split the proof of this fact in two propositions.
Remark 2.7. Observe that in the categories mod(k[G]) and Proj(S ∗ G) the reflexive symmetric powers
coincide with the usual symmetric powers. In fact, every finitely generated k[G]-module V is reflexive,
so it is canonically isomorphic to its double dual V ∗∗ . Since S is regular, for every finitely generated
q
projective S ∗ G-module N the symmetric powers SymS (N) are S-free, hence reflexive.
Proposition 2.8. For every k[G]-module V we have an isomorphism of S ∗ G-modules
q
q
SymS (S ⊗k V) S ⊗k Symk (V).
q
q
Proof. From [Nor84, (6.5.1)] we have an isomorphism of S-modules ψ : SymS (S ⊗k V) → S ⊗k Symk (V)
given by ψ((a1 ⊗ v1 ) ◦ · · · ◦ (aq ⊗ vq )) = (a1 · · · aq ) ⊗ (v1 ◦ · · · ◦ vq ). The only thing that we have to check is
that ψ is compatible with the action of G, and this is shown by the following commutative diagram
ψ
(a1 ⊗ v1 ) ◦ · · · ◦ (aq ⊗ vq )
g
(a1 · · · aq ) ⊗ (v1 ◦ · · · ◦ vq )
g
g(a1 ) ⊗ g(v1 ) ◦ · · · ◦ g(aq ) ⊗ g(vq )
ψ
g(a1 . . . aq ) ⊗ g(v1 ◦ · · · ◦ vq ) =
g(a1 ) · · · g(aq ) ⊗ g(v1 ) ◦ · · · ◦ g(vq ) .
We will use often the following well-known lemma. We write it here for ease of reference.
Lemma 2.9. Let (R, m, k) be a normal two-dimensional local domain, and let U = Spec R\{m} be the punctured
spectrum of R. For a torsion-free finitely generated R-module M we have an isomorphism
e
M∗∗ Γ(U, M).
10
HOLGER BRENNER AND ALESSIO CAMINATA
Remark 2.10. Let S = k~u, v and R = SG be as in the rest of the section. Consider the following
commutative diagram
U′ −−−−−→ Spec S
π
π
y
y
U −−−−−→ Spec R
where the map π is induced by the inclusion R ֒→ S, U is the punctured spectrum of R, and
U′ = π−1 (U) is the pull-back of U to Spec S, which is actually the punctured spectrum of S. For every
R-module M this gives us the following identification of sheaves on U′
e U) ∼
π∗ (M|
S ⊗R M
(2)
U′
.
Proposition 2.11. Let N be a projective S ∗ G-module, then we have an isomorphism of R-modules
∗∗
q
q
(SymS (N))G SymR (NG ) .
Proof. Let G′ (−) = (S ⊗R −)∗∗ be the right adjoint of the G-invariants functor G(−) = (−)G . Since G and
G′ are an equivalence of categories (Theorem 2.3), it is enough to show that
∗∗
q
q
G′ (SymS (N))G G′ SymR (NG )
,
that is
∗∗ ∗∗
G ∗∗
q
q
,
S ⊗R SymR (NG )
S ⊗R SymS (N)
(3)
as S ∗ G-modules.
Notice that the two double duals on the right hand side of (3) are different: the first one is the
double dual in mod(R) and the second is the double dual in mod(S).
In order to prove (3), we consider the left hand side
and the right hand
side separately. From the
G ∗∗
q
q
SymS (N).
second part of Lemma 2.2, the left hand side of (3) is S ⊗R SymS (N)
For the S-module on the right hand side of (3) we use Lemma 2.9, and we interpret it as the
evaluation of the sheaf
∗∗
q
S ⊗R SymR (NG )
∼
U′
′
on the punctured spectrum U of S. From the commutative diagram of Remark 2.10 we get the
isomorphism
!
!
∗∗
q
q
q
= π∗ SymR (NG ) ,
S ⊗R SymR (NG )
π∗ SymR (NG )∗∗
∼
∼
∼
U
U
U′
where we can remove the double dual over U, thanks to Lemma 2.9. Since taking symmetric powers
commute with sheafification and with the restriction map of sheaves we get
!
q
q f
q f
∗
G
G)
G|
π SymR (N )
π∗ SymX (N
π∗ SymX N
U ,
∼
U
U
where the second and the third symmetric powers are sheaf symmetric powers taken over X = Spec R.
We set Y = Spec S. Since sheaf symmetric powers commute with pullback we have
q
q f
fG |U .
G|
SymY π∗ N
π∗ SymX N
U
We apply again (2) and Lemma 2.9 to obtain
q
SymY
∼
fG |U Symq S ⊗R NG
π∗ N
Y
U′
!
=
q
SymY
∼
G ∗∗
(S ⊗R N )
U′
!
.
THE SYMMETRIC SIGNATURE
11
Since taking symmetric powers commutes with sheafification and with the restriction map of sheaves
we get
!
∗∗
q
q
q
G ∗∗
SymS S ⊗R NG
SymY (S ⊗R N )
SymS (N) ,
∼
∼
U′
∼
U′
U′
where the last isomorphism follows from Lemma 2.2. Taking global sections on U′ we obtain that
q
the right hand side of (3) is also isomorphic to SymS (N). Therefore we have an isomorphism of
S-modules as in (3). This is actually an isomorphism of S ∗ G-modules. In fact, the natural map of
G
G
q
q
q
q
R-modules SymR (NG ) → SymS (N) induces a natural map S ⊗R SymR (NG ) → S ⊗R SymS (N) ,
which is G-compatible since the action is just on S.
From Proposition 2.8 and Proposition 2.11 we immediately obtain the following.
Theorem 2.12. Let V be a k[G]-module, and let M = (S ⊗k V)G be the corresponding MCM R-module via
Auslander functor. Then we have
G
q
q
SymR (M)∗∗ S ⊗k Symk (V) .
q
q
In other words SymR (A(V))∗∗ A(Symk (V)).
3. Symmetric signature of Kleinian singularities
In this section we will compute the generalized symmetric signature (and differential symmetric
signature) of two-dimensional Kleinian singularities. These are invariant rings R := k~u, vG , where
k is an algebraically closed field, and G is one of the following finite small subgroups of SL(2, k): the
cyclic group Cn of order n, the binary dihedral group BDn of order 4n, the binary tetrahedral group
BT of order 24, the binary octahedral group BO of order 48, and the binary icosahedral group BI of
order 120. We always assume that the characteristic of k does not divide the order of the group G.
It is an old result of Klein that if k has characteristic zero, then every finite subgroup of SL(2, k) is
isomorphic to one of these groups. For this reason these groups are also called Klein groups.
The Kleinian singularities are hypersurface rings, that is they are isomorphic to a quotient ring
k~x, y, z/( f ) for some polynomial f , according to the following table.
G
singularity name
An−1
Dn+2
E6
E7
E8
cyclic
binary dihedral
binary tetrahedral
binary octahedral
binary icosahedral
|G|
n
4n
24
48
120
f
n
y − xz
x2 + yn+1 + yz2
x 2 + y 3 + z4
x2 + y3 + yz3
x 2 + y 3 + z5
3.1. The fundamental module of Kleinian singularities. Let (R, m, k) be a complete two-dimensional
normal non-regular domain with canonical module KR . In this setting Auslander [Aus86b] proved
that in the category Ref(R) there exists a unique non-split short exact sequence of the form
0 → KR → E → m → 0,
which is called the fundamental sequence of R.
The module E appearing in the middle term is also unique up to isomorphism and is called the
fundamental module or Auslander module of R (cf. [Yos90, Chapter 11]). It isa reflexive (hence MCM)
V2 ∗∗
module of rank 2. Moreover we have an isomorphism of R-modules
E KR .
The following example (cf. [Yos90, Example 11.8]) clarifies the name fundamental module.
Example 3.1. Let V be a k-vector space of dimension 2 with basis u, v and let G ⊆ GL(2, k) be a finite
subgroup. Then, G acts on the power series ring S = k~u, v and we consider the invariant ring
R = SG , which is a two-dimensional quotient singularity. The fundamental module of R is the image
via Auslander functor of the fundamental representation V of G, that is
E = (S ⊗k V)G .
12
HOLGER BRENNER AND ALESSIO CAMINATA
Theorem 3.2 (Yoshino-Kawamoto, [YK88]). Let R = k~x, y, z/( f ) be a complete two-dimensional normal
non-regular domain with canonical module, with f ∈ k~x, y, z. Then the fundamental module E is isomorphic
to the third syzygy of k, i.e.
E Syz3R (k).
Moreover, the following facts are equivalent.
(1) The fundamental module E is decomposable.
(2) R is a cyclic quotient singularity.
Observe that the Kleinian singularities satisfy the condition of the previous Theorem 3.2. Now we
prove that also the second syzygy of the residue field is isomorphic to the fundamental module.
Theorem 3.3. Let (R, m, k) be a two-dimensional Kleinian singularity. Then the second syzygy of the residue
field k is isomorphic to the fundamental module, that is
Syz2R (k) E.
From Example 3.1, it is clear that Theorem 3.3 is equivalent to the following statement.
Theorem 3.4. Let (R, mR , k) be a two-dimensional Kleinian singularity, and let V1 be the two-dimensional
fundamental representation of the acting group G of R. Then, the second syzygy of the residue field k is
isomorphic to the image of V1 via Auslander functor, that is
Syz2R (k) A(V1 )
Theorem 3.4 can be proved by case considerations. We illustrate the general strategy and we apply
it to the singularity An−1 in Example 3.5. The reader interested to the computations for Dn+2 , E6 , E7
and E8 may consult [Cam16].
Proof. Let S = k~u, v be the powers series ring over an algebraically closed field k, and let G be one of
the Klein subgroups of SL(2, k) acting on S through a faithful representation V1 . Assume that char k
and |G| are coprime. Then we have R = SG , and we denote by mR the maximal ideal of R. Let p1 , p2 , p3
be a minimal system of generators of mR as ideal in S.
We consider the following short exact sequence of S-modules, which is the beginning of an S-free
resolution of S/mR S
(4)
p1 ,p2 ,p3
0 → Syz1S (p1 , p2 , p3 ) → S3 −−−−−→ S → S/mR S → 0.
The group G acts linearly on S through the fundamental representation V1 , and this action extends
naturally to S3 , and its submodule Syz1S (p1 , p2 , p3 ). In other words, the sequence (4) is an exact
sequence of S ∗ G-modules. We apply the functor G(N) = NG , which is exact, to the sequence (4) and
we obtain an exact sequence
0 → M → R3 → R → R/mR → 0.
Since p1 , p2 , p3 are a system of generators for the maximal ideal mR of R, the last non-zero module
on the right is the residue field. It follows that the module M appearing on the left of the last sequence
is just the second syzygy of k, that is M = Syz2R (k). In other words we have that
Syz1S (p1 , p2 , p3 )G = Syz2R (k).
So, to understand which R-module Syz2R (k) is, we need to understand which is the action of G on
Syz1S (p1 , p2 , p3 ), that is its S ∗ G-module structure.
We know that as S-module Syz1S (p1 , p2 , p3 ) S2 and it is therefore generated by two elements
s1 (u, v), s2 (u, v) ∈ S3 . We need to keep track of the action of G through this isomorphism. The action
of G on Syz1S (p1 , p2 , p3 ) is inherited by the action on S, which is linear and given by matrices M = M g
in SL(2, k).
In order to understand how these matrices act on the generators s1 (u, v) and s2 (u, v) we procede
as follows. For each matrix M we apply the linear transformation (u, v) 7→ M(u, v)T to s1 (u, v) and
THE SYMMETRIC SIGNATURE
13
s2 (u, v). We obtain two elements s′1 (u, v) and s′2 (u, v) in S3 which belong to Syz1S (p1 , p2 , p3 ). Therefore
we can write them as linear combination of s1 and s2
!
!
s′1
s1
=N
s2
s′2
for some matrix N. In this way we obtain a collection of matrices N = N g for g ∈ G, which gives
us the representation corresponding to Syz1S (p1 , p2 , p3 ). Then one has to check that this is actually
isomorphic to the fundamental representation V1 .
Example 3.5. Let ξ be a primitive n-th root of unity in k, and consider the cyclic group Cn generated
by
!
ξ 0
A=
.
0 ξ−1
The maximal ideal mR of the invariant ring R = k~u, vCn is generated by polynomials p1 = un , p2 = vn ,
p3 = uv. Their syzygy module Syz1S (p1 , p2 , p3 ) is generated by
0
−v
−u
0
s1 =
and s2 =
.
n−1
n−1
v
u
We apply the linear transformation given by A: u 7→ ξu, v 7→ ξ−1 v, to s1 and s2
0
0
−ξu
s1 7→ s′1 =
= ξ −u = ξs1 ;
−1 n−1 n−1
n−1
(ξ ) v
v
−1
−v
−ξ v
= ξ−1 0 = ξ−1 s2 .
0
s2 7→ s′2 =
n−1 n−1
un−1
ξ u
Thus, the representation corresponding to Syz1S (p1 , p2 , p3 ) is exactly the fundamental representation.
The referee suggested to us an alternative proof of Theorem 3.4 which works for the singularities
Dn+2 , E6 , E7 and E8 . We sketch it in the following remark.
Remark 3.6. We use the same notation as in the proof of Theorem 3.4. First, we observe that by
Auslander correspondence the isomorphism Syz2R (k) A(V1 ) is equivalent to say that the top of
Ext2S (S/mR S, S) is V1∗ . By the equivariant local duality (cf. [HO10, (5.5)]), this is equivalent to say
that the socle of S/mR S is V1 . Now let x, y, z ∈ S be a minimal system of generators of mR as in the
table at page 11, and consider the quotient Λ = S/(y, z). Since y and z are invariants, Λ inherits a
natural G-action. By the relation f in the table, Λ is Artinian and hence it is a complete intersection.
Moreover, the degree relation deg x = deg y + deg z − 1 holds and implies that x · u = x · v = 0, because
deg(x · u) = deg(x · u) = deg y + deg z. So x must sit in the socle of Λ, which is one-dimensional and is
therefore equal to Λdeg x , the degree deg x component of Λ. Since x is invariant, the socle of Λ is the
trivial representation. The natural multiplication Λi × Λdeg x−i → Λdeg x k forces Homk (Λ, k) Λ as
S ∗ G-modules, and moreover we have isomorphisms in each degree, i.e. Λi Homk (Λ, k)deg x−i . We
quotient by x, and we have that the socle of Λ/xΛ S/mR S, which is two-dimensional, must agree with
Λdeg x−1 , which is also two-dimensional. So we obtain Λdeg x−1 Homk (Λ, k)1 Homk (V1 , k) V1 ,
since the fundamental representation of the Klein groups is self-dual, and we are done.
For Kleinian singularities, also the module of Zariski differentials Ω∗∗
is isomorphic to the fundaR/k
mental module.
Theorem 3.7 (Martsinkovsky, [Mar90]). Let (R, m, k) be a two-dimensional quotient singularity over an
algebraically closed field k, and assume that the characteristic of k does not divide the order of the acting group.
Then the module of Zariski differentials Ω∗∗
of R over k is isomorphic to the fundamental module E.
R/k
14
HOLGER BRENNER AND ALESSIO CAMINATA
3.2. The decomposition of Symq (V). Let G be a finite group of order n and let k be an algebraically
closed field such that char k does not divide n. Let µn (k) be the group of n-th roots of unity in k and
let µn (C) be the group of complex n-th roots of unity. Both µn (k) and µn (C) are cyclic groups of order
n, so we can fix an isomorphism φ : µn (k) → µn (C), which we name a lift. In the same way, we say
that a complex root of unity z ∈ µn (C) is a lift of a ∈ µn (k) if z = φ(a).
Let (V, ρ) be a k-representation of G of dimension r ≥ 1 and let g be an element of G. Then the
matrix ρ(g) is diagonalizable in k and its eigenvalues are elements of µn (k), since the order of g divides
n. Let λ1 , . . . , λr be these eigenvalues, counted with multiplicity. The Brauer character or simply the
character of (V, ρ) is the function χ : G → C given by
χV (g) = φ(λ1 ) + · · · + φ(λr ).
This definition depend on the choice of the isomorphism φ. Since there are in general many choices
for φ, we have a certain degree of arbitrariness. However once chosen φ, it will never be changed
and sometimes we will simply say that we lift the eigenvalues to C, meaning that the isomorphism
φ is fixed.
For a proof of the following classical results of character theory the reader may consult the books
of Feit [Fei82, Chapter IV], and of Fulton and Harris [FH91, Chapter 2].
Proposition 3.8. Let V and W be k-representations of G with characters χV and χW . Then the following facts
hold:
(1)
(2)
(3)
(4)
(5)
χV (e) = dimk V, where e is the identity of G;
χV (g−1 ) = χV (g), the complex conjugate of χV (g), for every g ∈ G;
χV (hgh−1) = χV (g) for every g, h ∈ G;
χV⊕W (g) = χV (g) + χW (g) for every g ∈ G;
χV⊗W (g) = χV (g) · χW (g) for every g ∈ G.
We define an Hermitian inner product on the set of characters of G by
1 X
hϕ, ψi :=
ϕ(g)ψ(g),
|G|
g∈G
for every ϕ, ψ characters, where ϕ(g) denotes the complex conjugation. This bilinear form satisfies
1 X
hχ, ψi =
χ(g)ψ(g) = hψ, χi,
|G|
g∈G
for every characters χ, ψ.
The characters of the irreducible representations of G are orthonormal with respect to this inner
product.
Theorem 3.9. Let G be a finite group and let k be an algebraically closed field such that char k does not divide
|G|. Then the following facts hold.
(1) A representation V over k is irreducible if and only if hχV , χV i = 1.
(2) If χ and ψ are the characters of two non-isomorphic irreducible k-representations then hχ, ψi = 0.
Corollary 3.10. Let k and G be as in Theorem 3.9. Let V be a k-representation of G and let
V = V1n1 ⊕ · · · ⊕ Vrnr
be its decomposition into irreducible representations Vi . If χV is the character of V and χVi is the character of
Vi then
ni = hχV , χVi i.
In particular, two representations are isomorphic if and only if they have the same character.
Thanks to Corollary 3.10, characters are an important tool to understand the decomposition of a
representation into irreducible representations. We will use them to obtain the decomposition of the
q
q-th symmetric powers Symk (V) of a two-dimensional faithful representation V of a Klein group. We
begin with an elementary statement.
THE SYMMETRIC SIGNATURE
15
Lemma 3.11. Let λ , 1 be a root of unity in C, and let f : N → C be the function
f (N) :=
q
N X
X
λ2t−q .
q=0 t=0
2
Then f (N) = o(N ), that is
| f (N)|
= 0.
N→+∞ N 2
lim
Proof. We have
P
P
PN Pq
N+1 q PN+1 −q
2t−N−1
2t−q
− N+1
q=0 λ
q=0 λ
t=0 λ
| f (N)| 2 q=0 t=0 λ
≤
=
0≤
N2
N2
N2
2
4/|1 − λ| + N + 2
≤
,
N2
which goes to 0 for N → +∞.
Example 3.12. Let (V, ρ) be a two-dimensional representation of a finite group G over an algebraically
closed field k. Since k is algebraically closed, for every element g ∈ G the matrix ρ(g) can be
diagonalized
!
λ 0
ρ(g) =
,
0 µ
with λ, µ ∈ k. The representation Symq (V) evaluated at the element g is given by the matrix
q
0
···
···
0
λ
0 λq−1 µ 0
···
0
.
.. .
..
..
.
.
q−1
···
· · · λµ
0
0
···
···
0
µq
In other words, if λ and µ are the eigenvalues of V at the element g, then the eigenvalues of Symq (V)
at g are {λt µq−t : t = 0, . . . , q}.
Lemma 3.13. Let k be an algebraically closed field and let G be finite group such that char k and |G| are
coprime. Let (V, ρ) be a two-dimensional representation of G whose image is contained in SL(2, k), then the
(Brauer) character of the representation Symq (V) is given by
χSymq (V) (g) =
q
X
2t−q
λg
,
t=0
where λ g is the lift to C of an eigenvalue of the matrix ρ(g).
Proof. Let g ∈ G and let λ g and µ g be the lift to C of the eigenvalues of ρ(g). Since ρ(g) ∈ SL(2, k), we
q
have µ g = λ−1
g . Then the lift of the eigenvalues of g in the representation Sym (V) are
q−t
2t−q
{λtg · µ g : t = 0, . . . , q} = {λ g
: t = 0, . . . , q},
q
so the formula for the character of Sym (V) follows immediately.
q
Remark 3.14. The character of the symmetric representation Sym (V) can be computed also using
Molien’s formula (cf. [Stu08, Theorem 2.2.1])
(5)
+∞
X
q=0
χSymq (V) (g)tq =
1
.
det(I − tρ(g))
Here t is an indeterminate, and the determinant on the right hand side is of a matrix with entries
conveniently lifted to the polynomial ring C[t]. Expanding the rational function on the right, we
obtain a formal power series which is equal to the formal power series on the left, and we can use
this equality to compute the character χSymq (V) . Notice that Molien’s formula holds also if dimk V > 2.
16
HOLGER BRENNER AND ALESSIO CAMINATA
Theorem 3.15. Let k be an algebraically closed field and let G be a Klein subgroup of SL(2, k), such that char k
does not divide |G|. Let (V, ρ) be a two-dimensional faithful k-representation of G with image contained in
SL(2, k) and let (Vi , ρi ) be an irreducible k-representation of G. We denote by αi,q (V) the multiplicity of Vi in
Symq (V) and by βq (V) = dimk Symq (V). Then
PN
dimk Vi
q=0 αi,q (V)
=
.
lim PN
N→+∞
|G|
q=0 βq (V)
Proof. Since dimk V = 2, the dimension of Symq (V) is q + 1. So the denominator in the limit above is
PN
1
q=0 βq (V) = 2 (N + 1)(N + 2). From Corollary 3.10 and Lemma 3.13 we have
X
|G|αi,q (V) = hχSymq (V) , χVi i =
χSymq (V) (g) · χVi (g)
g∈G
=
X
g∈G
q
X 2t−q
χVi (g)
λ g ,
t=0
where λ g is the lift to C of an eigenvalue of ρ(g). The order m g of the root of unity λ g coincides with
the order of ρ(g) in SL(2, k). Since V is faithful, this coincides also with the order of the group element
g in G. In particular, λ g = 1 if and only if g is the identity I of G. Thus, we can write the previous sum
as
q
q
q
X
X X
X 2t−q
X 2t−q
χVi (I)
χVi (g)
χVi (g)
1 +
λ g = dimk (Vi )(q + 1) +
λ g .
t=0
t=0
g∈G
g,I
g∈G
g,I
t=0
We sum for q running from 0 to a fixed natural number N and we obtain
|G|
N
X
αi,q (V) =
q=0
1
dimk (Vi )(N + 1)(N + 2) + o(N2 ),
2
thanks to Lemma 3.11. Therefore, the numerator of the limit is
N
X
αi,q (V) =
q=0
so the limit is equal to
dimk Vi
|G|
(N + 1)(N + 2) dimk Vi
+ o(N2 )
2|G|
as desired.
Remark 3.16. The multiplicity αi,q (V) can be computed also using Molien’s formula (5), as showed
by Springer in [Spr87]. One obtains
+∞
X
q=0
αi,q (V)tq =
1 X χVi (g−1 )
.
|G|
det(I − tρ(g))
g∈G
3.3. Symmetric signature of Kleinian singularities.
Theorem 3.17. Let k be an algebraically closed field and let G be a Klein subgroup of SL(2, k) such that char k
does not divide |G|. Let R be the corresponding Kleinian singularity and let Mi be an indecomposable MCM
R-module. Then the generalized symmetric signature and the generalized differential symmetric signature of
R with respect to Mi are
rankR Mi
sσ (R, Mi ) = sdσ (R, Mi ) =
.
|G|
Proof. Let E be the fundamental module of R. By Theorem 3.3, and by Theorem 3.7 we have
E Syz2R (k) Ω∗∗
R/k .
Therefore the two symmetric signatures, differential and syzygy, coincide for Kleinian singularities.
THE SYMMETRIC SIGNATURE
17
Let Vi be the irreducible k-representation of G such that Mi = A(Vi ), and let V be the fundamental
representation of G, that is the k-representation such that A(V) = E. Let αi,q (E) be the multiplicity of
q
q
Mi into SymR (E)∗∗ and let βq (E) = rankR SymR (E)∗∗. From Theorem 2.12 we have
q
A′ SymR (E)∗∗ = Symq (V).
It follows that βq (E) = dimk Symq (V) and αi,q (E) equals the multiplicity of the representation Vi in
Symq (V). Since V is the fundamental representation of a Klein group, it is faithful and its image is
contained in SL(2, k), so we can apply Theorem 3.15 to conclude the proof.
Corollary 3.18. Let k be an algebraically closed field and let G be a Klein subgroup of SL(2, k) such that char k
does not divide |G|. The symmetric signature and the differential symmetric signature of the corresponding
Kleinian singularity R are
1
.
sσ (R) = sdσ (R) =
|G|
Example 3.19. Consider the An−1 -singularity R = k~u, vCn over an algebraically closed field k, and
∗∗
q
let Sq = SymR (Syz2R (k)) as in the definition of symmetric signature. From the proofs of Theorem
3.15 and Theorem 3.17, we have that rankR Sq = q + 1, and
2(q+1)
q+1
n + O(1) for q even
q
q
(1 + (−1) ) + O(1) =
frkR S =
O(1)
n
for q odd.
This shows that the limit
frkR Sq
q→+∞ rankR Sq
does not exist. Specifically, in this setting the syzygy module splits as Syz2R (k) M1 ⊕ Mn−1 (with
notation as in Example 2.6). Therefore we have
lim
Sq
q
M
t=0
⊗q−t
M⊗t
1 ⊗ Mn−1
q
M
⊗2t−q
M1
.
t=0
4. Symmetric signature of cones over elliptic curves
In this section we prove that the differential symmetric signature of the coordinate ring R of a
plane elliptic curve over an algebraically closed field of characteristic different from 2 and 3 is zero.
Our intention is motivated by the following remark, which shows that the F-signature of such rings
(over a field of positive characteristic) is zero.
Remark 4.1. We consider R = k[x, y, z]/( f ), with k an algebraically closed field of positive characteristic, and f a homogeneous non-singular polynomial of degree 3. In other words, R is the coordinate
ring of a plane elliptic curve over k. The ring R is not strongly F-regular, and in particular its Fsignature s(R) is 0 by the result of Aberbach and Leuschke (Theorem 1.5). To see this, assume for
simplicity that f is in Weierstrass normal form, that is f = y2 z − x3 − axz2 − bz3 for some a, b ∈ k. This
is always possible if the characteristic of k is different from 2 and 3. Then the ideal I = (x, z) is not
tightly closed in R, since y2 < I, but y2 ∈ I∗ . Therefore R is not strongly F-regular and s(R) = 0.
The methods we use in this situation are different from those of previous sections, and are of
geometric nature. We will use the correspondence between graded MCM R-modules and vector
bundles over the smooth projective curve Y = Proj R to translate the problem into geometric language.
We will not make a general assumption on the characteristic of the base field k in this section, but we
will restrict to characteristic , 2, 3 only when needed. The letter q is used to denote a positive integer.
4.1. Generalities on vector bundles. We recall some preliminary facts concerning vector bundles
over curves, and in particular over elliptic curves. For further details and the proofs of these facts,
we refer to the books of Hartshorne [Har77], Le Potier [LeP97], and Mukai [Muk03, Chapter 10].
Let Y be a smooth projective curve over an algebraically closed field k. We denote by VB(Y) the
category whose objects are vector bundles over Y, and whose maps are morphisms between them.
The category VB(Y) is a Krull-Schmidt category (cf. [Ati56, Theorem 3]). Moreover there is an
18
HOLGER BRENNER AND ALESSIO CAMINATA
equivalence of categories between the category of vector bundles over Y and the category of locally
free sheaves on Y which respects the rank (cf. [Har77, Ex. II 5.18]). Having this identification in mind,
we will use indifferently the words vector bundle and locally-free sheaf. For example, we will use the
notation for the structure sheaf OY of Y also to denote the trivial bundle of rank one corresponding
to it.
Let S be a locally free sheaf of rank r over X. The
V degree of S is defined as the degree of the
corresponding determinant line bundle degS = deg r S. The slope of S is µ(S) = deg S/r. The
degree is additive on short exact sequences and moreover µ(S ⊗ T ) = µ(S) + µ(T ).
The sheaf S is called semistable if for every locally free subsheaf T ⊆ S the inequality µ(T ) ≤ µ(S)
holds. If the strict inequality µ(T ) < µ(S) holds for every proper subsheaf T ⊂ S, then S is called
stable. Stable vector bundles are indecomposable, but the converse does not hold in general.
The usual operations on vector spaces like direct sum, tensor product, symmetric and wedge
powers extend naturally to vector bundles and correspond to the same constructions in the category
of locally free sheaves. The dual of a vector bundle E is the bundle E∨ := Hom(E, OY ).
Remark 4.2. Since every vector bundle E is canonically isomorphic to its double dual E∨∨ , reflexive
symmetric powers Symq (−)∨∨ coincide with ordinary symmetric powers Symq (−) in the category
VB(Y).
The following are standard properties of symmetric powers of vector bundles (see also [Laz04,
Corollary 6.4.14]).
Proposition 4.3. Let Y be a smooth projective curve, let E be a vector bundle and L a line bundle over Y.
Then the following facts hold.
1) Symq (E ⊗ L) Symq (E) ⊗ L⊗q .
2) µ(Symq (E)) = q · µ(E).
3) If the field k has characteristic 0 and E is semistable, then Symq (E) is also semistable.
Let R be a normal standard-graded domain of dimension 2 over an algebraically closed field k,
that is R0 = k and R is generated by finitely many elements of degree 1. The normal assumption on
R implies that Rp is a regular ring for every prime ideal p , R+ . It follows that the projective variety
e the
Y = Proj R is a smooth projective curve over k. For every graded module M we denote by M
g are invertible (cf. [Har77, Proposition II.5.12])
corresponding coherent sheaf on Y. The sheaves R(n)
and denoted by OY (n). Moreover, every MCM graded R-module M is locally free on the punctured
e is in fact a vector bundle over Y. If we denote by
spectrum, so the associated coherent sheaf M
MCMZ (R) the category of finitely generated graded MCM R-modules, then we have a functor
e: MCMZ (R) → VB(Y),
whose properties we collect in the following proposition.
Proposition 4.4. Let R be a normal standard graded domain of dimension 2 over a field k, let Y = Proj R, and
let M and N be finitely generated graded MCM R-modules. Then the following facts hold.
e ⊕ N;
e
1) M]
⊕N M
∼∗∗ e
e
2) (M ⊗R N) M ⊗OY N;
∼
q
q
e
3) SymR (M)∗∗ SymY (M).
Now we give an appropriate definition of free rank for the category of vector bundles over Y.
Definition 4.5. Let E be a vector bundle over Y with a fixed very ample invertible sheaf OY (1). We
define the free rank of E as
n
M
n
frkOY (1) (E) := max n : ∃ a split surjection ϕ : E ։ F, with F =
OY (−di )
o
a splitting vector bundle of rank n .
i=1
THE SYMMETRIC SIGNATURE
19
Since the category VB(Y) has the KRS property, to compute the free rank of a bundle E, one should
count how many copies of twisted structure sheaves OY (−di ) appear in the decomposition of E into
indecomposable bundles. For this reason, the free rank of vector bundles frkOY (1) agrees with the
gr
graded free rank frkR and not with the free rank, as explained in the following corollary.
Corollary 4.6. Let R be a normal standard graded domain of dimension 2 over a field k, let Y = Proj R, and
let M be a finitely generated graded MCM R-module. Then the following facts hold.
e
(1) rankR M = rankOY M.
gr
e
(2) frkR M = frkOY (1) M.
Remark 4.7. Let I be an R+ -primary homogeneous ideal with homogeneous generators f1 , . . . , fn of
degrees di . The following presentation of R/I
n
M
f1 ,..., fn
R(−di ) −−−−→ R → R/I → 0
i=1
induces a short exact sequence of sheaves on Y
0→S→
n
M
OY (−di ) → OY → 0,
i=1
Ln
where R/I corresponds to 0 because I is R+ -primary. Since OY and
i=1 OY (−di ) are locally free, the
sheaf S is also locally free. The sheaf S is the kernel of the sheaf morphism f1 , . . . , fn and is denoted
by Syz( f1 , . . . , fn ) and called syzygy bundle. In fact, it is nothing but the vector bundle corresponding
to the MCM graded R-module Syz2R (R/I) = Syz1R ( f1 , . . . , fn ).
Syzygies of the irrelevant maximal ideal R+ play a special role, they correspond to the restriction
of the cotangent bundle ΩPn of Pn to the curve. In fact, it follows from the Euler sequence on Pn
[Har77, Theorem 8.13]
n
M
0 → ΩPn →
OPn (−1) → OPn → 0,
i=0
that ΩPn is a syzygy bundle. So if x0 , . . . , xn is a system of generators for OPn (1), we have the
isomorphism
Syz(x0 , . . . , xn ) ΩPn .
If we restrict these bundles to the curve Y, we obtain isomorphic bundles. In fact, we have just proved
the following proposition.
Proposition 4.8. Let Y be a smooth projective curve with an embedding in Pn given by ϕ : Y → Pn . Let
x0 , . . . , xn be a system of generators of OPn (1) and denote by ΩPn the cotangent bundle of Pn and by ΩPn |Y its
restriction to Y. Then, we have the following isomorphism of vector bundles on Y
Syz(x0 , . . . , xn )|Y ΩPn |Y .
It is a classical result, usually ascribed to Grothendieck, that the only indecomposable vector
bundles over a curve of genus zero are of the form OY (a) for some integer a. Therefore the structure of
the category VB(Y) is quite easy for such a curve. The next interesting case concerns a curve of genus
1, an elliptic curve. This case was treated by Atiyah [Ati57], who was able to give a description of
all indecomposable vector bundles of fixed rank and degree. Moreover he gave explicit formulas for
the multiplicative structure of the monoid VB(Y) with the tensor product operation if the base field
has characteristic zero.
In the following theorem we collect some results of Atiyah that we will need later on.
Theorem 4.9 (Atiyah, [Ati57]). Let Y be an elliptic curve over an algebraically closed field k with structure
sheaf OY . For every positive integer r there exists a unique (up to isomorphism) indecomposable vector bundle
Fr of rank r and degree 0 with Γ(Y, Fr ) , 0. Moreover, the following facts hold.
(1) Fr has only one section up to scalar multiplication, i.e. Γ(Y, Fr ) k.
(2) Fr F∨r .
20
HOLGER BRENNER AND ALESSIO CAMINATA
(3) There is a non-split exact sequence
0 → OY → Fr → Fr−1 → 0.
(4) Symq (F2 ) Fq+1 .
(5) If E is an indecomposable vector bundle of rank r and degree 0 then E L ⊗ Fr , where L is a line bundle
of degree zero, unique up to isomorphism and such that L det E.
We call the bundle Fr Atiyah bundle of rank r. Clearly we have that F1 = OY , the structure sheaf
is the unique line bundle of degree 0 with non-zero sections. The bundle F2 is given by the unique
non-trivial extension
0 → OY → F2 → OY → 0.
In other words, F2 is the non-zero element of Ext1 (OY , OY ).
The following two results are well known to experts. However, since we are unable to locate a
proof in the literature, for the sake of completeness we add a proof here.
Lemma 4.10. Let Y be an elliptic curve over an algebraically closed field k with structure sheaf OY . Then for
every positive integers r and q the vector bundle Symq (Fr ) is semistable.
Proof. Indecomposable bundles over an elliptic curve are semistable (cf. [Tu93, Lemma 29]), so if
char k = 0 the result follows from Proposition 4.3.
Now assume that char k = p > 0. The Atiyah bundle Fr is defined inductively by a non-split short
exact sequence 0 → OY → Fr → Fr−1 → 0, hence by a non-zero cohomology class c ∈ H1 (Y, F∨r−1 )
H1 (Y, Fr−1 ). By [Bre05, Lemma 2.4] the pe -multiplication [pe ] : Y → Y on the elliptic curve kills this
cohomology class for a suitable choice of e, that is [pe ]∗ (c) = 0. Therefore the pullback via [pe ] of the
previous sequence splits, and inductively this shows that [pe ]∗ (Fr ) is free. Hence its symmetric powers
are also free and in particular semistable. Then the semistability of [pe ]∗ (Symq (Fr )) Symq ([pe ]∗ (Fr ))
implies that Symq (Fr ) is semistable.
Lemma 4.11. Let Y be an elliptic curve over an algebraically closed field k, and let E be a semistable vector
bundle, then Symq (E) is semistable.
Proof. If char k = 0, then the result follows from Proposition 4.3. So we assume that char k = p > 0,
and we denote by r the rank of E. By a result of Oda [Oda71] (see also [Tu93]) the bundle E is strongly
semistable, so by [Miy87, Proposition 5.1] it follows that [n]∗ E is semistable for every integer n, where
[n] : Y → Y is the standard multiplication on the elliptic curve.
We consider the semistable bundle [r]∗ E. Its degree is a multiple of r, so by tensoring with an
appropriate line bundle L, we obtain a semistable vector bundle E′ = [r]∗ E ⊗ L of degree 0. It follows
that the indecomposable direct summands of E′ must have degree 0 too, so they are of the form Fs ⊗ L,
with L line bundle. With the help of a further multiplication map as in the proof of Lemma 4.10 we
may achieve that [n]∗ (E) is a direct sum of line bundles of degree 0. Hence its symmetric powers are
semistable and this descends to the symmetric powers of E.
4.2. Symmetric signature of cones over elliptic curves. Let f be a homogeneous non-singular polynomial f of degree 3 in k[x, y, z], and let R = k[x, y, z]/( f ). Then R is a normal standard graded
k-domain of dimension 2 and the projective curve Y = Proj R is a plane elliptic curve over k.
To compute the symmetric signature of R one should consider reflexive symmetric powers of the
second syzygy of the residue field, Syz2R (k). Thanks to Remark 4.7 and Proposition 4.4 this is equivalent
to consider the syzygy bundle Syz(x, y, z) over Y and taking ordinary symmetric powers in the
e and frkgr M = frkO (1) M
e
category VB(Y). We recall that by Corollary 4.6 we have rankR M = rankOY M
Y
R
for every MCM graded R-module. Therefore, we have that sσ (R) exists if and only if the limit
PN
q
q=0 frkOY (1) Sym (Syz(x, y, z))
lim PN
q
N→+∞
q=0 rankOY Sym (Syz(x, y, z))
exists, and in this case they coincide.
THE SYMMETRIC SIGNATURE
21
The same reasoning applies to the differential symmetric signature. In this case, we should consider
the sheaf associated to the module of Zariski differentials Ω∗∗
on Y and then take ordinary symmetric
R/k
powers in the category VB(Y). In other words, we consider the vector bundles
q g
SymY Ω
R/k .
q
Notice that we can forget about the double dual inside SymY (−), thanks to Proposition 4.4. The
differential symmetric signature of the coordinate ring R exists if and only if the limit
PN
q g
q=0 frkOY (1) SymY ΩR/k
(6)
lim P
q g
N
N→+∞
rankO Sym Ω
R/k
q=0
Y
Y
exists, and in this case they coincide.
Proposition 4.12. Let Y be a plane elliptic curve over a field k of characteristic , 2, 3 with coordinate ring
k[x, y, z]/( f ), where f is a homogeneous polynomial of degree 3. Then we have
!
∂f ∂f ∂f
Syz
,
,
(3) F2 .
∂x ∂y ∂z
∂f ∂f ∂f
Proof. It is enough to show that Syz ∂x , ∂y , ∂z (3) corresponds to a non-zero element of Ext1 (OY , OY )
with global sections, then the uniqueness of F2 will imply the desired isomorphism.
We consider the following injective map of sheaves, ϕ : OY → OY (1)⊕3 , given by ϕ(1) = (x, y, z).
From the Euler formula
∂f
∂f
∂f
x+
y+
z = 3 f,
∂x
∂y
∂z
∂f ∂f ∂f
which vanishes on Y, we obtain that the image of ϕ is contained in the syzygy bundle Syz ∂x , ∂y , ∂z (3).
Actually, since this image does not vanish anywhere, it defines a subbundle and hence also a quotient bundle, which is by rank and degree reasons the structure sheaf. In other words, we have the
following short exact sequence of vector bundles
!
∂f ∂f ∂f
ϕ
(7)
0 → OY −
→ Syz
,
,
(3) → OY → 0,
∂x ∂y ∂z
It remains to prove that the sequence (7) is non-split, equivalently (x, y, z) is the unique non-zero
global section. This can be done easily by explicit computations, for example assuming that f is in
Weierstrass normal form.
g
Corollary 4.13. Let char k , 2, 3, and let Ω
R/k be the sheaf version of the cotangent module ΩR/k of the cone
R of Y, then we have an isomorphism of vector bundles
g
Ω
R/k (−1) F2 .
Proof. We recall that the cotangent module ΩR/k is the graded R-module
!
∂f
∂f
∂f
ΩR/k =< dx, dy, dz > /
dx +
dy +
dz .
∂x
∂y
∂z
In other words, ΩR/k can be defined by the following short exact sequence of graded R-modules
ψ
(8)
where ψ(1) =
∂f
ϕ
0 → R(−2) −
→ R⊕3 −
→ ΩR/k → 0,
∂f
∂x , ∂y ,
∂f
∂z
, and ϕ sends the canonical basis to dx, dy, dz.
Sequence (8) induces a short exact sequence of locally free sheaves on Y
g
0 → OY (−2) → O⊕3
Y → ΩR/k → 0.
We dualize this sequence and we get
∨
⊕3
g
0→Ω
R/k → OY → OY (2) → 0,
22
HOLGER BRENNER AND ALESSIO CAMINATA
where the last map is given by ψ. We obtain that
g
Ω
R/k
!
∂f ∂f ∂f
,
,
(2).
Syz
∂x ∂y ∂z
∨
∨
g
So by the previous Proposition 4.12 we have Ω
R/k (1) F2 on Y, but the Atiyah bundle F2 is self-dual,
∨
g
so we have Ω
(−1)
F
F
.
2
R/k
2
We can compute the differential symmetric signature of the cone R.
Theorem 4.14. Let Y be a plane elliptic curve over an algebraically closed field k of characteristic , 2, 3 with
coordinate ring R. Then the differential symmetric signature of R is sdσ (R) = 0.
Proof. By the previous observations we should compute the limit (6). From Corollary 4.13, we have
g
Ω
R/k F2 ⊗ OY (1). Therefore from part 4) of Theorem 4.9 and from Proposition 4.3 we obtain
q
q g
SymY Ω
R/k SymY (F2 ) ⊗ OY (q) Fq+1 ⊗ OY (q)
q g
for all q ≥ 1. So the module SymY Ω
R/k is indecomposable, and in particular it has free rank 0 and
the claim follows.
Proposition 4.15. The vector bundle Syz(x, y, z) is stable of rank 2 and degree −9. Moreover det Syz(x, y, z) =
OY (−3).
Proof. The syzygy bundle fits into a short exact sequence
0 → Syz(x, y, z) →
(9)
3
M
x,y,z
OY (−1) −−−→ OY → 0.
i=1
So from the additivity of rank and degree, we immediately get that Syz(x, y, z) has rank 2 and degree
3
M
deg Syz(x, y, z) = deg
OY (−1) − deg OY deg Y = −3 deg Y = −9.
i=1
For the indecomposable property, we have that Syz(x, y, z) equals the restriction of the cotangent
bundle ΩP2 of P2 to Y by Proposition 4.8. This bundle is stable, and in particular indecomposable,
by the result of Brenner and Hein [BH06, Theorem 1.3]. Finally, taking determinants of sequence (9)
yields det Syz(x, y, z) = OY (−3).
Remark 4.16. Since Syz(x, y, z) has rank 2, we have that rankOY Symq (Syz(x, y, z)) = q + 1.
The difficult part is to determine the free rank of Symq (Syz(x, y, z)).
Corollary 4.17. Let q be odd, then the bundle Symq (Syz(x, y, z)) contains no direct summand of rank one. In
particular, frkOY (1) Symq (Syz(x, y, z)) = 0.
Proof. From Proposition 4.3, Proposition 4.15 and Lemma 4.11 we know that Symq (Syz(x, y, z)) is
semistable of slope − 29 q. Let L be a direct summand of rank one. Since Symq (Syz(x, y, z)) is semistable,
we have µ(L) = − 29 q, but µ(L) = deg L must be an integer. Since q is odd we get a contradiction.
Corollary 4.18. If the symmetric signature of R exists, then sσ (R) ≤ 21 .
Proof. For every q ∈ N, let aq = frkOY (1) Symq (Syz(x, y, z)), and let bq = rankOY Symq (Syz(x, y, z)). From
the previous Corollary 4.17 and Remark 4.16, we have aq = 0 for q odd, and bq = q + 1 for all q.
For every even q we have the following inequalities
aq + aq+1 = aq ≤ bq =
It follows that
PN
q=0 aq
PN
q=0 bq
1
2
PN
1
1
(bq + bq ) ≤ (bq + bq+1 ).
2
2
q=0 bq
≤ PN
q=0 bq
+ φ(N) =
1
+ φ(N),
2
THE SYMMETRIC SIGNATURE
where φ(N) =
proved.
b
PN N
q=0 bq
23
if N is even, and 0 otherwise. In both cases φ(N) → 0 if N → +∞, so the claim is
Question 4.19. Is it true that also sσ (R) = 0?
Acknowledgements
We would like to thank Winfried Bruns, Ragnar-Olaf Buchweitz, Helena Fischbacher-Weitz, Lukas
Katthän, and Yusuke Nakajima for their interest and many valuable comments. We thank the referee
for showing us how to simplify the proof of Theorem 3.15 and pointing out Remark 3.6.
References
[AL03] Ian Aberbach and Graham Leuschke, The F-signature and strong F-regularity, Math. Res. Lett., vol. 10, no. 1, pp. 51–56,
2003.
[Ati56] Michael Atiyah, On the Krull-Schmidt theorem with application to sheaves, Bull. Soc. Math. France, vol. 84, pp. 307–317,
1956.
[Ati57] Michael Atiyah, Vector bundles over an elliptic curve, Proc. London Math. Soc. (3), vol. 7, pp. 414–452, 1957.
[Aus86b] Maurice Auslander, Rational singularities and almost split sequences, Trans. Amer. Math. Soc., vol. 293, no. 2, pp.
511–531, 1986.
[AG60] Maurice Auslander and Oscar Goldman, The Brauer group of a commutative ring, Trans. Amer. Math. Soc., vol. 97, pp.
367–409, 1997.
[BDO08] Fedor Bogomolov and Bruno De Oliveira, Symmetric tensors and geometry of PN subvarieties, Geom. Funct. Anal.
vol. 18, no. 3, 637–656, 2008.
[Bre05] Holger Brenner, Tight closure and plus closure for cones over elliptic curves, Nagoya Math. J., vol. 177, pp. 31–45, 2005.
[BF09] Holger Brenner and Helena Fischbacher-Weitz, Symmetric powers of vector bundles and Hilbert-Kunz theory, unpublished paper, 2009.
[BH06] Holger Brenner and Georg Hein, Restriction of the cotangent bundle to elliptic curves and Hilbert-Kunz functions,
Manuscripta Math., vol. 119, no. 1, pp. 17–36, 2006.
[BD08] Igor Burban and Yuriy Drozd, Maximal Cohen-Macaulay modules over surface singularities, Trends in representation
theory of algebras and related topics, pp. 101–166, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich, 2008.
[Cam16] Alessio Caminata, The symmetric signature, Ph.D. Thesis, Universität Osnabrück, 2016.
[CK16] Alessio Caminata and Lukas Katthän, The symmetric signature of cyclic quotient singularities, preprint arXiv:1603.06427.
[CHR65] Stephen Chase, David Kent Harrison, and Alex Rosenberg, Galois theory and Galois cohomology of commutative rings,
Mem. Amer. Math. Soc., no. 52, pp. 15–33, 1965.
[Fei82] Walter Feit, The representation theory of finite groups, North-Holland Mathematical Library, vol. 25, 1982.
[FH91] William Fulton and Joe Harris, Representation theory. A first course, Graduate Texts in Mathematics, vol. 129, Springer,
1991.
[Har77] Robin Hartshorne, Algebraic geometry, Graduate Texts in Mathematics, vol. 52, Springer, 1977.
[HN15] Mitsuyasu Hashimoto and Yusuke Nakajima, Generalized F-signature of invariant subrings, J. Algebra, vol. 443, pp.
142–152, 2015.
[HO10] Mitsuyasu Hashimoto and Masahiro Ohtani, Equivariant Matlis and the local duality, J. Algebra, vol. 324, pp. 1447–
1470, 2010.
[Her78a] Jürgen Herzog, Ein Cohen-Macaulay-Kriterium mit Anwendungen auf Konormalenmodul und den Differentialmodul, Math.
Z., vol. 163, no. 2, pp. 149–162, 1978.
[Her78b] Jürgen Herzog, Ringe mit nur endlich vielen Isomorphieklassen von maximalen, unzerlegbaren Cohen-Macaulay-Moduln,
Math. Ann., vol. 233, pp. 21–34, 1978.
[HL02] Craig Huneke and Graham Leuschke, Two theorems about maximal Cohen-Macaulay modules, Math. Ann., vol. 324, no.
2, pp. 391–404, 2002.
[Kun69] Ernst Kunz, Characterizations of regular local rings for characteristic p, Amer. J. Math., vol. 91, pp. 772–784, 1969.
[Laz04] Robert Lazarsfeld, Positivity in Algebraic Geometry II, Positivity for Vector Bundles, and Multiplier Ideals, Ergebnisse der
Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics, vol. 49, Springer, 2004.
[LeP97] Joseph Le Potier, Lectures on vector bundles, Cambridge Studies in Advanced Mathematics, vol. 54, Cambridge
University Press, 1997.
[LW12] Graham Leuschke and Roger Wiegand, Cohen-Macaulay representations, Mathematical Surveys and Monographs, vol.
181, American Mathematical Society, 2012.
[Mar90] Alex Martsinkovsky, Almost split sequences and Zariski differentials, Trans. Amer. Math. Soc., vol. 319, pp. 285–307,
1990.
[Miy87] Yoichi Miyaoka, The Chern class and Kodaira dimension of a minimal variety, Algebraic Geometry, Sendai 1985, Adv.
Stud. Pure Math., vol. 10, pp. 449–476, 1987.
[Muk03] Shigeru Mukai, An introduction to invariants and moduli, Cambridge Studies in Advanced Mathematics, vol. 81,
Cambridge University Press, 2003.
[Nor84] Douglas G. Northcott, Multilinear Algebra, Cambridge University Press, 1984.
[Oda71] Tadao Oda, Vector bundles on a elliptic curve, Nagoya Math. J., vol. 43, pp. 41–72, 1971.
24
HOLGER BRENNER AND ALESSIO CAMINATA
[Sak78] Fumio Sakai, Symmetric powers and the cotangent bundle and classification of algebraic varieties, Lecture Notes in Math.,
vol. 732, pp. 545–563, 1978.
[Sei97] Gerhard Seibert, The Hilbert-Kunz function of rings of finite Cohen-Macaulay type, Arch. Math. (Basel), vol. 69, no. 4, pp.
286–296, 1997.
[SVB96] Karen Smith and Michel Van Den Bergh, Simplicity of rings of differential operators in prime characteristic, Proc. London
Math. Soc. (3), vol. 75, no. 1, pp. 32–62, 1997.
[Spr87] Tonny Albert Springer, Poincaré Series of Binary Polyhedral Groups and McKay’s Correspondence, Math. Ann., vol. 278,
pp. 99–116, 1987.
[Stu08] Bernd Sturmfels, Algorithms in invariant theory. Second edition, Texts and Monographs in Symbolic Computation,
Springer, Vienna, 2008.
[Tu93] Loring W. Tu, Semistable bundles over an elliptic curve, Adv. Math., vol. 98, no. 1, pp. 1–26, 1993.
[Tuc12] Kevin Tucker, F-signature exists, Invent. Math., vol. 190, no. 3, pp. 743–765, 2012.
[Wat74a] Kei-ichi Watanabe, Certain invariant subrings are Gorenstein. I, Osaka J. Math., vol. 11, pp. 1–8, 1974.
[Wat74b] Kei-ichi Watanabe, Certain invariant subrings are Gorenstein. II, Osaka J. Math., vol. 11, pp. 379–388, 1974.
[WY00] Kei-ichi Watanabe and Ken-ichi Yoshida, Hilbert-Kunz multiplicity and an inequality between multiplicity and colength,
J. Algebra, vol. 230, no. 1, pp. 295–317, 2000.
[WY04] Kei-ichi Watanabe and Ken-ichi Yoshida, Minimal relative Hilbert-Kunz multiplicity, Illinois J. Math., vol. 48, pp.
273–294, 2004.
[Yao05] Yongwei Yao, Modules with finite F-representation type, J. London Math. Soc. (2), vol. 72, pp. 53–72, 2005.
[Yao06] Yongwei Yao, Observations on the F-signature of local rings of characteristic p, J. Algebra, vol. 299, no. 1, pp. 198–218, 2006.
[Yos90] Yuji Yoshino, Cohen-Macaulay modules over Cohen-Macaulay rings, London Mathematical Society Lecture Note Series,
vol. 146, Cambridge University Press, Cambridge, 1990.
[YK88] Yuji Yoshino and Takuji Kawamoto, The fundamental module of a normal local domain of dimension 2, Trans. Amer. Math.
Soc., vol. 309, no. 1, pp. 425–431, 1988.
Holger Brenner, Institut für Mathematik, Universität Osnabrück, Albrechtstrasse 28a, 49076 Osnabrück, Germany
E-mail address: [email protected]
Alessio Caminata, Institut für Mathematik, Universität Osnabrück, Albrechtstrasse 28a, 49076 Osnabrück, Germany
Current address: Institut de Mathématiques, Université de Neuchâtel, Rue Emile-Argand 11, CH-2000 Neuchâtel, Switzerland
E-mail address: [email protected]
| 0math.AC
|
arXiv:1709.06286v1 [math.GR] 19 Sep 2017
A note on the normal subgroup lattice of
ultraproducts of finite quasisimple groups
Jakob Schneider and Andreas Thom
Abstract. In [3] it was stated that the lattice of normal subgroups
of an ultraproduct of finite simple groups is always linearly ordered.
This is false in this form in most cases for classical groups of Lie type.
We correct the statement and point out a version of ‘relative’ bounded
generation results for classical quasisimple groups and its implications
on the structure of the lattice of normal subgroups of an ultraproduct
of quasisimple groups.
Contents
1.
2.
3.
Introduction
Auxiliary geometric results
Relative bounded generation in universal finite quasisimple
groups
4. The lattice of normal subgroups of an algebraic ultraproduct of
classical finite quasisimple groups
Acknowledgments
References
1
3
5
8
14
14
1. Introduction
The purpose of this note is to correct statements from the work of Stolz
and the second author in [3] and generalize to a setting of quasisimple
groups. As stated, [3, Theorem 3.9] is not correct and our main result
(Theorem 1) should replace it. Note that already in [1], it was pointed out
that some of the techniques and results of [3] were flawed. Some corrections on results about bounded generation in the setting of unitary groups
on finite-dimensional Hilbert spaces can be found in [1]. The statement of
[3, Theorem 4.20] should be considered as open problem at the moment. In
this note we focus entirely on the case of finite groups.
1
2
JAKOB SCHNEIDER AND ANDREAS THOM
Let G be a group, g ∈ G and L ⊂ G. We write gG := {hgh−1 | h ∈ G}
for the conjugacy class of g and call N normal if g ∈ L implies g G ⊂ L.
We denote by L∗k the set of k-fold products of elements in L, i.e. L∗k =
{l1 · · · lk | l1 , . . . , lk ∈ L}. Using the results of [2] it is a simple matter
to prove the following result about bounded ‘relative’ generation for the
alternating groups An with n ≥ 5:
Lemma 1. There exists c > 0 such that for any S, T ⊆ An normal
subsets with |S|, |T | ≥ 1, T 6= {1} for any integer
k ≥ c max{log|S|/ log|T |, 1}
it holds that S ⊆
(T An )∗k .
A straightforward consequence of this fact is that for any σ, τ ∈ An either
σ ∈ (τ An )∗k or τ ∈ (σ An )∗k for an integer k ≥ c which easily implies that
the normal subgroups of an algebraic ultraproduct of alternating groups are
linearly ordered.
However, in the case of finite classical simple groups of Lie type, all of
the above is false in general. The prototype of a counterexample is given by
the following families of elements.
∼ PSLq (q) be the elements given
Example 1. Let h1 , h2 ∈ H := SLq (q) =
by h1 = diag(1, λ, . . . , λ) and h2 = diag(1, µ, . . . , µ) where λ, µ ∈ F∗q are
arbitrary such that λ 6∈ hµi and µ 6∈ hλi, e.g. take λ = ζ a and µ = ζ b
where hζi = F∗q and a, b > 1 are coprime with q − 1 = ab. Then it is
−1 H ∗k
easy to show by induction that each h ∈ (hH
has at least
1 ∪ (h1 ) )
q − k non-trivial eigenvalues in hλi. From the assumptions it follows that
−1 H ∗k
h2 6∈ (hH
for any such k ≤ q − 1. Similarly, it follows that
1 ∪ (h1 ) )
−1 H ∗k
H
h1 6∈ (h2 ∪ (h2 ) ) for k ≤ q − 1.
This example already implies that the normal subgroups of an algebraic
ultraproduct of finite simple groups of type SLq (q) are not linearly ordered.
In this note, we shall prove the ‘best possible’ result on bounded relative
generation in classical quasisimple groups of Lie type. Recall, a quasisimple group is a perfect central extension of a simple group. Every finite
non-abelian simple group has a universal quasisimple extension which is
called the Schur covering group. We will also describe the lattice of normal
subgroups of an algebraic ultraproduct of the universal finite quasisimple
groups.
In the study of ultraproducts of finite quasisimple groups, we may ignore
exceptional finite non-abelian simple groups, abelian groups, and families of
bounded rank. Indeed, in those cases the ultraproduct yields a finite group
or a quasisimple group defined over a pseudofinite field. Those groups and
their lattice of normal subgroups are much easier to analyse. Assuming the
classification of non-abelian finite simple groups, it follows that the remaining cases consist only of ultraproducts of the covering groups of alternating
groups with increasing rank and of ultraproducts of quasisimple groups of
ON THE NORMAL SUBGROUP LATTICE
3
Lie type with increasing rank. We will show (see Remark 10) that the lattice
of normal subgroups in the case of universal covers of alternating groups is
still a linear order. It remains to study ultraproducts of universal quasisimple finite groups of Lie type. Each of the groups of Lie type appearing as
factors in the ultraproduct is endowed with its natural rank function ℓrk and
projective rank function ℓpr , as discussed later in the paper. We will prove
the following theorem.
Theorem 1. Let (Hi )i∈I be a sequence of universal
quasisimple finite
Q
groups of Lie type of increasing rank. Set G := i Hi and define subgroups
N rk := {(gi ) ∈ G | limU ℓrk (gi ) = 0} and N pr := {(gi ) ∈ G | limU ℓpr (gi ) = 0}
of G. We also define N0 := {(gi ) ∈ G | gi = 1Hi along U } and suitable
subgroups N1 , A1 ≤ G, defined in Section 4. Then,
(i) The subgroup N pr contains all proper normal subgroups of G containing N0 . Especially, G/N pr is non-abelian simple and N pr /N0 is a
characteristic subgroup of G/N0 .
(ii) The normal subgroups of G lying between N0 and N rk are linearly
ordered. Moreover, each such normal subgroup is perfect.
(iii) Define mappings between the following two sets
{N E G | N1 ≤ N ≤ N pr }
Ψ
Φ
{M | M E G, N1 ≤ M ≤ N rk } × {A | A E G, N1 ≤ A ≤ A1 }
by Φ : N 7→ (N ∩ N rk , N ∩ A1 ) and Ψ : (M, A) 7→ M A. Then Φ and
Ψ are isomorphisms of posets and mutually inverse to each other.
Throughout, we use the notation and basic facts from [4].
2. Auxiliary geometric results
In this section we provide the necessary geometric results for the rest of
the article. In the following let H be a quasisimple group from the following
list:
(i) linear: SLn (q), n ≥ 2, (n, q) 6= (2, 2), (2, 3);
(ii) symplectic: Sp2m (q), m ≥ 2, (m, q) 6= (2, 2);
(iii) unitary: SUn (q), n ≥ 3, (n, q) 6= (3, 2);
(iv) orthogonal: Ω2m+1 (q), m ≥ 3, q odd; Ω±
2m (q), m ≥ 4.
Here we omit the groups Ω2m+1 (q) with q even as they are isomorphic to
the groups Sp2m (q). We have also omitted to mention the spin groups, i.e.
the double covers of Ω±
2n (q) and Ω2n+1 (q), as they can be treated directly in
a way similar to the double cover of the alternating groups, see Remark 10.
Let V be the natural representation of H (there are two such representations when H = Ω2n+1 (q) for q odd, corresponding to the two equivalence
classes of non-degenerate symmetric bilinear forms in this dimension) and
4
JAKOB SCHNEIDER AND ANDREAS THOM
n := dim V . For (ii)–(iv), denote by f the corresponding non-degenerate
alternating, conjugate-symmetric sesquilinear, or symmetric bilinear form.
In (iv) for q even, denote by Q the corresponding quadratic form inducing
the non-degenerate alternating form f .
A subspace U of V is called non-singular if f |U is non-degenerate. In
this case U ⊕ U ⊥ = V . The following fact will be used in the Section 4.
Lemma 2. If U ≤ V is a non-singular subspace with 2 ≤ dim U < n/2,
then there exists a decomposition W1 ⊕ W2 = U ⊥ and an element h ∈ H
such that h is the identity on W2 and interchanges U and W1 .
Proof. Explicit computations with standard bases show that U ⊥ always contains a subspace W1 isometric to U via an isometry θ : U → W1 with
respect to f or Q in the orthogonal case in characteristic two (here we need
dim U ⊥ > dim U ). Set W2 := (U ⊕ W1 )⊥ . Then h1 : V = U ⊕ W1 ⊕ W2 →
U ⊕ W1 ⊕ W2 = V given by (u, w1 , w2 ) 7→ (θ −1 (w1 ), θ(u), w2 ) is an isometric
involution of V .
Hence, in the symplectic case (ii) we may take h := h1 . In the unitary
case (iii), if det(h1 ) = ε ∈ {±1}, letting h2 be the linear map which scales a
non-isotropic vector of U by ε and fixes its perpendicular complement, then
h := h1 h2 works.
In the orthogonal case (iv) for q odd, define h2 as in the unitary case. If
the spinor norm of h1 h2 is ε ∈ {±1}, find an element s ∈ SO(U ) of spinor
norm ε and set h3 := s ⊕ idU ⊥ (for the existence of s we use that dim U ≥ 2).
Then h := h1 h2 h3 works.
In the orthogonal case for q even (forcing dim V to be even), if the
quasideterminant of h1 is ε ∈ {±1}, find an element s ∈ GO(U ) of quasideterminant ε and set h4 := s ⊕ idU ⊥ (again using dim U ≥ 2). Then h := h1 h4
is an appropriate choice. The proof is complete.
The following simple result about the existence of certain non-singular
subspaces will also be used in later sections.
Lemma 3. Let U ≤ V , dim U = l. Then there exists W ≤ U nonsingular such that dim W ≥ 2l − n.
Proof. Chose W ≤ U maximal non-singular. If there were two vectors
u, v ∈ W ⊥ ∩ U with f (u, v) 6= 0, then W ⊕ hu, vi > W would still be nonsingular. Hence f is zero on this subspace. Moreover, for dimension reasons
U = W ⊕ (W ⊥ ∩ U ). Together this implies U ⊥ ∩ U = W ⊥ ∩ U . Hence
dim(W ⊥ ∩ U ) = dim U − dim W ≤ dim U ⊥ = n − l,
so dim W ≥ 2l − n as desired.
ON THE NORMAL SUBGROUP LATTICE
5
We introduce the group of quasiscalars S(H) of H by
S(SLn (q)) := F∗ ∼
= Cq−1 ;
q
S(SUn (q)) := {x ∈ F∗q2 | xq+1 = 1} ∼
= Cq+1 ;
S(Sp2m ) := {±1};
S(Ω2m+1 (q)) := {±1} (q odd);
S(Ω±
2m (q)) := {±1}.
Of course −1 = 1 if char(Fq ) = 2. Our last auxiliary result will be used in
Section 4:
Lemma 4. For any λ ∈ S(H) there is a diagonalizable element in h ∈ H
with respect to V and a suitable basis such that all but two of its diagonal
entries are equal to λ.
Proof. Clearly, we may assume that λ 6= 1, since otherwise we can
always take h = id. If H = SLn (q), take h = diag(λ, . . . , λ, λ−(n−1) ) with
respect to any basis of V . If H = Sp2m (q), q odd, and λ = −1, take
− idV . If H = SUn (q), take h = diag(λ, . . . , λ, λ−(n−1) ) with respect to an
orthonormal basis e1 , . . . , en for f , i.e. f (ei , ej ) = δij for i, j = 1, . . . , n.
For Ω2m+1 (q), q odd, and λ = −1, take diag(−1, . . . , −1, 1) with respect
to a basis e1 , . . . , e2m+1 such that f (ei , ej ) = 0 if i 6= j and f (ei , ei ) = 1 for
i = 1, . . . , 2m and f (e2m+1 , e2m+1 ) = 1 or a non-square α ∈ F∗q (there are
two equivalence classes of such forms both giving a natural representation
of Ω2n+1 (q)). Clearly h has determinant and spinor norm equal to one.
For H = Ω±
2n (q), q odd, and λ = −1, we find a basis e1 , . . . , e2n such
that either f (ei , ej ) = δij for all i, j = 1, . . . , 2n, or f (ei , ej ) = 0 for
i 6= j, f (ei , ei ) = 1 for i = 1, . . . , 2n − 1 and f (e2n , e2n ) = α ∈ F∗q a
non-square (there are two equivalence classes of non-degenerate symmetric
bilinear forms corresponding to the two non-isomorphic groups Ω+
2n (q) and
−
Ω2n (q)). In either case we can take h = diag(−1, . . . , −1, 1, 1) with respect
to this basis, which has determinant and spinor norm one.
In all remaining cases we have S(H) = 1, so the proof is complete.
3. Relative bounded generation in universal finite quasisimple
groups
In this section we keep the notation from Section 2. In addition, we
define the following three length functions on our group H: The rank length
function ℓrk (h) := rk(1 − h)/n, the projective rank length function ℓpr (h) :=
min{ℓrk (λ−1 h) | λ ∈ F∗q or F∗q2 } (q 2 when H is unitary), and the conjugacy
length function ℓc (h) := log|hH |/ log|H|.
In the following we will prove a version of ‘relative bounded generation’
for all universal quasisimple groups from families of unbounded rank (for the
others there is no such notion). We start with the alternating case. Recall
that the normalized Hamming length defined by ℓH (σ) := |supp(σ)|/n for
6
JAKOB SCHNEIDER AND ANDREAS THOM
σ ∈ An is Lipschitz equivalent to ℓc by [3, Theorem 2.11]. We denote the
en .
universal central extension of An by A
en ,
Lemma 5. There exists a constant c > 0 such that for any σ ∈ A
e
en ) for any integer k ≥ c max{ℓH (σ)/ℓH (τ ), 1} we have σ ∈ (τ An )∗k .
τ 6∈ Z(A
en under the canonical map.
Here σ, τ ∈ An are the images of σ, τ ∈ A
Proof. We prove the lemma for An (n ≥ 5) and derive the corresponding result for its Schur covering group. Let σ, τ ∈ An , τ 6= 1. After
conjugating σ, we may assume that either supp(σ) ⊆ supp(τ ) or the opposite inclusion holds. In the first case, σ ∈ Alt(supp(τ )) = (τ An )∗c1 for some
integral constant c1 > 0 and in the second case σ ∈ Alt(supp(σ)) = (τ An )∗k
H
H
for any integer k ≥ c2 /ℓH
Alt(supp(σ)) (τ ) = c2 ℓ (σ)/ℓ (τ ) with some constant
c2 > 0; both times we use [2, Lemma 3.2]. So we may take c := max{c1 , c2 }.
en by the following arThe statement now extends to the Schur cover A
gument. Without loss of generality, we may assume n > 7 (we may neglect
en is a twofold cover. It follows from the standard conthis finite data), so A
struction of the two double covers of Sn that when a and b are lifts of the
transposition (12) and (34), then (ab)2 = z is the unique non-trivial central
en ) ⊆ (τ Aen )∗c3 for any nonelement in one of these. This shows that Z(A
en and some absolute integer constant c3 > 0, which implies the
central τ ∈ A
e
claim for An .
Now we turn to the universal quasisimple groups from families of unbounded rank. The proof of ‘relative bounded generation’ for these is actually very similar as for the Schur covering groups of the alternating groups.
We need the following fact which can be deduced by adapting the proof
of Lemma 4.1 of [2] to quasisimple groups and looking at Lemma 5.4, 6.4,
and the end of Section 7 of the same article.
Lemma 6. There is an absolute constant c > 0 (independent of H) such
that for h 6∈ H \ Z(H), for any k ≥ c/ℓpr (h) it holds that H = (hH )∗k .
Remark 1. It follows from Lemma 5.3, 6.3, and the end of Section 7
of [2] that ℓpr and ℓc are Lipschitz equivalent by a universal Lipschitz constant, so putting ℓc in place of ℓpr in the Lemma 6 would also yield a valid
statement.
Here is now the promised result for almost all other universal quasisimple
groups from families of unbounded rank:
Lemma 7. Let ε > 0 be arbitrary. There exists an absolute constant
C > 0 and a constant D > 0 only depending on ε such that the following
∗k for
holds: Let h1 ∈ H \ Z(H), h2 ∈ H, ℓrk (h1 ) ≤ 1 − ε. Then h2 ∈ (hH
1 )
rk
rk
all integers k ≥ max{Cℓ (h2 )/ℓ (h1 ), D}.
Proof. First assume that ε ≤ ℓrk (h1 ) ≤ 1 − ε. Then by Proposition
2.13 of [3] it holds that ℓpr (h1 ) ≥ ε, so by Lemma 6 there is D ∈ N only
∗D = H.
depending on ε such that (hH
1 )
ON THE NORMAL SUBGROUP LATTICE
7
So we may assume w.l.o.g. that ℓrk (h1 ) < ε ≤ 1/8. Assume additionally
that ℓrk (h2 ) ≤ 1/8 as well (we will remove this assumption at the end). Set
Ui := ker(1 − hi ) ≤ V and li := dim Ui (i = 1, 2).
At first we treat the special linear case, i.e. H = SLn (q): Replace h1 , h2 ∈
H by conjugates in Jordan normal form (where one off-diagonal entry might
be different from 0 or 1) as matrices with respect to a suitable basis e1 , . . . , en
with all 1×1 Jordan blocks corresponding to eigenvalue one in the upper left
corner. Assuming there are mi of these for hi , i.e. hi e1 = e1 , . . . , hi emi = emi ,
it is easy to see that mi ≥ 2li − n (i = 1, 2). Set m := min{m1 , m2 } and
W ′ := hem+1 , . . . , en i. Choose n − m ≤ j ≤ m minimal such that for
X := hem−j+1 , . . . , em i, the space Y := X ⊕ W ′ gives rise to a quasisimple
group K := SL(Y ). This is possible since by assumption on l1 and l2 we
have m ≥ 3n/4 > n/2 and SL(V ) is quasisimple. Since dim W ′ ≥ 1, there
is an absolute constant d > 1 such that j ≤ d dim W ′ = d(n − m). Then set
Y ′ := he1 , . . . , em−j i. The operators h1 , h2 write as hi = idY ′ ⊕ idX ⊕Ai =
idY ′ ⊕Bi (i = 1, 2) with respect to te decompositions V = Y ′ ⊕ X ⊕ W ′ =
Y ′ ⊕ Y . Assuming m1 ≤ m2 gives (1 + d)−1 /2 ≤ ℓrk (B1 ) ≤ 1/2, so by
Proposition 2.13 of [3] we have ℓpr (B1 ) ≥ (1 + d)−1 /2 implying the existence
∗D (by
of a constant D ∈ N with B2 ∈ (B1K )∗D which yields h2 ∈ (hH
1 )
Lemma 6). If m2 < m1 the same computation shows (1 + d)−1 /2 < ℓrk (B2 )
and ℓrk (B1 ) ≤ 1/2, so as previously ℓpr (B1 ) = ℓrk (B1 ). By Lemma 6 there
is c > 0 such that for all integers k ≥ c/ℓpr (B1 ) we have K = (B1K )∗k . Then
it holds that h2 ∈ (H1H )∗k and by
2(1 + d)cℓrk (h2 )/ℓrk (h1 ) = 2(1 + d)cℓrk (B2 )/ℓrk (B1 ) > c/ℓpr (B1 )
we are done in this case with C := 2(1 + d)c.
In the other cases, i.e. H 6= SLn (q), the proof is almost identical: Define
U := U1 ∩ U2 and l := dim U ≥ l1 + l2 − n. From Lemma 3 we get W ≤ U
non-singular with dim W ≥ 2l − n. Then we infer dim W ⊥ ≤ 2(n − l) from
dim W ≥ 2l − n. This implies
dim W ⊥ ≤ 2(n − l) ≤ 2(2n − (l1 + l2 )) ≤ 4(n − min{l1 , l2 })
and
4 min{l1 , l2 } − 3n ≤ 2(l1 + l2 ) − 3n ≤ 2l − n ≤ dim W.
As by assumption ℓrk (h1 ), ℓrk (h2 ) ≤ 1/8, implying that 7/8n ≤ l1 , l2 , we
obtain dim W ⊥ ≤ dim W . Now take X ≤ W a non-singular subspace such
that dim X ≥ dim W ⊥ is as small as possible such that Y := X ⊕ W ⊥ gives
rise to a classical quasisimple group K from the beginning of Section 2. As
dim W ⊥ ≥ 1, then dim X ≤ d dim W ⊥ for some absolute d > 1. With respect to the decompositions V = Y ⊥ ⊕ X ⊕ W ⊥ = Y ⊥ ⊕ Y , the operators
h1 , h2 write as hi = idY ⊥ ⊕ idX ⊕Ai = idY ⊥ ⊕Bi with isometric automorphism Ai of W ⊥ and Bi of Y (i = 1, 2). Now assume l1 ≤ l2 . Then
ℓrk (B1 ) =
rk(1 − A1 )
,
dim W ⊥ + dim X
8
JAKOB SCHNEIDER AND ANDREAS THOM
which can be bounded from above by 1/2 and from below by the chain
n − l1
n − l1
1
n − l1
≥
≥
=
.
2(1 + d)(n − l)
2(1 + d)(2n − (l1 + l2 ))
4(1 + d)(n − l1 )
4(1 + d)
Hence by Proposition 2.13 of [3] we get that ℓpr (B1 ) ≥ (1 + d)−1 /4, so by
Lemma 6 there is a constant D ∈ N such that B2 ∈ (B1K )∗D implying that
∗D
h2 ∈ (hH
1 ) . On the other hand, if l2 < l1 , by the same computation as
above, 0.25(1 + d)−1 < ℓrk (B2 ) and ℓrk (B1 ) ≤ 1/2, so by Proposition 2.13
of [3] ℓpr (B1 ) = ℓrk (B1 ). Applying Lemma 6 gives c > 0 such that for all
∗k for such
integers k ≥ c/ℓpr (B1 ) we have K = (B1K )∗k . But then h2 ∈ (hH
1 )
k, and since
4(1 + d)cℓrk (h2 )/ℓrk (h1 ) = 4(1 + d)cℓrk (B2 )/ℓrk (B1 ) > c/ℓpr (B1 )
we are done in this case with C := 4(1 + d)c.
We still need to eliminate the condition ℓrk (h2 ) ≤ 1/8. But this is again
just another straightforward application of Lemma 6, since the elements
h ∈ H with ℓrk (h) > 1/8 generate H ‘quickly’.
Remark 2. The condition ℓrk (h1 ) ≤ 1 − ε for a fixed ǫ > 0 cannot be
removed by Example 1. In that sense, the previous result is best possible.
Remark 3. The only universal quasisimple groups from families of unbounded rank which are not covered by Lemma 5 and 7 are the double covers
of the orthogonal groups in odd characteristic.
Defining ℓrk (h) := ℓrk (h) for h an element of the twofold cover of Ω±
2m (q)
or Ω2m+1 (q) (q odd), the statement of Lemma 7 also holds for these. This
e5 is embedded in both of them as the group of products of
is true since A
an even number of the reflections with respect to the vectors e1 − e2 , e2 −
e3 , e3 − e4 , e4 − e5 ∈ V , where e1 , . . . , e5 is an orthonormal system for f .
So we can argue as at the end of the proof of Lemma 5 to see that the
two-element kernel of the covering map is generated ‘quickly’ by conjugates
of any non-central element.
4. The lattice of normal subgroups of an algebraic ultraproduct
of classical finite quasisimple groups
Let (Hi )i∈I be a sequence
of groups from the list at the beginning of
Q
Section 2 and set G := i∈I Hi . Let ni be the dimension of the ‘natural’
representation Vi of Hi . Note that the sequence (ni ) is Lipschitz equivalent
to the sequence of Lie ranks of the Hi by a universal Lipschitz constant. For
some ultrafilter U on I for which limU ni = ∞, define the normal subgroup
N0 := {(gi ) ∈ G | gi = 1Hi along U } of G. In this section we give a complete
description of the lattice of normal subgroups of the algebraic ultraproduct
G/N0 of the groups Hi (i ∈ I) with respect to U .
Define the subgroups
N rk := {(gi ) ∈ G | lim ℓrk (gi ) = 0} and
U
N pr := {(gi ) ∈ G | lim ℓpr (gi ) = 0}
U
ON THE NORMAL SUBGROUP LATTICE
9
of G where ℓrk and ℓpr are the length functions defined at the beginning of
Section 3. As they are invariant length functions, it is clear that N rk and
N pr are normal in G and contain N0 . Moreover, as ℓpr ≤ ℓrk , we get that
N rk ≤ N pr .
The following result is an immediate consequence of Lemma 6:
Lemma 8. The subgroup N pr contains all proper normal subgroups of
G containing N0 . Especially, G/N pr is non-abelian simple and N pr /N0 is a
characteristic subgroup of G/N0 .
Proof. When g = (gi ) ∈ G\N pr , there exist ε > 0 and U ∈ U such that
ℓpr (gi ) ≥ ε for i ∈ U . Hence from Lemma 6 it follows that there is k ∈ N
such that (giHi )∗k = Hi for i ∈ U . This implies hhgiiN0 = G as wished.
Hence from now on, we may restrict to the normal subgroups above
N0 which are contained in N pr . Let us first characterize N pr among the
subgroups of G containing N rk . For this recall the definition of S(H) from
the end of Section 2.
Q
Lemma 9. The map ϕ : N pr → Z := U S(Hi ) defined by (gi ) 7→ (λi ),
where λi is arbitrary (from S(Hi )) if ℓpr (gi ) ≥ 1/4 and λi is the unique
λ ∈ F∗qi for which ℓrk (λ−1 gi ) < 1/4 otherwise, is a surjective homomorphism
with kernel N rk . Moreover, Z ∼
= N pr /N rk = Z(G/N rk ).
Proof. At first we check that λi is always in S(Hi ): This is clear when
ℓpr (gi ) ≥ 1/4. In the special linear case there is nothing to check, so consider
the remaining cases. Set Ui to be the eigenspace corresponding to eigenvalue
λi of gi and set dim Ui := li > 3/4ni . Then by Lemma 3 there is Wi ≤ Ui
non-singular such that dim Wi ≥ 2li − ni > ni /2 ≥ 1. The restriction gi |Wi
is a scalar from the group GI(Wi , fi |Wi ), so it must lie in S(Hi ).
Let us now show that ϕ is a homomorphism.
So let (gi ), (hi ) ∈ N pr and
Q
pick U ∈ U and sequences (λi ), (µi ) from i S(Hi ) such that
rk −1
ℓrk (λ−1
i gi ), ℓ (µi hi ) < 1/8
for i ∈ U (this is possible by the definition of N pr ). Then it follows that
ℓrk ((λi µi )−1 gi hi ) < 1/4 for i ∈ U by the triangle inequality, establishing
that ϕ is a homomorphism.
Also ϕ is surjective: Let λ = (λi ) ∈ Z. In any case, by Lemma 4 there
is a diagonal matrix in Hi with all but at most two entries of the diagonal
equal to λi . This is the desired preimage of λ.
Moreover, the kernel of ϕ consists of sequences g = (gi ) such that when
(λi ) is the image of g under ϕ, then λi = 1 for i ∈ U and some U ∈
U . But this means precisely that g ∈ N rk , since then 0 = limU ℓpr (gi ) =
rk
limU ℓrk (λ−1
i gi ) = limU ℓ (gi ).
Lastly, we have to show that N pr /N rk is the center of G/N rk . Since by
Lemma 8 the quotient G/N pr is non-abelian simple, it suffices to show that
[G, N pr ] ⊆ N rk . So let g = (gi ) ∈ G and h = (hi ) ∈ N pr and let (λi ) be
10
JAKOB SCHNEIDER AND ANDREAS THOM
rk
sequence in F∗qi such that (λ−1
i hi ) ∈ N . Let Ui be the eigenspace of hi
corresponding to eigenvalue λi and set Wi := Ui ∩ gi−1 Ui . Then for w ∈ Wi
one computes that [gi , hi ]w = gi−1 h−1
i gi hi w = w. As dim Wi /ni tends to one
along U , this verifies that [g, h] ∈ N rk . The proof is complete.
To describe the normal subgroups N of N pr containing N0 accurately,
we need some definitions: Define L to be the set of null sequences (ri ) along
U with ri ∈ [0, 1] (i ∈ I). Two such sequences (ri ) and (si ) are said to be
equivalent (write (ri ) ∼ (si ) and (ri ) for the corresponding equivalence class)
if limU ri /si ∈ (0, ∞). For two elements r and s of the quotient L/∼ write
r ≤ s if an only if limU ri /si < ∞, where (ri ) and (si ) are representatives
for r and s, respectively. It is routine to check that this is a definition and
turns (L/∼, ≤) into a linear order.
Now define the function ct : N pr → L/∼ by g = (gi ) 7→ (ℓpr (gi )) and
call it the convergence type of g. Denote by L the subset of elements r of
L/∼ for which either r ≥ (1/ni ) or r = (0).
Lemma 10. The image of the function ct : N pr → L/∼ is equal to L.
Proof. At first we prove that the image of ct lies in L: Namely when
r = (ri ) ∈ N pr and r 6= (0), then ri 6= 0 for all i ∈ U for some U ∈ U . But
then it follows that ri ≥ 1/ni for these i and hence r ≥ (1/ni ).
The surjectivity is a bit more subtle. We prove it here only for the
case that Hi = SLni (qi ); the other cases are analogous. Choose a nontrivial element g ∈ SL2 (q). Then if r = (ri ) satisfies ri ≥ 1/ni along U , set
gi := g⊕⌊ni ri ⌋ ⊕ idni −2⌊ni ri ⌋ if ni ≥ 2 ⌊ni ri ⌋ and choose an arbitrary element
otherwise. Letting g = (gi ), one sees immediately that ct(g) = r. Finally,
ct(1G ) = (0) completes the proof.
Remark 4. The rank length function ℓrk does not attain any possible
value on the classical quasisimple groups even for large ranks: E.g. 2nℓrk (g)
for g ∈ GOε2n (q) is even if and only if g ∈ Ωε2n (see [4, p. 77] at the end of
Section 3.8.1).
Due to the previous lemma, henceforth we shall consider ct as a function
with domain N pr and codomain L. We extend this function to subsets
S ⊆ N pr by setting ct(S) := {ct(g) | g ∈ S}. For a normal subgroup N of G
between N0 and N pr the set ct(N ) is then called the associated order ideal
(see Lemma 12 below).
Lemma 11. For such a subgroup it holds that ct(N ) = ct(N ∩ N rk ).
Proof. It is clear that ct(N ) ⊇ ct(N ∩ N rk ). Let us prove the converse
containment to finish the proof: Let g = (gi ) ∈ N ≤ N pr . Then there
is U ∈ U and λi ∈ F∗qi (i ∈ I) such that ℓpr (gi ) = ℓrk (λ−1
i gi ) < 1/4 for
i ∈ U and so when Ui is the eigenspace to λi for such an i, then li :=
dim Ui > 3/4ni . By Lemma 3 there is a non-singular subspace Wi ≤ Ui
with dim Wi ≥ 2li − ni > ni /2. Moreover, for all large ni we may also
ON THE NORMAL SUBGROUP LATTICE
11
assume that ni − dim Wi = dim Wi⊥ ≥ 2 (by modifying Wi a little). Now
Lemma 2 implies the existence of hi ∈ Hi (i ∈ I) and Wi1 , Wi2 ≤ Wi (i ∈ U )
non-singular such that Wi⊥ ⊕ Wi1 ⊕ Wi2 is a perpendicular direct sum and
hi restricts to the identity on Wi2 while it interchanges Wi⊥ and Wi1 . Then
−1 −1
−1 −1
−1 hi −1
it holds that [gi , hi ](h−1
i w) = gi hi gi w = λi hi gi w = λi gi (hi w) and
−1
⊥
[gi , hi ]w = gi−1 h−1
i gi hi w = λi gi w for w ∈ Wi and i ∈ U . This implies
that ℓrk ([gi , hi ]) = 2ℓpr (gi ) < 1/2. But then by Proposition 2.13 of [3] we
have ℓpr ([gi , hi ]) = ℓrk ([gi , hi ]) and by the end of Lemma 9, setting h := (hi ),
the commutator [g, h] is in N rk (and of course in N ). This element has the
same convergence type as g.
Remark 5. In view of Lemma 10, this implies that even the restriction
ct|N rk : N rk → L is surjective.
Now we are ready to justify the name ‘associated order ideal’ for ct(N )
for N normal in G lying between N0 and N pr .
Lemma 12. Let N be a normal subgroup between N0 and N pr . Then
ct(N ) is an order ideal of L. Moreover, the mappings
{N E G | N0 ≤ N ≤ N rk }
α
{order ideals of (L, ≤)}
β
defined by α : N 7→ ct(N ) and β : I 7→ {g ∈ N rk | ct(g) ∈ I} are isomorphisms of posets and mutually inverse to each other.
Proof. By Lemma 11, we may restrict to the case N ≤ N rk . Clearly,
ct(N ) is not empty as 1G ∈ N . If r ∈ ct(N ), Lemma 7 implies that if s ≤ r
then s ∈ ct(N ). As (L, ≤) was a linear order, ct(N ) is an order ideal.
Concerning the second part: Obviously, both maps are inclusion preserving. To show that β ◦ α : N 7→ ct(N ) 7→ {g ∈ N rk | ct(g) ∈ ct(N )} is
the identity, i.e. that ct(g) ∈ ct(N ) iff g ∈ N , is just another straightforward
application of Lemma 7. The map α ◦ β : I 7→ {g ∈ N rk | ct(g) ∈ I} 7→
{ct(g) ∈ L | ct(g) ∈ I} is the identity by Remark 5.
Remark 6. Similarly, one can prove that for g ∈ N rk and S ⊆ N rk it
holds that ct(g) is in the order ideal of (L, ≤) generated by ct(S) if and only
if g ∈ hhSiiN0 .
Corollary 1. The normal subgroups of G lying between N0 and N rk
are linearly ordered.
Proof. This holds by the correspondence of Lemma 12 as (L, ≤) is
linear and so its ideals are linearly ordered by inclusion.
Corollary 2. For N E N rk containing N0 it holds that N is perfect,
i.e. N = N ′ .
Proof. Pick g ∈ N and proceed as in the proof of Lemma 11. The
construction of h shows that ct(h) ≤ ct(g), so Remark 6 implies h ∈
12
JAKOB SCHNEIDER AND ANDREAS THOM
hhgiiN0 ⊆ N . Hence the commutator [g, h] lies in N ′ and ct([g, h]) = ct(g),
so ct(N ′ ) = ct(N ). To apply Lemma 12 to deduce that N = N ′ , we still
need that N0 ≤ N ′ . It is therefore enough to show that N0 is perfect. This
follows from the following simple application of Lemma 6: Let h ∈ H be
an element with ℓpr (h) ≥ ε for some fixed ε > 0, where H is one of the Hi .
Then the previously mentioned result implies that H ′ ⊇ (hH )∗k h−k = H for
a fixed integer k only depending on ε. Hence, if g = (gi ) ∈ N0 , applying this
in every component i for which gi 6= 1Hi yields that g ∈ N0′ as wished.
Remark 7. The argument at the end of the proof also shows that G
itself is perfect.
Remark 8. Especially, this implies that for N ≤ N pr it holds that
N ∩ N rk = [G, N ] = N ′ by the end of Lemma 9 and Corollary 2. Hence
N rk = (N pr )′ and so N rk /N0 is characteristic in G/N0 as well.
Using the correspondence of Lemma 12, we introduce the normal subgroup N1 as the normal subgroup of N rk such that ct(N1 ) = I1 , where
I1 := {(ri ) ∈ L | ni ri < C for some C > 0 along U }. It is obvious that I1
covers the trivial ideal I0 = {(0)} corresponding to N0 , i.e. there is no ideal
properly between them. Hence N1 covers N0 . Let us mention here that for
M, N normal subgroups of G between N0 and N rk it holds that M is covered
by N if and only if the corresponding order ideals I and J of (L, ≤) are of
the form I = {r ∈ L | r < s} and J = {r ∈ L | r ≤ s} for some s 6= (0).
Because we will need them later, we introduce the normal subgroups
A0 := {g ∈ N pr | ct(g) ∈ I0 } and A1 := {g ∈ N pr | ct(g) ∈ I1 }.
Now we complete the picture of the lattice of normal subgroups of G/N0
by looking at an arbitrary N E N pr with N0 ≤ N .
Let us at first assume that N1 ≤ N . Then the following lemma applies:
Lemma 13. Define mappings between the following two sets
{N E G | N1 ≤ N ≤ N pr }
Ψ
Φ
{M | M E G, N1 ≤ M ≤ N rk } × {A | A E G, N1 ≤ A ≤ A1 }
by Φ : N 7→ (N ∩ N rk , N ∩ A1 ) and Ψ : (M, A) 7→ M A. Then Φ and Ψ are
isomorphisms of posets and mutually inverse to each other.
Proof. Obviously, both maps are inclusion preserving. Moreover, Φ ◦
Ψ(M, A) = Φ(M A) = (M A ∩ N rk , M A ∩ A1 ) = (M (A ∩ N rk ), A(M ∩ A1 ))
by Dedekind’s modular law. However, it is easy to see from the definitions
that A ∩ N rk = M ∩ A1 = N1 , so Φ ◦ Ψ = id as N1 ≤ M, A.
Similarly, we compute Ψ ◦ Φ(N ) = Ψ(N ∩ N rk , N ∩ A1 ) = (N ∩ N rk )(N ∩
A1 ) ⊆ N . So it is enough to show that every element g = (gi ) ∈ N can be
written as a product of elements from N ∩N rk and N ∩A1 . Choose λi (i ∈ I)
pr
such that ℓrk (λ−1
i g) = ℓ (gi ). Let hi ∈ Hi be an element which has all but
ON THE NORMAL SUBGROUP LATTICE
13
two diagonal entries equal to λi (which exists by Lemma 4). Then setting
f := (gi h−1
i ) and h := (hi ), g = f h is the desired product decomposition.
Remark 9. Essentially, Lemma 13 (in combination with Lemma 12)
says that the lattice of normal subgroups of G/N1 is isomorphic to the product of the linear order of ideals of (L, ≤) different from I0 and the subgroup
lattice of the abelian group Z ∼
= A1 /(A1 ∩ N rk ) =
= N pr /N rk = N rk A1 /N rk ∼
A1 /N1 from Lemma 9. If Φ : N 7→ (M, A), then N 7→ (ct(M ), ϕ(A)) is the
described isomorphism of lattices, where ϕ is the map from Lemma 9.
In this situation, M = N ′ is the abelianization (by Remark 8 above)
and ϕ(A) = ϕ(M A) = ϕ(N ) ∼
= N/(N ∩ N rk ) = N/N ′ is the
= N N rk /N rk ∼
universal abelian quotient.
Now assume that N1 6≤ N . Then it follows from Lemma 11 that ct(N ) =
ct(N ∩ N rk ) = ct(N0 ) = {(0)} = I0 (as by assumption N0 ≤ N ). Hence we
have established the following result.
Lemma 14. If N is normal in G containing N0 and N1 6≤ N , then
N ≤ A0 .
Finally, we describe when a normal subgroup as in Lemma 14 is contained in a normal subgroup as in Lemma 13:
Lemma 15. A normal subgroup N containing N0 but not N1 is contained
in a normal subgroup K ≤ N pr containing N1 if and only if N ≤ A (or
equivalently ϕ(N ) ≤ ϕ(A), with ϕ the map from Lemma 9), where Φ(K) =
(M, A) is from Lemma 13.
Proof. Assume N ≤ K. Then by Lemma 14 it holds that N ≤ K ∩
A0 ≤ K ∩ A1 = A implying ϕ(N ) ≤ ϕ(A). Conversely, by Lemma 9,
the assumption ϕ(N ) ≤ ϕ(A) implies N N rk ≤ AN rk . Thus N N rk ∩ A1 =
N (N rk ∩A1 ) = N N1 ≤ AN rk ∩A1 = A(N rk ∩A1 ) = AN1 = A by Dedekind’s
modular law. So N ≤ A ≤ K, completing the proof.
We will end the article with a few remarks.
Remark 10. Again, we have not yet covered the case that the Hi (i ∈ I)
are double covers of Ω±
odd). In this case, define N0
2m (q) or Ω2m+1 (q) (q Q
as above and let A0 be the elements of G = i Hi which are central along
the ultrafilter U . Then G/N0 is a twofold cover of an ultraproduct G/M0 of
orthogonal groups with kernel M0 /N0 ∼
= C2 .
It follows from Remark 3 that if N0 ≤ N E G contains an element
non-central modulo N0 , i.e. it is not contained in A0 , then M0 ≤ N , so it
corresponds to a normal subgroup of G/M0 , which we have described.
Finally, when N0 ≤ N ≤ A0 with M0 6≤ N and M0 ≤ K ≤ G, then
N ≤ K iff N M0 ≤ K and this again can be decided by considering the
lattice of normal subgroups of G/M0 , which we discussed.
Using the same argument, one can show that the normal subgroups of
an ultraproduct of the double covers of simple alternating groups are still
linearly ordered by inclusion (since here A0 = M0 ).
14
JAKOB SCHNEIDER AND ANDREAS THOM
Acknowledgments
The results of this article are part of the PhD thesis of the first author.
This research was supported by ERC Consolidator Grant No. 681207.
References
[1] Philip Dowerk and Andreas Thom, Bounded Normal Generation and Invariant Automatic Continuity, arXiv:1506.08549 (2015), submitted. ↑1
[2] Martin W. Liebeck and Aner Shalev, Diameters of finite simple groups: sharp bounds
and applications, Annals of Mathematics (2001), 383–406. ↑2, 6
[3] Abel Stolz and Andreas Thom, On the lattice of normal subgroups in ultraproducts of
compact simple groups, Proceedings of the London Mathematical Society 108 (2014),
no. 1, 73–102. ↑1, 6, 7, 8, 11
[4] Robert Wilson, The finite simple groups, Vol. 251, Springer Science & Business Media,
2009. ↑3, 10
J. Schneider, TU Dresden, 01062 Dresden, Germany
E-mail address: [email protected]
A. Thom, TU Dresden, 01062 Dresden, Germany
E-mail address: [email protected]
| 4math.GR
|
Static Analysis for Regular Expression
Denial-of-Service Attacks
James Kirrage
Asiri Rathnayake
Hayo Thielecke
arXiv:1301.0849v1 [cs.PL] 4 Jan 2013
University of Birmingham, UK
Abstract. Regular expressions are a concise yet expressive language
for expressing patterns. For instance, in networked software, they are
used for input validation and intrusion detection. Yet some widely deployed regular expression matchers based on backtracking are themselves
vulnerable to denial-of-service attacks, since their runtime can be exponential for certain input strings. This paper presents a static analysis for
detecting such vulnerable regular expressions. The running time of the
analysis compares favourably with tools based on fuzzing, that is, randomly generating inputs and measuring how long matching them takes.
Unlike fuzzers, the analysis pinpoints the source of the vulnerability and
generates possible malicious inputs for programmers to use in security
testing. Moreover, the analysis has a firm theoretical foundation in abstract machines. Testing the analysis on two large repositories of regular
expressions shows that the analysis is able to find significant numbers of
vulnerable regular expressions in a matter of seconds.
1
Introduction
Regular expression matching is a ubiquitous technique for reading and validating input, particularly in web software. While pattern matchers are among the
standard techniques for defending against malicious input, they are themselves
vulnerable. The root cause of the vulnerability is that widely deployed regular
expression matchers, like the one in the Java libraries, are based on backtracking
algorithms, rather than the construction of a Deterministic Finite Automaton
(DFA), as used for lexers in compiler construction [11,2]. One reason for relying
on backtracking rather than a DFA construction is to support a more expressive
pattern specification language commonly referred to as “regexes”. Constructs
such as back-references supported by such regex languages go beyond regular
and even context-free languages and are known to be computationally expensive [1]. However, even if restricted to purely regular constructs, backtracking
matchers may have a running time that is exponential in the size of the input [6],
potentially causing a regular expression denial-of-service (ReDoS) attack [17].
It is this potentially exponential runtime on pure regular expressions (without
backreferences) that we are concerned about in this paper. Part of our motivation is that, for purely regular expressions, the attack could be defended against
by avoiding backtracking matchers and using more efficient techniques [7,24]
instead.
For a minimalistic example [6], consider matching the regular expression
a** against the input string a. . . a b, with n repetitions of a. A backtracking
matcher takes an exponential time [6] in n when trying to find a match; all
matching attempts fail in the end due to the trailing b. For such vulnerable
regular expressions, an attacker can craft an input of moderate size which causes
the matcher to take so long that for all practical purposes the matcher fails to
terminate, leading to a denial-of-service attack. Here we assume that the regular
expression itself cannot be manipulated by the attacker but that it is matched
against a string that is user-malleable.
While the regular expression a** as above is contrived, one of the questions
we set out to answer is how prevalent such vulnerable expressions are in the
real world. As finding vulnerabilities manually in code is time consuming and
error-prone, there is growing interest in automated tools for static analysis for
security [12,5], motivating us to design an analysis for ReDoS.
Educating and warning programmers is crucial to defending against attacks
on software. The standard coverage of regular expressions in the computer science curriculum, covering DFAs in courses on computability [11] or compiler
construction [2], is not necessarily sufficient to raise awareness about the possibility of ReDoS. Our analysis constructs a series of attack strings, so that
developers can confirm the exponential runtime for themselves.
This paper makes the following contributions:
1. We present an efficient static analysis for DoS on pure regular expressions.
2. The design of the tool has a firm theoretical foundation based on abstract
machines [18] and derivatives [4] for regular expressions.
3. We report finding vulnerable regular expressions in the wild.
In Section 2, we describe backtracking regular expression matchers as abstract machines, so that we have a precise model of what it means for a matching
attempt to take an exponential number of steps. We build on the abstract machine in designing our static analysis in Section 3, which we have implemented
in OCaml as described in Section 4. Experimental results in testing the analysis
on two large corpora of regular expressions are reported in Section 5. Finally,
Section 6 concludes with a discussion of related work and directions of further
research. The code of the tool and data sets are available at this URL:
http://www.cs.bham.ac.uk/~hxt/research/rxxr.shtml
2
Regular expression matching by backtracking
This and the next section present the theoretical basis for our analysis. Readers
primarily interested in the results may wish to skim them.
We start with the following minimal syntax for regular expressions:
e ::= e1 | e2
Alternation
e∗
Kleene star
e1 · e2
Concatenation
a
Constant, where a is an input symbol
The · in concatenation e1 · e2 is usually omitted, except when it is useful for emphasis, as in a syntax tree. Following the usual parser construction methods [2],
we can define a parser which is capable of transforming (parsing) a given regular
expression into an AST (abstract syntax tree) which complies with the above
grammar. As an example, the AST constructed by such a parser for the regular
expression (a | b)∗c can be visualized in the following manner:
p0
p
·
p1
p5
∗
c
p2
|
p3
p4
a
b
π(p)
p0 p1 · p5
p1
p2 ∗
p2 p3 | p4
p3
a
p4
b
p5
c
Notice that we have employed a pointer notation to illustrate the AST structure; this is quite natural given that in most programming languages, such an
AST would be defined using a similar pointer-based structure definition. Each
node of this AST corresponds to a unique sub-expression of the original regular
expression, the relationships among these nodes are given on the table to the
right. We have used the notation π(p) to signify the dereferencing of the pointer
p with respect to the heap π in which the above AST is constructed. A formal
definition of π was avoided in order to keep the notational clutter to a minimum,
interested readers may refer [18] for a more precise definition of π.
Having parsed the regular expression into an AST, the next step is to construct an NFA structure that allows us to define a backtracking pattern matcher.
While there are several standard NFA construction techniques [2], we opt for a
slightly different construction which greatly simplifies the rest of the discussion.
The idea is to associate a continuation pointer cont with each of the nodes in the
AST such that cont points to the following (continuation) expression for each
of the sub-expressions in the AST. In other words, cont identifies the “next subexpression” which must be matched after matching the given sub-expression.
More formally, cont is defined as follows:
Definition 1. Let cont be a function
cont : dom(π) → (dom(π) ∪ {null})
Such that,
–
–
–
–
If π(p) = (p1 | p2 ), then cont p1 = cont p and cont p2 = cont p
If π(p) = (p1 · p2 ), then cont p1 = p2 and cont p2 = cont p
If π(p) = (p1 )∗, then cont p1 = p
cont p0 = null, where p0 is the pointer to the root of the AST.
The following example illustrates the NFA constructed this way for the regular
expression (a | b)∗c:
null
p
p0
·
p1
p5
∗
c
p2
|
p3
p4
a
π(p)
p0 p1 · p5
p1
p2 ∗
p2 p3 | p4
p3
a
p4
b
p5
c
cont p
null
p5
p1
p1
p1
null
b
Here the dashed arrows identify the cont pointer for each of the AST nodes.
Readers familiar with Thompson’s construction [24,2] will realize that the resulting NFA is a slightly pessimized version of that resulting from Thompson’s
algorithm. The reason for this pessimization is purely of presentational nature; it
helps to visualize the NFA as an AST with an overlay of a cont pointer mesh so
that the structure of the original regular expression is still available in the AST
portion. Furthermore, this presentation allows the definitions and proofs to be
presented in an inductive fashion with respect to the structure of the expressions.
With the NFA defined, we present a simple non-deterministic regular expression matcher in the form of an abstract-machine called the PWπ machine:
Definition 2. A configuration of the PWπ machine consists of two components:
hp ; wi
The p component represents the current sub-expression (similar to a code pointer)
while w corresponds to the rest of the input string that remains to be matched.
The transitions of this machine are as follows:
hp ; wi → hp1 ; wi if π(p) = (p1 | p2 )
hp ; wi → hp2 ; wi if π(p) = (p1 | p2 )
hp ; wi → hq ; wi if π(p) = p1 ∗ ∧ cont p = q
hp ; wi → hp1 ; wi if π(p) = p1 ∗
hp ; wi → hp1 ; wi if π(p) = (p1 · p2 )
hp ; awi → hq ; wi if π(p) = a ∧ cont p = q
hp ; wi → hq ; wi if π(p) = ε ∧ cont p = q
The initial state of the PWπ machine is hp0 ; wi, where p0 is the root of the AST
corresponding to the input expression and w is the input string. The machine
may terminate in the state hnull ; w′′ i where it has matched the original regular
expression against some prefix w′ of the original input string w such that w =
w′ w′′ . Apart from the successful termination, the machine may also terminate if
it enters into a configuration where none of the above transitions apply.
The PWπ machine searches for a matching prefix by non-deterministically making a choice whenever it has to branch at alternation or Kleene nodes. While
this machine is not very useful in practice, it allows us to arrive at a precise
model for backtracking regular expression matchers. Backtracking matchers operate by attempting all the possible search paths in order; this allows us to model
them with a stack of PWπ machines. We call the resulting machine the PWFπ
machine:
Definition 3. The PWFπ machine consists of a stack of PWπ machines. The
transitions of the PWFπ machine are given below:
hp ; wi → hq ; w′ i
′
hp ; wi :: f → hq ; w i :: f
hp ; wi → hq1 ; wi
hp ; wi 6→
hp ; wi :: f → f
hp ; wi → hq2 ; wi
hp ; wi :: f → hq1 ; wi :: hq2 ; wi :: f
The initial state of the PWFπ machine is [hp0 ; wi]. The machine may terminate if one of the PWπ machines locates a match or if none of them succeeds
in finding a match. In the latter case the PWFπ machine has exhausted the entire search space and determined that the input string cannot be matched by the
regular expression in question.
The PWFπ machine allows us to analyze backtracking regular expression matchers at an abstract level without concerning ourselves about any implementation
specific details. More importantly, it gives an accurate cost model of backtracking matchers; the number of steps executed by the PWFπ machine corresponds
to the amount of work a backtracking matcher has to perform when searching
for a match. In the following sections we employ these ideas to develop and
implement our static analysis.
3
Static analysis for exponential blowup
The problem we are aiming to solve is this: given a regular expression e, represented as in Section 2, are there input strings x, y, and z, such that:
1. Reading x takes the machine to a pointer p0 that is the root of a Kleene star
expression.
2. Reading the input w takes the machine from p0 back to p0 , and in at least
two different ways, that is, along two different paths in the NFA.
3. Reading the input z when starting from p0 causes the match to fail.
r
x
p0
w
w
p0
w
p0
w
p0
w
w
p0
p0
p0
z
z
z
z
fail
fail
fail
fail
Fig. 3.1. The search tree for x w w y
We call x the prefix, w the pumpable string by analogy with pumping lemmas
in automata theory [11], and z the failure suffix.
From these three strings, malicious inputs can be constructed: the n-th malicious input is x wn z. Figure 3.1 illustrates the search tree that a backtracking
matcher has to explore when w is pumped twice. Because w can be matched in
two different ways, the tree branches every time a w is read from the input. All
branches fail in the end due to the trailing z, so that the matcher must explore
the whole tree.
To state the analysis more formally, we will need to define paths in the
matcher.
w
Definition 4. A path of pointers, t : p −→ q is defined according to the following
inductive rules:
ε
– For each pointer p, [p] : p −→ p is a path (identity).
w
– If t : p −→ q is a path and there exists a PWπ transition such that:
hq ; w′ w1 i → hq ′ ; w1 i
ww ′
Then t · [q ′ ] : p −→ q ′ is also a path.
w
Lemma 1. The path t : p −→ q (q 6= p) exists if and only if a PWπ run exists
such that:
hp ; ww′ i → · · · → hq ; w′ i
Lemma 1 associates a unique string w with each path of pointers (the substring matched by the corresponding PWπ run). However, note that the inverse
of this implication does not hold; there can be input strings for which we may
find more than one PWπ run. In fact, it is this property of paths that leads us
to the main theorem of this paper:
Theorem 1. For a given Kleene expression p0 where π(p0 ) = p1 ∗, if at least
w
w
two paths exist such that t1 : p1 −→ p0 , t2 : p1 −→ p0 and t1 6= t2 , then a
n
regular expression involving p0 exhibits o(2 ) runtime on a backtracking regular
expression matcher for input strings of the form xwn z where x is a sub-string
matching the prefix of p0 and z is such that xwn z fails to match the overall
expression.
While a formal proof of Theorem 1 is outside of the scope of this paper, we
sketch its proof with reference to Figure 3.1. The prefix x causes the PWFπ
machine to advance into a state where it has to match p0 against the remainder
of the input string, which leads to the branching of the search tree. Finally, the
suffix z at the end of the input causes each search path to fail, which in turns
forces the PWFπ machine to backtrack and explore the entire search tree before
concluding that a match cannot be found. For the complexity, note that each
additional pumping increases the size of the input by a constant (the length of
w) whereas it doubles the size of the binary subtree given by the w branches,
as well as the number of failed attempts to match z at the end. If there are
more than 2 ways to match the pumpable string, say b, then b rather than 2
becomes the base of the exponent, but 2 is still a lower bound. The matching of
the prefix x at the beginning contributes a constant to the runtime, which can
be disregarded relative to the exponential growth. Thus the lower bound for the
number of steps is exponential.
3.1
Generating the pumpable string
The most important step in generating an attack string for a vulnerable regular
expression is to generate the pumpable string w in xwn z (for some Kleene subexpression). In order to arrive at the machine for building the pumpable string,
we must first introduce several utility definitions. Note that in the remainder of
this discussion, p0 refers to a Kleene expression such that π(p0 ) = p1 ∗.
Definition 5. For a given pointer p, the operation p (called evolve) is defined
as:
ε
p = [q | ∃t.t : p −→ q ∧ ∃a.π(q) = a]
Notice that the result of p is a list of pointers.
Definition 6. The function Da (P ), (called derive) is defined on a list of pointers
P and an input symbol a according to the following rules:
Da ([ ]) = [ ]
Da (t)
Da (h :: t) = q :: Da (t)
Da ( h · t)
if π(h) = b, b 6= a
if π(h) = a ∧ cont h = q
otherwise.
The definition Da (P ) is analogous to Brzozowski’s derivatives of regular expressions [4]. In essence, the analysis computes derivatives of a Kleene expression
in order to find two different matcher states for the same input string.
Definition 7. A wP frame is defined as a pair (w, P ) where w is a string and
P is a list of pointers. A non-deterministic transition relation is defined on wP
frames as follows:
Da (P ) 6= [ ]
(w, P ) → (w a, Da (P ))
Definition 8. The HFπ machine has configurations of the following form:
hH ; f i
Here H (history) represents a set of (sorted) pointer lists and f is a list of wP
frames. A deterministic transition relation defines the behavior of this machine
as follows:
(w, P ) → (wx0 , P0 )
...
(w, P ) → (wxn , Pn )
∀i.xi ∈ Σ
Pi ∈
/H
hH ; (w, P ) :: f i → hH ∪ {P0 , . . . , Pn } ; f · [(wx0 , P0 ), . . . , (wxn , Pn )]i
The initial configuration of the HFπ machine is h∅ ; [(ε, [p1 ])]i and the machine
can terminate in either of the following two configurations:
hH ; [ ]i
ε
ε
hH ; (w, P ) :: f i where ∃p′ , p′′ ∈ P. ∃t′ , t′′ . t′ : p′ −→ p0 ∧ t′′ : p′′ −→ p0
In the former configuration the machine has determined the Kleene expression
in question to be non-vulnerable while in the latter it has derived the pumpable
string w.
3.2
Generating the Prefix and the Suffix
For a regular expression of the form e1 (e2 ∗) e3 , apart from a pumpable string w,
we must also generate a prefix x and a suffix z. The intention is that x would
lead the matcher to the point where it has to match e2 ∗, after which we can
pump many copies of w to increase the search space of the matcher. However, a
successful exploit also needs a suffix z which forces the matcher to fail and so to
traverse the entire search space.
Generating the prefix is quite straightforward since a depth-first search for
Kleene sub-expressions (on the AST) can be augmented such that a (minimal)
prefix is generated for each search-path. On the other hand, the suffix generation
is more involved. One would think z should be generated such that it fails to
match the continuation expression e3 of the original expression, but this intuition
is flawed since there is a possibility that z could be matched by e2 itself (while
it was meant for e3 ). Depending on e2 , it could be the case that no failure suffix
exists. One example we found is that e2 ends in .*, so that it can match anything.
In other cases, a failure suffix may exists, but depend in complicated ways on
e2 . We chose not to solve this problem in full generality, but rather to employ
heuristics that find failure suffixes for many practical expressions, as illustrated
in the results section.
4
Implementation of the static analysis
We implemented the HFπ machine described in Section 3 using the OCaml
programming language. OCaml is well suited to programming abstract syntax,
and hence a popular choice for writing static analyses. One of the major obstacles faced with the implementation is that in order to be able to analyze
real-world regular expressions, it was necessary to build a sophisticated parser.
In this regard, we decided to support the most common elements of the Perl /
PCRE standards, as these seem to be the most commonly used (and adapted)
syntaxes. It should be noted that the current implementation does not support
back-references or look-around expressions due to their inherent complexity; it
remains to be seen if the static analysis proposed in this work can be adapted
to handle such “regexes”. However, as it was explained earlier, exponential vulnerabilities in pattern specifications are not necessarily dependent on the use of
back-references or other advanced constructs (although one would expect such
constructs to further increase the search space of a backtracking matcher). A
detailed description of the pattern specification syntax currently supported by
the implementation has been included in the resources accompanying this paper.
The implementation closely follows the description of the HFπ machine presented in Section 3. The history component H is implemented as a set of sorted
integer lists, where a single sorted integer list corresponds to a list of nodes
pointed by the pointer list P of a wP frame (w, P ). This representation allows
for quick elimination of looping wP frames. While the size of H is potentially
exponential in the number of nodes of a given Kleene expression, for practical
regular expressions we found this size to be well within manageable levels (as
evidenced in the results section).
A matter not addressed in the current work is that the PWFπ machine (and
naive backtracking matching algorithms in general) can enter into infinite loops
for Kleene expressions when the enclosed sub-expression matches the empty
string (i.e the sub-expression is nullable). Although a complete treatment of this
issue and its solution (implemented by most of the well known backtracking
matchers) is beyond the scope of this paper, it should be mentioned that a
similar problem occurs in the HFπ machine during the p operation. We have
incorporated a method for detecting and terminating such infinite loops into the
OCaml code for the p function so that it terminates in all cases.
5
Experimental results
The analysis was tested on two corpora of regexes (Figure 1). The first of these
was extracted from an online regex library called RegExLib [19], which is a
community-maintained regex archive; programmers from various disciplines submit their solutions to various pattern matching tasks, so that other developers
can reuse these expressions for their own pattern matching needs. The second
corpus was extracted from the popular intrusion detection and prevention system
Snort [23], which contains regex-based pattern matching rules for inspecting IP
RegExLib Snort
Total patterns
2994 12499
Analyzable (only regular constructs)
2213
9408
Uses Kleene star
1103
2741
127
20
15
4
2066
9389
Pumpable Kleene and suffix found
Pumpable Kleene only
No pumpable Kleene
Max HFπ steps
509
256
Total classification time
(Intel Core 2 Duo 1.8 MHz, 4 GB RAM)
40 s
10 s
Table 1. Experimental results with RegExLib and Snort
packets across network boundaries. The contrasting purposes of these two corpora allow us to get a better view of the seriousness of exponential vulnerabilities
in practical regular expressions.
The regex archive for RegExLib was only available through the corresponding website [19]. Therefore, as the first step the expressions had to be scraped
from their web source and adapted so that they can be fed into our tool. These
adaptations include removing unnecessary white-space, comments and spurious
line breaks. A detailed description of these adjustments as well as copies of both
adjusted and un-adjusted data sets have been included with the resources accompanying this paper (also including the Python script used for scraping). The
regexes for Snort, on the other hand, are embedded within plain text files that
define the Snort rule set. A Python script (also included in the accompanying
resources) allowed the extraction of these regexes, and no further processing was
necessary.
The results of the HFπ static analysis on these two corpora of regexes are
presented in Table 1. The figures show that we can process around 75% of each
of the corpora with the current level of syntax support. Out of these analyzable
amounts, it is notable that regular expressions from the RegExLib archive use
the Kleene operator more frequently (about 50% of the analyzable expressions)
than those from the Snort rule set (close to 30%). About 11.5% of the Kleenebased RegExLib expressions were found to have a pumpable Kleene expression
as well as a suitable suffix, whereas for Snort this figure stands around 0.55%.
The vulnerabilities reported range from trivial programming errors to more
complicated cases. For an example, the following regular expression is meant to
validate time values in 24-hour format (from RegExLib):
^(([01][0-9]|[012][0-3]):([0-5][0-9]))*$
Here the author has mistakenly used the Kleene operator instead of the ? operator to suggest the presence or non-presence of the value. This pattern works
perfectly for all intended inputs. However, our analysis reports that this expression is vulnerable with the pumpable string “13:59” and the suffix “/”. This
result gives the programmer a warning that the regular expression presents a
DoS security risk if exposed to user-malleable input strings to match.
For a moderately complicated example, consider the following regular expression (again from RegExLib):
^([a-zA-z]:((\\([-*\.*\w+\s+\d+]+)|(\w+)\\)+)(\w+.zip)|(\w+.ZIP))$
This expression is meant to validate file paths to zip archives. Our tool identifies
this expression as vulnerable and generates the prefix “z:\ ”, the pumpable
string “\zzz\” and the empty string as the suffix. This is probably an unexpected
input in the author’s eye, and this is another way in which our tool can be
useful in that it can point out potential mis-interpretations which may have
materialized as vulnerabilities.
It is worth noting that the HFπ machine manages to classify both the corpora
(the analyzable portions) in a matter of seconds on modest hardware. This shows
that our static analysis is usable for most practical purposes, with the average
classification time for an expression in the range of micro-seconds. The two
extreme cases for which the machine took several seconds for the classification
are given below (only the respective Kleene expressions):
([\d\w][-\d\w]{0,253}[\d\w]\.)+
([^\x00]{0,255}\x00)*
Here counting expressions [-\d\w]{0,253} and [^\x00]{0,255} were expanded out during the parsing phase. The expansion produces a large Kleene
expression, which naturally requires more analysis during the HFπ simulation.
However, it should be noted that such expressions are the exception rather than
the norm.
Finally, it should be mentioned that all the vulnerabilities reported above
were individually verified using a modified version of the PWFπ machine (which
counts the number of steps taken for a particular matching operation). A sample
of those vulnerabilities was also tested on the Java regular expression matcher.
6
Conclusions
We have presented a static analysis to help programmers defend against regular
expression DoS attacks. Large numbers of regular expressions can be analysed
quickly, and developers are given feedback on where in their regular expressions
the problem has been identified as well as examples of malicious input.
As illustrated in Section 5, the prefix, pumpable string and failure suffix can
be quite short. If their length is, say, 3, 5 and 0 characters, then an attacker
only needs to spend a very small amount of effort in providing a malicious
input of length 3+5*100 characters to cause a matching time in excess of 2100
steps. Even if a matching step takes only a nanosecond, such a running time
takes, for all intents and purposes, forever. The attacker can still scale up the
attack by pumping a few times more and thereby correspondingly multiplying
the matching time.
The fact that the complexity of checking a regular expression for exponential
runtime may be computationally expensive in the worst case does not necessarily
imply that such an analysis is futile. Type checking in functional languages like
ML and Haskell also has high complexity [14,21], yet works efficiently in practice
because the worst cases rarely occur in real-world code. There are even program
analyses for undecidable problems like termination [3], so that the worst-case
running time is infinite; what matters is that the analysis produces results in
enough cases to be useful in practice. It is a common situation in program
analysis that tools are not infallible (having false positives and negatives), but
they are nonetheless useful for identifying points in code that need attention by
a human expert [9].
6.1
Related work
A general class of DoS attacks based on algorithmic complexities has been explored in [8]. In particular, the exponential runtime behavior of backtracking
regular expression matchers has been discussed in [6] and [20]. The seriousness
of this issue is further expounded in [22] and [16] where the authors demonstrate
the mounting of DoS attacks on an IDS/IPS system (Snort) by exploiting the
said vulnerability. The solutions proposed in these two works involve modifying
the regular expressions and/or the matching algorithm in order to circumvent
the problem in the context of IDS/IPS systems. We consider our work to be
quite orthogonal and more general since it is based on a compile-time static
analysis of regular expressions. However, it should be noted that both of those
works concern of regexes with back-references, which is a feature we are yet to
explore (known to be NP-hard [1]).
While the problem of ReDoS has been known for at least a decade, we are not
aware of any previous static analysis for defending against it. A handful of tools
exist that can assist programmers in finding such vulnerable regexes. Among
these tools we found Microsoft’s SDL Regex Fuzzer [15] and the RegexBuddy [13]
to be the most usable implementations, as other tools were too unstable to be
tested with complex expressions.
While RegexBuddy itself is not a security oriented software, it offers a debug
mode, which can be used to detect what the authors of the tool refer to as
Catastrophic Backtracking [10]. Even though such visual debugging methods
can assist in detecting potential vulnerabilities, it would only be effective if the
attack string is known in advance—this is where a static analysis method like
the one presented on this paper has a clear advantage.
SDL Fuzzer, on the other hand, is aimed specifically at analyzing regular
expression vulnerabilities. While details of the tool’s internal workings are not
publicly available, analyzing the associated documentation reveals that it operates fuzzing, i.e., by brute-forcing a sequence of generated strings through the
regular expression in question to detect long running times. The main disadvantage of this tool is that it can take a very long time for the tool to classify a
given expression. Tests using some of the regular expressions used in the results
section above revealed that it can take up to four minutes for the Fuzzer to
classify certain expressions. It is an inherent limitation of fuzzers for exponential runtime DoS attacks that the finding out if something takes a long time
by running it takes a long time. By contrast, our analysis statically analyzes
an expression without ever running it. It is capable of classifying thousands of
regular expressions in a matter of seconds. Furthermore, the output produced
by the SDL Fuzzer only reports the fact that the expression in question failed
to execute within a given time limit for some input string. Using this generated
input string to pin-point the exact problem in the expression would be quite
a daunting task. In contrast, our static analysis pin-points the exact Kleene
expression that causes the vulnerability and allows programmers to test their
matchers with a sequence of malicious inputs.
6.2
Directions for further research
In further work, we aim to broaden the coverage of our tool to include more
regexes. Given its basis in our earlier work on abstract machines [18] and derivatives [4], we aim for a formal proof of the correctness of our analysis. We intend
to release the source code of the tool as an open source project. More broadly,
we hope that raising awareness of the dangers of backtracking matchers will help
in the adoption of superior techniques for regular expression matching [7,24,18].
References
1. Alfred V. Aho. Algorithms for Finding Patterns in Strings. In Jan van Leeuwen,
editor, Handbook of theoretical computer science (vol. A), pages 255–300. MIT
Press, Cambridge, MA, USA, 1990.
2. Alfred V. Aho, Monica Lam, Ravi Sethi, and Jeffrey D. Ullman. Compilers Principles, Techniques and Tools. Addison Wesley, second edition, 2007.
3. J. Berdine, B. Cook, D. Distefano, and P. OHearn. Automatic termination proofs
for programs with shape-shifting heaps. In Computer Aided Verification, pages
386–400. Springer, 2006.
4. Janusz A. Brzozowski. Derivatives of Regular Expressions. J. ACM, 11(4):481–494,
1964.
5. B. Chess and G. McGraw. Static analysis for security. Security & Privacy, IEEE,
2(6):76–79, 2004.
6. Russ Cox. Regular Expression Matching Can Be Simple And Fast (but is slow in
Java, Perl, Php, Python, Ruby, ...). Available at http://swtch.com/~rsc/regexp/
regexp1.html, January 2007.
7. Russ Cox. Regular expression matching: the virtual machine approach. Available
at http://swtch.com/~rsc/regexp/regexp2.html, December 2009.
8. Scott A. Crosby and Dan S. Wallach. Denial of Service via Algorithmic Complexity
Attacks. In Proceedings of the 12th USENIX Security Symposium, Washington,
DC, August 2003.
9. Mark Dowd, John McDonald, and Justin Schuh. The Art of Software Security
Assessment: Identifying and Preventing Software Vulnerabilities. Addison Wesley,
2006.
10. Jan Goyvaerts. Runaway Regular Expressions: Catastrophic Backtracking. Available at http://www.regular-expressions.info/catastrophic.html, 2009.
11. John E. Hopcroft and Jeffrey D. Ullman. Introduction to Automata Theory, Languages and Computation. Addison-Wesley, 1979.
12. V.B. Livshits and M.S. Lam. Finding security vulnerabilities in java applications
with static analysis. In Proceedings of the 14th conference on USENIX Security
Symposium, volume 14, pages 18–18, 2005.
13. Just Great Software Co. Ltd. RegexBuddy. Available at http://www.regexbuddy.
com/, 2012.
14. H.G. Mairson. Deciding ML typability is complete for deterministic exponential
time. In Proceedings of the 17th ACM SIGPLAN-SIGACT symposium on Principles of programming languages, pages 382–401. ACM, 1989.
15. Microsoft. SDL Regex Fuzzer. Available at http://www.microsoft.com/en-gb/
download/details.aspx?id=20095, 2011.
16. Kedar Namjoshi and Girija Narlikar. Robust and Fast Pattern Matching for Intrusion Detection. In Proceedings of the 29th conference on Information communications, INFOCOM’10, pages 740–748, Piscataway, NJ, USA, 2010. IEEE Press.
17. The Open Web Application Security Project (OWASP). Regular Expression Denial
of Service - ReDoS. Available at https://www.owasp.org/index.php/Regular_
expression_Denial_of_Service_-_ReDoS, 2012.
18. Asiri Rathnayake and Hayo Thielecke. Regular Expression Matching and Operational Semantics. In Structural Operational Semantics (SOS 2011), Electronic
Proceedings in Theoretical Computer Science, 2011.
19. RegExLib.com. Regular Expression Library. Available at http://regexlib.com/,
2012.
20. Alex Roichman and Adar Weidman.
Regular Expression Denial
of Service.
Available at http://www.checkmarx.com/white_papers/
redos-regular-expression-denial-of-service/, 2012.
21. H. Seidl et al. Haskell overloading is DEXPTIME-complete. Information Processing Letters, 52(2):57–60, 1994.
22. Randy Smith, Cristian Estan, and Somesh Jha. Backtracking Algorithmic Complexity Attacks Against a NIDS. In Proceedings of the 22nd Annual Computer Security Applications Conference, ACSAC ’06, pages 89–98, Washington, DC, USA,
2006. IEEE Computer Society.
23. Sourcefire. Snort (IDS/IPS). Available at http://www.snort.org/, 2012.
24. Ken Thompson. Programming Techniques: Regular Expression Search Algorithm.
Communications of the ACM, 11(6):419–422, June 1968.
| 6cs.PL
|
Risk Minimization and Optimal Derivative Design in a Principal
arXiv:0710.5512v1 [cs.CE] 29 Oct 2007
Agent Game∗
Ulrich Horst
Santiago Moreno-Bromberg
Department of Mathematics
Department of Mathematics
Humboldt University Berlin
University of British Columbia
Unter den Linden 6
1984 Mathematics Road
10099 Berlin
Vancouver, BC, V6T 1Z2
[email protected]
[email protected]
March 15, 2018
Abstract
We consider the problem of Adverse Selection and optimal derivative design within a
Principal-Agent framework. The principal’s income is exposed to non-hedgeable risk factors
arising, for instance, from weather or climate phenomena. She evaluates her risk using a
coherent and law invariant risk measure and tries minimize her exposure by selling derivative
securities on her income to individual agents. The agents have mean-variance preferences
with heterogeneous risk aversion coefficients. An agent’s degree of risk aversion is private
information and hidden to the principal who only knows the overall distribution. We show
that the principal’s risk minimization problem has a solution and illustrate the effects of risk
transfer on her income by means of two specific examples. Our model extends earlier work of
Barrieu and El Karoui (2005) and Carlier, Ekeland and Touzi (2007).
Preliminary - Comments Welcome
AMS classification: 60G35, 60H20, 91B16, 91B70.
Keywords: Optimal derivative design, structured securities, adverse selection, risk transfer.
∗
We thank Guillaume Carlier, Pierre-Andre Chiappori, Ivar Ekeland, Andreas Putz and seminar participants
at various institutions for valuable comments and suggestions. Financial support through an NSERC individual
discovery grant is gratefully acknowledged.
1
1
Introduction
In recent years there has been an increasing interest in derivative securities at the interface of
finance and insurance. Structured products such as risk bonds, asset-backed securities and weather
derivatives are end-products of a process known as securitization that transforms non-tradable
risk factors into tradable financial assets. Developed in the U.S. mortgage markets, the idea of
pooling and underwriting risk that cannot be hedged through investments in the capital markets
alone has long become a key factor driving the convergence of insurance and financial markets.
Structured products are often written on non-tradable underlyings, tailored to the issuers
specific needs and traded “over the counter”. Insurance companies, for instance, routinely sell
weather derivatives or risk bonds to customers that cannot go to the capital markets directly
and/or seek financial securities with low correlation with stock indices as additions to diversified
portfolios. The market for such claims is generally incomplete and illiquid. As a result, many of
the standard paradigms of traditional derivative pricing theory, including replication arguments
do not apply to structured products. In an illiquid market framework, preference-based valuation
principles that take into account characteristics and endowment of trading partners may be more
appropriate for designing, pricing and hedging contingent claims. Such valuation principles have
become a major topic of current research in economics and financial mathematics. They include
rules of Pareto optimal risk allocation ([11], [16]), market completion and dynamic equilibrium
pricing ([14], [15]) and, in particular, utility indifference arguments ([2], [3], [5], [6], [9], ...). The
latter assumes a high degree of market asymmetry. For indifference valuation to be a pricing rather
than valuation principle, the demand for a financial security must come from identical agents
with known preferences and negligible market impact while the supply must come from a single
principal. When the demand comes from heterogeneous individuals with hidden characteristics,
indifference arguments do not always yield an appropriate pricing scheme.
In this paper we move away from the assumption of investor homogeneity and allow for
heterogeneous agents. We consider a single principal with a random endowment whose goal is to
lay off some of her risk with heterogeneous agents by designing and selling derivative securities
on her income. The agents have mean variance preferences. An agent’s degree of risk aversion
is private information and hidden to the principal. The principal only knows the distribution of
risk aversion coefficients which puts her at an informational disadvantage. If all the agents were
homogeneous, the principal, when offering a structured product to a single agent, could (perhaps)
extract the indifference (maximum) price from each trading partner. In the presence of agent
heterogeneity this is no longer possible, either because the agents would hide their characteristics
from the principal or prefer another asset offered by the principal but designed and priced for
another customer.
The problem of optimal derivative design in a Principal-Agent framework with informed agents
and an uninformed principal has first been addressed in a recent paper by of Carlier, Ekeland
2
and Touzi [7]. With the agents being the informed party, theirs is a screening model. The
literature on screening within the Adverse Selection framework can be traced back to Mussa and
Rosen [18], where both the principal’s allocation rule and the agents’ types are one-dimensional.
Armstrong [1] relaxes the hypothesis of agents being characterized by a single parameter. He
shows that, unlike the one-dimensional case, “bunching” of the first type is robust when the
types of the agents are multi-dimensional. In their seminal paper, Rochet and Choné [19] further
extend this analysis. They provide a characterization of the contracts, determined by the (nonlinear) pricing schedule, that maximize the principal’s utility under the constraints imposed by
the asymmetry of information in the models. Building on their work, Carlier, Ekeland and Touzi
[7] study a Principal-Agent model of optimal derivative design where the agents’ preferences are of
mean-variance type and their multi-dimensional types characterize their risk aversions and initial
endowments. They assume that there is a direct cost to the principal when she designs a contract
for an agent, and that the principal’s aim is to maximize profits.
We start from a similar set-up, but substitute the idea that providing products carries a cost
for the idea that traded contracts expose the principal to additional risk - as measured by a convex
risk measure - in exchange for a known revenue. This may be viewed as a partial extension of the
work by Barrieu and El Karoui ([2],[3]) to an incomplete information framework.
The principal’s aim is to minimize her risk exposure by trading with the agents subject to the
standard incentive compatibility and individual rationality conditions on the agents’ choices. In
order to prove that the principal’s risk minimization problem has a solution we first follow the
seminal idea of Rochet and Choné [19] and characterize incentive compatible catalogues in terms
of U -convex functions. When the impact of a single trade on the principal’s revenues is linear
as in Carlier, Ekeland and Touzi [7], the link between incentive compatibility and U -convexity is
key to establish the existence of an optimal solution. In our model the impact is non-linear as a
single trade has a non-linear impact on the principal’s risk assessment. Due to this non-linearity
we face a non-standard variational problem where the objective cannot be written as the integral
of a given Lagrangian. Instead, our problem can be decomposed into a standard variational
part representing the aggregate income of the principal, plus the minimization of the principal’s
risk evaluation, which depends on the aggregate of the derivatives traded. We state sufficient
conditions that guarantee that the principal’s optimization problem has a solution and illustrate
the effect of risk transfer on her exposure by means of two specific examples.
The remainder of this paper is organized as follows. In Section 2 we formulate our PrincipalAgent model and state the main result. The proof is given in Section 3. In Section 4 we illustrate
the effects of risk transfer on the principal’s position by two examples. In the first we consider
a situation where the principal restricts itself to type-dependent multiples of some benchmark
claim. This case can be solved in closed form by means of a standard variational problem. The
second example considers put options with type-dependent strikes. In both cases we assume that
3
the principal’s risk measure is Average Value at Risk. As a consequence the risk minimization
problem can be stated in terms of a min-max problem; we provide an efficient numerical scheme
for approximating the optimal solution. The code is given in an appendix.
2
The Microeconomic Setup
We consider an economy with a single principal whose income W is exposed to non-hedgeable
risk factors rising from, e.g., climate or weather phenomena. The random variable W is defined
on a standard, non-atomic, probability space (Ω, F, P) and it is square integrable:
W ∈ L2 (Ω, F, P).
The principal’s goal is to lay off parts of her risk with individual agents. The agents have
heterogenous mean-variance preferences1 and are indexed by their coefficients of risk aversion
θ ∈ Θ. Given a contingent claim Y ∈ L2 (Ω, F, P) an agent of type θ enjoys the utility
U (θ, Y ) = E[Y ] − θ Var[Y ].
(1)
Types are private information. The principal knows the distribution µ of types but not the
realizations of the random variables θ. We assume that the agents are risk averse and that the
risk aversion coefficients are bounded away from zero. More precisely,
Θ = [a, 1]
for some a > 0.
The principal offers a derivative security X(θ) written on her random income for any type θ.
The set of all such securities is denoted by
X := X = {X(θ)}θ ∈Θ | X ∈ L2 (Ω × Θ, P ⊗ µ), X is σ(W ) × B(Θ) measurable .
(2)
v(θ) = sup U (θ, X(θ ′ )) − π(θ ′ ) .
θ ′ ∈Θ
(3)
We refer to a list of securities {X(θ)} as a contract. A catalogue is a contract along with prices
π(θ) for every available derivative X(θ). For a given catalogue (X, π) the optimal net utility of
the agent of type θ is given by
Remark 2.1 No assumption will be made on the sign of π(θ); our model contemplates both the
case where the principal takes additional risk in exchange of financial compensation and the one
where she pays the agents to take part of her risk.
1
Our analysis carries over to preferences of mean-variance type with random initial endowment as in [7]; the
assumption of simple mean-variance preferences is made for notational convenience.
4
A catalogue (X, π) will be called incentive compatible (IC) if the agent’s interests are best
served by revealing her type. This means that her optimal utility is achieved by the security X(θ):
U (θ, X(θ)) − π(θ) ≥ U (θ, X(θ ′ )) − π(θ ′ ) for all
θ, θ′ ∈ Θ.
(4)
We assume that each agent has some outside option (“no trade”) that yields a utility of zero.
A catalogue is thus called individually rational (IR) if it yields at least the reservation utility for
all agents, i.e., if
U (θ, X(θ)) − π(θ) ≥ 0 for all θ ∈ Θ.
(5)
Remark 2.2 By offering only incentive compatible contracts, the principal forces the agents to
reveal their type. Offering contracts where the IR constraint is binding allows the principal to
exclude undesirable agents from participating in the market. It can be shown that under certain
conditions, the interests of the Principal are better served by keeping agents of “lower types” to
their reservation utility; Rochet and Choné [19] have shown that in higher dimensions this is
always the case.
R
If the principal issues the catalogue (X, π), she receives a cash amount of Θ π (θ) dµ(θ) and is
R
subject to the additional liability Θ X(θ)µ(dθ). She evaluates the risk associated with her overall
position
Z
(π(θ) − X(θ))dµ(θ)
W+
Θ
via a coherent and law-invariant risk measure ̺ : L2 (Ω, F, P) → R ∪ {∞} that has the Fatou
property. It turns out that such risk measures can be represented as robust mixtures of Average
Value at Risk.2 The principal’s risk associated with the catalogue (X, π) is given by
Z
(π(θ) − X(θ))dµ(θ) .
(6)
̺ W+
Θ
Her goal is to devise contracts (X, π) that minimize (6) subject to the incentive compatibility and
individual rationality condition:
Z
(π(θ) − X(θ))dµ(θ) | X ∈ X , X is IC and IR .
(7)
inf ̺ W +
Θ
We are now ready to state the main result of this paper. The proof requires some preparation
and will be carried out in the following section.
Theorem 2.3 If ̺ is a coherent and law invariant risk measure on L2 (P) and if ̺ has the Fatou
property, then the principal’s optimization problem has a solution.
For notational convenient we establish our main result for the spacial case dµ(θ) = dθ. The
general case follows from straight forward modifications.
We review properties of coherent risk measures on Lp spaces in the appendix and refer to the textbook by
Föllmer and Schied [12] and the paper of Jouini, Schachermayer and Touzi [17] for detailed discussion of law
invariant risk measures.
2
5
3
Proof of the Main Theorem
Let (X, π) be a catalogue. In order to prove our main result it will be convenient to assume that
the principal offers any square integrable contingent claim and to view the agents’ optimization
problem as optimization problems over the set L2 (P). This can be achieved by identifying the
price list {π(θ)} with the pricing scheme
π : L2 (P) → R
that assigns the value π(θ) to an available claim X(θ) and the value E[Y ] to any other claim
Y ∈ L2 . In terms of this pricing scheme the value function v defined in (3) satisfies
v(θ) =
sup {U (θ, Y ) − π(Y )} .
(8)
Y ∈L2 (P)
for any individually rational catalogue. For the remainder of this section we shall work with the
value function of the (8). It is U -convex in the sense of the following definition; it actually turns
out to be convex and non-increasing as we shall prove in Proposition 3.2 below.
Definition 3.1 Let two spaces A and B and a function U : A × B → R be given.
(i) The function f : A → R is called U -convex if there exists a function p : B → R such that
f (a) = sup {U (a, b) − p(b)} .
b∈B
(ii) For a given function p : B → R the U -conjugate pU (a) of p is defined by
pU (a) = sup {U (a, b) − p(b)} .
b∈B
(iii) The U-subdifferential of p at b is given by the set
∂U p(b) := a ∈ A | pU (a) = U (a, b) − p(b) .
(iv) If a ∈ ∂U p(b), then a is called a U-subgradient of p(b).
Our goal is to identify the class of IC and IR catalogues with a class of convex and nonincreasing functions on the type space. To this end, we first recall the link between incentive
compatible contracts and U -convex functions from Rochet and Choné [19] and Carlier, Ekeland
and Touzi [7].
Proposition 3.2 ([19], [7]) If a catalogue (X, π) is incentive compatible, then the function v
defined by (3) is proper and U-convex and X(θ) ∈ ∂U v(θ). Conversely, any proper, U-convex
function induces an incentive compatible catalogue.
6
Proof. Incentive compatibility of a catalogue (X, π) means that
U (θ, X(θ)) − π(θ) ≥ U (θ, X(θ ′ )) − π(θ ′ ) for all
θ, θ′ ∈ Θ,
so v(θ) = U (θ, X(θ)) − π(θ) is U-convex and X(θ) ∈ ∂U v(θ). Conversely, for a proper, U-convex
function v and X(θ) ∈ ∂U v(θ) let
π(θ) := U (θ, X(θ)) − v(θ).
By the definition of the U-subdifferential, the catalogue (X, π) is incentive compatible.
✷
The following lemma is key. It shows that the U -convex function v is convex and non-increasing
and that any convex and non-increasing function is U -convex, i.e., it allows a representation of
the form (8). This allows us to rephrase the principal’s problem as an optimization problem over
a compact set of convex functions.
Lemma 3.3 (i) Suppose that the value function v as defined by (8) is proper. Then v is convex
and non-increasing. Any optimal claim X ∗ (θ) is a U -subgradient of v(θ) and almost surely
−Var[X ∗ (θ)] = v ′ (θ).
(ii) If v̄ : Θ → R+ is proper, convex and non-increasing, then v̄ is U -convex, i.e., there exists a
map π̄ : L2 (P) → R such that
v̄(θ) =
sup {U (θ, Y ) − π̄(Y )} .
Y ∈L2 (P)
Furthermore, any optimal claim X̄(θ) belongs to the U -subdifferential of v̄(θ) and satisfies
−Var[X̄(θ)] = v̄ ′ (θ).
Proof.
(i) Let v be a proper, U -convex function. Its U -conjugate is:
v U (Y ) =
sup {E[Y ] − θVar[Y ] − v(θ)}
θ∈Θ
= E[Y ] + sup {θ(−Var[Y ]) − v(θ)}
θ∈Θ
= E[Y ] + v ∗ (−Var[Y ]),
where v ∗ denotes the convex conjugate of v. As a U -convex function, the map v is characterized by the fact that v = (v U )U . Thus
v(θ) = (v U )U (θ)
=
sup {U (θ, Y ) − E[Y ] − v ∗ (−Var[Y ])}
Y ∈L2 (P)
=
sup {E[Y ] − θVar[Y ] − E[Y ] − v ∗ (−Var[Y ])}
Y ∈L2 (P)
= sup {θ · y − v ∗ (y)}
y≤0
7
where the last equality uses the fact that the agents’ consumption set contains claims of
any variance. We deduce from the preceding representation that v is non-increasing. Furthermore v = (v ∗ )∗ so v is convex. To characterize ∂U v(θ) we proceed as follows:
∂U v(θ) = Y ∈ L2 | v(θ) = U (θ, X) − v U (Y )
= Y ∈ L2 | v(θ) = E[Y ] − θVar[Y ] − v U (Y )
= Y ∈ L2 | v(θ) = E[Y ] − θVar[Y ] − E[Y ] − v ∗ (−Var[Y ])
= Y ∈ L2 | v(θ) = θ(−Var[Y ]) − v ∗ (−Var[Y ])
= Y ∈ L2 | −Var[Y ] ∈ ∂v(θ)
The convexity of v implies it is a.e. differentiable so we may write
∂U v(θ) := Y ∈ L2 | v ′ (θ) = −Var[Y ]) .
(ii) Let us now consider a proper, non-negative, convex and non-increasing function v̄ : Θ → R.
There exists a map f : R → R such that
v̄(θ) = sup {θ · y − f (y)} .
y≤0
Since v̄ is non-increasing there exists a random variable Y (θ) ∈ L2 (P) such that −Var[Y (θ)] ∈
∂v̄(θ) and the definition of the subgradient yields
v̄(θ) = sup {θ(−Var[Y ]) − f (−Var[Y ])} .
Y ∈L2
With the pricing scheme on L2 (P) defined by
π̄(Y ) := −E[Y ] − f (−Var[Y ])
this yields
v̄(θ) = sup {U (θ, Y ) − π̄(Y )} .
Y ∈L2
The characterization of the subdifferential follows by analogy to part (i).
✷
The preceding lemma along with Proposition 3.2 shows that any convex, non-negative and
non-increasing function v on Θ induces an incentive compatible catalogue (X, π) via
X(θ) ∈ ∂U v(θ) and π(θ) = U (θ, X(θ)) − v(θ).
Here we may with no loss of generality assume that E[X(θ)] = 0. In terms of the principal’s
choice of v her income is given by
Z
θv ′ (θ) − v(θ) dθ.
I(v) =
Θ
8
Since v ≥ 0 is decreasing and non-negative the principal will only consider functions that satisfy
the normalization constraint
v(1) = 0.
We denote the class of all convex, non-increasing and non-negative real-valued functions on Θ
that satisfy the preceding condition by C:
C = {v : Θ → R | v is convex, non-increasing, non-negative and v(1) = 0.}
Conversely, we can associate with any IC and IR catalogue (X, π) a non-negative U -convex
function of the form (8) where the contract satisfies the variance constraint −Var[X(θ)] = v ′ (θ). In
view of the preceding lemma this function is convex and non-increasing so after normalization we
may assume that v belongs to the class C. We therefore have the following alternative formulation
of the principal’s problem.
Theorem 3.4 The principal’s optimization problem allows the following alternative formulation:
Z
′
X(θ)dθ − I(v) | v ∈ C, E[X(θ)] = 0, −Var[X(θ)] = v (θ) .
inf ̺ W −
Θ
In terms of our alternative formulation we can now prove a preliminary result. It states that
a principal with no initial endowment will not issue any contracts.
Lemma 3.5 If the principal has no initial endowment, i.e., if W = 0, then (v, X) = (0, 0) solves
her optimization problem.
Proof. Since ̺ is a coherent, law invariant risk measure on L2 (P) that has the Fatou property
it satisfies
̺(Y ) ≥ −E[Y ] for all Y ∈ L2 (P).
(9)
For a given function v ∈ C the normalization constraint E[X(θ)] = 0 therefore yields
Z
Z
X(θ)dθ − I(v) = −I(v).
X(θ)dθ − I(v) ≥ E
̺ −
Θ
Θ
Since v is non-negative and non-increasing −I(v) ≥ 0. Taking the infimum in the preceding
inequality shows that v ≡ 0 and hence X(θ) ≡ 0 is an optimal solution.
✷
3.1
Minimizing the risk for a given function v
In the general case we approach the principal’s problem in two steps. We start by fixing a function
v from the class C and minimize the associated risk
Z
X(θ)dθ
̺ W−
Θ
9
subject to the moment conditions E[X(θ)] = 0 and −Var[X(θ)] = v ′ (θ). To this end, we shall
first prove the existence of optimal contracts Xv for a relaxed optimization where the variance
constraint is replaced by the weaker condition
Var[X(θ)] ≤ −v ′ (θ).
In a subsequent step we show that based on Xv the principal can transfer risk exposures among
the agents in such a way that (i) the aggregate risk remains unaltered; (ii) the variance constraint
becomes binding. We assume with no loss of generality that v does not have a jump at θ = a.
3.1.1
The relaxed optimization problem
For a given v ∈ C let us consider the convex set of derivative securities
X v := X ∈ X | E[X(θ)] = 0, Var[X(θ)] ≤ −v ′ (θ) µ − a.e. .
Lemma 3.6
(10)
(i) All functions v ∈ C that are acceptable for the principal are uniformly bounded.
(ii) Under the conditions of (i) the set X v is closed and bounded in L2 (P ⊗ µ). More precisely,
kXk22 ≤ v(a)
for all
X ∈ X v.
Proof.
(i) If v is acceptable for the principal, then any X ∈ X v satisfies
Z
X(θ)dθ − I(v) ≤ ̺(W ).
̺ W−
Θ
From (9) and that fact that E[X(θ)] = 0 we deduce that
Z
X(θ)dθ − I(v) ≤ ̺(W )
−E[W ] − I(v) ≤ ̺ W −
Θ
so
−I(v) ≤ E[W ] + ̺(W ) =: K.
Integrating by parts twice and using that v is non-increasing and v(1) = 0 we see that
Z 1
v(θ)dθ ≥ av(a).
K ≥ −I(v) = av(a) + 2
a
This proves the assertion because a > 0.
(ii) For X ∈ X v we deduce from the normalization constraint v(1) = 0 that
Z Z
Z
2
2
kXk2 =
X (θ, ω)dP dθ ≤ − v ′ (θ)dθ ≤ v(a)
so the assertion follows from part (i).
10
✷
Since ̺ is a convex risk measure on L2 and because the set Xv of contingent claims is convex,
closed and bounded in L2 a general result from the theory of convex optimization yields the
following proposition.
Proposition 3.7 If the function v is acceptable for the principal, then there exists a contract
{Xv (θ)} such that
Z
Z
Xv (θ)dθ .
X(θ)dθ = ̺ W −
inf ̺ W −
X∈X v
Θ
Θ
The contract Xv along with the pricing scheme associated with v does not yield an incentive
compatible catalogue unless the variance constraints happen to be binding. However, as we are
now going to show, based on Xv the principal can find a redistribution of risk among the agents
such that the resulting contract satisfies our IC condition.
3.1.2
Redistributing risk exposures among agents
Let
∂X v = X ∈ X v | E[X(θ)] = 0, Var[X(θ)] = −v ′ (θ), µ − a.e.
be the set of all contracts from the class X v where the variance constraint is binding. Clearly,
Z
Z
X(θ)dθ .
Xv (θ)dθ ≤ inf ̺ W −
̺ W−
X∈∂ X v
Θ
Θ
Let us then introduce the set of types
Θv := θ ∈ Θ | Var[Xv (θ)] < −v ′ (θ) ,
for whom the variance constraint is not binding. If µ(Θv ) = 0, then Xv yields an incentive
compatible contract. Otherwise, we consider a random variable Ỹ ∈ X v , fix some type θ ∈ Θ and
define
Ỹ (θ)
.
(11)
Y := q
Var[Ỹ (θ)]
We may with no loss of generality assume that Y is well defined for otherwise the status
quo is optimal for the principal and her risk minimization problem is void. The purpose of
introducing Y is to offer a set of structured products Zv based on Xv , such that Zv together with
the pricing scheme associated with v yields an incentive compatible catalogue. To this end, we
choose constants α̃(θ) for θ ∈ Θv such that
Var[Xv (θ) + α̃(θ)Y ] = −v ′ (θ).
11
This equation holds for
α̃± (θ) = −Cov[Xv (θ), Y ] ±
q
Cov2 [Xv (θ), Y ] − v ′ (θ) − Var[Xv (θ)].
For a type θ ∈ Θv the variance constraint is not binding. Hence −v ′ (θ) − Var[Xv (θ)] > 0 so
that α+ (θ) > 0 and α− (θ) < 0. An application of Jensen’s inequality together with the fact that
kXv k2 is bounded shows that α± are µ-integrable functions. Thus there exists a threshold type
θ∗ ∈ Θ such that
Z
Z
+
α− (θ)dθ = 0.
α (θ)dθ +
Θv ∩(θ ∗ ,1]
Θv ∩(a,θ ∗ ]
In terms of θ∗ let us now define a function
(
α(θ) :=
α̃+ (θ), if
α̃− (θ), if
θ ≤ θ∗
θ > θ∗
and a contract
Zv := Xv + αY ∈ ∂X v .
(12)
R
Since αdθ = 0 the aggregate risks associated with Xv and Zv are equal. As a result, the contract
Zv solves the risk minimization problem
Z
X(θ)µ(dθ) .
(13)
inf ̺ W +
X∈∂ X v
Θ
Remark 3.8 In Section 4 we shall consider a situation where the principal restricts itself to a
class of contracts for which the random variable Xv can be expressed in terms of the function
v. In general such a representation will not be possible since v only imposes a restriction on the
contracts’ second moments.
3.2
Minimizing the overall risk
In order to finish the proof of our main result it remains to show that the minimization problem
Z
Zv (θ)µ(dθ) − I(v)
inf ̺ W −
v∈C
Θ
has a solution and the infimum is obtained. To this end, we consider a minimizing sequence
{vn } ⊂ C. The functions in C are locally Lipschitz continuous because they are convex. In fact
they are uniformly locally Lipschitz: by Lemma 3.6 (i) the functions v ∈ C are uniformly bounded
and non-increasing so all the elements of ∂v(θ) are uniformly bounded on compact sets of types.
As a result, {vn } is a sequence of uniformly bounded and uniformly equicontinuous functions
when restricted to compact subsets of Θ. Thus there exists a function v̄ ∈ C such that, passing
to a subsequence if necessary,
lim vn = v̄
n→∞
uniformly on compact sets.
12
A standard 3ǫ-argument shows that the convergence properties of the sequence {vn } carry over
to the derivatives so that
lim v ′
n→∞ n
= v̄ ′
almost surely uniformly on compact sets.
Since −θvn′ (θ) + vn (θ) ≥ 0 it follows from Fatou’s lemma that −I(v̄) ≤ lim inf n→∞ −I(vn ) so
Z
lim inf ̺ W −
Zvn (θ)µ(dθ) − I(vn )
n→∞
Θ
Z
Zvn (θ)µ(dθ) + lim inf −I(vn )
≥ lim inf ̺ W −
n→∞
n→∞
ZΘ
Zvn (θ)µ(dθ) − I(v̄)
≥ lim inf ̺ W −
n→∞
Θ
and it remains to analyze the associated risk process. For this, we first observe that for Zvn ∈ ∂Xvn
Fubini’s theorem yields
Z Z
Z
2
2
kZvn k2 =
Zvn dPdθ = − vn′ (θ)dθ = vn (a).
(14)
Since all the functions in C are uniformly bounded, we see that the contracts Zvn are contained in
an L2 bounded, convex set. Hence there exists a square integrable random variable Z such that,
after passing to a subsequence if necessary,
w − lim Zn = Z
(15)
n→∞
Let Zv̄ ∈ Xv̄ . Convergence of the functions vn implies kZvn k2 → kZv̄ k2 . Thus (15) yields
kZk2 = kZv̄ k2 along with convergence of aggregate risks:
Z
Z
Z(θ, ω)dθ weakly in L2 (P).
Zn (θ, ω)dθ →
kZk2 = kZv̄ k2 and
Θ
Θ
By Corollary I.2.2 in Ekeland and Témam (1976) [10], a lower semi-continuous convex function
f : X → R remains ,so with respect to the weak topology σ(X, X ∗ ), the Fatou property of the
risk measure ̺ guarantees that
Z
Z
Z(θ)µ(dθ)
Zv̄ (θ)µ(dθ)
≤ ̺ W−
̺ W−
Θ
Θ
Z
≤ lim inf ̺ W −
n→∞
We conclude that (Zv̄ , v̄) solves the Principal’s problem.
13
Θ
Z(θ)µ(dθ) .
4
Examples
Our main theorem states that the principal’s risk minimization problem has a solution. The
solution can be characterized in terms of a convex function that specifies the agents’ net utility.
Our existence result is based on a min-max optimization scheme whose complexity renders a rather
involved numerical analysis . In this section we consider some examples where the principal’s
choice of contracts is restricted to class of numerically more amenable securities. The first example
studies a situation where the principal offers type-dependent multiplies of some benchmark claim.
In this case the principal’s problem can be reduced to a constraint variational problem that can
be solved in closed form. A second example comprises put options with type dependent strikes.
Here we provide a numerical algorithm for approximating the optimal solution.
4.1
A single benchmark claim
In this section we study a model where the principal sells a type-dependent multiple of a benchmark claim f (W ) ≥ 0 to the agents. More precisely, the principal offers contracts of the form
X(θ) = α(θ)f (W ).
(16)
In order to simplify the notation we shall assume that the T-bond’s variance is normalized:
Var[f (W )] = 1.
4.1.1
The optimization problems
Let (X, π) be a catalogue where the contract X is of the form (16). By analogy to the general
case it will be convenient to view the agents’ optimization problem as an optimization problem
of the set of claims {γf (W ) | γ ∈ R} so the function α : Θ → R solves
sup {U (θ, γf (W )) − π(θ)} .
γ∈R
In view of the variance constraint on the agents’ claims the principal’s problem can be written as
Z p
inf {̺ (W − C(v)f (W )) − I(v) | v ∈ C} where C(v) =
−v ′ (θ)dθ.
Θ
p
Note that E[f (W )] > 0, so the term E[f (W )] −v ′ (θ) must be included in the income. Before
proceeding with the general case let us first consider a situation where in addition to being coherent
and law invariant, the risk measure ̺ is also comonotone. In this case each security the principal
sells to some agent increases her risk by the amount
p
̺ −(f (W ) − E[f (W )]) −v ′ (θ) + v(θ) − θv ′ (θ) ≥ 0.
This suggests that it is optimal for the principal not to sell a bond whose payoff moves into the
same direction as her initial risk exposure.
14
Proposition 4.1 Suppose that ̺ is comonotone additive. If f (W ) and W are comonotone, then
v = 0 is a solution to the principal’s problem.
Proof. If W and f (W ) are comonotone, then the risk measure in equation (4.1.1) is additive
and the principal needs to solve
Z
p
v(θ) + ̺ (f (W ) − E[f (W )]) −v ′ (θ) − θv ′ (θ) dθ.
̺ (W ) + inf
v∈C Θ
Since ̺ (f (W ) − E[f (W )]) ≥ 0 and −θv ′ (θ) ≥ 0 we see that
Z
p
v(θ) + ρ (F (W )) −v ′ (θ) − θv ′ (θ) dθ ≥ 0
Θ
and hence v ≡ 0 is a minimizer.
✷
In view of the preceding proposition the principal needs to design the payoff function f in
such a way that W and f (W ) are not comonotone. We construct an optimal payoff function in
the following subsection.
4.1.2
A solution to the principal’s problem
Considering the fact that ̺(·) is a decreasing function the principal’s goal must be to make the
quantity C(v) as small as possible while keeping the income as large as possible. In a first step
we therefore solve, for any constant A ∈ R the optimization problem
Z
p
(17)
E[f (W )] −v ′ (θ) − v(θ) + θv ′ (θ) dθ = A.
sup C(v) subject to
v∈C
Θ
The constraint variational problem (17) captures the problem of risk minimizing subject to
an income constraint. It can be solved in closed form. The associated Euler-Lagrange equation
is given by
!
λE[f ] − 1
d
−λθ + p
,
(18)
λ=
dθ
2 −v ′ (θ)
where λ is the Lagrange multiplier. The income constraint and boundary conditions are:
v ′ (a) = −
(λ′ )2
4λ2 a2
and v(1) = 0 where λ′ = (λE[f ] − 1).
Integrating both sides of equation (18) and taking into account the normalization condition
v(1) = 0, we obtain
1 λ′ 2
1
1
v(θ) =
−
.
8 λ
2θ − a 2 − a
Inserting this equation into the constraint yields
s
′ 2 Z 1
Z
λ
θ
1
1
1
1
λ′ 2 1 dθ
−
−
dθ.
A = E[f ]
+
λ
λ
8 2θ − a 2 − a
4 (2θ − a)2
a 2θ − a
a
15
In terms of
M :=
Z
a
1
1
θ
1
1
1
−
dθ
+
8 2θ − a 2 − a
4 (2θ − a)2
and N :=
Z
a
1
dθ
2θ − a
we have the quadratic equation
−M
which has the solution
s
λ′
λ
λ′
λ
2
2
=
s
λ′ 2
+ N E[f ]
− A = 0,
λ
N E[f ] −
p
(N E[f ])2 − 4AM
2M
We have used the root with alternating signs, as we require the problem to reduce to ̺(W ) for
A = 0.
Remark 4.2 We notice that the constraint variational problem (17) is independent of the risk
measure employed by the principal. This is because we minimized the risk pointwise subject to a
constraint on aggregate revenues.
In view of the preceding considerations the principal’s problem reduces to a one-dimensional
minimization problems over the Reals:
N p
N 2 E[f ]
2
+ f (W )
inf ̺ W − f (W )
(N E[f ]) − 4AM − A.
A
2M
2M
Once the optimal value A∗ has been determined, the principal offers the securities
λE[f ] − 1
f (W )
4θλ − 2λa
at a price
λE[f ] − 1
2
3Eλ(2θ − a) − a λE − 1 1
+
(4θλ − 2λa)2
λ2 2 − a
.
Example 4.3 Assume that the principal measures her risk exposure using Average Value at Risk
at level 0.05. Let W̃ be a normally distributed random variable with mean 1/2 and variance
1/20. One can think that W̃ represents temperature. Suppose that the principal’s initial income
is exposed to temperature risk and it is given by W = 0.1(W̃ − 1.1) with associated risk
̺(W ) = 0.0612.
Suppose furthermore that the principal sells units of a put option on W̃ with strike 0.5, i.e.,
f (W ) = (W − 0.5)+
16
By proceeding as above we approximated the principal’s risk as −0.6731 and she offers the security
X(θ) =
0.5459
f (W )
2θ − a
to the agent of type θ for a price
π(θ) =
4.2
1.1921
(1.1921)θ − (0.22)(2θ − a)
√
−
.
8(2 − a)
2(2θ − a)2
Put options with type dependent strikes
In this section we consider the case where the principal underwrites put options on her income
with type-dependent strikes. We assume that W ≤ 0 is a bounded random variable and consider
contracts of the form
X(θ) = (K(θ) − |W |)+
with
0 ≤ K(θ) ≤ kW k∞ .
The boundedness assumption on the strikes is made with no loss of generality and each equilibrium pricing scheme is necessarily non-negative. Note that in this case the risk measure can
be defined on L(P), so we only require convergence in probability to use the Fatou property. We
deduce that both the agents’ net utilities and the variance of their positions is bounded from
above by some constants K1 and K2 , respectively. Thus, the principal chooses a function v and
contract X from the set
{(X, v) | v ∈ C, v ≤ K1 , −Var[K(θ) − |W |] = v ′ (θ), |v ′ | ≤ K2 , 0 ≤ K(θ) ≤ kW k∞ }.
The variance constraint v ′ (θ) = −Var[(K(θ) − W )+ ] allows us to express the strikes in terms
of a continuous function of v ′ , i.e.,
K(θ) = F (v ′ (θ)).
The Principal’s problem can therefore be written as
Z
′
+
′
+
(F (v (θ)) − |W |) − E[(F (v (θ)) − |W |) ] d − I(v)
inf ̺ W −
Θ
where the infimum is taken over the set of all functions v ∈ C that satisfy v ≤ K1 and |v ′ | ≤ K2 .
Remark 4.4 Within our current framework the contracts are expressed in terms of the derivative
of the principal’s choice of v. This reflects the fact that the principal restricts itself to typedependent put options and is not always true in the general case.
17
4.2.1
An existence result
Let {vn } be a minimizing sequence for the principal’s optimization problem. The functions vn
are uniformly bounded and uniformly equicontinuous so we may with no loss of generality assume
that vn → v uniformly. Recall this also implies a.s. convergence of the derivatives. By dominated
convergence and the continuity of F, along with the fact that W is bounded yields
Z
Z
′
+
(F (v ′ (θ)) − |W |)+ d P-a.s.
(F (vn (θ)) − |W |) d −→
Θ
Θ
and
lim
n→∞
Z
E[(F (vn′ (θ)) − |W |)+ dθ =
Z
E[(F (v ′ (θ)) − |W |)+ dθ
Θ
Θ
This shows that the principal’s positions converge almost surely and hence in probability. Since
̺ is lower-semi-continuous with respect to convergence in probability we deduce that v solves the
principal’s problem.
4.2.2
An algorithm for approximating the optimal solution
We close this paper with a numerical approximation scheme for the principal’s optimal solution
within the pit option framework. We assume the set of states of the World is finite with cardinality
m. Each possible state ωj can occurs with probability pj . The realizations of the principal’s wealth
are denoted by W = (W1 , . . . , Wm ). Note that p and W are treated as known data. We implement
a numerical algorithm to approximate a solution to the principal’s problem when she evaluates
risk via the risk measure
m
X
X(ωj )pj qj ,
̺(X) = − sup
q∈Qλ j=1
where
−1
Qλ := q ∈ Rm
.
+ | p · q = 1, qj ≤ λ
We also assume the set of agent types is finite with cardinality n, i.e. θ = (θ 1 , . . . , θn ). The
density of the types is given by M := (M1 , . . . , Mn ). In order to avoid singular points in the
principal’s objective function, we approximate the option’s payoff function f (x) = (K − x)+ by
the differentiable function
if x ≤ K − ǫ,
0,
T (x, K) =
S(x, K), if K − ǫ < x < K + ǫ,
x − K,
if x ≥ K + ǫ.
where
S(x, K) =
x2 ǫ − K
K 2 − 2Aǫ + ǫ2
+
x+
.
4ǫ
2ǫ
4ǫ
18
The algorithm uses a penalized Quasi-Newton method, based on Zakovic and Pantelides [20],
to approximate a minimax point of
n
n
n
n
n−1
X
X
X
1 X X
1
Wi p i q i +
F (v, K, q) = −
T (Kj − |Wi |) pi qi −
T (Kj − |Wi |) pi
n
n
i=1
i=1
i=1
j=1
j=1
n
1X
1
vn − vn−1
vi+1 − vi
+
+
vn −
vi − θ i
n
θi+1 − θi
n
1 − θ n−1
i=1
where v = (v1 , . . . , vn ) stands for the values of a convex, non-increasing function, K = (K1 , . . . , Kn )
denotes the vector of type dependent strikes and the derivatives v ′ (θi ) are approximated by
v ′ (θ i ) =
vi+1 − vi
.
θi+1 − θ i
The need for a penalty method arises from the fact that we face the equality constraints
= −V ar[(K(θ) − |W |)+ ] and p · q = 1. In order to implement a descent method, these
constraints are relaxed and a penalty term is added. We denote by ng the total number of
constraints. The principal’s problem is to find
v ′ (θ)
min max F (v, K, q)
(v,K) q∈Qλ
subject to
G(v, K, q) ≤ 0
where G : R2n+m → Rng determines the constraints that keep (v, K) within the set of feasible
contracts and q ∈ Qλ . The Maple code for our procedure is given in the appendix for completeness.
Example 4.5 Let us illustrate the effects of risk transfer on the principal’s position in two
model with five agent types and two states of the world. In both cases W = (−1, −2), θ =
(1/2, 5/8, 3/4, 7/8, 1) and λ = 1.1 The starting values v0 , q0 and K0 we set are (4, 3, 2, 1, 0), (1, 1)
and (1, 1, 1, 1, 1) respectively.
i) Let p = (0.5, 0.5) and the types be uniformly distributed. The principal’s initial evaluation
of her risk is 1.52. The optimal function v and strikes are:
V1
0.1055
K1
1.44
V2
0.0761
K2
1.37
V3
0.0501
K3
1.07
V4
0.0342
K4
1.05
V5
0.0195
K5
1.05
The Principal’s valuation of her risk after the exchanges with the agents decreases to 0.2279.
ii) In this instance p = (0.25, 0.75) and M = (1/15, 2/15, 3/15, 4/15, 5/15). The principal’s
initial evaluation of her risk is 1.825. The values for the discretized v the type-dependent
strikes are:
19
1.5
0.11
0.1
1.45
0.09
1.4
0.08
1.35
0.07
1.3
0.06
1.25
0.05
1.2
0.04
1.15
0.03
1.1
1.05
0.02
1
1.5
2
2.5
3
3.5
4
4.5
0.01
5
1
1.5
(a) The type-dependent strikes.
2
2.5
3
3.5
4
4.5
5
(b) The optimal function v.
Figure 1: Optimal solution for underwriting put options, Case 1.
V1
0.0073
K1
1.27
V2
0.0045
K2
1.16
V3
0.0029
K3
1.34
V4
0.0026
K4
0.11
V5
0.0025
K5
0.12
The Principal’s valuation of her risk after the exchanges with the agents is 0.0922.
−3
1.4
7
x 10
6.5
1.2
6
1
5.5
0.8
5
0.6
4.5
4
0.4
3.5
0.2
0
3
1
1.5
2
2.5
3
3.5
4
4.5
2.5
5
(a) The type-dependent strikes.
1
1.5
2
2.5
3
3.5
4
4.5
(b) The optimal function v.
Figure 2: Optimal solution for underwriting put options, Case 2.
20
5
5
Conclusions
In this paper we analyzed a screening problem where the principal’s choice space is infinite dimensional. Our motivation was to present a nonlinear pricing scheme for over-the-counter financial
products, which she trades with a set of heterogeneous agents with the aim of minimizing the
exposure of her income to some non-hedgeable risk. In order to characterize incentive compatible
and individually rational catalogues, we have made use of U-convex analysis. To keep the problem tracktable we have assumed the agents have mean-variance utilities, but this is not necessary
for the characterization of the problem. Considering more general utility functions is an obvious
extension to this work. Our main result is a proof of existence of a solution to the principal’s
risk minimization problem in a general setting. The examples we have studied suggest that the
methodologies for approaching particular cases are highly dependent on the choice of risk measure, as well as on the kinds of contracts the principal is willing (or able) to offer. In most cases
obtaining closed form solutions is not possible and implementations must be done using numerical
methods. As a work in progress we are considering agents with heterogenous initial endowments
(or risk exposures), as well as a model that contemplates an economy with multiple principals.
A
Coherent risk measures on L2.
In this appendix we recall some properties and representation results for risk measures on L2
spaces; we refer to the textbook of Föllmer and Schied [12] for a detailed discussion of convex risk
measures on L∞ and to Cheridito and Tianbui [8] for risk measures on rather general state spaces.
Bäuerle and Müller [4] establish representation properties of risk law invariant risk measures on Lp
spaces for p ≥ 1. We assume that all random variables are defined on some standard non-atomic
probability space (Ω, F, P).
Definition A.1 (i) A monetary measure of risk on L2 is a function ̺ : L2 → R ∪ {∞} such
that for all X, Y ∈ L2 the following conditions are satisfied:
• Monotonicity: if X ≤ Y then ̺(X) ≥ ̺(Y ).
• Cash Invariance: if m ∈ R then ̺(X + m) = ̺(X) − m.
(ii) A risk measure is called coherent if it is convex and homogeneous of degree 1, i.e., if the
following two conditions hold:
• Convexity: for all λ ∈ [0, 1] and all positions X, Y ∈ L2 :
̺(λX + (1 − λ)Y ) ≤ λ̺(X) + (1 − λ)̺(Y )
21
• Positive Homogeneity: For all λ ≥ 1
̺(λX) = λ̺(X).
(iii) The risk measure is called coherent and law invariant, if, in addition,
ρ(X) = ρ(Y )
for any two random variables X and Y which have the same law.
(iv) The risk measure ̺ on L2 has the Fatou property if for any sequence of random variables
X1 , X2 , . . . that converges in L2 to a random variable X we have
ρ(X) ≤ lim inf ρ(Xn ).
n→∞
Given λ ∈ (0, 1], the Average Value at Risk of level λ of a position Y is defined as
1
AV @Rλ (Y ) := −
λ
Z
λ
qY (t)dt,
0
where qY (t) is the upper quantile function of Y . If Y ∈ L∞ , then we have the following characterization
AV @Rλ (Y ) = sup −EQ [Y ]
Q∈Qλ
where
dQ
1
Qλ = Q << P |
≤
.
dP
λ
Proposition A.2 For a given financial position Y ∈ L2 the mapping λ 7→ AV @Rλ (Y ) is decreasing in λ.
It turns out the Average Value of Risk can be viewed as a basis for the space of all lawinvariant, coherent risk measures with the Fatou property. More precisely, we have the following
result.
Theorem A.3 The risk measure ̺ : L2 → R is law-invariant, coherent and has the Fatou Property if and only if ̺ admits a representation of the following form:
Z 1
̺(Y ) = sup
AV @Rλ (Y )µ(dλ)
µ∈M
0
where M is a set of probability measures on the unit interval.
As a consequence of Proposition A.2 and Theorem A.3 we have the following Corollary:
22
Corollary A.4 If ̺ : L2 → R is a law-invariant, coherent risk measure with the Fatou Property
then
̺(Y ) ≥ −E[Y ].
An important class of risk measures are comonotone risk measures risk. Comonotone risk
measures are characterized by the fact that the risk associated with two position whose payoff
“moves in the same direction” is additive.
Definition A.5 A risk measure ̺ is said to be comonotone if
̺(X + Y ) = ρ(X) + ρ(Y )
whenever X and Y are comonotone, i.e., whenever
(X(ω) − X(ω ′ ))(Y (ω) − Y (ω ′ )) ≥ 0
P-a.s.
Comonotone, law invariant and coherent risk measures with the Fatou property admit a
representation of the form
Z 1
̺(Y ) =
AV @Rλ (Y )µ(dλ).
0
B
Maple code for the example of Section 4.2.2
with(LinearAlgebra)
n := 5:
m := 2:
ng := 2*m+4*n+1:
This section constructs the objective function f and its gradient.
x := Vector(2*n, symbol = xs):
q := Vector(m, symbol = qs):
f :=
add(−G[j] ∗ p[j] ∗ q[j], j = 1..m)+
add(add(T (x[i + n], W [j]) ∗ M [i] ∗ p[j] ∗ q[j], i = 1..n), j = 1..m)−
add(add(T (x[i + n], W [j]) ∗ M [i] ∗ p[j], i = 1..n), j = 1..m)+
add((x[i] − t[i] ∗ (x[i + 1] − x[i])/(t[i + 1] − t[i])) ∗ M [i], i = 1..n − 1)+
x[n] − t[n] ∗ (x[n] − x[n − 1]) ∗ M [n]/(t[n] − t[n − 1]) :
T T := (x, K)− > 1/4 ∗ x2 /eps + 1/2 ∗ (eps − K) ∗ x/eps + 1/4 ∗ (K 2 − 2 ∗ K ∗ eps + eps2 )/eps :
T := (x, K)− > piecewise(x ≤ K − eps, 0, x < K + eps, T T (x, K), K + eps ≤ x, x − K) :
gradfx := 0:
gradfq := 0:
23
for i from 1
do gradfx[i]
end do:
for i from 1
do gradfq[i]
end do:
to 2*n
:= diff(f, x[i])
to m
:= diff(f, q[i])
This section constructs the constraint function g and its gradient.
g := Vector(ng, symbol = tt):
for j from 1 to n
do g[j] := -x[j] end do:
for j from 1 to n-1
do g[j+n] := x[j+1]-x[j]
end do:
for i from 1 to n-1
do g[i+2*n-1] := add(T(x[i+n], W [j])2 ∗ p[j], j = 1..m) − add(T (x[i + n], W [j]) ∗ p[j], j = 1..m)2 + (x[i +
1] − x[i])/(t[i + 1] − t[i]) − eps2
end do:
g[3 ∗ n − 1] := add(T (x[2 ∗ n], W [j])2 ∗ p[j], j = 1..m) − add(T (x[2 ∗ n], W [j]) ∗ p[j], j = 1..m)2 + (x[n] −
x[n − 1])/(t[n] − t[n − 1]) − eps2 :
for i from 1 to n-1
do g[i + 3 ∗ n − 1] := −add(T (x[i + n], W [j])2 ∗ p[j], j = 1..m) + add(T (x[i + n], W [j]) ∗ p[j], j = 1..m)2 −
(x[i + 1] − x[i])/(t[i + 1] − t[i]) − eps2
end do:
g[4 ∗ n − 1] := −add(T (x[2 ∗ n], W [j])2 ∗ p[j], j = 1..m) + add(T (x[2 ∗ n], W [j]) ∗ p[j], j = 1..m)2 − (x[n] −
x[n − 1])/(t[n] − t[n − 1]) − eps2 :
g[4*n] := add(p[i]*q[i], i = 1 .. m)-1+eps3:
g[4*n+1] := -add(p[i]*q[i], i = 1 .. m)-1-eps3:
for i from 1 to m
do g[i+4*n+1] := -q[i]
end do:
for i to m
do g[i+m+4*n+1] := q[i]-lambda
end do: gradgx := 0:
gradgq := 0:
for i from 1 to ng
do for j from 1 to 2*n
do gradgx[i, j] := diff(g[i], x[j])
end do:
end do:
24
for i from 1 to ng
do for j from 1 to m
do gradgq[i, j] := diff(g[i], q[j])
end do:
end do:
This section constructs the slackness structures.
e := Vector(ng, 1):
s := Vector(ng, symbol = si):
z := Vector(ng, symbol = zi):
S :=DiagonalMatrix(s):
Z := DiagonalMatrix(z):
This section initializes the variable and parameter vectors.
x := Vector(2*n, symbol = xs):
q := Vector(m, symbol = qs):
p :=Vector(m, symbol = ps):
W := Vector(m, symbol = ws):
G := Vector(m,symbol = gs):
t := Vector(n, symbol = ts):
M := Vector(n, symbol = ms):
chi := convert([x, q, s, z], Vector):
This section constructs the Lagrangian and its Hessian matrix.
F :=convert([gradfx+(VectorCalculus[DotProduct])(Transpose(gradgx), z),
gradfq-(VectorCalculus[DotProduct])(Transpose(gradgq), z),
(VectorCalculus[DotProduct])((VectorCalculus[DotProduct])(Z, S), e)-mu*e, g+s], Vector):
DF := 0:
for i from 1 to 2*n+m+2*ng
do for j from 1 to 2*n+m+2*ng
do DF[i, j] :=diff(F[i], chi[j])
end do:
end do:
This section inputs the initial values of the variables and the values of the parameters.
xinit := (4,3, 2, 1, 0, 1, 1, 1, 1, 1):
qinit := (1, 1):
25
sinit := (.1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1, .1):
zinit := (1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1):
tinit := (1/2, 5/8, 3/4, 7/8,1):
pinit := (.5, .5):
ginit := (-1, -2):
minit := (1,1, 1, 1, 1):
winit := (1, 2):
tau := .1; mu := .1; rho := .5: lambda := 1.1: eps := .1: eps2 := .1: eps3 := .2:
xo := xinit; qo :=qinit; so := sinit; zo := zinit; po := pinit; go := ginit; mo := minit; wo := winit;
xs := xinit; qs := qinit; si := sinit; zi := zinit; ps := pinit; gs := ginit; ms := minit; ws := winit; ts := tinit:
The following section contains the executable code.
i := 0; j := 0
normF:=Norm(F,2):
while (normF > mu and j < 40)
do
printf(” Inner loop: #Iteration = %g /n”,j);
printf(” - Solve Linear System (11) ...”);
d := LinearSolve(DF, -Transpose(F)):
printf(”doneN”);
printf(” - Update Params ...”);
alphaS:=min(seq(select(type,-s[k]/d[2*n+m+k],positive), k = 1 .. Dimension(s)));
alphaZ:=min(seq(select(type,-z[k]/d[2*n+m+ng+k],positive), k = 1 .. Dimension(z)));
alphamax:=min(alphaS,alphaZ):
alphamax:=min(tau*alphamax,1);
printf(”done/n”);
chiold := chi;
chinew := chiold+alphamax*d;
xs := chinew[1 .. 2*n]: ys := chinew[2*n+1 .. 2*n+m]:
si := chinew[2*n+m+1 .. 2*n+m+ng]: zi := chinew[2*n+m+ng+1 .. 2*n+m+2*ng]:
normF:=Norm(F,2);
printf(”doneN”);
j:=j+1;
end do:
26
References
[1] Armstrong, M.: Multiproduct Nonlinear Pricing, Econometrica, 64, 51-75, 1996.
[2] Barrieu, P. & N. El Karoui: Optimal Design of Derivatives in Illiquid Framework,
Quantitative Finance, 2, 1-8, 2005.
[3] Barrieu, P. & N. El Karoui: Inf-convolution of risk measures and optimal risk transfer, Finance and Stochastics, 9, 269-298, 2005.
[4] Bäuerle, N. & A. Müller: Stochastic Orders and Risk Measures: Consistency and
Bounds, Insurance: Mathematics & Economics,38, 132-148, 2006.
[5] Becherer, D.: Utility indifference Hedging and Valuation via Reaction Diffusion Systems, Proceedings of the Royal Society, Series A, 460, 27-51, 2004.
[6] Becherer, D.: Bounded Solutions to Backward SDE’s with Jumps for Utility Optimization and Indifference Hedging, Annals of Applied Probability, to appear.
[7] Carlier, G., Ekeland, I & N. Touzi: Optimal Derivatives Design for Mean-Variance
Agents under Adverse Selection, Preprint, 2006.
[8] Cheridito, P. & L. Tianhui: Monetary Risk Measures on Maximal Subspaces Of Orlicz
Classes, Preprint, 2006.
[9] Davis, M.: Pricing Weather Derivatives by Marginal Value, Quantitative Finance, 1,
305–308, 2001.
[10] Ekeland, I., Témam, R., Convex Analysis and Variational Problems, Classics in Applied
Mathematics, 28, SIAM, 1976.
[11] Filipović, D. & M. Kupper: Equilibrium and Optimality for Monetary Utility Functions under Constraints, Preprint, 2006.
[12] Föllmer, H. & A. Schied: Stochastic Finance. An Introduction in Discrete Time, de
Gruyter Studies in Mathematics, 27, 2004.
[13] Guesnerie, R.: A contribution to the Pure Theory of Taxation, Econometrica, 49, 33-64,
1995.
[14] Horst U. & M. Müller: On the Spanning Property of Risk Bonds Priced by Equilibrium, Mathematics of Operations Research, to appear.
[15] Horst U., Pirvu, T. & G. Nunes dos Reis: On Securitization, Market Completion
and Equilibrium Risk Transfer, Working Paper, 2007.
27
[16] Hu, Y., Imkeller, P. & M. Müller: Market Completion and Partial Equilibrium,
International Journal of Theoretical & Applied Finance, to appear.
[17] Jouini, E., Schachermeyer, W. & N. Touzi: Law Invariant Risk Measures have the
Fatou Property, Advances in Mathematical Economics 9, 49-71, 2006.
[18] Mussa M. & S. Rosen: Monopoly and Product Quality, Journal of Economic Theory,
18, 301-317, 1978.
[19] Rochet, J.-C. & P. Choné: Iroining, Sweeping and Multidimensional Screening, Econometrica,66, 783-826, 1988.
[20] Zakovic, S. & C. Pantelides: An Interior Point Algorithm for Computing Saddle
Points of Constrainde Continuous Minimax, Annals of Operations Research, 99, 59-77,
2000.
28
| 5cs.CE
|
Experience-based Optimization: A Coevolutionary
Approach
arXiv:1703.09865v1 [cs.NE] 29 Mar 2017
Shengcai Liua , Ke Tanga,∗, Xin Yaob
a USTC-Birmingham Joint Research Institute in Intelligent Computation and Its
Applications (UBRI), School of Computer Science and Technology, University of Science
and Technology of China, Hefei, Anhui, China
b Department of Computer Science and Engineering, Southern University of Science and
Technology, Shenzhen, China
Abstract
This paper studies improving solvers based on their past solving experiences,
and focuses on improving solvers by offline training. Specifically, the key issues
of offline training methods are discussed, and research belonging to this category
but from different areas are reviewed in a unified framework. Exsiting training
methods generally adopt a two-stage strategy in which selecting the training instances and training instances are treated in two independent phases. This paper
proposes a new training method, dubbed LiangYi, which addresses these two issues simultaneously. LiangYi includes a training module for a population-based
solver and an instance sampling module for updating the training instances.
The idea behind LiangYi is to promote the population-based solver by training
it (with the training module) to improve its performance on those instances
(discovered by the sampling module) on which it performs badly, while keeping
the good performances obtained by it on previous instances. An instantiation
of LiangYi on the Travelling Salesman Problem is also proposed. Empirical
results on a huge testing set containing 10000 instances showed LiangYi could
train solvers that perform significantly better than the solvers trained by other
state-of-the-art training method. Moreover, empirical investigation of the be∗ Corresponding
author
Email addresses: [email protected] (Shengcai Liu), [email protected] (Ke
Tang), [email protected] (Xin Yao)
Preprint submitted to Journal of LATEX Templates
March 30, 2017
haviours of LiangYi confirmed it was able to continuously improve the solver
through training.
Keywords: experience, combinatorial optimization, offline training,
coevolution, dynamic sampling, parallel solvers
1. Introduction
Hard optimization problems (e.g., NP-hard problems) are ubiquitous in AI
research and real-world applications. To tackle them, numerous problem solvers
have been proposed over the last few decades [? ]. In general, such a solver is
designed for a certain problem domain rather than a single instance, because
when used in practice, it usually needs to solve many different instances belonging to that domain. Besides, many of such problem solvers are heuristic
methods, the performance (e.g., time complexity required to obtain the optimal
solution) of which can hardly be rigorously proved. As a result, the development of these solvers typically involves repeatedly testing it against a number
of problem instances and adjusting it based on the testing results [? ].
Given that both the design and the applications of a solver would involve
many problem instances, a natural question is whether a solver could leverage on the experience acquired from solving previous problem instances to
grow/enhance its capacity in solving new coming problem instances. For the
sake of brevity, a solver whose performance can improve as it solves more and
more problem instances is termed as an Experience-based Optimizer (EBO) in
this work.
The intuitions behind EBO are two-fold. First, any human expert in a
specific domain starts as a novice and his/her path to an expert mainly relies on
the gradual accumulation of problem-solving experience in this domain. Second,
exploiting past experience to facilitate the solving of new problems, from a more
technical point of view, concerns the generalization of past experience, which lies
in the heart of AI research, particularly the machine learning sub-area. The past
few decades have witnessed great progress on this issue, while most successes
2
were achieved on building a learner that can correctly map an input signal (e.g.,
an image) to a predefined output (e.g., a label). It is interesting to ask whether
similar idea could be developed to encompass more complex problem such as
NP-hard optimization problems, which may introduce new challenges as the
desired output will no longer be a label (or other types of variables), but a
solution to the optimization problem.
EBO could offer three advantages in practice. First, it enables an automated
process analogous to life-long learning of we human beings and thus alleviate
the tedious step-by-step fine-tuning or upgrade work that is now mostly done by
human experts. Second, as an EBO can improve its performance automatically,
it would be able to better exploit nowadays high-performance computing facilities to generate and test much more problem instances than a domain expert
can do manually, such that the risk of over-tuning the solver to a small set of
problem instances can be reduced. Finally, the underlying properties of realworld hard optimization problem instances, even if they are from exactly the
same problem class, may change over time. Since EBO dynamically updates
itself when solving more and more problem instances, it would better fit the
changing world.
Analogous to many machine learning techniques, an EBO may run in two
manners, i.e., the offline and online modes. For the offline mode, a set of problem
instances is fed to the EBO at a time and the solver is updated after collecting
the solutions it obtained on all the instances. For the online mode, problem
instances are fed to EBO one at a time and the solver is updated immediately
after solving an instance. Being conceptually different, the offline mode might
play a more important role in EBO since a set of training instances are usually
available when designing a solver. On comparison, the online mode is more
likely to occur after the EBO is deployed in a real-world application. Even
in this case, offline mode could also be adopted since EBO simply needs to
postpone its update until collecting more instances. For these considerations,
we focus on the offline mode as the first step to investigate EBO. Specifically,
this work consists of three main parts, as summarized below:
3
(1) An overview on the key issues of EBO: In the literature, there have been
several pioneer attempts to design mechanisms that enhance a solver based
on past experience. For examples, Transfer methods [? ? ? ], as the name
implies, transfer the useful information extracted from solved instances to
unsolved problem instances. Automatic algorithm configuration [? ? ],
Portfolio-based algorithm selection [? ? ? ], and Automatic portfolio construction [? ? ? ? ] seek to identify better parameter settings, algorithm
selectors, and portfolios of algorithms, respectively, based on historical
data. Although all these methods are within the scope of EBO and thus
relevant to one another, they were developed through independent paths
and have never been discussed in a unified context. We first bring together the existing literature on the offline scenario of EBO, and review
them under the unified umbrella of EBO, so as to make the key issues for
EBO clearer.
(2) A new offline training approach for EBO. A (and probably the most fundamental) form of EBO is to train it with many instances of the problem
of interests, so as to obtain a well-developed solver before deployment.
This scenario involves at least two questions, i.e., where comes the training instances and how the solver is adapted (trained) to the training instances. These two issues were usually treated through two independent
phases and seldom addressed simultaneously in the literature. We argue
that they are inter-correlated and propose a co-evolutionary framework,
namely LiangYi to address them as a whole.
(3) A case study of LiangYi on Travelling Salesman Problem (TSP). To assess
the potential of LiangYi, a specific instantiation of it is implemented based
on the Chained Lin-Kernighan (CLK) algorithm for the TSP. Empirical
studies are conducted to compare LiangYi to other state-of-the-art methods for fine-tuning CLK algorithm, as well as to investigate the properties
of LiangYi.
The rest of this paper is structured as follows. Section 2 first gives a definition
4
of the offline training in EBO and presents the key issues for describing the offline
training methods, and then review the existing methods. Section 3 presents the
approach LiangYi. Section 4 instantiates LiangYi on the TSP and reports the
empirical results. Section 5 concludes the paper and outlines directions for
future research.
2. Offline Training Methods in EBO
Given an optimization task T and a performance metric m, the training in
EBO is defined as making a solver s improve at optimization task T with respect
to performance metric m through experience E. This definition borrows some
basic concepts (i.e., T , m, E) from the definition of machine learning by Mitchell
[? ], yet each of them has a concrete meaning here. Specifically, the optimization
task T is conceptually an instance set containing all the target instances to which
the output solver is expected to be applied. The performance metric m is userspecified and it is often related to the computational resources consumed by
the solver (such as runtime or memory) or the quality of the solution found.
Basically, the experience E that is produced in the training is the processes of
the solver s solving training instances.
To improve the solver s at task T , a training method must consider two
things: How to use the solver s and the training instances to produce useful
experience and how to exploit the experience so as to enhance the solver s.
Generally, a solver s is comprised of multiple different parts (for example, a
search-based solver includes at least an initialization module and a search operator). It is conceivable that if the solver s is improved by the training, some
parts of s are necessarily changed during the training process. Based on these
analyses, we consider that an offline training method in EBO consists of three
essential parts:
• the form of the solver being trained;
• the settings of the training instances;
5
• the training algorithm that manipulates the solver and the training instances to produce experience, and exploits the experience to improve the
solver.
With this framework, we can describe offline training methods in EBO in a
unified way. In the combinatorial optimization field, there have been various
attempts by different communities to obtain solvers through training ways. The
next few sections review these researches with the presented framework in turn.
2.1. Automatic Algorithm Configuration Methods
The first class of methods are automatic algorithm configuration (AAC)
methods [? ? ]. AAC methods improve the solver (a parameterized algorithm)
at the optimization task by finding parameter values under which the solver
achieves high performance on the target instances. Specifically, AAC methods adopt a two-stage strategy. They first build a training set containing the
training instances that are representative of the target instances, and then run
the training algorithms to find high-performance parameter values on the training set. Due to the similarity between the training instances and the target
instances, the found parameter configurations are expected to perform well on
the target instances as well. A number of efficient methods have been developed
in the field of AAC, such as CALIBRA [? ], ParamILS [? ], GGA [? ], SMAC
[? ] and irace [? ]. With the framework presented previously, AAC methods
are expressed as follows:
• The solver s being trained is a parameterized algorithm.
1
• The efficacy of AAC methods depends greatly on the selection of the training instances, that is, the training instances should represent the target
1 Although there may be some significant differences between the parameterized algorithms
(in the aspects such as the types of the parameters, or the number of the parameters) that
different AAC methods can handle, we choose to ignore these details because what we want to
clarify here is which part of the solver is changed by the training, and the solver description,
i.e., a parameterized algorithm, is enough for this purpose. Such a simplicity principle also
applies in the reviews of other kinds of methods.
6
instances well so that the optimized performance on the training instances
could be favourably transferred to the target instances. The usual practice
in setting training instances for AAC methods [? ? ? ? ] is that the training instances are directly selected from some benchmarks, or are randomly
generated through some instance generators. Such practice is based on the
assumption that the selected benchmarks and generators could represent
the target scenarios to which the output solver will be applied to, particularly some practical application scenarios. This assumption however has
sparked some controversy [? ? ], which we will discuss more in Section 3.
• Essentially, in the training process, AAC methods run training algorithms
that keep testing different parameter configurations with the training set
and obtaining the testing results. The experience E produced in the training process is actually those tested parameter configurations and the corresponding testing results. The way of exploiting E so as to enhance the
solver is straightforward — using the best-performing one as the parameter configuration of the output solver (so the parameter configuration of
the solver is the part that is changed by the training).
When producing the experience, the training algorithms mainly need to
address such a problem: Which parameter configurations should be evaluated? Various AAC methods have proposed different solutions to this
problem. Among them the most notable ones include using the experimental design methods to generate candidate parameter configurations,
such as the case in CALIBRA [? ], searching the parameter space with a
meta-heuristic, such as the case in ParamILS [? ] (which uses an iterated
local search) and GGA [? ] (which uses a genetic algorithm), and adopting a model to bias the generation of candidate parameter configurations,
such as the case in SMAC [? ] (which builds an explicit regression model
to describe the dependence of algorithm performance on parameter settings) and irace [? ] (which builds a probability model over the parameter
space). In addition to determining the candidate parameter configurations
7
to be evaluated, some other important issues need to be addressed include
which instances are used to evaluate a parameter configuration and when
to terminate the evaluation of those poorly performing configurations. A
detailed review of these aspects in this area is beyond the scope of this
paper and one may refer to [? ? ] for a more comprehensive treatment
on the subject.
2.2. Portfolio-based Algorithm Selection Methods
The second class of methods are portfolio-based algorithm selection (P AS)
methods [? ? ? ]. Although there are different interpretations of this term
”portfolio” in the literature, we use it here to denote a solver that contains
several candidate algorithms and always selects one of them when solving a
problem instance. To improve the solver (an algorithm portfolio), unlike AAC
methods, P AS methods do not change the algorithms that constitute the solver,
but build a selector that can accurately select the best from the candidate algorithms for each instance. P AS methods adopt the same two-stage strategy as
AAC methods, except that the second-stage training algorithms in P AS methods are used to establish the selector. With the framework presented previously,
P AS methods are expressed as follows:
• The solver s being trained is an algorithm portfolio.
• The training instance settings for P AS methods are the same as AAS
methods.
• Although the training algorithms adopted by different P AS methods are
different in detail, their basic ideas are mostly the same. In the training process, the training algorithms first gather performance data of the
candidate algorithms on the training instances, and then build an algorithm selector based on the gathered data. The experience E produced in
the training is the performance data, which is obtained by running each
candidate algorithm on all training instances. The exploitation of E is
carried out in this way: First suitable features that characterize problem
8
instances are identified, and then the feature values of training instances
are computed; Once each training instance is represented by a vector of
feature values, the performance data is transformed into a set of training data, and based on these training data, machine learning techniques
are used to learn a mapping from instance features to algorithms. This
mapping is exactly the algorithm selector that will be used in the output
portfolio (so the algorithm selector of the solver is the part that is changed
by the training).
Various P AS methods build different models to do the mapping (selection), such as regression models [? ? ? ? ? ? ] (so-called empirical
performance models), classification models [? ? ? ? ? ? ] and ranking
models [? ].
2
For additional information one may refer to [? ] which
presents a survey of many approaches for algorithm selection from crossdisciplinary perspectives and [? ] which presents a survey focusing on the
contributions made in the area of combinatorial search problems.
2.3. Automatic Portfolio Construction Methods
The third class of methods are automatic portfolio construction (AP C)
methods, which seek to automatically build an algorithm portfolio based on
a parameterized algorithm from scratch. An AP C method can be seen as a
combination of an AAC method and a P AS method, that is, AP C methods simultaneously change the member algorithms and build the algorithm selectors.
There are two representative methods, dubbed Hydra [? ? ] and ISAC [? ? ], of
this category. Another class of portfolio construction methods, called automatic
parallel portfolio construction (AP P C) methods, differ from AP C methods in
that they seek to construct a parallel algorithm portfolio that runs all candidate
algorithms in parallel when solving an instance. In other words, AP P C methods also change the member algorithms of the solver, which makes them closely
2 Although many of the methods cited here are not originally proposed for combinatorial
optimization problems, the ideas behind them are very general and apply to the combinatorial
optimization problems as well.
9
related to AAC methods, but do not involve any algorithm selection. The representative methods of this category is ParHydra [? ]. Like AAC methods and
P AS methods, both AP C methods and AP P C methods also adopt a two-stage
strategy. With the framework presented previously, AP C methods and AP P C
methods are expressed as follows:
• The solver s being trained by AP C methods is an algorithm portfolio,
while the solver s being trained by AP P C methods is a parallel algorithm
portfolio.
• The training instance settings for AP C methods and AP P C methods are
the same as AAC methods.
• During the training process, AP C methods call an AAC method and a
P AS method as subroutines, while AP P C methods only call an AAC
method as a subroutine. Thus the experience collection and exploitation
in AP C methods and AP P C methods are done by their respective subroutines.
Different AP C methods mainly differ in how they specifically arrange the
calls of the AAC algorithm and the P AS algorithm. Hydra [? ] iteratively
runs them to build the portfolio. In the first iteration, it uses the AAC
algorithm to identify a parameter configuration with the best overall performance on the training set, and in each subsequent iteration, it uses the
AAC algorithm to find a parameter configuration that maximizes marginal
performance contribution across the configurations identified in the previous iterations. These marginal contributions are calculated relative to the
current algorithm selector, which is built by the P AS algortihm. A later
version of Hydra [? ] aims to find multiple algorithm configurations at
each iteration to speed up the training process. Unlike Hydra, ISAC [? ]
runs the P AS algorithm and the AAC algorithm sequentially. It first uses
a clustering algorithm (the P AS algorithm) to divide training instances
into clusters based on normalized instance features, thus building an algo-
10
rithm selector that is actually a 1-Nearest Neighbor model (this means at
runtime, for each instance the nearest cluster is determined and the corresponding algorithm is invoked). Then ISAC applies the AAC algorithm to
find a good algorithm configuration for each cluster. The later version of
ISAC [? ] extends ISAC by using the AAC algorithm on a set of different
parameterized algorithms to find a good algorithm configuration for each
instance cluster.
The representative AP P C method ParHydra [? ], like Hydra, also calls
the AAC algorithm iteratively. In each iteration ParHydra aims to find
an algorithm configuration that maximizes marginal contribution. Unlike
Hydra, ParHydra has no P AS algorithm involved and the marginal contribution is calculated relative to the current portfolio running in parallel.
A variant of ParHydra, dubbed ParHydrab , which is also proposed by [?
], extends ParHydra by using the AAC algorithm on a set of different
parameterized algorithms to identify an algorithm configuration in each
iteration.
2.4. Transfer Methods
The last class of methods are transfer methods, which explicitly extract useful information from solving processes of the training instances, and use it to
improve the performance of the solver on the target instances. The biggest difference between transfer methods and the methods reviewd above is that the
formaer collect experience based on every single instance, while the latetr collect based on a set of instances. Recall that AAC methods run the candidate
algorithms on a set of instances to evaluate them, and P AS methods run the
member algorithms on all training instances to collect the training data. On
comparasion, transfer methods extract information from each individual training instance, and the extraction of different instances is independent of each
other. Two specific transfer methods, dubbed XSTAGE [? ] and MEMETrans
11
[? ], are reviewed here 3 . Both of them enhance search-based solvers by introducing high-quality solutions during the solving processes. With the framework
presented previously, they are expressed as:
• The solver s being trained by XSTAGE is a multi-restart local search
algorithm in which a value function is used to determine the starting
point for each local search process.
The solvers s being trained by MEMETrans are search-based meta-heuristics.
• XSTAGE collects training instances and target instances from the same
benchmark, implicitly assuming these instances are similar enough to
make the transfer useful. MEMETrans collects training instances and
target instances from multiple heterogeneous benchmarks, and uses an
instance similarity measure to determine given a target instance, which
training instances will be used.
• Both XSTAGE and MEMETrans first run the solvers to solve the training
instances, producing the experience that is actually the solving processes
of the training instances. They differ in the exploitation of the experience. XSTAGE extracts a value functions from the solving process of
each training instance, and combine all of them into one value function
that determines starting solution for the solver. MEMETrans learns a
mapping from each solved training instance to corresponding optimal solution, and then combines all of them to be an initialization module that
helps generate high-quality initial solutions for the solver.
3. LiangYi
The two main issues of offline training are selecting the training instances
and training the solvers. The two-stage strategy, which is widely adopted by
3 In the paper [? ] that proposed XSTAGE, the term XSTAGE is used to denote the
composition of the training method and the solver. We use XSTAGE here only to denote
the training method. The term MEMETrans is created by us to denote the transfer method
proposed by [? ] since the paper does not give a term for this method.
12
existing training methods such as AAC methods [? ? ? ? ? ], P AS methods
[? ? ? ? ? ? ? ? ? ? ? ? ? ], AP C methods [? ? ? ? ? ] and transfer
methods [? ? ], treat them in two independent phases. However, intrinsically,
these two issues are correlated. From the perspective of developing solvers, the
greatest chance of the solver getting improved is on those problem instances
which the current solver cannot solve well 4 . Thus those hard instances for the
solver are best suited as the training instances. On the other hand, during the
training process, the solver is being adapted to the training instances; thereby
as the training proceeds, instances that are previously appropriate for the solver
may not be appropriate any more, in which case fixing the training instances
(as two-stage strategy methods do) is actually not helpful in improving the
solver, but will result in the waste of computing resources. A better strategy is
to dynamically changing the training instances during the training process to
keep them always appropriate (hard) for the solver, so that the solver can be
continuously improved.
As aforementioned, usually the training instances for two-stage methods are
directly selected from some benchmarks [? ? ], or are randomly generated
through some instance generators [? ? ? ? ? ? ? ? ? ? ], based on an
assumption that the benchmarks and generators could represent the target scenarios to which the solver is expected to be applied. However, such assumption
is not always tenable. On the contrary, in many cases, the commonly studied
benchmark instances and randomly generated instances lack diversity, are too
simple and rarely resemble real-world instances [? ? ]. Such risks could be
avoided by dynamically changing the training instances. First, this strategy selects training instances that are never easy for the solver. Second, this strategy
keeps changing the training instances, which naturally introduces the diversity.
Based on the above considerations, we propose a new training method,
4 This is also the key idea behind Hydra [? ? ], in which the training instances are given
different importance according to their hardness to the solver. However, such adjustment of
instance importance is still within a fixed training set.
13
dubbed LiangYi 5 . Basically, LiangYi is a competitive co-evolutionary [? ]
framework that alternately trains the solver and searches for new training instances. It maintains a set of algorithms and a set of instances, and each set
strives to improve itself against the other during the evolutionary process. The
details of this framework are elaborated below.
3.1. Design Principles
3.1.1. Basic Solver
The form of the solver being trained by LiangYi is a parallel portfolio that
runs all candidate algorithms in parallel when solving an instance 6 . In the coevolutionary framework, the parallel portfolio is called an algorithm population.
The reason for choosing an algorithm population other than a single algorithm
as the solver is simple: An algorithm population has the potential to achieve
better overall performance than a single algorithm [? ? ? ? ], since it is often
the case that for a problem domain there is no single best algorithm for all
possible instances, but different algorithms perform well on different problem
instances.
3.1.2. General Framework
Overall LiangYi consists of two components — The first component is a
training module for training the solver, and the second component is a sampling module which samples the target instance space for updating the training instances. These two modules are coupled in a co-evolutionary framework.
Specifically, LiangYi maintains two populations in which one population is an
algorithm population (the solver) and the other is an instance population (the
5 The name ”LiangYi” comes from the Taoism of Chinese philosophy. Generally, it means
two opposite elements of the world that interact and co-evolve with each other.
6 Employing parallel solvers to problem instances is an emerging area in training solvers
[? ]. Note that running all algorithms in parallel is different from an algorithm portfolio [?
? ? ], which typically involves some mechanism (e.g., selection [? ? ] or scheduling [? ])
to allocate computational resource to different algorithms. Running multiple algorithms in
parallel does not require any resource allocation as all involved algorithms are assigned with
the same amount of resource. This may lead to the waste of resource to some extent but can
keep the implementation simple.
14
training set). The training module and the sampling module are as two subprocedures that are performed alternatively by LiangYi to train the algorithm
population and to adjust the instance population, respectively. From the perspective of evolution, LiangYi is a process in which two populations are evolved
alternatively. At each iteration, LiangYi first evolves the algorithm population
(using the training module) to improve the performance of the algorithm population on the instance population while keeping the good performances obtained
by the algorithm population in previous iterations, and then evolves the instance
population (using the sampling module) to discover and include those instances
that cannot be solved well by the algorithm population, and then enters the
next iteration.
Intuitively, if we consider an instance in the target instance space ”covered”
by a solver as it can be solved well by the solver, the essence of LiangYi is
to enlarge the algorithm population’s coverage on the target instance space by
a) making the algorithm population cover the area that has not been covered
yet and b) keeping the algorithm population covering the area that has already
been covered. Figure 1 gives an intuitive visual example. In the example the algorithm population before training contains member algorithms that are neither
good nor complementary, and after training, the algorithm population obtain
new member algorithms that are both individually better and more complementary, which enlarges the coverage of the algorithm population a lot on the basis
of the original one.
3.2. Implementation Details
For the sake of brevity, henceforth we use abbreviations for the frequently
used terms. All the used abbreviations and corresponding terms are listed in
Table 1.
The pseudo code of LiangYi is outlined in Algorithm Framework 1. LiangYi
first randomly initializes AP as AP1 and IP as IP1 (Lines 1-2) and then enters
the iteration in which AP and IP are evolved alternatively (Lines 4-8). During
the k-th iteration, LiangYi first evolves AP from APk to APk+1 , then evolves
15
Coverage after training
Coverage before training
Training by CoEA
Target Instance Space
Target Instance Space
Figure 1: An example of LiangYi enlarging the coverage of the algorithm population on the
target instance space. The target instance space is the solid square. The algorithm population
contains 3 member algorithms. The coverage of each algorithm on the target insatnce space
is indicated by a black circle.
Table 1: Aberrations and corresponding terms
AP
IP
APk
IPk
alg
ins
EvolveAlg
EvolveIns
NAP
NIP
paramsAP
paramsIP
paramsR
P (solver, instanceset)
Aggr()
M
fAP (alg)
fIP (ins)
Algorithm population
Instance population
Algorithm population at the initial stage of the k-th
iteration of LiangYi
Algorithm population at the initial stage of the k-th
iteration of LiangYi
An algorithm belonging to the algorithm population
An instance belonging to the instance population
Training module of LiangYi
Sampling module of LiangYi
Number of algorithms in algorithm population
Number of instances in instance population
Parameter set that controls training module
Parameter set that controls sampling module
Parameter set that controls RemoveW orst procedure
Performance of the solver on the instance set
Aggregate function
Performance matrix of algorithm population on instance population
Fitness of algorithm alg
Fitness of instance Ins
16
IP from IPk to IPk+1 . After the k-th iteration, LiangYi enters the (k + 1)th iteration with AP as APk+1 and IP as IPk+1 . Finally, when LiangYi is
terminated, the AP is returned as the output solver (Line 9).
Algorithm Framework 1: LiangYi(NAP , NIP , paramsAP , paramsIP )
Input: Number of algorithms in AP , NAP , number of instances in IP , NIP ,
set of the parameters which control EvolveAlg, paramsAP , set of the parameters which control Evolveins, paramsIP
Output: Algorithm population APk
1. AP1 ← Randomly generate an algorithm population with NAP algorithms
2. IP1 ← Randomly generate an instance population with NIP instances
3. k ← 1
4. while not LiangY iT ermination() do
5.
APk+1 ← EvolveAlg(APk , IPk , k, paramsAP )
6.
IPk+1 ← EvolveIns(APk+1 , IPk , paramsIP )
7.
k ←k+1
8. end while
9. return APk
The training module and the sampling module in LiangYi are implemented
as two evolutionary procedures EvolveAlg and EvolveIns, respectively. In
general, any search method can be used in these two procedure. In this work,
evolutionary algorithms (EAs) are employed as the off-the-shelf tools, because
EAs are directly suitable for handling populations (of either algorithms or instances) and are less restricted by the properties of objective functions [? ]
(in comparison to other search methods such as gradient descent that requires
differentiable objective functions). Note that, in the process of LiangYi, when
training algorithm population or adjusting the instance population, the objective functions are a priori unknown and are changing in every iteration, so it
is a must that the two modules are able to cope with whatever objective landscape encountered. The behaviors of EvolveAlg and EvolveIns are controlled
by parameter sets paramsAP and paramsIP , respectively.
3.2.1. Preliminaries
Now we address three issues before going into the details of EvolveAlg and
EvolveIns. First of all, we explain how algorithms and instances are repre17
sented. Generally, when applying EAs, the choice of individual representation
would determine which genetic operators are applicable. In this paper. a fixedsize linear integer representation is chosen for both algorithms and instances,
namely, each algorithm and each instance is represented by an array of integers.
Although such representation might restrict the application scope of LiangYi,
it makes simple genetic operators [? ] (e.g., uniform crossover and uniform
mutation) directly applicable without any customization. For specific types of
algorithms and problems, more complex representations (e.g., tree-based representation) and corresponding genetic operators are needed, which will be a
subject of future research.
The second issue that needs to be addressed is to define the algorithms in
AP . Similar with automatic portfolio construction methods (see Section 2.3)
, LiangYi builds solver (the algorithm population) based on a parameterized
algorithm. In other words, each algorithm in AP is actually a parameter configuration which is a sample of the configuration space consisting of all possible
configurations of the parameterized algorithm.
The last issue that needs to be addressed is how to measure performance of
AP . For simplicity henceforth we use P (solver, instance set) to denote the performance of a solver on the instance set according to a given performance metric
m, in which the solver could be a single algorithm or an algorithm population
and the instance set could be a single instance or an instance population. Since
a population-based solver runs all member algorithms in parallel when solving
an instance, the performance of AP on an instance ins is the best performance
achieved by its member algorithms on ins (we assume a larger value is better
for m without loss of generality), i.e,
P (AP, ins) = max P (alg, ins).
alg∈AP
(1)
The performance of AP on IP is an aggregated value of the performance of AP
18
on each instance of IP , i.e.,
P (AP, IP ) = Aggr (P (AP, ins))
(2)
ins∈IP
where Aggr() is an aggregate function. The performance metric m and the
aggregate function Aggr() are user-specified.
3.2.2. Evolution of the Algorithm Population
The pseudo code of EvolveAlg is shown in Procedure 2. First of all, APk is
Procedure 2: EvolveAlg(APk , IPk , k, paramsAP )
Memory is a global cache which stores the final AP and final performance matrix
M of EvolveAlg of each iteration of LiangYi
Input: Algorithm population APk , instance population IPk , number of current iteration of LiangYi, k, parameter set paramsAP containing: crossover
probability croalg , mutation rate mualg , set of parameters which control
RemoveW orst, paramsR
Output: Algorithm population APk+1
1. M ← Test APk on IPk
2. while not EvolveAlgT ermination() do
3.
a1 , a2 ← Randomly select two individuals from APk
4.
anew ← Use a1 , a2 to generate an offspring by uniform crossover with
probability croalg and uniform mutation with rate mualg
5.
Mnew ← Test anew on IPk
6.
APk0 ← APk ∪ {anew }
7.
M 0 ← Concatenate M and Mnew
8.
APk , M ← RemoveW orst(APk0 , M 0 , k, paramsR )
9. end while
10. Add APk and M to M emory
11. APk+1 ← APk
12. return APk+1
tested on IPk and the result is represented by a NAP ×NIP matrix M (NAP and
NIP are the number of the algorithms in AP and the number of the instances
in IP , respectively), in which each row corresponds to an algorithm in APk
and each column corresponds to an instance in IPk , so peach entry in M is
the performance of the corresponding algorithm in APk on the corresponding
instance in IPk (Line 1). Then at each generation, two individuals are randomly
selected from APk and an offspring is generated by using uniform crossover and
19
mutation [? ] (Lines 3-4). The generated algorithm anew is then tested on
IPk and the result is represented by a 1 × NIP matrix Mnew (Line 5). The
last step is to decide the survivors of this generation (Lines 6-8). Together with
all the algorithms in APk , anew is put into a temporary algorithm population
AP 0 . The corresponding performance matrix M 0 to AP 0 , of which the size
is (NAP + 1) × NIP , is constructed by concatenating M and Mnew . Then
the procedure RemoveW orst is run to decide which algorithm in AP 0 will be
removed.
Basically, RemoveW orst firstly calculates the fitness of each algorithm in
AP 0 and then selects the algorithm with the lowest fitness to be removed. The
core of RemoveW orst is its fitness evaluation. The idea is that an algorithm will
be preferred only if it contributes to the algorithm population, and the more it
contributes, the more it is preferred. The contribution of a member algorithm
is actually the performance improvement it brings to the population, which
can be calculated as the population’s performance loss caused by removing the
algorithm. Formally, let C(AP, IP, alg) denote the contribution of algorithm
alg to the performance of AP on IP . If |AP | > 1, which means AP contains
other algorithms besides alg, C(AP, IP, alg) is calculated via Equation (3):
C(AP, IP, alg) = |P (AP, IP ) − P (AP − {alg}, IP )|,
(3)
where P (AP, IP ) and P (AP − {alg}, IP ) are calculated via Equation (2). If
|AP | = 1, which means AP only contains one algorithm (i.e., alg), then removing alg from AP will cause complete performance loss on IP . In this case,
C(AP, IP, alg) is calculated via Equation (4):
C(AP, IP, alg) = α|P (AP, IP )|,
(4)
where the parameter α > 0.
Based on Equation (4), for each member algorithm alg of the temporary
algorithm population APk0 , the performance contribution on instance popula-
20
tion IPk is C(APk0 , IPk , alg). A high contribution indicates that alg should be
reserved to next generation. However, directly using C(APk0 , IPk , alg) as the
fitness of alg is not appropriate. As aforementioned (see Section 3.1), the evolution of APk should not only improve the performance of APk on IPk , but also
keep the good performances obtained in previous iterations (on IP1 , ..., IPk−1 ).
Using C(APk0 , IPk , alg) to evaluate alg only considers the first target. Hence
the fitness of alg is calculated based on two types of contributions: The first one
is the current performance contribution, i.e., C(APk0 , IPk , alg), and the second
one is the historical performance contribution of alg on IP1 , ..., IPk−1 (if there
are).
To calculate the historical performance contribution of a member algorithm,
the concept age is introduced to describe how long the algorithm has been in
AP . Suppose a member algorithm alg was added to AP in the j-th iteration
of LiangYi (and now is in the k-th iteration), the age of alg is k − j. The
performances of alg on IPj ,...,IPk are known because alg has been tested on
them in corresponding iterations 7 . To calculate the historical performance contribution of alg on IPr (j ≤ r ≤ k − 1), the algorithms that satisfy the condition
age ≥ (k − r) are selected from the temporary algorithm population APk0 (our
target algorithm alg is also selected, since its age is k − j, which satisfies the
condition) to form a virtual algorithm population virtualAPr . The condition
age ≥ (k − r) indicates that these selected algorithms were added to AP during
or before the r-th iteration of LiangYi, so they have been tested on IPr . The
performances of these algorithms on IPr are represented by a |virtualAPr |×NIP
matrix virtualMr . If |virtualAPr | > 1, which means virtualAPr contains other
algorithms besides alg, the performance contribution of alg on IPr is calculated
7 Note that the performance contributions of alg on IP ,...,IP
1
j−1 are not considered in this
paper because the performances of alg on them are unknown. To obtain these performances,
we can store IP1 ,...,IPj−1 and test alg on them. However, this would make the computational
cost and the storage cost increase fast over time.
21
via Equation (3):
C(virtualAPr , IPr , alg) = |P (virtualAPr , IPr ) − P (virtualAPr − {alg}, IPr )|.
If |virtualAPr | = 1 , which means virtualAPr only contains alg, then removing
alg from virtualAPr will cause complete loss of performance on IPr . In this
case, C(virtualAPr , IPr , alg) is calculated via Equation (4):
C(virtualAPr , IPr , alg) = α|P (virtualAPr , IPr )|.
Now we have all the performance contributions of alg on IPj ,...,IPk . The
fitness of alg, denoted as fAP (alg), is calculated via Equation (5)
fAP (alg) =
βΣj≤r≤k−1 C(virtualAPr , IPr , alg) + C(APk0 , IPk , alg)
k−j+1
(5)
where k is the index of the current iteration of pLiangY i, j is the age of alg,
and β is a nonnegative parameter. The terms C(virtualAPr , IPr , alg)(j ≤
r ≤ k − 1) are historical performance contributions on IPj ,...,IPk−1 , while
C(APk0 , IPk , alg) is the current performance contribution on IPk . Thereby the
numerator in the fraction is actually a weighted sum of (k − j + 1) performance
contributions, in which the parameter β is used to balance between historical
performance contributions(on IP1 , ..., IPk−1 ) and current performance contribution (on IPk ).
The pseudo code of RemoveW orst is demonstrated in Procedure 3. First
the fitness of each member algorithm in APk0 is calculated (Lines 1-15). Specifically, for an algorithm alg which was added to AP in the j-th iteration of
LiangYi, k − j virtual algorithm populations, i.e., virtualAPj , ..., virtualAPk−1 ,
are constructed (according to the global cache Memory) to calculate its historical
performance contributions on IPj , ..., IPk−1 , via Equation (3) or Equation (4)
(Lines 2-12). Together with the current performance contribution calculated
via Equation (3) (Line 13), the historical contributions are used to calculate the
fitness of alg via Equation (5) (Line 14). After the fitness of each algorithm in
22
APk0 has been calculated, the algorithm with the lowest fitness will be removed
(Line 16).
Procedure 3: RemoveW orst(APk0 , M 0 , k, paramsR )
Input: Temporary algorithm population APk0 , temporary performance matrix
M 0 , number of current iteration of LiangYi, k, parameter set paramsR
containing α (used in Equation (4)) and β (used in Equation (5))
Output: Algorithm population APk , performance matrix M
1. for each algorithm alg in APk0 do
2.
age ← Query M emory for how many iterations alg has been staying
3.
j ← k − age
4.
for r ← j to k − 1 do
5.
V irtualAPr ← Select algorithms which satisfy the condition age ≥
(k − r) (according to M emory) from APk0 to form a virtual algorithm
population
6.
Mr ← Construct corresponding performance matrix to V irtualAPr
according to M emory
7.
if |virtualAPr | > 1 then
8.
C(virtualAPr , IPr , alg) ← Calculate the algorithm contribution of
alg on IPr via Equation (3)
9.
else
10.
C(virtualAPr , IPr , alg) ← Calculate the algorithm contribution of
alg on IPr via Equation (4)
11.
end if
12.
end for
13.
C(APk0 , IPk , alg) ← Calculate the algorithm contribution of alg on IPk
via Equation (3)
14.
fAP (alg) ← Calculate the fitness of alg via Equation (5)
15. end for
16. APk ← Remove the algorithm with the lowest fitness from APk0
17. M ← Remove the corresponding row in M 0 to the removed algorithm
18. return APk , M
3.2.3. Evolution of the Instance Population
As aforementioned (see Section 3.1) the evolution of IP aims at discovering
those instances that cannot be solved well by AP ; thus the fitness of an instance
in IP is measured by how AP performs on it — the worse the performance, the
higher the fitness.
The pseudo code of EvolveIns is demonstrated in Procedure 4. First of all,
APk+1 is tested on the IPk and the result is represented by a NAP ×NIP matrix
23
M (Line 1), and the fitness of each instance is calculated (Lines 2-4). The fitness
of an instance ins, denoted as fIP (ins), is calculated via Equation (6):
fIP (ins) = −P (APk+1 , ins),
(6)
where P (APk+1 , ins) is the performance of APk+1 on instance ins, calculated
via Equation (1). At each generation, NIP ∗ re new instances are generated
by repeatedly selecting two instances from IPk and creating two offsprings using uniform crossover with probability croins and uniform mutation with rate
muins [? ](Lines 6-11). These offsprings are then tested against the algorithm
population APk+1 and the fitness of each offspring is calculated (Lines 12-15).
At the end of this generation, all instances in IPk and the offsprings are put into
a candidate pool and the worst NIP ∗ re instances are removed (Lines 16-19).
4. Case Study: the Travelling Salesman Problem
The main purpose of this section is to empirically verify whether LiangYi is
an effective method for training solvers. We evaluated LiangYi on the Travelling Salesman Problem (TSP) [? ], one of the most well-known computationally
hard optimization problem. Specifically, the symmetric TSP, i.e., the distance
between two cities is the same in each opposite direction, with Euclidean distances in a two-dimensional space is considered here. In the remainder of this
section, we first give the target scenario (including the optimization task and the
performance metric) where LiangYi is applied, and then instantiates LiangYi
for the scenario. After that, we first compare LiangYi with other existing training methods, and then we investigate the properties of LiangYi to see whether
it is able to perform as expected in design.
4.1. Target Scenario
The optimization task T considered here are all TSP instances with problem
size equal to 500, i.e., the number of cities equals to 500. The number of the
24
Procedure 4: EvolveIns(APk+1 , IPk , paramsIP )
Input: Algorithm population APk+1 , instance population IPk , parameter set
paramsIP containing crossover probability mualg , mutation rate muins and
replacement ratio res
Output: Instance population IPk+1
1. M ← Test APk+1 on IPk
2. for each instance ins in IPk do
3.
fIP (ins) ← Calculate the fitness of ins via Equation (6)
4. end for
5. while not EvolveInsT ermination() do
6.
of f springs ← ∅
7.
for i ← 1 to NIP2∗res do
8.
ins1 , ins2 ← Select two parents from IPk with tournament selection
9.
insnew1 , insnew2 ← Use ins1 , ins2 to generate two offsprings by using
uniform crossover with probability croins and uniform mutation with
rate muins
10.
of f springs ← of f springs ∪ {insnew1 , insnew2 }
11.
end for
12.
Mof f springs ← Test APk+1 on of f springs
13.
for each instance ins in of f springs do
14.
fIP (ins) ← Calculate the fitness of ins via Equation (6)
15.
end for
16.
candidates ← IPk ∪ of f springs
17.
Mcandidates ← Concatenate M and Mof f spring
18.
IPk ← Remove the worst NIP ∗ re instances from candidates
19.
M ← Remove the corresponding columns to those removed instances from
Mcandidates
20. end while
21. IPk+1 ← IPk
22. return IPk+1
25
target instances are theoretically infinite. Training solvers on such kind of tasks
have not been studied before.
This work focuses on optimizing the applicability on the target instances,
i.e., the performance metric m is applicability. A solver is said to be applicable
to an instance if it can find a good enough solution to this instance within a given
time. For TSP, the goodness of a solution sol is measured by the percentage
by which the tour length of sol exceeds the tour length of the optimum sol? 8 ,
abbreviated as PEO(percentage excess optimum):
P EO =
lengh(sol) − lengh(sol? )
∗ 100%.
lengh(sol? )
With the definition of PEO, given a cut-off time t, a solver is said to be applicable
to an instance ins if the P EO of the best solution found by the solver in time t
is below a threshold θ. With the definition of the applicability of a solver to a
single instance, the applicability of a solver to an instance set is defined as the
proportion of the instances to which the solver is applicable.
In this paper very radical values for the cut-off time t and the PEO threshold
θ are adopted (t = 0.1s and θ = 0.05%) to see whether LiangYi is able to train
solvers that can work well under such harsh conditions.
4.2. Instantiating LiangYi
In order to instantiate LiangYi for the above scenario, there are several
issues to be addressed. The first issue is to specify the performance function
P (solver, instance set) used by LiangYi (see Section 3.2.1) so that LiangYi can
optimize the applicability appropriately. The performance of an algorithm alg
on an instance ins, i.e., P (alg, ins) in Equation (1), is specified as follow:
P (alg, ins) =
1, if alg is applicable to ins
0, otherwise.
8 The optimum sol? is obtained using Concorde [? ], a branch-and-cut based exact TSP
solver.
26
Table 2: The parameters of the CLK used
Parameters
Initialization Strategy
Perturbation Strategy
Search Depth
Search Width
Backtrack Strategy
Parameter Type
Categorical
Categorical
Numerical
Numerical
Categorical
# of Candidate Values
4
4
6
8
14
Intuitively, an algorithm population AP is said to be applicable to an instance
ins if any member algorithm of AP is applicable to ins. With P (alg, ins)
specified as above, this definition is equivalent to the definition given by Equation (1), namely, AP is applicable to ins if the best member algorithm of AP
is applicable to ins. The aggregate function aggr() in Equation (2) is specified
as returning mean value of the aggregated terms:
P
P (AP, IP ) =
P (AP, ins)
,
|IP |
ins∈IP
which essentially calculates the proportion of the instances to which AP is
applicable.
The second issue is to choose a parameterized algorithm for LiangYi to build
an algorithm population based on it. The choice of the parameterized algorithm
in this work is Chained Lin-Kernighan (CLK) [? ]. It is a variant of the LinKernighan heuristic [? ], one of the best heuristics for solving symmetric TSP.
CLK chains multiple runs of the Lin-Kernighan algorithm to introduce more
robustness in the resulting tour. Each run starts with a perturbed version of
the final tour of the previous run. We extended the original implementation of
CLK to allow a more comprehensive control on its components. The parameters
of the resulting algorithm are summarized in Table 2. To handle the randomness
of CLK, we adopt a simple way - fixing the random seed of CLK and turning it
into a deterministic algorithm.
The third issue is to concretely represent the algorithms and the instances
27
Table 3: The parameter settings of the instantiation of LiangYi for TSP
EvolveAlg
croalg = 0.6
mualg = 0.6
α=2
β=2
EvolveIns
croins = 1
muins = 0.8
res = 0.3
with fixed-size integer arrays (see Section 3.2.1). Each algorithm in AP is
represented by a list containing 5 integers, each of which indicates its value
for the corresponding parameter. Each instance in IP is represented by a list of
500 (x, y) coordinates on a 106 × 106 grids. The random initialization, uniform
crossover, and mutation are implemented as conventional ways for fixed-sizelist based representations [? ]. Specifically, the random initialization works
by uniform randomly selecting a value (i.e., two coordinates for the instance,
a parameter value for the algorithm) from candidate values for each entry of
the individual (the algorithm or the instance). The uniform crossover operates
by choosing for each entry of the offspring with equal probability either the
value of the entry from the first or the second parent. The mutation consists
of replacing the value of each entry of the offspring, with a probability, with
uniform randomly chosen one from the candidate values.
The last issue is to set the termination conditions and the parameters of
LiangYi. The termination condition for LiangYi is the number of iterations
reaching 3. In each iteration, procedure EvolveAlg will be run for 500 generations and EvolveIns will be run for 10 generations. The number of algorithms
in AP , i.e., NAP , is set to 6, and the number of instances in IP , i.e., NIP , is
set to 150. The parameter settings of LiangYi are listed in Table 3. In order
to build an algorithm population with complementary algorithms, the diversity
between these member algorithms need to be kept on a high level. For this reason, the mutation rate in evolvealg is set to a high value (0.4). For evolveins,
it is important to keep the instance population exploring the target instance
space instead of stagnating in some local areas, and therefore the mutation rate
28
Procedure 𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬
Procedure 𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬𝑬
𝐴𝐴𝐴𝐴1
𝐼𝐼𝐼𝐼1
𝐴𝐴𝐴𝐴2
𝐴𝐴𝐴𝐴2
𝐴𝐴𝐴𝐴2
𝐼𝐼𝐼𝐼1
𝐼𝐼𝐼𝐼2
𝐴𝐴𝐴𝐴3
𝐴𝐴𝐴𝐴3
𝐼𝐼𝐼𝐼2
1
𝐴𝐴𝐴𝐴3
𝐼𝐼𝐼𝐼2
𝐼𝐼𝐼𝐼3
𝐴𝐴𝐴𝐴4
𝐴𝐴𝐴𝐴4
𝐼𝐼𝐼𝐼3
2
# of iterations of CoEA
𝐼𝐼𝐼𝐼3
𝐼𝐼𝐼𝐼4
3
Figure 2: The training process of LiangYi in which AP and IP are evolved alternatively. Dotted rectangle and dashed rectangle represent procedure EvolveAlg and procedure EvolveIns,
respectively.
in evolveins is also set to a high value (0.7).
We applied the instantiation of LiangYi described above to the considered
scenario. The training process of LiangYi in which AP and IP are evolved
alternatively is depicted in Figure 2.
4.3. Comparative Study
In this section we compare LiangYi with other existing training methods
in the considered scenario. Since the algorithm population built by LiangYi
is actually a parallel portfolio, we chose ParHydra [? ] (see Section 2.3) , the
state-of-the-art automatic parallel portfolio construction method, as the method
to compare with.
4.3.1. Settings of ParHydra
ParHydra accepts a parameterized algorithm to be configured, a set of training instances, and a performance metric to be optimized. For the target scenario
considered, the performance metric is applicability. The parameterized algorithm fed to ParHydra is CLK, same as LiangYi. We used two different ways
to construct the training sets for ParHydra. The first training set IPrandom
29
was built according to the usual practice for two-stage methods — randomly
generating a set of instances. Specifically, each instance in IPrandom was generated by randomly choosing two coordinates for each city on a 106 × 106 grids.
The second training set IPtraining was built by collecting the instance populations which were produced by LiangYi and once served as the training instances
during the training process, i.e., IPtraining = IP1 ∪ IP2 ∪ IP3 . Since LiangYi
produces instance populations as by-products, it is interesting to see, how good
these instance populations are as training instances for existing methods like
ParHydra. Both sets contain 450 instances.
ParHydra is an iterative method, which builds the portfolio from scratch
and adds an algorithm to it in each iteration (See Section 2.3). Thus we set
the iteration number of ParHydra as 6 to keep in line with LiangYi in terms of
algorithm number. At each iteration of ParHydra, 15 copies of ParamILS [? ]
were run with different random seeds in parallel to identify an algorithm that
maximizes marginal performance contribution across the algorithms identified
in the previous iterations. The solvers output by ParHydra based on IPrandom
and IPtraining , are denoted as P Hrandom and P HLiangY i respectively.
4.3.2. Experimental Protocol
Since LiangYi and ParHydra are both stochastic methods, we ran each comparative method 20 times. Specifically, first we ran LiangYi 20 times. Then
we obtained 20 IPtraining from the 20 runs of LiangYi and randomly generated 20 different IPrandom , and then based on each of these 40 training set, we
ran ParHydra to obtain a solver. Finally we obtained 20 AP4 (the algorithm
population output by LiangYi), 20 P Hrandom and 20 P HLiangY i .
In order to adequately test the output solvers on target instance space, we
generated a huge testing set, denoted as IPtesting , which contains 10000 TSP
instances with the number of cities equal to 500. Specifically, each instance was
generated by randomly choosing two coordinates for each city from the interval
[0, 106 ). To our knowledge, this is the first time that a testing set of such a large
size (10000) is used to test TSP solvers.
30
All of our experiments were carried out on a workstation of a Xeon CPU
with 12 cores and 24 threads at 2.50GHz, running Ubuntu Linux 16.04. The
runtime requirements in CPU days were as follows: 120 days for CoEA training
(including 20 runs); 600 days for ParHydra training (including 20 runs); 4 days
for testing.
4.3.3. Experiment Results
The averaged testing results of each type of solvers are presented in Table 4. We performed 2-sample t-test to compare the results each type of solvers,
considering p-values below 0.05 to be statistically significant. The ranking in
Table 4 is AP4 , P HLiangY i , P Hrandom . P HLiangY i obtained better results
than P Hrandom , which indicates the training instances produced by LiangYi
are more diverse than random generated ones. It is a little surprising to see
AP4 obtained better results than P HLiangY i at first sight. Different from AP4 ,
P HLiangY i was trained based on the whole IPtraining , which was produced by
LiangYi iteration by iteration; thus it is expected that P HLiangY i would obtain
better results than AP4 on IPtraining (actually, their performances on IPtraining ,
i.e., P (AP4 , IPtraining ) and P (P HLiangY i , IPtraining ), are 0.6063 and 0.6644).
The reason why AP4 obtained better results than P HLiangY i on IPtesting is as
follow: The dynamic instance sampling strategy used by LiangYi is actually a
filter that only keeps those hard instances for AP to make the training focus
on them , which makes the actual coverage of AP on the target instances far
greater than its coverage on IPtraining (0.7001 > 0.6063) because those easy
target instances which are sampled by EvolveIns and are actually covered by
AP are all filtered out. Compared to LiangYi, ParHydra accepts all the training
instances and only focuses on the training set. The lack of the instance dynamic
sampling makes the performance of the output solver greatly depend on how
much the training set can represent the target instance space.
31
Table 4: Averaged performances of AP4 , P Hrandom , P HLiangY i on IPtesting . All results are
presented in terms of applicability, i.e., the proportion of the instances to which the solver
is applicable. We performed 2-sample t-test to compare their results, considering p-values
below 0.05 to be statistically significant. The second row provide the results of the test, where
0 W − D − L0 indicates how many competitors the corresponding solver is superior to, not
significantly different from or inferior to.
AP4
0.7001
2-0-0
P Hrandom
0.6600
0-0-2
P HLiangY i
0.6719
1-1-0
4.4. Investigating the Properties of LiangYi
As aforementioned, the idea behind LiangYi is to optimize the performance
of AP on target instances by a) improving its performances on those instances
on which it performs badly and b) keeping its good performances on those
instances on which it performs well. The main purpose of this section is to investigate whether LiangYi is able to accomplish the two objectives listed above.
Specifically, the verification is divided into two parts — the training part and
the testing part. In the training part we investigate that, in the training process,
whether LiangYi gives satisfactory answers to the following three questions:
(1) Whether is procedure evolvealg able to improve the performance of AP
on current IP ?
(2) Whether is procedure evolveins able to degrade the performance of AP
on current IP ?
(3) Whether is procedure evolvealg able to keep the performance of AP on
previous IP s?
The second question indicates whether the evolution of IP is able to discover and
include hard-to-solve instances to AP , and the first question indicates whether
the evolution of AP is able to improve the performance of AP on the hard
instances included in the current IP . The combination of these two checks the
whether LiangYi is able to accomplish the first objective. The third question
checks whether LiangYi is able to accomplish the second objective. In addition
to focusing on the three specific aspects, we also directly check if LiangYi is
32
1
0.9
0.8
Applicability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
3
# of iterations of CoEA
Figure 3: The performance of AP on IP at the beginning (EvolveAlg begins), the middle
(EvolveAlg ends) and the end (EvolveIns ends) of each iteration of LiangYi. The performances are represented in terms of applicability, i.e., the proportion of the instances to which
AP is applicable, and are averaged over 20 runs.
able to continuously improve AP in the training part. Specifically, we check
whether the performances of AP on IPtraining that are produced in the training
are improved by LiangYi. Similarly, in the testing part we also directly check
whether the performance of AP at the optimization task is being improved by
LiangYi.
4.4.1. Training Part
To answer the first question and the second question, the performances of
AP on IP during the training process averaged over 20 runs are plotted in
Figure 3. The results depicted in Figure 3 clearly show that, at each iteration
of LiangYi, EvolveAlg improves the performance of APk on IPk , and evolveins
degrades the performance of APk+1 on IPk , which gives positive answers to the
first two questions, thus confirming the first aspect of the idea behind LiangYi.
The third question is answered in this way: Since procedure evolvealg
evolved APk to APk+1 to improve the performance on IPk , we checked whether
the improvement from P (APk , IPk ) to P (APk+1 , IPk ), i.e., |P (APk , IPk ) −
P (APk+1 , IPk )|, was kept in subsequent iterations of LiangYi. Specifically,
33
Table 5: Performances of APk ,...AP4 on IPk . All the results are presented in terms of
applicability, i.e., the proportion of the instances to which AP is applicable, and are averaged
over 20 runs.
AP1
AP2
AP3
AP4
IP1
0.3800
0.8121
0.7421
0.6867
IP2
IP3
0.1456
0.6590
0.5667
0.1735
0.5654
we tested APk+2 , ..., AP4 on IPk to obtain their performances on IPk , i.e.,
P (APk+2 , IPk ), ..., P (A4 , IPk ), and calculated the performance drops from P (APk+1 , IPk )
to these performances, i.e., |P (APk+1 , IPk )−P (APk+2 , IPk )|,..., |P (APk+1 , IPk )−
P (A4 , IPk )|, then these performance drops were compared to the performance
improvement. The averaged performances (over 20 runs) of APk , ..., AP4 on
IPk are presented in Table 5. The averaged performance improvement on IP1
is P (AP2 , IP1 ) − P (AP1 , IP1 ) = 0.4321, and the two averaged performance
drops on IP1 are P (AP3 , IP1 ) − P (AP2 , IP1 ) = 0.0700 and P (AP4 , IP1 ) −
P (AP2 , IP1 ) = 0.1254, so the ratios between the performance drops and the
performance improvements on IP1 are 16.20% and 29.02%. Calculated in the
same way, the ratio on IP2 is 17.98%. All the ratios between the performance
drops and the corresponding performance improvements are below 30%.
In order to check whether the performances of AP on IPtraining are improved by LiangYi, the algorithm populations obtained from each iteration of
LiangYi, i.e., AP1 , AP2 , AP3 , AP4 , were tested on IPtraining . The testing results
averaged over 20 runs are depicted in Figure 4. A constant improvement of the
performances of APk on IPtraining , according to the increase of k, is shown.
4.4.2. Testing Part
The algorithm population obtained from each iteration of LiangYi, i.e.,
AP1 , ..., AP4 was tested on IPtesting . The testing results averaged over 20 runs
are depicted in Figure 5. Once again, a constant improvement of the performances of APk on IPtesting according to the increase of k is shown.
34
1
0.9
0.8
Applicability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
AP 1
AP 2
AP 3
AP 4
Algorithm population
Figure 4: The performances of AP1 , AP2 , AP3 , AP4 on IPtraining . The performances are
represented in terms of applicability, i.e., the proportion of the instances to which AP is
applicable, and are averaged over 20 runs.
0.7
0.65
Aplicability
0.6
0.55
0.5
0.45
0.4
AP 1
AP 2
AP 3
AP 4
Algorithm population
Figure 5: The performances of AP1 , AP2 , AP3 , AP4 on IPtesting . The performances are
represented in terms of applicability, i.e., the proportion of the instances to which AP is
applicable, and are averaged over 20 runs.
35
5. Conclusion
This paper first put forward the concept of Experience-based optimizer
(EBO) whose performance can improve as it solves more and more problem
instances, and summarized several previous research in a unified context, i.e.,
offline training of EBO. A new training method, dubbed LiangYi, was proposed.
The most novel feature of LiangYi is that, different from existing methods, it
addresses selecting training instances and training solvers simultaneously. A
specific instantiation of LiangYi on TSP was also proposed. Empirical results
showed the advantages of LiangYi in comparison to ParHydra, the state-ofthe-art AP C method, on a huge testing set containing 10000 instances. Moreover, through empirically investigating behaviours of LiangYi, we confirmed
that Liangyi is able to continuously improve the solver through training.
As discussed in the introduction, EBO is a far more broad direction than
merely offline training of problem solvers. Further investigations may include:
(1) Further improvements to LiangYi. Diversity preservation scheme, such as
speciation [? ] or negatively correlated search [? ] can be introduced into
LiangYi to explicitly promote cooperation between different algorithms
in AP . Another tack is to use machine learning techniques to accelerate
LiangYi. Specifically, regression models and classification models can be
used to predict the performances of algorithms in AP or instances in IP ,
without actually evaluating them, which is vary time-consuming.
(2) Online mode of EBO. Situations in which a solver faces a series of different problem instances coming sequentially pose new challenges. For
example, the objective in online mode is to maximize the cumulative performances on all the instances. Thus methods designed for this scenario
must consider making solvers perform well on current instances and improving solvers for future instances simultaneously. Besides, in a dynamic
environment, the underlying properties of instances may change overtime;
therefore the solvers may keep detecting the changes of environment and
adapt to new instances.
36
(3) Deep understanding of the fundamental issues of EBO is also worthy of
exploration. For example, LiangYi actually maintains two adversary sets
competing with one another, which is a typical scenario where game theory
can be applied. Besides, other more general issues in EBO include the
similarity measure between instances, a unified approach to information
extraction from solved instances, and theoretical proofs of the usefulness
of transmitting information between similar instances.
6. Acknowledgements
References
37
| 9cs.NE
|
Asymptotically Stable Drift and Minorization for Markov Chains
arXiv:1712.08867v1 [math.ST] 24 Dec 2017
with Application to Albert and Chib’s Algorithm
Qian Qin and James P. Hobert
Department of Statistics
University of Florida
December 2017
Abstract
The use of MCMC algorithms in high dimensional Bayesian problems has become routine. This has
spurred so-called convergence complexity analysis, the goal of which is to ascertain how the convergence
rate of a Monte Carlo Markov chain scales with sample size, n, and/or number of covariates, p. Recent
work suggests that, while often useful for establishing convergence rates when n and p are fixed, techniques based on drift and minorization may not be versatile enough to handle the more delicate task of
convergence complexity analysis. This article provides some general ideas for sharpening drift and minorization conditions, particularly in cases where n and/or p are large. The key ideas include developing
an appropriately “centered” drift function, and suppressing high-dimensionality in the construction of
minorization conditions. These concepts are employed in a thorough convergence complexity analysis of
Albert and Chib’s (1993) data augmentation algorithm for the Bayesian probit model. The main result is
that the geometric convergence rate of the underlying Markov chain is bounded below 1 both as n → ∞
(with p fixed), and as p → ∞ (with n fixed). Furthermore, the first computable bounds on the total
variation distance to stationarity are byproducts of the asymptotic analysis.
1 Introduction
Markov chain Monte Carlo (MCMC) has become an indispensable tool in Bayesian analysis, and it is now
well known that the ability to utilize an MCMC algorithm in a principled manner (with regard to choosing
Key words and phrases. Bayesian probit regression, Convergence complexity, Geometric ergodicity, High dimensional inference, Large p - small n, Markov chain Monte Carlo
1
the Monte Carlo sample size, for example) requires an understanding of the convergence properties of the underlying Markov chain (see, e.g., Flegal et al., 2008). Taking this a step further, in modern high dimensional
problems, it is also important to understand how the convergence properties of the chain change as the sample
size, n, and/or number of covariates, p, increases. Denote the data (i.e., responses and covariates) by D. If
we imagine n or p (or both) increasing, this leads to consideration of a sequence of data sets, {Dj }, and corresponding sequences of posterior distributions and Monte Carlo Markov chains. A natural question to ask
is “What can we say about the convergence properties of the Markov chains as j → ∞?” There is currently
a great deal of interest in questions like this in the MCMC community (see, e.g. Rajaratnam and Sparks,
2015; Durmus and Moulines, 2016; Johndrow et al., 2016; Yang et al., 2016; Yang and Rosenthal, 2017).
Rajaratnam and Sparks (2015) call this the study of convergence complexity, and we will follow their lead.
Asymptotic analysis of the convergence properties of a sequence of Markov chains associated with increasingly large finite state spaces has a long history in the computer science literature, dating back at least
to Sinclair and Jerrum (1989). While the techniques developed in computer science have been successfully
applied to a few problems in statistics (see, e.g. Yang et al., 2016), they are generally not applicable in situations where the state space is high-dimensional and uncountable, which is the norm for Monte Carlo Markov
chains in Bayesian statistics. In this paper, we show how methods based on drift and minorization can be
used to study the convergence complexity of such Markov chains.
Let π : X → [0, ∞) denote an intractable probability density function (pdf), where X ⊂ Rd , and let
R
Π(·) denote the corresponding probability measure, i.e., for measurable A, Π(A) = A π(x) dx. (More
general spaces will be considered later.) Let K(x, ·), x ∈ X, denote the Markov transition function (Mtf)
of an irreducible, aperiodic, Harris recurrent Markov chain with invariant probability measure Π. (See
Meyn and Tweedie (2012) for definitions.) The chain is called geometrically ergodic if there exist M : X →
[0, ∞) and ρ ∈ [0, 1) such that
kK m (x, ·) − Π(·)kTV ≤ M (x)ρm for all x ∈ X and all m ∈ N ,
(1)
where k·kTV denotes total variation norm, and K m (x, ·) is the m-step Mtf. The important practical benefits
of basing one’s MCMC algorithm on a geometrically ergodic Markov chain have been well-documented
by, e.g., Roberts and Rosenthal (1998), Jones and Hobert (2001), Flegal et al. (2008) and Łatuszyński et al.
(2013). Define the geometric convergence rate of the chain as
ρ∗ = inf ρ ∈ [0, 1] : (1) holds for some M : X → [0, ∞) .
Clearly, the chain is geometrically ergodic if and only if ρ∗ < 1.
2
Establishing the convergence rate of a practically relevant Monte Carlo Markov chain can be quite challenging. A key tool for this purpose has been the technique developed by Rosenthal (1995), which allows
for the construction of an upper bound on ρ∗ using drift and minorization conditions. This method, which is
described in detail in Section 2, has been used to establish the geometric ergodicity of myriad Monte Carlo
Markov chains (see, e.g., Fort et al., 2003; Marchev and Hobert, 2004; Roy and Hobert, 2010; Vats, 2017).
Since the drift and minorization (hereafter, d&m) method is still the most (and arguably the only) reliable tool
for bounding ρ∗ for practically relevant Monte Carlo Markov chains on uncountable state spaces, it is important to know whether it remains useful in the context of convergence complexity analysis. Unfortunately, it
turns out that most of the upper bounds on ρ∗ that have been produced using techniques based on d&m converge to 1 (often exponentially fast), becoming trivial, as n and/or p grow (see, e.g., Rajaratnam and Sparks,
2015). One example of this is Roy and Hobert’s (2007) analysis of Albert and Chib’s (1993) data augmentation algorithm for the Bayesian probit model, which establishes geometric ergodicity of the underlying
Markov chain, but also leads to an upper bound on ρ∗ that converges to 1 as n → ∞.
There are, of course, many possible explanations for why the d&m-based upper bounds on ρ∗ converge
to 1. It could be that the associated Monte Carlo Markov chains actually have poor asymptotic properties,
or, if not, perhaps Rosenthal’s (1995) method is simply not up to the more delicate task of convergence
complexity analysis. We show that, in the case of Albert and Chib’s (1993) chain, neither of these potential
explanations is correct. Indeed, our careful d&m analysis of Albert and Chib’s (1993) chain (hereafter,
A&C’s chain) leads to upper bounds on ρ∗ that are bounded away from 1 in both the large n, small p case,
and the large p, small n case. We believe that this is the first successful convergence complexity analysis of
a practically relevant Monte Carlo Markov chain using Rosenthal’s (1995) method.
Recently, Yang and Rosenthal (2017) used a modified version of Rosenthal’s (1995) technique to successfully analyze the convergence complexity of a Gibbs sampler for a simple Bayesian linear mixed model.
We note that, because one of the variance components in their model is assumed known, it is actually straightforward to sample directly from the posterior distribution using a univariate rejection sampler (Jones, 2001,
Section 3.9). Thus, while Yang and Rosenthal’s (2017) results are impressive, and their methods may suggest
a way forward, the Monte Carlo Markov chain that they analyzed is not practically relevant.
One of the main contributions of this article is a general approach for sharpening the upper bounds
that are produced via d&m, particularly in cases where n and/or p is large. The key ideas include “centering” the drift function to a region in the state space that the chain visits frequently, and suppressing
high-dimensionality in the construction of minorization conditions. In addition, for two-block Gibbs chains,
we introduce a technique for constructing asymptotically stable minorization conditions that is based on
3
the well known fact that the two marginal Markov chains and the joint chain all share the same geometric
convergence rate.
Before describing our results for A&C’s chain, we introduce an alternative definition of geometric ergodicity. Let L2 (Π) denote the set of signed measures µ that are absolutely continuous with respect to Π,
R
and satisfy X (dµ/dΠ)2 dΠ < ∞. As in Roberts and Rosenthal (1997), we say that the Markov chain with
Mtf K is L2 -geometrically ergodic if there exists ρ < 1 such that for each probability measure ν ∈ L2 (Π),
there exists a constant Mν < ∞ such that
kνK m (·) − Π(·)kTV ≤ Mν ρm for all m ∈ N ,
where νK m (·) =
R
XK
m (x, ·) ν(dx).
We define the L2 -geometric convergence rate, ρ∗∗ , to be the infimum
of all ρ ∈ [0, 1] that satisfy this definition. Not surprisingly, ρ∗ and ρ∗∗ are closely related (see, e.g.,
Roberts and Tweedie, 2001).
We now provide an overview of our results for Albert and Chib’s (1993) Markov chain, starting with a
brief description of their algorithm. Let {Xi }ni=1 be p-dimensional covariates, and let {Yi }ni=1 be a corresponding sequence of binary random variables such that Yi |Xi , B ∼ Bernoulli(Φ(XiT B)) independently,
where B is a p × 1 vector of unknown regression coefficients, and Φ is the standard normal distribution
function. Consider a Bayesian analysis using a prior density for B given by
ω(β) ∝ exp
n
o
1
− (β − v)T Q(β − v) ,
2
where v ∈ Rp , and Q ∈ Rp×p is either a positive definite matrix (proper Gaussian prior), or a zero matrix
(flat improper prior). Assume for now that the posterior is proper. (Propriety under a flat prior is discussed
in Section 3.) As usual, let X denote the n × p matrix whose ith row is XiT , and let Y = (Y1 Y2 · · · Yn )T
denote the vector of responses. The intractable posterior density is given by
)
( n
Y
Y
1−Y
i
i
ω(β) .
Φ(XiT β)
1 − Φ(XiT β)
πB|Y,X (β|Y, X) ∝
(2)
i=1
The standard method for exploring (2) is the classical data augmentation algorithm of Albert and Chib
(1993), which simulates a Harris ergodic (irreducible, aperiodic, and Harris recurrent) Markov chain, {Bm }∞
m=0 ,
that has invariant density πB|Y,X . In order to state the algorithm, we require a bit of notation. For θ ∈ R,
σ > 0, and i ∈ {0, 1}, let TN(θ, σ 2 ; i) denote the N(θ, σ 2 ) distribution truncated to (0, ∞) if i = 1, and
to (−∞, 0) if i = 0. The matrix Σ := X T X + Q is necessarily non-singular because of propriety. If the
current state of A&C’s Markov chain is Bm = β, then the new state, Bm+1 , is drawn using the following
two steps:
4
Iteration m + 1 of the data augmentation algorithm:
1. Draw {Zi }ni=1 independently with Zi ∼ TN(XiT β, 1; Yi ), and let Z = (Z1 Z2 · · · Zn )T .
2. Draw
Bm+1 ∼ Np Σ−1 X T Z + Qv , Σ−1 .
The convergence rate of A&C’s chain has been studied by several authors. Roy and Hobert (2007)
proved that when Q = 0, the chain is always geometrically ergodic. (Again, we’re assuming posterior
propriety.) A similar result for proper normal priors was established by Chakraborty and Khare (2017). Both
results were established using a technique that does not require construction of a minorization condition (see
Meyn and Tweedie, 2012, Lemma 15.2.8), and, consequently, does not yield an explicit upper bound on ρ∗ .
Thus, neither paper addresses the issue of convergence complexity. However, in Section 5 we prove that
Roy and Hobert’s (2007) drift function cannot be used to construct an upper bound on ρ∗ that is bounded
away from 1 as n → ∞. Johndrow et al. (2016) recently established a convergence complexity result for the
intercept only version of the model (p = 1 and Xi = 1 for i = 1, 2, . . . , n) with a proper (univariate) normal
prior, under the assumption that all the responses are successes (Yi = 1 for i = 1, 2, . . . , n). Their results,
which are based on Cheeger’s inequality, imply that ρ∗∗ → 1 as n → ∞, indicating that the algorithm is
inefficient for large samples.
The results established herein provide a much more complete picture of the convergence behavior of
A&C’s chain. Three different regimes are considered: (i) fixed n and p, (ii) large n, small p, and (iii) large
p, small n. Our analysis is based on two different drift functions that are both appropriately centered (at
the posterior mode). One of the two drift functions is designed for case (ii), and the other for case (iii).
We establish d&m conditions for both drift functions, and these are used in conjunction with Rosenthal’s
(1995) result to construct two explicit upper bounds on ρ∗ . They are also used to construct two computable
upper bounds on the total variation distance to stationarity (as in (1)), which improves upon the analyses of
Roy and Hobert (2007) and Chakraborty and Khare (2017).
The goal in case (ii) is to study the asymptotic behavior of the geometric convergence rate as n → ∞,
when p is fixed. To this end, we consider a sequence of data sets, Dn := {(Xi , Yi )}ni=1 . So, each time n
increases by 1, we are are given a new p × 1 covariate vector, Xi , and a corresponding binary response,
Yi . To facilitate the asymptotic study, we assume that the (Xi , Yi ) pairs are generated according to a random
mechanism that is governed by very weak assumptions (that are consistent with the probit regression model).
5
We show that there exists a constant ρ < 1 such that, almost surely, lim supn→∞ ρ∗ (Dn ) ≤ ρ. Apart from
this general result, we are also able to show that, in the intercept only model considered by Johndrow et al.
(2016), the A&C chain is actually quite well behaved as long as the proportion of successes is bounded
away from 0 and 1. To be specific, let {Yi }∞
i=1 denote a fixed sequence of binary responses, and let p̂n =
P
n−1 ni=1 Yi . Our results imply that, as long as 0 < lim inf n→∞ p̂n ≤ lim supn→∞ p̂n < 1, there exists
ρ < 1 such that both ρ∗ (Dn ) and ρ∗∗ (Dn ) are eventually bounded above by ρ, and there is a closed form
expression for ρ.
In case (iii), n is fixed and p → ∞. There are several important differences between cases (ii) and (iii).
First, in case (ii), since p is fixed, only a single prior distribution need be considered. In contrast, when
p → ∞, we must specify a sequence of priors, {(Qp , vp )}∞
p=1 , where vp is a p × 1 vector, and Qp is a p × p
positive definite matrix. (When p > n, a positive definite Q is required for posterior propriety.) Also, in
case (iii), there is a fixed vector of responses (of length n), and it is somewhat unnatural to consider the new
columns of X to be random. Let Dp := {(vp , Qp , Xp , Y )} denote a fixed sequence of priors and data sets,
where Y is a fixed n × 1 vector of responses, and Xp is an n × p matrix. We show that, under a natural
T
regularity condition on Xp Q−1
p Xp , there exists a ρ < 1 such that ρ∗ (Dp ) ≤ ρ for all p.
The remainder of the paper is laid out as follows. In Section 2, we formally introduce the concept of
stable d&m conditions, and describe our general ideas for constructing such. The centered drift functions
that are used in our analysis of A&C’s chain are described in Section 3. In Section 4, we provide results
for A&C’s chain in the case where n and p are both fixed. Two sets of d&m conditions are established,
and corresponding exact total variation bounds on the distance to stationarity are provided. The heart of the
paper is Section 5 where it is shown that the geometric convergence rate of A&C’s chain is bounded away
from 1 as n → ∞ for fixed p, and as p → ∞ for fixed n. A good deal of technical material is relegated to an
Appendix.
2 Asymptotically Stable Drift and Minorization
Let X be a set equipped with a countably generated σ-algebra B(X). Suppose that K : X × B(X) → [0, 1]
R
is an Mtf with invariant probability measure Π(·), so that Π(A) = X P (x, A)Π(dx) for all measurable A.
Assume that the corresponding Markov chain is Harris ergodic. Recall the definitions of geometric ergodicty
and geometric convergence rate from Section 1. The following result has proven extremely useful for establishing geometric ergodicity in the context of Monte Carlo Markov chains used to study complex Bayesian
posterior distributions.
6
Theorem 1 (Rosenthal (1995)). Suppose that K(x, ·) satisfies the drift condition
Z
V (x′ )K(x, dx′ ) ≤ λV (x) + L,
x∈X
(3)
X
for some V : X → [0, ∞), λ < 1 and L < ∞. Suppose that it also satisfies the minorization condition
K(x, ·) ≥ εQ(·)
whenever V (x) < d
(4)
for some ε > 0, probability measure Q(·) on X, and d > 2L/(1 − λ). Then assuming the chain is started
according to the probability measure ν(·), for any 0 < r < 1, we have
kνK m (·) − Π(·)kTV ≤ (1 − ε)rm
#m
"
Z
1 + 2L + λd 1−r
L
r
V (x)ν(dx)
+
.
{1 + 2(λd + L)}
+ 1+
1−λ
1+d
X
The function V is called the drift (or Lyapunov) function, and {x ∈ X : V (x) < d} is called the small
set associated with V . We will refer to ε as the minorization number. Manipulation of the total variation
bound in Theorem 1 leads to
m
kνK (·) − Π(·)kTV
≤ 2+
L
+
1−λ
where
r
ρ̂ := (1 − ε) ∨
1 + 2L + λd
1+d
1−r
Z
X
V (x)ν(dx) ρ̂m ,
1 + 2(λd + L)
r
,
(5)
and r ∈ (0, 1) is arbitrary. Then, ρ̂ is an upper bound on the geometric convergence rate ρ∗ . It’s easy to
verify that when λ < 1, L < ∞, and ε > 0, there exists r ∈ (0, 1) such that ρ̂ < 1.
The bound ρ̂ has a reputation for being too conservative. This is partly due to the fact that there are toy
examples where the true ρ∗ is known, and ρ̂ is quite far off (see, e.g., Rosenthal, 1995). There also exist
myriad analyses of practical Monte Carlo Markov chains where the d&m conditions (3) and (4) have been
established (proving that the underlying chain is indeed geometrically ergodic), but the total variation bound
of Theorem 1 is useless because ρ̂ is so near unity. Of course, the quality of the bound ρ̂ depends on the
choice of drift function, and the sharpness of (3) and (4). Our results for the A&C chain suggest that poorly
chosen drift functions and/or loose inequalities in the d&m conditions are to blame for (at least) some of the
unsuccessful applications of Theorem 1. We now introduce the concept of asymptotically stable d&m.
Consider a sequence of geometrically ergodic Markov chains, {Ψ(j) }∞
j=1 , with corresponding geometric
(j)
convergence rates given by ρ∗ . (In practice, j is usually the sample size, n, or number of covariates, p.)
We are interested in the asymptotic behavior of the rate sequence. For example, we might want to know if it
7
is bounded away from 1. Suppose that for each chain, Ψ(j) , we have d&m conditions defined through λ(j) ,
(j)
L(j) , and ε(j) , and thus an upper bound on the convergence rate, ρ̂(j) ∈ [ρ∗ , 1). The following simple result
provides conditions under which these upper bounds are unstable, that is, lim supj→∞ ρ̂(j) = 1.
(jl ) }∞ , that satisfies one
Proposition 2. Suppose that there exists a subsequence of {Ψ(j) }∞
j=0 , call it {Ψ
l=0
or more of the following three conditions: (i) λ(jl ) → 1, (ii) L(jl ) → ∞, while ε(jl ) is bounded away from 1
and λ(jl ) is bounded away from 0, (iii) ε(jl ) → 0. Then the corresponding subsequence of the upper bounds,
ρ̂(jl ) , converges to 1.
Proof. Throughout the proof, we drop the superscripts on λ, L, and ε. If (i) holds, then
1 + 2L + λd
≥λ→1.
1+d
Since 1 + 2(λd + L) ≥ 1, regardless of the asymptotic behavior of r ∈ (0, 1), we have
1 + 2L + λd 1−r
lim inf
{1 + 2(λd + L)}r ≥ 1.
j→∞
1+d
Hence, ρ̂(jl ) → 1. Now assume that (ii) holds. Then 1 + 2(λd + L) → ∞. Meanwhile, for large enough l,
1 + 2L + λd
≥ λ > c,
1+d
where c is a positive constant. Hence, to ensure that
1 + 2L + λd 1−r
{1 + 2(λd + L)}r
1+d
is bounded, one must let r → 0. However, this would result in (1 − ε)r → 1. Finally, assume that (iii) holds.
Then (1 − ε)r → 1 regardless of the asymptotic behavior of r ∈ (0, 1).
If one (or more) of the conditions in Proposition 2 holds for some subsequence of {Ψ(j) }∞
j=0 , then we
say that the d&m conditions are (asymptotically) unstable (in j). On the other hand, if (i′ ) λ(j) is bounded
away from 1, (ii′ ) L(j) is bounded above, and (iii′ ) ε(j) is bounded away from 0, then the sequence ρ̂(j) can
(j)
be bounded away from 1, thus giving an asymptotically nontrivial upper bound on ρ∗ . We say that the drift
conditions are stable if (i′ ) and (ii′ ) hold, and likewise, that the minorization conditions are stable if (iii′ )
holds.
Before proceeding to discuss techniques for achieving stability, we note that there exist some elementary
transformations on V that could affect the quality of ρ̂, and even the stability of d&m. For instance, if (3) is
satisfied, then
Z
L−1 V (x′ )K(x, dx′ ) ≤ λL−1 V (x) + 1 .
X
8
Thus, if L → ∞, we can simply re-scale each drift function by its corresponding L−1 , which results in
L ≡ 1. However, such re-scaling cannot be used to stabilize minorization conditions. In the above example,
suppose that for some d′ > 2/(1 − λ), K(x, ·) ≥ εQ(·) whenever L−1 V (x) > d′ . Then K(x, ·) ≥ εQ(·)
whenever V (x) > Ld′ , where Ld′ > 2L/(1 − λ). Thus, if one cannot establish a stable minorization condition associated with V (x), then there is no hope of finding a stable minorization associated with L−1 V (x).
Stability could also be affected by adding a scalar to the drift function, i.e., replacing V (x) with some
V (x) + c. One can easily show that subtracting a positive number from V (while preserving the positiveness
of V ) decreases the volume of the small set, and thus increases the minorization number, ε. In fact, it always
improves the quality of ρ̂. Needless to say, we will only deal with instability that cannot be prevented by
these trivial transformations. In particular, throughout the paper, we consider only drift functions whose
infimums are 0, that is, we make the elementary transformation of V (x) 7→ V (x) − inf x′ ∈X V (x′ ).
Choosing an appropriate drift function is extremely important for establishing stable d&m conditions.
The total variation bound in Theorem 1 is based on a theoretical coupling argument that involves two (dependent) copies of the Markov chain in question, one started in stationarity, and the other started from an
arbitrary initial distribution. The total variation distance between the two chains after m iterations is bounded
above by P (T > m), where T is the coupling time. In the coupling construction, there is a chance to couple
each time the chains enter the small set, {x ∈ X : V (x) ≤ d}. Loosely speaking, we want the chains to
visit the small set as often as possible since each visit provides a chance to couple, and fast coupling means
small values for P (T > m), which, in turn, leads to small total variation distance. This argument suggests
that one should choose a drift function that is “centered” with respect to the Markov chain in the sense that
the drift function should take small values in the part of the state space where the chain is likely to spend the
bulk of its time. Of course, if the chain is well suited to the problem, it should spend the bulk of its time in
the high posterior probability regions of X. A properly centered drift function is likely to give a small L, and
in turn a small d, which is often necessary for a large ε.
This intuition is not new, and many authors do attempt to center their drift functions in some way. (Although, in many cases, the centering is not delicate enough to ensure that the drift function reflects the dynamics of the chain.) However, we believe that our work is one of the first to illustrate the importance of centering in a quantitative fashion through a genuinely non-trivial example. Indeed, in Section 5 we show that,
for A&C’s chain, in the large n, small p case, the un-centered drift function employed by Roy and Hobert
(2007) cannot possibly lead to stable d&m, while a centered version of the same drift function does.
Suppose now that V is properly centered and stable drift has been established. Obtaining stable minorization can still be difficult due to the intractability of K(x, ·). We now describe a common difficulty
9
encountered in the convergence analysis of two-block Gibbs samplers and data augmentation algorithms.
For simplicity, we’ll investigate Mtfs that admit density functions with respect to the Lebesgue measure, but
most of the material in the remainder of this section can be easily generalized. Suppose that X ⊂ Rp , and
K(x, ·) is associated with a Markov transition density (Mtd) of the form
k(x, x′ ) =
Z
s(x′ |z)h(z|x) dz ,
(6)
Rn
where n is the sample size, z = (z1 , z2 , . . . , zn )T ∈ Rn is a vector of latent data, and s : X × Rn → [0, ∞)
and h : Rn × X → [0, ∞) are conditional densities associated with some joint density on X × Rn . In practice,
it is often the case that h(z|x) can be factored as follows:
h(z|x) =
n
Y
hi (zi |x) ,
i=1
where, for each i, hi : R × X → [0, ∞) is a univariate (conditional) pdf. Usually, s(·|z) and hi (·|x) are
tractable, but there is no closed form for k(x, x′ ). However, there is a well-known argument for establishing
a minorization condition in this case. Suppose that whenever V (x) < d,
hi (t|x) > ǫi νi (t) ,
i = 1, 2, . . . , n
where ǫi > 0 and νi : R → [0, ∞) is a pdf. Then, whenever V (x) < d, we have
!Z
n
n
Y
Y
′
′
νi (zi ) dz .
s(x |z)
ǫi
k(x, x ) >
Rn
i=1
Since
R
Rn
s(x′ |z)
Qn
i=1 νi (zi ) dz
(7)
i=1
is a pdf on X, (7) gives a minorization condition with ε =
Qn
i=1 ǫi .
Un-
fortunately, this quantity will almost always converge to 0 as n → ∞. Consequently, if our sequence of
Markov chains is indexed by n, then we have unstable minorization. This problem is well known (see, e.g.,
Rajaratnam and Sparks, 2015).
The instability of the minorization described above is due to the fact that the dimension of z is growing
with n. However, it is often the case that s(x|z) depends on z only through f (z), where f : Rn → Y is
a function into some fixed space Y, say Y ⊂ Rq , where q does not depend on n. Then integrating along
f (z) = γ in (6) yields
′
k(x, x ) =
where
s̃(x′ |f (z))
=
s(x′ |z),
Z
Y
and
Z
C
h̃(γ|x) dγ =
s̃ x′ |γ h̃ γ|x dγ ,
Z
h(z|x) dz
{z:f (z)∈C}
10
for all x ∈ X and any measurable C ⊂ Y. Note that this new representation of k(x, x′ ) no longer contains n
explicitly, and the high dimensionality problem for z is resolved. However, we now have a new problem.
Namely, h̃(γ|x) is likely to be quite intractable. Fortunately, there is also a way to circumvent this difficulty.
Consider a Markov chain driven by the Mtd
k̃(γ, γ ′ ) =
Z
h̃(γ ′ |x)s̃(x|γ) dx ,
γ, γ ′ ∈ Y .
X
We call this the “flipped” version of the chain defined by k(x, x′ ). It is well known that these two chains
share the same geometric convergence rate (see, e.g., Roberts and Rosenthal, 2001; Diaconis et al., 2008).
Thus, we can analyze the original sequence of Markov chains indirectly through the corresponding flipped
chains. Recall that the drift condition for the original chain is
Z
Define Ṽ (γ) =
R
XV
V (x′ )
X
Z
s̃(x′ |γ ′ )h̃(γ ′ |x) dγ ′ dx′ ≤ λV (x) + L .
(8)
Y
(x)s̃(x|γ) dx. Multiplying both sides of (8) by s̃(x|γ), and integrating out x yields the
following drift condition for the flipped chain:
Z
Ṽ (γ ′ )k̃(γ, γ ′ ) dγ ′ ≤ λṼ (γ) + L .
Y
Moreover, since λ and L are the same as before, this drift is stable.
Now, as mentioned above, s̃(x|f (z)) = s(x|z) is often tractable. Suppose that there exists some ε̃ > 0
and pdf ν̃ : X → [0, ∞) such that s̃(x|γ) ≥ ε̃ν̃(x) when Ṽ (γ) < d, then we have the following minorization
condition for the flipped chain:
′
k̃(γ, γ ) ≥ ε̃
Z
h̃(γ ′ |x)ν̃(x) dx
whenever Ṽ (γ) < d .
X
This minorization is stable as long as ε̃ is bounded away from 0 asymptotically. The flipped chain argument
is not always applicable, but when it is, it simplifies calculations greatly. Furthermore, we will demonstrate
that the flipped chain argument can also be useful for establishing minorization conditions that are stable
in p. We end this section with a result that allows us to use information about a flipped chain to get total
variation bounds for the original.
Obviously, if we have d&m conditions for a flipped chain, then this provides us with an upper bound on
the geometric convergence rate of the corresponding original chain. However, using the information about
the flipped chain to get total variation bounds for the corresponding original chain requires a bit of work. The
following result follows immediately from Proposition 24, which is stated and proven in Subsection B.1.
11
Corollary 3. Suppose we have d&m conditions for a flipped chain (which is driven by k̃). Let λ, L, and ε
denote the drift and minorization parameters, and let ρ̂f denote the corresponding bound on the geometric
convergence rate. Then, if the original chain is started at x ∈ X, we have, for m ≥ 1,
Z
L
m
.
Ṽ (γ)h̃(γ|x) dγ ρ̂m−1
+
kK (x, ·) − Π(·)kTV ≤ 2 +
f
1−λ
Y
In the next section, we begin our analysis of the A&C chain.
3 Albert and Chib’s Markov Chain and the Centered Drift Functions
3.1 Basics
As in the introduction, let {Xi }ni=1 be p-dimensional covariates and let {Yi }ni=1 be a sequence of binary
random variables such that Yi |Xi , B ∼ Bernoulli(Φ(XiT B)) independently, where B is a p × 1 vector of
unknown regression coefficients. Suppose that, having observed the data, D := {(Xi , Yi )}ni=1 , we wish to
perform a Bayesian analysis using a prior density for B given by
ω(β) ∝ exp
n
o
1
− (β − v)T Q(β − v) ,
2
where v ∈ Rp , and Q is a p × p inverse covariance matrix that is either positive definite, or vanishing. Let X
denote the n × p matrix whose ith row is XiT , and let Y = (Y1 Y2 · · · Yn )T denote the vector of responses.
The posterior distribution (2) is proper precisely when
c(X, Y ) :=
Z
n
Y
Rp i=1
Yi
1−Yi
ω(β) dβ < ∞ .
Φ(XiT β)
1 − Φ(XiT β)
When Q is positive definite, ω(β) is a proper normal density, and the posterior is automatically proper. If
Q = 0, then propriety is not guaranteed. Define X∗ as the n × p matrix whose ith row is −XiT if Yi = 1,
and XiT if Yi = 0. Chen and Shao (2000) proved that when the prior is flat, i.e., Q = 0, the following two
conditions are necessary and sufficient for posterior propriety:
(C1) X has full column rank;
(C2) There exists a vector a = (a1 a2 · · · an )T ∈ Rn such that ai > 0 for all i, and X∗T a = 0.
Until further notice, we will assume that the posterior is proper.
12
A&C’s algorithm to draw from the intractable posterior is based on the following latent data model.
Given X and B, let {(Yi , Zi )}ni=1 be a sequence of independent random vectors such that
Yi |Zi , X, B is a point mass at 1R+ (Zi )
Zi |X, B ∼ N (XiT B, 1) .
Clearly, under this hierarchical structure, Yi |Xi , B ∼ Bernoulli(Φ(XiT B)), which is consistent with the
original model. Thus, if we let πB,Z|Y,X (β, z|Y, X) denote the corresponding (augmented) posterior density
(where Z = (Z1 Z2 · · · Zn )T ), then it’s clear that
Z
πB,Z|Y,X (β, z|Y, X) dz = πB|Y,X (β|Y, X) ,
Rn
which is the target posterior from (2). Albert and Chib’s algorithm is simply a two-variable Gibbs sampler
based on πB,Z|Y,X (β, Z|Y, X). Indeed, the Mtd, k : Rp × Rp → R+ , is defined as
Z
′
′
πB|Z,Y,X (β ′ |z, Y, X)πZ|B,Y,X (z|β, Y, X) dz .
kAC (β, β ) := kAC (β, β ; Y, X) =
Rn
One can show that
B|Z, Y, X ∼ Np Σ−1 X T Z + Qv , Σ−1 ,
where, again, Σ = X T X + Q. Moreover, the density πZ|B,Y,X (z|β, Y, X) is a product of n univariate
densities, where
Zi |B, Y, X ∼ TN(XiT B, 1; Yi ).
Obviously, these are the conditional densities that appear in the algorithm described in the Introduction.
3.2 A centered drift function
Roy and Hobert (2007) and Chakraborty and Khare (2017) both used the drift function V0 (β) = kΣ1/2 βk2 .
While this drift function is certainly amenable to analysis, it is not “centered” in any practical sense. Indeed,
V0 (β) takes on it’s minimum when β = 0, but, in general, there is no reason to expect A&C’s Markov
chain to make frequent visits to the vicinity of the origin. This heuristic is borne out by a result in Section 5
showing that V0 cannot lead to stable d&m in the large n, small p case. As an alternative to V0 (β), we
consider drift functions of the form
V (β) = kM (β − β ∗ )k2 ,
(9)
where M = M (X, Y ) is a matrix with p columns, and β ∗ = β ∗ (X, Y ) is a point in Rp that is “attractive” to
A&C’s Markov chain. A candidate for β ∗ would be a fixed point for the chain, that is, a value BF satisfying
Z
β kAC (BF , β) dβ = BF .
(10)
Rp
13
Indeed, this equation suggests that the chain tends to “stay around” the point BF , and hence BF might be an
effective point at which to center the drift function. The following result shows that there is, in fact, a unique
solution to (10).
Proposition 4. The posterior density, πB|Y,X (β|Y, X), possesses a unique mode, and this mode is the unique
solution to (10).
Proof. Note that log πB|Y,X (β|Y, X) = log πY |B,X (Y |β, X) + log ω(β) + constant. It is well-known that
log ω(β) is concave, and that log πY |B,X (Y |β, X), as a function of β, is strictly concave. Therefore, since
the posterior density is proper, it must possess a unique mode in Rp . Now, the mode is the solution to the
following equation:
d
log ω(β) + log πY |B,X (Y |β, X) = 0 ,
dβ
where
(11)
1
log ω(β) = − (β − v)T Q(β − v) + constant ,
2
and
log πY |B,X (Y |β, X) =
n
X
i=1
log Φ(XiT β)1{1} (Yi ) + 1 − Φ(XiT β) 1{0} (Yi ) + constant .
Thus, (11) is equivalent to the following equation
n
X
φ(XiT β)
φ(XiT β)
1{1} (Yi ) −
1{0} (Yi ) Xi − (Qβ − Qv) = 0 .
Φ(XiT β)
1 − Φ(XiT β)
i=1
(12)
On the other hand, a fixed point solves the following equation in β:
Z
Rp
Z
Rn
β ′ πB|Z,Y,X (β ′ |z, Y, X)πZ|B,Y,X (z|β, Y, X) dz dβ ′ = β .
(13)
Recall that B|Z, Y, X ∼ Np (Σ−1 (Qv + X T Z), Σ−1 ). Thus, (13) is equivalent to
Qv +
n
X
Xi E(Zi |B = β, Y, X) = Σβ .
(14)
i=1
By (27) in Subsection A.2, we have
E(Zi |B = β, Y, X) = XiT β +
φ(XiT β)
φ(XiT β)
1
(Y
)
−
1{0} (Yi ) ,
i
{1}
Φ(XiT β)
1 − Φ(XiT β)
where φ(·) is the pdf of the standard normal distribution. Routine calculation shows that (14) is equivalent
to (12).
14
Remark 5. The posterior density, πB|Y,X (β|Y, X), is invariant for A&C’s Markov chain, so it shouldn’t be
too surprising that the chain tends to “hang around” the posterior mode. On the other hand, the fact that
the posterior mode is an exact (unique) fixed point for the chain is a bit unexpected.
We will consider two different versions of (9), both centered at BF . The first, which will be used in the
large n-small p case, is simply a centered version of V0 given by
V1 (β) = Σ1/2 (β − BF )
2
.
In the large p-small n case, we assume that Q is positive definite (which is necessary for posterior propriety)
and that X is full row rank, and we use the following drift function
V2 (β) =
XΣ−1 X T
−1/2
X(β − BF )
2
.
In the next section, we establish two sets of d&m conditions for the A&C chain based on V1 and V2 .
4 Results for the Albert and Chib Chain Part I: Fixed n and p
4.1 Drift inequalities for V1 and V2
Define g : R → R as
θφ(θ)
+
g(θ) =
Φ(θ)
φ(θ)
Φ(θ)
2
.
For any β ∈ Rp , let D(β) denote an n × n diagonal matrix with ith diagonal element
1 − g XiT β 1{1} (Yi ) − g − XiT β 1{0} (Yi ) .
Lemma 6. If V (β) = kM (β − BF )k2 , where M is any matrix with p columns, then
Z
o
n
2
V (β ′ )kAC (β, β ′ ) dβ ′ ≤ sup M Σ−1 X T D(BF + t(β − BF ))X(β − BF ) + 2 tr M Σ−1 M T .
Rp
t∈(0,1)
Proof. Note that
Z
V (β ′ ) πB|Z,Y,X (β ′ |z, Y, X) dβ ′ = M Σ−1 (X T z + Qv) − M BF
Rp
Moreover,
Z
Rn
2
+ tr M Σ−1 M T
.
kM Σ−1 (X T z + Qv) − M BF k2 πZ|B,Y,X (z|β, Y, X) dz
2
= M Σ−1 X T E (Z|B = β, Y, X) + Qv − M BF
+ tr M Σ−1 X T var (Z|B = β, Y, X) XΣ−1 M T
15
.
(15)
For two symmetric matrices of the same size, M1 and M2 , we write M1 ≤ M2 if M2 − M1 is non-negative
definite. By Lemma 23 in Subsection A.2, var(Z|B = β, Y, X) ≤ In . It follows that
M Σ−1 X T var (Z|B = β, Y, X) XΣ−1 M T ≤ M Σ−1 X T XΣ−1 M T ≤ M Σ−1 M T .
(16)
Therefore,
Z
′
Rp
′
′
V (β )kAC (β, β ) dβ =
Z
′
V (β )
Rp
Z
Rn
πB|Z,Y,X (β ′ |z, Y, X)πZ|B,Y,X (z|β, Y, X) dz dβ ′
≤ M Σ−1 X T E (Z|B = β, Y, X) + Qv − M BF
2
+ 2 tr M Σ−1 M T
.
(17)
Now, for α ∈ Rp , define
µ(α) = M Σ−1 X T E (Z|B = BF + α, Y, X) + Qv − M BF .
Since BF is a fixed point, we have
µ(0) = M
Z
Rp
βkAC (BF , β) dβ − M BF = 0 .
By the mean value theorem for vector-valued functions (see, e.g., Rudin, 1976, Theorem 5.19), for any
α ∈ Rp ,
kµ(α)k2 ≤
sup
t∈(0,1)
∂µ(tα)
∂t
!2
.
Now, by (27) in Subsection A.2,
n
X
∂
∂µ(tα)
Xi E(Zi |B = BF + tα, Y, X)
= M Σ−1
∂t
∂t
= M Σ−1
i=1
n
X
i=1
∂ T
Xi (BF + tα)
∂t
d
φ(θ)
φ(θ)
×
− 1{0} (Yi )
θ + 1{1} (Yi )
dθ
Φ(θ)
1 − Φ(θ)
Xi
θ=XiT (BF +tα)
= M Σ−1 X T D(BF + tα) Xα .
Hence,
kµ(α)k2 ≤ sup
M Σ−1 X T D(BF + tα)Xα
2
.
t∈(0,1)
The result then follows from (17) and (18) by taking α = β − BF .
We now use Lemma 6 to establish explicit drift inequalities for V1 and V2 . We begin with V1 .
16
(18)
Proposition 7. For all β ∈ Rp , we have
Z
kΣ−1/2 X T D(BF + tα)Xαk2
′
′
′
V1 (β) + 2p .
V1 (β )kAC (β, β ) dβ ≤
sup sup
kΣ1/2 αk2
t∈(0,1) α6=0
Rp
Proof. Taking M = Σ1/2 in Lemma 6 yields
Z
Rp
V1 (β ′ )kAC (β, β ′ ) dβ ′ ≤ sup
Σ−1/2 X T D(BF + t(β − BF ))X(β − BF )
2
+ 2p
t∈(0,1)
≤
kΣ−1/2 X T D(BF + tα)Xαk2
sup sup
V1 (β) + 2p .
kΣ1/2 αk2
t∈(0,1) α6=0
In Subsection B.2, we prove that
kΣ−1/2 X T D(BF + tα)Xαk2
<1.
kΣ1/2 αk2
t∈(0,1) α6=0
sup sup
For a symmetric matrix M , let λmin (M ) and λmax (M ) denote the smallest and largest eigenvalues of
M , respectively. Here is the analogue of Proposition 7 for V2 .
Proposition 8. Assume that X has full row rank. Then for all β ∈ Rp , we have
Z
Rp
V2 (β ′ )kAC (β, β ′ ) dβ ′ ≤ λ2max XΣ−1 X T V2 (β) + 2n .
Proof. Taking M = XΣ−1 X T
Z
Rp
−1/2
X in Lemma 6 and applying Lemma 23 yields
V2 (β ′ )kAC (β, β ′ ) dβ ′
≤ sup
XΣ−1 X T
t∈(0,1)
≤ sup λ2max
t∈(0,1)
1/2
D(β + t(β − BF ))X(β − BF )
XΣ−1 X T
1/2
2
+ 2n
D(BF + t(β − BF )) XΣ−1 X T
≤ λ2max XΣ−1 X T V2 (β) + 2n .
1/2
V2 (β) + 2n
4.2 Drift and minorization for the Albert and Chib chain based on V1
In this subsection, we exploit the flipped chain idea described in Section 2. In particular, V1 is used to
establish d&m conditions for a flipped chain that has the same geometric convergence rate as A&C’s chain.
Later, in Section 5, we will use these results to prove asymptotic stability as n → ∞.
17
Note that πB|Z,Y,X (β|Z, Y, X) depends on the n-dimensional vector Z only through X T Z, which is a
one-to-one function of the following p-dimensional vector:
Γ := Σ1/2 Σ−1 X T Z + Qv − BF
.
Hence, we can represent the Mtd of the A&C chain as follows:
Z
′
kAC (β, β ) =
Rp
πB|Γ,Y,X (β ′ |γ, Y, X)πΓ|B,Y,X (γ|β, Y, X) dγ .
Recalling the discussion in Section 2, this maneuver seems to represent progress since we have replaced n
with p. However, it is difficult to establish a minorization condition using this version of kAC (β, β ′ ) because
πΓ|B,Y,X (γ|β, Y, X) lacks a closed form. On the other hand, this chain has the same geometric convergence
rate as the flipped chain defined by the following Mtd:
′
′
k̃AC (γ, γ ) := k̃AC (γ, γ ; Y, X) =
Z
Rp
πΓ|B,Y,X (γ ′ |β, Y, X)πB|Γ,Y,X (β|γ, Y, X) dβ .
Constructing a minorization condition for this Mtd is much less daunting since
B|Γ, Y, X ∼ N Σ−1/2 Γ + BF , Σ−1 .
Here is the main result of this subsection.
Proposition 9. Let Ṽ1 (γ) = kγk2 . The Mtd k̃AC satisfies the drift condition
Z
Rp
Ṽ1 (γ ′ )k̃AC (γ, γ ′ ) dγ ′ ≤ λṼ1 (γ) + L ,
where L = p(1 + λ), and
kΣ−1/2 X T D(BF + tα)Xαk2
.
kΣ1/2 αk2
t∈(0,1) α6=0
λ = sup sup
Moreover, for d > 2L/(1 − λ), k̃AC satisfies
k̃AC (γ, γ ′ ) ≥ εq(γ ′ ) ,
where q : Rp → [0, ∞) is a pdf, and ε = 2−p/2 e−d .
Proof. We begin with the drift. It’s easy to verify that
Z
Rp
V1 (β)πB|Γ,Y,X (β|γ, Y, X) dβ = Ṽ1 (γ) + p .
18
We now use the techniques described at the end of Section 2 to convert the drift inequality in Proposition 7
into a drift inequality for the flipped chain. We know that
Z
Z
′ ′
′
′
′
πB|Γ,Y,X (β |γ , Y, X) πΓ|B,Y,X (γ |β, Y, X) dγ dβ ′ ≤ λV1 (β) + 2p .
V1 (β )
Rp
Rp
Multiplying both sides of the above inequality by πB|Γ,Y,X (β|γ, Y, X) and integrating with respect to β
yields
Z
Rp
Ṽ1 (γ ′ )k̃AC (γ, γ ′ ) dγ ′ ≤ λṼ1 (γ) + p(1 + λ) .
As explained in Section 2, to establish the minorization condition, it suffices to show that there exists a pdf
ν(β) := ν(β|Y, X) such that
πB|Γ,Y,X (β|γ, Y, X) ≥ εν(β)
(19)
whenever Ṽ1 (γ) ≤ d. Recall that
B|Γ, Y, X ∼ N Σ−1/2 Γ + BF , Σ−1 .
Define
ν1 (β) =
inf
γ:Ṽ1 (γ)≤d
πB|Γ,Y,X (β|γ, Y, X)
2
1
|Σ|1/2
1/2
−1/2
exp − Σ
β − BF − Σ
γ
.
= inf
p/2
2
γ:Ṽ1 (γ)≤d (2π)
Then ν(β) = ν1 (β)/
R
Rp
ν1 (β ′ ) dβ ′ is a pdf, and whenever Ṽ1 (γ) ≤ d,
Z
′
′
ν1 (β ) dβ ν(β) .
πB|Γ,Y,X (β|γ, Y, X) ≥
Rp
This is (19) with
ε=
Z
Rp
−p/2
ν1 (β) dβ = (2π)
1
inf exp − kβ − γk2
2
2
Rp γ:kγk ≤d
Z
dβ .
Finally, since kβ − γk2 ≤ 2(kβk2 + kγk2 ),
ε ≥ (2π)−p/2
Z
inf
2
Rp γ:kγk ≤d
exp − kβk2 − kγk2 dβ = 2−p/2 e−d .
Mainly, Proposition 9 will be used to establish asymptotic stability results for A&C’s chain in the large n,
small p case. Indeed, since the Markov chains defined by kAC and k̃AC have the same geometric convergence
rate, ρ̂ calculated using (5) with λ, L, and ε from Proposition 9 is an upper bound on ρ∗ = ρ∗ (X, Y ) for the
19
A&C chain. On the other hand, Proposition 9 can also be used in conjunction with Theorem 1 and Corollary 3
to get computable bounds on the total variation distance to stationarity for the A&C chain for fixed n and p.
(m)
In order to state the result, we require a bit of notation. For an integer m ≥ 1, let kAC : Rp × Rp → R+ be
the chain’s m-step Mtd. For β ∈ Rp , let ϕ(β) = E (Z|B = β, Y, X) ∈ Rn . Then results in Subsection A.2
show that the ith element of ϕ(β) is given by
XiT β +
φ(XiT β)
φ(XiT β)
1
(Y
)
−
1{0} (Yi ) .
{1} i
Φ(XiT β)
1 − Φ(XiT β)
Proposition 10. If ρ̂ is calculated using (5) with λ, L, and ε from Proposition 9, then for m ≥ 1 and β ∈ Rp ,
Z
(m)
kAC (β, β ′ ; Y, X) − πB|Y,X (β ′ |Y, X) dβ ′ ≤ H(β) ρ̂m−1 ,
Rp
where
H(β) = 2 +
L
+ tr XΣ−1 X T + Σ1/2 Σ−1 X T ϕ(β) + Qv − BF
1−λ
Proof. We simply apply Corollary 3. First, putting M = Σ1/2 in (15), we have
Z
Z
Σ1/2 Σ−1 X T z + Qv − BF
Ṽ1 (γ) πΓ|B,Y,X (γ|β, Y, X) dγ =
Rn
Rp
2
2
.
πZ|B,Y,X (z|β, Y, X) dz
2
= Σ1/2 Σ−1 X T ϕ(β) + Qv − BF
o
n
+ tr Σ−1/2 X T var(Z|B = β, Y, X)XΣ−1/2 .
Finally, a calculation similar to (16) shows that
o
n
tr Σ−1/2 X T var(Z|B = β, Y, X)XΣ−1/2 ≤ tr XΣ−1 X T
,
and the result follows.
Calculating λ in Proposition 9 calls for maximization of a function on (0, 1)×Rp , which may be difficult.
p
Here we provide an upper bound on λ that’s easy to compute when p is small. Let {Sj }2j=1 denote the open
orthants of Rp . For instance, if p = 2, then S1 , S2 , S3 and S4 are the open quadrants of the real plane. Define
W (Sj ) = W (Sj ; X, Y ) as follows
W (Sj ) =
X
Xi 1{0} (Yi )XiT +
Xi ∈Sj
X
Xi 1{1} (Yi )XiT .
Xi ∈−Sj
The following result is proven in Subsection B.3.
Proposition 11. An upper bound on λ1/2 in Proposition 9 is
2
λmax Σ−1/2 X T XΣ−1/2 −
min p λmin Σ−1/2 W (Sj )Σ−1/2 .
π 1≤j≤2
If this upper bound is strictly less than 1 (which is always true when Q is positive definite), then one can
replace λ with its square in Proposition 9.
20
4.3 Drift and minorization for the Albert and Chib chain based on V2
The A&C chain has the same convergence rate as the flipped chain defined by the following Mtd:
Z
′
′
πZ|B,Y,X (z ′ |β, Y, X)πB|Z,Y,X (β|z, Y, X) dβ .
ǩAC (z, z ) := ǩAC (z, z ; Y, X) =
Rp
In this subsection, we use V2 to establish d&m conditions for this chain, and these will be used later to prove
asymptotic stability as p → ∞. First, for z ∈ Rn , define
w(z) = XΣ−1 X T
−1/2 −1 T
X Σ (X z + Qv) − BF
.
Now define V̌2 : Rn → [0, ∞) as V̌2 (z) = kw(z)k2 .
Proposition 12. The Mtd ǩAC satisfies the drift condition
Z
V̌2 (z ′ )ǩAC (z, z ′ )dz ′ ≤ λV̌2 (z) + L ,
(20)
Rn
where λ = λ2max XΣ−1 X T
and L = n(1 + λ). Moreover, for d > 2L/(1 − λ), ǩAC satisfies
ǩAC (z, z ′ ) ≥ εq(z) ,
where q : Rn → [0, ∞) is a pdf, and ε = 2−n/2 e−d .
Proof. It’s easy to verify that
Z
Rp
V2 (β ′ )πB|Z,Y,X (β ′ |z, Y, X) dβ ′ = V̌2 (z) + n .
We know from Proposition 8 that
Z
Z
′ ′
′
′
′
πB|Z,Y,X (β |z , Y, X) πZ|B,Y,X (z |β, Y, X) dz dβ ′ ≤ λ2max XΣ−1 X T V2 (β)+2n .
V2 (β )
Rn
Rp
Multiplying both sides of the above inequality by πB|Z,Y,X (β|z, Y, X) and integrating with respect to β
yields (20).
We now move on the the minorization condition. Note that πZ|B,Y,X (z|B, Y, X) depends on B only
through XB, which is a one-to-one function of the n-dimensional vector
A := XΣ−1 X T
−1/2
X(B − BF ) .
Hence, ǩAC (z, z ′ ) can be re-expressed as
Z
′
πZ|A,Y,X (z ′ |α, Y, X)πA|Z,Y,X (α|z, Y, X) dα ,
ǩAC (z, z ) =
Rn
21
where A|Z, Y, X ∼ N (w(Z), In ). To get the minorization condition, we will construct a pdf ν(α) :=
ν(α|Y, X) such that
πA|Z,Y,X (α|z, Y, X) ≥ εν(α)
whenever V̌2 (z) ≤ d. Define
ν1 (α) =
Then ν(α) = ν1 (α)/
−n/2
ε = (2π)
inf
z:V̌2 (z)≤d
R
Rn
πA|Z,Y,X (α|z, Y, X) =
1
inf (2π)−n/2 exp − kα − wk2 .
2
w:w 2 ≤d
ν1 (α′ ) dα′ is a pdf, and
1
inf exp − kα − wk2 dα ≥ (2π)−n/2 e−d
2
2
Rn w:w ≤d
Z
Z
exp(−kαk2 ) dα = 2−n/2 e−d .
Rn
The next result is the analogue of Proposition 10. The proof is omitted as it is essentially the same as the
proof of the said proposition.
Proposition 13. Assume that X has full row rank. Let λ, L, and ε be as in Proposition 12. If ρ̂ is calculated
using (5), then for m ≥ 1 and β ∈ Rp ,
Z
Rp
(m)
kAC (β, β ′ ; Y, X) − πB|Y,X (β ′ |Y, X) dβ ′ ≤ H(β) ρ̂m−1 ,
where
H(β) = 2 +
L
+ tr XΣ−1 X T +
1−λ
XΣ−1 X T
−1/2
X Σ−1 X T ϕ(β) + Qv − BF
2
.
Let ρ∗ = ρ∗ (X, Y ) denote the geometric convergence rate of the A&C chain. In the next section, which
is the heart of the paper, we develop general convergence complexity results showing that, under weak
regularity conditions, ρ∗ is bounded away from 1 both as n → ∞ (for fixed p), and as p → ∞ (for fixed n).
5 Results for the Albert and Chib Chain Part II: Asymptotics
5.1 Large n, small p
In this section, we consider the case where p is fixed and n grows. In particular, we are interested in what
happens to the geometric convergence rate of the A&C chain in this setting. Recall that the prior on B is
ω(β) ∝ exp{−(β − v)T Q(β − v)/2} .
22
Since p is fixed, so is the prior. Hence, the hyperparameters v and Q will remain fixed throughout this
subsection. In Subsection 3.1, we introduced the data set D := {(Xi , Yi )}ni=1 . We now consider a sequence
of data sets, Dn := {(Xi , Yi )}ni=1 . So, each time n increases by 1, we are are given a new p × 1 covariate
vector, Xi , and a corresponding binary response, Yi . In order to study the asymptotics, we assume that the
(Xi , Yi ) pairs are generated according to a random mechanism that is consistent with the probit regression
model. In particular, we make the following assumptions:
p
(A1) The pairs {(Xi , Yi )}∞
i=1 are iid random vectors such that Yi |Xi ∼ Bernoulli(G(Xi )), where G : R →
(0, 1) is a measurable function;
(A2) EX1 X1T is finite and positive definite;
(A3) For j ∈ {1, 2, . . . , 2p } and β 6= 0,
P 1Sj (X1 ) + 1Sj (−X1 ) X1T β 6= 0 > 0 .
Assumption (A1) contains the probit model as a special case. Thus, our results concerning the asymptotic
behavior of A&C’s Markov chain do not require strict adherence of the data to the probit regression model.
The only reason for assuming (A2) is to guarantee that, almost surely, X will eventually have full column
rank. This is a necessary condition for posterior propriety when Q = 0. While (A3) is rather technical, it
is clearly satisfied if X1 follows a distribution admitting a pdf (with respect to Lebesgue measure) that is
positive over some open ball {γ ∈ Rp : kγk2 ≤ c}, where c > 0. It is shown in Subsection B.4 that (A3)
also allows for an intercept. That is, even if the first component of X1 is 1 (constant), then as long as the
remaining p − 1 components satisfy the density condition described above, (A3) is still satisfied.
It was assumed throughout Section 3 that the posterior distribution is proper. Of course, for a fixed data
set, this is check-able. All we need in the large n, small p regime is a guarantee that the posterior is proper
for all large n, almost surely. The following result is proven in Subsection B.5.
Proposition 14. Under assumptions (A1)-(A3), almost surely, the posterior distribution is proper for all
sufficiently large n.
For fixed n, Dn := {(Xi , Yi )}ni=1 represents the first n (random) covariate vectors and responses. Let
ρ∗ (Dn ) denote the (random) geometric convergence rate of the corresponding A&C chain. Here is one of
our main results.
Theorem 15. If (A1)-(A3) hold, then there exists a constant ρ < 1 such that, almost surely,
lim sup ρ∗ (Dn ) ≤ ρ .
n→∞
23
Proof. Let ρ̂(Dn ) denote the upper bound on ρ∗ (Dn ) that is based on λ, L, and ε from Proposition 9. We
prove the result by showing that, almost surely, lim supn→∞ ρ̂(Dn ) ≤ ρ < 1. Note that L = p(1 + λ)
and ε = 2−p/2 e−d (where d > 2L/(1 − λ)). Thus, control over λ provides control over L and ε as well.
In particular, to prove the result it suffices to show that there exists a constant c ∈ [0, 1), such that, almost
surely,
lim sup λ(Dn ) ≤ c .
(21)
n→∞
Noting that Σ−1/2 X T XΣ−1/2 ≤ Ip , we have, by Proposition 11,
(
!
)
−1/2
−1/2
X
X
2
1
Σ
Σ
1
λ1/2 ≤ 1−
Xi 1{0} (Yi )XiT +
Xi 1{1} (Yi )XiT
.
min λmin
π 1≤j≤2p
n
n
n
n
Xi ∈Sj
Xi ∈−Sj
(22)
Fix j ∈ {1, 2, . . . , 2p }. By (A1) and the strong law, almost surely,
!
−1/2
X
Σ −1/2
1
1 X
Σ
Xi 1{0} (Yi )XiT +
Xi 1{1} (Yi )XiT
lim
n→∞ n
n
n
n
Xi ∈Sj
Xi ∈−Sj
−1/2
−1/2
= EX1 X1T
EX1 X1T 1Sj (X1 )1{0} (Y1 ) + EX1 X1T 1Sj (−X1 )1{1} (Y1 ) EX1 X1T
o
−1/2 n
−1/2
= EX1 X1T
EX1 X1T 1Sj (X1 )(1 − G(X1 )) + EX1 X1T 1Sj (−X1 )G(X1 ) EX1 X1T
.
It follows from (26) and Lemma 22 that, almost surely,
!
)
(
−1/2
X
Σ
1
Σ −1/2 1 X
Xi 1{0} (Yi )XiT +
Xi 1{1} (Yi )XiT
lim λmin
n→∞
n
n
n
n
Xi ∈Sj
Xi ∈−Sj
n
o
T
T
T
≥ λ−1
EX
X
EX
X
1
(X
)(1
−
G(X
))
+
EX
X
1
(−X
)G(X
)
. (23)
λ
1 1
1 1 Sj
1
1
1 1 Sj
1
1
min
max
T
By (A2), λ−1
max EX1 X1 > 0. Hence by (22) and (23), to show that (21) holds, almost surely, it is enough
to show that
n
o
λmin EX1 X1T 1Sj (X1 )(1 − G(X1 )) + EX1 X1T 1Sj (−X1 )G(X1 ) > 0 .
By (A3), for any β 6= 0,
P 1Sj (X1 )X1T β 6= 0 or 1Sj (−X1 )X1T β 6= 0 > 0 .
(24)
Since 0 < G(X1 ) < 1, (24) implies that, for any β 6= 0,
P β T X1 X1T 1Sj (X1 ) 1 − G(X1 ) β + β T X1 X1T 1Sj (−X1 )G(X1 ) β > 0 > 0
As a result, for any β ∈ Rp such that kβk2 = 1,
o
n
β T EX1 X1T 1Sj (X1 )(1 − G(X1 )) + EX1 X1T 1Sj (−X1 )G(X1 ) β > 0 .
24
(25)
It follows from (A2) that the left-hand side of (25) is continuous in β. Hence, we can take infimum on both
sides of the inequality with respect to β and retain the greater-than symbol, which yields
n
o
λmin EX1 X1T 1Sj (X1 )(1 − G(X1 )) + EX1 X1T 1Sj (−X1 )G(X1 ) > 0 ,
and the result follows.
Theorem 15 shows that, under weak regularity conditions on the random mechanism that generates
Dn := {(Xi , Yi )}ni=1 , A&C’s MCMC algorithm scales well with n. Johndrow et al. (2016) studied the
convergence rate of the A&C chain as n → ∞ for a particular fixed sequence of covariates and responses.
Suppose that p = 1, Q > 0 is a constant, v = 0, and X1 = X2 = · · · = Xn = 1. Johndrow et al.
(2016) showed that if all the Bernoulli trials result in success, i.e., Y1 = Y2 = · · · = Yn = 1, then
limn→∞ ρ∗∗ (Dn ) = 1. That is, in terms of L2 -geometric convergence rate, the convergence is arbitrarily
slow for sufficiently large n. As we now explain, our results can be used to show that, in Johndrow et al.’s
(2016) setting, almost any other sequence of responses leads to well behaved convergence rates. Let {Yi }∞
i=1
Pn
−1
denote a fixed sequence of binary responses, and define p̂n = n
i=1 Yi . It follows from Propositions 9
and 11 that the A&C chain satisfies d&m conditions with
2
2
n
2 n
2
λ=
−
≤ 1 − {p̂n ∧ (1 − p̂n )} ,
p̂n ∧ (1 − p̂n )
n+Q πn+Q
π
L = 1 + λ, and ε = 2−1/2 e−d for d > 2L/(1 − λ). For any fixed n, suppose that there exist c1 , c2 ∈ (0, 1)
such that c1 ≤ p̂n ≤ c2 , then (using (5)) one can find ρ < 1, which depends only on c1 ∧ (1 − c2 ), such that
ρ∗ (Dn ) ≤ ρ . It now follows that the geometric convergence rates, ρ∗ (Dn ), are eventually bounded away
from 1 so long as 0 < lim inf n→∞ p̂n ≤ lim supn→∞ p̂n < 1. (It is important to note that Dn is not random
here.) Moreover, an analogous result holds for ρ∗∗ (Dn ). Here’s a formal statement.
Corollary 16. For the intercept-only model described above, if 0 < lim inf n→∞ p̂n ≤ lim supn→∞ p̂n < 1,
then for any
δ∈
2
0, 1 − 1 −
π
2 !
lim inf p̂n ∧ 1 − lim sup p̂n
,
n→∞
n→∞
lim supn→∞ ρ∗ (Dn ) ≤ ρ < 1, and lim supn→∞ ρ∗∗ (Dn ) ≤ ρ < 1, where ρ equals ρ̂ in (5) with
2
2
λ= 1−
lim inf p̂n ∧ 1 − lim sup p̂n
+δ ,
π n→∞
n→∞
L = 1 + λ , and ε = 2−1/2 e−d , where d > 2L/(1 − λ).
25
Proof. It suffices to prove the result for ρ∗∗ , since the argument for ρ∗ has already been provided. Fix an
arbitrary δ. By Proposition 10, when n is sufficiently large, for any β ∈ R and m ≥ 1,
Z
(m)
kAC (β, β ′ ; Y, X) − πB|Y,X (β ′ |Y, X) dβ ′ ≤ H(β) ρm−1 ,
R
where H(β) is given in the said proposition. Let Π be the probability measure corresponding to the posterior
density, πB|Y,X (β|Y, X). Then for any probability measure ν ∈ L2 (Π) and m ≥ 1,
Z Z
(m)
kAC (β, β ′ ; Y, X)ν(dβ)−πB|Y,X (β ′ |Y, X) dβ ′
R
R
Z Z
(m)
kAC (β, β ′ ; Y, X) − πB|Y,X (β ′ |Y, X) dβ ′ ν(dβ)
≤
R R
Z
H(β) ν(dβ) ρm−1 .
≤
R
One can verify that H(β) can be bounded by polynomial functions. As a result, by Theorem 2.3 in
R
R
Chen and Shao (2000), R H 2 (β)πB|Y,X (β|Y, X) dβ < ∞. Then by Cauchy-Schwarz, R H(β)ν(dβ) <
∞. Therefore, ρ∗∗ (Dn ) ≤ ρ for all sufficiently large n .
Recall that the convergence rate analyses of Roy and Hobert (2007) and Chakraborty and Khare (2017),
which establish the geometric ergodicity of the A&C chain for fixed n and p, are based on an un-centered
drift function. We end this subsection with a result showing that, while this un-centered drift may be adequate
for non-asymptotic results, it simply does not match the dynamics of the A&C chain well enough to get a
result like Theorem 15. The following result is proven is Subsection B.6.
Proposition 17. Assume that (A1) and (A2) hold, and that there exists β∗ ∈ Rp such that β∗ 6= 0 and
G(γ) = G∗ (γ T β∗ ) for all γ ∈ Rp , where G∗ : R → (0, 1) is a strictly increasing function such that
G∗ (0) = 1/2. Then, almost surely, any drift and minorization based on V0 (β) = kΣ1/2 βk2 is necessarily
unstable in n.
Remark 18. In the above proposition, if G∗ (θ) = Φ(θ) for all θ ∈ R, then the probit model is correctly
specified, and the true parameter is β∗ .
5.2 Large p, small n
In this subsection, we consider the case where n is fixed and p grows. In contrast with the strategy of
the previous subsection, here we consider a deterministic sequence of data sets. Also, since p is changing,
we need to specify a sequence of prior parameters {(Qp , vp )}∞
p=1 . Let Dp := {(vp , Qp , Xp , Y )} denote a
sequence of priors and data sets, where Y is a fixed n × 1 vector of responses, Xp is an n × p matrix, vp is
26
a p × 1 vector, and Qp is a p × p positive definite matrix. (Note that positive definite-ness of Qp is required
for posterior propriety.) So, each time p increases by 1, we are are given a new n × 1 column vector to add
to the current design matrix. For the rest of this subsection, we omit the p subscript. We also assume that the
following conditions hold for all p:
(B1) X has full row rank;
(B2) There exists a finite, positive constant, c, such that λmax (XQ−1 X T ) < c .
Assumption (B1) is equivalent to XΣ−1 X T being non-singular. Assumption (B2) regulates the eigenvalues of the prior variance, Q−1 . More specifically, it requires that the prior drives B towards v. For
illustration, if X = (1, 1, . . . , 1), then (B2) holds if for some τ > 0, Q−1 = diag(τ /p, τ /p, . . . , τ /p),
or Q−1 = diag(τ, τ /22 , . . . , τ /p2 ). Assumption (B2) is satisfied by the generalized g-priors used by,
e.g., Gupta and Ibrahim (2007), Ai-Jun and Xin-Yuan (2009), and Baragatti and Pommeret (2012). It can
be shown (see, e.g., Chakraborty and Khare, 2017) that (B2) is equivalent to the existence of a constant
c < 1 such that
λmax XΣ−1 X T < c .
While (B2) may seem like a strong assumption, we will provide some evidence later in this subsection
suggesting that it may actually be necessary. Here is our main result concerning the large p, small n case.
Theorem 19. If (B1) and (B2) hold, then there exists a constant ρ < 1 such that ρ∗ (Dp ) ≤ ρ for all p.
Proof. The proof is based on Proposition 12. Indeed, as in the proof of Theorem 15, it suffices to show that
there exists a c < 1 such that λ(Dp ) = λ2max XΣ−1 X T < c for all p. But this follows immediately from
(B2).
An important feature of Theorem 19 is that it holds for any sequence of prior means, {vp }∞
p=1 . This is
achieved by adopting a drift function that is centered around a point that adapts to the prior mean. Although
Gaussian priors with non-vanishing means are not commonly used in practice, it is interesting to see that
Albert and Chib’s algorithm can be robust under location shifts in the prior, even when the dimension of the
state space is high.
The following result, which is proven in Subsection B.7, shows that (B2) is not an unreasonable assumption.
Proposition 20. If n = 1 and v = 0, then as XΣ−1 X T tends to 1,
1 − ρ∗∗ = O 1 − XΣ−1 X T .
27
In particular, ρ∗∗ is not bounded away from 1 if (B2) does not hold.
Acknowledgment. The second author’s work was supported by NSF Grant DMS-15-11945.
Appendices
A Useful Results
A.1 Hermitian matrices
For a Hermitian matrix M , let λmin (M ) and λmax (M ) denote the smallest and largest eigenvalues of M ,
respectively. The following result is part of a famous result due to H. Weyl (see, e.g., Horn and Johnson,
2012, Section 4.3).
Lemma 21. [Weyl’s Inequality] Let M1 and M2 be Hermitian matrices of the same size, then
λmax (M2 ) + λmin (M1 − M2 ) ≤ λmax (M1 ) ≤ λmax (M2 ) + λmax (M1 − M2 ) ,
and
λmin (M2 ) + λmin (M1 − M2 ) ≤ λmin (M1 ) ≤ λmin (M2 ) + λmax (M1 − M2 ) .
Let M1 and M2 be non-negative definite matrices of the same size. It is well known that
λmax (M1 M2 M1 ) ≤ λ2max (M1 ) λmax (M2 ) ,
and
λmin (M1 M2 M1 ) ≥ λ2min (M1 ) λmin (M2 ) .
(26)
The following result follows immediately from Lemma 21.
Lemma 22. Let {Mj }∞
j=1 be a sequence of Hermitian matrices of the same size such that Mj → M . Then
M is Hermitian, limj→∞ λmax (Mj ) = λmax (M ), and limj→∞ λmin (Mj ) = λmin (M ).
A.2 Truncated normal distributions
Fix θ ∈ R. Let Z+ ∼ TN(θ, 1; 1) and Z− ∼ TN(θ, 1; 0). Then we have
EZ+ = θ +
φ(θ)
Φ(θ)
and
28
EZ− = θ −
φ(θ)
.
1 − Φ(θ)
(27)
Also,
θφ(θ)
−
varZ+ = 1 −
Φ(θ)
φ(θ)
Φ(θ)
2
θφ(θ)
varZ− = 1 +
−
1 − Φ(θ)
and
φ(θ)
1 − Φ(θ)
2
.
(28)
Note that
dEZ+
= varZ+
dθ
and
dEZ−
= varZ− ,
dθ
which shows that both expectations are increasing in θ. Recall that
φ(θ) 2
θφ(θ)
.
+
g(θ) =
Φ(θ)
Φ(θ)
Thus, we have the shorthand varZ+ = 1 − g(θ) and varZ− = 1 − g(−θ). The following result is Theorem 3
in Horrace (2015).
Lemma 23 (Horrace (2015)). The function g(θ) is decreasing in θ. When θ < 0, g(θ) ∈ (2/π, 1), and when
θ > 0, g(θ) ∈ (0, 2/π). In particular, varZ+ , varZ− ∈ (0, 1).
B Proofs
B.1 Corollary 3 via Proposition 24
We begin by resetting the stage from Section 2. Let X ⊂ Rp , Y ⊂ Rq , and let s̃ : X × Y → [0, ∞) and
h̃ : Y × X → [0, ∞) be the conditional pdfs defined in Section 2. Let {Ψm }∞
m=0 be a Markov chain on X
whose Mtd is given by
′
k(x, x ) =
and let
{Ψ̃m }∞
m=0
Z
s̃ x′ |γ h̃ γ|x dγ ,
Y
be a Markov chain on Y whose Mtd is given by
Z
′
h̃ γ ′ |x s̃ (x|γ) dx .
k̃(γ, γ ) =
X
Let K(x, ·) and Π denote the Mtf and invariant probability measure for {Ψm }∞
m=0 , respectively, and define
K̃(γ, ·) and Π̃ analogously. Assume that both chains are Harris ergodic (i.e., irreducible, aperiodic, and
Harris recurrent). The following result provides a total variation distance connection between the two chains.
Proposition 24. Suppose that there exists R : Y × Z+ → (0, ∞) such that for Ψ̃0 ∼ ν̃ and m ≥ 0, we have
kν̃ K̃ m (·) − Π̃(·)kTV ≤ Eν̃ R(Ψ̃0 , m) .
Then for Ψ0 ∼ ν and m ≥ 1, we have
m
kνK (·) − Π(·)kTV ≤ Eν
Z
Y
29
R(γ, m − 1)h̃(γ|Ψ0 ) dγ
.
Proof. Denote the m-step Mtds of the two chains as k(m) (x, x′ ) and k̃(m) (γ, γ ′ ), respectively. Moreover, let
π(x) and π̃(γ) be the respective stationary densities for the two chains, and note that
π(x) =
Z
s̃(x|γ)π̃(γ) dγ .
Y
The key is to relate the two m-step Mtds. Indeed, for m ≥ 1, we have
k
(m)
′
(x, x ) =
Z
s̃(x′ |γ ′ )k̃(m−1) (γ, γ ′ )h̃(γ|x) dγ dγ ′ .
Y2
It follows that for any m ≥ 1,
kνK m (·) − Π(·)kTV =
Z
Eν k(m) (Ψ0 , x′ ) − π(x′ ) dx′
XZ
k(m) (Ψ0 , x′ ) − π(x′ ) dx′
≤ Eν
Z
ZX Z
(m−1)
′
′
′ ′
= Eν
k̃
(γ, γ )h̃(γ|Ψ0 ) dγ − π̃(γ ) dγ ′ dx′
s̃(x |γ )
X
Y
Y
Z Z
k̃(m−1) (γ, γ ′ )h̃(γ|Ψ0 ) dγ − π̃(γ ′ ) dγ ′ .
≤ Eν
Y
Y
Letting ν̃ be the probability measure associated with h̃(γ|Ψ0 ) yields
Z
m
m−1
kνK (·) − Π(·)kTV ≤ Eν kν̃K
− Π̃kTV ≤ Eν
R(γ, m − 1)h̃(γ|Ψ0 ) dγ .
Y
B.2 An inequality related to Proposition 7
Let
kΣ−1/2 X T D(BF + tα)Xαk2
.
kΣ1/2 αk2
t∈(0,1) α6=0
λ = sup sup
Proposition 25. λ < 1.
Proof. If Q is positive definite, then by Lemmas 23 and 21,
λ ≤ sup λ2max (Σ−1/2 X T D(β)XΣ−1/2 ) ≤ λ2max (Ip − Σ−1/2 QΣ−1/2 ) < 1 .
β∈Rp
For the remainder of the proof, assume that Q = 0.
We begin by introducing some notations. For i = 1, 2, . . . , n, let Xi∗ = Xi 1{0} (Yi )−Xi 1{1} (Yi ). Define
A1 , A2 , . . . , A2n to be the subsets of {1, 2, . . . , n}. For each Aj , define
Tj = α 6= 0 : (Xi∗ )T α ≥ 0 for all i ∈ Aj , (Xi∗ )T α ≤ 0 for all i ∈ Acj .
30
Let C = {j : Tj 6= ∅}. It’s easy to see that
S
j∈C
Tj = Rp \ {0}.
Note that XΣ−1 X T ≤ In . Thus, for any α ∈ Rp and t ∈ (0, 1),
kΣ−1/2 X T D(BF + tα)Xαk2 ≤ kD(BF + tα)Xαk2 .
Suppose that α ∈ Tj , then by Lemma 23, for all i ∈ Aj and t > 0, the ith diagonal of D(BF + tα) satisfies
1 − g(XiT (BF + tα))1{1} (Yi ) − g(−XiT (BF + tα))1{0} (Yi ) = 1 − g((−Xi∗ )T (BF + tα)) ∈ (0, 1 − c0 ],
where
c0 = min g(−(Xi∗ )T BF ) ∈ (0, 1) .
1≤i≤n
As a result, whenever α ∈ Tj ,
kD(BF + tα)Xαk2 ≤ kXαk2 − c0 (2 − c0 )
X
kXiT αk2
i∈Aj
for all t ∈ (0, 1). It follows that
X kX T αk2
i
.
1/2 αk2
α∈Tj
kΣ
i∈A
λ ≤ 1 − c0 (2 − c0 ) min inf
j∈C
j
Let Rj (α) =
P
i∈Aj
kXiT αk2 /kΣ1/2 αk2 . Note that for any constant c 6= 0, Rj (cα) = Rj (α). Hence, to
show that λ < 1, it suffices to verify that for each j ∈ C, inf α∈Tj∗ Rj (α) > 0, where Tj∗ = {α ∈ Tj : kαk =
1}. Since Tj∗ is compact, and Rj (α) is continuous in Tj∗ , it’s enough to show that for all α ∈ Tj∗ , Rj (α) > 0.
Lesaffre and Kaufmann (1992) showed that BF exists if and only if the following condition holds:
(C0) For every α ∈ Rp \ {0}, there exists an i ∈ {1, 2, . . . , n} such that either XiT α 1{1} (Yi ) < 0 or
XiT α1{0} (Yi ) > 0.
Hence, for any α ∈ Rp , there exists i ∈ {1, 2, . . . , n} such that (Xi∗ )T α > 0. This implies that for any
j ∈ C and α ∈ Tj∗ , there exists i ∈ Aj such that kXiT αk > 0. As a result, Rj (α) > 0 for any α ∈ Tj∗ .
B.3 Proposition 11
Proposition 11. An upper bound on λ1/2 in Proposition 9 is
2
min p λmin Σ−1/2 W (Sj )Σ−1/2 .
λmax Σ−1/2 X T XΣ−1/2 −
π 1≤j≤2
31
Proof. For θ ∈ {0, 1} and γ, β ∈ Rp , define
u(γ; θ, β) =
By Lemma 23, for any β ∈ Rp ,
By Lemma 21,
1 if θ = 1 and γ T β > 0
1 if θ = 0 and γ T β < 0
0 otherwise .
2
1 − u(Xi ; Yi , β) .
1 − 1{1} (Yi )g XiT β − 1{0} (Yi )g −XiT β ≤ 1 −
π
kΣ−1/2 X T D(BF + tα)Xαk
kΣ1/2 αk
t∈(0,1) α6=0
≤ sup λmax Σ−1/2 X T D(β)XΣ−1/2
λ1/2 = sup sup
β∈Rp
≤ λmax Σ−1/2 X T XΣ−1/2
2
−
inf λmin
π β∈Rp
(
Σ−1/2
Now, since u(Xi ; Yi , 0) = 0,
n
X
i=1
Xi 1 − u(Xi ; Yi , β) XiT
!
Σ−1/2
)
.
1 − u(Xi ; Yi , 0) ≥ 1 − u(Xi ; Yi , β)
for any β ∈ Rp . Therefore,
#
)
( n
"
X
2
Xi 1 − u(Xi ; Yi , β) XiT Σ−1/2
inf λmin Σ−1/2
π β∈Rp
i=1
#
)
( n
"
X
T
2
−1/2
−1/2
= inf λmin Σ
.
Xi 1 − u(Xi ; Yi , β) Xi Σ
π β6=0
i=1
For j =
1, 2, . . . , 2p ,
let
Sj′
be the closure of Sj . Note that if a vector β ∈ Sj′ \ {0} for some j, then for any
θ ∈ {0, 1} and γ ∈ Sj ∪ (−Sj ),
u(γ; θ, β) =
1
{1} (θ)
{0} (θ)
S2p
if γ ∈ −Sj .
= Rp , by Lemma 21,
#
)
( n
"
X
T
2
−1/2
−1/2
Xi 1 − u(Xi ; Yi , β) Xi Σ
inf λmin Σ
π β6=0
i=1
#
)
( n
"
X
2
=
Xi 1 − u(Xi ; Yi , β) XiT Σ−1/2
min
inf λmin Σ−1/2
π 1≤j≤2p β∈Sj′ \{0}
i=1
X
X
2
min p λmin Σ−1/2
Xi 1{0} (Yi )XiT +
Xi 1{1} (Yi )XiT Σ−1/2 .
≥
π 1≤j≤2
Since
′
j=1 Sj
1
if γ ∈ Sj
Xi ∈Sj
Xi ∈−Sj
The result follows immediately.
32
B.4 Assumption (A3) allows for an intercept
Suppose that X1 is a p-dimensional random vector with p > 1, such that the first component of X1 is 1, and
the remaining p − 1 components follow a distribution that admits a pdf with respect to Lebesgue measure
that is positive over the open ball Bc := {γ ∈ Rp−1 : kγk2 < c}, where c > 0. We will demonstrate that
(A3) is satisfied. Let β ∈ Rp \ {0} be arbitrary. The set
γ ∈ Rp : γ T β = 0 ∩ (1, γ T )T : γ ∈ Bc
is in a hyperplane, whose dimension is at most p−2. Meanwhile, the support of X1 ’s distribution has positive
Lebesgue measure on the (p − 1)-dimensional hyperplane (1, γ T )T : γ ∈ Rp−1 . Thus P(X1T β 6= 0) = 1.
On the other hand, it’s easy to verify that P X1 ∈ Sj ∪ (−Sj ) > 0. Thus,
P 1Sj (X1 ) + 1Sj (−X1 ) X1T β 6= 0 = P 1Sj (X1 ) + 1Sj (−X1 ) 6= 0 and X1T β 6= 0 > 0 ,
so (A3) holds.
B.5 Proposition 14
Proposition 14. Under assumptions (A1)-(A3), almost surely, the posterior distribution is proper for sufficiently large n.
Proof. It suffices to consider the case that Q = 0. Fix n and p. Chen and Shao (2000) proved that the
posterior is proper if the following two conditions are satisfied:
(C1) X has full column rank;
(C3) The maximum likelihood estimator of β exists in Rp .
(These two conditions are equivalent to (C1) and (C2) in Subsection 3.1.) As mentioned in Subsection B.2,
(C3) holds if and only if the following condition is satisfied:
(C0) For every β ∈ Rp \ {0}, there exists an i ∈ {1, 2, . . . , n} such that either XiT β 1{1} (Yi ) < 0 or
XiT β1{0} (Yi ) > 0.
We can therefore prove the result by showing that, almost surely, (C1) and (C0) will hold for all sufficiently
large n. As mentioned previously, (A2) ensures that, almost surely, (C1) will hold for all sufficiently large n.
p
Now recall that {Sj }2j=1 denotes the set of open orthants in Rp . As before, let Sj′ denote the closure of Sj .
It follows from (A1) and (A3) that, for every j ∈ {1, 2, . . . , 2p }, at least one of the following two statements
must hold:
33
• P X1 ∈ Sj , Y1 = 1 > 0 and P X1 ∈ Sj , Y1 = 0 > 0 ,
• P X1 ∈ −Sj , Y1 = 1 > 0 and P X1 ∈ −Sj , Y1 = 0 > 0 .
Thus, almost surely, when n is sufficiently large (say n > N ), for each j ∈ {1, 2, . . . , 2p } there exists
i′ , i′′ ∈ {1, 2, . . . , n} such that either Xi′ , Xi′′ ∈ Sj , (Yi′ , Yi′′ ) = (1, 0), or Xi′ , Xi′′ ∈ −Sj , (Yi′ , Yi′′ ) =
p
(1, 0). Now suppose that n > N , and let β ∈ Rp \ {0} be arbitrary. Since ∪2j=1 (Sj′ \ {0}) = Rp \ {0}, there
′
exists an integer j(β) such that β ∈ Sj(β)
\ {0}. One can pick i′ and i′′ such that either Xi′ , Xi′′ ∈ Sj(β) ,
(Yi′ , Yi′′ ) = (1, 0), or Xi′ , Xi′′ ∈ −Sj(β) , (Yi′ , Yi′′ ) = (1, 0). Note that for any γ ∈ Sj(β) , γ T β > 0.
Depending on whether Xi′ , Xi′′ ∈ Sj(β) or Xi′ , Xi′′ ∈ −Sj(β), (C0) holds with either i = i′′ or i = i′ .
B.6 Proposition 17
Proposition 17. Assume that (A1) and (A2) hold, and that there exists β∗ ∈ Rp such that β∗ 6= 0 and
G(γ) = G∗ (γ T β∗ ) for all γ ∈ Rp , where G∗ : R → (0, 1) is a strictly increasing function such that
G∗ (0) = 1/2. Then, almost surely, any drift and minorization based on V0 (β) = kΣ1/2 βk2 is necessarily
unstable in n.
Proof. We prove the result for the flat prior, and leave the extension to the reader. Note that V0 (β) =
β T X T Xβ, and assume that there exist λ := λ(Dn ) < 1 and L := L(Dn ) < ∞ such that
Z
V0 (β ′ )kAC (β, β ′ ) dβ ′ ≤ λV0 (β) + L .
Rp
Suppose further that there exists d := d(Dn ) > 2L/(1 − λ) such that, for all β with V0 (β) < d,
kAC (β, β ′ ) ≥ εq(β ′ ) ,
where ε := ε(Dn ) > 0, and q(β ′ ) := q(β ′ |Y, X) is a pdf on Rp . We will show that ε → 0 as n → ∞ almost
surely.
Our first step is to show L → ∞. Let β = 0, then V0 (β) = 0, and
Z
V0 (β ′ )kAC (β, β ′ ) dβ ′ = p+tr (X T X)−1 X T var(Z|B = 0, Y, X)X(X T X)−1
Rp
+ E(Z T |B = 0, Y, X)X(X T X)−1 X T E(Z|B = 0, Y, X) .
Hence, when β = 0, a lower bound for
R
Rp
V0 (β ′ )kAC (β, β ′ ) dβ ′ is given by
!T
−1
n
1X
XT X
V0 (β) + p + n
Xi E(Zi |B = 0, Y, X)
n
n
i=1
34
!
n
1X
Xi E(Zi |B = 0, Y, X) .
n
i=1
(29)
Note that {Xi E(Zi |B = 0, Y, X)}ni=1 are iid. By (27), we have
r
r
r
2
2
2
1{1} (Y1 )−
1{0} (Y1 ) X1 =
E 2G∗ (X1T β∗ ) − 1 X1 .
E X1 E(Z1 |B = 0, Y, X) = E
π
π
π
Now, since EX1 X1T is positive definite and β∗ 6= 0, β∗T EX1 X1T β∗ > 0, indicating that P(X1T β∗ 6=
0) > 0. It’s easy to see that (2G∗ (X1T β∗ ) − 1)β∗T X1 > 0 whenever X1T β∗ 6= 0, which implies that
β∗T E 2G∗ (X1T β∗ ) − 1 X1 > 0. As a result, E 2G∗ (X1T β∗ ) − 1 X1 6= 0. Then, by the strong law
and (A2), almost surely,
lim
n→∞
!T
!
−1
n
n
1X
XT X
1X
Xi E(Zi |B = 0, Y, X)
Xi E(Zi |B = 0, Y, X)
n
n
n
i=1
i=1
−1
2
T
EX1 X1T
E 2G∗ (X1T β∗ ) − 1 X1 > 0 .
= E 2G∗ (X1T β∗ ) − 1 X1
π
It then follows from (29) that there exists a positive constant c such that almost surely, L > cn for any large
enough n.
Note that d > 2L. By (A2), X T X is of order n. Thus, almost surely, there exists a positive constant b
such that for all sufficiently large n,
β ∈ Rp : kβk2 ≤ b2 ⊂ β ∈ Rp : V0 (β) < d .
Let γ ∈ Rp be a fixed vector such that 0 < kγk2 ≤ b2 . Then both γ and −γ are in the small set β ∈
Rp : V0 (β) < d for all sufficiently large n. We now investigate the conditional densities kAC (γ, β ′ ) and
kAC (−γ, β ′ ). Let (B1 , B2 ) be a random vector such that the conditional distribution of B2 given B1 = β is
precisely the Mtd kAC (β, β ′ ). Then
T
−1
E(B2 |B1 = β) = (X X)
n
X
Xi E(Zi |B = β, Y, X) .
i=1
Thus,
2
E(B2 |B1 = γ) − E(B2 |B1 = −γ)
T −1 X
n
X X
1
=
Xi E(Zi |B = γ, Y, X) − E(Zi |B = −γ, Y, X)
n
n
2
.
(30)
i=1
Now using (27), we have
E(Zi |B = γ, Y, X) − E(Zi |B = −γ, Y, X)
φ(XiT γ)
φ(XiT γ)
φ(XiT γ)
φ(XiT γ)
T
T
−
+
1{1} (Yi ) + 2Xi γ −
1{0} (Yi )
= 2Xi γ +
Φ(XiT γ) 1 − Φ(XiT γ)
1 − Φ(XiT γ) Φ(XiT γ)
= 2XiT γ +
φ(XiT γ)
φ(XiT γ)
−
.
Φ(XiT γ) 1 − Φ(XiT γ)
35
Another appeal to the strong law yields
n
1X
Xi E(Zi |B = γ, Y, X) − E(Zi |B = −γ, Y, X)
lim
n→∞ n
i=1
φ(X1T γ)
φ(X1T γ)
T
−
X1 .
= E 2X1 γ +
Φ(X1T γ) 1 − Φ(X1T γ)
As in Subsection A.2, let Z+ ∼ TN(θ, 1; 1) and Z− ∼ TN(θ, 1; 0). Note that
2θ +
φ(θ)
φ(θ)
−
= EZ+ + EZ− .
Φ(θ) 1 − Φ(θ)
Since both expectations on the right-hand side are (strictly) increasing functions of θ, it follows that the
left-hand side is positive when θ > 0 and negative when θ < 0. Thus,
φ(X1T γ)
φ(X1T γ)
T
2X1 γ +
−
γ T X1 > 0
Φ(X1T γ) 1 − Φ(X1T γ)
(31)
whenever X1T γ 6= 0. Since EX1 X1T is positive definite and γ 6= 0, P(X1T γ 6= 0) > 0. Taking expectation
on both sides of (31) yields
T
γ E
2X1T γ
φ(X1T γ)
φ(X1T γ)
−
X1 > 0 ,
+
Φ(X1T γ) 1 − Φ(X1T γ)
and it follows that
E
2X1T γ
φ(X1T γ)
φ(X1T γ)
−
X1 6= 0 .
+
Φ(X1T γ) 1 − Φ(X1T γ)
By (30) and (A2), almost surely,
lim
n→∞
2
E(B2 |B1 = γ) − E(B2 |B1 = −γ)
φ(X1T γ)
φ(X1T γ)
T −1
T
= EX1 X1
E 2X1 γ +
−
X1
Φ(X1T γ) 1 − Φ(X1T γ)
2
=: cγ > 0 .
Next, consider var(B2 |B1 = γ) and var(B2 |B1 = −γ). Note that for any β ∈ Rp ,
var(B2 |B1 = β) = (X T X)−1 + (X T X)−1 X T var(Z|B = β, Y, X) X(X T X)−1 .
It follows from Lemma 23 that var(Zi |B = β, Y, X) < 1 for all i ∈ {1, 2, . . . , n}. Therefore,
var(B2 |B1 = β) ≤ 2(X T X)−1 .
Then by Chebyshev’s inequality, for any δ > 0,
P
B2 − E(B2 |B1 = γ)
2
2
> δ B1 = γ ≤ tr (X T X)−1 ,
δ
36
and
P
2
B2 − E(B2 |B1 = −γ)
2
> δ B1 = −γ ≤ tr (X T X)−1 .
δ
Note that by (A2), almost surely, tr {(X T X)−1 } = O(n−1 ). We have shown in the previous paragraph that
with probability 1, kE(B2 |B1 = γ) − E(B2 |B1 = −γ)k2 converges to a positive constant cγ . Now, fix
δ ∈ (0, cγ /4), and let C1 = β ∈ Rp : kβ − E(B2 |B1 = γ)k2 > δ , C2 = β ∈ Rp : kβ − E(B2 |B1 =
−γ)k2 > δ . With probability 1, for all sufficiently large n, C1 ∪ C2 = Rp . As a result, almost surely,
Z
Z
Z
min kAC (γ, β), kAC (−γ, β) dβ ≤ lim
kAC (−γ, β) dβ = 0 .
kAC (γ, β) dβ +
lim
n→∞
n→∞ Rp
C2
C1
Therefore,
ε≤
Z
inf
Rp
{β:V0 (β)<d}
with probability 1.
kAC (β, β ′ ) dβ ′ → 0
B.7 Proposition 20
Proposition 20. If n = 1 and v = 0, then as XΣ−1 X T tends to 1,
1 − ρ∗∗ = O 1 − XΣ−1 X T .
In particular, ρ∗∗ is not bounded away from 1 if (B2) does not hold.
Proof. Without loss of generality, assume that Y = Y1 = 1. The A&C chain has the same convergence rate
as the flipped chain defined by the following Mtd:
′
′
ǩAC (z, z ) := ǩAC (z, z ; Y, X) =
Z
Rp
πZ|B,Y,X (z ′ |β, Y, X)πB|Z,Y,X (β|z, Y, X) dβ .
The flipped chain is reversible with respect to πZ|Y,X (z|Y, X). Roberts and Rosenthal’s (1997) Theorem 2
and Liu et al.’s (1994) Lemma 2.3 imply that
ρ∗∗ ≥ γ0 :=
cov(Z, Z ′ |Y, X)
,
var(Z|Y, X)
where Z ′ |Z = z, Y, X ∼ ǩAC (z, ·), and Z|Y, X ∼ πZ|Y,X (·|Y, X). Hence, it suffices to show that
1 − γ0 = O 1 − XΣ−1 X T
37
.
We begin by giving an explicit form of πZ|Y,X (z|Y, X). Note that
Z
πZ|B,Y,X (z|β, Y, X)πB|Y,X (β|Y, X) dβ
Z
1 T
1
2
exp − (z − Xβ) − β Qβ dβ
∝ 1(0,∞) (z)
2
2
Rp
1
∝ 1(0,∞) (z) exp − (1 − ψ)z 2 ,
2
πZ|Y,X (z|Y, X) =
Rp
where ψ = XΣ−1 X T . Thus, Z|Y, X ∼ TN(0, (1 − ψ)−1 ; 1). It follows immediately from (27) and (28)
that
′
E(Z|Y, X) = E(Z , Y, X) =
r
2
(1 − ψ)−1/2 and var(Z|Y, X) =
π
2
1−
(1 − ψ)−1 .
π
To calculate cov(Z, Z ′ |Y, X), first note that
cov(Z, Z ′ |Y, X) = E(ZZ ′ |Y, X) −
2
(1 − ψ)−1 .
π
Now Z|B, Y, X ∼ TN(XB, 1; 1) and XB|Z, Y, X ∼ N (ψZ, ψ). Therefore,
φ(XB)
′
2
Y, X
E(ZZ |Y, X) = ψE Z |Y, X + E Z
Φ(XB)
≥ ψE Z 2 |Y, X
= ψ(1 − ψ)−1 .
Hence,
Collecting terms, we have
2
cov(Z, Z ′ |Y, X) ≥ 1 −
(1 − ψ)−1 − 1 .
π
γ0 ≥ 1 −
(1 − ψ)
,
(1 − 2/π)
and the result follows.
References
A I -J UN , Y. and X IN -Y UAN , S. (2009). Bayesian variable selection for disease classification using gene
expression data. Bioinformatics 26 215–222.
A LBERT, J. H. and C HIB , S. (1993). Bayesian analysis of binary and polychotomous response data. Journal
of the American statistical Association 88 669–679.
38
BARAGATTI , M. and P OMMERET, D. (2012). A study of variable selection using g-prior distribution with
ridge parameter. Computational Statistics & Data Analysis 56 1920–1934.
C HAKRABORTY, S. and K HARE , K. (2017). Convergence properties of Gibbs samplers for Bayesian probit
regression with proper priors. Electronic Journal of Statistics 11 177–210.
C HEN , M. H. and S HAO , Q. M. (2000). Propriety of posterior distribution for dichotomous quantal response
models. Proceedings of the American Mathematical Society 129 293–302.
D IACONIS , P., K HARE , K. and S ALOFF -C OSTE , L. (2008). Gibbs sampling, exponential families and
orthogonal polynomials (with discussion). Statistical Science 23 151–200.
D URMUS , A. and M OULINES , E. (2016). High-dimensional Bayesian inference via the unadjusted Langevin
algorithm. arXiv:1605.01559 .
F LEGAL , J. M., H ARAN , M. and J ONES , G. L. (2008). Markov chain Monte Carlo: Can we trust the third
significant figure? Statistical Science 23 250–260.
F ORT, G., M OULINES , E., ROBERTS , G. and ROSENTHAL , J. (2003). On the geometric ergodicity of
hybrid samplers. Journal of Applied Probability 40 123–146.
G UPTA , M. and I BRAHIM , J. G. (2007). Variable selection in regression mixture modeling for the discovery
of gene regulatory networks. Journal of the American Statistical Association 102 867–880.
H ORN , R. A. and J OHNSON , C. R. (2012). Matrix Analysis. 2nd ed. Cambridge University Press.
H ORRACE , W. C. (2015). Moments of the truncated normal distribution. Journal of Productivity Analysis
43 133–138.
J OHNDROW, J. E., S MITH , A., P ILLAI , N. and D UNSON , D. B. (2016). Inefficiency of data augmentation
for large sample imbalanced data. arXiv:1605.05798 .
J ONES , G. L. (2001). Convergence Rates and Monte Carlo Standard Errors for Markov Chain Monte Carlo
Algorithms. Ph.D. thesis, University of Florida.
J ONES , G. L. and H OBERT, J. P. (2001). Honest exploration of intractable probability distributions via
Markov chain Monte Carlo. Statistical Science 312–334.
Ł ATUSZY ŃSKI , K., M IASOJEDOW, B. and N IEMIRO , W. (2013). Nonasymptotic bounds on the estimation
error of MCMC algorithms. Bernoulli 19 2033–2066.
39
L ESAFFRE , E. and K AUFMANN , H. (1992). Existence and uniqueness of the maximum likelihood estimator
for a multivariate probit model. Journal of the American Statistical Association 87 805–811.
L IU , J. S., W ONG , W. H. and KONG , A. (1994). Covariance structure of the Gibbs sampler with applications to the comparisons of estimators and augmentation schemes. Biometrika 81 27–40.
M ARCHEV, D. and H OBERT, J. P. (2004). Geometric ergodicity of van Dyk and Meng’s algorithm for the
multivariate student’s t model. Journal of the American Statistical Association 99 228–238.
M EYN , S. P. and T WEEDIE , R. L. (2012). Markov Chains and Stochastic Stability. 2nd ed. Cambridge
University Press.
R AJARATNAM , B. and S PARKS , D. (2015). MCMC-based inference in the era of big data: A fundamental
analysis of the convergence complexity of high-dimensional chains. arXiv:1508.00947 .
ROBERTS , G. O. and ROSENTHAL , J. S. (1997). Geometric ergodicity and hybrid Markov chains. Electronic Communications in Probability 2 13–25.
ROBERTS , G. O. and ROSENTHAL , J. S. (1998). Markov chain Monte Carlo: Some practical implications
of theoretical results (with discussion). Canadian Journal of Statistics 26 5–31.
ROBERTS , G. O. and ROSENTHAL , J. S. (2001). Markov chains and de-initializing processes. Scandinavian
Journal of Statistics 28 489–504.
ROBERTS , G. O. and T WEEDIE , R. L. (2001). Geometric L2 and L1 convergence are equivalent for reversible Markov chains. Journal of Applied Probability 38A 37–41.
ROSENTHAL , J. S. (1995). Minorization conditions and convergence rates for Markov chain Monte Carlo.
Journal of the American Statistical Association 90 558–566.
ROY, V. and H OBERT, J. P. (2007). Convergence rates and asymptotic standard errors for Markov chain
Monte Carlo algorithms for Bayesian probit regression. Journal of the Royal Statistical Society, Series B
69 607–623.
ROY, V. and H OBERT, J. P. (2010). On Monte Carlo methods for Bayesian multivariate regression models
with heavy-tailed errors. Journal of Multivariate Analysis 101 1190–1202.
RUDIN , W. (1976). Principles of Mathematical Analysis. 3rd ed. McGraw-Hill Publishing Co.
40
S INCLAIR , A. and J ERRUM , M. (1989). Approximate counting, uniform generation and rapidly mixing
Markov chains. Information and Computation 82 93–133.
VATS , D. (2017). Geometric ergodicity of Gibbs samplers in Bayesian penalized regression models. Electronic Journal of Statistics 11 4033–4064.
YANG , J. and ROSENTHAL , J. S. (2017). Complexity results for MCMC derived from quantitative bounds.
arXiv:1708.00829 .
YANG , Y., WAINWRIGHT, M. J. and J ORDAN , M. I. (2016). On the computational complexity of highdimensional Bayesian variable selection. The Annals of Statistics 44 2497–2532.
41
| 10math.ST
|
Intelligent Computer Numerical Control Unit for Machine Tools
Joze BALIC
University of Maribor, Faculty of Mechanical Engineering
Laboratory for Intelligent Manufacturing
Post: SI-2000, City: Maribor Country: Slovenia
E-mail: [email protected] http://www.uni-mb.si
Abstract
The paper describes a new CNC control unit for machining centres with
learning ability and automatic intelligent generating of NC programs on the bases of a
neural network, which is built-in into a CNC unit as special device. The device
performs intelligent and completely automatically the NC part programs only on the
bases of 2D, 2,5D or 3D computer model of prismatic part. Intervention of the operator
is not needed. The neural network for milling, drilling, reaming, threading and
operations alike has learned to generate NC programs in the learning module, which is
a part of intelligent CAD/CAM system.
Keywords: Intelligent CNC & CAD/CAM, Neural Networks, Machining Centres
1. Description of the problem
Figure 1 shows the conventional CNC unit, connected with the machining centre [1]. The
input of data respectively NC program of the part is usually done through DNC (Direct
Numerical Control) connection, with punched tape through the tape reader, or manual,
through interface for manual input. The manual input is used only for corrections of NC
program and for changing technological parameters.
DNC
Manual input
Interface
Reader
Processor
Decoding
CNC machine tool
Position
memory
X, Y, Z
Function
memory
M, S, T
Interpolation
program
Function
program
Comparison
Position
measuring
Interface
Position
meter
Tools
W
Motor
M
Position loop
Amplifier
Geometric data
Conventional CNC unit
Figure 1: The conventional CNC unit for machine tool
Extended and advanced version of this paper appears in Neural-Network-Based Numerical Control for
Milling Machine, Journal of Intelligent and Robotic System, Volume 40, Issue 4, Aug 2004; Pages: 343358
1
For each individual part the new NC part program must be set up, as the control
does not remember already performed operations and can not automatically change
program parameters (except some technological parameters, as are cutting conditions,
corrections of lengths of tools, displacement of offset and zero points).
Also the use of modern CAD/CAM systems does not solve this problem. These
systems enable that a new NC part program can be performed faster and more reliable.
Some systems enable also saving of certain processing strategies, but the intervention
of skilled NC programmer is still necessary. The programs performed on such way are
not possible to use directly in CNC control of the machine tool, however they must be
adapted in so-called post-processor phase. The task of this phase is that generally valid
file of the tool path (CLDATA) is modified for each individual machine tool
respectively CNC control. Each NC part program and each change must go over this
post-processor phase.
2. State of the art and existing solutions
There are only some solutions known from the literature, which describe the
broader field of research presenting in this paper.
Literature [2] describes the NC control unit with integrated function of learning.
The NC control unit performs the teaching NC part program, which is compared with
the inserted NC part program and performs then the resulting NC part program. On this
way the operator of the machine tool can choose the “teaching mode” and changes the
actual NC part program according to the suggestion from teaching program. The
solution requires the intervention of the skilled worker during the NC programming.
Literature [3] describes a device for generating the tool path on NC machine
tools and adequate NC control. The device at first recognizes the geometric feature
characteristics of CAD model of the part and on the base of preserved processing
procedures (machining cycles, sub-programs) chooses the most suitable tool path. The
device can choose only that machining procedures, which have been previously defined
as typical processing for particularly sub-programs.
Literature [4] describes the dialog orientated program system for programming
of CNC machine tools. Hereby it is possible to choose interactively different control
programs and procedures, which are then automatically composed into a NC part
program. The intervention of the operator, programmer is necessary.
Literature [5] describes a learning method of a purpose made device. For this
reason a special man-machine interface, which enables a dialog with the user and
learning, is built-in into the control unit of the machine.
Literature [6] describes a device and a method for generating of NC programs.
A special device saves the data about parts, belonging coordinates, characteristically
junctions and time of assembly for single electronic components. The solution enables
shortening of the time for the composition of NC programs and reduction of mistakes at
preparing of programs.
In the paper [7] an adaptive controller with optimisation function of the milling
process is described. It used neural network to adjust the learning procedure and for online modelling of the milling process. The efficiency of NN based controller is higher
than of the conventional CNC controller.
In the paper [8] the new concept of CNC control unit is proposed. It consists
from feature based NC unit and a basic control unit. The feature-based unit is used as
exchange for geometrical data between basic NC unit and CAD/CAM system. It can be
connected in Internet and used in virtual manufacturing.
2
Common for all described solutions is, that manual intervention of skilled
operator, programmer is still necessary for preparing of NC part program for CNC
machine tools. The systems can’t prepare NC programs for parts, which are not saved
in the databases and can’t choose and use the machining strategies automatically.
3. Intelligent computer numerical control unit
3.1 Description
The new control unit for CNC machining centres - milling, drilling and
operations like drilling is shown on Figure 2. It has the capability of learning and
intelligent, automated generating of NC part programs on the bases of neural network,
which is built-in into a
Conventional
CAD/CAM system
NN
teaching
module
CAD
part
model
NN device
Manual input
Internal interface
Decoding
Position
memory
X, Y, Z
Function
memory
M, S, T
Interpolation
program
Function
program
Comparison
Position
measuring
Interface
W
M
Amplifier
Figure 2: Neural Network Based CNC Machine Tool Control Unit
The neural network for milling, drilling, reaming, threading and operations like
this have learned to generate NC part programs in system for learning, which is not a
constituent part of a new CNC control unit, and works independent.
The new control unit then completely automatically, without any intervention of
the operator, on the bases of a CAD model of a part, generate NC part programs for
different products, which are designed for the production on CNC machining centres
and for which previously NC part programs have not been done.
CNC control unit consists of a modified microcomputer, which includes a
module for decoding and a module for manual input of commands, which are mostly of
technological nature (feed rate, revolution speed, switch on/off of cooling liquid etc.).
Computer includes also the internal interface, which has the function of data
transmission from the NN device into the position memory.
3.2 Description of the NN device
NN device (Figure 3) consist of:
- A module for recognizing of geometric and technological features from CAD
model of the part,
3
- Computer CAD model of the part, which contains geometric and technological
features,
- NN device for milling, which has been learned for NC programming of
milling, drilling and machining operations like drilling and
- NC control program for the part, which has been sent to control device, as a
CAD model of the part.
Conventional
CAD/CAM system
NN teaching
module
CAD
part
model
Identification of geometrical
and technological features
Features based CAD
model of the part
NN milling module
NC control program
Figure 3: Schematic lay-out of NN device
In the phase of learning the NN device is connected to the module for learning
of the neural network, which takes the data from the usual, commercial available
CAD/CAM system, which is used for programming of NC/CNC machine tools. NN
device automatically generates NC programs for CAD models of each individual
product, which it becomes from the commercial CAD/CAM programming system.
CNC control unit after this invention can receive the data for teaching of NN device
from a special module for teaching, which is not the constituent part of the control unit.
The task of the module for teaching is to teach the neural network, which is built-in into
NN device, the principles and the technology of NC programming of all machining
operations on CNC machining centres, mostly milling, drilling and operations alike.
In principle for this purpose different systems of neural network and different,
also commercial developed software, can be used [9,10,11]. If we want to meet special
criteria at machining (costs, time, quality of cutting, tool life, high speed cutting etc.)
special developed neural network must be used.
4. Mode of operation
4.1 Intelligent automatic mode
In the mode of intelligent, completely automated processing the CNC control unit
receives the data about CAD model of the product from the usual, commercially
available CAD/CAM system for programming of CNC machines [13,14,15]. The
model is then transmitted to NN device, which first of all recognizes geometric and
technological features of CAD model of the part [16,17]. After recognizing the
geometric and technological features are classified. On this way the new CAD model of
4
the part is built again, which is now based on this characteristic features. Such model is
then transmitted to NN device for milling, which on the base of learned intelligent
procedures determines the most suitable machining operations and cutting parameters
(cutting speed, feed-rate and the depth of cutting), with respect to chosen conditions
(machining time, surface quality, machining costs).
The exit from the NN device for milling is NC part program for the processed
part, which includes all geometric data about the mode (linear G01 or circular G02/G03
interpolation, etc.) and coordinates of cutting tool path (for instance milling cutter),
technological data (revolution speed Sxxxx, feed-rate Fxx, depth of cutting) and
auxiliary data (coordinates of reference, zero and starting points, direction of rotation of
the main spindle M02/M03, change of cutting tools M06, etc.).
The data are then transmitted to internal interface, which split the data in NC
program on tool path data (coordinates of movement in axis X, Y, Z and/or rotation A,
B, C around coordinate axis X, Y, Z and on data about functions (M, S, T).
The geometrical data are obtained from the NC part program for each individual
part and are treated in the position control circuit.
4.2 Neural network learning mode
In the learning mode the learned NC programming system, which is based on the
principle of neural network, is put into the NN device, through the system for learning.
The method of teaching the NN device is performed in a special NN device for
teaching. The operation principle of NN device is shown more detailed on Figure 4.
Engineering drawings of
parts 1,2,3,...n
Conventional
CAD/CAM system
Teaching NC program
NN teaching module
Testing
NO
IF
YES
NN device
Figure 4: Procedure of learning and generation of neural network
In the learning phase, the NN device is connected to the teaching module designed for
instructing the neural network (NN). The teaching module takes the data from
conventional, commercially available CAD/CAM system for programming the
NC/CNC machine tools. By means of a conventional CAD/CAM system, the teaching
NC programs are prepared for different parts, defined in engineering drawings module,
5
and are sent to the teaching module. In the decision module, subsequent to the testing
module, the decision is taken on the success of teaching. In case that the decision is NO
the repetition of the teaching process takes place. If on the other hand the NN device
has learned enough the generated neural network is sent to the NN device.
The neural network built-in in the NN device consists of three layers: the input
layer, the hidden layer and the output layer [12]. On the input layer, the X-Y-Z sets of
coordinate points appear, representing the coordinate point values obtained from the
modified CAD model for individual machining operations types. The coordinate point
values are determined according to special procedure. Through the intermediate hidden
layer the input coordinates are transposed into output layer in a form of a set of
coordinate points Xi, Yi, Zi, representing the position values of the tool path for
individual machining operations.
4.3 Machining operations
The control unit can learn how to generate the NC part programs for the
following machining procedures: milling (rough face milling, rough contour milling,
final milling after the contour and in Z-plane, final contour milling, final contour 3D
milling, milling on Z-plane, final contour milling on equidistant and milling of
pockets), drilling (normal drilling, deep drilling and centering), reaming, sinking and
threading.
5. Industrial applicability
The NN device can be built-in into each CNC control unit of the milling
machine as shown in Figure 1. The standard parallel transmission of data is used. If it is
not possible to programming the internal interface the NN device must be connected to
existing DNC interface, which each CNC control possess. The NN device for teaching
is connected to NN device with standard serial interface. The CAD model of the part is
sends to NN device through standard communication interface.
For teaching of NN device through the module for teaching different commercial
available CAD/CAM programming systems can be used as are Unigraphics Solution, IDeas, Catia, HyperMill etc. [13,14,15].
6. Conclusion
The new control unit solves the problem of automatic and intelligent generating
of NC programs for CNC machining centres. It describes the control device of a CNC
machining centre for face milling (rough), contour milling (rough), final milling after
the contour and in Z-plain, final contour 3D milling, contour final milling, milling on
Z-plain, final contour milling on equidistant, milling of pockets, normal drilling, deep
drilling, centering, reaming, sinking and threading with the capability of teaching and
automatic intelligent generating of NC part programs on the bases of neural network,
which is built-in into a special NN device. The neural network for milling, drilling and
operations alike can learn to generate NC part programs in module for learning. The
device then completely automatically, without the intervention of the operator, only on
the bases of 2D, 2,5D or 3D computer model of prismatic part, which has not been
previously programmed, generate NC part programs. In solving the machining
problems, the NN can serve as ideal tool, for helping NC programmer making the right
decision, or at least serve as an orientation tool. Time saving could be also achieved,
because not so many post-machining operations would be necessary.
6
References
1. M. Weck. M., Werkzeugmaschinen Fertigungssystem, Berlin [etc.], Springer,
1998.
2. Kato. K., Momochi. T., Numerical controller for machining tool with learning
function - combines learning program with entered program to produce
resulting process, Patent DE4011591 (JP19890098177), 1989.
3. Kadono. M., Tool path data generation apparatus for NC machine tool and
numerical controller provided with it, Patent Nr. US2001/0000805 A1, 2001.
4. Waldenmaier. P., Wirth. H., Monz. J., Dialogue orientated programming system
for a CNC machine tool, Patent EP0726509A1, 1996.
5. Kamioka T., Mochizuki. H., Learning promotion method on tool and learning
promotion type machine, Patent JP2001034155, 2001.
6. Nemoto. K., Kyoichi. M., Yamaguchi. M., Sugimoto. H., Hasegawa. H., NC
data generation device and its method, Patent JP11242510, 1999.
7. Liu. Y., Wang. C., Neural network adaptive control and optimisation in the
milling process, Int. J. Adv. Manuf. Technol., 15 (1999), 791-795.
8. Chang. I., Deng. J., Chan. S., A next generation machining system based on NC
feature unit and real-time tool path generation, Int. J. Adv. Manuf. Technol., 16
(2000) 889-901.
9. NeuroSolutions, NeuroSolutions for Excel, Custom Solution Wizard, Neuro
Dimension, Inc., Gainesville, FL, URL: http://www.nd.com/, 2000.
10. aiNet V1.25 Neural Network Application, Celje, Slovenia, Europe, 1997
11. NeuroShell Predictor, Ward Systems Group, Inc. Frederick, MD
http://www.wardsystems.com/products.asp?p=predictor, 2000.
12. Balic. J., Korosec. M., Intelligent tool path generation for milling of free
surfaces using neural networks, Int. j. mach. tools manuf., [Print ed.], 10 (2002),
1171-1179.
13. Unigraphics, http://www.eds.com/products/plm/unigraphics, 2000.
14. Parametric Technology and Corporation, Pro/MFG&Pro/NC-CHECK User's
Guide Release 2000i, http://www.ptc.com, 2000.
15. HyperMill, OPEN MIND Technologies AG, http://www.openmind.de/, 2003.
16. Brezocnik. M., Pahole. I., Balic. J, Feature recognition from boundary model of
a part (intelligent CAD-CAP interface), International Conference Design to
Manufacture in Modern Industry, Bled, Slovenia, Proceedings. Maribor, Faculty
of Mechanical Engineering, 1995, 395-404.
17. Balic. J., Brezocnik. M., A model of intelligent CAD/CAM interface in CIM,
Proceedings of the 13th International conference on production research,
Jerusalem, Israel, Freund Publishing House, Ltd, London, 1995, 29-31, 1995.
7
| 5cs.CE
|
arXiv:1711.08228v1 [cs.AI] 22 Nov 2017
November 2016
An influence-based fast preceding
questionnaire model for elderly
assessments
Weiping Li a,3 , Jingbo Zhang a,4 , Zhonghai Wu a,5 and
Wei Tan b,6
a School of Software and Microelectronics, Peking University, Beijing, China
b IBM T. J. Watson Research Center, Yorktown Heights, NY, USA
Tong Mo
a,1 ,
Rong Zhang
a,2 ,
AbstractTo improve the efficiency of elderly assessments, an influence-based fast
preceding questionnaire model (FPQM) is proposed. Compared with traditional
assessments, the FPQM optimizes questionnaires by reordering their attributes. The
values of low-ranking attributes can be predicted by the values of the high-ranking
attributes. Therefore, the number of attributes can be reduced without redesigning
the questionnaires. A new function for calculating the influence of the attributes
is proposed based on probability theory. Reordering and reducing algorithms are
given based on the attributes’ influences. The model is verified through a practical
application. The practice in an elderly-care company shows that the FPQM can
reduce the number of attributes by 90.56% with a prediction accuracy of 98.39%.
Compared with other methods, such as the Expert Knowledge, Rough Set and C4.5
methods, the FPQM achieves the best performance. In addition, the FPQM can also
be applied to other questionnaires.
Keywords. Questionnaire, reorder, reduce, fast preceding questionnaire model,
elderly assessment
1. Introduction
Questionnaires have been widely used in various fields, including elderly assessments.
Several questionnaires have been developed and are currently in extensive use to assess
health-related quality of life (HRQOL) [20]. Aging is an increasingly serious social phenomenon in China, and there is a strong need for care services. Assessments of the el1 Corresponding
Author: Tong Mo, School of Software and Microelectronics, Peking University, Beijing
100086, China; E-mail: [email protected]
2 Rong Zhang, School of Software and Microelectronics, Peking University, Beijing 100086, China; E-mail:
[email protected]
3 Weiping Li, School of Software and Microelectronics, Peking University, Beijing 100086, China; E-mail:
[email protected]
4 Jingbo Zhang, School of Software and Microelectronics, Peking University, Beijing 100086, China; E-mail:
[email protected]
5 Zhonghai Wu, School of Software and Microelectronics, Peking University, Beijing 100086, China; E-mail:
[email protected]
6 Wei Tan, IBM T. J. Watson Research Center, Yorktown Heights, NY, USA; E-mail: [email protected]
T. Mo et al. / An influence-based fast preceding questionnaire model
derly are essential for providing personalized services. Existing assessment methods are
usually based on the Barthel Index [18] and the national industry standard for the ability
assessment of elderly adults [28]. Many investigation attributes are needed to systematically obtain information about the elderly. The elderly are asked about multiple attributes
in succession. These assessment methods are inefficient, and the order of the attributes
is not reasonable. When there is a relationship between attributes, some unknown attributes can be predicted by known attributes, and a more reasonable order should be
determined [8, 9, 19].
Classical Test Theory (CTT), Rasch Analysis (RA), decision rule, and experts
[4, 10, 15, 21, 22, 25] have been applied to reduce the length of health questionnaires.
However, actually, these removed attributes have provided additional information. A reasonable order of these attributes of the questionnaires can also be considered. Correlation, multiple regression, factor analysis, cluster analysis and structural equation modelling, and hierarchical multiple regression [2, 7, 24, 27] can be used to determine the
relationships among the attributes of health questionnaires. Certain attributes can indeed
be predicted by other attributes using hierarchical logistic regression, correlation analysis, and binary logistic stepwise regression [1, 3, 6, 11, 14]. However, only one attribute,
not multiple attributes simultaneously, could be predicted in one study. Meanwhile, the
involved attributes in each study are incomplete.
A solid mathematical definition of the question is given. The fast preceding questionnaire model (FPQM) is proposed to solve the problem in five steps. First, the influence of one attribute on all other attributes is defined and calculated. Second, we traverse
every investigation attribute and chooses the attribute with the largest influence as the
best attribute to split. Third, we create the FPQM with the best attribute. We traverse
every value of the attribute, and the sub-dataset corresponding to the value can be used
to obtain the sub-model recursively. Then, the sub-model is attached to the full model,
and the full model is obtained when the recursion ends. Fourth, the created FPQM is
used for the real investigation. The value is directly asked for at the beginning of the
real investigation because there is no prior information about the respondent. Certain investigation attributes can be inferred after sufficient information has been accumulated.
At that time, the confidence level is greater than the given threshold, which means that
questions about the attribute do not need to be asked. Fifth, we calculate the evaluation
metrics and evaluate the model FPQM.
This paper is organized as follows. Section 2 reviews related work. The fast preceding questionnaire model (FPQM) is introduced in Section 3. First, a solid mathematical
definition of the question is given. Then, an influence calculation formula, the best attribute to split choosing algorithm (BASCA), the fast preceding questionnaire model creating algorithm (FPQMCA), the model used for real investigation algorithm (MURIA),
and the model evaluation algorithm (MEA) are presented. Section 4 shows the experimental results, therein presenting the experimental data;, evaluation metrics; the overall
results of the FPQM; the comparison experiment with Expert Knowledge, Rough Set,
and C4.5; and the factor analysis, which includes the number of elderly, number of investigation attributes, and threshold. Section 5 concludes the paper.
T. Mo et al. / An influence-based fast preceding questionnaire model
2. Related work
2.1. Attribute reduction
Luis Prieto [22] presents a parallel reduction in a 38-attribute questionnaire, the Nottingham Health Profile (NHP), to empirically compare Classical Test Theory (CTT) and
Rasch Analysis (RA) results. The CTT results in 20 attributes (4 dimensions), whereas
RA results in 22 attributes (2 dimensions). Moreover, the attribute-total correlation
ranges from 0.45-0.75 for NHP20 and from 0.46-0.68 for NHP22, while the reliability
ranges from 0.82-0.93 and from 0.87-94, respectively.
Ephrem Fernandez [10] reduces and reorganizes the McGill Pain Questionnaire
(MPQ) using a 3-step decision rule for affective and evaluative descriptors of Pain. With
a minimum absolute frequency of 17 and a minimum relative frequency of 1/2 as the
threshold values, the words of the MPQ are reduced from 78 to less than 20 on average. This reduction leads to a negligible loss of information transmitted. Moreover, Wasuwat Kitisomprayoonkul [15] develops the Thai Short-Form McGill Pain Questionnaire
(Th-SFMPQ).
RC Rosen [25] develops an abridged five-attribute version (IIEF-5) of the 15attribute International Index of Erectile Function (IIEF) to diagnose the presence and
severity of erectile dysfunction (ED). The five attributes are selected based on the ability
to identify the presence or absence of ED and on adherence to the National Institute of
Health’s definition of ED. The IIEF-5 possesses favorable properties for detecting the
presence and severity of ED.
X. Badia [4] achieves a qualitative and quantitative reduction in the 179 expressions of the bone metastasis quality of life questionnaire (BOMET-QOL) with respect
to clarity, frequency and importance with 15 experts. This phase, which is performed in
two steps, results in the 35-attribute version of the BOMET-QOL. The initial reduction
yields a 25-attribute questionnaire via factorial analysis. Similarly, the BOMET-QOL-25
is reduced to an integrated version of 10 attributes through a sample of 263 oncology
patients. The BOMET-QOL is an accurate, reliable and precise 10-attribute instrument
for assessing HRQOL.
Tamar E.C. Nijsten [21] tests and reduces Skindex-29 to Skindex-17 using Rasch
Analysis. The Rasch Analysis of the combined emotion and social functioning subscale
of Skindex-29 results in a 12-attribute psychosocial subscale. A total of five of the seven
attributes are retained in a symptom subscale. Classical psychometric properties, such
as the response distribution, attribute–rest correlation, attribute complexity, and internal
consistency, of the two subscales of Skindex-17 are at least adequate. Skindex-17 is a
Rasch-reduced version of Skindex-29, with two independent scores that can be used for
the measurement of health-related quality of life (HRQOL) for dermatological patients.
[4, 10, 15, 21, 22, 25] remove some attributes directly and develop qualitative and
quantitative reductions in questionnaires about health using Classical Test Theory (CTT),
Rasch Analysis (RA), decision rules, or experts. These questionnaires include the Nottingham Health Profile (NHP), McGill Pain Questionnaire (MPQ), 15-attribute International Index of Erectile Function (IIEF), bone metastasis quality of life questionnaire
(BOMET-QOL), and Skindex-29. However, these removed attributes can provide additional information, and their values can be predicted by the remaining attributes with reduction methods. Meanwhile, a more reasonable order of these attributes is not considered.
T. Mo et al. / An influence-based fast preceding questionnaire model
2.2. Relationships among attributes
Alexandra-Lelia Dima [7] studies the interrelations between acceptance, emotions, illness perceptions and health status. The confirmatory analysis (employing a variety of statistical procedures, from correlation to multiple regression, factor analysis, cluster analysis and structural equation modelling) largely confirms the expected relations within
and between domains and is also informative regarding the most suitable data reduction methods. An additional exploratory analysis focuses on identifying the comparative characteristics of acceptance, emotions, and illness perceptions in predicting health
status metrics.
Arnow, Bruce A [2] provides estimates of the prevalence and strength of association between major depression and chronic pain in a primary care population and examines the clinical burden associated with the two conditions alone and together. Data are
collected by questionnaires assessing major depressive disorder (MDD), chronic pain,
pain-related disability, somatic symptom severity, panic disorder, other anxiety, probable alcohol abuse, and health-related quality of life (HRQL). The instruments include
the Patient Health Questionnaire, SF-8, and the Graded Chronic Pain Questionnaire. The
conclusions are that chronic pain is common among those with MDD, and Comorbid
MDD and disabling chronic pain are associated with greater clinical burden than is MDD
alone.
A. Elizabeth Rippentrop [24] seeks to better understand the relationships among
religion/spirituality and physical health, mental health, and pain in 122 patients with
chronic musculoskeletal pain. Hierarchical multiple regression analyzes reveal significant associations between components of religion/spirituality and physical and mental
health. Forgiveness, negative religious coping, daily spiritual experiences, religious support, and self-rankings of religious/spiritual intensity significantly predict mental health
status. Religion/spirituality is unrelated to pain intensity and life interference due to pain.
Religion/spirituality may have both costs and benefits for the health of those with chronic
pain.
Susan W. Vines [27] determines the relationships between pain perceptions, immune
function, depression and health behaviors and examines the effects of chronic pain on
immune function using depression and health behaviors as covariates. Pain perceptions
show positive significant correlations with depression (P = 0.01) and total percent of NK
cells (P = 0.04). Depression and health behaviors are negatively correlated (P = 0.01).
Positive associations are observed for depression and 2 PHA mitogen levels (P<0.05).
The immune function of patients with chronic pain is significantly higher than in the nopain comparison group. Pain perceptions may have a deleterious effect on enumerative
NK cell measures and depression levels.
Most of the attributes mentioned in [2, 7, 24, 27] are included in Table 4, such as
acceptance, emotions, illness perceptions, and health status; depression, chronic pain,
and clinical burden; religion/spirituality and physical health, mental health, and pain; and
pain perceptions, immune function, depression and health behaviors. The only two differences between the investigation attributes in Table 4 and the attributes mentioned in
the literature are the expressions. The attributes mentioned in the literature are more conceptual. Applied methods include correlation, multiple regression, factor analysis, cluster analysis, structural equation modelling, and hierarchical multiple regression. The literature proves that relationships among these attributes do exist. However, the attributes
T. Mo et al. / An influence-based fast preceding questionnaire model
covered by the relationships in each study are incomplete, and the relationships have not
been well utilized to provide results of interest such as in prediction.
2.3. Prediction
Kersh, BC [14] uses psychosocial and health status variables independently to predict
health care seeking for fibromyalgia. Subjects are administered 14 measures, which produce six domains of variables: background demographics and pain duration; psychiatric
morbidity; and personality, environmental, cognitive, and health status factors. These
domains are input into 4 different hierarchical logistic regression analyzes to predict
the status as patient or non-patient. The full regression model is statistically significant
(P<0.0001) and correctly identifies 90.7% of the subjects, with a sensitivity of 92.4%
and a specificity of 87.2%.
Maki Aoyama [1] uses physical and functional factors in activities of daily living
to predict falls in community-dwelling older women. Correlation analysis investigating
associations among the scores of assessment scales and actual measurements of muscle strength and balance shows that there are significant correlations between handgrip
strength and the Falls Efficacy Scale, Functional Reach test, Timed Up and Go test, Berg
Balance Scale, Motor Fitness Scale, and Motor Functional Independence Measure in
fallers and non-fallers. A binary logistic stepwise regression analysis reveals that only
an inability of “being able to go up and down the staircase” in the Motor Fitness Scale
remains a significant variable to predict falls.
Aydeniz, Ali [3] predicts falls in the elderly with physical, functional and sociocultural parameters. Falls are common in patients with weakness, fatigue, dizziness, and
swelling in the legs and in subjects with appetite loss. Fallers have lower functional status than do non-fallers (p=0.028). In addition, fallers have more depressive symptoms
than do non-fallers (p=0.019). Quality of life (NHP), especially physical activity, energy
level and emotional reaction, subgroups are different (p=0.016, 0.015, and 0.005, respectively). Disability and mental status are similar in groups (p=0.006). Musculoskeletal
problems, functional status and social status might be contributors to falls.
Jennifer L. Gatz [11] uses depressive symptoms to predict Alzheimer’s disease and
dementia. The Total Center for Epidemiologic Studies Depression (CES-D) score is a
significant predictor of AD and dementia when categorized as a dichotomous variable
according to the cutoff scores of 16 and 17; a CES-D cutoff of 21 is a significant predictor of AD and a marginally significant predictor of dementia. When analyzed as a continuous variable, the CES-D score is marginally predictive of AD and dementia. Neither
participant-reported history of depression nor participant-reported duration of depression
is significant in predicting AD or dementia.
Madeline Cruice [6] predicts social participation in older adults with personal factors, communication and vision. Assessments are individually conducted in a face-toface interview situation with the primary researcher, who is a speech pathologist. Social
participation is shown to be associated with vision, communication activities, age, education and emotional health. Naming and hearing impairments are not reliable predictors of social participation. It is concluded that professionals interested in maintaining
and improving the social participation of older people should strongly consider these
predictors in community-directed interventions.
Most of the attributes mentioned in [1,3,6,11,14] are also included in Table 4 such as
psychosocial and health status variables and health care; physical, functional and socio-
T. Mo et al. / An influence-based fast preceding questionnaire model
cultural parameters and falls; depressive symptoms, Alzheimer’s disease and dementia;
and personal factors, communication and vision, and social participation. Health care,
falls, Alzheimer’s disease, dementia, and social participation are predicted using hierarchical logistic regression, correlation analysis, and binary logistic stepwise regression,
respectively. The literature here is illustrative of the fact that certain attributes can indeed
be predicted by other attributes. However, one study only predicts one attribute, not multiple attributes simultaneously. Meanwhile, the involved attributes in each study are also
incomplete. A sufficient prediction between complete attributes can be studied.
The relation and motivation between all the related works and this research are emphasized and explained here. [4, 10, 15, 21, 22, 25] show that certain attributes in Table 4
can be reduced directly. The information of these attributes is redundant and contained
in other attributes. [2, 7, 24, 27] provide further evidence that there is an inherent relationship between these attributes. [1, 3, 6, 11, 14] further show that certain attributes can
be predicted by other attributes because of the underlying relationship. All the related
works form the foundation of this research. The proposed fast preceding questionnaire
model (FPQM) can achieve state-of-the-art performance only when there is an inherent relationship between attributes. If the relationship does not exist at all, no methods
can predict certain attributes by others. The relationship is the foundation of all possible
methods, including the FPQM. [4,10,15,21,22,25] reduce some attributes directly, while
the FPQM predicts the values of the attributes and preserves these attributes. In addition,
In [1, 3, 6, 11, 14], one study only predicts one attribute, while the FPQM can predict
multiple attributes simultaneously.
3. Fast preceding questionnaire model (FPQM)
A solid mathematical definition of question is given, and the fast preceding questionnaire
model (FPQM) is proposed to solve the problem in five steps.
Step 1: Calculate the influence
We calculate, in order, the confidence level of the attribute by taking a value under
the condition of another attribute taking a value, the influence of the attribute taking a
value on another attribute, the influence of the attribute on another attribute, the influence
of the attribute on all other attributes, and the attribute that has the largest influence on
all other attributes.
Step 2: Choose the best attribute to split
We traverse every investigation attribute, every other attribute, every value of the
attribute, and every value of the other attribute to calculate every influence. Finally, the
influence of the investigation attribute on all other attributes can be calculated. Then, we
logically choose the best attribute that has the largest influence on all other attributes.
Step 3: Create the FPQM
After the best attribute to split is chosen, we traverse every value of the attribute, and
the sub-model can be obtained with sub-dataset corresponding to the value recursively.
Then, we attach the sub-model to the full model, and the full model is obtained when the
recursion ends.
Step 4: Use the FPQM for real investigation
Now, the FPQM can be used to investigate a new respondent. At the beginning
of the real investigation, there is no prior information about the respondent; therefore,
T. Mo et al. / An influence-based fast preceding questionnaire model
we ask for the value directly. After sufficient information has been accumulated, some
investigation attributes can be inferred. If the confidence level is larger than the given
threshold, then the attribute does not need to be asked about.
Step 5: Evaluate the model
After the FPQM is used to investigate the new respondent, the evaluation metrics
can be calculated. The FPQM can be evaluated based on these metrics.
The five steps are also presented in the form of a graph, as shown in Figure 1. Step
1: Calculate the influence with Eq. (1)-(5). Step 2: Choose the best attribute to split
with Algorithm 1 and the calculated influence in Step 1. Step 3: Create the FPQM with
Algorithm 2, and at every recursion step, call Algorithm 1 to choose the best attribute to
split. Step 4: Use the FPQM for the real investigation with Algorithm 3 after the FPQM
is created in Step 3. Step 5: Evaluate the model with Algorithm 4.
Figure 1. Illustration of the five steps.
3.1. Problem definition
Definition 3.1. Let U = {U1 ,U2 , ...,Um } be the collection of all individuals who are
investigated in the training dataset, where Ui (1 6 i 6 m) is the i-th individual being
investigated.
Definition 3.2. Let Ũ = {U˜1 , U˜2 , ..., U˜m̃} be the collection of all individuals in the testing
dataset, where Ũi (1 6 i 6 m̃) is the i-th individual and Ũ ∩U = φ .
Definition 3.3. Let V = {V1 ,V2 , ...,Vn } be the collection of investigation attributes, where
V j is the j-th investigation attribute.
Definition 3.4. Let W = {W1 ,W2 , ...,Wn } be the collection of all possible values on all
investigation attributes, where W j is the collection of all possible values on V j .
Definition 3.5. N = {N1 , N2 , ..., Nn }, where N j is the number of W j . N will be used in
the time complexity analysis of the following four algorithms.
Definition 3.6. Let D = {Di j }(1 6 i 6 m, 1 6 j 6 n) be the matrix of all real values of
all individuals, and D is the training dataset, where Di j is the real value of individual Ui
on the investigation attribute V j and Di j ∈ W j .
T. Mo et al. / An influence-based fast preceding questionnaire model
Definition 3.7. Let R = {Ri j }(1 6 i 6 m̃, 1 6 j 6 n) be the matrix of all real values,
where R is the testing dataset, in which Ri j is the real value of Ũi on V j and Ri j ∈ W j . R
can also be represented as R = {Ri }(1 6 i 6 m̃); Ri = {Ri j }(1 6 j 6 n), where Ri is the
vector of real values of Ũi .
Definition 3.8. Let R0 = {R0i j }(1 6 i 6 m̃, 1 6 j 6 n) be the matrix of all final values,
where R0i j is the final value of Ũi on V j and R0i j ∈ W j . R0 can also be represented as
R0 = {R0i j }(1 6 i 6 m̃); R0i = {R0i j }(1 6 j 6 n), where R0i is the vector of the final values
of Ũi .
Definition 3.9. Let K = {Ki j }(1 6 i 6 m̃, 1 6 j 6 n) be the matrix of indication values
showing whether the final value R0i j is equal to the real value Ri j . Ki j ∈ {0, 1}, where
Ki j = 1 when R0i j = Ri j and Ki j = 0 when R0i j 6= Ri j .
Definition 3.10. Let P = {Pi j }(1 6 i 6 m̃, 1 6 j 6 n) be the matrix of the confidence
levels of Ũi taking the final values R0i j . Pi j ∈ [0, 1], where Pi j is the confidence level of Ũi
taking R0i j on V j . P can also be represented as P = {Pi }(1 6 i 6 m̃); Pi = {Pi j }(1 6 j 6 n),
where Pi is the vector of the confidence levels of Ũi .
Definition 3.11. Let I = {Ii j }(1 6 i 6 m̃, 1 6 j 6 n) be the matrix of indication values
showing whether the value R0i j is predicted from other already known attributes. Ii j ∈
{0, 1}, where Ii j = 1 denotes that R0i j is predicted and Ii j = 0 denotes that R0i j = Ri j by
asking Ũi on V j directly. Note that Ii j = 1 when Pi j > σ , and Ii j = 0 when Pi j < σ , where
σ is the given threshold. I can also be represented as I = {Ii }(1 6 i 6 m̃); Ii = {Ii j }(1 6
j 6 n), where Ii is the vector of the indication values of Ũi .
Definition 3.12. Let O = {O1 , O2 , ..., Om̃ } be the collection of all reasonable orders in
which individuals should be investigated, where Oi is a reasonable order in which Ũi
should be investigated. Oi = { j1 , j2 , ..., jn } is a single substitution of {1, 2, ..., n}. Oi1 is
different from Oi2 (which it most likely is) when i1 is different from i2 .
Definition 3.13. Let M be the fast preceding questionnaire model. M is a tree structure,
and M will determine R0 , I, and O.
Definition 3.14. Let C be the collection of all appearing confidence levels when creating
M, where C will determine P.
Definition 3.15. Let S = {U,V,W, D, R, R0 , K, P, I, O, M,C} be the system of the fast preceding questionnaire model.
Definition 3.16. Let S̄ be the space of all possible questionnaire models.
Definition 3.17. Average accuracy rate (AAR): AAR =
1
m̃
∑m̃
i=1
∑nj=1 Ki j ∗Ii j
.
∑nj=1 Ii j
AAR describes
how accurate the model can be. AAR can also be represented as AAR =
ARi =
∑nj=1 Ki j ∗Ii j
∑nj=1 Ii j
1
m̃
∑m̃
i=1 ARi ;
, where ARi is the accuracy rate of Ũi .
∑n
Ii j
j=1
Definition 3.18. Average reduction rate (ARR): ARR = m̃1 ∑m̃
. ARR describes
i=1
n
how well the model can accelerate the questionnaire. ARR can also be represented as
ARR =
1
m̃
∑m̃
i=1 RRi ; RRi =
∑nj=1 Ii j
,
n
where RRi is the reduction rate Ũi .
T. Mo et al. / An influence-based fast preceding questionnaire model
2
(β +1)∗ARi ∗RRi
Definition 3.19. Average Fβ -Measure: AFβ = m̃1 ∑m̃
i=1 β 2 ∗ARi +RRi , where β is a given
parameter. AFβ describes a balance between AAR and ARR.
Definition 3.20. The problem is defined as the following.
max AFβ sub ject to S ∈ S̄
3.2. Influence calculation formula
To create the model from the training dataset D, the influence of one investigation attribute on all others should be calculated. The influence calculation formula is given in
Definition 3.21-3.25 when the depth of the created model reaches t. The influence of the
investigation attribute depends on the influence of the values.
Definition 3.21. The confidence level of the investigation attribute V jtk taking the value
2
vtjk under the condition of the investigation attribute V jtk taking the value vtjk when the
2
1
1
previous k-1 layer values V 1 ,V 2 , ...,V t−1 are already known.
P(V jtk = vtjk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 )
2
=
2
1
1
N(V jt =vtj ,V jt =v jk ;V 1 ,V 2 ,...,V t−1 )
k2
k2
k1
1
N(V jt =vtj ;V 1 ,V 2 ,...,V t−1 )
k1
(1)
k1
where V jtk ∈ V , V jtk ∈ V , vtjk ∈ W jk , and vtjk ∈ W jk . V 1 is the 1st-layer value of
1
1
1
2
2
2
all investigation attributes V , V 2 is the 2nd layer value of V , etc. N(V jtk = vtjk ,V jtk =
2
2
1
v jk ;V 1 ,V 2 , ...,V t−1 ) is the number of individuals in U who take the value vtjk , vtjk on
1
1
2
the investigation attributes V jtk , V jtk , respectively. N(V jtk = vtjk ;V 1 ,V 2 , ...,V t−1 ) is the
1
1
2
1
number of individuals in U who take the value vtjk on V jtk .
Definition 3.22. The influence of
values V 1 ,V 2 , ...,V t−1
V jtk
1
taking
1
1
vtjk
1
V jtk
2
on
when the previous k-1 layer
are already known.
INF(V jtk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 )
2
=
1
1
∑
vtj ∈W jk
k
P(V jtk = vtjk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 )2
2
2
(2)
1
1
2
2
where P(V jtk = vtjk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 ) is defined in Definition 3.21.
2
2
1
1
Notice that INF(V jtk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 ) is not defined as INF(V jtk |V jtk =
1
1
2
v jk ;V 1 ,V 2 , ...,V t−1 ) = ∑vtj
1
∑vtj
k2
∈W jk
2
k2
2
∈W jk
2
2
2
1
1
P(V jtk = vtjk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 ) = 1 is always true.
2
2
1
1
1
P(V jtk = vtjk |V jtk = v jk ;V 1 ,V 2 , ...,V t−1 ) because
T. Mo et al. / An influence-based fast preceding questionnaire model
Definition 3.23. The influences of V jtk on V jtk when the previous k-1 layer values
2
1
V 1 ,V 2 , ...,V t−1 are already known.
INF(V jtk |V jtk ;V 1 ,V 2 , ...,V t−1 )
2
1
P(V jtk = vtjk ;V 1 ,V 2 , ...,V t−1 )∗
= ∑
1
vtj ∈W jk
k1
1
1
INF(V jtk |V jtk = vtjk ;V 1 ,V 2 , ...,V t−1 )
2
1
(3)
1
where P(V jtk = vtjk ;V 1 ,V 2 , ...,V t−1 ) is the confidence level of V jtk taking vtjk .
1
1
1
1
Definition 3.24. The influence of V jtk on all other investigation attributes ∆ when the
2
previous k-1 layer values V 1 ,V 2 , ...,V t−1 are already known.
INF(∆|V jtk ;V 1 ,V 2 , ...,V t−1 )
1
=
∑
V jt ∈V,k2 6=k1
k2
INF(V jtk |V jtk ;V 1 ,V 2 , ...,V t−1 )2
2
1
(4)
where ∆ = V − {V jtk }.
1
Definition 3.25. The investigation attribute V∗t that has the largest influence on all other
investigation attributes ∆.
V∗t = arg max INF(∆|V jtk ;V 1 ,V 2 , ...,V t−1 )
V jt ∈V
1
(5)
k1
3.3. Best attribute to split choosing algorithm (BASCA)
When creating the FPQM, it is necessary to choose the best attribute to split. When the
depth of the created model reaches t, we traverse every investigation attribute; then, we
traverse every other investigation attribute and calculate the influence of the investigation
attribute on all other attributes. Lines 5, 9, 11, 13, and 14 are calculated with Eqs (1)-(5).
Then, we logically choose V∗t as the best attribute that has the largest influence on ∆. The
pseudocode is shown in Algorithm 1 when the depth of the created model reaches t.
V∗t = BASCA(D,V ). The BASCA returns the investigation attribute V∗t that has the
largest influence on ∆. C is a global variable, and the BASCA can also be called to obtain
C.
T. Mo et al. / An influence-based fast preceding questionnaire model
Algorithm 1 Best attribute to split choosing algorithm (BASCA)
Input: D, V
Output: V∗t
1:
for (V jtk ∈ V ) do
1
2:
for (V jtk ∈ V & V jtk 6= V jtk ) do
2
3:
1
2
for each value vtjk of V jtk do
1
4:
1
for each value vtjk of V jtk do
2
2
5:
Calculate P(vtjk |vtjk ;V 1 ,V 2 , ...,V t−1 )
6:
Add P(vtjk |vtjk ;V 1 ,V 2 , ...,V t−1 ) to C
2
1
end for
Calculate INF(V jtk |vtjk ;V 1 ,V 2 , ...,V t−1 )
1
2
end for
Calculate INF(V jtk |V jtk ;V 1 ,V 2 , ...,V t−1 )
2
1
end for
Calculate INF(∆|V jtk ;V 1 ,V 2 , ...,V t−1 )
1
end for
Calculate V∗t
return V∗t
2
7:
8:
9:
10:
11:
12:
13:
14:
15:
1
3.4. Fast preceding questionnaire model creating algorithm (FPQMCA)
Now, the FPQM can be created with the above groundwork. After V∗ is chosen as the
best attribute to split, we traverse every value v∗ of V∗ , and the sub-model can be obtained
with the sub-dataset corresponding to v∗ recursively. Then, we attach the sub-model to
the full model. Algorithm 2 is the pseudocode.
L = List(0), and L is initialized with a zero vector. Then, the FPQMCA can be called
to obtain the fast preceding questionnaire model M. M = FPQMCA(D,V, L).
3.5. Model used for real investigation algorithm (MURIA)
With the created FPQM, the new person in the testing dataset can be investigated quickly.
At the beginning of the real investigation, there is no information about the respondent;
therefore, all we can do is ask about the attribute directly. After sufficient information has
been accumulated, some investigation attributes can be inferred; otherwise, we continue
asking about attributes.
Let Ind be the index indicating whether the current investigation attribute is the top
attribute. Ind = 0 indicates that it is the top attribute, and Ind = 1 indicates that it is not.
If the current investigation attribute is the top attribute, there is no information about the
respondent; therefore, the attribute cannot be predicted.
R0 , I, O, and P are global variables that have be defined in Definitions 3.8, 3.11, 3.12,
and 3.10, respectively. The pseudocode is shown in Algorithm 3.
Ind = 0 is the initial value. We traverse every testing dataset in R as Ri , and we call
the algorithm LN = MURIA(Ri ,V, M, σ , Ind) to obtain R0 , I, O, and P.
T. Mo et al. / An influence-based fast preceding questionnaire model
Algorithm 2 Fast preceding questionnaire model creating algorithm (FPQMCA)
Input: D, V , Index list: L, β
Output: Fast preceding questionnaire model: M
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Create a node M
Let N(L(0)) be the number of 0s in L
if N(L(0)) == 1 then
Let V (L(0)) be the investigation attribute whose corresponding index in L is 1
return M as a leaf node labeled with V (L(0))
else
V∗ = BASCA(D,V )
Label node M with V∗
Let L(V −1 (V∗ )) be the corresponding index in L of V∗
L(V −1 (V∗ )) = 1
for each value v∗ of V∗ do
Let Dv∗ be the set of data tuples in D satisfying v∗ on V∗
Mv∗ = FPQMCA(Dv∗ ,V, L)
Attach the node Mv∗ to node M
end for
end if
return M
3.6. Model evaluation algorithm (MEA)
Now, the FPQM should be evaluated to determine its performance. Various evaluation
metrics are calculated by the model evaluation algorithm (MEA). First, K can be calculated with R and R0 . Then, AAR, AAR, and AFβ can be obtained with K and I. A larger
AFβ indicates a better FPQM. Algorithm 4 is the pseudocode.
When β is given, (AAR, ARR, AFβ ) = MEA(R, R0 , I, β ). AAR, ARR, AFβ can be obtained by calling the MEA.
3.7. An example
Here is an example to illustrate the Definitions 3.1-3.25 and Algorithms 1-4.
Table 1. Example Training Dataset
Investigation attributes
Education
Income
Social Skills
Work Ability
Communication
U1
U2
U3
U4
0
1
1
1
1
2
0
0
0
0
1
1
1
0
0
1
1
1
1
0
U = {U1 ,U2 ,U3 ,U4 ,U5 }
V = {V1 ,V2 ,V3 ,V4 ,V5 } = {Education, Income, Social Skills,Work Ability,Communication}
W = {W1 ,W2 ,W3 ,W4 ,W5 } = {{0, 1}, {0, 1, 2}, {0, 1}, {0, 1}, {0, 1}}
N = {N1 , N2 , N3 , N4 , N5 } = {2, 3, 2, 2, 2}
T. Mo et al. / An influence-based fast preceding questionnaire model
Algorithm 3 Model used for real investigation algorithm (MURIA)
Input: Ri , V , M, threshold of confidence level: σ , Index: Ind
Output: Last node: LN
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
Let M(1) be the top investigation attribute index of M. Let V (M(1)) be the top
investigation attribute of M. Let Ri (M(1)) be the real value of Ri on V (M(1)). Let
Ri 0 (M(1)) be the final value of Ri 0 on V(M(1)). Let Ii (M(1)) be the indication value
of Ii on V (M(1)). Let Pi (M(1)) be the confidence level of Pi on V (M(1)). Let Oi ( j1 )
be the current attribute of Oi .
if Ind == 0 then
Ri 0 (M(1)) = Ri (M(1))
Ii (M(1)) = 0
Pi (M(1)) = 1
Ind = 1
end if
Oi ( j1 ) = M(1)
Let Ms be the sub-model of M when Ũi takes Ri 0 (M(1)).
Let Ms (1) be the top investigation attribute index of Ms . Let V (Ms (1)) be the top
investigation attribute of Ms . Let Ri (Ms (1)) be the real value of Ri on V (Ms (1)). Let
Ri 0 (Ms (1)) be the final value of Ri 0 on V (Ms (1)). Let Ii (Ms (1)) be the indication
value of Ii on V (Ms (1)). Let Pi (Ms (1)) be the confidence level of Pi on V (Ms (1)).
Let Oi ( j2 ) be the next attribute of Oi . Let W (Ms (1)) be possible values on Ms (1).
Let C(v(Ms (1)) | Ri 0 (M(1))) be the confidence level of Ms (1) taking v(Ms (1)) under
the condition of M(1) taking Ri 0 (M(1)).
C∗ =
max
C(v(Ms (1)) | Ri 0 (M(1)))
v(Ms (1))∈W (Ms (1))
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
if C∗ > σ then
Ri 0 (Ms (1)) = v(Ms (1))
Ii (Ms (1)) = 1
Pi (Ms (1)) = C∗
Oi ( j2 ) = Ms (1)
else
Ri 0 (Ms (1)) = Ri (Ms (1))
Ii (Ms (1)) = 0
Pi (Ms (1)) = 1
Oi ( j2 ) = Ms (1)
end if
if Ms ! = a tree then
return Ms
else
return MURIA(Ri ,V, Ms , σ , Ind)
end if
01011
1 2 0 0 1
D=
1 0 1 0 1
10110
When t = 1, no values of the investigation attributes are known yet. By Eq. (1),
T. Mo et al. / An influence-based fast preceding questionnaire model
Algorithm 4 Model evaluation algorithm (MEA)
Input: R, R0 , I, β
Output: AAR, ARR, AF β
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
for i ← 1 to m̃ do
for j ← 1 to n do
if Ri j 0 6= Ri j then
Ki j = 0
else
Ki j = 1
end if
end for
end for
Let Sum1, Sum2, Sum3, and Sum4 be temporary variables
Sum1 = 0
Sum2 = 0
Sum3 = 0
for i ← 1 to m̃ do
Sum4 = 0
Sum5 = 0
for j ← 1 to n do
Sum4 = Sum1 + Ki j ∗ Ii j
Sum5 = Sum2 + Ii j
end for
ARi = Sum4/Sum5
RRi = Sum5/n
2
23:
24:
25:
26:
27:
28:
29:
30:
31:
i ∗RRi
Fi = (ββ +1)∗AR
2 ∗AR +RR
i
i
Sum1 = Sum1 + ARi
Sum2 = Sum2 + RRi
Sum3 = Sum3 + Fi
end for
AAR = Sum1/m̃
ARR = Sum2/m̃
AFβ = Sum3/m̃
return AAR, ARR, AFβ
Table 2. Example Testing Dataset
Investigation attributes
Education
Income
Social Skills
Work Ability
Communication
U˜1
U˜2
1
0
1
1
0
1
1
0
0
1
P(V21 = 0|V11 = 0) =
N(V21 = 0,V11 = 0) 0
= =0
1
N(V11 = 0)
(6)
T. Mo et al. / An influence-based fast preceding questionnaire model
P(V21 = 1|V11 = 0) =
N(V21 = 1,V11 = 0) 1
= =1
1
N(V11 = 0)
(7)
P(V21 = 2|V11 = 0) =
N(V21 = 2,V11 = 0) 0
= =0
1
N(V11 = 0)
(8)
By Eq. (2),
P(V21 |V11 = 0) = P(V21 = 0|V11 = 0)2 + P(V21 = 1|V11 = 0)2 + P(V21 = 2|V11 = 0)2
= 02 + 12 + 02 = 1
(9)
Similar to Eqs (6)-(9),
P(V21 |V11 = 1) = P(V21 = 0|V11 = 1)2 + P(V21 = 1|V11 = 1)2 + P(V21 = 1|V11 = 1)2
=
12 5
22
+ 02 +
=
3
3
9
(10)
By Eq. (3),
P(V21 |V11 ) = P(V11 = 0) ∗ P(V21 |V11 = 0) + P(V11 = 1) ∗ P(V21 |V11 = 1)
=
1
3 5 2
∗1+ ∗ =
4
4 9 3
(11)
Similar to Eqs (6)-(11),
P(V31 |V11 ) =
2
3
P(V41 |V11 ) =
2
3
P(V51 |V11 ) =
2
3
(12)
By Eq. (4),
P(∆|V11 ) = P(V21 |V11 ) + P(V31 |V11 ) + P(V41 |V11 ) + P(V51 |V11 )
=
2 2 2 2 8
+ + + =
3 3 3 3 3
(13)
Similar to Eqs (6)-(13),
7
2
5
P(∆|V41 ) =
2
P(∆|V21 ) =
11
4
5
P(∆|V51 ) =
2
P(∆|V31 ) =
(14)
T. Mo et al. / An influence-based fast preceding questionnaire model
By Eq. (5),
V∗1 = arg max INF(∆|V j1k ) = V21
V j1 ∈V
1
(15)
k1
Eqs (6)-(15) also shows how Algorithm 1 performs.
As for Algorithm 2, a node M is created at the beginning. L is initialized as a zero
vector: L = List(0). N(L(0)) = n and not 1, and V∗1 = BASCA(D,V ) with Algorithm
1. We label node M as V∗1 , L(V −1 (V∗1 )) = 1. V∗1 = V21 (Income) and W2 = {0, 1, 2};
therefore,
thereare three branches from the attribute Income. On the branch Income = 0,
1101
Dv ∗ =
, Mv∗ = FPQMCA(Dv∗ ,V, L). Attach the node Mv∗ to node M. When
1110
N(L(0)) == 1, Algorithm 2 terminates, and the fast preceding questionnaire model is
created. Notice that only the first recursion step is shown here to ensure that the example
is not too long.
The created model is shown in Figure 2.
Figure 2. The obtained FPQM for the example training dataset.
11010
.
10110
After the model is created, Algorithm 3 can be used for a real investigation
on
the testing dataset. V∗1 = V21 ; therefore, M(1) = 2 and V (M(1)) = V21 . R1 = 1 1 0 1 0 .
Ind = 0, then R1 0 (M(1)) = R1 0 (2) = R1 (2) = 1, I1 (2) = 0, P1 (2) = 1, Ind = 1. O1 (1) =
2. Ms (1) = 1, V (Ms (1)) = V11 . C∗ =
max
C(v(Ms (1)) | R1 0 (2)) = 1 and
Ũ = {U˜1 , U˜2 }, R =
v(Ms (1))∈W (Ms (1))
v(Ms (1)) = 0. C∗ > σ = 0.8, then R1 0 (1) = 0, I1 (1) = 1, P1 (1) = 1, O1 (2) = 1. Ms = a
tree; therefore,
return MURIA(R1 ,V, Ms , σ , 1). When Algorithm 3 terminates, we obtain
R01 = 0 1 0 1 1 , P1 , I1 , and O1 .
AfterAlgorithm 3 finishes
the
investigating
testingdatasetR, we canobtain
0
1
0
1
1
1
1
1
1
1
10111
21345
R0 =
,P=
,I=
,O=
.
10110
1 1 1 0.5 1
10101
13024
M is the fast preceding questionnaire model, C is the collection of all confidence
levels that appear when creating M. S = {U,V,W, D, R, R0 , K, P, I, O, M,C}, where S̄ is
the space of all possible questionnaire models.
T. Mo et al. / An influence-based fast preceding questionnaire model
AAR, ARR, and AFβ can be calculated by Algorithm 4, which can be applied for
11110
model evaluation. With Lines 1-9, K =
. With Lines 10-30, AAR = 0.7, ARR =
11111
0.75, and AF0.5 = 0.7114 with β = 0.5.
The problem is defined as the following.
max AFβ sub ject to S ∈ S̄
3.8. The comparison with Decision Tree
The fast preceding questionnaire model (FPQM) is similar to Decision Tree to some extent, but they are different models. There are many types of Decision Tree algorithms.
The notable models include ID3, C4.5, C5.0, CART, CHAID, MARS, and Conditional
Inference Trees. ID3, C4.5 and CART are chosen as representative Decision Tree algorithms. The FPQM is compared with these three Decision Tree algorithms in all aspects [5, 13, 26].
3.8.1. The comparison on measures
The algorithms for constructing decision trees choose an attribute at each step that best
splits the dataset. Different Decision Tree algorithms use different metrics for determining the ”best”. ID3 uses information gain, C4.5 uses a gain ratio, and CART uses the
Gini index, whereas the FPQM uses Influence, which is defined in Definitions 3.21-3.25.
3.8.2. The comparison on solved problems
ID3, C4.5, and CART can solve both classification and regression problems. On the other
hand, the FPQM is designed to optimize questionnaires by reordering the attributes and
reducing the number of attributes by predicting the low-ranking attributes. The solved
problems between the FPQM and Decision Tree are entirely different.
3.8.3. The comparison on model forms
The model form of the FPQM looks very similar to that of Decision Tree algorithms;
both are in tree forms, as shown in Figure 3.
3.8.4. The comparison on attribute types
The attribute types that the FPQM and Decision Tree can handle are different. ID3 can
only handle nominal types, C4.5 and CART can handle both nominal and numeric types,
and the FPQM can only handle nominal attribute types.
3.8.5. The comparison on splits
The split methods of the FPQM and Decision Tree are different. ID3 and C4.5 split the
data in multiple ways, CART splits the data as a 2-way split, and the FPQM splits the
data in multiple ways. CART can only create binary trees, whereas ID3, C4.5 and the
FPQM can create general trees.
The comparisons between the FPQM with Decision Tree in all aspects are listed in
Table 3.
T. Mo et al. / An influence-based fast preceding questionnaire model
Table 3. The comparison with Decision Tree
Decision Tree
Aspect
ID3
C4.5
CART
FPQM
Measure
Solved problem
Model form
Attribute type
Split
Information gain
Classification
Tree
Nominal
Multi-way
Gain ratio
Classification
Tree
Nominal/Numeric
Multi-way
Gini index
Classification/Regression
Tree
Nominal/Numeric
2-way
Influence
Reorder & Reduce
Tree
Nominal
Multi-way
3.9. The time complexity of the FPQM
There are many loops and recursive steps in the four algorithms constituting the FPQM;
the time complexity of the FPQM is analyzed below. The derivation process of the time
complexity is described in detail in Appendices.
3.9.1. The time complexity of the BASCA
The time complexity of the BASCA is
TBASCA (n, N̄) = O(n2 N̄ 2 )
(16)
where N̄ is the average of N j in N, N̄ = 1n ∑nj=1 N j , and n is the number of V .
3.9.2. The time complexity of the FPQMCA
The time complexity of the FPQMCA is
TFPQMCA (n, N̄) = O(N̄ n+1 + n2 N̄)
(17)
3.9.3. The time complexity of the MURIA
The time complexity of the MURIA is
TMURIA (n, N̄) = O(nN̄)
(18)
3.9.4. The time complexity of the MEA
The time complexity of the MEA is
TMEA (n, m̃) = O(m̃n)
where m̃ is the number of Ũ.
(19)
T. Mo et al. / An influence-based fast preceding questionnaire model
3.9.5. The time complexity of the FPQM
The time complexity of the FPQM is
TFPQM (n, N̄, m̃) = TBASCA (n, N̄) + TFPQMCA (n, N̄) + TMURIA (n, N̄) + TMEA (n, m̃)
= O(n2 N̄ 2 ) + O(N̄ n+1 + n2 N̄) + O(nN̄) + O(m̃n)
= O(n2 N̄ 2 ) + O(N̄ n+1 ) + O(m̃n)
(20)
4. Experiments
4.1. Experimental data
The experiments are based on actual data from the Lime Family Limited Company (Lime
Family). Lime Family focuses on the pension service field and is the nationwide leading
provider of high-quality home care services. Lime Family originally required 45 investigation attributes to assess a newcomer. The investigation attributes are based on the
Barthel Index [18] and an ability assessment for elderly adults [28]. A total of 45 investigation attributes seems excessive for the elderly individuals who are asked to give the
value of the attributes one by one, and the order of the attributes is not reasonable. Therefore, every assessment requires approximately 15-20 minutes, which is also excessive,
resulting in customer churn in practice.
4.1.1. Noise
There are three main classes of noise: spurious readings, measurement error, and background data. Some values of the Height and Weight attributes are 0, and some values
of the Age attribute are 2015. These are the spurious reading type of noise data; it is easy
for people to see that they are incorrect values. Height, Weight and Age are numeric
attributes, and thus, these noise data are replaced with a sample mean.
4.1.2. Missing values
Certain values of the Religion, Ground Walking, and Up Down Stairs attributes are
missing. Religion is a nominal attribute, and Ground Walking and Up Down Stairs
are ordinal attributes; thus, these missing values are replaced with a sample mode.
After preprocessing, including discretization, handling missing values, addressing
noise data, and textual data processing, all data are converted into nominal data. The
investigation attributes can be observed in Table 4.
The entire experiment is performed using the 64-bit Python language on a MacBook
Pro, with a 2.2 GHz Intel Core i7 CPU and 16 GB of 1600 MHz DDR3 memory.
4.2. Evaluation metrics
To evaluate the performance of the FPQM, the accuracy rate, average accuracy rate, standard deviation of the accuracy rate, reduction rate, average reduction rate, and standard
deviation of the reduction rate are defined as follows.
T. Mo et al. / An influence-based fast preceding questionnaire model
Table 4. Investigation attributes
No.
Attribute Name
No.
Attribute Name
No.
Attribute Name
0
Sex
15
Transfer Bed Chair
30
Financial Affairs Capability
1
Age
16
Ground Walking
31
Cognitive Function
2
Ethnic
17
Up Down Stairs
32
Attacks
3
Education
18
Witted
33
Depressive Symptoms
4
Religion
19
Height
34
Consciousness Level
5
Marital Status
20
Weight
35
Vision
6
Housing Condition
21
Glasses
36
Hearing
7
Income
22
Hearing Aid
37
Communication
8
Eating
23
Go Shopping
38
Life Skills
9
Bathe
24
Outings
39
Work Ability
10
Modify
25
Food Cooking
40
Time Spatial Orientation
11
Clothing
26
Maintain Housework
41
People Orientation
12
Stool Control
27
Washing Clothes
42
Social Skills
13
Pee Control
28
Use Phone Ability
43
Event
14
Toilet
29
Medication
44
Occurrences Times
Accuracy rate (AR):
ARi =
∑nj=1 Ki j ∗Ii j
∑nj=1 Ii j
(21)
ARi is the accuracy rate of individual Ũi .
Average accuracy rate (AAR):
m̃
AAR =
1
m̃
∑
i=1
∑nj=1 Ki j ∗Ii j
∑nj=1 Ii j
(22)
AAR is the average accuracy rate of Ũ and describes how accurate the model can be.
Standard deviation of the accuracy rate (SAR):
s
SAR =
m̃
1
m̃
∑ (ARi − AAR)2
(23)
i=1
SAR is the standard deviation of the accuracy rate and describes the volatility of the
accuracy rate.
Reduction rate (RR):
RRi =
RRi is the reduction rate of Ũi .
∑nj=1 Ii j
n
(24)
T. Mo et al. / An influence-based fast preceding questionnaire model
Average reduction rate (ARR):
m̃
ARR =
1
m̃
∑
∑nj=1 Ii j
n
(25)
i=1
ARR is the average reduction rate of Ũ and describes how the model can accelerate
the questionnaire.
Standard deviation of the reduction rate (SRR):
s
m̃
1
m̃
SRR =
∑ (RRi − ARR)2
(26)
i=1
SRR is the standard deviation of the reduction rate and describes the volatility of the
reduction rate.
Fβ -Measure (F):
Fβi =
(β 2 +1)∗ARi ∗RRi
β 2 ∗ARi +RRi
(27)
Fβi is the Fβ -Measure of Ũi .
Average Fβ -Measure (AF):
m̃
AFβ =
1
m̃
2
i ∗RRi
2 ∗AR +RR
∑ (ββ +1)∗AR
i
i
(28)
i=1
AFβ is the average Fβ -Measure of Ũ and describes a balance between AAR and ARR.
When Ii j = 0 ∀i, j, ARR = 0 and AAR = 1; therefore, AFβ = 0. When Ii j = 1 ∀i, j,
ARR = 1, and both AAR and AFβ will be low. A proper Ii j should be chosen to maximize
AFβ with proper AAR and ARR. AFβ describes a balance between AAR and ARR. β = 0.5
is chosen in this experiment because ARR is as important as AAR, and the goal is to
improve ARR and AAR simultaneously.
Standard deviation of Fβ -Measure (SF):
s
SF =
m̃
1
m̃
∑ (Fβi − AFβ )2
(29)
i=1
SF is the standard deviation of the Fβ -Measure and describes the volatility of the
Fβ -Measure, which is the volatility of the overall performance.
T. Mo et al. / An influence-based fast preceding questionnaire model
Figure 3. A part of the obtained FPQM.
4.3. The overall results
The depth of the obtained FPQM is 45 because there are 45 investigation attributes in
total. There is a corresponding rule with every node in the FPQM, and there are a total
of 6072 rules. A part of the obtained FPQM is shown in Figure 3.
A reasonable order for the investigation attributes can be obtained from the FPQM.
Transfer Bed Chair is the first investigation attribute. When there is a new individual
Ũi of Ũ, if the Transfer Bed Chair value of Ũi equals 1, we enter the corresponding branch of the FPQM, and the second investigation attribute is Outings. This will
continue recursively until the last investigation attribute and a reasonable order for Ũi
are obtained during this process. The order will be as follows: Transfer Bed Chair,
Outings etc. Every elderly man has his own exclusive reasonable questionnaire order,
which is made especially for that individual [16].
The FPQM is evaluated using the metrics AR, RR, and F, which are calculated with
Eqs (21), (24), and (27), respectively. The AR, RR, and F of the FPQM are shown in
Figure 4. The Y-axis indicates the values of AR, RR, and F, and the x-axis indicates the
id of the elderly individual, thereby representing each investigated person. AR is more
volatile; accordingly, F is also volatile. The FPQM is only a preceding questionnaire
model, and the time required to obtain the pre-result is very short. To ensure 100% accuracy, the obtained pre-result needs to be verified by the elderly individual, but verification
is certainly much faster than asking directly.
4.4. Comparison experiments
To demonstrate that the FPQM is a good model for fast preceding questionnaires, Expert
Knowledge, Rough Set [12,17], and C4.5 [13,23] are applied to solve the same problem.
The results of the FPQM and those from these three methods are compared.
4.4.1. Expert Knowledge
AR, RR and F are shown in Figure 5 when using the Expert Knowledge method. Expert
Knowledge achieves a high AR, but the RR is very low. This result conforms with the
characteristics of experts, who ensure high AR but do not greatly affect RR. AR is more
volatile than RR between elderly individuals; accordingly, F is also volatile.
T. Mo et al. / An influence-based fast preceding questionnaire model
Figure 4. AR, RR and F of the FPQM.
Figure 5. AR, RR and F of the Expert Knowledge method.
4.4.2. Rough Set
The Rough Set method is also applied to solve the same problem in four steps. Step
1: First, take one attribute as the decision attribute and the other directions as the condition attributes. The correlation degree between each condition attribute and the decision attribute is calculated, and then, the attribute for which its correlation degree is less
than a given threshold is deleted. Second, calculate the correlation degree between each
T. Mo et al. / An influence-based fast preceding questionnaire model
condition with all the other conditions. Delete the condition attributes whose correlation
degree with this particular condition attribute is larger than the degree with the decision
attribute. Step 2: Generate the decision rules with the reliability and coverage degrees.
Let the decision equivalent class (DEC) be the collection of elderly individuals who have
the same assessment result. Calculate the reliability and coverage degree for each DEC
and delete those classes whose degree is less than a given threshold. The classes that may
generate rarely appearing rules and uncertain rules can be deleted in this manner. The
decision rules can be created using each of the remaining decision equivalent classes.
Step 3: Sort the investigation attributes with the coverage degree. Merge the rules according to the coverage degree and create the assessment sequence of attributes by sorting
the coverage degrees in descending order. Step 4: Use the merged rules and assessment
sequence to simulate the assessment process for the new elderly individual in Ũ.
When using the Rough Set method, AR, RR and F are shown in Figure 6. Rough Set
achieves a high AR, but RR is slightly low. Both AR and RR are very volatile; therefore,
F is also quite volatile.
Figure 6. AR, RR and F of the Rough Set method.
4.4.3. C4.5
Now, we solve the same problem by applying the C4.5 method. We choose each investigation attribute of the 45 attributes in total of V as the terminal node, and other attributes
are internal nodes. Using the C4.5 method, a decision tree is obtained, and there are 45
decision trees with 45 attributes. Then, we evaluate the new individual Ũi of Ũ with the
45 obtained decision trees. Initially, there are no attributes that can be inferred, and we
ask the elderly individual Ũi for the real value Ri j directly. After accumulating sufficient
real values {Ri j }, j ∈ {1, 2, ..., j∗ }, some attributes can be inferred as the leaf nodes of
some of the 45 obtained decision trees; then, we can infer their final values R0i j . Other-
T. Mo et al. / An influence-based fast preceding questionnaire model
wise, we continue to ask about Ri j directly. Finally, all final values R0i j of the 45 attributes
can be obtained.
AR, RR and F using the C4.5 method are shown in Figure 7. C4.5 achieves very
low AR, RR and F. The volatility of AR is slightly high, but C4.5 achieves a significantly
lower volatility in RR and F.
Figure 7. AR, RR and F of the C4.5 method.
4.4.4. Comparison results
The FPQM, Expert Knowledge, Rough Set, and C4.5 methods are compared with respect
to AR, RR and F individually to demonstrate which method achieves the best result.
Figure 8 shows the AR comparison for the FPQM, Expert Knowledge, Rough Set,
and C4.5 methods. The FPQM achieves the best AR performance, and the volatility is
the lowest. Expert Knowledge achieves a slightly better AR than the Rough Set and C4.5
methods, but these three methods have almost the same volatility.
The RR comparison of the FPQM, Expert Knowledge, Rough Set, and C4.5 methods
is shown in Figure 9. Apparently, the FPQM also achieves the best RR, which is far
higher than those of the other three methods. The Rough Set method achieves the second
best value and has the greatest volatility. By comparison, RR of the Expert Knowledge
and C4.5 methods are the worst and approximately equal.
F is an integrated measurement of AR and RR. Figure 10 shows the F comparison
for the FPQM, Expert Knowledge, Rough Set, and C4.5 methods. The FPQM achieves
the best AR and RR, as will F naturally, and the volatility is the lowest. Because the
other three methods have almost the same AR volatility, the F volatility is mainly determined by RR. The F performance of the Rough Set method is the second best but is the
most volatile. The Expert Knowledge and C4.5 methods are the worst and approximately
equal, but they do not exhibit a very high volatility.
T. Mo et al. / An influence-based fast preceding questionnaire model
Figure 8. The AR comparison for the FPQM, Expert Knowledge, Rough Set, and C4.5 methods.
Figure 9. The RR comparison of the FPQM, Expert Knowledge, Rough Set, and C4.5 methods.
Table 5 shows the mean and standard deviation of the four methods with respect to
AR, RR, and F. The mean reflects the overall performance, and the standard deviation
reflects the volatility in digital form. The mean of AR is denoted as AAR, as mentioned
previously; similar notation is used for RR and F. These values can be calculated with
Eqs (22), (25), and (28). The standard deviation of AR is denoted as SAR; similar notation
is used for RR and F. These values can be calculated with Eqs (23), (26), and (29).
T. Mo et al. / An influence-based fast preceding questionnaire model
Figure 10. The F comparison of the FPQM, Expert Knowledge, Rough Set, and C4.5 methods.
Table 5. Mean and standard deviation of the results under the four methods
Metric
Method
Mean
Standard deviation
AR
FPQM
Expert Knowledge
Rough Set
C4.5
0.9839
0.9203
0.7589
0.6850
0.0560
0.1993
0.2223
0.2072
RR
FPQM
Expert Knowledge
Rough Set
C4.5
0.9056
0.0769
0.3335
0.0916
0.0221
0.0262
0.1003
0.0312
F
FPQM
Expert Knowledge
Rough Set
C4.5
0.9666
0.2800
0.5984
0.2778
0.0461
0.0870
0.1758
0.0637
The FPQM has the largest mean and the smallest standard deviation with respect
to AR, RR and F. This means that the FPQM achieves the best performance and has
the lowest volatility. The Expert Knowledge method has the second best mean for AR,
0.9203, and it has the smallest mean for RR, which conforms to the characteristics of
expert knowledge. The Rough Set method is the most volatile with respect to AR, RR
and F, and it achieves the second best mean for F and RR. The performance of the C4.5
method is almost the worst.
4.4.5. Result analysis
Some of the improvements in the results are not highly significant when the FPQM is
compared with other methods, such as the Rough Set method. We can explain this as
T. Mo et al. / An influence-based fast preceding questionnaire model
follows. The Rough Set method is also an excellent method for attribute reduction. Both
the Rough Set and FPQM methods can capture the internal relationships among the attributes listed in Table 4, and predict multiple attributes simultaneously. Therefore, the
FPQM does not achieve highly significant improvements compared with the Rough Set
method.
4.5. Factor analysis
Three factors are analyzed: the number of elderly individuals, the number of investigation
attributes, and the threshold σ . The evaluation metrics AAR, ARR, AF, SAR, SRR, and
SF are used to show the changes in the results. AAR, ARR, and AF can be calculated with
Eq. (22), (25), and (28), respectively. Eq. (23), (26), and (29) can be used to calculate
SAR, SRR, and SF, respectively.
As illustrated in Figure 11, ARR decreases steadily and with minimal volatility when
the number of the elderly individuals increases. AAR and AF are volatile and present a
slight wavelike decrease overall. SRR increases continuously with a small slope. SF is
highly similar to SAR because SRR is so small that SF is mainly determined by SAR.
Figure 11. AAR, ARR, AF, SAR, SRR, and SF with increasing number of elderly individuals.
With increasing number of investigation attributes, AAR, ARR, and AF present a
fluctuating increasing trend, as shown in Figure 12. The increased level of ARR is higher
than that of AAR and AF, and ARR shows slightly larger volatility. SAR, SRR, and SF
decrease with some volatility, and SRR falls fairly steadily in the second half of the curve.
SRR has minimal impact on SF, and SF has the same general trend as SAR. Briefly, more
investigation attributes produce better results.
Notice that some parts of the AAR, ARR, AF, SAR, SRR, and SF curves are flat,
which simply embodies the intermittent nature of the threshold, as illustrated in Figure 13. Only the threshold σ achieves a level whereby a significant change can oc-
T. Mo et al. / An influence-based fast preceding questionnaire model
Figure 12. AAR, ARR, AF, SAR, SRR, and SF with increasing number of investigation attributes.
cur. Changes mainly occur when σ is in the range of 0.3-0.65. If σ is too small,
Ii j = 1 ∀1 6 i 6 m̃, 1 6 i 6 n. On the other hand, Ii j = 0 ∀1 6 i 6 m̃, 1 6 i 6 n when σ
is too large. When the threshold σ increases, AAR, AF, and SRR present an increasing
trend, whereas ARR, SAR and SF decrease. SRR simply shows a slow growth. ARR and
SRR have minimal impact on AF and SF, respectively. Therefore, AF and SF show the
same general trends as AAR and SAR, respectively. Therefore, to achieve better results, it
is suggested to choose a slightly larger threshold σ .
5. Conclusion
Traditional questionnaires are used in a manner whereby respondents are asked one question after another. The main two problems are inefficiency and that the order of the questions is not reasonable. In this paper, the fast preceding questionnaire model (FPQM)
is proposed to solve these problems in five steps, as shown in Figure 1. The influence
calculation formula, best attribute to split choosing algorithm (BASCA), fast preceding
questionnaire model creating algorithm (FPQMCA), model used for real investigation
algorithm (MURIA), and model evaluation algorithm (MEA) are all presented. The experimental section presents the experimental data; the evaluation metrics; the overall results of the FPQM; the comparison experiments with the Expert Knowledge, Rough Set,
and C4.5 methods; and factor analysis, which includes the number of elderly individuals,
the number of investigation attributes, and the threshold. When the FPQM is applied by
the Lime Family company, after asking about certain attributes, most of the remaining
attributes could be inferred automatically with a high accuracy and reduction rate and
low volatility. To ensure 100% correctness, the model can be considered as a preceding
questionnaire. Then, the elderly individual can perform verification, which is much faster
than asking questions directly.
T. Mo et al. / An influence-based fast preceding questionnaire model
Figure 13. AAR, ARR, AF, SAR, SRR, and SF with increasing threshold σ .
Further work will focus on three points. After the elderly individual is well assessed,
a thorough study should be performed on determining an appropriate health care plan
to be recommended automatically according to the assessment result. Then, the effect
of the health care plan should be assessed as well. Moreover, as the amount data on
elderly individuals continues increasing, certain distributed platforms, such as Hadoop
and Spark, will be considered.
Acknowledgements
This work is supported by the National High Technology Research and Development
Program (“863” Program) of China under Grant No. 2015AA016009 and the National
Natural Science Foundation of China under Grant No. 61232005
The authors wish to thank Lei Yang from the Lime Family Company, who provided
us with the evaluation data for the elderly individuals.
Declaration: The data presented herein were only used for academic research. The
customer ids have been anonymized, and the privacy of the elderly will not be invaded.
References
[1]
M. Aoyama, Y. Suzuki, J. Onishi and M. Kuzuya, Physical and functional factors in activities of daily
living that predict falls in community-dwelling older women, in: Geriatrics & Gerontology International
11(3) (2011), 348–357.
[2] B. A. Arnow, E. M. Hunkeler, C. M. Blasey, J. Lee, M. J. Constantino, B. Fireman, H. C. Kraemer,
R. Dea, R. Robinson and C. Hayward, Comorbid depression, chronic pain, and disability in primary
care, in: Psychosomatic Medicine 68(2) (2006), 262–268.
[3] A. Aydeniz, Ö. Altındağ, J. İrdesel, E. Madenci, F. Toraman, N. Turhan, İ. Yağcı, N. Eskiyurt, P. Yazgan
and Y. G. Kutsal, Physical, functional and sociocultural parameters that predict fall in elderly: Multicen-
T. Mo et al. / An influence-based fast preceding questionnaire model
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
ter study., in: Journal of Physical Medicine & Rehabilitation Sciences/Fiziksel Tup ve Rehabilitasyon
Bilimleri Dergisi 18(3).
X. Badia, A. Vieta and M. Gilabert, The bone metastases quality of life questionnaire, in: Handbook of
Disease Burdens and Quality of Life Measures, Springer (2010), 195–207.
S. Cinaroglu, Comparison of Performance of Decision Tree Algorithms and Random Forest: An Application on OECD Countries Health Expenditures, in: .
M. Cruice, L. Worrall and L. Hickson, Personal factors, communication and vision predict social participation in older adults, in: Advances in Speech Language Pathology 7(4) (2005), 220–232.
A.-L. Dima, Living with chronic pain-a longitudinal study of the interrelations between acceptance,
emotions, illness perceptions and health status, Ph.D. thesis (2010).
F. J. Drummond, L. Sharp, A.-E. Carsin, T. Kelleher and H. Comber, Questionnaire order significantly
increased response to a postal survey sent to primary care physicians, in: Journal of Clinical Epidemiology 61(2) (2008), 177–185.
K. K. Faulkner and R. Cogan, Measures of shame and conflict tactics: Effects of questionnaire order, in:
Psychological Reports 66(3 suppl) (1990), 1217–1218.
E. Fernandez and G. J. Boyle, Affective and evaluative descriptors of pain in the McGill pain questionnaire: Reduction and reorganization, in: The Journal of Pain 2(6) (2001), 318–325.
J. L. Gatz, S. L. Tyas, P. S. John and P. Montgomery, Do depressive symptoms predict Alzheimer’s
disease and dementia?, in: The Journals of Gerontology Series A: Biological Sciences and Medical
Sciences 60(6) (2005), 744–747.
A. Gyorgy and T. Linder, Codecell convexity in optimal entropy-constrained vector quantization, in:
IEEE Transactions on Information Theory 49(7) (2003), 1821–1828.
J. Han, J. Pei and M. Kamber, Data Mining: Concepts and Techniques, Elsevier (2011).
B. C. Kersh, L. A. Bradley, G. S. Alarcón, K. R. Alberts, A. Sotolongo, M. Y. Martin, L. A. Aaron, D. F.
Dewaal, M. L. Domino, W. F. Chaplin et al., Psychosocial and health status variables independently
predict health care seeking in fibromyalgia, in: Arthritis Care & Research 45(4) (2001), 362–371.
W. Kitisomprayoonkul, J. Klaphajone and A. Kovindha, Thai short-form McGill pain questionnaire, in:
Journal-Medical Association of Thailand 89(6) (2006), 846.
M. Kroh, F. Winter and J. Schupp, Using person-fit measures to assess the impact of panel conditioning
on reliability, in: Public Opinion Quarterly (2016), nfw025.
Y. Liu and W. Zhu, On three types of covering-based rough sets via definable sets, in: 2014 IEEE
International Conference on Fuzzy Systems (FUZZ-IEEE), IEEE, 1226–1233.
F. I. Mahoney, Functional evaluation: The Barthel index, in: Maryland State Medical Journal 14 (1965),
61–65.
E. Mccoll, M. P. Eccles, N. S. Rousseau, I. N. Steen, D. W. Parkin and J. M. Grimshaw, From the generic
to the condition-specific? Instrument order effects in quality of life assessment, in: Medical Care 41(7)
(2003), 777–790.
I. McDowell, Measuring Health: A Guide to Rating Scales and Questionnaires, Oxford University Press
(2006).
T. E. Nijsten, F. Sampogna, M.-M. Chren and D. D. Abeni, Testing and reducing skindex-29 using Rasch
analysis: Skindex-17, in: Journal of Investigative Dermatology 126(6) (2006), 1244–1250.
L. Prieto, J. Alonso and R. Lamarca, Classical test theory versus Rasch analysis for quality of life
questionnaire reduction, in: Health and Quality of Life Outcomes 1(1) (2003), 1.
J. R. Quinlan, C4. 5: Programs for Machine Learning, Elsevier (2014).
A. E. Rippentrop, E. M. Altmaier, J. J. Chen, E. M. Found and V. J. Keffala, The relationship between
religion/spirituality and physical health, mental health, and pain in a chronic pain population, in: Pain
116(3) (2005), 311–321.
R. Rosen, J. Cappelleri, M. Smith, J. Lipsky and B. Pena, Development and evaluation of an abridged,
5-item version of the international index of erectile function (IIEF-5) as a diagnostic tool for erectile
dysfunction, in: International Journal of Impotence Research 11(6) (1999), 319–326.
S. Singh and P. Gupta, Comparative study ID3, cart and C4. 5 decision tree algorithm: a survey, in: Int J
Adv Inf Sci Technol [Internet] 27 (2014), 97–103.
S. W. Vines, S. Gupta, T. Whiteside, D. Dostal-Johnson and A. Hummler-Davis, The relationship between chronic pain, immune function, depression, and health behaviors, in: Biological Research for
Nursing 5(1) (2003), 18–29.
S. Wang, X. Feng and Y. Liu, MZ/T 039-2013, Ability Assessment for Elder Adults, Standards Press of
T. Mo et al. / An influence-based fast preceding questionnaire model
China (2014).
Appendices
A. The time complexity of the FPQM
A.1. The time complexity of the BASCA
As shown in Algorithm 1, there are two loops on the attributes (Line 1 and Line 2) and
two loops on the values of the attributes (Line 3 and Line 4) in the BASCA. As shown
in Definition 3.3, V = {V1 ,V2 , ...,Vn }; therefore, there are n investigation attributes in V .
V jtk has N jk1 possible values, where N jk1 is defined in Definition 3.5. Therefore, the time
1
complexity of Lines 1-13 is
1
TBASCA
(n, N̄) = O(n2 N̄ 2 )
(30)
where N̄ is the average of N j in N, N̄ = n1 ∑nj=1 N j , and n is the number of V .
Line 14 calculates V∗t with Eq. (5), and there are n investigation attributes in V ;
therefore, the time complexity of Line 14 is
2
TBASCA
(n, N̄) = O(n)
(31)
Thus, the time complexity of the BASCA is
1
2
TBASCA (n, N̄) = TBASCA
(n, N̄) + TBASCA
(n, N̄) = O(n2 N̄ 2 )
(32)
A.2. The time complexity of the FPQMCA
The time complexity of Lines 1-5 is
1
(n, N̄) = O(1)
TFPQMCA
(33)
The time complexity of Line 7 is
2
TFPQMCA
(n, N̄) = O(n2 N̄ 2 )
(34)
The time complexity of Lines 8-10 is
3
TFPQMCA
(n, N̄) = O(1)
(35)
T. Mo et al. / An influence-based fast preceding questionnaire model
The time complexity of Lines 11-15 is
4
TFPQMCA
(n, N̄) = N̄ ∗ TFPQMCA (n − 1)
(36)
The time complexity of the FPQMCA is
1
2
3
4
TFPQMCA (n, N̄) = TFPQMCA
(n, N̄) + TFPQMCA
(n, N̄) + TFPQMCA
(n, N̄) + TFPQMCA
(n, N̄)
= O(1) + O(n2 N̄ 2 ) + O(1) + N̄ ∗ TFPQMCA (n − 1, N̄)
= O(n2 N̄ 2 ) + N̄ ∗ TFPQMCA (n − 1, N̄)
(37)
Eq. (37) is the recursion formula of TFPQMCA (n, N̄).
TFPQMCA (n, N̄) = O(n2 N̄ 2 ) + N̄ ∗ TFPQMCA (n − 1, N̄)
= O(n2 N̄ 2 ) + N̄ ∗ O (n − 1)2 N̄ 2 + N̄ ∗ TFPQMCA (n − 2, N̄)
= O(n2 N̄ 2 ) + O (n − 1)2 N̄ 3 + N̄ 2 ∗ TFPQMCA (n − 2, N̄)
= O(n2 N̄ 2 ) + O (n − 1)2 N̄ 3 ) + ... + O(12 N̄ n+1 ) + N̄ n−1 ∗ T (1, N̄)
n
= ∑ O(i2 N̄ n−i+2 ) + N̄ n−1 ∗ T (1, N̄)
i=1
n
=O
∑ (i2 N̄ n−i+2 )
+ N̄ n−1 ∗ T (1, N̄)
(38)
i=1
Let S = ∑ni=1 (i2 N̄ n−i+2 ).
n−1
(N̄ − 1)S = N̄ n+2 + ∑ (i + i + 1)N̄ n−i+2 − n2 N̄ 2
(39)
i=1
n+2
N̄(N̄ − 1)S − (N̄ − 1)S = N̄ n+3 + 2 ∑ N̄ i − (n2 + 2n − 1)N̄ 3 + n2 N̄ 2
(40)
n+3 2N̄ 4 (1 − N̄ n−1 )
1
N̄
+
− (n2 + 2n − 1)N̄ 3 + n2 N̄ 2
2
(N̄ − 1)
1 − N̄
(41)
i=4
S=
Plugging Eq. (41) into Eq. (38), T (1, N̄) = O(1); thus,
T. Mo et al. / An influence-based fast preceding questionnaire model
TFPQMCA (n, N̄)
n+3 2N̄ 4 (1 − N̄ n−1 )
1
2
3
2 2
+ N̄ n−1 ∗ T (1, N̄)
=O
N̄
N̄
+
−
(n
+
2n
−
1)
N̄
+
n
(N̄ − 1)2
1 − N̄
n+3 2N̄ 4 (1 − N̄ n−1 )
1
2
3
2 2
=O
+ O(N̄ n−1 )
N̄
N̄
+
−
(n
+
2n
−
1)
N̄
+
n
(N̄ − 1)2
1 − N̄
1
N̄ 4 N̄ n−1
= O 2 N̄ n+3 +
− n2 N̄ 3 + n2 N̄ 2 + O(N̄ n−1 )
N̄
N̄
= O(N̄ n+1 + n2 N̄)
(42)
The time complexity of the FPQMCA is
TFPQMCA (n, N̄) = O(N̄ n+1 + n2 N̄)
(43)
A.3. The time complexity of the MURIA
The time complexity of Lines 1-10 is
1
TMURIA
(n, N̄) = O(1)
(44)
The time complexity of Line 11 is
2
(n, N̄) = O(N̄)
TMURIA
(45)
The time complexity of Lines 12-24 is
3
TMURIA
(n, N̄) = O(1)
(46)
The time complexity of Lines 25-27 is
4
TMURIA
(n, N̄) = TMURIA (n − 1, N̄)
(47)
The time complexity of the MURIA is
1
2
3
4
(n, N̄) + TMURIA
(n, N̄) + TMURIA
(n, N̄)
TMURIA (n, N̄) = TMURIA
(n, N̄) + TMURIA
= O(1) + O(N̄) + O(1) + TMURIA (n − 1, N̄)
= O(N̄) + TMURIA (n − 1, N̄)
Eq. (48) is the recursion formula of TMURIA (n, N̄).
(48)
T. Mo et al. / An influence-based fast preceding questionnaire model
TMURIA (n, N̄) = O(N̄) + TMURIA (n − 1, N̄)
= 2O(N̄) + TMURIA (n − 2, N̄)
= (n − 1)O(N̄) + TMURIA (1, N̄)
(49)
TMURIA (1, N̄) = O(1); therefore, the time complexity of the MURIA is
TMURIA (n, N̄) = (n − 1)O(N̄) + O(1) = O(nN̄)
(50)
A.4. The time complexity of the MEA
The time complexity of Lines 1-9 is
1
TMEA
(n, m̃) = O(m̃n)
(51)
The time complexity of Lines 14-27 is
2
TMEA
(n, m̃) = O(m̃n)
(52)
The time complexity of Lines 10-13 and 28-30 is
3
TMEA
(n, m̃) = O(1)
(53)
Thus, the time complexity of the MEA is
1
2
3
TMEA (n, m̃) = TMEA
(n, m̃) + TMEA
(n, m̃) + TMEA
(n, m̃) = O(m̃n)
(54)
A.5. The time complexity of the FPQM
The time complexity of the FPQM is
TFPQM (n, N̄, m̃) = TBASCA (n, N̄) + TFPQMCA (n, N̄) + TMURIA (n, N̄) + TMEA (n, m̃)
= O(n2 N̄ 2 ) + O(N̄ n+1 + n2 N̄) + O(nN̄) + O(m̃n)
= O(n2 N̄ 2 ) + O(N̄ n+1 ) + O(m̃n)
(55)
| 2cs.AI
|
Longest Common Prefixes with k-Errors and
Applications
arXiv:1801.04425v1 [cs.DS] 13 Jan 2018
Lorraine A.K. Ayad, Panagiotis Charalampopoulos, Costas S. Iliopoulos, and
Solon P. Pissis
Department of Informatics, King’s College London, London, UK
[lorraine.ayad,panagiotis.charalampopoulos,
costas.iliopoulos,solon.pissis]@kcl.ac.uk
Abstract. Although real-world text datasets, such as DNA sequences,
are far from being uniformly random, average-case string searching algorithms perform significantly better than worst-case ones in most applications of interest. In this paper, we study the problem of computing the
longest prefix of each suffix of a given string of length n over a constantsized alphabet that occurs elsewhere in the string with k-errors. This
problem has already been studied under the Hamming distance model.
Our first result is an improvement upon the state-of-the-art average-case
time complexity for non-constant k and using only linear space under the
Hamming distance model. Notably, we show that our technique can be
extended to the edit distance model with the same time and space complexities. Specifically, our algorithms run in O(n logk n log log n) time on
average using O(n) space. We show that our technique is applicable to
several algorithmic problems in computational biology and elsewhere.
1
Introduction
The longest common prefix (LCP) array is a commonly used data structure
alongside the suffix array (SA). The LCP array stores the length of the longest
common prefix between two adjacent suffixes of a given string as they are stored
(in lexicographical order) in the SA [20]. A typical use combining the SA and
the LCP array is to simulate the suffix tree functionality using less space [2].
However, there are many practical scenarios where the LCP array may be
applied without making use of the SA. The LCP array provides us with essential
information regarding repetitiveness in a given string and is therefore a useful
data structure for analysing textual data in areas such as molecular biology,
musicology, or natural language processing (see [21] for some applications).
It is also quite common to account for potential alterations within textual
data (sequences). For example, they can be the result of DNA replication or
sequencing errors in DNA sequences. In this context, it is natural to define the
longest common prefix with k-errors. Given a string x[0 . . n − 1], the longest
common prefix with k-errors for every suffix x[i . . n − 1] is the length of the
longest common prefix of x[i . . n − 1] and any x[j . . n − 1], where j 6= i, with
applying up to k substitution operations [21]. Some applications are given below.
Interspersed Repeats. Repeated sequences are a common feature of genomes.
One type in particular, interspersed repeats, are known to occur in all eukaryotic
genomes. These repeats have no repetitive pattern and appear irregularly within
DNA sequences [15]. Single nucleotide polymorphisms result in the existence of
interspersed repeats that are not identical [19]. Identifying these repeats has
been linked to genome folding locations and phylogenetic analysis [24].
Genome Mappability Data Structure. In [3] the authors showed that using the
longest common prefixes with k-errors they can construct, in O(n) worst-case
time, an O(n)-sized data structure answering the following type of queries in
O(1) time per query: find the smallest m such that at least µ of the substrings
of x of length m do not occur more than once in x with at most k errors. This
is a data structure version of the genome mappability problem [8, 21, 4].
Longest Common Substring with k-Errors. The longest common substring with
k-errors problem has received much attention recently, in particular due to its applications in computational biology [27, 18, 26]. We are asked to find the longest
substrings of two strings that are at distance at most k. The notion of longest
common prefix with k-errors is thus closely related to the notion of longest common substring with k-errors. We refer the interested reader to [1, 10, 11, 14, 25].
All-Pairs Suffix/Prefix Overlaps with k-Errors. Finding approximate overlaps
is the first stage of most genome assembly methods. Given a set of strings and
an error-rate ǫ, the goal is to find, for all pairs of strings, their suffix/prefix
matches (overlaps) that are within distance k = ⌈ǫℓ⌉, where ℓ is the length of
the overlap [23, 28, 16]. By concatenating the strings to form one single string x
and then computing longest common prefixes with k-errors for x only against the
prefixes of the strings we have all the information we need to solve this problem.
Our Model. We assume the standard word-RAM model with word size w =
Ω(log n). Although real-world text datasets are far from being uniformly random,
average-case string searching algorithms perform significantly better than worstcase ones in most applications of interest. We are thus interested in the averagecase behaviour of our algorithms. When we state average-case time complexities
for our algorithms, we assume that the input is a string x of length n over an
alphabet Σ of size σ > 1 with the letters of x being independent and identically
distributed random variables, uniformly distributed over Σ. In the context of
molecular biology we typically have Σ = {A,C,G,T} and so we assume σ = O(1).
Related Works. The problem of computing longest common prefixes with kerrors was first studied by Manzini for k = 1 in [21]. We distinguish the following
techniques that can be applied to solve this and other related problems.
Non-constant k and ω(n) space: In this case, we can make use of the wellknown data structure by Cole et al [7]. The size of the data structure is
n)k
O(n (c log
), where c > 1 is a constant.
k!
2
Constant k and O(n) space: In this case, we can make use of the technique
by Thankachan et al [25] which builds heavily on the data structure by Cole
et al. The working space is exponential in k but O(n) for k = O(1).
Non-constant k and O(n) space: In this case, there exists a simple O(n2 k)time worst-case algorithm to solve the problem. The best-known averagecase algorithm was presented in [3]. It requires O(n(σR)k log log n(log k +
log log n)) time on average, where R = ⌈(k + 2)(logσ n + 1)⌉.
Other related works: In [14] it was shown that a strongly subquadratic-time
algorithm for the longest common substring with k-errors problem, for k =
Ω(log n) and binary strings, refutes the Strong Exponential Time Hypothesis. Thus subquadratic-time solutions for approximate variants of the problem have been developed [14]. A non-deterministic algorithm is also known [1].
Our Contribution. In this paper, we continue the line of research for non-constant
k and O(n) space to investigate the limits of computation in the average-case
setting; in particular in light of the worst-case lower bound shown in [14]. We
make the following threefold contribution.
1. We first show a non-trivial upper bound of independent interest: the expected
length of the maximal longest common prefix with k-errors between a pair
of suffixes of x is O(logσ n) when k ≤ logloglogn n .
2. By applying this result, we significantly improve upon the state-of-the-art
algorithm for non-constant k and using O(n) space [3]. Specifically, our algorithm runs in O(n logk n log log n) time on average using O(n) space.
3. Notably, we extend our results to the edit distance model with no extra cost
thus solving the genome mappability data structure problem, the longest
common substring with k-errors problem, and the all-pairs suffix/prefix overlaps with k-errors problem in strongly sub-quadratic time for k ≤ logloglogn n .
2
Preliminaries
We begin with some basic definitions and notation. Let x = x[0]x[1] . . . x[n−1] be
a string of length |x| = n over a finite ordered alphabet Σ of size |Σ| = σ = O(1).
For two positions i and j on x, we denote by x[i . . j] = x[i] . . . x[j] the substring
(sometimes called factor ) of x that starts at position i and ends at position j.
We recall that a prefix of x is a substring that starts at position 0 (x[0 . . j]) and
a suffix of x is a substring that ends at position n − 1 (x[i . . n − 1]).
Let y be a string of length m with 0 < m ≤ n. We say that there exists an
occurrence of y in x, or, more simply, that y occurs in x, when y is a substring
of x. Every occurrence of y can be characterised by a starting position in x. We
thus say that y occurs at the starting position i in x when y = x[i . . i + m − 1].
The Hamming distance between two strings x and y, with |x| = |y|, is defined
as dH (x, y) = |{i : x[i] 6= y[i], i = 0, 1, . . . , |x|−1}|. If |x| 6= |y|, we set dH (x, y) =
∞. The edit distance between x and y is the minimum total cost of a sequence
of edit operations (insertions, deletions, substitutions) required to transform x
into y. It is known as Levenshtein distance for unit cost operations. We consider
3
this special case here. If two strings x and y are at (Hamming or edit) distance
at most k we say that x and y have k-errors or have at most k errors.
We denote by SA the suffix array of x. SA is an integer array of size n storing
the starting positions of all (lexicographically) sorted non-empty suffixes of x,
i.e. for all 1 ≤ r < n we have x[SA[r − 1] . . n − 1] < x[SA[r] . . n − 1] [20]. Let
lcp(r, s) denote the length of the longest common prefix between x[SA[r] . . n − 1]
and x[SA[s] . . n − 1] for positions r, s on x. We denote by LCP the longest
common prefix array of x defined by LCP[r] = lcp(r − 1, r) for all 1 ≤ r < n,
and LCP[0] = 0. The inverse iSA of the array SA is defined by iSA[SA[r]] = r,
for all 0 ≤ r < n. It is known that SA, iSA, and LCP of a string of length n,
over a constant-sized alphabet, can be computed in time and space O(n) [22, 9].
It is then known that a range minimum query (RMQ) data structure over the
LCP array, that can be constructed in O(n) time and O(n) space [5], can answer
lcp-queries in O(1) time per query [20]. The lcp queries are also known as longest
common extension (LCE) queries.
The permuted LCP array, denoted by PLCP, has the same contents as the
LCP array but in different order. Let i− denote the starting position of the
lexicographic predecessor of x[i . . n−1]. For i = 0, . . . , n−1, we define PLCP[i] =
LCP[iSA[i]] = lcp(iSA[i− ], iSA[i]]), that is, PLCP[i] is the length of the longest
common prefix between x[i . . n − 1] and its lexicographic predecessor. For the
starting position j of the lexicographically smallest suffix we set PLCP[j] = 0.
For any k ≥ 0, we define lcpk (y, z) as the largest ℓ ≥ 0 such that y[0 . . ℓ − 1]
and z[0 . . ℓ − 1] exist and are at Hamming distance at most k; note that this is
defined for a pair of strings. We analogously define the permuted LCP array with
k-errors, denoted by PLCPk . For i = 0, . . . , n − 1, we have that
PLCPk [i] =
max
j=0,...,n−1, j6=i
lcpk (x[i . . n − 1], x[j . . n − 1]).
The main computational problem in scope can be formally stated as follows.
PLCP with k-Errors
Input: A string x of length n and an integer 0 < k < n
Output: PLCPk and Pk ; Pk [i] 6= i, for i = 0, . . . , n − 1, is such that x[i . . i +
ℓ − 1] ≈k x[Pk [i] . . Pk [i] + ℓ − 1], where ℓ = PLCPk [i]
We assume that k ≤
log n
log log n
throughout, since all relevant time-complexities
k
contain an n log n factor and any larger k would force this value to be Ω(n2 ):
n logk n ≤ cn2 ⇔ k log log n ≤ log(cn) ⇔ k ≤
3
log c + log n
, c ≥ 1.
log log n
Computing PLCPk
In this section we propose a new algorithm for the PLCP with k-Errors
problem under both the Hamming and the edit distance (Levenshtein distance)
4
models. This algorithm is based on a deeper look into the expected behaviour
of the longest common prefixes with k-errors. This in turn allows us to make
use of the y-fast trie, an efficient data structure for maintaining integers from
a bounded domain. We already know the following result for errors under the
Hamming distance model.
Theorem 1 ([3]). Problem PLCP with k-Errors for 1 ≤ k ≤ logloglogn n can be
solved in average-case time O(n(σR)k log2 log n), where R = ⌈(k+2)(logσ n+1)⌉,
using O(n) extra space.
In the rest of this section, we show the following result for errors under both
the Hamming and the edit distance models.
Theorem 2. Problem PLCP with k-Errors can be solved in average-case
k
time O(n ck! logk n log log n), where c is a constant, using O(n) extra space.
For clarity of presentation, we first do the analysis and present the algorithm
under the Hamming distance model in Sections 3.1 and 3.2. We then show how
to extend our technique to work under the edit distance model in Section 3.3.
3.1
Expectations
The expected maximal value in the LCP array is 2 logσ n + O(1) [13]. We can
thus obtain a trivial O(k logσ n) bound on the expected length of the maximal
longest common prefix with k-errors for arbitrary k and σ. By looking deeper
into the expected behaviour of the longest common prefixes with k-errors we
show the following result of independent interest for when k ≤ logloglogn n .
Theorem 3. Let x be a string of length n over an alphabet of size σ > 1 and
1 ≤ k ≤ logloglogn n be an integer.
(a) The expected length of the maximal longest common prefix with k-errors between a pair of suffixes of x is O(logσ n).
(b) There exists a constant α such that the expected number of pairs of suffixes
of x with a common prefix with k-errors of length at least α logσ n is O(1).
Proof (a). Let us denote the ith suffix of x by xi = x[i . . n − 1]. Further let us
define the following random variables:
Xi,j = lcpk (xi , xj ) and Y =
Claim. Pr(Xi,j ≥ m) ≤
max
0≤i<j≤n−1
Xi,j .
m
1
m−k .
k σ
Proof (of Claim). Each possible
set of positions where a substitution is allowed
is a subset of one of the m
subsets
of m of size k. For each of these subsets, we
k
can disregard what happens in the k chosen positions; in order to yield a match
with k-errors, the remaining m − k positions must match and each of them
matches with probability σ1 . The claim follows by applying the Union-Bound
(Boole’s inequality).
⊓
⊔
5
By applying the Union-Bound again we have that
Pr(Y ≥ m) = Pr(
[
i<j
{Xi,j ≥ m}) ≤
X
Pr(Xi,j
i<j
m
1
≥ m) ≤ n
,
m−k
k σ
2
for m ≥ k and Pr(Y ≥ m) = 1 for m ≤ k. The expected value of Y is given by:
∞
X
E[Y ] =
m=1
α(logσ +k)
Pr(Y ≥ m) =
X
m=1
|
Pr(Y ≥ m) +
{z
≤α(logσ n+k)
}
∞
X
m=α(logσ n+k)+1
Pr(Y ≥ m).
(Note that we bound the first summand using that Pr(Y ≥ m) ≤ 1 for all m.)
m
rm,k
≤ 23 for m ≥ 6k.
Claim. Let rm,k =
. We have that rm−1,k
k
Proof (of Claim).
rm,k
m
6k
6k
3
=
≤
=
≤ .
rm−1,k
(m − k − 1)
(6k − k − 1)
(5k − 1)
2
⊓
⊔
By assuming β = α(logσ n + k) + 1 ≥ 6k, for some α > 1, we apply the above
claim to bound the second summand as follows.
m−6k
∞
∞
∞
X
X
X
1
3
1
2 6k
2 m
≤
n
Pr(Y ≥ m) ≤
n
2
σ m−k
k
k σ m−k
m=β
=
∞
X
m=β
n2
m=β
m=β
5k β−k X
m
5k m−k
∞
3
6k
3
3
2
6k
2
≤ n2
3
2σ
3
2σ
2σ
k
k
m=0
≤ An6
3
2σ
β−k
≤ An6
3
2σ
α logσ n
=
An6
n
α(1−1/ log3/2 σ)
for some constant A since σ ≥ 2 and (6k)k ≤ 23 log n k k ≤ n3 logk n = O(n4 ).
Then 1 − 1/ log3/2 σ > 0 and we can thus pick an α large enough such that this
sum is O(n−ǫ ) for any ǫ > 0.
⊓
⊔
Proof (b). Let Ii,j,m be the indicator random variable for the event {Xi,j ≥ m}.
We then have that
X
X
X
Pr(Xi,j ≥ m),
E[Ii,j,m ] =
Ii,j,m ] =
E[
i<j
i<j
i<j
which we have already shown is O(1) if m = α logσ n for some α > 1.
6
⊓
⊔
p
i
a0
j0
ae
je
X
X
X
X
i+m−1
Fig. 1: The (e+1)th error is any possible substitution at a position je < p ≤ i+m.
3.2
Improved Algorithm for Hamming Distance
The y-fast trie, introduced in [29], supports insert, delete and search (exact,
predecessor and successor queries) in time O(log log U ) with high probability,
using O(n) space, where n is the number of stored values and U is size of the
universe. We consider each substring of x of length at most λ = α log n for
a constant α satisfying Theorem 3 (b) as a λ-digit number; note that by our
assumptions this number fits in a computer word. We thus have U = σ λ and
hence log log U = O(log log n + log α) = O(log log n).
We initialise PLCPk and Pk for each i based on the longest common prefix
of x[i . . n − 1] (i.e. not allowing any errors) that occurs elsewhere using the SA
and the LCP array; this can be done in O(n) time. For each pair of suffixes that
share a prefix of at least λ we perform (at most) k LCE queries to find their
longest common prefix allowing for k-errors; by Theorem 3 these pairs are O(1).
We then initialise the y-fast trie by inserting x[i . . i + λ − 1] to it for each
position i of x with i ≤ n − λ. (For the rest of the positions, for which we reach
the end of x, we insert x[i . . n − 1] after some trivial technical considerations.)
This procedure takes time O(n log log n) in total.
We then want to find a longest prefix of the σ k λk strings of length at most
λ that are at Hamming distance at most k from x[i . . i + λ − 1] that occurs
elsewhere in x as well as an occurrence of it. If this prefix is of length λ, we
find all positions t in x for which dH (x[i . . i + λ − 1], x[t . . t + λ −1]) ≤ k and
treat each of them individually. We generate a subset of the σ k λk strings; we
avoid generating some that we already know do not occur in x. We only want to
allow the first error at position p, where i ≤ p ≤ i + PLCP0 [i]. Let us denote the
substitution at position j with letter a by (j, a). Suppose that the longest prefix
of x[i . . n − 1] after substitutions (j0 , a0 ), . . . , (je , ae ) that occurs elsewhere in
the string is of length m. We then want to allow the (e + 1)th error at positions
je < p ≤ i + m; inspect Figure 1 for an illustration. It should be clear that we
obtain each possible sequence of substitutions at most once.
We view each string z created after at most k substitution operations as a
number; the aim is to find its longest prefix that occurs elsewhere in x. To this end
we perform at most three queries over the y-fast trie: an exact; a predecessor; and
a successor query. If the exact query is unsuccessful, then either the predecessor
or the successor query will return a factor z ′ of x that attains the maximal longest
common prefix that any factor of x has with z. Note that it may be the case
that z ′ only occurs at position i; however in this case lcpk (z, z ′ ) will be smaller
7
or equal than the value currently stored at PLCPk [i] due to how we generate
each such string z. Hence we do not perform an invalid update of PLCPk [i].
Having found z ′ , we can then compute the length of the longest common
prefix between z and z ′ in constant time using standard bit-level operations. For
clarity of presentation we assume |z| = |z ′ | = λ. An XOR operation between z
and z ′ provides us with an integer d specifying the positions of errors (bits set
on when d is viewed as binary). If d 6= 0, we take δ = ⌊log d⌋, which provides us
with the index of the leftmost bit set on which in turn specifies the length of the
σ⌉−δ−1
longest common prefix between z and z ′ ; specifically lcp0 (z, z ′ ) = ⌊ λ⌈log
⌋.
⌈log σ⌉
′
If z = z we perform LCE queries between all suffixes of the text that have z
as a prefix and x[j . . j + λ − 1]; by Theorem 3 we expect this to happen O(1)
times in total, so the cost is immaterial.
We have α log n positions where we need to consider the k errors, yielding an
k
k
n
overall time complexity of O(nσ k α log
log log n) = O(n (ασ)
k! log n log log n).
k
We thus obtain the following result.
Theorem 2. Problem PLCP with k-Errors can be solved in average-case
k
time O(n ck! logk n log log n), where c is a constant, using O(n) extra space.
k
≤ (ασ)ασ = O(1) and hence the required time
Remark 1. We have that (ασ)
k!
is bounded by O(n logk n log log n).
Remark 2. If α log n > w, where w is the word size in the word-RAM model,
we can make use of the deterministic data structure presented in [6] (Theorem 1
therein), which can be built in O(n) time for a string x of length n and answers
predecessor queries (i.e. given a query string p, it returns the lexicographically
σ
largest suffix of x that is smaller than p) in time O( |p| log
+log |p|+log log σ). In
w
particular, the queries in scope can be answered in time O(log log n) per query.
3.3
Edit Distance
We next consider computing PLCPk under the edit distance model; however in
this case we observe that x[i . . n − 1] and x[i + j . . n − 1] are at edit distance j
for i − k ≤ j ≤ i + k. We hence alter the definition so that PLCPk [i] refers to
the longest common prefix of x[i . . n − 1] with k-errors occurring at a position
j∈
/ Si,k = {i − k, . . . , i + k}.
The proof of Theorem 3 can be extended to allow for k-errors under the
3k
edit distance. In this case we have that Pr(Xi,j ≥ m) ≤ m
k σm−k ; this can be
seen by following the same reasoning as in the first claim of the proof with two
extra considerations: (a) each deletion/insertion operation conceptually shifts
the letters to be matched (giving the 3k factor); (b) the letters to be matched
are m minus the number of deletions and substitutions and hence at least
√ m − k.
The extra 3k factor gets consumed by (2/3)5k later in the proof since 2 5 3/3 < 1.
On the technical side, we modify the algorithm of Section 3.2 as follows:
8
x[j . . n − 1]
j
j+k
n−1
i
i+k
x[i . . n − 1]
n−1
Fig. 2: We need to perform k extension steps in 2k + 1 diagonals of the dynamic
programming matrix for x[i . . n − 1] and x[j . . n − 1].
1. At each position, except for σ − 1 substitutions, we also consider σ insertions
and 1 deletion. This yields a multiplicative 2k factor in the time complexity.
We keep counters ins for insertions and del for deletions; for each length we
obtain, we add del and subtract ins.
2. When querying for a string z while processing position i we now have to
check that we do not return a position j ∈ Si,k . We can resolve this by
spending O(k) time for each position i; when we start processing position
i, we create an array of size O(k) that stores for each position j ∈ Si,k a
position fj ∈
/ Si,k with the maximal longest common prefix with x[j . . n − 1]
using the SA and the LCP array. When a query returns a position j ∈ Si,k
we instead consider fj .
3. We replace the LCE queries used to compute values in PLCPk longer than λ
(that required O(k) time in total) by the Landau-Vishkin technique [17] to
perform extensions. For an illustration inspect Figure 2. We initiate 2k + 1
diagonal paths in the classical dynamic programming matrix for x[i . . n − 1]
and x[j . . n−1]. The ith diagonal path above and the ith diagonal path below
the main diagonal are initialised to i errors. The path starting at the main
diagonal is initialised to 0 errors. We first perform an LCE query between
x[i . . n − 1] and x[j + d . . n − 1], for all 0 ≤ d ≤ k, and an LCE query between
x[i + d . . n − 1] and x[j . . n − 1], for all 1 ≤ d ≤ k. Then, for all 1 ≤ d ≤ k, we
try to extend a path with exactly d− 1 errors to a path with exactly d errors.
We perform an insertion, a deletion, or a substitution with a further LCE
query and pick the farthest reaching extension. The bottom-most extension
of any diagonal when d = k specifies the length of the longest common prefix
with k-errors. The whole process takes time O(k 2 ).
9
4
Genome Mappability Data Structure
The genome mappability problem has already been studied under the Hamming
distance model [8, 21, 4]. We can also define the problem under the edit distance
model. Given a string x of length n and integers m < n and k < m, we are
asked to count, for each length-m substring x[i . . i + m − 1] of x, the number
occ of other substrings of x occurring at a position j ∈
/ Si,k = {i − k, . . . , i + k}
that are at edit distance at most k from x[i . . i + m − 1]. We then say that
this substring has k-mappability equal to occ. Specifically, we consider a data
structure version of this problem [3]. Given x and k, construct a data structure,
which, for a query value µ given on-line, returns the minimal value of m that
forces at least µ length-m substrings of x to have k-mappability equal to 0.
Theorem 4 ([3]). An O(n)-sized data structure answering genome mappability
queries in O(1) time per query can be constructed from PLCPk in time O(n).
By combining Theorem 2 with Theorem 4 we obtain the first efficient algorithm for the genome mappability data structure under the edit distance model.
5
Longest Common Substring with k-Errors
In the longest common substring with k-errors problem we are asked to find
the longest substrings of two strings that are at distance at most k. The Hamming distance version has received much attention due to its applications in
computational biology [27, 18, 26]. Under edit distance, the problem is largely
unexplored. The average k-error common substring is an alignment-free method
based on this notion for measuring string dissimilarity under Hamming distance;
we denote the induced distance by Distk (x, y) for two strings x and y (see [27]
for the definition). Distk (x, y) can be computed in time O(|x| + |y|) from arrays
Λx,y and Λy,x , defined as
Λx,y [i] =
max
(lcpk (x[i . . |x| − 1], y[j . . |y| − 1])).
0≤j≤|y|−1
A worst-case and a more practical average-case algorithm for the computation
of Λx,y have been presented in [25, 26]. This measure was extended to allow for
wildcards (don’t care letters) in the strings in [12]. Here we provide a natural
generalisation of this measure: the average k-error common substring under the
edit distance model. The sole change is in the definition of Λx,y [i]: except for
substitution, we also allow for insertion and deletion operations.
The algorithm of Section 3.3 can be applied to compute Λx,y under the edit
distance model within the same complexities. We start by constructing the y-fast
trie for y. We then do the queries for the suffixes of x; we now also check for an
exact match (i.e. for x[i . . i + α log(|x| + |y|) − 1]). We obtain the following result.
Theorem 5. Given two strings x and y of length at most n and a distance
threshold k, arrays Λx,y and Λy,x and Distk (x, y) can be computed in averagek
case time O(n ck! logk n log log n), where c is a constant, using O(n) extra space.
10
Remark 3. By applying Theorem 5 we essentially solve the longest common
substring with k-errors for x and y within the same complexities.
6
All-Pairs Suffix/Prefix Overlaps with k-Errors
Given a set of strings and an error-rate ǫ, the goal is to find, for all pairs of
strings, their suffix/prefix matches (overlaps) that are within distance k = ⌈ǫℓ⌉,
where ℓ is the length of the overlap [23, 28, 16].
Using our technique but only inserting prefixes of the strings in the y-fast
trie and querying for all starting positions (suffixes) in a similar manner as in
Section 3.1, we obtain the following result.
Theorem 6. Given a set of strings of total length n and a distance threshold k, the length of the maximal longest suffix/prefix overlaps of every string
against all other strings within distance k can be computed in average-case time
k
O(n ck! logk n log log n), where c is a constant, using O(n) extra space.
References
1. Amir Abboud, Ryan Williams, and Huacheng Yu. More applications of the polynomial method to algorithm design. In SODA, SODA ’15, pages 218–230. Society
for Industrial and Applied Mathematics, 2015.
2. Mohamed Ibrahim Abouelhoda, Stefan Kurtz, and Enno Ohlebusch. Replacing
suffix trees with enhanced suffix arrays. Journal of Discrete Algorithms, 2(1):53–
86, 2004.
3. Hayam Alamro, Lorraine A. K. Ayad, Panagiotis Charalampopoulos, Costas S.
Iliopoulos, and Solon P. Pissis. Longest common prefixes with k-mismatches and
applications. In SOFSEM, volume 10706 of LNCS, pages 636–649. Springer International Publishing, 2018.
4. Mai Alzamel, Panagiotis Charalampopoulos, Costas S. Iliopoulos, Solon P. Pissis,
Jakub Radoszewski, and Wing-Kin Sung. Faster algorithms for 1-mappability of
a sequence. In COCOA, volume 10628 of LNCS, pages 109–121. Springer International Publishing, 2017.
5. Michael A. Bender and Martín Farach-Colton. The LCA problem revisited. In
LATIN, volume 1776 of LNCS, pages 88–94. Springer-Verlag, 2000.
6. Philip Bille, Inge Li Gørtz, and Frederik Rye Skjoldjensen. Deterministic indexing
for packed strings. In CPM, volume 78 of LIPIcs, pages 6:1–6:11. Schloss Dagstuhl
- Leibniz-Zentrum fuer Informatik, 2017.
7. Richard Cole, Lee-Ad Gottlieb, and Moshe Lewenstein. Dictionary matching and
indexing with errors and don’t cares. In STOC, STOC ’04, pages 91–100. ACM,
2004.
8. Thomas Derrien, Jordi Estellé, Santiago Marco Sola, David Knowles, Emanuele
Raineri, Roderic Guigó, and Paolo Ribeca. Fast computation and applications of
genome mappability. PLoS ONE, 7(1), 2012.
9. Johannes Fischer. Inducing the LCP-array. In WADS, volume 6844 of LNCS,
pages 374–385. Springer-Verlag, 2011.
10. Tomás Flouri, Emanuele Giaquinta, Kassian Kobert, and Esko Ukkonen. Longest
common substrings with k mismatches. Inf. Process. Lett., 115(6-8):643–647, 2015.
11
11. Szymon Grabowski. A note on the longest common substring with k-mismatches
problem. Inf. Process. Lett., 115(6-8):640–642, 2015.
12. Sebastian Horwege, Sebastian Lindner, Marcus Boden, Klas Hatje, Martin Kollmar, Chris-Andre Leimeister, and Burkhard Morgenstern. Spaced words and
kmacs: fast alignment-free sequence comparison based on inexact word matches.
Nucleic Acids Research, 42(Webserver-Issue):7–11, 2014.
13. S. Karlin, G. Ghandour, F. Ost, Tavare S., and L. J. Korn. New approaches
for computer analysis of nucleic acid sequences. In Proceedings of the National
Academy of Sciences USA, volume 80, pages 5660–5664, 1983.
14. Tomasz Kociumaka, Jakub Radoszewski, and Tatiana A. Starikovskaya. Longest
common substring with approximately k mismatches. CoRR, abs/1712.08573,
2017.
15. Roman Kolpakov, Ghizlane Bana, and Gregory Kucherov. mreps: efficient and
flexible detection of tandem repeats in DNA. Nucleic Acids Research, 31(13):3672–
3678, 2003.
16. Gregory Kucherov and Dekel Tsur. Improved filters for the approximate suffixprefix overlap problem. In SPIRE, volume 8799 of LNCS, pages 139–148. Springer,
2014.
17. Gad M. Landau and Uzi Vishkin. Efficient string matching in the presence of
errors. In IEEE, editor, FOCS, pages 126–136. IEEE Computer Society, 1985.
18. Chris-Andre Leimeister and Burkhard Morgenstern. Kmacs: the k -mismatch average common substring approach to alignment-free sequence comparison. Bioinformatics, 30(14):2000–2008, 2014.
19. Kung-Hao Liang. Bioinformatics for Biomedical Science and Clinical Applications.
Woodhead Publishing Series in Biomedicine. Woodhead Publishing, 2013.
20. Udi Manber and Eugene W. Myers. Suffix arrays: A new method for on-line string
searches. SIAM J. Comput., 22(5):935–948, 1993.
21. Giovanni Manzini. Longest common prefix with mismatches. In SPIRE, volume
9309 of LNCS, pages 299–310. Springer, 2015.
22. Ge Nong, Sen Zhang, and Wai Hong Chan. Linear suffix array construction by
almost pure induced-sorting. In DCC, IEEE, pages 193–202, 2009.
23. Kim R. Rasmussen, Jens Stoye, and Eugene W. Myers. Efficient q-gram filters for
finding all epsilon-matches over a given length. Journal of Computational Biology,
13(2):296–308, 2006.
24. Arian FA Smit. Interspersed repeats and other mementos of transposable elements
in mammalian genomes. Current Opinion in Genetics & Development, 9(6):657–
663, 1999.
25. Sharma V. Thankachan, Alberto Apostolico, and Srinivas Aluru. A provably efficient algorithm for the k -mismatch average common substring problem. Journal
of Computational Biology, 23(6):472–482, 2016.
26. Sharma V. Thankachan, Sriram P. Chockalingam, Yongchao Liu, Alberto Apostolico, and Srinivas Aluru. ALFRED: A practical method for alignment-free distance computation. Journal of Computational Biology, 23(6):452–460, 2016.
27. Igor Ulitsky, David Burstein, Tamir Tuller, and Benny Chor. The average common
substring approach to phylogenomic reconstruction. Journal of Computational
Biology, 13(2):336–350, 2006.
28. Niko Välimäki, Susana Ladra, and Veli Mäkinen. Approximate all-pairs suffix/prefix overlaps. Inf. Comput., 213:49–58, 2012.
29. Dan E. Willard. Log-logarithmic worst-case range queries are possible in space
theta(n). Inf. Process. Lett., 17(2):81–84, 1983.
12
| 8cs.DS
|
Article
A Cost/ Speed/ Reliability Tradeoff to Erasing†
Manoj Gopalkrishnan
Received: date ; Accepted: date ; Published: date
Academic Editor: name
1
arXiv:1410.1710v3 [cs.SY] 22 Apr 2016
†
Tata Institute of Fundamental Research; [email protected]
This paper is an extended version of the paper entitled ”Cost / speed / reliability tradeoff in erasing
a bit”, presented at the 14th International Conference on Unconventional and Natural Computation,
Auckland, New Zealand
Abstract: We present a KL-control treatment of the fundamental problem of erasing a bit. We
introduce notions of reliability of information storage via a reliability timescale τr , and speed of
erasing via an erasing timescale τe . Our problem formulation captures the tradeoff between speed,
reliability, and the Kullback-Leibler (KL)
to erase a bit. We show that rapid erasing of
cost required
τe
a reliable bit costs at least log 2 − log 1 − e− τr
r
> log 2, which goes to 12 log 2τ
τe when τr >> τe .
Keywords: Erasing; Information; Thermodynamics; Kullback-Leibler; Optimal control; Reliability;
Speed; Tradeoff
1. Motivation
Biological systems are remarkably ordered at multiple scales and dimensions, from the spatial
order witnessed in the packing of DNA inside the nucleus, the arrangement of cells to form tissues,
and organs, and whole organisms, to the temporal order witnessed in the execution of various cellular
processes. Superficially such order might appear to violate the second law of thermodynamics which
requires an increase in disorder with overwhelmingly high probability. In fact, there is no violation
since biological systems expend energy to bring about and maintain this increase in order.
We would like to understand this “energy to order” conversion quantitatively. What are the
fundamental limits to this conversion? Order can be measured in terms of information, by counting
the number of bits required to describe that order. From this point of view, understanding how
much energy is required to create order becomes an instance of the investigation of the connection
between information processing and thermodynamics. The basic information processing operation
that increases order is the operation of “erasing” or resetting a bit to state 0. To fix ideas, imagine
erasing random chalk marks from a blackboard, to leave it in a neat and ordered state.
Szilard [1] and later Landauer [2] have argued from the second law of thermodynamics that
erasing at temperature T requires at least k B T log 2 units of energy, where k B is Boltzmann’s constant.
The Szilard engine is a simple illustration of this result. Imagine a single molecule of ideal gas in a
cylindrical vessel. If this molecule is in the left half of the vessel, think of that as encoding the bit “0,”
and the bit “1” otherwise. Erasing this Brownian bit corresponds to ensuring that the molecule lies on
the left half, for example by compressing the ideal gas to half its volume. For a heuristic analysis we
may use the ideal gas law PV = k B T, integrating the expression dW = − PdV for work from limits
V to V/2 to obtain W = k B T log 2. More rigorous and general versions of this calculation are known,
which also clarify why this is a lower bound [3–5].
In practice, one finds that both man-made and biological instrumentation often require energy
substantially more than k B T log 2 to perform erasing [6,7]. John von Neumann remarked on this
large gap in his 1949 lectures at the University of Illinois [8]. (Bennett [9] has remarked that DNA
polymerases come close to the k B T log 2 bound. To copy a single base, a DNA polymerase hydrolyzes
Entropy 2016, xx, 1–x; doi:10.3390/——
www.mdpi.com/journal/entropy
Entropy 2016, xx, 1–x
a triphosphate molecule to a monophosphate, which provides close to 18k B T at temperature T =
300K. Note that this is still almost two orders of magnitude away from k B T log 2. Further, it is not
clear whether the comparison is valid at all since copying and erasing are different operations.)
How does one explain this large gap? Note that the result of k B T log 2 holds only in the
isothermal limit, which takes infinite time. In practice, we want erasing to be performed rapidly, say
in time τe , which requires extra entropy production. For intuition, suppose one wants to compress a
gas in finite time τe . The gas heats up, and pushes back, increasing the work required.
Several groups [10–12] have recognized that rapid erasing requires entropy production which
pushes up the cost of erasing beyond k B T log 2, and have obtained bounds for this problem. A
grossly oversimplified, yet qualitatively accurate, sketch of these various results is obtained by
considering the energy cost of compressing the Szilard engine rapidly. Specializing a result from
finite-time
thermodynamics
[13] to the case of the Szilard engine, one obtains an energy cost
k log 2
1 + στe −B k log 2 k B T log 2 where σ is the coefficient of heat conductivity of the vessel.
B
The bounds obtained by such considerations depend on technological parameters like the heat
conductivity σ, and not just on fundamental constants of physics and the requirement specifications
of the problem. If one varies over the technological parameters as well, e.g. allowing σ → ∞, the
energy cost tends to k B T log 2. Does there exist a more fundamental analysis for the cost of erasing
that is independent of technological parameters, and improves on k B T log 2? This is the open question
we address in this paper.
Our contribution: We follow up on von Neumann’s suggestion [8] that the gap was “due to
something like a desire for reliability of operation.” Swanson [14] and Alicki [15] have also looked
into issues of reliability. We introduce the notion of “reliability timescale” τr , and explicitly consider
the three-way trade-off between speed, reliability, and cost.
The other novelty of our approach is in bringing the tools of Kullback-Leibler (KL) control [16,17]
to bear on the problem of erasing a bit. The intuitive idea is that the control can reshape the dynamics
as it pleases, but pays for the deviation from the uncontrolled dynamics. The cost of reshaping
the dynamics is a relative entropy or KL divergence between the controlled and the uncontrolled
dynamics, expressed as measures on path space.
We find theoptimal control
for rapid erasing of a reliable bit, and argue that it requires cost of at
τe
r
least log 2 − log 1 − e− τr > log 2, which goes to 21 log 2τ
τe when τr >> τe . Importantly, our answer
does not depend on any technological parameters, but only on the requirement specifications τr and
τe of the problem.
2. The Erasing Problem
As a model of a bit, consider a two-state continuous-time Markov chain with states 0 and 1 and
the passive or uncontrolled dynamics given by transition rates k01 from state 0 to state 1 and k10 from
state 1 to state 0.
0
k01
k10
1
The transition rates k01 and k10 model spontaneous transitions between the states when no one is
looking at the bit or trying to erase it. The time independence of these rates represents the physical
fact that the system is not being driven.
Such finite Markov chain models often arise in physics by “coarse-graining.” For example, for
the case of the Szilard engine, the transition rate k10 models the rate at which the molecule enters the
left side, conditioned on it currently being on the right side.
Apart from their importance in approximating the behavior of real physical systems, finite
Markov chains are also important to thermodynamics from a logical point of view. They may be
viewed as finite models of a mathematical theory of thermodynamics. The terms “theory” and
2
Entropy 2016, xx, 1–x
“model” are to be understood in their technical sense as used in mathematical logic. We develop
this remark no further here since doing so would take us far afield.
Suppose the distribution at time t is ( p0 (t), p1 (t)) with p1 (t) = 1 − p0 (t). Then the time evolution
of the bit is described by the ODE
ṗ0 (t) = −k01 p0 (t) + k10 (1 − p0 (t)).
(1)
Setting π0 = k10 /(k01 + k10 ) and the reliability timescale τr := 1/(k01 + k10 ), this admits the solution
p0 (t) = π0 + e−t/τr ( p0 (0) − π0 )
(2)
Here τr represents the time scale on which memory is preserved. The smaller the rates k01 and k10 ,
the larger is the value of τr , and the slower the decay to equilibrium, so that the system remembers
information for longer.
Fix a required erasing time τe . Fix p(0) = π0 . We want to control the dynamics with transition
rates u01 (t) and u10 (t) to achieve p(τe ) = (1, 0), where
ṗ0 (t) = −u01 (t) p0 (t) + u10 (t)(1 − p0 (t))
(3)
∗ ( t ) and u∗ ( t ) to achieve this objective, according
We want to find the cost of the optimal protocol u01
10
to a cost function which we introduce next. In particular, when k01 = k10 , the equilibrium distribution
π = (π0 , 1 − π0 ) takes the value (1/2, 1/2) and we can interpret this task as erasing a bit of reliability
τr = 1/(k01 + k10 ) in time τe .
2.1. Kullback Leibler Cost
Define the path space P := {0, 1}[0,τe ] of the two-state Markov chain. This is the set of all paths
in the time interval [0, τe ] that jump between states 0 and 1 of the Markov chain. Each path can also
be succinctly described by its initial state, and the times at which jumps occur. We can also effectively
think of the path space as the limit as h → 0 of the space Ph := {0, 1}{0,h,2h,...,Nh=τe } corresponding to
the discrete-time Markov chain that can only jump at clock ticks of h units.
Once the rates u01 (t), u10 (t) and the initial distribution p(0) = p for the Markov chain are fixed,
there is a unique probability measure µu,p on path space which intuitively assigns to every path the
probability of occurrence of that path according to the Markov chain evolution (Equation 3) with
initial conditions p.
h for a single path i =
For pedagogic reasons, we first describe the discrete-time measure µu,p
(i0 , i1 , . . . , i N ) ∈ Ph . First we describe the transition probabilities of the discrete-time Markov chain.
For a, b ∈ {0, 1} with a 6= b, for all times t, define uhaa (t) := 1 − hu ab (t) and uhab (t) := hu ab (t) as the
probability of jumping to a and to b respectively in the time step t, conditioned on being in state a.
Then the probability of the path i under control u is given by:
N −1
µu,p (i ) := pi0
∏ uihj ,ij+1 ( jh)
j =0
3
Entropy 2016, xx, 1–x
Figure 1. The discrete-time path space Ph . A specific path is labeled in red.
h by sending
We describe the continous-time case now. We could obtain the measure µu,p from µu,p
h → 0, but it can also be described more directly. Fix i0 ∈ {0, 1}, and consider the set of paths S =
Si0 ,t1 ,t2 ,...,tn starting at i0 with jumps occurring around times t1 < t2 < · · · < tn within infinitesimal
intervals dt1 , dt2 , . . . , dtn and leading to the trajectory (i0 , i1 , . . . , in ) ∈ {0, 1}n+1 . Setting t0 = 0:
µu,p (S) := pi0
n −1 − R t j +1 u
i
t
∏e
j
j i j +1
(s)ds
ui j i j+1 (t j+1 )dt j+1
j =0
R t1
where pi0 is the probability of starting at i0 , e− 0 ui0 i1 (s)ds is the probability of not jumping in the time
interval (0, t1 ), ui0 i1 (t1 )dt1 is the probability of jumping from i0 to i1 in the interval (t1 , t1 + dt1 ) and
so on. This is the well-known Feynam-Kac formula for this Markov chain.
Specializing to u01 (t) = k01 and u10 (t) = k10 , we obtain the probability measure µk,p induced on
P by the passive dynamics (Equation 1) with initial conditions p.
We declare the Kullback Leibler (KL) cost D (µu,p k µk,p ) as the cost for implementing the control
u. More generally, for a physical system with path space P , passive dynamics corresponding to a
measure ν on P , and a controlled dynamics with a control corresponding to a measure µ on P , we
declare D (µkν) as the cost for implementing the control. This cost function has been widely used in
control theory [16–27]. In Section 4 we will explore some other interpretations of this cost function.
3. Solution to the Erasing Problem
Out of all controls u(t), we want to find a control u∗ (t) that starts from
k10
k01
p (0) = π =
,
(k01 + k10 k01 + k10
and achieves p(τe ) = (1, 0) while minimizing the relative entropy D (µu∗ ,π k µk,π ).
To find: u∗ = arg inf D (µu,π k µk,π )
u
Subject to: µu,π (τe ) = (1, 0)
Question ?? can be described within the framework of a well-studied problem in optimal control
theory that has a closed-form solution [16,17,28]. Following Todorov [16], we introduce the optimal
cost-to-go function v(t) = (v0 (t), v1 (t)). We intend vi (t) to denote the expected cumulative cost for
starting at state i at time t < τe , and reaching a distribution close to (1, 0) at time τe .
To discourage the system from being in state 1 at time τe , define v1 (τe ) = +∞ and v0 (τe ) = 0.
Suppose the control performs actions u01 (t) and u10 (t) at time t. Fix a small time h > 0. Define the
transition probability uijh (t) as the probability that a trajectory starting in state i at time t will be found
4
Entropy 2016, xx, 1–x
in state j at time t + h. When i 6= j, uijh (t) ≈ huij (t), whereas uiih (t) ≈ 1 − uijh (t) ignoring terms of size
O(h2 ). We define k hij similarly.
Let “log” denote the natural logarithm. To derive the law satisfied by the optimal cost-to-go v(t),
we approximate v(t) by the backward recursion relations:
"
v0 (t) = min E vi (t + h) + log
u01 (t)
"
v1 (t) = min E vi (t + h) + log
u10 (t)
h (t)
u0i
#
h
k0i
h (t)
u1i
(4)
#
h
k1i
h ( t ), u h ( t )), and the second is over i ∼
h
h
where the first expectation is over i ∼law (u00
law ( u10 ( t ), u11 ( t )),
01
and the approximation ignores terms of size O(h2 ). As h → 0 the second terms E log
uhji (t)
k hji
approach
the relative entropy cost in path space over the time interval (t, t + h).
In words, Equation 4 says that the cost-to-go from state 0 at time t equals the cost of the control
u(t) plus the expected cost-to-go in the new state i reached at time t + h. The cost of the control is
measured by relative entropy of the control dynamics relative to the passive dynamics, over the time
interval (t, t + h).
Define the desirability z0 (t) = e−v0 (t) and z1 (t) = e−v1 (t) . Define
h
h
G0 [z](t) = k00
z0 (t) + k01
z1 ( t ),
h
h
G1 [z](t) = k10
z0 (t) + k11
z1 ( t ).
We can rewrite Equation 4 as
"
log z0 (t) = log G0 [z](t + h) − min E log
h ( t ) G [ z ]( t + h )
u0i
0
h z (t + h)
k0i
i
u01 (t)
"
log z1 (t) = log G1 [z](t + h) − min E log
#
h ( t ) G [ z ]( t + h )
u1i
1
u10 (t)
#
(5)
h z (t + h)
k1i
i
Since the last term is the relative entropy of (uhj0 (t), uhj1 (t)) relative to the probability distribution
(khj0 z0 (t + h)/Gj [z](t + h), khj1 z1 (t + h)/Gj [z](t + h)), its minimum value is 0, and is achieved by the
protocol u∗ given by:
u∗ji (t)
k ji
= lim
h →0
e− v i ( t + h )
e− v i ( t )
= −v (t)
Gj [z](t + h)
e j
(6)
when i 6= j.
It remains to solve for z(t) and the optimal cost. From Equation 5, at the optimal control u∗ the
desirability z(t) must satisfy the equation − log z(t) = − log G [z](t + h) + 0, so that:
z0 ( t )
z1 ( t )
!
=
1 − k01 h
k10 h
k01 h
1 − k10 h
5
!
z0 ( t + h )
z1 ( t + h )
!
Entropy 2016, xx, 1–x
which simplifies to
dz
dt
= −Kz in the limit h → 0, where K =
−k01
k10
k10
−k10
is the infinitesimal
Kt
generator
of the Markov chain. This equation has the formal solution z(τe − t) = e z(τe ) where
1
z(τe ) = 0 . In the symmetric case k01 = k10 ,
1/2
1/2
z(t) =
!
+e
t
− τeτ−
r
1/2
−1/2
!
where τr = 1/(k01 + k10 ). Substituting t = 0 and taking logarithms, we find the cost-to-go function
at time 0:
!
−τe /τr
log
1
+
e
log 2
v (0) =
−
log 2
log 1 − e−τe /τr
When ν(0) = (1/2, 1/2) with k01 = k10 , the cost Cerase (τr , τe , T ) required for erasing a bit of
reliability τr = 1/(k01 + k10 ) in time τe is at least:
log 2 −
1
log 1 − e−2τe /τr
2
(7)
Note that Cerase ≥ log 2 with equality when τe /τr → ∞, since 1 − e−2τe /τr ≤ 1.
r
From Equation 7, Cerase ≥ 21 log 2τ
τe when τr >> τe .
4. Interpreting the KL cost
One motivation for our cost function comes from the field of KL control theory. We now compare
other possible meanings to this cost function.
4.1. Path space Szilard-Landauer correspondence
The correspondence between information and thermodynamics was revealed in the work of
Szilard, and clarified by Landauer. More rigorous and general treatments of this correspondence
have been worked out recently [3–5]. We first recall this result, and then show how our cost function
is a formal extension of this result.
Consider a physical system with finite state space S and energy E : S → R. (More general state
spaces S can be handled by replacing the sum by an appropriate integral. For our present purposes,
it suffices to assume S is finite.) Define the Gibbs distribution π at temperature T by
π (i ) =
e− Ei /k B T
∑ j∈S e
− Ej /k B T
for all i ∈ S. Define the free energy
F ( p) :=
1
∑ pi Ei − k B T ∑ pi log pi
i ∈S
i ∈S
where p is a probability distribution.
p
Define the relative entropy D ( pkq) = ∑i∈S pi log q i with Euler’s constant for the base of
i
the logarithm. Following Jaynes [29], assume that equilibrium π corresponds to a maximally
uninformative state of the system, so that we have zero information about the system when it is
at equilibrium. Recall that a nat is the unit of information when logarithms are taken to the base of
Euler’s constant. 1 bit = log 2 nats. Then the relative entropy D ( pkπ ) has an axiomatic identification
with the amount of information in nats that we know about the system when it is in a nonequilibrium
state p [4].
6
Entropy 2016, xx, 1–x
The following identity is easily verified:
F ( p) − F (π ) = k B TD ( pkπ ).
(8)
The conceptual significance of this simple identity is that it supplies a dictionary between
thermodynamics and information theory [4]. In particular, erasing a bit corresponds to increasing
relative entropy which in turn corresponds — via the identity — to increasing available free energy
F ( p) − F (π ) by k B T log 2, recovering the classical result of Szilard as an alternative statement of the
second law of thermodynamics. In the other direction, charging a battery corresponds to increasing
available free energy which in turn corresponds — via Identity 8 — to erasing of information. This
relates the energy efficiency of charging a battery to the energy required to erase a bit.
Now consider our cost function D (µkν). The relative entropy D (µkν) counts the number of nats
erased by the control in path space, relative to the passive dynamics. Since the Szilard-Landauer
principle asserts that erasing one bit requires at least k B T log 2 units of energy, our cost function may
be viewed as a Path Space Szilard-Landauer Principle, formally extending Identity 8 to path space.
4.2. Thermodynamic interpretation
We wish to compare the cost D (µkν) with the usual thermodynamic expected work ∆W. We will
quickly outline how thermodynamic quantities can be defined for a two-state Markov chain.
4.2.1. Thermodynamics on a two-state Markov chain
The ideas we present here are well-known in the nonequilibrium thermodynamics community,
for example see Propp’s thesis [30]. The construction can be carried out more generally, but the
generalization is not necessary for our present purposes.
1. Consider again the two-state continuous-time Markov chain with passive dynamics given by
transition rates k01 and k10 .
k01
0
1
k10
Let E0 and E1 denote the internal energy of states “0” and “1” respectively. Then the equilibrium
distribution is given by π0 ∝ e− E0 /k B T and π1 ∝ e− E1 /k B T . We also have k01 π0 = k10 π1 from
detailed balance. Together this yields
E0 − E1 = k B T log
k01
.
k10
(9)
2. Now consider the same two-state system with a control applied to it by means of a field of
potential φ(t) = (φ0 (t), φ1 (t)) so that the potential energy in state i becomes Ei + φi (t). The
transition rates due to the control become u01 (t) and u10 (t). By a reasoning similar to how we
derived Equation 9, we get
E0 + φ0 − E1 − φ1 = k B T log
u01
.
u10
Combining with Equation 9 this yields
φ0 − φ1 = k B T log
7
u01 k10
u10 k01
(10)
Entropy 2016, xx, 1–x
3. Given a distribution p = ( p0 , p1 ) on the states, we can define the following thermodynamic
quantities:
• Expected internal energy E( p) = p0 E0 + p1 E1 .
• Entropy S( p) = − p0 log p0 − p1 log p1 .
• Nonequilibrium free energy F ( p) = E( p) − k B TS( p).
4. Given a transition from state i to state j in the presence of the control field, we can define the
following thermodynamic quantities:
• Heat dissipated Qij (t) = Ei + φi (t) − Ej − φj (t).
• Work done by the control Wij (t) = φi (t) − φj (t). This expression for work can be
traced back to Sekimoto [31], and is commonly employed in the field of Stochastic
Thermodynamics to describe the work done by switching on a control field [32].
p (t)
• Entropy increase of the system Sij (t) = log pi (t) .
j
The first law of thermodynamics manifests as Wij = Qij + Ei − Ej .
5. Suppose the system is described at time t by a distribution p(t) = ( p0 (t), p1 (t)). Define the
Current Jij (t) = pi (t)uij (t) − p j (t)u ji (t) so that ṗ0 (t) = − J01 (t).
6. We can further compute
dE
dt
dW
dt
dQ
dt
dS
dt
dF
dt
= J01 (t)( E1 (t) − E0 (t)) = k B T J01 (t) log
k10
k01
= J01 (t)W01 (t) = J01 (t)(φ0 (t) − φ1 (t)) = k B T J01 (t) log
u01 (t)k10
u10 (t)k01
= J01 (t) Q01 (t) = J01 (t)( E0 + φ0 (t) − E1 − φ1 (t)) = k B T J01 (t) log
= J01 (t)S01 (t) = J01 (t) log
= k B T J01 (t) log
u01 (t)
u10 (t)
p0 ( t )
p1 ( t )
k10 p1 (t)
k01 p0 (t)
7. Define Total Entropy Production Stot (t) to be the total entropy produced from time 0 to time t.
In other words, Stot (0) = 0 and
1 dQ dS
dStot (t)
=
+ .
dt
k B T dt
dt
After simplification,
dStot (t)
p (t)u01 (t)
= ( p0 (t)u01 (t) − p1 (t)u10 (t)) log 0
≥0
dt
p1 (t)u10 (t)
(11)
which is a statement of the second law of thermodynamics.
8. The following identity is immediate
dW
dF dStot
=
+
dt
dt
dt
and is another form of the first law.
4.2.2. Thermodynamic cost for rapid erasing of a reliable bit
How much does it cost for rapid erasing of a reliable bit, with the cost function equal to ∆W? We
claim that it costs k B log 2. In particular, neither the reliability timescale τr nor the erasing timescale
τe appear in this answer.
8
Entropy 2016, xx, 1–x
Suppose we can erase a (τr , τe ) bit for work W. First note that dW
dt is a function of k 01 /k 10 ,
u01 (t)/u10 (t) and J01 (t) as in (6). In particular, simultaneously sending the rates k01 and k10 as low
as possible while keeping their ratio the same has no effect on the work. So if we can erase a (τr , τe )
bit for work W, then we can erase a ( Aτr , τe ) bit for work W, for an arbitrarily large constant A. In
particular it is enough for us to demonstrate a protocol when τr = 1.
Now note that dW
dt depends only on the ratio u01 ( t ) /u10 ( t ) and not on the actual values of the
rates. We can also erase a (τr , τe /2) bit for work W by taking the (τr , τe )-protocol (uij (t)) and defining
new rates vij (t) = 2uij (2t). Since v01 (t)/v10 (t) = u01 (2t)/u10 (2t), it follows from a simple calculation
that the work required does not change.
By taking a limit of this time scaling argument, we only need to erase a (1, +∞) bit. Here the
infinite-time isothermal protocol, which proceeds by raising the ‘1’ well infinitesimally, waiting for
the system to equilibrate, and repeating, erases for a total work of k B T log 2 since that is the free
energy difference between the initial and final state, and there is no extra dissipation. This establishes
our claim.
A more detailed version of this calculation can be found in [33]. This work assumes that there
is a maximum energy limit Emax to which a state can be raised, so that there will be some small
error to erasing. It also makes another assumption about thermalization timescale which translates in
our setting to assuming that there is a maximum value to the rates u01 (t) and u10 (t). With these
assumptions, they show that the cost of rapid erasing is slightly more than k B log 2 and goes to
k B T log 2 very quickly as the timescale of thermal relaxation becomes smaller and Emax goes to infinity.
4.2.3. Link between KL-cost and thermodynamic work
We will now characterize entropy production in terms of time reversal. This will allow us to
make a link between KL-cost and the thermodynamic work W.
We first recall the notion of time reversal of a Markov chain. Usually time-reversal is defined for
time-homogeneous Markov chains. However, for the purposes of characterizing entropy production
in terms of time reversal, we will work with a definition that applies to time-inhomogeneous Markov
chains also. Instead of giving this definition in full generality, we work with a Markov chain with
a finite state space. This is sufficient for our purposes, and allows us to avoid dealing with certain
technical issues.
Note that given a matrix U with positive non-diagonal entries and U · 1 = 0, there is a
nonnegative vector v such that v · U = 0. This can be shown by applying the Perron-Frobenius
theorem to the exponential matrix eU .
Definition 1 (Time-reversal). Consider a continuous-time time-inhomogeneous Markov chain with
state space [n] = {1, 2, . . . , n} described by a time-dependent transition matrix U (t) = (uij (t))i,j∈[n]
(so that at time t, uij (t) denotes the rate of jumping to state j given that the system is in state i). Let
π (t) be a sequence of stationary probability distributions on [n], i.e., π (t)U (t) = 0 and πi ≥ 0 for
all i ∈ [n] and ∑i πi = 1. Then the time-reversal Markov chain is described by the time-dependent
transition matrix Û (t) = (ûij (t))i,j∈[n] where
ûij (t) =
πi (t)uij (t)
πi ( t )
A Markov chain is reversible iff Û = U.
The justification for considering time-reversal comes from Bayes’ rule. Reversible Markov chains
are well-known to be characterized by the conditions of existence of a detailed balanced equilibrium,
as well as by the Kolmogorov chain conditions. In particular, two-state Markov chains are always
reversible.
9
Entropy 2016, xx, 1–x
For the special case of Equation 3 in particular, given a distribution q at time τe , the time reversal
Markov chain evolves in time according to the ODE:
q̇0 (t) = −u01 (t)q0 (t) + u10 (t)(1 − q0 (t)),
(12)
q(τe ) = q.
We see that the difference is that in Equation 3, the boundary condition was specified at time 0,
whereas here the boundary condition is specified at time τe .
We define the time-reversed measure µrev
u,q as the measure on path space corresponding to the
process described by Equation 12. Strictly speaking, we should write µrev
u,q,τe to denote the time at
which the boundary condition is provided to the differential equation, but we will avoid this by using
the convention that we’re always going to set the boundary condition at time τe when considering
the time-reversal.
The following result is key to our comparison.
Theorem 2. Run the control dynamics Equation 3 forward from initial condition p(0) upto time τe to obtain
the distribution p(τe ). Consider the measure µrev
. Then the total entropy production Stot (τe ) from time 0
u,p(τe )
to time τe equals
Stot (τe ) = D (µu,p(0) kµrev
u,p(τe ) ).
Proof. We will show that the time derivative of the RHS equals the rate of entropy production. This
will prove the theorem.
Fix a time t ∈ [0, τe ]. Let the probability distribution at time t be represented by p(t) =
( p0 (t), p1 (t)). Let f ij := pi (t)uij (t) denote the flow rate from state i to state j 6= i at time t.
Then pi (t + h) = pi (t) + h( f ji − f ij ) + o (h) where i 6= j and o (h) denotes terms g(h) such that
limh→0 g(h)/h → 0.
We will consider the probabilities of the four Markov chain transitions 0 → 0, 0 → 1, 1 → 0 and
1 → 1 in the interval (t, t + h) in the limit h → 0 according to µu,p(0) and according to µrev
. Up to
u,p(τ )
e
terms of size o (h2 ), we have for i 6= j:
µu,p(0) (i → j) = hpi (t)uij (t) = h f ij
µrev
u,p(τe ) (i → j ) = hp j ( t + h ) u ji ( t ) = h ( f ji + h ( f ij − f ji ))
The increment in the relative entropy in the time interval (t, t + h) equals, upto o (h2 ) terms:
h f 01
1 − h f 01
+ h f 01 log
1 − h f 10 − h2 ( f 01 − f 10 )
h f 10 + h2 ( f 01 − f 10 )
h f 10
1 − h f 10
+ h f 10 log
+ (1 − h f 10 ) log
h f 01 + h2 ( f 10 − f 01 )
1 − h f 01 − h2 ( f 10 − f 01 )
(1 − h f 01 ) log
The off-diagonal terms contribute:
(1 − h f 01 ) log
1 − h f 01
1 − h f 10
+ (1 − h f 10 ) log
1 − h f 10 − h2 ( f 01 − f 10 )
1 − h f 01 − h2 ( f 10 − f 01 )
≈ (1 + o (h))(h f 10 − h f 01 + o (h2 )) + (1 + o (h))(h f 01 − h f 10 + o (h2 ))
≈ o ( h2 )
10
Entropy 2016, xx, 1–x
Divide by h, and take the limit h → 0. We can ignore the off-diagonal terms. The diagonal terms sum
to the rate of entropy production dSdttot as in Equation 11, and we are done.
By the First Law of Thermodynamics and Theorem 2,
∆W = ∆F + k B TD (µu,p kµrev
u,p(τe ) )
(13)
where the increase in free energy of the system
∆F = k B T ( D ( p(τe )kπ ) − D ( p(0)kπ ))
by Equation 8. Now to compare our cost function with ∆W.
Theorem 3. The KL-cost equals change in free energy by k B T plus a path-space relative entropy term that
resembles entropy production:
k B TD (µu,p k µk,p ) = ∆F + k B TD (µu,p kµrev
k,p(τe ) )
(14)
where p(τe ) is — as in Equation 13 — the solution to the control dynamics Equation 3 at time τe .
Proof. Using Equation 8, we can rewrite the claim as
D (µu,p k µk,p ) + D ( p(0)kπ ) = D (µu,p kµrev
k,p(τe ) ) + D ( p ( τe )k π )
Both LHS and RHS equal D (µu,p k µk,π ). The assertion for the LHS is straightforward. The assertion
for the RHS is true because time-reversal dynamics was defined to keep the stationary distribution π
remain stationary under time reversal.
Comparing 13 and 14, a KL control treatment replaces the total entropy production
D (µu,p kµrev
) in 13 by the new term D (µu,p kµrev
) which compares the control dynamics with the
u,p(τe )
k,p(τe )
time reversal of the passive dynamics. This suggests an interpretation as follows. If we applied the
control during some time interval [0, τe ], and remembered what control we applied, then the entropy
production is correctly given by D (µu,p kµrev
). However, the information that a control was applied
u,p(τe )
also needs to be stored somewhere. If we forget that a control was applied, and if application of the
control is very rare, then our default model for the dynamics should be much closer to the passive
dynamics. In this case, entropy production may be closer to the value D (µu,p kµrev
).
k,p(τ )
e
4.3. Large deviations interpretation
Our cost function D (µkν) also admits a large deviation interpretation which was, remarkably,
already noted by Schrödinger in 1931 [34–37]. Motivated by quantum mechanics, Schrödinger asked:
conditioned on a more or less astonishing observation of a system at two extremes of a time interval,
what is the least astonishing way in which the dynamics in the interval could have proceeded?
Specializing to our problem of erasing, suppose an ensemble of two-state Markov chain with passive
dynamics given by Equation 1 was observed at time 0 and at time τe . Suppose the empirical state
distribution over the ensemble was found to be the equilibrium distribution π at time 0, and (1, 0) at
time τe respectively. This would be astonishing because no control has been applied, yet the ensemble
has arrived at a state of higher free energy. Conditioned on this rare event having taken place, what
is the least unlikely measure µ∗ on path space via which the process took place?
By a statistical treatment of multiple single particle trajectories, Schrödinger found that the
likelihood of an empirical measure µ on path space falls exponentially fast with the relative entropy
11
Entropy 2016, xx, 1–x
D (µkν) where ν is the measure induced by the passive dynamics. In particular, the least unlikely
measure µ∗ is that measure which — among all µ whose marginals at time 0 and time τe respect the
observations — minimizes D (µkν). So for the problem of erasing where k01 = k10 , the measure µ
varies over all measures that have marginal (1/2, 1/2) at time 0 and marginal (1, 0) at time τe , and µ∗
is that measure among all such µ that minimizes D (µkµk,(1/2,1/2) ). Thus our optimal control produces
in expectation the least surprising trajectory among all controls that perform rapid erasing.
4.4. Gibbs measure
Equation 6 is not accidental for this example, but is in fact a general feature when the cost
function is relative entropy [28]. More abstractly, the Radon-Nikodym derivative (i.e., “probability
d µ∗
density”) d ν of the measure µ∗ induced on path space by the optimal control u∗ with respect to
the measure ν induced by the passive dynamics is a Gibbs measure, with the cost-to-go function
v(t) playing the role of an energy function. In other words, mathematically our problem is precisely
the free energy minimization problem so familiar from statistical mechanics. There is also a possible
physical interpretation: we are choosing paths in P as microstates, instead of points in phase space.
The idea of paths as microstates has occurred before [38].
5. Concluding remarks
Since charging a battery can also be thought of as erasing a bit [4], our result may also hold
insights into the limits of efficiencies of rapidly charging batteries that must simultaneously hold
their energy for a long time.
So long as the noise is Markovian, we conjecture that the KL cost for erasing the two-state
Markov chain is a lower bound for more general cases – for example for bits with Langevin
dynamics [39], which is a stochastic differential equation expressing Newton’s laws of motion with
Brownian noise perturbations.
Acknowledgments: I thank Sanjoy Mitter, Vivek Borkar, Nick S. Jones, Mukul Agarwal, and Krishnamurthy
Dvijotham for helpful discussions. I thank Abhishek Behera for drawing Figure 1.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Leo Szilard. On the reduction of entropy in a thermodynamic system by the interference of intelligent
beings. Z Phys, 53:840–856, 1929.
Rolf Landauer. Irreversibility and heat generation in the computing process. IBM Journal of Research and
Development, 5:183–191, 1961.
Massimiliano Esposito and Christian Van den Broeck. Second law and Landauer principle far from
equilibrium. EPL (Europhysics Letters), 95(4):40004, 2011.
Manoj Gopalkrishnan. The Hot Bit I: the Szilard-Landauer correspondence. CoRR, abs/1311.3533, 2013.
David Reeb and Michael M Wolf. An improved Landauer principle with finite-size corrections. New Journal
of Physics, 16(10):103011, 2014.
Simon B Laughlin, Rob R de Ruyter van Steveninck, and John C Anderson. The metabolic cost of neural
information. Nature neuroscience, 1(1):36–41, 1998.
Trevor Mudge. Power: A first-class architectural design constraint. Computer, 34(4):52–58, 2001.
John von Neumann. Theory of self-reproducing automata, page 66. University of Illinois Press, Urbana, 1966.
lecture delivered at University of Illinois in December, 1949.
Charles H. Bennett. The thermodynamics of computation – a review. Int. J. Theo. Phys., 21(12):905–940,
1982.
Erik Aurell, Krzysztof Gawȩdzki, Carlos Mejía-Monasterio, Roya Mohayaee, and Paolo
Muratore-Ginanneschi. Refined second law of thermodynamics for fast random processes. Journal
of statistical physics, 147(3):487–505, 2012.
Giovanni Diana, G Baris Bagci, and Massimiliano Esposito. Finite-time erasing of information stored in
fermionic bits. Physical Review E, 87(1):012111, 2013.
12
Entropy 2016, xx, 1–x
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
Patrick R Zulkowski and Michael R DeWeese. Optimal finite-time erasure of a classical bit. Physical Review
E, 89(5):052140, 2014.
Peter Salamon and Abrahan Nitzan. Finite time optimizations of a Newton’s law Carnot cycle. The Journal
of Chemical Physics, 74:3546, 1981.
JA Swanson. Physical versus logical coupling in memory systems. IBM Journal of Research and Development,
4(3):305–310, 1960.
Robert Alicki. Information is not physical. arXiv preprint arXiv:1402.2414, 2014.
Emanuel Todorov. Efficient computation of optimal actions. Proceedings of the national academy of sciences,
106(28):11478–11483, 2009.
Wendell H Fleming and Sanjoy K Mitter. Optimal control and nonlinear filtering for nondegenerate
diffusion processes. Stochastics: An International Journal of Probability and Stochastic Processes, 8(1):63–77,
1982.
Hilbert J Kappen. Path integrals and symmetry breaking for optimal control theory. Journal of statistical
mechanics: theory and experiment, 2005(11):P11011, 2005.
Hilbert J Kappen. Linear theory for control of nonlinear stochastic systems. Physical review letters,
95(20):200201, 2005.
Evangelos A Theodorou. Iterative path integral stochastic optimal control: Theory and applications to motor
control. PhD thesis, University of Southern California, 2011.
Evangelos Theodorou and Emanuel Todorov. Relative entropy and free energy dualities: Connections to
path integral and KL control. In CDC, pages 1466–1473, 2012.
Freek Stulp, Evangelos A Theodorou, and Stefan Schaal. Reinforcement learning with sequences of motion
primitives for robust manipulation. Robotics, IEEE Transactions on, 28(6):1360–1370, 2012.
Krishnamurthy Dvijotham and Emanuel Todorov. A unified theory of linearly solvable optimal control.
Artificial Intelligence (UAI), page 1, 2011.
Hilbert J Kappen, Vicenç Gómez, and Manfred Opper. Optimal control as a graphical model inference
problem. Machine learning, 87(2):159–182, 2012.
Bart van den Broek, Wim Wiegerinck, and Bert Kappen. Graphical model inference in optimal control of
stochastic multi-agent systems. J. Artif. Intell. Res.(JAIR), 32:95–122, 2008.
Wim Wiegerinck, Bart van den Broek, and Hilbert Kappen. Stochastic optimal control in continuous
space-time multi-agent systems. arXiv preprint arXiv:1206.6866, 2012.
Matanya Benasher Horowitz. Efficient methods for stochastic optimal control. PhD thesis, California Institute
of Technology, 2014.
Paul Dupuis and Richard S Ellis. A weak convergence approach to the theory of large deviations, volume 902.
John Wiley & Sons, 2011.
Edwin T Jaynes. Information theory and statistical mechanics. Physical review, 106(4):620, 1957.
Michael Benjamin Propp. The thermodynamic properties of Markov processes. PhD thesis, Massachusetts
Institute of Technology, 1985.
Ken Sekimoto. Kinetic characterization of heat bath and the energetics of thermal ratchet models. Journal
of the physical society of Japan, 66(5):1234–1237, 1997.
Udo Seifert. Stochastic thermodynamics, fluctuation theorems and molecular machines. Rep. Prog. Phys.,
75(12):126001, 2012.
Browne, Cormac, Andrew JP Garner, Oscar CO Dahlsten, and Vlatko Vedral. Guaranteed energy-efficient
bit reset in finite time. Physical review letters, 113(10):100603, 2014.
E Schrödinger. Uber die umkehrung der naturgesetze, sitzung ber preuss. Akad. Wissen., Berlin Phys. Math,
144, 1931.
Arne Beurling. An automorphism of product measures. Annals of Mathematics, pages 189–200, 1960.
Hans Föllmer. Random fields and diffusion processes. In École d’Été de Probabilités de Saint-Flour XV–XVII,
1985–87, pages 101–203. Springer, 1988.
Robert Aebi. Schrödinger diffusion processes. Springer, 1996.
AD Wissner-Gross and CE Freer. Causal entropic forces. Physical review letters, 110(16):168702, 2013.
R. Zwanzig. Nonequilibrium Statistical Mechanics. Oxford University Press, USA, 2001.
13
Entropy 2016, xx, 1–x
c 2016 by the author; licensee MDPI, Basel, Switzerland. This article is an open
access article distributed under the terms and conditions of the Creative Commons by
Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).
14
| 3cs.SY
|
COARSENINGS, INJECTIVES AND HOM FUNCTORS
arXiv:1210.5325v1 [math.AC] 19 Oct 2012
FRED ROHRER
Abstract. It is characterized when coarsening functors between categories of graded modules preserve injectivity of objects, and when they
commute with graded covariant Hom functors.
Introduction
Throughout, groups and rings are understood to be commutative. By Ab,
Ann and Mod(R) for a ring R we denote the categories of groups, rings, and
R-modules, respectively. If G is a group, then by a G-graded ring we mean
a pair (R, (Rg )g∈G ) consisting of a ring R and a family (Rg )g∈G of subgroups
of the additive group of R whose direct sum equals the additive group of R
such that for g, h ∈ G it holds Rg Rh ⊆ Rg+h . If (R, (Rg )g∈G ) is a G-graded
ring, then by a G-graded R-module we mean a pair (M, (Mg )g∈G ) consisting
of an R-module M and a family (Mg )g∈G of subgroups of the additive group
of M whose direct sum equals the additive group of M such that for g, h ∈ G
it holds Rg Mh ⊆ Mg+h . If no confusion can arise then we denote a G-graded
ring (R, (Rg )g∈G ) just by R, and a G-graded R-module (M, (Mg )g∈G ) just by
M . Accordingly, for a G-graded ring R, a G-graded R-module M and g ∈ G
we denote by Mg the component of degree g of M . Given G-graded rings R
and S, by a morphism of G-graded rings from R to S we mean a morphism of
rings u : R → S such that u(Rg ) ⊆ Sg for g ∈ G, and given a G-graded ring R
and G-graded R-modules M and N , by a morphism of G-graded R-modules
from M to N we mean a morphism of R-modules u : M → N such that
u(Mg ) ⊆ Ng for g ∈ G. We denote by GrAnnG and GrModG (R) for a G-graded
ring R the categories of G-graded rings and G-graded R-modules, respectively,
with the above notions of morphisms. In case G = 0 we canonically identify
GrAnnG with Ann and GrModG (R) with Mod(R) for a ring R.
Let ψ : G ։ H be an epimorphism in Ab and let R be a G-graded ring.
We consider the ψ-coarsening R[ψ] of R, i.e., the H-graded ring whose underlying ring is the ring underlying R and whose component of degree h ∈ H is
L
g∈ψ−1 (h) Rg . An analogous construction for graded modules yields the ψcoarsening functor •[ψ] : GrModG (R) → GrModH (R[ψ] ), coinciding for H = 0
2010 Mathematics Subject Classification. Primary 13A02; Secondary 13C11.
Key words and phrases. Graded ring, graded module, coarsening, refinement, injective
module, small module, steady ring, graded Hom functor.
The author was supported by the Swiss National Science Foundation.
1
2
FRED ROHRER
with the functor that forgets the graduation. The aim of this note is to study
functors of this type.
Let M be a G-graded R-module. Some properties of M behave well under
coarsening functors – e.g., M is projective (in GrModG (R)) if and only if M[ψ]
is projective (in GrModH (R[ψ] )) –, but others do not. An example is injectivity.
For H = 0 it is well known that if M[ψ] is injective then so is M , but that
the converse does not necessarily hold. However, the converse does hold if
G is finite, as shown by Nǎstǎsescu, Raianu and Van Oystaeyen ([9]). We
generalize this to arbitrary H by showing that the converse holds if Ker(ψ) is
finite, and we moreover show that this is the best possible without imposing
further conditions on R or M (Theorem 2.4). One should note that finiteness
of Ker(ψ) is fulfilled if G is of finite type and ψ is the canonical projection
onto G modulo its torsion subgroup. Such coarsenings can be used to reduce
the study of graduations by groups of finite type to that of (often easier)
graduations by free groups of finite rank.
A further interesting question is whether coarsening functors commute with
graded Hom functors. The G-graded covariant Hom functor G HomR (M, •) of
M maps a G-graded R-module N onto the G-graded R-module
M
G
HomR (M, N ) =
HomGrModG (R) (M, N (g))
g∈G
(where •(g) denotes shifting by g). There is a canonical monomorphism of
G
H Hom
functors hM
R[ψ] (M[ψ] , •[ψ] ). For H = 0 this is
ψ : HomR (M, •)[ψ]
an isomorphism if and only if G is finite or M is small, as shown by Gómez
Pardo, Militaru and Nǎstǎsescu ([5]). We generalize this to arbitrary H by
showing that hM
ψ is an isomorphism if and only if Ker(ψ) is finite or M is small
(Theorem 3.7). A surprising consequence is that if hM
ψ is an isomorphism for
M
some epimorphism ψ with infinite kernel then hϕ is an isomorphism for every
epimorphism ϕ (Corollary 3.8).
The proofs of the aforementioned results are similar to and inspired by
those in [9] and [5]. In particular, they partially rely on the existence of
adjoint functors of coarsening functors, treated in the first section.
1. Coarsening functors and their adjoints
Let ψ : G ։ H be an epimorphism in Ab.
We first recall the definition of coarsening and refinement functors for rings
and modules, and the construction of some canonical morphisms of functors.
(1.1) A) For a G-graded ring R there is an H-graded ring R[ψ] with underlyL
ing ring the ring underlying R and with H-graduation ( g∈ψ−1 (h) Rg )h∈H . For
u : R → S in GrAnnG there is a u[ψ] : R[ψ] → S[ψ] in GrAnnH with underlying
map the map underlying u. This defines a functor •[ψ] : GrAnnG → GrAnnH ,
called ψ-coarsening.
COARSENINGS, INJECTIVES AND HOM FUNCTORS
3
B) For an H-graded ring S there is a G-graded ring S [ψ] with G-graduation
L
(Sψ(g) )g∈G , so that its underlying additive group is
g∈G Sψ(g) , and with
multiplication given by the maps Sψ(g) × Sψ(h) → Sψ(g)+ψ(h) for g, h ∈ G
induced by the multiplication of S. For v : S → T in GrAnnH there is a
[ψ]
v [ψ] : S [ψ] → T [ψ] in GrAnnG with vg = vψ(g) for g ∈ G. This defines a functor
•[ψ] : GrAnnH → GrAnnG , called ψ-refinement.
(1.2) A) For a G-graded ring R, the coproduct in Ab of the canonical inL
[ψ]
jections Rg
for g ∈ G is a monomorphism
f ∈ψ−1 (ψ(g)) Rf = (R[ψ] )g
G
[ψ]
αψ (R) : R (R[ψ] ) in GrAnn , and the coproduct in Ab of the restrictions
M
Rf ։ Rg
(R[ψ] )[ψ]
g =
f ∈ψ−1 (ψ(g))
Q
of the canonical projections f ∈ψ−1 (ψ(g)) Rf ։ Rg for g ∈ G is an epimorphism δψ (R) : (R[ψ] )[ψ] ։ R in GrAnnG . Varying R we get a monomorphism
αψ : IdGrAnnG (•[ψ] )[ψ] and an epimorphism δψ : (•[ψ] )[ψ] ։ IdGrAnnG .
B) For an H-graded ring S, the coproduct in Ab of the codiagonals
M
((S [ψ] )[ψ] )h =
Sh ։ Sh
g∈ψ−1 (h)
for h ∈ H is an epimorphism βψ (S) : (S [ψ] )[ψ] ։ S in GrAnnH . If Ker(ψ) is
finite, so that ψ −1 (h) is finite for h ∈ H, then the coproduct in Ab of the
L
Q
[ψ]
diagonals Sh g∈ψ−1 (h) Sh = g∈ψ−1 (h) Sg = ((S [ψ] )[ψ] )h for h ∈ H is a
monomorphism γψ (S) : S (S [ψ] )[ψ] in GrAnnH . Varying S we get an epimorphism βψ : (•[ψ] )[ψ] ։ IdGrAnnH and — if Ker(ψ) is finite — a monomorphism
γψ : IdGrAnnH (•[ψ] )[ψ] .
(1.3) A) Let R be a G-graded ring. For a G-graded R-module M there is
an H-graded R[ψ] -module M[ψ] with underlying R[0] -module the R[0] -module
L
underlying M and with H-graduation ( g∈ψ−1 (h) Mg )h∈H . For u : M → N
in GrModG (R) there is a u[ψ] : M[ψ] → N[ψ] in GrModH (R[ψ] ) with underlying
map the map underlying u. This defines an exact functor
•[ψ] : GrModG (R) → GrModH (R[ψ] ),
called ψ-coarsening.
B) Let S be an H-graded ring. For an H-graded S-module M there is a Ggraded S [ψ] -module M [ψ] with G-graduation (Mψ(g) )g∈G , so that its underlying
L
additive group is g∈G Mψ(g) , and with S [ψ] -action given by the maps
Sψ(g) × Mψ(h) → Mψ(g)+ψ(h)
for g, h ∈ G induced by the S-action of M . For u : M → N in GrModH (S)
[ψ]
there is a u[ψ] : M [ψ] → N [ψ] in GrModG (S [ψ] ) with ug = uψ(g) for g ∈ G.
This defines an exact functor •[ψ] : GrModH (S) → GrModG (S [ψ] ), called ψrefinement.
4
FRED ROHRER
C) For a G-graded ring R, composing
•[ψ] : GrModH (R[ψ] ) → GrModG ((R[ψ] )[ψ] )
with scalar restriction GrModG ((R[ψ] )[ψ] ) → GrModG (R) by means of αψ (R)
(1.2 A)) yields an exact functor GrModH (R[ψ] ) → GrModG (R), by abuse of
language again denoted by •[ψ] and called ψ-refinement.
(1.4) A) Let R be a G-graded ring. For a G-graded R-module M , the coL
[ψ]
product in Ab of the canonical injections Mg f ∈ψ−1 (ψ(g)) Mf = (M[ψ] )g
for g ∈ G is a monomorphism α′ψ (M ) : M (M[ψ] )[ψ] in GrModG (R), and
L
[ψ]
the coproduct in Ab of the restrictions (M[ψ] )g = f ∈ψ−1 (ψ(g)) Mf ։ Mg of
Q
the canonical projections f ∈ψ−1 (ψ(g)) Mf ։ Mg for g ∈ G is an epimorphism
δψ′ (M ) : (M[ψ] )[ψ] ։ M in GrModG (R). Varying M we get a monomorphism
α′ψ : IdGrModG (R) (•[ψ] )[ψ] and an epimorphism δψ′ : (•[ψ] )[ψ] ։ IdGrModG (R) .
B) For an H-graded R[ψ] -module M , the coproduct in Ab of the codiagonals
L
((M [ψ] )[ψ] )h = g∈ψ−1 (h) Mh ։ Mh for h ∈ H is an epimorphism
βψ′ (M ) : (M [ψ] )[ψ] ։ M
in GrModH (R[ψ] ). If Ker(ψ) is finite, so that ψ −1 (h) is finite for h ∈ H, then
the coproduct in Ab of the diagonals
Y
M
Mh
Mh =
Mg[ψ] = ((M [ψ] )[ψ] )h
g∈ψ−1 (h)
g∈ψ−1 (h)
for h ∈ H is a monomorphism γψ′ (M ) : M (M [ψ] )[ψ] in GrModG (R). Varying
M we get an epimorphism βψ′ : (•[ψ] )[ψ] ։ IdGrModH (R[ψ] ) and — if Ker(ψ) is
finite — a monomorphism γψ′ : IdGrModH (R[ψ] ) (•[ψ] )[ψ] .
(1.5) Examples A) If H = 0 then •[ψ] coincides with the forgetful functor
GrAnnG → Ann (for rings) or GrModG (R) → Mod(R[0] ) (for modules) that
forgets the graduation.
Z Z
B) Let ψ : /2 → 0 and let S be a ring. The underlying additive group of
is the group S ⊕ S, its components of degree 0 and 1 are S × 0 and 0 × S,
respectively, and its multiplication is given by (a, b)(c, d) = (ac + bd, ad + cb)
for a, b, c, d ∈ S.
S [ψ]
C) Let A be a ring and let R be the G-graded ring with R0 = A and Rg = 0
for g ∈ G\0. Then, R[ψ] is the H-graded ring with (R[ψ] )0 = A and (R[ψ] )h = 0
for h ∈ H \ 0, and GrModG (R) and GrModH (R[ψ] ) are canonically isomorphic
to the product categories Mod(A)G and Mod(A)H , respectively. Under these
isomorphisms, •[ψ] and •[ψ] correspond to functors Mod(A)G → Mod(A)H with
L
(Mg )g∈G 7→ ( g∈ψ−1 (h) Mg )h∈H and Mod(A)H → Mod(A)G with (Mh )h∈H 7→
(Mψ(g) )g∈G , respectively. Using this it is readily checked that for an H-graded
R[ψ] -module M it holds (M [ψ] )[ψ] = M ⊕ Ker(ψ) .
COARSENINGS, INJECTIVES AND HOM FUNCTORS
5
For modules, ψ-coarsening is left adjoint to ψ-refinement ([9, 3.1]), and for
H = 0 the same holds for rings ([11, 1.2.2]). We recall now the result for
modules and generalize the one for rings to arbitrary H.
(1.6) Proposition a) For a G-graded ring R there is an adjunction
•[ψ]
•[ψ]
GrModG (R) −−→ GrModH (R[ψ] ), GrModH (R[ψ] ) −−→ GrModG (R)
with unit α′ψ and counit βψ′ .
b) There is an adjunction
•[ψ]
•[ψ]
GrAnnG −−→ GrAnnH , GrAnnH −−→ GrAnnG
with unit αψ and counit βψ .
Proof. Straightforward.
In [9, 3.1] it was shown that if Ker(ψ) is finite then ψ-refinement for modules
is left adjoint to ψ-coarsening. For H = 0 this was sharpened by the result
that ψ-coarsening has a left adjoint if and only if G is finite ([3, 2.5]). We now
generalize this to arbitrary H and prove moreover the corresponding statement
for rings. We will need the following remark on products of graded rings and
modules.
(1.7) A) The category GrAnnG has products, but •[ψ] does not necessarily
Q
commute with them. The product R = i∈I R(i) of a family (R(i) )i∈I of Ggraded rings in GrAnnG is a G-graded ring as follows. Its components are
Q
L
Q
(i)
(i)
i∈I Rg for g ∈ G, so that its underlying additive group is
g∈G
i∈I Rg .
(i)
(i)
(i)
For i ∈ I, the multiplication of R(i) is given by maps Rg × Rh → Rg+h for
Q
Q
Q
(i)
(i)
(i)
g, h ∈ G, and their products i∈I Rg × i∈I Rh → i∈I Rg+h for g, h ∈ G
define the multiplication of R.
B) Let R be a G-graded ring. The category GrModG (R) has products, but
Q
•[ψ] does not necessarily commute with them. The product M = i∈I M (i)
of a family (M (i) )i∈I of G-graded R-modules in GrModG (R) is a G-graded
Q
(i)
R-module as follows. Its components are i∈I Mg for g ∈ G, so that its
L
Q
(i)
underlying additive group is
g∈G
i∈I Mg . For i ∈ I, the R-action of
(i)
(i)
M (i) is given by maps Rg × Mh → Mg+h for g, h ∈ G, and their products
Q
Q
(i)
(i)
Rg × i∈I Mh → i∈I Mg+h for g, h ∈ G define the R-action of M .
(1.8) Theorem a) If R is a G-graded ring, then
•[ψ] : GrModG (R) → GrModH (R[ψ] )
has a left adjoint if and only if Ker(ψ) is finite, and then
•[ψ]
•[ψ]
GrModH (R[ψ] ) −−→ GrModG (R), GrModG (R) −−→ GrModH (R[ψ] )
is an adjunction with unit γψ′ and counit δψ′ .
6
FRED ROHRER
b) •[ψ] : GrAnnG → GrAnnH has a left adjoint if and only if Ker(ψ) is finite,
and then
•[ψ]
•[ψ]
GrAnnH −−→ GrAnnG , GrAnnG −−→ GrAnnH
is an adjunction with unit γψ and counit δψ .
Proof. If Ker(ψ) is finite then γψ′ (for modules) and γψ (for rings) are defined
(1.4 B), 1.2 B)). In both cases it is straightforward to check that •[ψ] is left
adjoint to •[ψ] .
We prove now the converse statement for modules, analogously to [11, 2.5.3].
Suppose •[ψ] has a left adjoint and thus commutes with products ([7, 2.1.10]).
We consider the family of G-graded R-modules (M (g) )g∈Ker(ψ) with M (g) =
(g)
R(−g) for g ∈ G, so that eg = 1R ∈ Mg \ 0 for g ∈ G. For h ∈ Ker(ψ) we
Q
Q
(h)
(g)
denote by π (h) : g∈Ker(ψ) M (g) → M (h) and ρ(h) : g∈Ker(ψ) M[ψ] → M[ψ]
the canonical projections. There is a unique morphism ξ in GrModH (R[ψ] )
such that for h ∈ Ker(ψ) the diagram
Q
( g∈Ker(ψ) M (g) )[ψ]
❘❘❘
❘❘❘
❘❘❘
❘)
(π (h) )[ψ]
ξ
/
(h)
M[ψ]
Q
(g)
g∈Ker(ψ) M[ψ]
♠♠
♠♠♠
v♠♠♠ ρ(h)
in GrModH (R[ψ] ) commutes. This ξ is an isomorphism since •[ψ] commutes
with products. Taking components of degree 0 we get a commutative diagram
L
Q
ξh
∼
=
(g)
f ∈Ker(ψ) g∈Ker(ψ)
Mf
/
Q
L
g∈Ker(ψ) f ∈Ker(ψ)
❚❚❚❚
❚❚❚❚
*
(g)
Mf
❥❥❥❥
t ❥❥❥
❥
Q
(g)
Q
f ∈Ker(ψ) g∈Ker(ψ)
Mf
in Ab, where the unmarked morphisms are the canonical injections (1.7 B)).
(g)
(g)
For g ∈ Ker(ψ) we set xgg = eg ∈ Mg \ 0 and xgf = 0 ∈ Mf for f ∈
g
Ker(ψ) \ {g}. If g ∈ Ker(ψ) then {f ∈ Ker(ψ) | xf 6= 0} has a single element,
Q
L
(g)
so that ((xgf )f ∈Ker(ψ) )g∈Ker(ψ) ∈ g∈Ker(ψ) f ∈Ker(ψ) Mf . If f ∈ Ker(ψ) then
xff = ef 6= 0, hence (xgf )g∈Ker(ψ) 6= 0, implying
{f ∈ Ker(ψ) | (xgf )g∈Ker(ψ) 6= 0} = Ker(ψ).
As ξh is an isomorphism it follows
((xgf )g∈Ker(ψ) )f ∈Ker(ψ) ∈
M
Y
(g)
Mf .
f ∈Ker(ψ) g∈Ker(ψ)
(xgf )g∈Ker(ψ)
Thus, Ker(ψ) = {f ∈ Ker(ψ) |
6= 0} is finite.
Finally, the converse statement for rings is obtained analogously by considering the algebra K[G] of G over a field K, furnished with its canonical Ggraduation, and the family (R(g) )g∈Ker(ψ) of G-graded rings with R(g) = K[G]
COARSENINGS, INJECTIVES AND HOM FUNCTORS
7
for g ∈ G. Denoting by {eg | g ∈ G} the canonical basis of K[G] and consid(g)
ering the elements eg ∈ Rg \ 0 for g ∈ G we can proceed as above.
2. Application to injective modules
We keep the hypothesis of Section 1. The symbols •[ψ] and •[ψ] refer always to
coarsening functors for graded modules over appropriate graded rings.
In this section we apply the foregoing generalities to the question on how
injective graded modules behave under coarsening functors. A lot of work
on this question, but mainly in case H = 0, was done by Nǎstǎsescu et al.
(e.g. [3], [8], [9]).
(2.1) Proposition Let R be a G-graded ring and let M be a G-graded Rmodule. If M[ψ] is injective then so is M .
Proof. Analogously to [10, A.I.2.1].
The converse of 2.1 does not necessarily hold; see [10, A.I.2.6.1] for a counand H = 0. But in [9, 3.3] it was shown that the
terexample with G =
converse does hold if G is finite and H = 0. We generalize this to the case of
arbitrary G and H such that Ker(ψ) is finite, and we moreover show that this
is the best we can get without imposing conditions on R and M . Our proof
is inspired by [3, 3.14]. We first need some remarks on injectives and cogenerators, and a (probably folklore) variant of the graded Bass-Papp Theorem;
we include a proof for lack of reference.
Z
(2.2) A) A functor between abelian categories that has an exact left adjoint
preserves injective objects ([12, 3.2.7]).
B) In an abelian category C, a monomorphism with injective source is a
section, and a section with injective target has an injective source ([7, 8.4.4–
5]). If C fulfils AB4∗ then an object A is an injective cogenerator if and only
if every object is the source of a morphism with target AL for some set L ([7,
5.2.4]). This implies ([12, 3.2.6]) that if A is an injective cogenerator and L is
a nonempty set then AL is an injective cogenerator.
C) If R is a G-graded ring then GrModG (R) is abelian, fulfils AB5, and has
a generator. Hence, it has an injective cogenerator ([7, 9.6.3]).
D) Let R be a G-graded ring and let M be a G-graded R-module. Analogously to [1, X.1.8 Proposition 12] one sees that M is a cogenerator if and only
if every simple G-graded R-module is the source of a nonzero morphism with
target M . As M is simple if M[ψ] is so, it follows that if M is a cogenerator
then so is M[ψ] .
(2.3) Proposition A G-graded ring R is noetherian1 if and only if E ⊕N is
injective for every injective cogenerator E in GrModG (R).
1as a G-graded ring, i.e., ascending sequences of graded ideals are stationary
8
FRED ROHRER
Proof. Analogously to [2, 4.1] one shows that R is noetherian if and only if
countable sums of injective G-graded R-modules are injective. Thus, if R is
noetherian then E ⊕N is injective for every injective cogenerator E. Conversely,
suppose this condition to hold, let (Mi )i∈N be a countable family of injective
G-graded R-modules, and let E be an injective cogenerator in GrModG (R)
(2.2 C)). For i ∈
there exist a nonempty set Li and a section Mi E Li
Q
the canonical projection L ։ Li
(2.2 B)). Let L = i∈N Li . For i ∈
induces a monomorphism E Li E L . Composition yields a section Mi E L
L
we get a section j :
(2.2 B)). Taking the direct sum over i ∈
i∈N Mi
(E L )⊕N . Now, E L is an injective cogenerator (2.2 B)), so (E L )⊕N is injective
L
by hypothesis, and as j is a section thus so is i∈N Mi (2.2 B)). By the first
sentence of the proof this yields the claim.
N
N
N
(2.4) Theorem Ker(ψ) is finite if and only if
•[ψ] : GrModG (R) → GrModH (R[ψ] )
preserves injectivity for every G-graded ring R.
Proof. Finiteness of Ker(ψ) implies that •[ψ] preserves injectivity by 1.3 C),
1.8 a) and 2.2 A). For the converse we suppose that •[ψ] preserves injectivity
for every G-graded ring and assume that Ker(ψ) is infinite. Let A be a nonnoetherian ring and let R be the G-graded ring with R0 = A and Rg = 0 for
g ∈ G \ 0. Then, R[ψ] is the H-graded ring with (R[ψ] )0 = A and (R[ψ] )h ) =
0 for h ∈ H \ 0, and in particular non-noetherian. Let E be a injective
cogenerator in GrModH (R[ψ] ) (2.2 C)). It holds (E [ψ] )[ψ] = E ⊕ Ker(ψ) (1.5 C)),
and this H-graded R[ψ] -module is injective by 2.2 A), 1.6 a), 1.3 A) and the
hypothesis. Now, infinity of Ker(ψ), 2.2 B) and 2.3 yield the contradiction
that R[ψ] is noetherian.
(2.5) If Ker(ψ) is infinite and torsionfree we can construct more interesting
examples of G-graded rings R such that •[ψ] does not preserve injectivity
than in the proof of 2.4. Indeed, let A be the algebra of Ker(ψ) over a field,
furnished with its canonical Ker(ψ)-graduation. Let R be the G-graded ring
with Rg = Ag for g ∈ Ker(ψ) and Rg = 0 for g ∈ G \ Ker(ψ), so that R[ψ]
is the H-graded ring with (R[ψ] )0 = A[0] and (R[ψ] )h = 0 for h ∈ H \ 0.
The invertible elements of R are precisely its homogeneous elements different
from 0 ([4, 11.1]), so that the G-graded R-module R is injective. If g ∈
Ker(ψ) \ 0 then x = 1 + eg ∈ A (where eg denotes the canonical basis element
of A corresponding to g) is a nonhomogeneous non-zerodivisor of A[0] ([4,
8.1]), hence free and not invertible. So, there is a morphism of A[0] -modules
hxiA[0] → A[0] with x 7→ 1 that cannot be extended to A[0] , and thus the
H-graded R[ψ] -module R[ψ] is not injective.
The above result can be used to show that graded versions of covariant right
derived cohomological functors commute with coarsenings with finite kernel
(cf. [14]).
COARSENINGS, INJECTIVES AND HOM FUNCTORS
9
3. Application to Hom functors
We keep the hypotheses of Section 2. Let R be a G-graded ring and let M be
a G-graded R-module. If no confusion can arise we write Hom(•, ) instead
of HomGrModG (R) (•, ) for the Hom bifunctor with values in Ab.
As a second application of the generalities in Section 1 we investigate when
coarsening functors commute with covariant graded Hom functors. For H =
0 a complete answer was given by Gómez Pardo, Militaru and Nǎstǎsescu
([5], see also [6]). We generalize their result to arbitrary H, leading to the
astonishing observation that if a covariant graded Hom functor commutes
with some coarsening functor with infinite kernel then it commutes with every
coarsening functor.
As in [5], the notion of a small module turns out to be important. We start
by recalling it and then prove a generalization of [5, 3.1] on coarsening of small
modules and of steady rings.
(3.1) Let I be a set, let N = (Ni )i∈I be a family of G-graded R-modules
L
and let ιj : Nj
i∈I Ni denote the canonical injection for j ∈ I. The
monomorphisms Hom(M, ιj ) in Ab for j ∈ I induce a morphism λM
I (N) in Ab
such that the diagram
Q
L
/ Hom(M,
Hom(M, i∈I Ni )
i∈I Ni )
Hom(M,ι❤
j )❤❤❤4
O
❤❤
❤
λM
❤
❤
I (N)
❤
4❤
L
/
/
Hom(M, Nj )
i∈I Hom(M, Ni )
∼
=
/
Q
i∈I
O
Hom(M, Ni ),
where the unmarked monomorphisms are the canonical injections and the
unmarked isomorphism is the canonical one, commutes for j ∈ I. It follows
that λM
I (N) is a monomorphism. If N is constant with value N then we write
M
λI (N ) instead of λM
I (N). Varying N we get a monomorphism
M
M
λM
Hom(M, •i ) Hom(M,
•i )
I :
i∈I
i∈I
of covariant functors from GrModG (R)I to Ab. If I is finite then λM
I is an
isomorphism.
(3.2) A) If N is a G-graded R-module then M is called N -small if λM
I (N )
is an isomorphism for every set I. Furthermore, M is called small if λM
I is
an isomorphism for every set I, and this holds if and only if M is N -small for
every G-graded R-module N ([5, 1.1 i)]).
B) If M is of finite type then it is small. The G-graded ring R is called steady
if every small G-graded R-module is of finite type. Noetherian G-graded rings
are steady, but the converse does not necessarily hold ([5, 3.5], [13, 7◦ ; 10◦ ]).
Furthermore, for every group G there exists a G-graded ring that is not steady
([5, p. 3178]).
(3.3) Proposition a) M is small if and only if M[ψ] is small.
10
FRED ROHRER
b) If N is a G-graded R-module, then M is
M[ψ] is N[ψ] -small.
L
g∈G N (g)-small
if and only if
Proof. Immediately from 1.6 a), [5, 1.3 i)–ii)], and the facts that •[ψ] and •[ψ]
L
commute with direct sums and that g∈G N (g) = (N[ψ] )[ψ] .
(3.4) Proposition If R[ψ] is steady then so is R; the converse holds if
Ker(ψ) is finite.
Proof. If R[ψ] is steady and N is a small G-graded R-module then N[ψ] is small
(3.3 a)), hence of finite type, and thus N is of finite type, too. Conversely,
suppose Ker(ψ) is finite and R is steady, and let N be a small H-graded R[ψ] module. Since •[ψ] commutes with direct sums it follows that N [ψ] is small
(1.8 a), [5, 1.3 ii)]), hence of finite type, and thus (N [ψ] )[ψ] is of finite type,
too. The canonical epimorphism βψ′ (N ) : (N [ψ] )[ψ] ։ N (1.4 B)) shows now
that N is of finite type.
Next, we look at graded covariant Hom functors and characterize when they
commute with coarsening functors, thus generalizing [5, 3.4].
(3.5) The G-graded covariant Hom functor G HomR (M, •) maps a G-graded
R-module N onto the G-graded R-module
M
G
HomR (M, N ) =
HomGrModG (R) (M, N (g)).
g∈G
For a G-graded R-module N and g ∈ G we have a monomorphism
HomGrModG (R) (M, N (g)) HomGrModH (R[ψ] ) (M[ψ] , N[ψ] (ψ(g))), u 7→ u[ψ]
in Ab, inducing a monomorphism
hψ (M, N ) : G HomR (M, N )[ψ] H HomR[ψ] (M[ψ] , N[ψ] )
in GrModH (R[ψ]). Varying N we get a monomorphism
G
H
hM
ψ : HomR (M, •)[ψ] HomR[ψ] (M[ψ] , •[ψ] ).
(3.6) Lemma If N is a G-graded R-module such that hM
ψ (N ) is an isomorM
phism then λKer(ψ) ((N (g))g∈Ker(ψ) ) is an isomorphism.
Proof. Let u : M →
L
in GrModG (R). As
M
N (g))[ψ] =
N[ψ]
g∈G N (g)
(
M
g∈Ker(ψ)
g∈Ker(ψ)
H
we can consider the codiagonal
g∈Ker(ψ) N[ψ] → N[ψ] in GrMod (R[ψ] ) as
L
a morphism ∇ : ( g∈Ker(ψ) N (g))[ψ] → N[ψ] . Composition with u[ψ] yields
∇ ◦ u[ψ] ∈ H HomR[ψ] (M[ψ] , N[ψ] )0 . By our hypothesis there exist a finite subset
L
COARSENINGS, INJECTIVES AND HOM FUNCTORS
11
E ⊆ Ker(ψ) and for e ∈ E a ve : M → N (e) in GrModG (R) such that for x ∈ M
P
it holds ∇(u(x)) = re∈E ve (x). For g ∈ G this implies
X
X
∇(u(Mg )) ⊆
Ng+e =
N (e)g .
e∈E
e∈E
Let x = (xg )g∈G ∈ M with xg ∈ Mg for g ∈ G. For g ∈ G there exL
L
(g)
ists (nh )h∈Ker(ψ) ∈ ( h∈Ker(ψ) N (h))g =
h∈Ker(ψ) N (h)g with u(xg ) =
P
(g)
(g)
(nh )h∈Ker(ψ) , but it holds ∇(u(xg )) ∈ e∈E N (e)g and therefore nh = 0 for
L
L
h ∈ Ker(ψ)\E. This implies ∇(u(x)) ∈ e∈E N (e), thus u(M ) ⊆ e∈E N (e),
and hence the claim.
(3.7) Theorem hM
ψ is an isomorphism if and only if M is small or Ker(ψ)
is finite.
Proof. If Ker(ψ) is finite then this is readily seen to hold. We suppose Ker(ψ)
M
is infinite. If M is small then hM
0 is an isomorphism ([5, 3.4]), thus h[ψ] is
an isomorphism, too. Conversely, we suppose hM
[ψ] is an isomorphism and
prove that M is small. Let (Li )i∈N be a family of G-graded R-modules, let
L
L = i∈N Li , let li : Li L denote the canonical injection for i ∈ , and let
L
f : M → L in GrModG (R). Let N = g∈Ker(ψ) L(g), and let ng : L(g) N
denote the canonical injection for g ∈ Ker(ψ). It is readily checked that
N (g) ∼
= N for g ∈ Ker(ψ). As Ker(ψ) is infinite we can without loss of
⊆ Ker(ψ). Choosing for g ∈
an isomorphism N →
generally suppose
L
N (g) we get a monomorphism v : N ⊕N g∈Ker(ψ) N (g). Furthermore, we
L
L
⊕N , hence a morphism
get a monomorphism u = i∈N n0 ◦ li :
i∈N Li N
L
v ◦ u ◦ f : M → g∈Ker(ψ) N (g). By 3.6 there exists a finite subset E ⊆ Ker(ψ)
L
L
with v(u(f (M ))) ⊆ g∈E N (g) ⊆ g∈Ker(ψ) N (g). By construction of v there
′
with u(f (M )) ⊆ N ⊕E ⊆ N ⊕N . Thus, by
exists a finite subset E ′ ⊆
L
L
construction of u, we have f (M ) ⊆ i∈E ′ Li ⊆ i∈N Li = L. Therefore, M
is small.
N
N
N
N
At the end we get the surprising corollary mentioned before.
(3.8) Corollary If there exists an infinite subgroup F ⊆ G such that, denotM
ing by π : G ։ G/F the canonical projection, hM
π is an isomorphism, then hψ
is an isomorphism for every epimorphism ψ : G ։ H in Ab, and in particular
hM
0 is an isomorphism
Proof. Immediately from 3.7.
References
[1] N. Bourbaki, Éléments de mathématique. Algèbre. Chapitre 10. Masson, Paris, 1980.
[2] S. U. Chase, Direct products of modules. Trans. Amer. Math. Soc. 97 (1960), 457–473.
[3] S. Dǎscǎlescu, C. Nǎstǎsescu, A. Del Rio, F. Van Oystaeyen, Gradings of finite support.
Application to injective objects. J. Pure Appl. Algebra 107 (1996), 193–206.
[4] R. Gilmer, Commutative semigroup rings. Chicago Lectures Math., Univ. Chicago Press,
Chicago, 1984.
12
FRED ROHRER
[5] J. L. Gómez Pardo, G. Militaru, C. Nǎstǎsescu, When is HOMR (M, −) equal to
HomR (M, −) in the category R-gr? Comm. Algebra 22 (1994), 3171–3181.
[6] J. L. Gómez Pardo, C. Nǎstǎsescu, Topological aspects of graded rings. Comm. Algebra
21 (1993), 4481–4493.
[7] M. Kashiwara, P. Schapira, Categories and sheaves. Grundlehren Math. Wiss. 332,
Springer, Berlin, 2006.
[8] C. Nǎstǎsescu, Some constructions over graded rings: applications. J. Algebra 120
(1989), 119–138.
[9] C. Nǎstǎsescu, S. Raianu, F. Van Oystaeyen, Modules graded by G-sets. Math. Z. 203
(1990), 605–627.
[10] C. Nǎstǎsescu, F. Van Oystaeyen, Graded ring theory. North-Holland Math. Library 28,
North-Holland, Amsterdam, 1982.
[11] C. Nǎstǎsescu, F. Van Oystaeyen, Methods of graded rings. Lecture Notes in Math.
1836, Springer, Berlin, 2004.
[12] N. Popescu, Abelian categories with applications to rings and modules. London Math.
Soc. Monogr. Ser. 3, Academic Press, London, 1973.
[13] R. Rentschler, Sur les modules M tels que Hom(M, −) commute avec les sommes directes. C. R. Acad. Sci. Paris Sér. A-B 268 (1969), 930–933.
[14] F. Rohrer, Coarsening of graded local cohomology. To appear in Comm. Algebra.
Universität Tübingen, Fachbereich Mathematik, Auf der Morgenstelle 10,
72076 Tübingen, Germany
E-mail address: [email protected]
| 0math.AC
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.